paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1611.08751 | 1 | 1611 | 2016-11-26T22:24:29 | Functional Alignment with Anatomical Networks is Associated with Cognitive Flexibility | [
"q-bio.NC"
] | Cognitive flexibility describes the human ability to switch between modes of mental function to achieve goals. Mental switching is accompanied by transient changes in brain activity, which must occur atop an anatomical architecture that bridges disparate cortical and subcortical regions by underlying white matter tracts. However, an integrated perspective regarding how white matter networks might constrain brain dynamics during cognitive processes requiring flexibility has remained elusive. To address this challenge, we applied emerging tools from graph signal processing to decompose BOLD signals based on diffusion imaging tractography in 28 individuals performing a perceptual task that probed cognitive flexibility. We found that the alignment between functional signals and the architecture of the underlying white matter network was associated with greater cognitive flexibility across subjects. Signals with behaviorally-relevant alignment were concentrated in the basal ganglia and anterior cingulate cortex, consistent with cortico-striatal mechanisms of cognitive flexibility. Importantly, these findings are not accessible to unimodal analyses of functional or anatomical neuroimaging alone. Instead, by taking a generalizable and concise reduction of multimodal neuroimaging data, we uncover an integrated structure-function driver of human behavior. | q-bio.NC | q-bio | Functional Alignment with
Anatomical Networks is Associated with
Cognitive Flexibility
John D. Medaglia1, Weiyu Huang2, Elisabeth A. Karuza1,
Sharon L. Thompson-Schill1, Alejandro Ribeiro2, and Danielle S. Bassett2,3
1Department of Psychology, University of Pennsylvania, Philadelphia, PA, 19104 USA
2Department of Electrical & Systems Engineering, University of Pennsylvania, Philadelphia, PA,
3Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, 19104 USA
19104 USA
6
1
0
2
v
o
N
6
2
]
.
C
N
o
i
b
-
q
[
1
v
1
5
7
8
0
.
1
1
6
1
:
v
i
X
r
a
1
Article summary
Human cognitive flexibility emerges from complex dynamics in anatomical networks. How-
ever, no concise measure integrating anatomical and functional organization in support of
cognitive flexibility has been identified. We examine cognitive "switch costs" – a measure
of cognitive flexibility – in 28 healthy individuals during functional neuroimaging. We con-
struct individualized anatomical networks from diffusion spectrum imaging data, allowing us
to reconstruct the white matter networks for each subject. We apply a novel approach from
graph signal processing to distill functional brain signals into aligned and liberal components
relative to underlying anatomy. We show that relatively alignment of brain signals with
underlying anatomical network organization supports faster cognitive switching. Moreover,
we show that the signals most aligned with the network from moment to moment are also the
most flexible. This indicates that anatomical networks provide organization to the flexible
brain and facilitate greater mental agility and establishes a potential neural biomarker that
integrates brain structure and function.
2
Abstract
Cognitive flexibility describes the human ability to switch between modes of mental function
to achieve goals. Mental switching is accompanied by transient changes in brain activ-
ity, which must occur atop an anatomical architecture that bridges disparate cortical and
subcortical regions by underlying white matter tracts. However, an integrated perspec-
tive regarding how white matter networks might constrain brain dynamics during cognitive
processes requiring flexibility has remained elusive. To address this challenge, we applied
emerging tools from graph signal processing to decompose BOLD signals based on diffusion
imaging tractography in 28 individuals performing a perceptual task that probed cognitive
flexibility. We found that the alignment between functional signals and the architecture of
the underlying white matter network was associated with greater cognitive flexibility across
subjects. Signals with behaviorally-relevant alignment were concentrated in the basal gan-
glia and anterior cingulate cortex, consistent with cortico-striatal mechanisms of cognitive
flexibility. Importantly, these findings are not accessible to unimodal analyses of functional
or anatomical neuroimaging alone. Instead, by taking a generalizable and concise reduction
of multimodal neuroimaging data, we uncover an integrated structure-function driver of hu-
man behavior.
3
Introduction
Cognitive flexibility is involved in virtually every complex behavior from mental arithmetic
to processing visual stimuli. For example, when navigating complex environments, humans
can flexibly switch between two foci of attention or between two processing modalities, in
order to effectively respond to sensory inputs. While a hallmark of human cognition, flexible
switching is also associated with a measurable cost: moving from one task to another induces
a natural extension in the time it takes a person to respond [1]. In patients with neurological
syndromes, this cost is even greater, to the point where it can hamper a patient's ability to
engage in the basic activities of daily living [2], impacting long-term cognitive outcomes [3].
In healthy individuals, cognitive flexibility varies considerably, and individual differences in
this trait contribute to mental facets ranging from the development of reasoning ability [4]
to quality of life into late age [5]. Clarifying the nature of cognitive flexibility in the human
brain is critical to understand the human mind.
The physiological origins of cognitive flexibility are thought to lie in corticobasal ganglia-
thalamo-cortical loops [6]: regions of the fronto-parietal and cingulo-opercular systems are
activated by cognitive switching tasks [7, 8, 9, 10]. In switching paradigms, the anterior cin-
gulate is thought to contribute negative feedback detection following switches, whereas the
lateral prefrontal cortex maintains rules and inhibits incorrect responses. Both of these re-
gions anatomically connect to subcortical regions, which are postulated to mediate processes
that both suppress prepotent motor responses and transition between behavioral outputs to
meet task goals [11]. Interactions between cortical systems and motor outputs are thought
to be anatomically mediated by subcortical circuits [7, 12, 13, 14, 9]. Yet, understanding ex-
actly how this circuit supports task switching has remained challenging, particularly because
it requires a conceptual integration of regional activity, inter-regional anatomical connectiv-
ity, and observable measures of behavior. While regional activity and behavioral markers
of cognitive flexibility are relatively straightforward to estimate, it is less straightforward to
integrate these features with the white matter structure (the connectome [15]) that guides
the propagation of functional signals [16, 17, 18].
As a result, these two research enterprises – cognitive neuroscience and connectomics –
have largely developed in parallel without significant cross-talk. Ideally, frameworks that
include a concise correspondence between brain network structure, function, and cognitive
measures have the potential to produce more comprehensive understanding of human cog-
nition [19, 20]. Conceptually, underlying white matter network organization in the brain
4
physically mediates communication among brain regions. However, analytic frameworks
that explicitly use white matter structure to constrain cognitively relevant functional signals
are lacking. Such approaches may allow investigators to adjudicate the relative contributions
of well-described systems in the brain [21, 8] to specific cognitive variables by integrating
neurophysiological dynamics and anatomy.
To address this challenge, we aim to identify the multimodal integration of network
anatomy and functional signals that supports cognitive switching. We introduce a novel
approach that allows us to examine the distinct contributions of functional signals in the
context of anatomically linked regions in human brain networks. In a cohort of 28 healthy
adult human subjects, we collected high-resolution diffusion spectrum imaging (DSI) data
as well as BOLD fMRI data acquired during the performance of a novel cognitive switching
paradigm built on a set of shapes that could be perceived as composed of different features at
the local versus global scales [22] (see Fig. 1). From the DSI data, we constructed anatom-
ical brain networks in which 111 cortical, subcortical, and cerebellar regions [23, 24] were
linked pairwise by the density of streamlines reconstructed by state-of-the-art tractography
algorithms. Next, we used the eigenspectrum of these anatomical networks to measure the
relative separation of framewise regional BOLD signals from the underlying white matter
(see Fig. 2 and Methods). Specifically, each regional signal was decomposed into a portion
that aligned tightly with the anatomical network ("aligned") and a portion that did not align
tightly with the network ("liberal"). Here, alignment and liberality refer to signal deviations
from the underlying anatomical network. We examine the observed BOLD activity for each
measurement in time to identify where and to what extent BOLD signals across the brain are
organized by white matter networks. Conceptually, this technique allows us to identify to
what degree individual BOLD signals deviate weakly versus strongly from underlying white
matter anatomy.
We anticipate that functional alignment with anatomical networks is an individually vari-
able feature that facilitates cognitive flexibility. We hypothesize that moment-to-moment
alignment in human brain networks facilitates switching performance, indicating interindi-
vidual variability in the degree of organization of information processing by anatomical
topology. This allows us to investigate a theoretical dynamic property supporting cognitive
flexibility and provides a mapping to neuroanatomical theories of cognitive switching. In the
current study, our task involves proactive switching following detection of a contextual cue
[11]. Whereas prior literature has focused on region-specific mechanisms associated with this
5
process, the current approach allows us to examine the role of local neural processes across
the brain's distributed anatomical network. We postulate that functional alignment serves
as a mechanism for cognitive flexibility across the brain's complex anatomical white matter
network. In particular, well-aligned signals may represent efficiently organized signals with
markedly less interference, whereas liberal signals do not leverage the inherent organization
of underlying white matter networks. This approach to multimodal neuroimaging could dis-
tinguish whether structure and function operate synergistically or in opposition to promote
cognitive flexibility. Accordingly, we examine whether alignment with anatomically medi-
ated expectations in subcortical regions, cortical systems, or both constitutes the basis for
effective cognitive switching. This allows us to determine (i) whether alignment forms a basis
for cognitive flexibility and (ii) whether regions such as those in the subcortical system that
serve an anatomically central network role also serve a greater or less functionally central
role in this mental skill.
Figure 1: Cognitive task requiring perceptual switching. (A) Example stimuli based on Navon
local-global features. Subjects were trained to respond to the larger (or "global") shape if the stimulus
was green and to the smaller (or "local") shapes if it was white. (B) An example of the non-switching
condition for responses. Subjects viewed a sequence of images and were instructed to respond as
quickly and accurately as possible.
(C) An example of the switching condition between stimuli
requiring global and local responses. Here, trials with a red exclamation point are switches from the
previous stimulus.
6
Figure 2: Multimodal approach to the study of cognitive switching using emerging graph
signal processing tools. (A) A notion of signal independence on a schematic modular network.
Left: An aligned signal on top of a given graph is one in which the magnitude of functional signals,
represented by the directionality of the colored cones, corresponds tightly to that expected by the
network's organization. In this toy example, one cluster of nodes contains similar positive signals,
and the other cluster contains similar negative signals. Right: A liberal signal on top of a given
graph is one in which signals diverge significantly from the underlying network. (B) We construct
a white matter graph (adjacency matrix) from 111 anatomically-defined regions where connections
are the streamline density between region pairs. (C) From BOLD fMRI data acquired during the
performance of the Navon task, we extract regional mean time series which we treat as graph signals.
(D) For each subject, we decompose the graph signals into aligned and liberal components using the
underlying eigenspectrum of the white matter graph. We spatially map aligned and liberal signals,
and we determine their relationship to switch costs estimated from behavioral performance on the
task. Cb = cerebellum. See Methods for details.
Functional alignment in liberal subcortical systems is
associated with faster cognitive switching
Aligned signals were concentrated within default mode, fronto-parietal, cingulo-opercular,
and subcortical systems across subjects, whereas the liberal signals were concentrated largely
in the subcortical system (Fig. 4). The significance of these concentrations was confirmed
statistically using a non-parametric permutation test (α = 0.05) in which we shuffled the
7
values of alignment (or liberality) uniformly at random across brain regions before computing
the mean alignment (or liberality) value within each system [25]. Then, we calculated the
correlation between aligned and liberal BOLD signals across the brain and cognitive switch
costs (response times during trials with a color-cued switch versus non-switching trials).
We observed that variability in aligned signals was not associated with switch costs (R =
0.15, p = 0.43, accounting for 2% of the variance), while variability in liberal signals accounted
for 32% of variance (see Fig. 3). Among the liberal signals, lower values of liberality (that
is, relative alignment) were also associated with lower switch costs both during fixation
(R = 0.62, p = 0.0006) and during nonswitching (R = 0.71, p = 0.0001) perceptual blocks.
Figure 3: Lower independence is associated with lower switch costs. (A) Liberal signals
were concentrated especially in subcortical regions and cingulate cortices. (B) Reduced liberality (in-
creased alignment) is associated with reduced switch costs across subjects. (C) Aligned signals were
concentrated especially in subcortical, default mode, fronto-parietal, and cingulo-opercular systems.
(D) Variability in aligned signals was not significantly associated with switch costs across subjects.
L = left hemisphere, R = right hemisphere.
It is interesting to note that signals extracted from subcortical areas can be decomposed
into both highly anatomically-aligned portions and highly liberal portions (see Methods).
The values of both alignment and liberality were significantly greater than expected in sub-
cortical structures. Yet, other systems show a disposition toward high alignment (fronto-
parietal, cingulo-opercular, and default mode systems), or no disposition toward alignment
or liberality (visual, auditory, and dorsal & ventral attention). We quantified the relationship
between aligned and liberal signals by calculating the Pearson correlation coefficient between
alignment and liberality values over all brain regions. We observed that the two types of
8
Figure 4: Non-parametric permutation test for signal concentration within cognitive
systems. (A) Liberal signal concentrations sorted from highest ( top) to lowest concentration across
all regions in all subjects. (B) Liberal signals were most concentrated in subcortical regions. (C)
Aligned signal concentrations sorted from highest ( top) to lowest concentration across all regions
in all subjects. (D) Aligned signals were most concentrated in fronto-parietal, cingulo-opercular,
default mode, and subcortical systems.
9
signals were significantly but not perfectly correlated across the brain: mean over subjects
and tasks R = 0.69, p (cid:28) 0.001 (see also Supplement). This analysis highlights an interesting
measure of multimodal complexity that differs across brain areas: subcortical regions can
be thought of as displaying high multimodal complexity because they contain both highly
anatomically-aligned signals and highly liberal signals.
Anatomy and function together define a neural trait un-
derlying cognitive flexibility
Finally, we asked whether the observed regional variation in aligned and liberal signals
could be estimated reliably across subjects and tasks, supporting its utility as a marker for
behavior. To address this question, we compared the regional patterns of alignment between
subjects and between task conditions (fixation, non-switching, and switching). We observed
that the regional pattern of signal alignment was consistent across the three task conditions
for a given subject (intraclass correlation: R = 0.99, p (cid:28) 0.001), as was the regional pattern
of liberal signals (intraclass correlation: R = 0.99, p (cid:28) 0.001; see also Supplement). These
results indicate that regional patterns of signal alignment and liberality are highly stable
phenotypes of a subject across the three task conditions.
Taken together, our results demonstrate that individuals whose most liberal functional
signals were more aligned with white matter architecture could switch perceptual focus faster.
In other words, relative alignment with anatomy was associated with greater cognitive flexi-
bility, a finding that highlights the importance of simultaneously considering both functional
and anatomical neuroimaging in the study of higher order cognitive processes. These findings
complement prior studies of executive function that have focused on node-level, edge-level,
and module-level features of brain networks [26, 13] (see also Supplement). Here, we consider
brain function as a series of time-evolving states [25, 27] that are organized in relation to
the underlying pattern of white matter tracts. The state-based focus of our approach also
offers insights into the differential extent to which specific cognitive systems deviate from
tract anatomy, underscoring anatomical contributions to the organization of brain dynamics
across subjects.
Our results contextualize previous models of cortico-striatal cognitive switching mecha-
nisms [11, 28] within a connectomic perspective. As a complement to prior findings implicat-
ing individual prefrontal and striatal systems in cognitive switching, our results highlight the
10
importance of anatomical network organization, and the central role of subcortical functional
dynamics atop that structure. Specifically, we found that modest functional alignment of
the most liberal signals in the subcortical and cingulate cortices was associated with faster
cognitive switching. This observation is particularly interesting in the context of prior work
showing that subcortical and anterior cingulate regions manage multiple inputs and outputs
among cortical systems during task transitions [29, 9], potentially requiring more diverse
signal organization relative to anatomical networks. Interestingly, we observe that the levels
of alignment and liberality were consistent within subjects across all three conditions of the
cognitive task – fixation, non-switching, and switching blocks – suggesting that it may be
a trait rather than a state of an individual. While prior studies have similarly identified
consistent network architecture across fixation and task [30], these similarities have not pre-
viously been identified in a model-based integration of functional dynamics and white matter
anatomy. The observed inter-task consistency of functional signal alignment and liberality
was complemented by broad inter-subject variability. We speculate that this variability rep-
resents differences in the amount of energy utilized to enact perceptual switches [31], with
greater independence requiring greater energy. Thus, future studies could examine the po-
tentially subject-specific role of cognitive control systems in regulating energy-efficient state
transitions in local-global processing.
With respect to recent dynamic network analyses of executive function, this analysis
contributes a crucial anatomically-grounded perspective. The current approach represents
a framework in which to understand the dual features of anatomical organization and func-
tional processes supporting cognitive flexibility in the human connectome. Here, high func-
tional dependence in fronto-parietal, cingulo-opercular, default mode, and subcortical sys-
tems is not associated with intersubject switching variability. Critically, our results indicate
that regions that participate in highly flexible systems [32] in temporal network analysis
demonstrate high dependence on underlying anatomical networks across frames of BOLD
data during fixation, low cognitive control conditions, and task conditions (see Supplement
for further analysis and discussion). Previous studies identify dynamic network roles for
fronto-parietal and cingulo-opercular regions in cognitive switching, and our results indicate
that moment to moment signal configurations in highly flexible systems are strongly orga-
nized by structure across time (see Supplement for addition analysis and discussion). In the
context of this highly organized cortical activity, the current results suggest that subcortical
systems contain highly liberal signals. The extent to which subcortical systems exhibit rela-
11
tive alignment may form a flexible integrative role across the many computations supported
by cortical systems. The relationship between anatomically-bound momentary signal orga-
nization and functional reconfigurations in temporal networks may more generally provide
a fruitful area for future research.
Notably, our results do not explain the potential cognitive role of highly aligned signals.
It is possible that the role of these signals may be better explained in the context of other
cognitive control processes [33]. We speculate that anatomically aligned signals in fronto-
parietal, cingulo-opercular, default mode, and subcortical systems organize the dynamic
signals contributing to cortical mechanisms of cognitive control, attention, and resting and
preparedness processes, respectively. It would be interesting to test whether highly aligned
signals in association cortices and subcortical structures are associated with domain-general
performance variability across modalities [34]. In cognitive switching specifically, the extent
to which signal liberality relates to performance on tasks involving other sensory modalities,
transitioning between internally and externally focused attention, and divergent thinking
remains to be established.
In conclusion, our study supports the utility of network science in clarifying mechanisms
of executive function specifically and cognition more generally [19, 20]. Recent literature
firmly establishes that white matter organization is a critical, but incomplete determinant of
functional signals in brain networks [17, 35]. Conceptually, the current approach acknowl-
edges that without structure, functional signals lack a mediating organization. By examining
functional signal alignment within underlying white matter networks, we identify an impor-
tant definition of dynamic contributions to cognitive switching that powerfully discriminates
between the contributions of subcortical and other systems in the brain. Similar applications
to other large multimodal neuroimaging datasets could contribute to biomarker analyses in
psychiatric disease and neurological disorders, many of which are associated with deficits in
executive function [36, 37, 38].
12
Methods
Subjects
A total of 30 subjects were recruited. All subjects were screened for prior history of psychi-
atric or neurological illness. One subject was excluded due to near-chance performance on
the task (accuracy = 52%). One additional subject was excluded due to technical problems
on the day of scanning. The final sample included 28 individuals (mean age = 25.6, St.D. =
3.5, 70% caucasian, 13 females). All subjects volunteered with informed consent in writing
in accordance with the Institutional Review Board/Human Subjects Committee, University
of Pennsylvania.
Behavioral task
All participants completed a local-global perception task based on classical Navon figures
[22]. Local-global stimuli were comprised of four shapes – a circle, X, triangle, or square –
that were used to build the global and local aspects of the stimuli. On all trials, the local
feature did not match the global feature, ensuring that subjects could not use information
about one scale to infer information about another. Stimuli were presented on a black
background in a block design with three block types (See Fig. 2). In the first block type,
subjects viewed white local-global stimuli. In the second block type, subjects viewed green
local-global stimuli. In the third block type, stimuli switched between white and green across
trials uniformly at random with the constraint that 70% of trials included a switch in each
block. In all blocks, subjects were instructed to report only the local features of the stimuli
if the stimulus was white and to report only the global feature of the stimuli if the stimulus
was green. Blocks were administered in a random order. Subjects responded using their
right hand with a four-button box. All subjects were trained on the task outside the scanner
until proficient at reporting responses using a fixed mapping between the shape and button
presses (i.e., index finger = "circle", middle finger = "X", ring finger = "triangle", pinky
finger = "square"). In the scanner, blocks were administered with 20 trials apiece separated
by 20 s fixation periods with a white crosshair at the center of the screen. Each trial was
presented for a fixed duration of 1900 ms separated by an interstimulus interval of 100 ms
during which a black screen was presented.
13
Diffusion spectrum imaging acquisition and processing
Diffusion spectrum images (DSI) were acquired on a Siemens 3.0T Tim Trio for all subjects
along with a T1-weighted anatomical scan at each scanning session. We followed a parallel
strategy for data acquisition and construction of streamline adjacency matrices as in previous
work [25, 39]. DSI scans sampled 257 directions using a Q5 half-shell acquisition scheme
with a maximum b-value of 5,000 and an isotropic voxel size of 2.4 mm. We utilized an axial
acquisition with the following parameters: repetition time (TR) = 5 s, echo time (TE)=
138 ms, 52 slices, field of view (FoV) (231, 231, 125 mm). We acquired a three-dimensional
SPGR T1 volume (TE = minimal full; flip angle = 15 degrees; FOV = 24 cm) for anatomical
reconstruction. All subjects volunteered with informed consent in writing in accordance with
the Institutional Review Board/Human Subjects Committee, University of Pennsylvania.
DSI data were reconstructed in DSI Studio (www.dsi-studio.labsolver.org) using q-space
diffeomorphic reconstruction (QSDR)[40]. QSDR first reconstructs diffusion-weighted im-
ages in native space and computes the quantitative anisotropy (QA) in each voxel. These
QA values are used to warp the brain to a template QA volume in Montreal Neurological
Institute (MNI) space using the statistical parametric mapping (SPM) nonlinear registration
algorithm. Once in MNI space, spin density functions were again reconstructed with a mean
diffusion distance of 1.25 mm using three fiber orientations per voxel. Fiber tracking was
performed in DSI studio with an angular cutoff of 35◦, step size of 1.0 mm, minimum length
of 10 mm, spin density function smoothing of 0.0, maximum length of 400 mm and a QA
threshold determined by DWI signal in the colony-stimulating factor. Deterministic fiber
tracking using a modified FACT algorithm was performed until 1,000,000 streamlines were
reconstructed for each individual.
Anatomical scans were segmented using FreeSurfer[41] and parcellated using the connec-
tome mapping toolkit [42]. A parcellation scheme including n = 129 regions was registered
to the B0 volume from each subject's DSI data. The B0 to MNI voxel mapping produced
via QSDR was used to map region labels from native space to MNI coordinates. To extend
region labels through the grey-white matter interface, the atlas was dilated by 4 mm [43].
Dilation was accomplished by filling non-labelled voxels with the statistical mode of their
neighbors' labels.
In the event of a tie, one of the modes was arbitrarily selected. Each
streamline was labeled according to its terminal region pair.
Finally, we included a cerebellar parcellation [24]. We used FSL to nonlinearly register
the individual's T1 to MNI space. Then, we used the inverse warp parameters to warp the
14
cerebellum atlas to the individual T1. We registered the subject's DSI image to the T1.
We used the inverse parameters from this registration to map the individualized cerebellar
parcels into the subject's DSI space. Finally, we merged the cerebellar label image with the
dilated cortical and subcortical parcellation image.
From these data and parcellation, we constructed an anatomical connectivity matrix, A
whose element Aij represented the number of streamlines connecting different regions [17],
divided by the sum of volumes for regions i and j [44]. Prior to data analysis, all cerebellum-
to-cerebellum edges were removed from each individual's matrix because cerebellar lobules
are demonstrably not anatomically connected directly to one another [45].
Functional imaging acquisition and processing
fMRI images were acquired during the same scanning session as the DSI data on a 3.0T
Siemens Tim Trio whole-body scanner with a whole-head elliptical coil by means of a single-
shot gradient-echo T2* (TR = 1500 ms; TE = 30 ms; flip angle = 60◦; FOV = 19.2 cm,
resolution 3mm x 3mm x 3mm). Preprocessing was performed using FEAT v. 6.0 (fMRI
Expert Analysis Tool) a component of the FSL software package [46]. To prepare the func-
tional images for analyses, we completed the following steps: skull-stripping with BET to
remove non-brain material, motion correction with MCFLIRT (FMRIB's Linear Image Reg-
istration Tool; [46]), slice timing correction (interleaved), spatial smoothing with a 6-mm 3D
Gaussian kernel, and high pass temporal filtering to reduce low frequency artifacts. We also
performed EPI unwarping with fieldmaps in order to improve subject registration to stan-
dard space. Native image transformation to a standard template was completed using FSL's
affine registration tool, FLIRT [46]. Subject-specific functional images were co-registered to
their corresponding high-resolution anatomical images via a Boundary Based Registration
technique (BBR [47]) and were then registered to the standard MNI-152 structural template
via a 12-parameter linear transformation. Finally, each participant's individual anatomical
image was segmented into grey matter, white matter, and CSF using the binary segmentation
function of FAST v. 4.0 (FMRIB's Automated Segmentation Tool [48]). The white matter
and CSF masks for each participant were then transformed to native functional space and
the average timeseries were extracted. Images were spatially smoothed using a kernel with a
full-width at half-maximum of 6 mm. These values were used as confound regressors on our
time series along with 18 translation and rotation parameters as estimated by MCFLIRT
[49].
15
Functional decomposition into anatomical networks
We analyze the signal defined on a connected, weighted, and symmetric graph, G = (V, A),
where V = {1, . . . , n} is a set of n vertices or nodes representing individual brain regions and
A ∈ Rn×n is defined as above. Because the network A is symmetric, it has a complete set of
orthonormal eigenvectors associated with it [50, 51]. For this reason, it has an eigenvector
decomposition, A = VΛVT , in which Λ is the set of eigenvalues, ordered so that λ0 ≤ λ1 ≤
. . . ≤ λn−1, and V = {vk}n−1
k=0 is the set of associated eigenvectors. Following [51, 52], we
use the eigenvector matrix to define the Graph Fourier Transform (GFT) of the graph signal
x ∈ Rn, defined as
x = VT x.
(1)
Given x = [x0, . . . , xn−1]T , we can express our original signal as x =(cid:80)n−1
k=0 xkvk, a sum of
the eigenvector components vk. The contribution of vk to the signal x is the GFT compo-
nent xk. Note that the smoothness of vk on the network can be evaluated in the quadratic
k Avk = (cid:80)
form vT
composition. The quantity (cid:80)
i,j∈V Aijvk(i)vk(j) and that vT
k Avk = λk is given by the eigenvector de-
i,j∈V Aijvk(i)vk(j) will be negative when the signal is varied
(highly connected regions possess signals of different signs), and positive when the signal is
smooth (highly connected regions possess signals of same signs); for these reasons, this quan-
tity can be thought of as a measure of smoothness (alignment). Consequently, these GFT
coefficients xk for small values of k indicate how much variables that are highly misaligned
(liberality) with structure contribute to the observed brain signal x. GFT coefficients xk
for large values of k describe how much signals that are aligned with the anatomical network
contribute to the observed brain signal x. The inverse (i)GFT of x with respect to A is
defined as
x = Vx.
(2)
Given a graph signal x with GFT x, we can isolate the liberal components corresponding
to the lowest KL eigenvectors by applying a graph filter HL that only keeps components
with k < KL and sets other components to 0. The signal xL then contains the "liberal"
components of x (those with a low alignment with network structure). Apart from the graph
low-pass filter HL, we also consider a middle graph regime HM, which keeps only components
16
in the range of KL ≤ k < n−KA, and an "aligned" graph regime HA, such that only network-
aligned components with n− KA ≤ k are kept. Therefore, the liberal regime takes the lowest
KL components, the alignment regime takes the highest KA components, and the middle
regime captures the middle n − KL − KM components (here, we use the components with
the 10 lowest alignment values to represent the liberal regime and the components with the
10 highest alignment values to represent the aligned regime; see Supplement for robustness
of results to parameter selection). As such, if we use xM and xA to respectively denote the
signals represented by the middle and highly aligned regimes, we have that the original signal
can be written as the sum x = xL + xM + xA. This formulation gives a decomposition of
the original signals x into liberal, moderately aligned, and highly aligned components that
respectively represent signals that have high, medium, and low signal deviation with respect
to the anatomical connectivity between brain regions.
Prior work has consistently demonstrated that the aligned and liberal components aid in
better estimation of unknown movie ratings in recommendation systems [53], better predic-
tion of cancer using gene interaction networks [54, 55], and learning in neuroimaging data,
where learning-related processes are demonstrably expressed in low and high components in
fMRI data, and where the middle component xM is demonstrably less reliable and behav-
iorally uninformative [56]. In the supplement, we perform a similar analysis to [56] but with
the current data to examine the stability of our low and high alignment measurements to
parameter selection. The data indicates that the low and high alignment components in the
current data are stable. Mathematically, this is expected in general in applications of the
current approach because eigenvalues at the extreme low and high end are isolated from the
middle values, which leads to robustness in the high and low ranges of the decomposition
[57].
We note that this approach allows signals on the anatomical network to contain both
aligned and liberal components represented in the same region at a single TR. This flexibility
occurs because the anatomical network of n nodes has n2 entries (i.e., connection information
is encoded in the anatomical adjacency matrix for any node i to any node j). Rather
than examining a single BOLD signal measurement as n independently observed values, the
GFT considers the signal to be a composite of contributions to the signal across subject's
anatomical network topology. The decomposition occurs across the entire set of signals
(here, the vector of BOLD magnitudes across regions at a single TR), where there are only n
entries. The GFT applied here leverages the fact that the n entries in a given vector are not
17
Figure 5: Signal decomposition into anatomy. BOLD signals are decomposed into aligned
and liberal signal components. Left of equation: a schematic BOLD signal on a simple anatomical
network. Here, two signals are stronger in the high direction than the low direction. Right of
equation: the signals across the network are decomposed into an aligned and liberal component.
The original signals can be reconstructed from a basis set including a weighted part of the signal
that is aligned with the anatomical network and another part that is liberal with respect to the
anatomical network.
isolated, but are signals on top of the complex anatomical network. In the current approach,
instead of focusing on the single BOLD value observed at each region as a discrete entity, the
decomposition is sensitive to the observation of pairwise differences among BOLD signals
relative to that expected by the anatomical network. Some portion of each given region's
BOLD signal is estimated to be liberal with respect to the network, which is represented
by xL, and some portion is estimated to be aligned with the network, which is represented
by xA (See Fig. 5). This mathematical separation establishes the notions of alignment and
liberality of the BOLD signals in the anatomical network. All individual regions in the brain
will have some degree of alignment and some degree of liberality given the complexity of
BOLD signal patterns across the network, unless the observed BOLD signals in all regions
are perfectly aligned or perfectly misaligned with the subject's anatomical network. This
highlights an important strength of the use of the Graph Fourier Transform to examine
functional signal liberality in anatomical brain networks: the signal can be understood as
a network level composite of aligned and liberal signals, and the extent to which individual
regions contribute to these properties can be examined as the variation in the weights of the
region's contribution to each of the aligned and liberal components.
18
=+xAxLxBOLD measurementAlignedLiberal Relating signals to behavior
Following the signal decomposition into aligned and liberal signals, we associated the sig-
nal concentrations with median switch cost (median response time during switching trials
versus no-switching trials) performance for all accurate trials. To do so, we computed a
partial Pearson's correlation between the observed signal value for each subject with their
median switch cost using the average framewise displacement across BOLD measurements as
a second-level control for the influences of motion. Specifically, to examine the relationship
between alignment and switch costs across subjects, we computed the partial correlation for
the mean of xA for each subject with subjects' switch costs, controlling for average frame-
wise displacement. Then, to examine the relationship between liberality and switch costs,
across subjects, we computed the partial correlation for the mean of xL for each subject
with subjects' switch costs, controlling for average framewise displacement. We additionally
repeated these analyses including age and sex and found similar slopes of the associations
between the liberality values and switch costs (see Supplement).
System permutation test
To examine the spatial significance of system-level concentration of aligned and liberal
signals, we performed a non-parametric permutation test for each signal class. Separately for
each of xL and xA, we shuffled the observed mean signal concentration values across regions
in 10,000 permutations for aligned and liberal signals and computed a null distribution of
system mean signal concentrations for each system. Signals were judged to be significantly
concentrated in a system if the mean signal concentration in the system was greater or less
than 95% of the null permutations.
19
Supplementary Information and Analyses
Behavior
Across the 28 subjects, individuals were 94% (St.D. = 1%) accurate across all trials. The
average median response time for accurate non-switching trials was 0.89 s (St.D. = 0.12 s).
The average median response time for accurate switching trials across subjects was 1.22 s
(St.D. = 0.18 s). The average switch cost across subjects was 0.32 s (St.D. = 0.08 s).
Signal alignment as a trait-level variable in human brain networks
In the main text, we report intraclass correlations that indicate that alignment and liberality
form "trait" variables due to the consistency of aligned and liberal signal patterns regardless
of the task period during which the signals were sampled. To examine the stability of
aligned and liberal signals across all individual conditions, we compute a Pearson's correlation
coefficient between the signal concentration vector across regions for each condition pair
among aligned and liberal signal vectors for the fixation, no switch, and switching blocks. We
observe that the aligned and liberal signals are very highly stable for each pairwise correlation
within each signal type, but only moderately correlated across signal types (within aligned:
mean R = 0.99, p (cid:28) 0.001, within liberal: mean R = 0.99, p (cid:28) 0.001, between aligned and
liberal: mean R = 0.69, p (cid:28) 0.001; See. Fig. 1). These results establishe that during periods
of resting fixation, no switching demands, and switching demands, alignment and liberality
are highly stable subject-level traits.
Signal variability in systems
In the main text, we note that signals that are highly aligned with the underlying anatom-
ical networks were concentrated in fronto-parietal, cingulo-opercular, default mode, and
subcortical systems. This is especially interesting given numerous findings identifying the
fronto-parietal and cingulo-opercular systems to contain variable signals over time and more
flexible changes in network organization during both rest and task performance [32, 26]. To
examine the BOLD signal variability in systems across TRs in the current study, we com-
pute the average standard deviation of BOLD time series values over trials for all regions
within each system. Then, we conduct a non-parametric permutation test in which we shuf-
fle the assignment of regions to systems (n = 10, 000) and compute a null distribution of
20
Figure 6: Aligned and liberal signals are consistent traits in human brain networks.
Correlations among conditions decomposing BOLD fMRI signals using the eigenspectrum of indi-
viduals' anatomical networks. Aligned and liberal signals are highly stable at the subject level. Ali
= Aligned, Lib = Liberal.
average time series standard deviations over TRs for each system. We judge a system to
be significantly variable or invariant over time if its observed average standard deviation is
greater than or less than 97.5% of null permutations. We find that the fronto-parietal and
cingulo-opercular systems are more variable across TRs than the null expectation, whereas
the ventral attention, somatosensory, and cerebellar systems are less variable. Considered
together with our main results examining aligned and liberal signal concentrations, these re-
sults indicate that in cognitive control systems that are functionally dynamic, signal variance
is strongly organized by aligning with subject's anatomical networks from TR to TR.
System flexibility is associated with signal alignment
In the main manuscript, we focus on the trial-to-trial signal alignment with anatomy across
the brain as a basic spatiotemporal unity contributing to brain signal organization. We re-
mark that previous analyses have focused on a notion of "flexibility" in dynamic networks.
This notion is represented by creating a three-dimensional tensor from consecutive or over-
lapping association matrices (e.g., using correlation, coherence, or mutual information) and
observing the changes in node community allegiance over time. Flexible nodes are those that
exhibit a relatively high degree of changes in community assignments. Distinct from the cur-
rent analysis, flexibility occurs at a longer temporal scale since tensors are constructed from
21
Lib Fixation Lib No Switch Lib Switch Ali Switch Ali No Switch Ali Fixation Pearson's Correlation Coefficient 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Figure 7: System-level variability across TRs. Plot of system-level variability in BOLD time
series across TRs. Boxes and whiskers illustrate the distribution of system average standard devi-
ation of BOLD time series over TRs for each system. Green circles illustrate the observed average
standard deviations across TRs in each system. Asterisks denote systems in which the observed
means were higher or lower than 97.5% of permuted values.
many adjacency matrices over time, which in turn are computed over many TRs of time
series data. In this context, it is interesting to consider whether the anatomical alignment of
BOLD signals at the TR level is associated with the expression of flexibility at the temporal
network level.
To explore the relationship between flexible systems and signal alignment with anatomy,
we explicitly computed flexibility and associated these values within each participant's align-
ment values. Specifically, we followed [58] to examine the dynamic community structure of
the current data. We constructed a multilayer temporal network for each subject by com-
puting the correlation between all region pairs. Correlations whose significance did not pass
false discovery rate correction with a threshold of 0.05 were set to zero [59] for each of 10 non-
overlapping windows (40 TRs apiece) over the entire task session. We maximized multilayer
modularity Qml following [60]:
(cid:88)
ijlr
(cid:26)(cid:18)
Aijl − γl
kilkjl
2ml
(cid:19)
(cid:27)
δlr + δijCjlr
δ(gil, gjr),
Qml =
1
2µ
where the adjacency matrix of layer l (i.e., time window number l) has components Aill, γl
is the resolution parameter of layer l, gil gives the community assignment of node i in layer
22
- subcortical cingulo-opercular default mode dorsal attention fronto-parietal somatosensory auditory ventral attention visual cerebellum other * * * * * System signal standard deviation over TRs l, gjr gives the community assignment of node j in layer r, Cjlr is the connection strength
between node j in layer r and node j in layer l (see the discussion below), kil is the strength
of node i in layer l, 2µ =(cid:80)
jr kjr, kjl = kjl + cjl, and cjl =(cid:80)
r Cjlr. For simplicity, we set the
resolution parameter γl to unity and we have set all non-zero Cjlr to a constant C, which
we term the inter-layer coupling. Here, we compute community assignments for C = 1. For
each subject, we optimized the quality function 100 times and identified a representative
consensus partition for all nodes over all windows [61].
Within each subject's consensus partition, we defined the flexibility of a node fi to be the
number of times that node changed modular assignment throughout the session, normalized
by the total number of changes that were possible (i.e., by the number of consecutive pairs
of layers in the multilayer framework) [59]. We then defined the flexibility of each system
examined in the main text as the mean flexibility over all nodes in the subsystem. Finally,
we calculated the Pearson correlation coefficient between the mean system flexibility and
mean system alignment. Notably, we found that mean flexibility and mean alignment were
significantly positively correlated with one another across systems (R = 0.15, p = 0.015).
This result indicates that trial-level signal alignment across the brain is positively associated
with the expression of temporal system flexibility over much larger temporal scales across
brain networks.
Signal alignment in anatomy is similar to function
To provide further intuition for our observation that signal alignment is greatest in the most
flexible systems, we highlight that the seemingly counterintuitive observation – the regions
whose signals are aligned with the underlying structural network are also the regions that
display the greatest network flexibility across time points – is mathematically justifiable.
The reason comes from the close relationship between the graph Fourier transform (GFT)
and Principal Component Analysis (PCA) [62, 63], when the network represents pairwise
correlations between brain regions across time points. In this case, from PCA, the signal
distribution that explains most of the variation of signals across time points is represented
by the eigenvectors associated with large eigenvalues of the correlation matrix (functional
connectivity network). Simultaneously, recall from GFT that the eigenvectors associated
with large eigenvalues of the correlation matrix depict signal distributions that align well
with the network [51, 53]. This implies that the signal distribution that aligns with the
network representing pairwise correlations between brain regions also varies the most across
23
time points. Stated simply, it can be derived mathematically that the most flexible systems
are most aligned with the functional network constructed from those signals [56].
Thus, crucial to the current result, if the correlation matrix's eigenspectrum is similar to
the anatomical network, the most anatomically aligned functional signals will also be related
to the most functionally aligned and flexible signals. In order to evaluate this relationship
empirically, we examined the correlation between signals aligned with the structural networks
and signals aligned with the functional networks. For each subject in the experiment, we
constructed the subject's correlation matrix across all TRs and performed the same decom-
position of each BOLD TR from our primary analysis using each subject's (1) anatomical
network and then repeated this analysis using (2) the correlation matrix across all BOLD
TRs. Next, we examined the similarity between the signals aligned with the functional cor-
relation network and those aligned with the anatomical network. A significant correlation
would indicate that the decomposition of the aligned portion of signals is organized with the
functional correlation network similarly to its decomposition into the anatomical network,
suggesting that the relationships between flexibility and alignment is expressed both in the
subject's anatomical and mean functional correlation network.
We find a strong correlation between the anatomically and functionally aligned signals
(R = 0.814, p = 1.613 × 10−27). This empirical relationship indicates that signals most
aligned with the functional network are those most aligned with the anatomical network.
Thus, the eigenspectrum of the functional network that contributes to the tendency for flexi-
ble functional signals to align is related to the anatomical eigenspectrum cross subjects. This
highlights a fundamental relationship guiding BOLD signal dynamics in anatomical and func-
tional network analysis. Both anatomical and functional correlation networks demonstrate
an eigenspectrum that organizes flexible activity in the human brain.
24
Aligned and liberal signal associations with behavior across task
conditions
In the main text, we focus on the relationship between variability in signal alignment and
variability in switch costs due to our emphasis on cognitive control in human brain networks.
For reference, here we include a full table of results associating median response times in each
condition (no switching, switching, and switch costs) with aligned and liberal signals across
the 28 subjects. The results indicate that within the current signal decomposition using
anatomical networks, only the liberal signals demonstrate a relationship with behavioral
variability and specifically with switch costs.
Table 1: Pearson's partial correlation coefficients: all median response time associa-
tions with signal liberality and alignment.
Behavior
Lib Fix Lib NS Lib S Ali Fix Ali NS Ali S
No Switch
0.2223
0.3503
0.2207
0.0328
0.1134
-0.0907
Switch
0.4264
0.5563
0.4026
0.1962
0.2298
0.0090
Switch Cost
0.6181
0.7137
0.5669
0.3912
0.3431
0.1599
Pearson's correlation coefficients using average framewise displacement as a covariate. Rows
contain partial correlation coefficients associating each median response time at the subject
level with their graph alignment. Columns represent the signal type including the fixation,
no switch, and switch block signals. Lib = Liberal, Ali = Aligned, Fix = fixation, NS = no
switch, S = switch.
Table 2: Pearson's partial correlation p-values: all median response time associations
with signal liberality and alignment.
Behavior
Lib Fix
Lib NS
Lib S Ali Fix Ali NS Ali S
No Switch
0.2650
0.0733
0.2686
0.8709
0.5734
0.6526
Switch
0.0266
0.0026
0.0374
0.3267
0.2488
0.9644
Switch Cost
0.0006 <0.0001
0.0020
0.0436
0.0798
0.4258
Rows contain p-values corresponding to correlation coefficients in Table 1. Columns represent
the signal type including the fixation, no switch, and switch block signals. Lib = Liberal,
Ali = Aligned, Fix = fixation, NS = no switch, S = switch.
25
Inclusion of age and sex as covariates
In the main text, we relate alignment and liberality to behavior using subjects' average
framewise displacement as covariates. To examine whether our results are stable when
accounting for subjects' age and sex, we recompute partial correlations for each task condition
and alignment and liberality values using average framewise displacement, age, and sex as
covariates. Our results remain significant with similar correlation values across conditions
when including these covariates.
Table 3: Pearson's partial correlation coefficients: all median response time associa-
tions with signal liberality and alignment.
Behavior
Lib Fix Lib NS Lib S Ali Fix Ali NS Ali S
No Switch
0.3508
0.4990
0.4116
0.1168
0.1842
-0.0042
Switch
0.5009
0.6443
0.5390
0.2319
0.2390
0.0527
Switch Cost
0.6028
0.7052
0.5984
0.3459
0.2629
0.1237
Pearson's correlation coefficients using average framewise displacement, age, and sex as co-
variates. Rows contain partial correlation coefficients associating each median response time
at the subject level with their graph liberality. Columns represent the signal type includ-
ing the fixation, no switch, and switch block signals. Lib = Liberal, Ali = Aligned, Fix =
fixation, NS = no switch, S = switch.
Table 4: Pearson's partial correlation p-values: all median response time associations
with signal liberality and alignment.
Behavior
Lib Fix Lib NS Lib S Ali Fix Ali NS Ali S
No Switch
0.0856
0.0111
0.0409
0.5783
0.3780
0.9842
Switch
0.0107
0.0005
0.0054
0.2646
0.2499
0.8023
Switch Cost
0.0014
0.0001
0.0016
0.0903
0.2042
0.5559
Rows contain p-values corresponding to correlation coefficients in Table 3. Columns represent
the signal type including the fixation, no switch, and switch block signals. Lib = Liberal,
Ali = Aligned, Fix = fixation, NS = no switch, S = switch.
26
Robustness of aligned and liberal signals to parameter selection
In the main text, we report results where we represent the liberal signals xL and aligned
signals xH using cut points for KL and KH of the ten lowest (of KL) and highest (KH) values.
To test the robustness of aligned and liberal signals to variations in these parameters, we
decompose the observed BOLD signals into xL and xH under the choice of KL and KH
from five below to five above the choice of parameters studied in the main manuscript. We
then compute the correlation coefficient among xL and xH across all TRs for all subjects
to examine the stability of aligned and liberal signals. We find that the observed signal
decomposition is stable across parameter choices:
for xL (liberal signals), the values are
correlated at a mean of R = 0.95 (St.D. 0.03); for xH (aligned signals), the values are
correlated at a mean of R = 0.99 (St.D. 0.02). Thus, the signals examined in the main text
are highly stable over the choice of definition for the range of aligned and liberal signals, and
especially for liberal signals.
Null permutation test for the significance of the anatomical orga-
nization of signals
In the current analysis, the aligned signals are found to be the most concentrated in systems
including those known to be functionally flexible in prior studies [32, 26]. It is interesting to
examine the complexity of anatomical contributions to these signals. We consider whether
anatomically aligned signals are organized by the simple contributions of nearest neighbor
nodes in anatomical networks (i.e., only the nodes that share direct connections to each
node) versus the specific configuration of the entire anatomical network. This allows us to
identify whether the BOLD signals inclusive of the most flexible systems in the brain are
aligned with anatomy as a function of the contribution of the entire network, which would
potentially suggest that functionally flexible systems are organized by weighted contributions
from the entire brain.
To test this possibility, we conduct a non-parametric permutation test. Specifically, we
randomize each individual's network 2000 times preserving the strength and degree sequence
of nodes in the network. Then, we decompose the observed BOLD value for each region into
its corresponding node in the randomized network across all TRs. We then calculate over
subjects the average variance in original BOLD signals accounted for by the eigenvectors
associated with the aligned signals for each random permutation. We summarize these
27
values as the mean variance accounted for across all null permutations and compare it to
the mean variance accounted for by the original anatomical networks. We observe that the
variance accounted for in the real networks exceeds that accounted for in 100% of the null
permutations. This indicates that the BOLD signals are organized by anatomical features
at a greater topological length across the networks than the direct connections among nodes
(See Fig. 9). This finding indicates that signals that depend on underlying anatomy represent
complex contributions from anatomical network topology.
The behavioral relevance of this anatomically-aligned organization may be established
in future studies.
In particular, when the underlying network is constructed from fMRI
measurements (functional similarity networks constructed from correlations), it is known
that the components of BOLD signals that are the most aligned to the network also explain
the most variation across time [56]. Here also we find that signals aligned with anatomical
connectivity contain BOLD measurements that are the most variable over time in cingulo-
opercular and fronto-parietal systems. Functional connectivity (e.g., correlation, coherence,
and mutual information) networks evince similar but not completely overlapping topology
when compared to anatomical networks [64]. Thus, analyzing the alignment of measured cor-
relations on anatomical networks may provide an important extension to understand signals
in the time domain in the human connectome. Mathematically, the same phenomenon for
functional as well as anatomical networks implies that the eigenspectrum of the anatomical
network is not that different from functional connectivity. Speculatively, this could provide
a meaningful and efficient method to reliably recover anatomical networks from functional
connectivity networks.
Alternate measures of function and anatomy
Do anatomy-decomposed functional BOLD signals better related to switch costs than other
measures with similar intuitions? To test this possibility, we consider four comparison anal-
yses involving the subcortical system from which the behaviorally relevant signals in our
primary analyses originated. First, we compute the mean BOLD signal variance over all sub-
cortical regions for each TR. This correlation represents a measure of differences in BOLD
signals at each time measurement without considering its anatomical network context. We
then relate the average BOLD signal variance over TRs to switch costs and find no signifi-
cant relationship (R = −0.09, p = 0.68). Second, we compute the mean node strength – the
sum of connections to the node – for each region in the subcortical system. This represents
28
Figure 8: Aligned signals are organized by topologically distant anatomical organization.
Left: in the null permutations where the edges of the anatomical networks were randomized, a mean
of 0.455% variance is accounted for by the eigenvectors corresponding to the aligned parts of the
signal. In the observed data, a mean of 0.547% of variance is accounted for by the anatomical
networks in their true configurations.
the overall influence of the subcortical regions in the broader anatomical network. We cal-
culate the Pearson correlation coefficient between the mean subcortical node strength and
switch costs across subjects, and we find no significant relationship (R = −0.15, p = 0.48).
Finally, within each subject, we compute the mean Pearson correlation coefficient between
each BOLD signal TR and node strengths within the subcortical system, then we compute
the Pearson correlation coefficient between these values and switch costs across subjects,
finding no significant relationship (R = 0.17, p = 0.41). Taken together, these results
demonstrate that simple measures of BOLD variation, local anatomical network influences,
and the relationship between the two are insufficient to account for switch cost variability
over subjects. Indeed, decomposing signals in the context of the entire anatomical network
provides superior associative value for the cognitive control measure.
29
Number of null observations % variance explained Null Real 0 20 40 60 80 100 120 140 160 180 0.42 0.44 0.46 0.48 0.50 0.52 0.54 0.56 Figure 9: Alternate measures of framewise BOLD variance, anatomy, and function-
anatomy relationships in subcortical systems do not associate with switch costs. (A)
The relationship between mean BOLD variance in the subcortical system and switch costs. (B) The
relationship between mean anatomical node degree in the subcortical system and switch costs. (C)
The relationship between the correlation between BOLD signals and anatomical node degree across
regions in the subcortical systems and switch costs.
Anatomical network null model for behavioral correlations with
switch costs.
To examine the importance of the specific anatomical configuration of brain networks for
the association between liberal signals and switch costs, we performed a null permutation
test. Specifically, we performed 200 permutations in which we generated a random network
preserving the strength and degree distributions of the original anatomical networks for each
subject. Then, we repeated the BOLD signal decomposition into the null anatomical network
for each subject and correlated the liberal signals with switch costs in each permutation
to generate a null distribution of correlations. From this null distribution, we computed
the proportion of null correlation values greater than the observed correlation when using
the real anatomical networks. We find that the observed correlation is greater than 99%
of correlations (Fig. 10). This result indicates that the specific configuration of the true
anatomical network drives the correlation between liberal signal alignment and switch costs.
30
01000200030000.20.250.30.350.40.450.5500100015000.20.250.30.350.40.450.5-0.04-0.0200.020.040.20.250.30.350.40.450.5Switch cost (s)Switch cost (s)Switch cost (s)Mean BOLD varianceMean node anatomical strengthMean R between BOLD, anatomical strengthABC Figure 10: A null permutation test demonstrates that specific anatomical network or-
ganization is required to identify behaviorally relevant BOLD signals. Blue bars in the
histogram illustrate the Pearson's correlation coefficients between switch costs and liberal signals
over null permutations. Orange line designates the observed correlation value of the real network.
Relationships between motion and signal alignment.
In the results reported in the main text and supplement, we included a motion covariate
(average framewise displacement across BOLD TRs) in all analyses. Motion is not signifi-
cantly correlated with switch cost (R = 0.26, p = 0.16), but is correlated significantly with
liberal signals (R = 0.77, p = 1.10×10−6), justifying its inclusion as a covariate in behavioral
analyses. Importantly, when motion is not included as a covariate in the partial correlation
between liberal signals and switch costs, the correlation (R = 0.59 , p = 0.001) is highly
similar to that observed when including motion as a covariate (R = 0.57, p = 0.002). This
finding indicates that the behaviorally relevant portion of the liberal signals is not driven by
motion.
References
[1] Rogers, R. D. & Monsell, S. Costs of a predictible switch between simple cognitive
tasks. Journal of experimental psychology: General 124, 207 (1995).
[2] Szczepanski, S. M. & Knight, R. T. Insights into human behavior from lesions to the
prefrontal cortex. Neuron 83, 1002–1018 (2014).
31
Number of null observations Pearson's R 0.20.30.40.50.60.7051015202530Null Real [3] Clark, L. R. et al. Specific measures of executive function predict cognitive decline
in older adults. Journal of the International Neuropsychological Society 18, 118–127
(2012).
[4] Richland, L. E. & Burchinal, M. R. Early executive function predicts reasoning devel-
opment. Psychological science 24, 87–92 (2013).
[5] Davis, J. C., Marra, C. A., Najafzadeh, M. & Liu-Ambrose, T. The independent con-
tribution of executive functions to health related quality of life in older women. BMC
geriatrics 10, 16 (2010).
[6] Gunaydin, L. A. & Kreitzer, A. C. Cortico-basal ganglia circuit function in psychiatric
disease. Annual review of physiology 78, 327–350 (2016).
[7] Casey, B. et al. Early development of subcortical regions involved in non-cued attention
switching. Developmental science 7, 534–542 (2004).
[8] Cole, M. W. et al. Multi-task connectivity reveals flexible hubs for adaptive task control.
Nature neuroscience 16, 1348–1355 (2013).
[9] Heyder, K., Suchan, B. & Daum, I. Cortico-subcortical contributions to executive
control. Acta psychologica 115, 271–289 (2004).
[10] Luk, G., Green, D. W., Abutalebi, J. & Grady, C. Cognitive control for language
switching in bilinguals: A quantitative meta-analysis of functional neuroimaging studies.
Language and cognitive processes 27, 1479–1488 (2012).
[11] Hikosaka, O. & Isoda, M. Switching from automatic to controlled behavior: cortico-
basal ganglia mechanisms. Trends in cognitive sciences 14, 154–161 (2010).
[12] Hosoda, C., Hanakawa, T., Nariai, T., Ohno, K. & Honda, M. Neural mechanisms of
language switch. Journal of Neurolinguistics 25, 44–61 (2012).
[13] Leunissen, I. et al. Subcortical volume analysis in traumatic brain injury: the impor-
tance of the fronto-striato-thalamic circuit in task switching. Cortex 51, 67–81 (2014).
[14] Yehene, E., Meiran, N. & Soroker, N. Basal ganglia play a unique role in task switching
within the frontal-subcortical circuits: evidence from patients with focal lesions. Journal
of Cognitive Neuroscience 20, 1079–1093 (2008).
32
[15] Sporns, O., Tononi, G. & Kotter, R. The human connectome: A structural description
of the human brain. PLoS Computational Biology 1, e42 (2005).
[16] Alstott, J., Breakspear, M., Hagmann, P., Cammoun, L. & Sporns, O. Modeling the
impact of lesions in the human brain. PLoS computational biology 5, e1000408 (2009).
[17] Hermundstad, A. M. et al. Structural foundations of resting-state and task-based func-
tional connectivity in the human brain. Proceedings of the National Academy of Sciences
110, 6169–6174 (2013).
[18] Honey, C. J., Kotter, R., Breakspear, M. & Sporns, O. Network structure of cere-
bral cortex shapes functional connectivity on multiple time scales. Proceedings of the
National Academy of Sciences 104, 10240–10245 (2007).
[19] Medaglia, J. D., Lynall, M.-E. & Bassett, D. S. Cognitive network neuroscience. Journal
of cognitive neuroscience (2015).
[20] Sporns, O. Contributions and challenges for network models in cognitive neuroscience.
Nature Neuroscience 17, 652–660 (2014).
[21] Power, J. D. et al. Functional network organization of the human brain. Neuron 72,
665–678 (2011).
[22] Navon, D. Forest before trees: The precedence of global features in visual perception.
Cognitive psychology 9, 353–383 (1977).
[23] Cammoun, L. et al. Mapping the human connectome at multiple scales with diffusion
spectrum mri. Journal of Neuroscience Methods 203, 386–397 (2012).
[24] Diedrichsen, J., Balsters, J. H., Flavell, J., Cussans, E. & Ramnani, N. A probabilistic
mr atlas of the human cerebellum. Neuroimage 46, 39–46 (2009).
[25] Gu, S. et al. Controllability of structural brain networks. Nature Communications 6,
8414 (2015).
[26] Braun, U. et al. Dynamic reconfiguration of frontal brain networks during executive
cognition in humans. Proceedings of the National Academy of Sciences 112, 11678–
11683 (2015).
33
[27] Mayhew, S. D. et al. Global signal modulation of single-trial fmri response variability:
Effect on positive vs negative bold response relationship. NeuroImage 133, 62–74 (2016).
[28] Sekutowicz, M. et al. Striatal activation as a neural link between cognitive and percep-
tual flexibility. NeuroImage 141, 393–398 (2016).
[29] Liston, C., Matalon, S., Hare, T. A., Davidson, M. C. & Casey, B. Anterior cingulate
and posterior parietal cortices are sensitive to dissociable forms of conflict in a task-
switching paradigm. Neuron 50, 643–653 (2006).
[30] Cole, M. W., Bassett, D. S., Power, J. D., Braver, T. S. & Petersen, S. E. Intrinsic and
task-evoked network architectures of the human brain. Neuron 83, 238–251 (2014).
[31] Watanabe, T., Masuda, N., Megumi, F., Kanai, R. & Rees, G. Energy landscape and
dynamics of brain activity during human bistable perception. Nature communications
5 (2014).
[32] Mattar, M. G., Betzel, R. F. & Bassett, D. S. The flexible brain. Brain 139, 2110–2112
(2016).
[33] Miyake, A., Friedman, N. P., Emerson, M. J., Witzki, A. H. & Howerter, A. The unity
and diversity of executive functions and their contributions to complex "frontal lobe"
tasks: A latent variable analysis. Cognitive Psychology 41, 49–100 (2000).
[34] Fedorenko, E. The role of domain-general cognitive control in language comprehension.
Frontiers in psychology 5, 335 (2014).
[35] Hermundstad, A. M. et al. Structurally-constrained relationships between cognitive
states in the human brain. PLoS Comput Biol 10, e1003591 (2014).
[36] Belleville, S., Bherer, L., Lepage, ´E., Chertkow, H. & Gauthier, S. Task switching
capacities in persons with alzheimer's disease and mild cognitive impairment. Neu-
ropsychologia 46, 2225–2233 (2008).
[37] Kehagia, A. A., Barker, R. A. & Robbins, T. W. Neuropsychological and clinical
heterogeneity of cognitive impairment and dementia in patients with parkinson's disease.
The Lancet Neurology 9, 1200–1213 (2010).
[38] Kinnunen, K. M. et al. White matter damage and cognitive impairment after traumatic
brain injury. Brain awq347 (2010).
34
[39] Betzel, R. F., Gu, S., Medaglia, J. D., Pasqualetti, F. & Bassett, D. S. Optimally
controlling the human connectome: the role of network topology. Scientific Reports
(2016).
[40] Yeh, F.-C., Wedeen, V. J. & Tseng, W.-Y. I. Estimation of fiber orientation and spin
density distribution by diffusion deconvolution. Neuroimage 55, 1054–1062 (2011).
[41] Fischl, B. Freesurfer. Neuroimage 62, 774–781 (2012).
[42] Cammoun, L. et al. Mapping the human connectome at multiple scales with diffusion
spectrum mri. Journal of neuroscience methods 203, 386–397 (2012).
[43] Cieslak, M. & Grafton, S. Local termination pattern analysis: a tool for comparing
white matter morphology. Brain imaging and behavior 8, 292–299 (2014).
[44] Hagmann, P. et al. Mapping the structural core of human cerebral cortex. PLoS Biol
6, e159 (2008).
[45] Voogd, J. & Glickstein, M. The anatomy of the cerebellum. Trends in cognitive sciences
2, 307–313 (1998).
[46] Jenkinson, M., Beckmann, C. F., Behrens, T. E., Woolrich, M. W. & Smith, S. M. Fsl.
Neuroimage 62, 782–790 (2012).
[47] Greve, D. N. & Fischl, B. Accurate and robust brain image alignment using boundary-
based registration. Neuroimage 48, 63–72 (2009).
[48] Zhang, Y., Brady, M. & Smith, S. Segmentation of brain mr images through a hid-
den markov random field model and the expectation-maximization algorithm. IEEE
transactions on medical imaging 20, 45–57 (2001).
[49] Jenkinson, M., Bannister, P., Brady, M. & Smith, S. Improved optimization for the ro-
bust and accurate linear registration and motion correction of brain images. Neuroimage
17, 825–841 (2002).
[50] Chung, F. Spectral graph theory, vol. 92 (American Mathematical Soc., 1997).
[51] Sandryhaila, A. & Moura, J. M. Discrete signal processing on graphs. IEEE transactions
on signal processing 61, 1644–1656 (2013).
35
[52] Shuman, D. I., Narang, S. K., Frossard, P., Ortega, A. & Vandergheynst, P. The
emerging field of signal processing on graphs: Extending high-dimensional data analysis
to networks and other irregular domains. IEEE Signal Processing Magazine 30, 83–98
(2013).
[53] Ma, J., Huang, W., Segarra, S. & Ribeiro, A. Diffusion filtering for graph signals and its
use in recommendation systems. In Acoustics, Speech and Signal Processing (ICASSP),
2016 IEEE Int. Conf. on, 4563–4567 (Shanghai, China, 2016).
[54] Segarra, S., Huang, W. & Ribeiro, A. Diffusion and superposition distances for signals
supported on networks. Signal Inform. Process. over Network., IEEE Trans. on 1,
20–32 (2015).
[55] Huang, W., Segarra, S. & Ribeiro, A. Diffusion distance for signals supported on
networks. In Proc. Asilomar Conf. on Signals Systems Computers, 1219–1223 (Asilomar
CA, 2015).
[56] Huang, W. et al. Graph frequency analysis of brain signals. J. Sel. Topics Signal
Process. 10, 1189–1203 (2016).
[57] Spielman, D. Spectral graph theory. Lecture Notes, Yale University 740–0776 (2009).
[58] Bassett, D. S. et al. Robust detection of dynamic community structure in networks.
Chaos: An Interdisciplinary Journal of Nonlinear Science 23, 013142 (2013).
[59] Bassett, D. S. et al. Dynamic reconfiguration of human brain networks during learning.
Proceedings of the National Academy of Sciences 108, 7641–7646 (2011).
[60] Mucha, P. J., Richardson, T., Macon, K., Porter, M. A. & Onnela, J.-P. Community
structure in time-dependent, multiscale, and multiplex networks. Science 328, 876–878
(2010).
[61] Bassett, D. S. & Siebenhuhner, F. Multiscale network organization in the human brain.
Multiscale Analysis and Nonlinear Dynamics: From Genes to the Brain 179–204 (2013).
[62] Leonardi, N. et al. Principal components of functional connectivity: a new approach to
study dynamic brain connectivity during rest. NeuroImage 83, 937–950 (2013).
[63] Viviani, R., Gron, G. & Spitzer, M. Functional principal component analysis of fmri
data. Human brain mapping 24, 109–129 (2005).
36
[64] Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis of
structural and functional systems. Nature Reviews Neuroscience 10, 186–198 (2009).
37
Acknowledgements
JDM acknowledges support from the Office of the Director at the National Institutes of
Health through grant number 1-DP5-OD-021352-01. DSB acknowledges support from the
John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Army
Research Laboratory and the Army Research Office through contract numbers W911NF-
10-2-0022 and W911NF-14-1-0679, the National Institute of Health (2-R01-DC-009209-11,
1R01HD086888-01, R01-MH107235, R01-MH107703, R01MH109520, 1R01NS099348 and
R21-M MH-106799), the Office of Naval Research, and the National Science Foundation
(BCS-1441502, CAREER PHY-1554488, BCS-1631550, and CNS-1626008). The content is
solely the responsibility of the authors and does not necessarily represent the official views
of any of the funding agencies.
Competing Interests
The authors declare that they have no competing financial interests.
Correspondence
Correspondence and requests for materials should be addressed to D.S.B. (email: [email protected]).
38
|
1103.0668 | 1 | 1103 | 2011-03-03T12:37:57 | Spike Onset Dynamics and Response Speed in Neuronal Populations | [
"q-bio.NC"
] | Recent studies of cortical neurons driven by fluctuating currents revealed cutoff frequencies for action potential encoding of several hundred Hz. Theoretical studies of biophysical neuron models have predicted a much lower cutoff frequency of the order of average firing rate or the inverse membrane time constant. The biophysical origin of the observed high cutoff frequencies is thus not well understood. Here we introduce a neuron model with dynamical action potential generation, in which the linear response can be analytically calculated for uncorrelated synaptic noise. We find that the cutoff frequencies increase to very large values when the time scale of action potential initiation becomes short. | q-bio.NC | q-bio |
Spike Onset Dynamics and Response Speed in Neuronal Populations
MPI for Dynamics and Self-Organization, Faculty of physics, Georg-August University,
and Bernstein center for computational neuroscience, Gottingen, Germany
Wei Wei and Fred Wolf
Recent studies of cortical neurons driven by fluctuating currents revealed cutoff frequencies for
action potential encoding of several hundred Hz. Theoretical studies of biophysical neuron models
have predicted a much lower cutoff frequency of the order of average firing rate or the inverse
membrane time constant. The biophysical origin of the observed high cutoff frequencies is thus not
well understood. Here we introduce a neuron model with dynamical action potential generation, in
which the linear response can be analytically calculated for uncorrelated synaptic noise. We find
that the cutoff frequencies increase to very large values when the time scale of action potential
initiation becomes short.
PACS numbers: 87.19.ll, 05.40.-a, 87.19.ls
Keywords: Action potential, Onset rapidness, Linear response, Cut-off frequency
In the cerebral cortex of the brain information is en-
coded in the action potential (AP) firing rates of a large
ensemble of nerve cells. Recent experiments have ob-
served a surprisingly high cutoff frequency for the action
potential encoding of cortical neurons driven by fluctuat-
ing input currents [1 -- 4]. In a seminal paper Kondgen et
al. showed that the transmission function of layer 5 pyra-
midal neurons for a noisy sinusoidal signal does not de-
cay until about 200 Hz [1]. Later experiments confirmed
such high cutoff frequencies for signals coded by both the
mean current and noise strength [2] and in other types
of cortical neurons [3, 4]. For an early observation of fast
response see [5]. Previous theoretical studies of biophysi-
cal neuron models, however, predicted cutoff frequencies
of the order of the average firing rate or the inverse mem-
brane time constant (below 20 Hz), much lower than the
experimentally observed values [6 -- 8]. Thus, the origin of
the high cutoff frequencies found in cortical neurons is
currently not well understood. Numerical investigation
of neuron models with dynamical AP generation, like the
exponential integrate-and-fire (EIF) model or the gener-
alized theta neurons, suggested that details of AP gen-
eration can influence the dynamical response of neuronal
populations [6 -- 9]. What is missing, however, is a trans-
parent understanding of how and when the population
cutoff frequency can dissociate from the basic single neu-
ron timescale set by the mean firing rate and the time
constant of membrane potential relaxation.
In this work we present an analytically solvable model
which explicitly describes the dynamical AP initiation
process. A neuron initiates an AP if the membrane po-
tential passes an unstable fixed point, the voltage thresh-
old.
In the leaky integrate-and-fire (LIF) model, for
which the linear response is known analytically, the un-
stable fixed point coincides with the absorbing bound-
ary and a spike is triggered immediately when the mem-
brane potential reaches this threshold [10, 11]. As a con-
sequence, boundary induced artifacts dominate the re-
sponse for high signal frequencies in the LIF model [6 -- 8].
One important advantage of our new model is that such
boundary induced artifacts can be separated out mathe-
matically, isolating the physically meaningful part of the
response function. We first present the linear response
for both encoding paradigms with white noise. We find
that for a wide range of parameter settings the cutoff
frequency is directly proportional to the AP onset rapid-
ness for a noise coded signal. It therefore dissociates from
the membrane time constant and can become arbitrarily
large. For the mean current coded signal, however, the
cutoff frequency is confined by the membrane time con-
stant in the white noise case. We show by numerical
simulation that this confinement can be broken when a
finite correlation time in the synaptic noise is taken into
account and high cutoff frequencies can be obtained for
a large AP onset rapidness.
Interestingly, experiments
showed that the AP onset rapidness of cortical neurons
is very large both in vitro and in vivo [12, 13], which
may thus explain the occurrence of high cutoff frequen-
cies. Our results provide a relationship between the spike
onset dynamics and the population cutoff frequency that
can be directly tested in physiological experiments.
The simplest voltage dynamics that exhibits both a
stable fixed point (the resting potential) and an unsta-
ble fixed point (the voltage threshold) has a piecewise
linear membrane current, composed of a leak current for
low potential and a linear spike generating current for
high potential [Fig. 1(a)]. The model is defined by the
following Langevin equation
τm v = −v + Θ(v − v0) (r + 1)(v − v0) + µ + ση(t) , (1)
where v is the membrane potential relative to the rest-
ing potential, τm is the membrane time constant, Θ(v)
is the Heaviside step function, r is the AP onset rapid-
ness which sets an effective time constant τm/r for the
AP initiation process. The larger is r, the faster is the
spike onset [Fig. 1(b)].
In biophysical models, the on-
set rapidness will be set largely by intrinsic properties of
the voltage dependent sodium channels, e.g., the gating
′
′
)i = τmδ(t − t
charge and the slope of the activation curve [12]. µ is the
mean input current and σ is the amplitude of synaptic
noise. η(t) is a Gaussian white noise satisfying hη(t)i = 0
). The crossing point v0 of
and hη(t)η(t
the two pieces sets the rheobase current, which we use as
the unit of voltage, v0 = 1. The threshold potential is
vt = (1 + 1/r)v0. When the membrane potential reaches
vb, the truncation point of the AP upstroke, it is reset
to a voltage vr and stays there for an absolute refractory
period τr. For convenience we take τm as the unit of time
in analytical calculation.
The Fokker-Plank equation corresponding to Eq. (1)
has the following form
∂tP1 + ∂v(cid:18) − v + µ −
∂tP2 + ∂v(cid:18)r(v − vt) + µ −
1
2
σ2∂v(cid:19)P1 = 0 ,
1
2
σ2∂v(cid:19)P2 = 0,
(2)
where P1(v, t) and P2(v, t) are the probability densities
of membrane potential v for −∞ < v ≤ 1 and 1 < v ≤ vb
respectively. The stationary firing rate can be found eas-
ily when the boundary conditions are specified. From the
reset assumption, we impose an absorbing boundary at
vb and, as a result the probability density is zero there,
P2(vb, t) = 0. The firing rate is given by the probabil-
ity current at vb, ν(t) = − 1
2 σ2∂vP2(vb, t). At the reset
voltage vr, the probability density is continuous, while
its first derivative has a discontinuity from the reset con-
dition: ∂vP1(v+
r , t) − ∂vP1(v−r , t) = ∂vP2(vb, t − τr). In
addition, the density and its first derivative should be
continuous at v = 1. When µ and σ are constants, the
system is homogeneous and the stationary solution of
Eq. (2), denoted as P01(v) and P02(v) respectively, can
be found. The stationary firing rate ν0 is then obtained
from the normalization condition of the density [14].
The stationary firing rate ν0 of the model Eq.
(1)reduces to that for the LIF model for r → ∞. Figure
1(c) and D show the dependence of ν0 on the mean input
µ and the amplitude of noise σ, respectively. The firing
rate ν0 increases monotonically with r, µ, and σ and is
relatively insensitive to r when r > 10. For the dynam-
ical response, however, the r dependence is much more
pronounced. The instantaneous firing rate of an ensem-
ble of model neurons responds much faster for larger r to
a step change in the noise level [14].
Linear response. -- When the input current to a neu-
ron is weakly modulated, linear response theory can be
applied to study the dynamical response properties of an
ensemble of neurons. To this end, we consider a sinu-
soidal signal ε cos(ωt), where ε is small. When the signal
is encoded in the mean current, µ → µ + ε cos(ωt) or in
the noise amplitude, σ → σ + εcos(ωt), the instantaneous
firing rate can be written as ν(t) = ν0 + εν1c(ω)cos(ωt−
φc(ω)) or ν(t) = ν0 + εσν1n(ω)cos(ωt − φn(ω)). Here
ν1(ω) are the complex response functions. The absolute
(a)
v
1
τm
r
τm
0
1
vb
v
(b)
)
t
(
V
4
3
2
1
0
−1
0
(d)
20
r=1
r=10
r=100
r= ∞
σ=0.3
(c)
20
)
z
H
(
0
ν
15
10
5
0
0
)
z
H
(
0
ν
0.5
µ
1
15
10
5
0
0
2
10
1
r=1
r=10
r=100
r=1
r=10
r=100
r= ∞
µ=0
5
t(ms)
0.5
σ
FIG. 1: (color online) (a) Illustration of the model. (b) V (t)
trajectories for identical noise and three different values of
r. (c) and (d) show the dependence of stationary firing rate
on mean input current and noise strength in the noise driven
regime. (τm = 10 ms, vr = 0, vb = 10, τr = 0)
value ν1(ω) are the transmission functions and the phase
angles φ1(ω) = arg(ν1(ω)) give the phase lags, which
completely characterize the linear response. Note that
we refer to both signal channels when the subscripts c
and n are omitted.
It is known that the absorbing boundary condition
at vb can induce severe artifacts in the dynamical re-
sponse. The potential vb marks a "point of no re-
turn," which is not present in a biophysical dynamical
model of AP initiation. As a consequence, the trans-
mission function for a noise coded signal in the LIF
model, for instance, does not decay at high signal fre-
quencies [5, 11].
Ideally one would thus wish to sepa-
rate the response function into a physiologically mean-
ingful part νphy
(ω) and a part containing all artifacts
1
such that νphy
(ω). νphy
(ω) = ν1(ω) − νabs
(ω) must have
1
the following properties: i) νphy
(ω) approaches the static
susceptibility when ω → 0, specifically, νphy
1c (ω) → ∂ν0
and νphy
(ω) → 0 when ω → ∞;
iii) no essential dependence on the truncation point vb.
The artifactual part from the absorbing boundary should
have the following properties:
i) negligible contribution
for signal frequency in the physiologically relevant range
f ≤ 1kHz, e.g.
(ω), where f = ω/2π;
ii) strong dependence on the truncation point vb. As we
will show next such an isolation of the physiologically
meaningful response is possible in the model Eq. (1).
(ω) ≪ νphy
1n (ω) → 1
∂σ ; ii) νphy
νabs
∂ν0
∂µ
σ
1
1
1
1
1
1
In the model Eq. (1), the linear response can be ob-
tained analytically by expanding the probability density
in Eq. (2) to first order in ε and using the Green's func-
tion method. We find that ν1n(ω) decomposes naturally
into two parts, ν1(ω) = νLow
(ω) + νHigh
(ω), with
1
1
νLow
1c
(ω) =
νLow
1n (ω) =
iω
(1 − iω)(1 + iω/r)
iω(iω − 1)
(2 − iω)(2 + iω/r)
(a)
c
1
ν
e
d
u
t
i
l
p
m
A
100
10−2
ν
High→
1c
← ν
Low
1c
(b)
c
φ
g
a
L
e
s
a
h
P
2
0
−2
100
102
f(Hz)
(c)
n
1
ν
e
d
u
t
i
l
p
m
A
100
10−2
ν
High→
1n
← ν
Low
1n
100
102
f(Hz)
(d)
n
φ
g
a
L
e
s
a
h
P
100
2
0
−2
100
102
f(Hz)
102
f(Hz)
FIG. 2: (color online) The normalized function ν1(ω)/ν1(0.1)
and phase lag for a mean coded signal and noise coded signal
with r = 1, µ = 0, and ν0 = 5 Hz. Other parameters are the
same as in Fig. 1. Lines, theory; Symbols, simulation.
where D = 1
2 σ2, ∆0 = (1 − vr)(2µ − 1 − vr)/4D and
∆1 = (1 − vb)(2µ − 1 + r(vb − vt))/4D. ψ1, Φ1, and Υ1
are parabolic cylinder functions, and Y1, Y2 are a combi-
nation of parabolic cylinder functions, whose definition
together with the νHigh
(ω) parts are given in the sup-
plement [14]. Any prime represents the derivative with
respect to v. Here and in the following, the functions
adopt their values at v = 1 if not denoted explicitly. Fig-
ure 2 illustrates the linear response with r = 1 as an
example.
1
Removing boundary induced artifacts. -- The decompo-
sition of ν1(ω) into two additive components has exactly
the features required for the separation of artifacts. Us-
ing asymptotic expansion of the parabolic cylinder func-
tions we find that for a finite signal frequency, νHigh
can
be approximated by
1
νHigh
1c
(ω) ≃
νHigh
1n
(ω) ≃ −
ν0
r(vb − vt) + µ
ν0
iω
r + iω
,
(r(vb − vt) + µ)2
iω(1 + iω/r)
2 + iω/r
. (4)
1
when vb ≫ vt. So νHigh
(ω) are strongly dependent on
vb and approach zero when vb → ∞ for finite signal
frequency. When ω → 0, νHigh
(ω) are negligible com-
pared with νLow
(ω) and are strongly suppressed when
vb is large. That νHigh
(ω) captures all artificial con-
1
1
1n
01) − (1 + iω/r)Φ1(vr)e∆0+iωτr
,
′
(1 + 1/r)(ψ1P01 − √DΦ1P
(1 + 1/r)(cid:0)
ψ1(vr)e∆0+iωτr + (Y1ψ ′
Φ1P01 + 2Υ1P
(1−iω)√D
iω
1 ψ1)e∆1
′
1 − Y ′
01(cid:1) + ν0
1 − Y ′
ψ1(vr)e∆0+iωτr + (Y1ψ ′
1 ψ1)e∆1
D (2 + iω/r)Υ(vr)e∆0+iωτr
3
,
(3)
1
1c
1n
νLIF
1c
(ω) → ν0
D = limω→∞
(ω) → ν0√D
√ω eiπ/4 = limω→∞
tributions imposed by the absorbing boundary condi-
tion is finally confirmed from the high frequency behav-
ior, νHigh
(ω) and
νHigh
νLIF
1n (ω), since the high fre-
1n
quency behavior in the LIF model is determined solely
by the absorbing boundary. As a consequence, neglect-
ing νHigh
(ω) in the response function eliminates any
boundary induced instantaneous response components.
These results establish that νLow
(ω) capture the be-
havior of ν1(ω) for low and intermediate frequencies
and decay to zero in the large frequency limit. When
ω → 0, νLow
(ω) is
negligible there. νLow
(ω) exhibits only a weak depen-
dence on vb through a frequency dependence phase lag
φ0 = ω
, characterizing the time lag to the trun-
cation point vb of the AP upstroke. Therefore, we have
νLow
(ω). The physiologically meaningful pre-
1
dictions of the model can thus be revealed by examining
νLow
1
1n (ω) → ν1n(0) = 1
∂σ , since νHigh
(ω) in isolation.
r log vb+µ
√rD
(ω) = νphy
∂ν0
1n
σ
1
1
1
1
Cutoff frequency and AP onset rapidness. -- Figure 3
shows the behavior of νLow
(ω) with increasing r and how
the cutoff frequency fc changes with r for different ν0
[15]. We see that fc increases linearly with the onset
rapidness r for noise coded signals when the firing rate
is not very low (> 1 Hz here); while for mean coded
signals fc saturate for large r. The increase of fc with
firing rate ν0 results from the stochastic double resonance
phenomenon: the transmission function will develop a
peak for some optimal signal frequency before decaying
when ν0 is relatively large [16].
Figure 3D suggests that the cutoff frequency for a
noise coded signal follows fc ∝ r and dissociates from
τm. This is directly confirmed by the large frequency
approximation of νLow
1n (ω),
exp(− π
2 + iω/r (cid:18) − P
4 ω/r)
′
′
(5)
01 +
ipiω/r
2√D
ν Low
1n (ω) ∝
01(v0) → − ν0
P01(cid:19) ,
where P01(v) ≡ √rP01(v). Because P01(v0) → p π
D for r ≫ 1, the decay of νLow
ν0√D
1n (ω)
and P
depends essentially only on ω/r. This implies that the
cutoff frequency for a noise coded signal dissociates from
τm and becomes proportional to the onset rapidness r,
fc = Ar, where A depends on ν0 through the effect dis-
cussed above. This demonstrates that fast onset APs can
enhance the cutoff frequency and therefore the response
speed significantly. Note that a linear relationship fc ∝ r
2
(a)
c
1
l
w
o
ν
e
d
u
t
i
l
p
m
A
(c)
n
1
l
w
o
ν
e
d
u
t
i
l
p
m
A
100
10−2
10−4
100
100
10−2
10−4
100
r=1
r=10
r=100
r=1000
r=∞
102
f(Hz)
r=1
r=10
r=100
r=1000
r=∞
102
f(Hz)
(b)
(b)
)
z
H
(
)
c
1
ν
(
c
f
300
250
200
150
100
50
104
0
50
r
100
(d)
(d)
)
z
H
(
)
n
1
ν
(
c
f
3000
2500
2000
1500
1000
500
104
0
50
r
100
1
FIG. 3: (color online) (a) and (c) The normalized transmis-
sion function ν Low
(ω)/ν1(0.1) for different r with µ = 0,
ν0 = 5 Hz.
(b) and (d) The variation of cutoff frequency
with r for different firing rates: ν0 = 1, 5, 10, 20, 30, 40 Hz
from lower to upper curves. Other parameters are the same
as in Fig. 1.
(a)
100
c
1
ν
e
d
u
t
i
l
p
m
A
10−2
100
r=10
τ
s
τ
s
τ
s
τ
s
τ
s
=0
=1
=5
=10
=20
102
f(Hz)
(b)
100
c
1
ν
e
d
u
t
i
l
p
m
A
10−2
100
τ
=0
s
τ
=1
s
τ
=5
s
τ
=10
s
τ
=20
s
r=100
102
f(Hz)
FIG. 4: Variation of the normalized transmission function
ν1c(ω)/ν1c(0.1) with increasing correlation time τs (with
unit ms) of the synaptic noise for r = 10, 100. Parameter
used: µ = 0, vb = 10, ν0 = 5Hz.
was previously conjectured by Naundorf et al. based on
dimensional analysis [18].
1
1c
For a current coded signal, however, νLow
(ω) ∝
√ω exp(− π
4 ω/r) for large r. Therefore the linear response
is confined by the membrane time constant in the white
noise case, as seen also from Fig. 3A. Real synaptic in-
puts, however, have a finite correlation time and can be
approximated better as a colored noise. As shown in
Fig. 4, the confinement by the membrane time constant
is broken under such conditions and cutoff frequencies of
several hundred Hz can be reached for large r [14].
Our results identify AP onset rapidness as a critical de-
terminant of population cutoff frequency and reveal how
this cutoff frequency can dissociate from the basic single
neuron time constants set by the mean firing rate and
4
membrane time constant. The confinement of the mean
response for white noise constitutes an interesting predic-
tion of the model, which should be tested experimentally
by using a very short correlation time (≤ 1 ms) of the
synaptic noise. The origin of the large onset rapidness
seen in cortical neurons is a matter of ongoing debate
[17, 18].
Its value can be modified in real neurons by
applying drugs like TTX or knockout of sodium channel
subtypes [12, 19]. Measurements of dynamical response
for cortical neurons under such manipulations are thus
predicted to provide important insight into the mecha-
nism of fast population coding.
We thank D. Battaglia and B. Lindner for suggestions
and carefully reading an early version of the manuscript.
We are grateful to T. Geisel, M. Gutnick, M. Huang, M.
Monteforte, T. Moser and T. Tchumatchenko for discus-
sions. This work was supported by BMBF (01GQ07113),
GIF (906-17.1/2006), BCCN II (01GQ1005B), and DFG
(SFD899).
[1] H. Kondgen et al., Cereb. Cortex 18, 2086 (2008).
[2] C. Boucsein et al., J. Neurosci. 29, 1006 (2009).
[3] M. Higgs, W. Spain, J. Neurosci. 29, 1285 (2009).
[4] T. Tchumatchenko et al., to be published.
[5] G. Silberberg et al., J. Neurophysiol. 91, 704 (2004).
[6] N. Fourcaud-Trocm´e et al., J. Neurosci. 23, 11628 (2003).
[7] N. Fourcaud-Trocm´e, N. Brunel, J. Comput. Neurosci.
18, 311 (2005).
[8] B. Naundorf, T. Geisel, F. Wolf, J. Comput. Neurosci.
18, 297 (2005).
[9] M. Richardson, Phy. Rev. E 76, 021919 (2007).
[10] N. Brunel et al., Phys Rev Lett 86, 2186 (2001).
[11] B. Lindner, L. Schimansky-Geier, Phys Rev Lett 86,
2934 (2001).
[12] B. Naundorf, F. Wolf, M. Volgushev, Nature 440, 1060
(2006).
[13] L. Badel et al., J. Neurophysiol. 99, 656 (2008).
[14] See supplemental material at http://link.aps.org/ sup-
plemental/10.1103/PhysRevLett.106.088102.
∂D and saturates to a constant ν0
[15] For LIF model, ν1n(ω) starts at zero signal frequency
with a value ∂ν0
D at large
frequency. So in our model there will be a transition be-
tween these two plateaus before the decaying of transmis-
sion function, since in large r limit it reduces to the LIF
model. To characterize the decaying property and avoid
the transition region, we extract the cut-off frequencies
at which the normalized transmission function decays to
1/√10, which is smaller than the value usually adopted.
[16] H. Plesser and T Geisel, Phys. Rev. E 59, 7008 (1999).
[17] D. McCormick, Y. Shu, Y. Yu, Nature 445:E1-E2(2007).
[18] B. Naundorf, F. Wolf, M. Volgushev, Nature 445:E2-E3
(2007).
[19] M. Royeck et al., J. Neurophysiol. 100, 2361 (2008).
|
1507.05970 | 1 | 1507 | 2015-07-21T20:05:42 | Chaotic Neuronal Oscillations in Spontaneous Cortical-Subcortical Networks | [
"q-bio.NC"
] | Oscillatory activities are widely observed in specific frequency bands of recorded field potentials in different brain regions, and play critical roles in processing neural information. Understanding the structure of these oscillatory activities is essential for understanding the brain function. So far many details remain elusive about their rhythmic structures and how these oscillations are generated. We show that many oscillatory activities in spontaneous cortical-subcortical networks, such as delta, spindle, gamma, high-gamma and sharp wave ripple bands in different brain regions, are genuine chaotic time series which can be reconstructed as chaotic attractors through appropriately selected embedding delay and dimension. The reconstructed attractors are approximated by a simple radial basis function enabling high precision short-term prediction. Simultaneously recorded oscillatory activities in multiple brain regions differ greatly in term of temporal phase and amplitude but can be approximated by the same function. Our results suggest that neural oscillations are produced by deterministic chaotic systems. The occurrence of neural oscillation events is predetermined, and the brain possibly knows when and where the information will be processed and transferred in the future time as a result of the deterministic dynamic. | q-bio.NC | q-bio |
Chaotic Neuronal Oscillations in Spontaneous Cortical-
Subcortical Networks
Pengsheng Zheng
[email protected]
[email protected]
Abstract
Oscillatory activities are widely observed in specific frequency bands of recorded
field potentials in different brain regions, and play critical roles in processing neu-
ral information. Understanding the structure of these oscillatory activities is es-
sential for understanding the brain function. So far many details remain elusive
about their rhythmic structures and how these oscillations are generated. We show
that many oscillatory activities in spontaneous cortical-subcortical networks, such
as delta, spindle, gamma, high-gamma and sharp wave ripple bands in different
brain regions, are genuine chaotic time series which can be reconstructed as chaotic
attractors through appropriately selected embedding delay and dimension. The re-
constructed attractors are approximated by a simple radial basis function enabling
high precision short-term prediction. Simultaneously recorded oscillatory activities
in multiple brain regions differ greatly in term of temporal phase and amplitude
but can be approximated by the same function. Our results suggest that neural os-
cillations are produced by deterministic chaotic systems. The occurrence of neural
oscillation events is predetermined, and the brain possibly knows when and where
the information will be processed and transferred in the future time as a result of
the deterministic dynamic.
1
Introduction
Our brains are foretell machines with advanced predictive powers which are suggested
to emerge from various neuronal oscillations (Buzsaki, 2006). These neuronal oscilla-
tory activities prevail in cortical filed potentials in different behavior contexts demon-
strating a number of characteristic oscillatory frequency bands, which are crucial for
effective information processing within and across distinct brain structures (Buzsaki,
2006; Engel et al., 2001; Singer, 1999). Neural oscillations and synchronization have
been linked to many cognitive functions such as sleep and consciousness, perception,
motor coordination and memory (Fell and Axmacher, 2011; Schnitzler and Gross, 2005;
Fries, 2005).
Recent findings indicated that network oscillations temporally link neurons into as-
semblies and facilitate synaptic plasticity (Buzsaki and Draguhn, 2004). The gamma
cycle has been suggested to serve as a computational mechanism for the implementation
of a temporal coding scheme that enabled fast processing and flexible routing of activity,
supporting selection and binding of distributed responses (Fries et al., 2007). Power-
frequency and spatial coherence analyses showed that gamma oscillations in hippocam-
pal CA1 area split into distinct fast and slow frequency components routing the flow of in-
formation (Colgin et al., 2009). Spontaneously occurring oscillatory events like thalamo-
cortical sleep spindles and hippocampal sharp wave ripple complexes (SPW-R, 140-
250Hz) during slow wave sleep (SWS) were crucial for hippocampal-cortical interactions
during memory consolidation (Wilson and Mcnaughton, 1994; Siapas and Wilson, 1998;
1
Sirota et al., 2003; Diba and Buzsaki, 2007; Eschenko et al., 2008; Girardeau et al., 2009;
Logothetis et al., 2012). Understanding how these oscillations are generated and their
rhythmic structures are fundamental and critical for the understanding of brain function.
In this paper, we use two open access data sets, in one of which rats were implanted
with silicon probes to monitor both LFP and unit firing in different hippocampal regions,
such as CA1, entorhinal cortex layer 2 (EC2) and layer 5 (EC5) while they were sleep-
ing (Mizuseki et al., 2014). In the other data set, the recordings were made from anterior
thalamus (AT), hippocampus CA1 and mPFC of sleeping mice (Peyrache and Buzski,
2015). The simultaneously recorded field potentials in different regions teem with com-
plex rhythmic structures which differ widely in term of phase, amplitude and temporal
resolution. However, we show that these diverse rhythmic structures come from deter-
ministic systems, and there is no random components in it.
2 Reconstruction of neuronal oscillatory activity
We firstly band-pass filter the LFP to certain frequency range. Unless otherwise stated,
we use rat CA1 spindle (12-15Hz) band activity during SWS (ec014.716.dat (Mizuseki et al.,
2014)) as an example, and the results were extended to many other oscillatory activities
(see Fig.S1-S12 in supplementary material (SM)). The spindle band activity gives us the
time sequence of the form [X(t0), X(t1), · · · , X(t)], where X(t) denotes the oscillation
voltage at time t (Fig.1A). Following the delay coordinate embedding method (Takens,
1981), the reconstructed attractor of the original system is given by the vector sequence
P (t) = [X(t), X(t − τ ), X(t − 2τ ), · · · , X(t − (d − 1)τ )],
(1)
where τ and d are the embedding delay and embedding dimension respectively. The
delay embedding theorem states that this procedure provides a one-to-one image of the
original system for a large enough d (Takens, 1981).
The embedding delay (τ ) is estimated by mutual information method which measures
how much one knows about X(t−τ ) from X(t) (Fraser and Swinney, 1986). In principal,
τ has to be large enough so that the information we get from measuring X(t − τ ) is
significantly different from that of X(t), but τ should not be larger than the typical
time in which the system loses information of its initial state. If τ is chosen larger than
it should, then the reconstructed attractor will be folded and points will look more or
less random since they will be uncorrelated. In practice, the τ can be estimated as the
point where the average mutual information (in bits) is close to but before its first local
minimum (τ = 0.015 in Fig.1B).
The embedding dimension (d) is the minimum dimension of the space in which
the trajectory does not statistically cross itself. As shown in Fig.1C, the embedding
dimension is d = 5 suggested by Cao's method (Cao, 1997). We then reconstruct the
5-dimensional attractor and project it to 3-D (Fig.1D) and 2-D (Fig.1E) spaces, which
are reminiscent of a chaotic attractor (compare Lorenz attractor Fig.S19 in SM).
After reconstructing the attractor, we perform the determinism test to verify if the
studied time series are indeed from a deterministic system (Kaplan and Glass, 1992).
The determinism test constructs the vector field of the system directly from the time
series, and subsequently tests if the reconstructed vector field assures uniqueness of
solutions in the phase space (Kaplan and Glass, 1992), which enables us to distinguish
between deterministic chaos and irregular random behavior that often resembles chaos.
The calculated determinism measure for reconstructed spindle attractor is 0.99 (1 for
perfectly deterministic), which means the system is deterministic.
2
2
4
6
10
Time (s)
8
B
C
)
d
(
1
E
0.8
0.6
0.4
0.2
12
14
16
18
20
0.015
0.03
0.045
τ
1
2
3
4
5
6
7
8
9
10
d
0.5
A
V
m
0
−0.5
0
t
i
B
6
4
2
0
D
)
τ
2
−
t
(
X
0.5
0.4
0.3
0.2
0.1
0
−0.1
−0.2
−0.3
−0.4
−0.5
E
)
τ
2
−
t
(
X
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.6
−0.4
−0.2
0
X(t−τ)
0.2
0.4
0.6
0.5
0
X(t)
0.5
0.4
0.3
0.2
0.1
0
−0.1
−0.2
−0.3
−0.4
−0.5
−0.5
X(t−τ)
Figure 1: Reconstruction of spindle band activity. A: Time series of CA1 spindle. B:
Average mutual information in dependence on embedding delay (τ ). C: E1(d), the
Cao's-variation from d to d + 1, as a function of embedding dimension. D, E: Recon-
structed attractor projected to 3-D and 2-D spaces respectively.
3 Neuronal oscillatory activity approximation and predic-
tion
Following the line of traditional method for chaotic time series prediction, thousands
of input P (t) and output X(t + τ ) pairs are collected as t increases from time 0 to tp,
and tp represents the ending time of training time series. It then reduces to a function
approximation problem. Since the system is deterministic, it should be able to find a
function to approximate all the input-output pairs. We here use Radial Basis Function
Network (RBF) to approximate the delicate input-output function (Park and Sandberg,
1991). The RBF network has simple one-hidden layer structure and fast training process,
and most importantly it has a more intuitive mathematical interpretation than the
multilayer perceptron as follows,
X(t + τ ) = LW · Radbas(kIW − P (t)k · b1) + b2,
(2)
where IW is input weight matrix, LW is weight matrix of output layer, b1 and b2 are
bias terms of input and output layers respectively, k·k represents the Euclidean distance,
and
2
Radbas(x) = ex
.
Here we use 'newrb' function in the Matlab neural network toolbox (Beale et al., 2013)
to find proper parameters IW , LW , b1 and b2 for equation (2). The RBF network
reaches a small goal of training error (MSE = 1e − 6) in a few iterations implying all the
input-output pairs are well-fitted by the same network. The weight matrixes and bias
terms of RBF network are frozen after training.
3
1
A
C
0.3
0.2
0.1
0
-0.1
-0.2
-0.3
2
C
E
0.1
0.05
0
-0.05
-0.1
5
C
E
0.3
0.2
0.1
0
-0.1
-0.2
-0.3
A
0
0.5
1
1.5
MSE=8.76e-06
2
2.5
3
3.5
B
0
0.5
1
1.5
MSE=1.21e-06
2
2.5
3
3.5
C
0
0.5
1
1.5
MSE=6.24e-06
2
2.5
3
3.5
s
Figure 2: Short-term prediction of spindle band activity. A-C, Comparison of predicted
(red) and actual (blue) signals in CA1, EC2 and EC5 regions respectively. Note that
all the predictions are performed by the network designed for CA1 spindle. MSE is the
mean squared error.
The prediction is applicable in real time, we are now able to predict the future value
X(t + τ ), t > tp, by using P (t) and the frozen equation (2). For short-term prediction,
P (t) is always from real signal, but none of these data is exposed to the training of
RBF network. Figure 2A shows an example of the predicted X(t + τ ), the real signal
(blue curve) disappears under the predicted one (red curve) as a result of high-accuracy
prediction (MSE = 1.95e − 8). Many other spindle band activities collected from the
same rat at different time were also tested. The network learned the data structure
from a piece of signal and generalized to unknown signals. In other words, the spindle
band oscillatory activity, no matter in the past or future, can be fitted into the same
network (i.e. frozen equation (2)), which again proves the spindle activity is dominated
by a deterministic system, and there is no irregular or random components in it.
With the aforementioned method, we further verify that EC2 and EC5 spindle band
activities are all developed by a deterministic chaotic system with the same embed-
ding delay and dimension as that of the CA1 region. More interestingly, the EC2 and
EC5 spindle band activities are accurately predicted by the RBF network intently de-
signed for CA1 spindle time series (Fig.2B-C). Similarly, an RBF network trained by EC
spindle band time series is capable of predicting CA1 spindle band activity. These all
suggest that spindle band activities in different regions are dominated by deterministic
dynamical systems which share similar or even identical structure.
How could we reconcile the tremendous pointwise differences of simultaneously recorded
multi-region signals with the facts that they were approximated by the same function?
Take two identical Lorenz systems for example, the difference in initial states leads to
vastly different trajectories (See Fig.S20 in SM). This implies the dynamical systems
4
v
m
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2
−0.25
0.05
0.1
0.15
t(s)
0.2
0.25
0.3
0.05
0.1
0.15
t(s)
0.2
0.25
0.3
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
0.05
0.1
0.15
t(s)
0.2
0.25
0.3
0.05
0.1
0.15
t(s)
0.2
0.25
0.3
Figure 3: Examples of long-term iterative prediction of CA1 spindle activity. Red and
blue curves represent the predicted and real signals respectively. Note that none of these
data is exposed to the training of the network.
in different regions might be identical in structure but operated with different initial
states.
Same argument, Gamma band activities in these three regions have the same em-
bedding delay and dimension, and can be approximated by the same network intently
designed for one region. The results can also extend to Delta, High-Gamma and SPW-R
band activities with different dynamical systems developed for different frequency bands
(see Fig.S1-S12 in the SM).
In long-term prediction, the predicted value X(t + τ ) can be further applied to the
prediction of X(t + 2τ ), and so forth. Unlike the short-term prediction, P (t) (t >
tp + τ ) now is composed of predicted values rather than real signals. Figure 3 shows
four examples of long-term iterative prediction of CA1 spindle band activity by the
same frozen network.
In all cases, the network achieves high-precision prediction at
the beginning, and the performances extend for a certain amount of time followed by
decreasing precision as time elapses. This is consistent with the observation of typical
chaotic systems in which long term prediction is only theoretically plausible within a
limited period of time due to its exponential divergence nature.
All the results are further verified on the open access data set recorded in Hippocampal-
mPFC-thalamic network of sleeping mice (Peyrache and Buzski, 2015), an RBF network
is designed with the time series from one brain region and then applied to the prediction
of signals in other brain regions. Figure 4 shows the results of short-term prediction of
spindle band activity. More technical details and the results of other frequency bands
are reported in Fig.S16-S17 (SM). Unless otherwise stated, all the results are retrieved
during SWS state, but the network designed during SWS state is capable of predicting
the signal in REM state (Fig.S18 SM). Matlab code is provided in the SM for easy and
fast test of users' own neural signal under different conditions.
5
C
P
H
0.2
0.1
0
−0.1
−0.2
0
C
F
P
m
0.1
0.05
0
−0.05
−0.1
0
T
A
0.2
0.1
0
−0.1
−0.2
0
0.5
0.5
0.5
1
1
1
1.5
1.5
1.5
2
MSE=1.50e−06
2
MSE=3.53e−07
2
MSE=7.18e−07
2.5
2.5
2.5
3
3
3
3.5
3.5
3.5
4
4
4
(s)
Figure 4: Short-term prediction of spindle band activity in sleeping mouse (Mouse12-
120807). Comparison of predicted (red) and actual (blue) time series in hippocampal
(HPC) CA1, mPFC and AT regions respectively. Note that all the predictions are
performed by the network designed for HPC CA1 spindle.
4 Conclusion and discussion
Many cognitive functions of the brain are suggested to be emerged from neural oscil-
lations, but where do these complex oscillations come from? Here we show that these
diverse rhythmic activities, in different brain regions but in the same frequency band,
come from a unified deterministic system, and there is no random components in it. The
occurrence of certain kinds of neural oscillation events is predetermined by the deter-
ministic dynamic rather than randomly or statistically formed. In this sense, the brain
has perfect control of local and global information flow, which is ideal for information
transfer, avoidance and collaboration between distinct regions.
There is a long history of chaotic system study in physics and many interdisciplinary
areas with numerous proposed theorems and applications. Now the whole theory is
potentially open to neural oscillation research, and many interdisciplinary topics are
expected to be explored in the future. On the other hand, neural oscillations have been
suggested to play critical roles in many cognitive functions such as sleep, conscious-
ness, perception and memory. The results of paper provide potential new perspectives
to many of these researches. For example, among all the topics of chaos theory, the
synchronization of multiple chaotic attractors is one of the hottest topics. As men-
tioned above, two identical Lorenz systems display vastly different trajectories when
they are operated at different initial states (Fig.S21 in SM). However, the two systems
can be fast synchronized if a negative feedback is applied (Fig.S15 in SM). The syn-
chronization level is further adjustable by tuning the strength of the negative feedback
which might be analogue to the changes of network connectivity wildly observed in
epilepsy (David et al., 2008; Zhang et al., 2011; Clemens et al., 2013).
Integration of
the results with the epilepsy research might be one of the interesting studies in the
future.
6
References
Beale, M., Demuth, H., and Hagan, M. (2013). Neural Network Toolbox User's Guide.
The MathWorks Inc.
Buzsaki, G. (2006). Rhythms of the brain. Oxford University Press.
Buzsaki, G. and Draguhn, A. (2004). Neuronal oscillations in cortical networks. Science,
304(5679):1926 -- 1929.
Cao, L. Y. (1997). Practical method for determining the minimum embedding dimension
of a scalar time series. Physica D, 110(1-2):43 -- 50.
Clemens, B., Puskas, S., Besenyei, M., Spisak, T., Opposits, G., Hollody, K., Fogarasi,
A., Fekete, I., and Emri, M. (2013). Neurophysiology of juvenile myoclonic epilepsy:
EEG-based network and graph analysis of the interictal and immediate preictal states.
Epilepsy Research, 106(3):357 -- 369.
Colgin, L. L., Denninger, T., Fyhn, M., Hafting, T., Bonnevie, T., Jensen, O., Moser,
M. B., and Moser, E. I. (2009). Frequency of gamma oscillations routes flow of
information in the hippocampus. Nature, 462(7271):353 -- U119.
David, O., Guillemain, I., Saillet, S., Reyt, S., Deransart, C., Segebarth, C., and De-
paulis, A. (2008). Identifying neural drivers with functional MRI: An electrophysio-
logical validation. Plos Biology, 6(12):2683 -- 2697.
Diba, K. and Buzsaki, G. (2007). Forward and reverse hippocampal place-cell sequences
during ripples. Nature Neuroscience, 10(10):1241 -- 1242.
Engel, A. K., Fries, P., and Singer, W. (2001). Dynamic predictions: Oscillations and
synchrony in top-down processing. Nature Reviews Neuroscience, 2(10):704 -- 716.
Eschenko, O., Ramadan, W., Molle, M., Born, J., and Sara, S. J. (2008). Sustained in-
crease in hippocampal sharp-wave ripple activity during slow-wave sleep after learning.
Learning & Memory, 15(4):222 -- 228.
Fell, J. and Axmacher, N. (2011). The role of phase synchronization in memory pro-
cesses. Nature Reviews Neuroscience, 12(2):105 -- 118.
Fraser, A. M. and Swinney, H. L. (1986). Independent coordinates for strange attractors
from mutual information. Physical Review A, 33(2):1134 -- 1140.
Fries, P. (2005). A mechanism for cognitive dynamics: neuronal communication through
neuronal coherence. Trends in Cognitive Sciences, 9(10):474 -- 480.
Fries, P., Nikolic, D., and Singer, W. (2007). The gamma cycle. Trends in Neurosciences,
30(7):309 -- 316.
Girardeau, G., Benchenane, K., Wiener, S. I., Buzsaki, G., and Zugaro, M. B. (2009).
Selective suppression of hippocampal ripples impairs spatial memory. Nature Neuro-
science, 12(10):1222 -- 1223.
Kaplan, D. T. and Glass, L. (1992). Direct test for determinism in a time-series. Physical
Review Letters, 68(4):427 -- 430.
7
Logothetis, N. K., Eschenko, O., Murayama, Y., Augath, M., Steudel, T., Evrard, H. C.,
Besserve, M., and Oeltermann, A. (2012). Hippocampal-cortical interaction during
periods of subcortical silence. Nature, 491(7425):547 -- 553.
Mizuseki, K., Sirota, A., Pastalkova, E., Diba, K., Peyrache, A., and Buzski, G.
(2014). Extracellular recordings from multi-site silicon probes used for cluster-
ing of neuron responses in rat hippocampal and entorhinal regions. CRCNS.org,
http://dx.doi.org/10.6080/K0QJ7F75.
Park, J. and Sandberg, I. W. (1991). Universal approximation using radial-basis-function
networks. Neural Computation, 3(2):246 -- 257.
Peyrache, A. and Buzski, G. (2015). Extracellular recordings from multi-site silicon
probes in the anterior thalamus and subicular formation of freely moving mice. CR-
CNS.org, http://dx.doi.org/10.6080/K0G15XS1.
Schnitzler, A. and Gross, J. (2005). Normal and pathological oscillatory communication
in the brain. Nature Reviews Neuroscience, 6(4):285 -- 296.
Siapas, A. G. and Wilson, M. A. (1998). Coordinated interactions between hippocampal
ripples and cortical spindles during slow-wave sleep. Neuron, 21(5):1123 -- 1128.
Singer, W. (1999). Neuronal synchrony: A versatile code for the definition of relations?
Neuron, 24(1):49 -- 65.
Sirota, A., Csicsvari, J., Buhl, D., and Buzsaki, G. (2003). Communication between
neocortex and hippocampus during sleep in rodents. Proceedings of the National
Academy of Sciences of the United States of America, 100(4):2065 -- 2069.
Takens, F. (1981). Detecting strange attractors in turbulence. Lecture Notes in Mathe-
matics, 898:366 -- 381.
Wilson, M. A. and Mcnaughton, B. L. (1994). Reactivation of hippocampal ensemble
memories during sleep. Science, 265(5172):676 -- 679.
Zhang, Z. J., Valiante, T. A., and Carlen, P. L. (2011). Transition to seizure: From
"macro"- to "micro"-mysteries. Epilepsy Research, 97(3):290 -- 299.
8
|
1804.00153 | 2 | 1804 | 2018-04-06T09:06:53 | Spontaneous activity emerging from an inferred network model captures complex temporal dynamics of spiking data | [
"q-bio.NC"
] | The combination of new recording techniques in neuroscience and powerful inference methods recently held the promise to recover useful effective models, at the single neuron or network level, directly from observed data. The value of a model of course should critically depend on its ability to reproduce the dynamical behavior of the modeled system; however, few attempts have been made to inquire into the dynamics of inferred models in neuroscience, and none, to our knowledge, at the network level. Here we introduce a principled modification of a widely used generalized linear model (GLM), and learn its structural and dynamic parameters from ex-vivo spiking data. We show that the new model is able to capture the most prominent features of the highly non-stationary and non-linear dynamics displayed by the biological network, where the reference GLM largely fails. Two ingredients turn out to be key for success. The first one is a bounded transfer function that makes the single neuron able to respond to its input in a saturating fashion; beyond its biological plausibility such property, by limiting the capacity of the neuron to transfer information, makes the coding more robust in the face of the highly variable network activity, and noise. The second ingredient is a super-Poisson spikes generative probabilistic mechanism; this feature, that accounts for the fact that observations largely undersample the network, allows the model neuron to more flexibly incorporate the observed activity fluctuations. Taken together, the two ingredients, without increasing complexity, allow the model to capture the key dynamic elements. When left free to generate its spontaneous activity, the inferred model proved able to reproduce not only the non-stationary population dynamics of the network, but also part of the fine-grained structure of the dynamics at the single neuron level. | q-bio.NC | q-bio |
Spontaneous activity emerging from an inferred network
model captures complex temporal dynamics of spiking data
Cristiano Capone*,1, 2, a) Guido Gigante*,1, 3 and Paolo Del Giudice1, 2
1)Istituto Superiore di Sanit`a, Rome, Italy
2)INFN, Sezione di Roma 1, Italy
3)Mperience s.r.l., Rome, Italy
The combination of new recording techniques in neuroscience and powerful inference
methods recently held the promise to recover useful effective models, at the single
neuron or network level, directly from observed data. The value of a model of
course should critically depend on its ability to reproduce the dynamical behavior
of the modeled system; however, few attempts have been made to inquire into
the dynamics of inferred models in neuroscience, and none, to our knowledge, at
the network level. Here we introduce a principled modification of a widely used
generalized linear model (GLM), and learn its structural and dynamic parameters
from ex-vivo spiking data. We show that the new model is able to capture the most
prominent features of the highly non-stationary and non-linear dynamics displayed
by the biological network, where the reference GLM largely fails. Two ingredients
turn out to be key for success. The first one is a bounded transfer function that
makes the single neuron able to respond to its input in a saturating fashion; beyond
its biological plausibility such property, by limiting the capacity of the neuron
to transfer information, makes the coding more robust in the face of the highly
variable network activity, and noise. The second ingredient is a super-Poisson
spikes generative probabilistic mechanism; this feature, that accounts for the fact
that observations largely undersample the network, allows the model neuron to
more flexibly incorporate the observed activity fluctuations. Taken together, the
two ingredients, without increasing complexity, allow the model to capture the key
dynamic elements. When left free to generate its spontaneous activity, the inferred
model proved able to reproduce not only the non-stationary population dynamics
of the network, but also part of the fine-grained structure of the dynamics at the
single neuron level.
Dynamic models, in neuroscience and in general, embody in a mathematical form the
causal relationships between variables deemed essential to describe the system of interest;
of course, the value of the model is measured both by its ability to match observations,
and by its predictive power. Typically, along this route, parameters appearing in the model
are assigned through a mix of insight from experiments and trial-and-error and, in turn,
the study of the model dynamics helps understanding the relevance of each parameter in
determining the dynamic regimes accessible to the system. Another approach, initially
quite detached from dynamic modeling, is rooted in the domain of statistical inference,
and a seminal example in neuroscience was offered by the application of maximum-entropy
inference of Ising-like models to multi-electrode recordings of neural activity; in this case, as
well as in the later extension to kinetic Ising-like models, the Montecarlo/Glauber dynamics
of the model is only meant as a means to sample the probability distribution of interest and
it is not claimed to offer a detailed model of the actual dynamics at work in the system.
Recently, Generalized Linear Models (GLM) (which incorporate kinetic Ising models as a
special case) have been recognized as flexible and powerful inference models1,2. In time,
efforts have been made to make contact between the two approaches, e.g., in the case
of neuroscience, by endowing the coupling structure of the inference model with features
motivated by biological plausibility3–5; it has also been recognized that a GLM is close
to a stochastic version of the spike-response model. Recently, the repertoire of the driven
a)Electronic mail: [email protected]
*authors contributed equally
2
dynamics of GLM models of single neurons has been explored6; however, to our knowledge, a
largely open issue is to endow inference models with predictive power in terms of the system
dynamics. Our approach to this problem is to explore the free, spontaneous dynamics of the
inferred model in its relation with the one of the biological system generating the data and,
in doing this, to identify the role of different elements of the inference model in determining
the spontaneous dynamics of the neuronal network.
Indeed, an obvious but persisting
problem in the application of inference models to neuroscience has been how to assess the
meaning and value of the inferred parameters; a recurring example is offered by the inferred
synaptic couplings3: in the literature, a cautionary remark is usually included, recognizing
that the inferred couplings (whether in the form of synaptic efficacies of more complicated
synaptic kernels) are to be meant as 'effective', leaving of course open the problem of what
exactly this means (especially in the face of the dramatic subsampling of the underlying
biological network). The value of the inferred model cannot be assessed directly by a detailed
correspondence between its elements and corresponding elements of the biological system;
using the output of a GLM to decode input stimuli is an interesting recent approach7 Our
attitude is that spontaneous activity is a good testing ground to assess whether the inferred
model does indeed capture essential features of the biological system, and we choose a case
in which the spontaneous activity of the biological network is highly non-stationary and
irregular: a cultured neuronal network generating spontaneously a wide spectrum of activity
fluctuations, from population bursts, to neuronal avalanches and noisy oscillations8–20. Such
richness is extremely hard to reproduce with a GLM, and to our knowledge no models
proposed so far were able to cope with it. For this reason we chose population bursting
activity as a challenging benchmark to test our model. In our GLM approach we introduced
novel ingredients inspired by biological observation, such as activity dependent negative
feedback over different time-scales for the single neuron8 (spike frequency adaptation, SFA),
a bounded transfer function, and a generative probabilistic mechanism for spike generation
with super-Poisson fluctuations. We show that these ingredients are crucial to endow the
system with non-stationary spontaneous activity and also to account for detailed dynamic
features such as temporal correlations and spatio-temporal evolution of network bursts.
RESULTS
Bounded firing rate, super-Poisson spike generation, and their impact on the spontaneous
activity of the inferred model
The model we introduce is an extension of GLMs as widely employed for inferring the
structure of neuronal networks. With GLMs it shares the three fundamental assumptions
that spike generation is a probabilistic process driven by the instantaneous current afferent
to the neuron; that, secondly, such current is a linear function of the activity of the neurons
in the network and, finally, that there is a non-linear transfer function linking the current
to the probabilistic spike generation. The parameters of the model, in this framework,
are usually obtained through the numerical maximization of the likelihood that the model
itself generated the observed data. The main points of departure between more standard
GLMs and the present model lay in the choice of the specific non-linear transfer function
and probabilistic generative model; another notable difference is in the introduction, at
the single neuron level, of an activity-dependent self-inhibition current, mimicking spike-
frequency adaptation (SFA) effects widely observed in real neurons.
Fig. 1, panel A, gives a schematic view of the model. The current Hi(t) instantaneously
felt by neuron i sums spikes emitted by other neurons j in the recent past (t− ∆t) through
synaptic kernels kij(∆t), a constant external field hi, and a set of 5 SFA signals (collectively
indicated as ci(t)), that integrate, over different time-scales, the spiking activity Si of neuron
i itself. This current provides input to a non-linear transfer-function f [Hi], which in turn
determines the expected firing rate λi(t) for the spike generation model at the next time
bin t + dt. The transfer function and generative model most commonly employed in the
3
FIG. 1: Spontaneous activity of the inferred model reproduces non-stationary network
dynamics. A Schematic representation of the model. The current Hi(t) felt by neuron i sums
spikes emitted by other neurons j in the recent past (t − ∆t) through synaptic kernels kij(∆t), a
constant external field hi, and spike-frequency adaptations signals that integrate, over different
time-scales, the spiking activity Si of neuron i itself. This current, fed into a non-linear
transfer-function f [Hi], determines the expected number λi(t) of spikes Si(t + dt) that the neuron
will probabilistically generate at the next time bin. B Rastergrams of ex-vivo cultured cortical
neurons show a strongly non-stationary dynamics (top row; red line: whole network activity).
The proposed Sig-NegBin model (central row, blue line) closely follows the data used to drive its
dynamics (red line, t between 20 and 30 s); when left free to run autonomously at t = 30 s, the
model displays large irregular bursts of activity, qualitatively mimicking the behavior observed in
the recordings. The standard Exp-Poisson model (bottom row, blue line), on the other hand,
whilst still able to follow the data in driven mode, fails to produce any large activity fluctuation,
settling instead into a noisy, stationary state. C Comparison between burst amplitudes (total
spike counts, left), bursts durations (center) and IBI durations (right) histograms both for the
data (top row) and the autonomous activity of the Sig-NegBin model (bottom row). Mean and
standard deviation of each distribution are reported with red and green lines respectively. Results
for the standard Exp-Poisson model are not shown because of the extreme sparsity of burst
events in that case. D Average area under the inferred synaptic kernels (absolute value) as a
function of the distance separating the electrodes; blue and red circles are for excitatory and
inhibitory kernels respectively; dashed lines: exponential fits; the strength (area under the kernel)
of inferred excitatory synapses decays faster than that of the inhibitory ones.
4
literature are, respectively, the exponential function and the Poisson distribution; in the
following we will label this choice 'Exp-Poisson'; our model, on the other hand, makes use
of a sigmoid function (see Materials and Methods) and of a Negative-Binomial distribution
with shape parameter rNB set to 0.2, and will be referred to in the following as 'Sig-NegBin'
model.
The synaptic kernels kij(∆t), following4, are parametrized as a weighted sum of 4 'raised
cosine' functions peaking at times spanning a range between 10 ms (equal to the chosen
time bin dt) and 30 ms, with support extending in the past for 15 dt = 150 ms (kij(∆t) ≡ 0
for ∆t > 150 ms). The sigmoid transfer function has two free parameters: the maximum
firing rate attainable by the neuron and an asymmetry parameter that determines how fast
this maximum is approached for high currents compared to the other asymptote at 0 for
strongly negative currents. Finally each of the 5 SFA signals ci, that are set equal for each
neuron, has associated a characteristic time-scale τc for the integration of the spike train
Si(t) and a weight gc with which ci(t) enters the sum of the current Hi(t). These latter
parameters, the weights defining each of the kij(∆t) functions, and the free parameters of
the sigmoid function are all determined through the likelihood maximization procedure on
data (see Materials and Methods for further details). In the case of the Exp-Poisson model,
the only optimized parameters are the kij(∆t) weights, both for i (cid:54)= j and i = j ('post-spike
filter'); the kii functions are parametrized as a weighted sum of 10 raised cosines peaking
at times spanning from 10 ms to 50 ms and again extending to 150 ms in the past; the Sig-
NegBin model does not include post-spike filters; thus the total number of free parameters
for the Sig-NegBin is actually lower than that for the Exp-Poisson model.
Once the parameters are fitted by maximizing the likelihood, both models can be simu-
lated to generate new spike trains according to two basic simulation modalities. The first
one ('driven' mode) feeds the model with the actual spike trains observed in the data and
at each time t uses the model's output to make a 1-step prediction of the network activity
at step t + dt; thus the generated activity never re-enters the network dynamics; the prob-
abilistic distance between the 1-step prediction and the actual recorded network activity is
ultimately what is minimized during the optimization procedure. In the second modality
('free' mode) the models are left free to evolve and the generated spike trains are fed back
into the network, driving its dynamics at subsequent times.
We trained both the Sig-NegBin and the Exp-Poisson model on data from multi-electrode
recordings of ex-vivo cultured cortical neurons10, that show a clear non-stationary dynamics
(Fig. 1B, top row) in the rastergrams and in the whole network activity (red line), with
large and irregular bursts of synchronous activity interleaved by periods of low, noisy activity
('inter-burst intervals' or IBIs). At the end of the training, both models show, in driven
mode, an activity that follows quite closely the experimental spike trains (Fig. 1B, middle
and bottom rows; red line: data; blue line and rastergrams: simulation). However, when the
simulation is switched to free mode at time t = 30 s, the behavior of the two models clearly
diverges: whilst the Sig-NegBin is able to autonomously sustain a highly non-stationary
activity, the Exp-Poisson only exhibits stationary fluctuations. Note that, in free mode, the
activity of the Sig-NegBin model does not closely follow the activity of the real network
anymore (top row, last 10 s); this is not surprising: the choice of a probabilistic model
as a GLM stems from the assumption that the network dynamics is, to a large extent,
influenced by self-induced random fluctuations and, thus, non-deterministic; it is therefore
expected that, at least after a transient phase, the model and the data will drift away. Yet
the activity spontaneously generated by the model clearly resembles the one seen in data
in an intuitive statistical way, showing irregular bursts interspersed with interval of relative
quiescence.
Such resemblance is made more quantitative in Fig. 1C. Bursts of the network global
activity are detected through the algorithm developed in8, both for the data (top row) and
the autonomous activity of the Sig-NegBin model (bottom row). The comparison between
burst amplitudes (total spike counts, left), bursts durations (center) and IBI durations
(right) histograms shows a semi-quantitative agreement between the dynamics of the bio-
logical network and the inferred model (mean and standard deviation of each distribution
5
FIG. 2: Bounded spike rate is necessary to generate non-stationary spontaneous activity in the
inferred model. A Population activity (blue line) and rastergram generated by the model where
the bounded transfer function is replaced by an exponential and inference procedure is carried
out from scratch. Red line represent the data driving the dynamics for the first 10 s (20 ≤ t ≤ 30
s); afterwards, when left free to generate its spontaneous activity, the model is unable to sustain
any non-stationary bursting activity and settles in a quiescent state. B-C transfer function
(respectively sigmoid and exponential) f [·] (top-right panel) distribution of input currents Hi(t)
to the single neuron across different times (bottom panel) and distribution of firing rates
λi(t) = f [Hi(t)] ( top-left panel)
are reported with red and green lines respectively). The main discrepancy between model
and data is visible in the IBI distribution; "doublets" of bursts in the data, clearly separated
by the rest of the IBI distribution, are absent in the model. Based on previous modeling
work on similar data21, the discrepancy is likely due to elements, we did not include in our
minimal model, such as short-term synaptic facilitation and depression.
We remark that the analogous distributions for the Exp-Poisson model are not reported
because of the extreme sparsity of burst events in that case. Moreover we observe that,
despite the major difference in the exhibited dynamic behavior, the inferred synaptic kernels
for our model and the Exp-Poisson model are correlated (cPearson = 0.49 with a P < 0.05;
correlation between the total areas under each synaptic kernel, as a measure of 'synaptic
strength'). The inferred synaptic kernels also show a correlation with the distance separating
the electrodes where the activity of the pre- and post-synaptic neurons where recorded
(Fig. 1D; blue and red circles are for excitatory and inhibitory kernels respectively; dashed
lines: exponential fits).
It is worth noting that the connections' decay with distance is
faster for excitatory than for inhibitory inferred synapses, consistently with what is found
in activity correlations in the neocortex by using 2D multielectrode arrays22.
The choice of a sigmoidal, saturating transfer function, besides seeming just natural in a
GLM meant to model neural activities, turns out to be key in achieving the above results.
We repeated the inference procedure with a hybrid Sig-NegBin model, in which the sigmoidal
transfer function is replaced with an exponential (Fig. 2A). As above, when during the first
6
10 s the model is simulated (blue line and rastergrams) in driven mode, it shows a good
agreement with the driving data (red line); when the driving is turned off, instead, the
model settles in a quiescent state, absent any large bursts of activity. In free mode, the
saturation of the sigmoidal function f [·] (Fig. 2B, top-right panel) allows a wide distribution
of input currents Hi(t) to the single neuron across different times (Fig. 2B, bottom panel)
to give rise to a distribution of firing rates λi(t) = f [Hi(t)] (Fig. 2B, top-left panel) that is
strongly bimodal (notice the logarithmic scale on the y-axis). On the other hand, the fast
growth of the exponential function (Fig. 2C, top-right panel) pushes the inference procedure
towards somewhat narrower, more conservative distribution of currents (Fig. 2C, bottom
panel), resulting in a unimodal distribution of firing rates. The comparison of the two
models in driven and free modes suggests that the exponential transfer function makes the
system very selectively susceptible to the fluctuations experienced during learning, while the
sigmoid transfer function allows for susceptibility to a wider range of fluctuations, which
we believe is the root of the ability to spontaneously generate network bursts.
The recorded neurons from the cultured network constitute of course a dramatic sub-
sampling of the network (Fig. 3A, leftmost sketch; though, in fact, the activity of several
neurons will be collected by each electrode). The effects of the unobserved neurons on the
behavior of the observed part of the network could be in principle very complex23; yet we
argue that the adoption of a Negative Binomial distribution for the spike counts generation,
instead of a Poisson distribution as commonly assumed, captures a relevant part of those
effects and is in fact critical to account for the non-stationarity of the spontaneous activity
of the network.
(cid:80)
One way to approach the problem is to assume that at each time t the generated
spike count of a generic neuron Si(t) is determined by both the field Hi(t) = hi +
j αij(τ ) Sj(t − τ ) (the sum of external fields and the one due to the activity of
the other observed neurons), and the field due to inputs from unobserved, 'hidden' neu-
rons, hhid
(Fig. 3A), such that the observed spike count probability distribution would be
obtained by integrating over the (unknown) distribution of hhid
(cid:80)
:
τ
i
i
(cid:90)
p(SiHi) =
i p(SiHi, hhid
dhhid
i
) p(hhid
i
)
(1)
by λi(t) = f(cid:2)Hi(t) + hhid
(cid:3).
i
where we still assume that p(Si(t)Hi(t), hhid) is a Poisson distribution, with mean given
It is intuitive that p(Si(t)Hi(t)) will now generate super-Poissonian fluctuations; and it
turns out that larger fluctuations are critical for the ability of the Sig-NegBin model to
spontaneously generate bursting activity, as shown above (Fig. 3B upper panel; blue line:
data; red line: Sig-NegBin model in free mode).
In fact, an inferred Sig-Poisson model
shows very sparse bursting (Fig. 3B, middle panel). But it is not just that a broader spike
generation distribution favors larger fluctuations in the network activity. We simulated the
Sig-NegBin model replacing the Negative Binomial with a Poisson spike generator (Fig. 3B,
bottom panel); such 'Poisson replay' of the inferred Sig-NegBin model generates frequent
bursts, each lasting for long time on average. Thus the inferred synaptic couplings of the
Sig-NegBin model, unlike the ones for the Sig-Poisson, appear to be compatible with a
bistable network dynamics, between UP and DOWN states, as also suggested by the results
for the sigmoid transfer function (Fig. 2B). Therefore, the main role of a super-Poisson
spike generation distribution seems to be to destabilize the UP state, making the inferred
model once again more robust to the variability of the free network's activity and thus able
to self-sustain a highly non-stationary activity.
Taking together the results in Fig. 2 and Fig. 3, the suggested picture is that on the one
hand the sigmoid gain function allows the inference procedure to explore and use ranges of
couplings that are effectively forbidden for an exponential gain function, and are essential
to spontaneously generate large sudden increase of activity (burst onset); on the other
hand, the NegBin generator plays the dynamic role of efficiently destabilizing, with large
fluctuations of the spike counts, the high-activity states which are spontaneously generated
by the large recurrent excitation. The two ingredients are robustly coupled to spontaneously
7
FIG. 3: Rationale and implications of choosing a negative binomial as generative model of spike
counts. A Sketch showing the effects of the unobserved neurons on the fluctuations in the
observed part of the network. B Spontaneous bursting activity generated by different models
(blue line: data; red line: model in free mode). Upper panel: Sig-NegBin model; middle panel:
Sig-Poisson; bottom panel: Sig-NegBin model after replacing the Negative Binomial with a
Poisson spike generator. C Comparison between the spike counts generated by a negative
binomial distribution (r = 0.22) with a mean value equal to the average spike count of the neuron
(blue line), the Monte Carlo sampled p(SiHi) with best matching values for the first two
moments of spike counts (red line) and Poisson distribution for the same average spike count
(black line).
produce bursting activity.
Albeit representing just one of the forms that p(Si(t)Hi(t)) can take, the NegBin distri-
bution has been suggested to adequately model fluctuations in observed neural activity24,25,
which makes it a natural candidate. Yet, a NegBin would result from Eq. 1 assuming an
exponential transfer function and a log-Gamma distribution for hhid
: the first condition is
patently in contradiction with our choice of a saturating transfer function and the second
one is not expected to hold in general. We nevertheless tested the adequacy of the NegBin
assumption for the Sig-NegBin model as follows. We assumed that p(hhid
) in Eq. 1 is a
scaled and shifted version of the distribution of the observed fields H; in turn we estimated
such distribution from the inferred synaptic couplings and the observed experimental spike
counts, for the neuron with maximal average activity, for which we expected the differences
in the spike count distribution would both show up more clearly, and matter more for the
dynamics. From this, we performed a Monte Carlo estimate of p(SiHi) for different values
i
i
i
of the mean and the standard deviation defining p(hhid
). Fig. 3C shows a NegBin that is
very close to the one used in the Sig-NegBin model (r = 0.22) for a mean value equal to the
average spike counts of the neuron (blue line); the sampled p(SiHi) with best matching
values for the first two moments is reported in red; the two distributions are very similar,
especially if compared to the Poisson distribution for the same average spike count (black
line). Thus spike counts compatible with a NegBin distribution are naturally accounted
for by a plausible assumption on the effect of sub-sampling underlying Eq. 1, with marked
deviations from a Poisson distribution. We repeated the above procedure sampling hhid
from a Gaussian distribution, finding similar results; this suggests that, provided suitable
values for the mean and variance are chosen, the precise shape of the p(hhid
) has a minor
effect.
i
8
i
The inferred model captures detailed temporal and spatial aspects of the neural dynamics
Although the spike-frequency mechanism with which the Sig-NegBin model is endowed
allows for 5 independent time-scales, it is interesting to note that the inferred values consis-
tently aggregate around just two values, at about 100 ms and 2 s respectively, for different
temporal segments of the same network's activity (Fig. 4A). This result is consistent with
previous work on the same data aiming to infer the main time-scales of the bursting dy-
namics via a completely different approach8.
To assess how relevant the adaptation mechanism is for the model dynamics, we performed
inference with and without SFA. Then we simulated the model in driven mode until just
after the end of a burst of large amplitude, where adaptation effects are expected to be
more pronounced, leaving the model free to spontaneously evolve afterwards. In Fig.4B the
SFA and non-SFA dynamics are reported (top and bottom, respectively); black line is the
model activity averaged over 500 simulations of the dynamics, while gray shading marks
the plus/minus one standard deviation range; red line is the population activity from data.
For each realization we collected the next IBI, that is the time interval to the next burst
spontaneously generated by the model. The IBI histogram (inset) attains a maximum very
close to zero when SFA is not present; such maximum shifts toward higher values with
SFA; the two distributions also have different averages (2.3 s vs 3.2 s) and coefficients of
variation (0.98 vs 0.65). Thus SFA significantly increases the average interval to the next
burst, reducing at the same time its variability.
We also found in data, consistently with the results reported in19, a positive correlation
between burst amplitude and the length of the previous IBI, an effect that could be possibly
attributed to some adaptation mechanism, such as SFA. We estimated this correlation from
simulations of the models inferred, with and without SFA, on different recordings; Fig. 4C
compares the result of the two models with that found in their experimental counterparts.
It is seen that the model without SFA is unable to recover correlations that are significantly
different from zero, while the model with SFA produces correlations that are consistent with
that measured in the data.
We also addressed the ability of the model to account for the temporal structure and
spatial organization of the network activity. We first considered the average order of activa-
tion of different spatial locations within a burst (see Methods for details): time-rank is the
index of the time bin when a neuron spikes first since the burst onset. By evaluating the
average rank for both each neuron in the model and each electrode in the data, we found
them strongly correlated (c = 0.9). In Fig. 5A we report the spatial distribution of average
time-ranks for data (left) and simulations (right).
Looking farther into the burst temporal structure, we asked to what extent the time-ranks
in the model and in the data are comparable on a single burst basis. For each burst in the
data, we selected and compare the simulated burst with the closest time-rank pattern.
Fig 5B and C show two examples of such comparison. To provide a global comparison,
capturing the correlation in time-ranks between data and simulations, we proceeded as
follows:
independently for each neuron, we shuffled the vector of its time-ranks across
9
FIG. 4: Effect of inferred time-scales and SFA currents. A Inferred τSFA for different recordings
for the same preparation; the five inferred time-scales cluster around two values for all recordings,
one about 100ms, the other around 2s. B Comparison between the model dynamics with and
without SFA. In black, and gray-shading, are represented the average and variance over 500
simulations. Red line is the average activity from data during a large chosen burst. Insets:
histograms of the time to the subsequent burst. C Correlation between the burst amplitude and
the time since the preceding burst.
all the simulated bursts, thereby destroying spatial correlations in the time-ranks, while
preserving the average ranks (Fig. 5A, right).
shuf f and dmin
For each burst in the data we took the closest simulated burst and the closest shuffled
simulated one, computed the corresponding distances (dmin
sim ) and the distribution
of their differences (Fig. 5D). It is seen that, for the large majority of bursts, the simulated
burst is closer to the data burst than the surrogate. We therefore can conclude that inferred
model captures most of the spatial development of neural activity at the single burst level.
To inquire into the relationship between the inferred synaptic structure and the ensuing
network dynamics, we also asked whether for a neuron with low average time-rank (early
spiking), the efficacy of its outgoing synapses correlates with the time-ranks of its post-
synaptic neurons. Fig. 5E shows indeed a high negative correlation between the efficacy of
outgoing synapses and the time-rank of post-synaptic neurons (c = −0.8).
DISCUSSION
Several criteria can be identified to evaluate a model, yet the ability to reproduce the
behavior of the system under analysis is certainly among the most relevant. We have seen
that the addition of few computational ingredients can enable a probabilistic, generative
network to autonomously sustain a rich, highly non-stationary dynamics, as observed in the
experimental data employed to infer the model's parameters.
In recent years, the field of statistical inference applied to neuronal dynamics has mainly
focused on devising models and procedures that could reliably recover the real or effec-
tive synaptic structure of a network of neurons, under different dynamical and stimulation
conditions26–28, even though, more often than not, an experimental ground truth was not
readily available.
Although, quite surprisingly, the study of the dynamics of the inferred models has been
largely neglected, we are not the first to address the issue. In4 the authors demonstrated,
on an in-vitro population of ON and OFF parasol ganglion cells, the ability of a GLM
to accurately reproduce the dynamics of the network;7 studied the response properties of
lateral intraparietal area neurons at the single trial, single cell level; the capability of GLM
to capture a broad range of single neuron response behaviors was analyzed in6. In all these
works, however, the focus was on the response of neurons to stimuli of different spatio-
temporal complexity; even where network interactions were accounted for, they proved to
be, for the overall dynamics displayed by the ensemble, an important, yet not decisive
10
FIG. 5: Spatial microscopic network dynamics, data vs model. A Spatial distribution of the
average time-rank of neuron activations during bursts for data and model. Averages are
computed over all bursts in the data and a simulation of comparable length. B Spatial
distribution of time-rank for 2 single bursts in data. C Spatial distribution of time-rank for the 2
bursts from simulation, most similar to those showed in B. D Distribution of the difference of the
similarity between bursts from data, and both bursts from simulation and bursts from surrogate
simulated data. E Scatterplot of post-synaptic efficacies of the neuron with smallest time-rank vs
the rank of its post-synaptic targets.
correction. To our knowledge, no published study to date has focused on the autonomous
dynamics of GLM networks inferred from neuronal data.
Important progress has also been made on the ability of GLMs to predict single neuron
spiking in an ensemble of neurons, with accuracy higher than that provided by methods
based on the instantaneous state of the ensemble itself2 and the PSTH4. In the lingo of the
present paper, though, these attempts were based on simulations of the inferred network
in driven mode, with at most the activity of a single neuron free to reenter just that single
11
neuron's dynamics. To our knowledge, our work for the first time has attempted a direct,
microscopic comparison between the activity autonomously generated by an inferred GLM
network and the one observed in the biological data. The free mode of simulation opens the
way, in principle, to multi-time-step, whole network activity predictions. Being the model
probabilistic and the data noisy, it is of course expected that the ability of the model, in
free simulation mode after having been driven by the data for some time, to follow the
experimentally observed spike trains should quickly deteriorate. In fact, this is what we
found with our model: a burst becomes practically unforeseeable going some 100-200 ms
in the past. This result is in contrast with the impressive findings reported in29, in which
the authors, on recordings similar to the ones used in this work, through a new model-
free method based on state-space reconstruction, were able to predict the occurrence of
a burst even 1 s in advance. Whilst it is not clear why this is the case, we applied the
state-reconstruction method on our data with negative results, in agreement with what we
found with our model.
We presented several semi-quantitative comparisons between the spike trains generated
by the model network, without any time-varying external input, and the data from which
the model's parameters have been inferred. Such tests, albeit successful, represent of course
just an example of the many possible ways to compare two bursting dynamics. Yet we
stress how a similar model, that has been widely and successfully applied to the analysis of
neuronal data and has a slightly greater number of parameters too, failed in large part to
reproduce the most prominent features of the dynamics displayed by the data.
Two interrelated questions pertain to how representative the chosen experimental system
is of neuronal network dynamics in general and the capability of the presented framework
to generalize to other kinds of non-stationary neural activities. As for the first question,
our primary concern in selecting an ex-vivo bursting culture has been to study a system
that, absent any external stimuli, would show a non-trivial collective behavior. As dis-
cussed above, the capability of GLMs to mimic the response properties of single neurons
or a network to stimuli has already been well investigated. On the other hand, numerous
studies in literature witness to the complexity of spatial and temporal behavior of neu-
ronal cultures8–20. As for the second question, we remark that our primary goal was to
devise a 'proof of concept' extension to show that a GLM can autonomously generate a
complex non-stationary activity, and not to establish general conditions for such capability.
We therefore chose a bursting regime as a testbed for the model not because it is exem-
plary, but because it represented a clear challenge: a collective, non-linear phenomenon
that demands the model to adapt to a broad spectrum of time- and space-scales. There
were of course sensible alternatives to choose from, the most relevant being, e.g., the UP
and DOWN dynamics in brain slices; although we have not tackled the issue directly, our
results (see Fig. 3B, bottom plot) hint at a inherent capability of the model to sustain a
noisy bistable dynamics, as more in general suggested in30. Nevertheless, the extent to
which generative models are able to adapt to different regimes, and how dynamic network
effects and stimulation characteristics interact in a generative model like the one studied
here are open issues, that will probably attract much interest in the near future.
Given our goal to establish the capability of a GLM to reproduce a complex, non-linear
dynamics, we found two ingredients - a non-symmetric sigmoidal transfer function and
a Negative Binomial spike-generation model - that proved to be minimal: absent one of
them, the model's performances were drastically impaired. Even if we do not claim these
specific ingredients to be necessary in general, we are persuaded they represent instances
of mechanisms that are both biologically plausible and computationally desirable for the
model. Although a sigmoidal, saturating transfer function would appear to be a natural
choice for a model meant to reproduce neural activities, surprisingly this option has never
been explored, to our knowledge, to perform network inference on neuronal recordings.
While the use of an exponential transfer function in GLMs is grounded in general statistical
requirements31, in our case an asymmetric sigmoid appears to be the single most important
factor explaining the success of the proposed model. As already noted, the saturation of the
sigmoidal function allows naturally for a bimodal distribution of firing rates and thus makes
12
model's behavior more robust in the face of the intrinsic fluctuations of network activity.
Therefore, if the specific form we chose is just one possible instance among many plausible
ones, a saturating behavior is expected to benefit the model beyond the immediate scope
of the present work; a certain degree of asymmetry also appears to be beneficial according
to our results, allowing for differential sensitivity to high and low input levels.
It is interesting to note how this finding contrast with the success of non-saturating
transfer functions in deep learning literature, where the introduction of rectified linear units,
in place of the standard logistic ones, has represented one of the major breakthroughs of
recent years32. Such units exhibit 'intensity equivariance', that is the ability to readily
generalize to data points that differ only for a scale factor; whilst such property is clearly
valuable when dealing with data such natural images and sounds, when applying machine
learning techniques to very noise and sparse data, such as in the case studied here, bounded
transfer functions are probably beneficial exactly for the opposite reason: they filter out
most of the incoming information to gain a poorer but more robust coding in the output.
We have seen that, if the role of a sigmoid gain function seems to facilitate a bistable
behavior of the network, the negative binomial's one is to efficiently destabilize, with large
fluctuations of the spike counts, the high-activity state. Recent experimental evidence
has emerged supporting the Negative Binomial distribution as a candidate for the spike
counts variability in real neurons25; moreover, a negative binomial spike generation has
already been adopted in a model combining GLM with flexible graph-theoretic priors for
the connectivity33. Although, then, our second ingredient already finds experimental and
theoretical support in the literature, we provide a new hypothesis on why the super-Poisson
statistics arises: as the effect of input fluctuations generated by the activity of neurons that
have not been recorded. And our hypothesis directly hints to the Negative Binomial as
just one possible way to model such effect, where over-dispersed spike counts are, instead, a
necessary signature of it; it is this more general feature, in our opinion, that proved to be so
important for the success of the proposed model. Interestingly, our findings resonate with
the recently proposed role of fluctuating unobserved variables in the emergence of criticality
in a wide range of systems34.
Our hypothesis presupposes that the recorded neurons have been chosen at random from
the whole network; this is probably not the case in many instances, where a local set of
units is sampled instead; such non-random sub-sampling can potentially lead to strong
deviations from expected behavior23; this of course could produce systematic effects in the
inference. Besides, our model cannot incorporate the effects of slow changes in single neuron
excitability, as observed experimentally in35,36, and analyzed in a statistical model in25.
METHODS
Experimental data
As originally described in10, cortical neurons were obtained from newborn rats within 24
hours after birth, following standard procedures. Briefly, the neurons were plated directly
onto a substrate-integrated multielectrode array (MEA). The cells were bathed in MEM
supplemented with heat-inactivated horse serum (5%), glutamine (0.5 mM), glucose (20
mM), and gentamycin (10 µg/ml) and were maintained in an atmosphere of 37◦ C, 5%
CO2/95% air in a tissue culture incubator as well as during the recording phases. The data
analyzed here was collected during the third week after plating, thus allowing functional
and structural maturation of the neurons. MEAs of 60 Ti/Au/TiN electrodes, 30 µm in
diameter, and spaced 200 µm from each other (Multi Channel Systems, Reutlingen, Ger-
many) were used. The insulation layer (silicon nitride) was pretreated with poly-D-lysine.
All experiments were conducted under a slow perfusion system with perfusion rates of ∼100
µl/h. A commercial 60-channel amplifier (B-MEA-1060; Multi Channel Systems) with fre-
quency limits of 1-5000 Hz and a gain of 1024× was used. The B-MEA-1060 was connected
to MCPPlus variable gain filter amplifiers (Alpha Omega, Nazareth, Israel) for additional
13
amplification. Data was digitized using two parallel 5200a/526 analog-to-digital boards (Mi-
crostar Laboratories, Bellevue, WA). Each channel was sampled at a frequency of 24000 Hz
and prepared for analysis using the AlphaMap interface (Alpha Omega). Thresholds (8×
root mean square units; typically in the range of 10-20 µV ) were defined separately for each
of the recording channels before the beginning of the experiment. The electrophysiological
data are freely available from S. Marom (http://marom.net.technion.ac.il/).
ACKNOWLEDGEMENTS
We are grateful to Shimon Marom for having shared his data, and for several illuminating
discussions along the way.
1W. Truccolo, U. T. Eden, M. R. Fellows, J. P. Donoghue, and E. N. Brown, "A point process framework
for relating neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects,"
Journal of neurophysiology 93, 1074–1089 (2005).
2W. Truccolo, L. R. Hochberg, and J. P. Donoghue, "Collective dynamics in human and monkey sensori-
motor cortex: predicting single neuron spikes," Nature Neuroscience 13, 105–111 (2010).
3C. Capone, C. Filosa, G. Gigante, F. Ricci-Tersenghi, and P. Del Giudice, "Inferring synaptic structure
in presence of neural interaction time scales," PloS one 10, e0118412 (2015).
4J. W. Pillow, J. Shlens, L. Paninski, A. Sher, A. M. Litke, E. Chichilnisky, and E. P. Simoncelli, "Spatio-
temporal correlations and visual signalling in a complete neuronal population," Nature 454, 995–999
(2008).
5I. H. Stevenson, J. M. Rebesco, N. G. Hatsopoulos, Z. Haga, L. E. Miller, and K. P. Kording, "Bayesian
inference of functional connectivity and network structure from spikes," IEEE Transactions on Neural
Systems and Rehabilitation Engineering 17, 203–213 (2009).
6A. I. Weber and J. W. Pillow, "Capturing the dynamical repertoire of single neurons with generalized
linear models," Neural Computation 29, 3260–3289 (2017).
7I. M. Park, M. L. Meister, A. C. Huk, and J. W. Pillow, "Encoding and decoding in parietal cortex
during sensorimotor decision-making," Nature neuroscience 17, 1395–1403 (2014).
8G. Gigante, G. Deco, S. Marom, and P. Del Giudice, "Network events on multiple space and time scales
in cultured neural networks and in a stochastic rate model," PLoS computational biology 11, e1004547
(2015).
9T. Baltz, A. Herzog, and T. Voigt, "Slow oscillating population activity in developing cortical networks:
models and experimental results." J Neurophysiol 106, 1500–1514 (2011).
10D. Eytan and S. Marom, "Dynamics and effective topology underlying synchronization in networks of
cortical neurons," The Journal of Neuroscience 26, 8465–8476 (2006).
11M. Giugliano, P. Darbon, M. Arsiero, H.-R. Luscher, and J. Streit, "Single-neuron discharge properties
and network activity in dissociated cultures of neocortex," Journal of Neurophysiology 92, 977–996 (2004).
12T. Gritsun, J. le Feber, J. Stegenga, and W. L. Rutten, "Experimental analysis and computational mod-
eling of interburst intervals in spontaneous activity of cortical neuronal culture," Biological Cybernetics
105, 197–210 (2011).
13I. P. I. Park, D. X. D. Xu, T. B. DeMarse, and J. C. Principe, "Modeling of Synchronized Burst in
Dissociated Cortical Tissue: An Exploration of Parameter Space," (2006).
14D. A. Wagenaar, J. Pine, and S. M. Potter, "An extremely rich repertoire of bursting patterns during
the development of cortical cultures," BMC neuroscience 7, 11 (2006).
15J. M. Beggs and D. Plenz, "Neuronal avalanches in neocortical circuits," The Journal of Neuroscience 23,
11167–11177 (2003).
16T. Petermann, T. C. Thiagarajan, M. A. Lebedev, M. A. Nicolelis, D. R. Chialvo, and D. Plenz, "Sponta-
neous cortical activity in awake monkeys composed of neuronal avalanches," Proceedings of the National
Academy of Sciences 106, 15921–15926 (2009).
17D. Plenz and T. C. Thiagarajan, "The organizing principles of neuronal avalanches: cell assemblies in the
cortex?" Trends in neurosciences 30, 101–110 (2007).
18D. Plenz and H. G. Schuster, Criticality in neural systems (Wiley-VCH New York, NY, 2014).
19F. Lombardi, H. J. Herrmann, D. Plenz,
and L. de Arcangelis, "Temporal correlations in neuronal
avalanche occurrence," Scientific reports 6, 24690 (2016).
20M. Yaghoubi, T. de Graaf, J. G. Orlandi, F. Girotto, M. A. Colicos, and J. Davidsen, "Neuronal avalanche
dynamics indicates different universality classes in neuronal cultures," Scientific reports 8, 3417 (2018).
21T. Masquelier and G. Deco, "Network bursting dynamics in excitatory cortical neuron cultures results
from the combination of different adaptive mechanism," PloS one 8, e75824 (2013).
22A. Peyrache, N. Dehghani, E. N. Eskandar, J. R. Madsen, W. S. Anderson, J. A. Donoghue, L. R.
Hochberg, E. Halgren, S. S. Cash, and A. Destexhe, "Spatiotemporal dynamics of neocortical excitation
and inhibition during human sleep," Proceedings of the National Academy of Sciences 109, 1731–1736
(2012).
23A. Levina and V. Priesemann, "Subsampling scaling," Nature communications 8, 15140 (2017).
14
24A. Onken, S. Grunewalder, M. H. Munk, and K. Obermayer, "Analyzing short-term noise dependencies
of spike-counts in macaque prefrontal cortex using copulas and the flashlight transformation," PLoS
computational biology 5, e1000577 (2009).
25R. L. Goris, J. A. Movshon, and E. P. Simoncelli, "Partitioning neuronal variability," Nature neuroscience
17, 858–865 (2014).
26Y. Roudi, B. Dunn, and J. Hertz, "Multi-neuronal activity and functional connectivity in cell assemblies,"
Current opinion in neurobiology 32, 38–44 (2015).
27J. Tyrcha, Y. Roudi, M. Marsili, and J. Hertz, "The effect of nonstationarity on models inferred from
neural data," Journal of Statistical Mechanics: Theory and Experiment 2013, P03005 (2013).
28O. M. A. D. U. F. Trang-Anh Nghiem, Bartosz Telenczuk, "Maximum entropy models reveal the corre-
lation structure in cortical neural activity during wakefulness and sleep," ArXiv (2018).
29S. Tajima, T. Mita, D. J. Bakkum, H. Takahashi, and T. Toyoizumi, "Locally embedded presages of
global network bursts," arXiv preprint arXiv:1703.04176 (2017).
30V. Rostami, P. P. Mana, S. Grun, and M. Helias, "Bistability, non-ergodicity, and inhibition in pairwise
maximum-entropy models," PLOS Computational Biology 13, e1005762 (2017).
31J. K. Lindsey, Applying generalized linear models (Springer Science & Business Media, 2000).
32V. Nair and G. E. Hinton, "Rectified linear units improve restricted boltzmann machines," in Proceedings
of the 27th international conference on machine learning (ICML-10) (2010) pp. 807–814.
33S. Linderman, R. P. Adams, and J. W. Pillow, "Bayesian latent structure discovery from multi-neuron
recordings," in Advances in Neural Information Processing Systems (2016) pp. 2002–2010.
34D. J. Schwab, I. Nemenman, and P. Mehta, "Zipfs law and criticality in multivariate data without
fine-tuning," Physical review letters 113, 068102 (2014).
35A. Gal, D. Eytan, A. Wallach, M. Sandler, J. Schiller, and S. Marom, "Dynamics of excitability over
extended timescales in cultured cortical neurons," The Journal of Neuroscience 30, 16332–16342 (2010).
36A. Gal and S. Marom, "Self-organized criticality in single-neuron excitability," Physical Review E 88,
062717 (2013).
|
1604.01680 | 4 | 1604 | 2016-11-09T09:50:39 | The Complex Hierarchical Topology of EEG Functional Connectivity | [
"q-bio.NC"
] | Understanding the complex hierarchical topology of functional brain networks is a key aspect of functional connectivity research. Such topics are obscured by the widespread use of sparse binary network models which are fundamentally different to the complete weighted networks derived from functional connectivity. We introduce two techniques to probe the hierarchical complexity of topologies. Firstly, a new metric to measure hierarchical complexity; secondly, a Weighted Complex Hierarchy (WCH) model. To thoroughly evaluate our techniques, we generalise sparse binary network archetypes to weighted forms and explore the main topological features of brain networks- integration, regularity and modularity- using curves over density. By controlling the parameters of our model, the highest complexity is found to arise between a random topology and a strict 'class-based' topology. Further, the model has equivalent complexity to EEG phase-lag networks at peak performance. Hierarchical complexity attains greater magnitude and range of differences between different networks than the previous commonly used complexity metric and our WCH model offers a much broader range of network topology than the standard scale-free and small-world models at a full range of densities. Our metric and model provide a rigorous characterisation of hierarchical complexity. Importantly, our framework shows a scale of complexity arising between 'all nodes are equal' topologies at one extreme and 'strict class-based' topologies at the other. | q-bio.NC | q-bio | The Complex Hierarchical Topology of EEG Functional Connectivity
Keith Smith∗,a,b, Javier Escuderoa
aInstitute for Digital Communications, School of Engineering, University of Edinburgh, West Mains Rd, Edinburgh, EH9 3FB, UK,
bAlzheimer Scotland Dementia Research Centre, University of Edinburgh, 7 George Square, Edinburgh, EH8 9JZ, UK
e-mail: [email protected], [email protected]
6
1
0
2
v
o
N
9
]
.
C
N
o
i
b
-
q
[
4
v
0
8
6
1
0
.
4
0
6
1
:
v
i
X
r
a
Abstract
Background : Understanding the complex hierarchical topology of functional brain networks is a key aspect of func-
tional connectivity research. Such topics are obscured by the widespread use of sparse binary network models which are
fundamentally different to the complete weighted networks derived from functional connectivity.
New Methods: We introduce two techniques to probe the hierarchical complexity of topologies. Firstly, a new metric
to measure hierarchical complexity; secondly, a Weighted Complex Hierarchy (WCH) model. To thoroughly evaluate our
techniques, we generalise sparse binary network archetypes to weighted forms and explore the main topological features
of brain networks- integration, regularity and modularity- using curves over density.
Results: By controlling the parameters of our model, the highest complexity is found to arise between a random
topology and a strict 'class-based' topology. Further, the model has equivalent complexity to EEG phase-lag networks
at peak performance.
Comparison to existing methods: Hierarchical complexity attains greater magnitude and range of differences between
different networks than the previous commonly used complexity metric and our WCH model offers a much broader range
of network topology than the standard scale-free and small-world models at a full range of densities.
Conclusions: Our metric and model provide a rigorous characterisation of hierarchical complexity.
Importantly,
our framework shows a scale of complexity arising between 'all nodes are equal' topologies at one extreme and 'strict
class-based' topologies at the other.
Keywords: Functional connectivity, Hierarchical complexity, Brain networks, Electroencephalogram, Network
simulation
1. Introduction
Graph theory is an important tool in functional con-
nectivity research for understanding the interdependent
activity occurring over multivariate brain signals [1–3]. In
this setting, Complete Weighted Networks (CWNs) are
produced from all common recording platforms including
the Electroencephalogram (EEG), the Magnetoencephalo-
gram (MEG) and functional Magnetic Resonance Imaging
(fMRI), where every pair of nodes in the network share a
connection whose weight is the output of some connectiv-
ity measure. Complex hierarchical structures are known to
exist in real networks [4], including brain networks [1, 5],
for this reason it is important to find methods to specifi-
cally evaluate hierarchical complexity of network topology.
Here we introduce methods specific to this end.
Complexity is understood neither to mean regularity,
where obvious patterns and repetition are evident, nor
randomness, where no pattern or repetition can be estab-
lished, but attributed to systems in which patterns are
irregular and unpredictable such as in many real world
phenomena [6]. Particularly, the brain is noted to be such
a complex system [7] and this is partly attributed to its
hierarchical structure [5]. Hierarchical complexity is thus
concerned with understanding how the hierarchy of the
system contributes to its complexity. Here we introduce a
new metric aptly named hierarchical complexity, R, which
is based on targeting the structural consistency at each
hierarchical level of network topology. We compare our
metric with network entropy [8] and find that we can of-
fer a greater magnitude and density range for establishing
differences in complexity of different graph topologies.
Alongside this, we introduce the Weighted Complex Hi-
erarchy (WCH) model which simulates hierarchical struc-
tures of weighted networks. This model works by mod-
ifying uniform random weights by addition of multiples
of a constant, which is essentially a weighted preferential
selection method with a highly unpredictable component
provided by the original random weights. We show that it
follows very similar topological characteristics of networks
formed from EEG phase-lag connectivity. Intrinsic to our
model is a strict control of weight ranges for hierarchi-
cal levels which offers unprecedented ease, flexibility and
rigour for topological comparisons in applied settings and
for simulations in technical exploration for brain network
analysis. This also provides an unconvoluted alternative
to methods which randomise connections [9, 10] or weights
Preprint submitted to Journal of Neuroscience Methods
November 10, 2016
[11] of the original network.
Any rigorous evaluation of brain networks should ad-
dress their inherent complete weighted formulation [12].
However, the current field has largely lacked any concerted
effort to build an analytical framework specifically tar-
geted at CWNs, preferring instead to manipulate the func-
tional connectivity CWNs into sparse binary form (e.g.
[10, 13, 14] as well as wide-spread use of the Watts-Strogatz
[9] and Albert-Barabasi [15] models) and using the pre-
existing framework built around other research areas which
have different aims and strategies in mind [16].
In our
methodological approach we propose novel generalisations
of pre-existing sparse binary models to CWN form and
thus allow a full density range comparison of our tech-
niques. Due to the intrinsic properties of these graph types
we find minimal and maximal topologies which can help to
shed light on a wide variety of topological forms and their
possible limitations [8] in a dense weighted framework.
Further, as part of our study we seek after straightfor-
ward metrics to evaluate other main aspects of network
topology for comparisons [8, 17] and, in this search, found
it necessary to revise key network concepts of integration-
segregation [2, 9, 18] and scale-freeness [15, 19]. We pro-
vide here these revisions: i) That the clustering coefficient,
C, is enough to analyse the scale of integration and seg-
regation, finding it unnecessary and convoluted to use the
characteristic path length, L, as a measure of its opposite,
as generally accepted [1, 9]. ii) We provide mathematical
justification that the degree variance, V , and thus network
irregularity [20] is a strong indicator of the scale-free factor
of a topology.
Our study of hierarchical complexity, using a compre-
hensive methodological approach, provides mathematical
quantification of the hierarchical complexity of EEG func-
tional connectivity networks and reveals new insights into
key aspects of network topology in general. Our model
provides improved comparative abilities for future clinical
and technical research.
2. Network Science: Proposed methods and key
revisions
We adopt the notation in [21] so that a graph, G(V, W),
is a set of n nodes, V, connected according to an n × n
weighted adjacency matrix, W. Entry Wij of W corre-
sponds to the weight of the connection from node i to
node j and can be zero. An unweighted graph is one in
which connections are distinguished only by their existence
or non-existence, so that, without loss of generality, all ex-
isting connections have weight 1 and non-existent connec-
tions have weight 0. The graph is undirected if connec-
tions are symmetric, which gives symmetric W. A simple
graph is unweighted, undirected, with no connections from
a node to itself and with no more than one connection be-
tween any pair of nodes. This corresponds to a graph
with a symmetric binary adjacency matrix with zero di-
agonal. Such graphs are easy to study and measure [16].
ki =(cid:80)n
The degree, ki, of node i is defined as the number of its
adjacent connections, which is the number of non zero en-
tries of the ith column of W. Then, for a simple graph,
j=1 Wij. For a graph with 2m edges, the connec-
tion density, P , of a graph is P = 2m/n(n − 1).
A CWN is represented by a symmetric adjacency ma-
trix with zero diagonal (no self-loops) and weights, Wij ∈
[0, 1], elsewhere. To analyse CWNs it is beneficial to con-
vert it to simple form by binarising the adjacency matrix
using a threshold, where a percentage of strongest con-
nections are set to 1 and the remaining values set to 0.
This stays true to the network activity [12] whilst reducing
computational complexity and weight issues found with
weighted metrics [2]. Hereafter, all mathematics will refer
to simple graphs.
In this section we present the contributions of this
study. We first present the hierarchical complexity metric
and the WCH model, which are the key novel contributions
of this paper. Thereafter we detail revisions and clarifica-
tion of integration and segregation as a scale evaluated by
C and scale-freeness as a factor evaluated by V . Finally,
we outline the generalisation of key network archetypes
to CWN form, full details of which can be found in the
supplementary material.
2.1. Hierarchical Complexity Metric
The ideas of order and complexity are well known in
the discussion of networks (indeed, real world networks
are often called complex networks [1, 3, 22]).
In math-
ematics, the graphs studied derive from some theoretical
principles. These can involve set patterns, without ran-
dom fluctuations of connections, such as regular networks,
fractal networks, star networks and grid networks. On the
other hand much interest is shown in more randomly gen-
erated topologies, such as random graphs and other graphs
involving random processes, as these express something of
the more erratic and irregular quality of connections in
networks constructed from real world phenomena [9, 23].
However, real world phenomena differ from random pro-
cesses in that there is a clear organisational behaviour ap-
parent throughout the hierarchical structure, both within
hierarchical levels and between hierarchical levels [4, 5].
Although this structure is perhaps impossible to retrace,
because its formation inevitably involves many unknown
generative processes, we can provide methods for its anal-
ysis.
Hierarchies in networks are generally determined by
degrees of nodes, where a small group of highly connected
nodes create a rich club [22] on the top hierarchical level
and nodes with generally lesser connectivity exist on a pe-
ripheral lower levels. Further, it is seen that a node's re-
lationship within the context of the network is greatly de-
termined by the other nodes to which it is connected [24].
Thus, to understand the hierarchical complexity of a net-
work we propose to study the behaviour of nodes of a given
degree by looking at the degrees of nodes in their neigh-
bourhoods. We define D as the set of degrees of a graph,
2
disorganised or more complexly organised. For example,
in Fig.1.B the two degree nodes all have the same degree
sequences- {3, 4}- whereas the three degree nodes are split
into two different degree sequences- {1, 2, 2} and {1, 1, 4}-
and finally the neighbourhood degree sequences of the four
degree nodes are all different- {1, 1, 1, 4}, {1, 2, 2, 4} and
{2, 3, 3, 3}. So the complexity of just the two degree nodes
is 0, the complexity of just the three degree nodes is 2((2−
1.5)2 + (1 − 1.5)2 + (2 − 3)2 + (4 − 3)2)/(4 × 3 × 3) =
5/36 and the complexity of just the four degree nodes is
(2(1 − 4/3)2 + (2 − 4/3)2 + 2((1 − 2)2 + (3 − 2)2) + 2(4 −
11/3)2 + (3 − 11/3)2)/(4 × 3 × 2) = (16/3)/24 = 8/36, the
complexity over all three levels being the average- 13/108.
This measure is thus minimal for graphs in which, for
each k and k(cid:48), every k-degree node is connected to exactly
the same number of k(cid:48)-degree nodes. This property, for ex-
ample, is seen in ring lattices, and quasi-star graphs and is
close to minimal in the line graph, fractal graphs and grid
lattices. Furthermore, the degrees of random networks are
known to have a fairly small spread which is a factor pe-
nalised by our complexity value. Thus random networks
should obtain low values of our complexity measure. On
the other hand, R values of real networks are expected to
be higher given the high spread and degree fluctuations
of those networks caused by hub nodes promoting a high
degree irregularity while the spontaneous nature of real-
world connections should promote a high variability of the
neighbourhood degree sequences.
2.2. Weighted Complex Hierarchy Model
The foundation of our model is the random CWN model.
The most general random network is the Erdos-R´enyi (E-
R) random network [23] which is formed by assigning a
probability, p, to the question of the existence or non-
existence of connections on a network with n nodes. Such
a construct is,
in fact, an ensemble of graphs denoted
G(n, p). A sample of this ensemble is obtained by generat-
ing a random value for every possible connection and ap-
plying the probability value p as a threshold to see whether
or not that connection should exist in our sample. The ran-
dom CWN model is thus simply a symmetric matrix with
zero diagonal and randomly generated values Wij ∈ [0, 1]
elsewhere.
If we threshold the CWN at weight T = p,
we recover a binary Erdos-R´enyi random graph from the
random graph ensemble G(n, p).
Starting from an Erdos-R´enyi CWN we randomly dis-
tribute the nodes into hierarchy levels based on some dis-
crete cumulative distribution function, p, by generating a
random number, r, between 0 and 1 for each node and
putting the node in the level for which r − p is first less
than 0. We then distribute ls additional weight to all
connections of adjacent nodes in the lth level, for some
suitably chosen s. The parameters of this model are then
(n, s, l, p). The parameter n is the number of nodes in
the network. The parameter s is the strength parameter,
which is constant since the random generation of the ini-
tial weights is enough to contribute to weight randomness.
Figure 1: A. Example of a node degree neighbourhood. Here is shown
a part of a network relating to the neighbourhood of the blue node.
The blue node has neighbourhood degree sequence {1, 2, 3, 4, 4}, i.e.
the ordered degrees of the orange nodes. Grey connections indicate
all the additional connections of the orange nodes in the network.
B. Example for graph complexity. Here is shown a 20 node network
with varying 'orderedness' at different degree levels. C. Diagram
of the construction of the WCH model. Above is the probability
distribution function for a geometric distribution with p = 0.6 for
a three level hierarchy. Below is a graphic displaying the additional
weight added between nodes in given hierarchy levels.
G. Similar to the idea of node degree sequences [25], we
can construct neighbourhood degree sequences, specific to
each node in the graph. That is, for a node, i, of degree
k ∈ D we have a sequence
si = {di,1, di,2, , di,k} s.t. di,1 ≤ di,2 ≤ ··· ≤ di,k ∈ D,
where di,j is the degree of the jth node connected to node
i (see Fig. 1.A). For all nodes of a given degree, k, the
corresponding neighbourhood degree sequences have equal
length, k.
We define the hierarchical complexity, R, of a network
as the average variance of the k-degree neighbourhood de-
gree sequences and can be expressed as:
(cid:88)
Dk(cid:54)=∅
R =
1
D
1
krk(rk − 1)
k(cid:88)
j=1
(cid:32)(cid:88)
i∈Dk
(cid:33) ,
(ski(j) − µkj)2
(1)
where D is the number of distinct degrees in the graph, Dk
is the set of nodes of degree k, ski(j) is the jth element
of the ith k-length sequence, µkj is the mean value of el-
ement j over all k-length sequences and rk is the number
of nodes of degree k, which is added to the denominator
for normalisation of hierarchy levels.
Organisation of the graph at the level of k-degree nodes
can be seen by comparing the jth elements of their neigh-
bourhood sequences. If all of the jth elements of all the
sequences are equal, that is si = sj for all si, sj of length k,
then there is a high degree of order present in the k-degree
nodes of the graph. If these sequences differ widely how-
ever, then it can be said that the k-degree nodes are either
3
The parameter l is the number of levels of the hierarchy,
with a default setting of a random integer between 2 and
5. The vector p is the cumulative probability distribution
vector denoting the probabilities that a given node will be-
long to a given level where the default, which we use here,
is a geometric distribution with p = 0.6 in hierarchical
levels (0, 1, 2, . . . , l) where the nodes with highest connec-
tivity (top hierarchical level) are at the tail end of the
distribution. Fig. 1.C plots an example of the geometric
distribution for a three level hierarchy. The text inside the
box plots, above, indicates the additional weights given to
connections adjacent to nodes inside the given level. The
graphic below explains the additional weights provided by
the strength parameter of connections between nodes in
different levels as well as in the same level. For example, a
connection between a level 1 node and a level 2 node has
additional strength 3s which consists of one s provided by
the node in Level 1 and 2s provided by the node in Level 2.
At s = 0, we have the E-R random network and at s = 1
the weights of the network are linearly separable by the hi-
erarchical structure producing a strict 'class-based' topol-
ogy. Between these values a spontaneous 'class-influenced'
topology emerges.
2.3. Revision of concepts from network science
Here we present justifications for metrics as measures
of key topological factors- the global clustering coefficient,
C, for degree of segregation and the degree variance, V ,
for irregularity, linked to scale-freeness.
2.3.1. Integration-Segregation
The concept of integration in brain networks is closely
tied in to the small world phenomenon [26], where real
world networks are found to have an efficient 'trade off'
between segregative and integrative behaviours [27]. The
most widely used topological metrics in network science-
C and the characteristic path length, L- are commonly
noted as measures of these quantities, respectively. Here,
L is defined as the average of the shortest paths between
each pair of nodes and C is defined as the probability that
a path of length 2, or triple, in the graph has a shortest
path of length 1. That is,
C =
closed triples
triples
,
(2)
where a closed triples is such that, for triple {Wik, Wkj},
Wij = 1, for i, j, k distinct.
Since integration implies a non-discriminative behaviour
in choice, we argue that the random graph ensemble [23],
defined by its equal probability of existent connections be-
tween all pairs of nodes, is the most exemplary model of an
integrated network. Anything which deviates from equal
probability is a discriminative factor which favours certain
connections or nodes over others, likely leading to more
segregated activity. Further, it is clear that integration
and segregation are opposite ends of the same spectrum-
4
something which is not integrated must be segregated and
vice versa. Having one metric to inform on where a net-
work lies on that spectrum is therefore sufficient.
Thus, here we propose C as the topological measure
to evaluate levels of integration (and so segregation) of a
given network. Firstly, we note that values of C for random
graphs and small-world graphs are often much more dis-
tinguishable than those of L [9] and it is certainly assumed
that these graphs have very different levels of integration.
Secondly, since the random network is optimally integrated
and E[Cran] = E[Pran] [16], where Pran is the connection
density of the random network, then the larger the devi-
ation from 1 of the value γ = C/E[Cran] = C/E[Pran] =
C/P , the more segregated is the network. We will include
both L and C in our analysis in order to provide evidence
to back the above proposal.
2.3.2. Regularity and Scale-Freeness
Another topological factor of small world networks is
noted as a scale-free nature characterised by a power law
degree distribution [28]. To understand this aspect of net-
work topology another factor of network behaviour is for-
mulated distinguishing between 'line' like and 'star' like
graphs [2, 29].
Here, we show that characterisation of scale-freeness is
closely connected to the regularity of a network. Regular
graphs have been studied for over a century [30]. They
are defined as graphs for which every node has the same
degree. An almost regular graph is a graph for which the
highest and lowest degree differs by only 1. Thus a highly
irregular graph can be thought of as any graph whose ver-
tices have a high variability. Such behaviour can be cap-
tured simply by the variance of the degrees present in the
graph, that is
(3)
where D = {ki}i∈V , is the set of node degrees on a given
graph [20].
V = var(D),
graph with degrees k = {k1, k2, . . . , kn}, and (cid:80)n
For regular graphs V = 0 by definition, but more prob-
ing is necessary to distinguish high V topology. For a
i=1 ki =
2m, on multiplying out the brackets V simplifies to
(cid:19)2
− ki
(cid:18) 2m
n(cid:88)
i=1
n
− 2mP,
V =
=
1
(n − 1)
(cid:107)k(cid:107)2
(n − 1)
2
2 = (cid:80)n
i=1 k2
where P = 2m/n(n − 1) is the connection density and
(cid:107)k(cid:107)2
i , is the squared (cid:96)2 norm of k. This tells
us that V is proportional to the sum of the squares of
the degrees of the graph, (cid:107)k(cid:107)2
2, and, for fixed number of
connections, m, V in fact depends only on (cid:107)k(cid:107)2
2. Now,
it is known that (cid:107)k(cid:107)2
2 is maximal in quasi-star graphs
and quasi-complete graphs [31]. Essentially, the quasi-star
graph has a maximal number of maximum degree nodes in
the graph for the given connection density and the quasi-
complete graph has a maximal number of isolated, or zero-
degree, nodes in the graph. This tells us that, for low P ,
high V denotes the presence of a few high degree nodes
and a majority of relatively low degree nodes, i.e. scale-
free-like graphs. Thus, due to the restriction placed on
possible degree distributions by the number of edges (the
small number of edges in sparse networks means the num-
ber of high degree nodes is very limited), the irregularity
of degrees is a strong indicator of the strength of 'decay'
of the given distribution, relating to how 'scale-free' the
graph is.
2.4. Complete Weighted Network Archetypes
In the supplementary material we detail the method
to generalise sparse binary network archetypes to CWN
form. The pre-requisit of such a generalisation is that we
require obvious higher density versions of lower density
forms which can be arranged in adjacency matrix form
such that each non-zero entry, Wij, of the lower density
adjacency matrix exists as a non-zero entry, Wij in the
higher density adjacency matrix. This is indeed the case
for the Regular Ring Lattice, Star, Grid Lattice and Frac-
tal Modular CWNs (see Fig. 2 A,B,C,D respectively). We
explain these higher and lower density forms of the bi-
narised CWN in terms of weight categories where, if we
choose an appropriate threshold, T , we can recover all
edges in the same and all higher weight categories and
none of the edges existing in all lower categories.
3. Methods
Here we apply methods to graphs of 64 nodes, typical
of medium density EEG. For analysis we employ connec-
tion density thresholds at integer percentages of strongest
weighted connections, rounded to the nearest whole num-
ber of connections. We then implement metric algorithms
on each of these binary networks and plot the obtained
values on a curve against connection density, similar as in
e.g. [32, 33]. This generates metric curves plotted against
connection density which provides a detailed analysis of
the CWN topology. Other methods exist to analyse CWNs
such as weighted metrics [11] or density integrated metrics
[34], but these metrics still give only singular values for a
given network which belies little of topological behaviour
at different scales of connectivity strength.
For random and WCH CWNs we use sample sizes of
100 for each network and for the EEG functional connec-
tivity CWNs we have a sample size of 109 [35]. On the
metric curves for these we plot the median with the in-
terquartile range shaded in. For ordered networks there is
only one network per type by definition.
Our analytical framework is composed of a mixture
of entirely new concepts and novel generalisations of ex-
It is constituted of the
isting concepts to CWN form.
Figure 2: A. A 12 node ring lattice of degree 6, comprising the
three strongest weight categories of the ring lattice CWN. B. The
quasi-star with 4 nodes of degree n − 1 and n − 4 nodes of degree
4, also comprising the first four categories of the star CWN. C. The
grid lattice weight categorisation (relating to the grey node) in a 30
node network (see supplementary material). Colours of edges denote
category: black, blue, green, orange and red edges are in weight cate-
gories 1, 2, 3, 4 & 5, respectively. The increasingly lighter boundaries
thus represent 'catchment' areas around the node by increasing cat-
egory. Centring these 'catchment' areas around a given node gives
the respective categorisation of edges adjacent to the new node. D.
Fractal modular CWN weight categorisation on 30 nodes. Edges
shown (black) are 1st weight category edges.
In this instance, in-
creasingly lighter background represents areas within which all pairs
of nodes become connected by edges when the network is subject to
the threshold corresponding to the respectively increasing category
(see supplementary material).
following elements: four metrics, R, C, V , Q characteris-
ing four important and distinct topological features; five
CWN archetypal models- Random, Star, Regular Lattice,
Fractal Modular, Grid Lattice; the WCH model.
3.1. Metrics
In [8, 17] an 'architecture' of network topology is pro-
posed involving the three most widely studied properties
of brain networks- integration (and segregation) [2, 9, 18],
'scale-freeness' [15, 19] and modularity [5, 36]. For our
analysis in comparison with hierarchical complexity, R,
we choose a straightforward metric for each of these topo-
logical factors- C for integration, V for scale-freeness and
Q for modularity [36] where
(cid:18)
(cid:88)
(cid:19)
Q =
1
2m
i,j
Wij − kikj
2m
δ(ci, cj),
(4)
where ci is the module containing node i and δ() is the
Kronecker delta function. Highly efficient algorithms have
been created [37, 38] aiming to maximise the value of Q
for a given network. To compute the modularity of our
networks, we use the undirected modularity function [37]
in the Brain Connectivity Toolbox [18].
5
3.2. Comparison for Hierarchical Complexity
degree distribution qi = kipi/(cid:80)
We compare our hierarchical complexity metric with
a commonly used metric for analysing the entropy of the
network degrees [8]. This is defined using the normalised
j kjpj, where ki is the de-
gree of node i and pi is the proportion of nodes in the
graph with the same degree as node i which relates to
probabilities of going to/ coming from neighbouring nodes
in directed graph problems. Then the entropy of graph G
is a straightforward derivation of Shannon's entropy equa-
tion [39] for the degrees of the graph:
H(G) = − n(cid:88)
qi log(qi).
(5)
Thus, Network entropy encodes the eccentricity of the
graph degrees.
i=1
3.3. Comparisons for the WCH model
We implement comparisons with the Watts-Strogatz
small-world model [9] which randomly rewires a set pro-
portion of edges starting from a regular lattice. We use
the full range of parameters for initial degree specification
(2 up to 62) and random rewiring parameters from 0.05 in
steps of 0.05 up to 0.95. For each combination of parame-
ters, 100 realisations of the model were computed and C,
V , Q, and R were measured. We further compare with
Albert-Barabasi's scale-free model [15] which begins with
a graph consisting of core of highly connected nodes to
which the rest of the nodes are added one by one with a
set degree but paired by edges to randomly selected nodes.
We use an initial number of nodes of 15 and the additional
node's degree from 3 up to 14 in order to reach larger den-
sities.
3.4. EEG networks
We use an eyes open, resting EEG data set with 64
nodes. We report on networks created from the beta (12.5-
32Hz) band using coherence and the debiased Weighted
Phase-Lag Index (dWPLI) in order to account for differ-
ent possible types of EEG networks while reducing redun-
dancy of similar topological forms found between the fre-
quency bands (see supplementary material).
The dataset, recorded using the BCI2000 instrumenta-
tion system [40], was freely acquired from Physionet [35].
The signals were recorded from 64 electrodes placed in the
main in accordance with the international 10-10 system.
We took the eyes open resting state condition data, con-
sisting of 1 minute of continuously streamed data which
were partitioned into 1s epochs and averaged for each of
109 volunteers.
FieldTrip [41] was used for pre-processing, frequency
analysis and connectivity analysis to obtain the adjacency
matrices of complete weighted networks. The 64 channels
were re-referenced using an average reference, the multi-
taper method was implemented from 0 seconds onwards
6
using Slepian sequences and 2Hz spectral smoothing. A
0.5Hz resolution was obtained using one second of zero
padding. We chose to analyse the matrices obtained from
both the coherence and the debiased Weighted Phase-Lag
Index (dWPLI) [42] to look for differences between net-
work topologies of zero and non-zero phase lag dependen-
cies in the channels [43]. We treat the data of all tasks
as a single dataset to allow for the variability of the EEG
network topologies since we are not interested here in the
tasks themselves but on the behaviour of general EEG net-
works obtained from dWPLI and coherence.
3.5. Statistical analysis
Due to the polynomial formulation of the complexity
measure, producing a non-normal distribution, we com-
pare metric distributions using the Wilcoxon rank sum
test. The z-score is used to ascertain the magnitude and
direction of the relationship of the distributions.
4. Results
4.1. Metric Comparisons
Fig. 3 shows the metric curves (i.e. metric plotted
against network density) for C, V , Q, R, L and H for
all archetypes as well as for the EEG dWPLI (red shade)
and coherence (blue shade) networks. From these plots we
see experimental evidence of maximal and minimal topolo-
gies for the given topological characteristics. These max-
imal and minimal topologies are explained as the curves
whose lines are consistently lowest or highest over all densi-
ties. Fractal Modular networks (purple lines) are maximal
for both C and Q (top left and centre left, respectively).
This is to be expected since the modules are complete
sub-networks with very few connections between modules,
maximising Q. Further this minimises the number of open
triples in the graph, maximising C, by restricting open
triples to relating only to those few connections which do
extend between modules. The star CWN (orange lines)
acts as a maximal topology for V , as expected from the
theory explained in Section 2, while being a minimal topol-
ogy for L (bottom left). Regular graphs, such as the ring
lattice network (blue lines), give 0 degree variance and hi-
erarchical complexity, thus are minimal topologies of these
features. The results of Fig. 3 for 30 node networks, found
in the supplementary material, follow the same relation-
ships, providing evidence that these features are indepen-
dent of network size.
Comparing the plots in Fig. 3 of C (top left) with
L (bottom left) and R (centre right) with H (bottom
right), it is immediately clear that L and H show extreme
behaviour at low densities while remaining consistent at
higher densities. This exemplifies how these metrics are
aimed at analysis of sparse networks, where it appears that
values can take a much greater range than for higher den-
sity networks.
To explore these comparisons further we perform sta-
tistical analysis with Wilcoxon rank sum tests on the dif-
ferences of distributions of metric values of EEG dWPLI
and E-R random networks as well as of EEG dWPLI and
EEG coherence networks (Fig 4). The results show that C
(right) and R (left) attain a greater range over edge den-
sity, P , of significant differences than their counterparts,
L and H. Particularly, R distinguishes differences from
1% up to 44% densities in the EEG dWPLI and coherence
comparison (solid blue line), whilst entropy only can dis-
tinguish differences from 1% up to 27% (solid yellow line).
Further, the z-scores indicate that in the range 1-27%, the
differences found in R are greater than those found using
H. Comparing the EEG dWPLI networks with E-R ran-
dom networks (Fig 4, left, dashed lines), both metrics find
differences at all levels, but the magnitude of difference
found by R (blue) is consistently greater than those found
by H (yellow). Thus, our metric outperforms entropy in
both magnitude and range of differences found.
Similarly, C finds a greater range and magnitude of dif-
ferences than L, Fig 4 right. In fact, C discerns differences
at all connection densities for the two comparisons, while
L fails to find differences after 62% in comparing dWPLI
and coherence networks (solid yellow line) and after 73%
in comparing dWPLI and random networks dashed yellow
line). Furthermore, L displays inverse differences at low
densities (1-12%) compared to higher densities in the dW-
PLI vs random comparison (dashed yellow line). This in-
consistency is undesirable for translatability of integrative
behaviour of network types from sparse networks to more
dense networks. C does not suffer from such behaviour,
displaying a constant relationship of metric values through
the full range of densities (solid and dashed blue lines).
Given these results, for the rest of our analysis, we will
drop L and H and focus on the four proposed metric, C, V ,
Q and R. We must emphasise that this is taken purely in
terms of the simplicity of explaining a general topological
factor and does not mean that L and H are not useful for
other purposes.
4.2. Weighted Complex Hierarchy Null Model
Fig. 5 shows the mean results of C (top left), V (top
right), Q (bottom left) and R (bottom right) over 100
realisations of each of the WCH models. We include a
reduced number of strength parameters in the figure (s =
0.1, 0.2, . . . , 0.7) than those computed (s = 0.05, 0.1, . . . , 0.75)
for greater clarity. Above 0.75 the parameter begins to sat-
urate as the weights of the hierarchy levels tend to linear
separability (linear separability occurs when s = 1 since
0s, 1s, 2s, ... then places the edge weights, originally in
[0, 1], in disjoint ranges [0, 1], [1, 2], [2, 3], ...). We see that
WCH networks (grey shaded lines) exhibit curve behaviour
similar to the EEG networks and E-R random graphs (as
in Fig. 3). The scale-free model (red error bars) also ex-
hibits a similar behaviour, however in stark contrast, the
small-world model (blue error bars) exhibits very different
7
Figure 3: Topological metric values for integration (C), regularity
(V ), modularity (Q), hierarchical complexity (R), characteristic path
length (L) and network entropy (H) against network density, P .
Curves relate to network models as indicated in the legend (bottom
right).
Figure 4: Positive (negative) values indicate the contrasted distribu-
tions exhibit the relationship provided in the legend (or its opposite).
Zero indicates p-value insignificant at 5% level. a) The hierarchical
complexity, R (blue), compared with network entropy, H (yellow).
b) The clustering coefficient , C (blue), compared with characteristic
path length, L (yellow). P is the network density.
behaviours than those of the EEG or WCH networks, ex-
hibiting a strong unsuitability for comparisons with EEG
networks with much higher modularity and highly right
skewed V curve (Fig. 5, top left) towards high densities
as well as a similar right skew in R (bottom left) which
is opposite to the left skew found for WCH and EEG net-
work types. Although the scale-free model exhibits similar
Figure 6: The z-statistics of distributions with significant differences
from a Wilcoxon rank sum test. Positive (negative) values indicate
the contrasted distributions exhibit the relationship provided in the
legend (or its opposite). Zero indicates p-value insignificant at 5%
level. P is the connection density.
works do indeed exhibit greater complexity than the WCH
model. The strong exception to this is an inability to dis-
tinguish significant differences between the maximal com-
plexity s = 0.3 WCH model and dWPLI networks within
7-23% densities (bold yellow line). Also, as the weight pa-
rameter increases, the high plateaus previously mentioned
begin to take effect as in the medium ranges of density the
R values of the dWPLI networks and WCH model becomes
more indistinguishable, with greater complexity found in
the range 55-57% in the WCH model with s = 0.4 (green
line).
4.3. Null model approaching EEG phase-lag networks
Fig. 7 shows the values of the four topological features-
complexity, integration, regularity and modularity for EEG
dWPLI networks and the WCH network with strength pa-
rameter 0.2. We see clearly that these networks behave
very similarly with respect to the given metrics. The most
obvious difference is that the modularity, Q, of dWPLI
EEG networks is higher (bottom left). Also, as previously
discussed, the dWPLI network complexity is greater than
the WCH model, but it is still by far the most comparable
model for complexity of those presented here.
5. Discussion
5.1. Complexity as revealed by weighted complex hierarchy
model
The behaviour demonstrated by the WCH model with
respect to R indicates that high complexity arises from
a hierarchical structure in which a greater degree of vari-
ability is present in the rankings of weights with respect
to hierarchy level. Too little difference between levels and
the hierarchy is too weak to maintain complex interactions,
too much difference between levels and the complexity of
the hierarchy is dampened by a more ordered structure
produced from the tendency towards linear separability
of the edge weights enforced by the strength parameter.
Thus, we provide evidence that topological complexity is
Figure 5: The topological characterisation of network models by clus-
tering coefficient, C, degree variance, V , modularity, Q and hierar-
chical complexity, R, plotted against network density, P . Grey lines
indicate mean values of the weighted complex hierarchy model with
increasing light shade indicating increasing strength parameter from
s = 0.1 in steps of 0.1 up to s = 0.7. Red errorbars indicate values of
the Albert Barabasi scale-free model. Blue errorbars indicate values
of the Watts Strogatz small-world model with increasingly light blue
indicating increasing proportion of edges being randomly rewired.
tendencies in topological metrics to the WCH and EEG
networks, its range of values and densities is clearly very
limited and so, therefore, its ability for topological refine-
ment.
By increasing the strength parameter of the WCH model
we change the topology in a smooth fashion with decreas-
ing integration, regularity and modularity (Fig. 5, top
left, top right and bottom left, respectively). Interestingly,
R (bottom left) rises with increasing strength parameter
from s = 0.05 up to s = 0.3 where it takes its maximum
values at densities ranging from 1-30% before falling again
from s = 0.35 until s = 0.7. Further, above s = 0.3,
the curves begin to deviate significantly from those of the
EEG dWPLI networks, exhibiting greater plateaus of high
complexity (lighter grey lines) which are more comparable
with the EEG coherence networks.
Interestingly, the complexity of the EEG dWPLI net-
works appears to attain maximal values of R of all the
networks studied here (Fig. 3). The only model which
comes close is the WCH model (Fig. 7, bottom right).
To clarify this observation we perform Wilcoxon rank sum
tests on R values of the EEG dWPLI networks against
that of the WCH model with strength parameters ranging
from s = 0.2 up to s = 0.4, i.e. two steps before and after
the maximal complexity setting of s = 0.3. The results
are displayed in Fig. 6. In the vast majority of instances
of strength parameter and density, the EEG dWPLI net-
8
are built independently from the brain networks [10, 15] or
are constructed by the randomisation of connections of the
networks being compared [9, 18], run into problems with
density specification (in the case of independent models)
and reproducibility (in both types of model). With the
WCH model, we can simply create a bank of simulated
CWNs which can be used throughout the study in exactly
the same way as we use the functional connectivity CWNs.
As an example of the power and elegance of the pro-
posed model, say we want to find maximum spanning trees
[45] of our brain networks and compare with a null model,
then we simply take the maximum spanning trees of our
null model. In contrast, in [14] they use a convoluted re-
verse engineering process by assigning random weights to
the connections of Watts-Strogatz small world networks
(which are themselves of limited comparability to brain
networks) and computing the MST from these resulting
sparse weighted networks.
Further, as seen in Fig. 1.C, for technical studies which
rely on network simulations, the WCH model is built on
parameters which can be altered to subtly change the re-
sulting topology. This allows for sensitive analysis of a
new techniques ability to distinguish subtle topological dif-
ferences. Such paradigms are evident in clinical studies
where, for example, one may try to distinguish between
healthy and ill patients [29, 33] or between different cog-
nitive tasks [46], so that this null model offers simulations
which are directly relatable to clinical settings.
5.3. EEG coherence and WPLI networks
We see there is a large difference in the integration,
modularity and complexity of the EEG coherence and dW-
PLI networks (Fig. 3, top left, centre left and centre right,
respectively). The EEG coherence networks (blue shade)
behave similarly to the ring (blue lines) and grid lattice
(yellow lines) networks, agreeing with the volume conduc-
tion effects that dominate zero-lag dependency measures
[43], i.e. the closer the nodes are the stronger the weights
are. The dWPLI networks (red shade) on the other hand
have a more integrated and less modular nature, which re-
flects the notion that phase-based functionality mitigates
volume conduction effects and is thus less confined by
anatomical structure [43].
The very high complexity of the dWPLI networks (and
very possibly phase-lag measures in general [47]) provides
evidence to support that phase-based connectivity does
indeed largely overcome the volume conduction effect and
therefore maintains a richer complexity echoing the com-
plex interactions of brain functionality [1].
With regards to how the WCH model advances our
understanding of dWPLI and coherence network differ-
ences, we note that the high segregation of the coherence
networks (Fig. 3 top left) is approached by the WCH
model with high values of strength parameter (Fig. 5, top
left) and is comparable with regular lattice and grid lat-
tice CWN curves (Fig. 3, top left, blue and yellow lines,
respectively), denoting a move to a more strict class-based
Figure 7: Clustering coefficient, C, degree variance, V , modularity,
Q, and complexity, R, against connection density, P , of binarised
weighted networks, for WCH model (red, median (m) ± interquartile
range (iqr)) and EEG dWPLI neworks (blue, m ± iqr).
not driven by integration, arising as a middle ground be-
tween regular and random systems as previously conjec-
tured [7, 9], but, driven by hierarchical complexity, arising
in the middle ground between weak hierarchical topology
or 'all nodes are equal' systems, such as random or regular
networks, and strong hierarchical topology, such as star or
strict class-based systems including grid lattice and frac-
tal modular networks (see Fig. 5). Thus the hierarchical
structure can be seen as a key aspect of the complexity
inherent in complex systems.
Impressively, the dWPLI EEG networks display a gen-
erally greater hierarchical complexity than that expressed
by our model which is specifically designed to probe com-
plex interactions in hierarchical structures. Thus we pose
such complexity as a key aspect of brain function as mod-
elled by phase-based connectivity.
5.2. Weighted complex hierarchy as null model
There are two clear reasons why the WCH model is
a good fit for functional connectivity networks from EEG
recordings. Not only does it create several hub like nodes
giving a high degree variability, but furthermore it sim-
ulates the rich club phenomena found in complex brain
networks [22, 44], as the higher the hierarchy levels of two
nodes, the stronger the weight of the connection will be
between them, see Fig. 1.C.
One of the greatest benefits of this model over others is
that it simulates brain networks previous to network pro-
cessing steps because it creates CWNs rather than sparse
networks. This means that any and all techniques one
wants to use on the brain networks can be applied el-
egantly and in parallel with this single null model free
from any complications. Particularly, methods which cre-
ate sparse binary networks directly, whether these models
9
topology. This is also reflected in the hierarchical com-
plexity (bottom right of corresponding figures), where the
lower complexity peaking at a later density to dWPLI (Fig.
3, centre right) is mimicked in the behaviour of increasing
strength parameter in the WCH model (Fig. 5, bottom
right). This provides further evidence of the relevance and
flexibility of the WCH model. In contrast there is an evi-
dent lack of ability to make similar comments with respect
to the popular small world and scale-free models. This
criticism can be extended towards network models which
randomise connections while maintaining degree distribu-
tions [18], since such an enforced topological attribute does
not allow one to analyse how that very important attribute
is actually constructed. Future work will provide extension
to modular structures in our model to focus on what roles
modularity plays on these aspects, since Q and V behave
contrastingly to this extrapolation.
5.4. Dense scale-free networks
A striking feature seen is in the degree variance curves
where a highly symmetric parabolic curve is noted with
a central maximum value for random graphs, WCH net-
works and EEG networks. This feature reveals to us a
'scale-free' paradigm at all density levels and not just the
classic sparse network scale-free at low densities. In other
words, the scale-free nature found in brain networks is first
and foremost encoded in the connectivity weights, which,
through selective binarisation, therefore can reveal to us
the scale-free property as expressed at different density
ranges. As the density of the network increases one obtains
more even distributions of high and low density nodes, in-
dicated by the high values of V , and, eventually, towards
high densities the symmetry of V values with low densi-
ties tells us that the scale-free network is characterised by
a small number of low degree nodes and a majority of high
degree nodes, i.e. the inverse (or complement) of the low
density behaviour.
5.5. Topological randomness
If we define a uniformly random topology as that which
exhibits a uniform distribution of topological values over
the space enveloped by the minimal and maximal topolo-
gies, it is very apparent that E-R random networks do not
satisfy this criteria, but, instead, have a restricted topol-
ogy at all density levels where the interquartile range is
much smaller in comparison with that of the EEG net-
works and the proposed null model. We thus see that
uniformly distributed random weights do not lead to a
uniformly random topology in this sense, but instead to a
very particular optimally integrated, moderately regular,
lowly modular and low complexity topology at all densi-
ties. Based on this evidence and previous discussion of
random networks in the methods section, we suggest that
E-R random networks should be re-understood as opti-
mally integrated networks.
Following from this the randomisation of connections
used widely in null models is not a topologically randomis-
ing process but, more accurately, a topologically integra-
tive process. Such a feature is then not necessarily typical
of network topology and thus one must be cautious to use
this as a null model unless one wants to specifically target
integrative behaviour. Further, the practice of normalisa-
tion of graph values by E-R random graph values [48] must
also be used with due caution. The basis of such a normal-
isation is to contrast a networks values with those of the
'average' network topology [49], rather than contrasting
with a highly specific topology which behaves very differ-
ently to real world networks [50]. This evidence provides
further justification for the adoption of our WCH model
as a relevant and powerful replacement to these models.
6. Conclusion
We introduced a metric for measuring the hierarchical
complexity of a network and a highly flexible and elegant
WCH model. These provided key insights into what distin-
guishes functional brain networks from both ordered and
spontaneous forms as generally the most complex kind of
topology and the important role that hierarchical struc-
ture plays in this. Further, we showed that phase-based
connectivity topology was more complex than amplitude
influenced connectivity topology, which we extrapolated
as due to the more ordered structure enforced by volume
conduction effects. In our analysis we constructed a frame-
work for CWNs for brain functional connectivity to replace
the framework for sparse networks adopted from other net-
work science research areas. This included the synthesis of
concepts from the literature in a succinct manner and the
generalisation of sparse binary archetypes to CWN form.
The perspective allowed by this comprehensive analysis
provided new evidence regarding key factors of network
topology in general. Importantly we provided evidence of
the non-topologically random nature of uniformly random
weighted networks. From this it follows that our model is
more relevant and appropriate than prevalent connection
randomisation processes. Also, a scale-free paradigm was
extended to all network densities. Particularly, these in-
sights help towards a comprehensive understanding of the
framework within which functional connectivity networks
are set and thus provide invaluable information and tools
for future clinical and technical research in neuroscience.
Matlab codes for all synthesis and analysis of the networks
as introduced in this paper are publicly available on pub-
lication at http://dx.doi.org/10.7488/ds/1520.
7. Acknowledgements
Keith Smith is funded by the Engineering and Physical
Sciences Research Council (EPSRC).
10
References
[1] E. Bullmore, O. Sporns, "Complex brain networks: graph the-
oretical analysis of structural and functional systems", Nature
Neuroscience Review, 10: 186-198, 2009.
[2] C.J. Stam, "Modern network science of neurological disorders",
Nature Neuroscience Review, 15: 683-695, 2014.
[3] D. Papo, J. M. Buldu, S. Boccaletti, E. T. Bullmore, "Complex
network theory and the brain", Phil. Trans. R. Soc. B, 369:
20130520, 2014.
[4] E. Ravasz and A.L. Barab´asi, "Hierarchical organization in com-
plex networks", Phys. Rev. E, 67: 026112, 2003.
[5] D. Meunier, R. Lambiotte & E. T. Bullmore, "Modular and hi-
erarchically modular organisation of brain networks", Frontiers
in Neurscience, doi: 10.3389/fnins.2010.00200, 2010.
[6] M. Costa, A.L. Goldberger, C.-K. Peng, "Multiscale entropy
analysis of biological signals", Phys. Rev. E, 71: 021906, 2005.
[7] G. Tononi, O. Sporns, G.M. Edelman, "A measure for brain
complexity: Relating functional segregation and integration in
the nervous system", Proc Natl Acad Sci, 91(11):5033-7, 1994.
[8] R.Sol´e & S. Valverde, "Information theory of complex networks:
on evolution and architectural constraints", Lect. Notes Phys.,
650: 189-207, 2004.
[9] D.J. Watts & S.H. Strogatz, "Collective dynamics of small-
world networks",Letters to Nature, 393: 440-442, 1998.
[10] O. Sporns, "Small-world connectivity, motif composition, and
complexity of fractal neuronal connections", BioSystems, 85:
5564, 2006.
[11] M. Rubinov, O. Sporns, "Weight-conserving characterization of
complex functional brain networks", Neuroimage, 56(4):2068-
2079, 2011.
[12] F. D.V. Fallani, J. Richiardi, M. Chavez, S. Achard, "Graph
analysis of functional brain networks: practical issues in trans-
lational neuroscience", Phil. Trans. R. Soc. B: Biological Sci-
ences, 369 (1653): 20130521, 2014.
[13] C. Li, H. Wang, W. de Haan, C.J. Stam, P. Van Mieghem, "The
correlation of metrics in complex networks with applications in
functional brain networks", J. Stat. Mech., doi:10.1088/1742-
5468/ 2011/11/P11018, 2011.
[14] P. Tewarie, E. van Dellen, A. Hillebrand, C.J. Stam, "The min-
imum spanning tree: an unbiased method for brain network
analysis", Neuroimage, 104: 177-188, 2015.
[15] A.L. Barabasi & R. Albert, "Emergence of Scaling in Random
Networks", Science, 286: 509-512,1999.
[16] M.E.J. Newman, "Networks", Oxford University Press, Oxford,
2010.
[17] O. Sporns, "Networks of the Brain", MIT Press, MA, 2010.
[18] M. Rubinov, O. Sporns, "Complex network measures of brain
connectivity: Uses and interpretations", NeuroImage, 52:1059-
1069, 2010.
[19] V.M. Eguiluz, D.R. Chialvo, G.A. Cecchi, M. Baliki, A. V. Ap-
karian, "Scale Free Brain Functional Networks", Physical Re-
view Letters, 94: 018102, 2005.
[20] T.A.B. Snijders, "The degree variance: an index of graph het-
erogeneity", Social Networks, 3(3): 163-174, 1981.
[21] A. Sandryhaila, J.M.F. Moura, "Discrete Signal Processing on
Graphs", IEEE Transactions on Signal Processing, 61(7):1644-
1656, 2013.
[22] J.J. McAuley, L.F. Costa, T.S. Caetano, "The rich-club phe-
nomena across complex network hierarchies", Appl. Phys. Lett.,
91, doi: 10.1063/1.2773951, 2007.
[23] P. Erdos & A. R´enyi, "On Random Graphs", Publicationes
Mathematicae Debrecen, 6:290-297, 1959.
[24] P. Bonacich, P. Lloyd, "Eigenvector-like measures of central-
ity for asymmetric relations", Social Networks, 23(3): 191-201,
2001.
[25] M. Molloy & B. Reed, "A critical point for random graphs with
a given degree sequence", Random Structures & Algorithms, 6
(2-3): 161-180, 1995.
[26] S. Milgram, "The Small World Problem", Psychology To-
day,1(1): 61-67, 1967.
[27] E. Bullmore, O. Sporns, "The economy of brain network organ-
isation", Nature, 13:336-349, 2012.
[28] A.L. Barabasi, R. Albert & H. Jeong, "Mean-field theory for
scale-free random networks", Physica A, 272(1-2): 173-187,
1999.
[29] B.M. Tijms, A.M. Wink, W. de Haan, W.M. van der Flier, C.J.
Stam, P. Scheltens, F. Barkhof, "Alzheimer's disease: connect-
ing findings from graph theoretical studies of brain networks",
Neurobiology of Ageing, 34: 2023-2036, 2013.
[30] J. Petersen, "Die Theorie der regularen Graphs", Acta Math.,
15: 193-220, 1891.
[31] B.M. ´Abrego, S. Fernndez-Merchant, M.G. Neubauer, W.
Watkins, "Sum of squares of degrees in a graph", Journal of
Inequalities in Pure and Applied Mathematics, 10(3): 64, 2009.
[32] C.J. Stam, B.F. Jones, G. Nolte, M. Breakspear, P. Schel-
tens, "Small-World Networks and Functional Connectivity in
Alzheimer's Disease", Cerebral Cortex, 17:92-99, 2007.
[33] M. Lynall, D. S. Bassett, R. Kerwin, P.J. McKenna, M.
Kitzbichler, U. Muller & E. Bullmore, "Functional Connectiv-
ity and brain networks in schizophrenia", J Neurosci, 30(28):
9477-9487, 2010.
[34] C.E. Ginestet, T.E. Nichols, E.T. Bullmore, A. Simmons,
"Brain network analysis:
separating cost from topology us-
ing cost-integration", PloS One, 6, e21570, doi: 10.1371/jour-
nal.pone.0021570, 2011.
[35] A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff,
P.C. Ivanov, R.G. Mark, J.E. Mietus, G.B. Moody, C.K.
Peng, H.E. Stanley, "PhysioBank, PhysioToolkit, and Phys-
ioNet: Components of a new research resource for complex
physiologic signals", Circulation, 101(23): e215-e220, dataset:
doi:10.13026/C28G6P, 2000.
[36] M.E.J. Newman, M. Girvan, "Finding and evaluating commu-
nity structure in networks", Phys. Rev. E, 69(2): 026113, 2004.
[37] M.E.J. Newman, "Modularity and community structure in net-
works", Phys Rev E, 23: 8577-8582, 2006.
[38] V.D. Blondel, J-L. Guillaume, R. Lambiotte, E. Lefebvre, "Fast
unfolding of communities in large networks", J. Stat. Mech., doi:
10.1088/1742-5468/2008/10/P10008, 2008.
[39] C.E. Shannon, "A mathematical theory of communication",
Bell System Technical Journal, 27: 623656, 1948.
[40] G. Schalk, D.J. McFarland, T. Hinterberger, N. Birbaumer, J.R.
Wolpaw, "BCI2000: A General-Purpose Brain-Computer Inter-
face (BCI) System", IEEE Transactions on Biomedical Engi-
neering, 51(6):1034-1043, 2004.
[41] R. Oostenveld, P. Fries, E. Maris and J-M. Schoffelen, "Field-
Trip: Open Source Software for Advanced Analysis of MEG,
EEG, and Invasive Electrophysiological Data", Computational
Intelligence and Neuroscience, Volume 2011, 156869, 9 pages,
2011.
[42] M. Vinck, R. Oostenveld, M. van Wingerden, F. Battaglia,
C.M.A. Pennartz, "An improved index of phase-synchronization
for electrophysiological data in the presence of volume-
conduction, noise and sample-size bias", NeuroImage, 55:1548-
1565, 2011.
[43] E. van Diessen, T. Numan, E. van Dellen, A.W. van der Kooi,
M. Boersma, D. Hofman, R. van Lutterveld, B.W. van Dijk,
E.C.W. van Straaten, A. Hillebrand, C.J. Stam, "Opportunities
and methodological challenges in EEG and MEG resting state
functional brain network research", Clinical Neurophysiology,
doi:10.1016/j.clinph.2014.11.018, 2014.
[44] M.P. van den Heuvel, O. Sporns, "Rich-Club Organisation of the
Human Connectome", Journal of Neuroscience, 31(44): 15775-
15786, 2011.
[45] C.J. Stam, P. Tewarie, E. Van Dellen, E.C.W. van Straaten, A.
Hillebrand, P. Van Mieghem, "The trees and the forest: Char-
acterization of complex brain networks with minimum spanning
trees", International Journal of Psychophysiology, 92: 129-138,
2014.
[46] K. Smith, H. Azami, J. Escudero, M.A. Parra, J.M. Starr,
"Comparison of Network Analysis Approaches on EEG Con-
nectivity in Beta during Visual Short-Term Memory Tasks",
11
Proceedings of the IEEE EMBC15, 2207 - 2210, 2015.
[47] J. Dauwels, F. Viallate, T. Musha, A. Cichocki, "A compar-
ative study of synchrony measures for the early diagnosis of
Alzheimer's disease based on EEG", NeuroImage, 49(1): 668-
693, 2010.
[48] M.D. Humphries, K. Gurney, "Network 'Small-world-ness':
A Quantitative Method for Determining Canonical Network
Equivalence", PLoS One, 3(4): e0002051, 2008.
[49] B. Bollob´as, "Random Graphs", ch.8 of "Modern Graph The-
ory", Graduate Texts in Mathematics, Springer New York, 184:
215-252, doi:0.1007/978-1-4612-0619-4 7, 1998.
[50] M.E.J. Newman, "Random graphs as models of networks", ch.2
from S. Bornholdt, H.G. Schuster,"Handbook of Graphs and
Networks: From the Genome to the Internet", Wiley, UK, DOI:
10.1002/3527602755, 2006.
12
|
1106.2048 | 1 | 1106 | 2011-06-09T18:30:52 | Old equations for a new system: A possible use of Navier-Stokes equations to model the circulation of spikes in the nervous system | [
"q-bio.NC"
] | In the present work we discuss a possible application of Navier-Stokes-based models to the quantitative description of the circulation of nervous impulses throughout the nervous system. In previous works we have shown that the discharge from Basal Ganglia neurons from patients with Parkinson's disease share mathematical features with the velocity fields of fluids under turbulence regimes. In the present work we try to build the fundaments for a physical analogy between both kinds of systems. | q-bio.NC | q-bio | Old equations for a new system:
A possible use of Navier-Stokes equations to model the circulation of spikes in
the nervous system
Daniela Sabrina Andres
Department of Physiology, Medicine School, University of Buenos Aires
CONICET
Institute for Neurological Research Raúl Carrea, Movement Disorders Section, Neuroscience
Department, FLENI
Montañeses 2325, C1428AQK, Buenos Aires, Argentina
[email protected]
Abstract
In the present work we discuss a possible application of Navier-Stokes-based models to the
quantitative description of the circulation of nervous impulses throughout the nervous system.
Velocity and momentum of spikes are not usually considered when modeling information
transmission in the nervous system, however in previous works we have shown that the discharge
from Basal Ganglia neurons from patients with Parkinson’s disease share mathematical features
with the velocity fields of fluids under turbulence regimes. In particular the properties that are
similar are related to the behavior of energy, and therefore fluid dynamics’ concepts might be
useful in the future for the understanding of the nervous system from a mathematical perspective.
In the present work we try to build the fundaments for a physical analogy between both kinds of
systems.
Introduction
To present date, the structure of the neural
code, or even if there is one, remains unknown.1,2
Several approaches have been proposed to solve
the riddle. Different works have discussed the idea
of information being coded in a rate code or a time
code.3,4 In the first case the exact time of
occurrence of a single spike could be random,
while in the second case a deterministic time
pattern is thought to code for information. In a
recent work of our group, it has been proposed that
information could also be coded in a scale code,
with different scales of the neuronal discharge
transmitting either complementary or redundant
information.5 Of course, different combinations of
the previous might be possible, and
the
universality of the code is also under discussion.
1
Different neural systems might use different codes
and even switch to different forms of coding under
different circumstances. 6 ,7
In the present work we will discuss a
general frame to analyze the circulation of nervous
impulses
throughout
the nervous system. In
previous works we have shown that properties
measured from the brain from patients with
Parkinson’s disease (PD) have similarities with
turbulent fluids.5,8,9 We have proposed that Navier-
Stokes-like equations might be useful for the
modeling of
the
in
transmission
information
central nervous system. In the present paper we
present a preliminary description of a physical
analogy and a possible way
in which
the
formalism of Navier-Stokes (NS) might apply for
the study of information transmission in a neural
system.
A brief discussion about the possible
consequences of the present work might be
appropriate. In clinical neurology the question
about the way in which neurons code information
the understanding and
becomes relevant for
modeling of
information
transmission
in
the
central nervous system. In particular in the case of
Parkinson’s disease, the application of Deep Brain
Stimulation therapy (DBS) affects the transmission
of nervous impulses through the basal ganglia-
loop.10,11 DBS
thalamic-cortical
(BG-Th-Ctx)
consists of chronically implanting a stimulation
electrode in a chosen neural center and it has
proven to succeed in the treatment of different
conditions.12 However,
to present date
the
mechanisms of action of DBS therapy are not fully
understood.13 ,14 This poor understanding of the
exact mechanisms of action of DBS has important
practical
consequences. Today,
there
exist
limitations in the outcome prediction of patients
with PD who are treated with DBS, and the
indications of this therapy are limited to the most
severe cases.15 Therefore the fact that we currently
lack a mathematical model of DBS poses
important limitations to this treatment. Based on
some experimental evidence presented in previous
works 5 ,8,9 we will discuss a possible analogy
between the circulation of spikes in the BG-Th-Ctx
loop and the dynamics of fluids.
Fundaments of the model
The present model is centered on the
description of the velocity of transmission of the
nervous impulse in the nervous system. Although
velocity is frequently not used to describe the
behavior of neural systems, it might possibly be a
crucial variable
for
the understanding of
transmission.16 To understand
information
the
situation that we are trying to represent, imagine a
simple neural system like the one in figure 1.
Suppose that a spatially homogeneous stimulus S1
is applied to three adjacent neurons, N1-3, forming
layer one (LI) of a neural network. All three
neurons receive the stimulus at the same time,
called ti, and we propose that the three of them will
respond with the generation of a single spike to the
stimulus. The problem that concerns us is at what
times, called t1, t2 and t3, the spikes originated by
S1 will arrive to the next layer of the network
travelling through the different neurons. (For
simplification we will consider here that the next
layer lies at a constant distance from layer one,
although the spatial architecture of physiological
neural systems adds a complexity level to the
problem that we are analyzing. This fact could be
modeled with the use of non-Bravais lattices, i.e.
integration meshes
without
translational
symmetry.) Assuming that spikes travel at the same
velocity through all the axons of the layer, t1, t2
and t3 will depend on the time consumed in
summation and therefore on the level of excitation
of each neuron. We assume that all the cells in a
layer have the same time and space constants and
firing threshold. Therefore changes in membrane
potential will occur at the same rate in each cell
when the depolarizing stimulus is applied and, if
the stimulus is enough to produce a supra-
Figure 1) Although the velocity of spikes is a quantity that is not commonly discussed, it might well describe the circulation of
impulses in the nervous system. Suppose that three adjacent neurons are affected simultaneously by a spatially homogeneous
stimulus at an initial time t, and assume that all three neurons have the same time and space constants and firing threshold and a
constant axonal length. If the three neurons respond with the generation of a single spike to the stimulus, then the times t1, t2 and t3
at which the generated spikes will arrive to the next neuronal layer do not need to be the same, but will depend on the exact level of
excitation at of each neuron at time t.
2
threshold depolarization, then the time at which
the response will be generated depends on how far
the membrane potential lies from the threshold at
the moment the cell received the input. It can be
easily seen that this difference in the initial
excitation level of each neuron will translate into a
difference between t1, t2 and t3, a phenomenon that
can actually be interpreted as a change of velocity
of the impulse if we consider the network globally.
This acceleration effect could be potentiated or
attenuated in the next layer, adding variability to
the velocity of transmission of spikes in the
nervous system and finally generating the complex
interspike intervals time series that are usually
measured in neural systems.
So far we have described a simple fact
about neural transmission that is often not taken
into account. This is also true for another simple
fact that needs to be considered prior to the
discussion of the present model. Given two
neurons N1 and N2, any of them will have a
greater probability of applying a positive
acceleration to an impulse I if the other one is
doing so. This is equivalent to saying that the
velocity through a given neuron will be minor if
the adjacent neuron
is not applying any
acceleration to the impulse. This can be explained
with our current knowledge of extracellular
modulation of synaptic activity, a
function
conducted primarily by astrocytes. 17-19 The
importance of the finite extracellular space and its
chemical composition in neural conduction is
being acknowledged progressively, and
is a
possible mechanism of coupling between adjacent
cells. 20,21 When considering a large network of
neurons this tendency of adjacent neurons to show
similar acceleration values will have the effect of
helping a spatially homogeneous stimulus
to
remain so. In other words, while acceleration and
deformation in time and space of the stimulus
travelling through the network is possible, the
stimulus will essentially resist to this deformation.
We will build an analogy between this fact and the
viscosity of fluids, and will call this property of
stimuli travelling through large ensembles of
neurons, µ .
Having stated
the velocity of
that
transmission of an impulse from one layer to the
next one is not necessarily constant and that a
spatially homogeneous stimulus naturally resists
deformation, we have the bases needed to apply
the present NS-based model to the transport of
nervous stimuli through the neural network that we
are considering.
We can write the well known Navier-
Stokes equations as follows:
P
µ
∇+
2
r
v
(2)
(1)
−∇=
r
=
0
g
r
∂⇒
v
ρ
∂
t
Where µ is the viscosity, P stands for
pressure and v for velocity. We will consider g = 0
in order to ignore the gravity term. To make use of
this equation for an interpretation of the movement
of nervous impulses in the nervous systems, we
will define ρ as the density of impulses I in a given
volume of nervous tissue. Pressure can also be
considered as the energy of the system per unit of
volume. For that reason the best analogy possible
in the nervous system should be a function of the
excitation level of the neuron. In time, the level of
excitation can be described as a function of the
sum of synaptic weights
to make
in order
comparisons with other models. Since the general
state of excitation depends on the sum of synaptic
weights of all the synapses present in the system,
we can write excitation, e, as a function of this
sum:
=1i
= ∑
eP
m
iw
(3)
What was said to this point allows us to
state that the propagation of the nervous impulse
through the nervous system requires energy per
volume unit. We have already stated that µ will
account for the resistance to deformation of a
spatially homogeneous stimulus. These ideas are
just being formulated in a conceptual way in the
present work, and experimental research needs to
follow in order to describe in a quantitative way
these properties.
A few words need to be said in reference
to a possible comparison to experimental data.
Interspike intervals (ISIS) are a common, robust
and reliable measure used for the analysis of
experimental data in neural systems. Since the
detection of spikes is a robust procedure itself,
these time series have low levels of noise, making
them suitable for
the analysis of non-linear
properties. Interspike intervals time series are
3
constructed using the differences between the
times of appearance of spikes, and are therefore a
measure of time. A practical difficulty that we will
encounter when trying to apply NS-based model to
the circulation of impulses in the nervous system is
that most works based on NS models analyze
velocity fields. For that reason a way to compare
time and velocity would be desirable, and some
transformation between both variables might be
needed. This transformation can be performed
making use of single cell recordings. When
analyzing single unit data, it can be considered that
every spike is generated at the same point of space.
This allows us to take distance as 1, and so a
relationship of x-1 relates both variables. Of course,
some normalization would be needed when
working with velocity fields in more than one
dimension. Then the inverse relationship would be
to the norm of vector velocity, and the relationship
would be:
),(
txI
ISI
v
=
−=
t
i
∆
x
∆
t
,
t
+
1
i
∆
y
∆
t
,
∆
z
∆
t
(4)
(5)
(6)
∆
x
LI
−
LII
∆=
y
LI
−
LII
∆=
z
LII
LI
−
=
1
(7)
⇒
v
LI
−
LII
=
3
ISI
(8)
Here I(x,t) describes the impulse as a function of
time and space and ∆xLI-LII, ∆yLI-LII and ∆zLI-LII refer
to the distance in three dimensional space between
neuronal layers I and II (we are essentially
considering a cube of side =1).
Discussion
During the last decades our understanding
of BG physiology and pathophysiology has been
challenged by the introduction of DBS therapy. 22-
23 The observations made in the context of
stimulation treatment of human patients with PD
as well as functional neurosurgery showed many
inconsistencies and paradoxes with the classical
Albin-De Long’s model of the BG, opening the
way for new models, different than this classical
“box and arrow” view.24 There are many important
points that the classical model fails to account for.
First of all, it has been shown that high frequency
DBS of the GPi is effective in the treatment of
4
hyperkinetic (chorea) as well as hypokinetic (PD)
disorders in exactly the same way. 25-26 This fact is
contradictory with the notion that hyperkinetic and
hypokinetic diseases are based on opposite
pathogenic mechanisms. Secondly, stimulation of
almost any nuclei of the BG-Th-Ctx loop yields
positive results in the treatment of at least some
PD symptoms. This includes stimulation of GPi,
GPe, ventrolateral Thalamus, subthalamic nucleus,
motor cortex and even zona incerta. 27-31 Finally, it
has been described that HF stimulation of the GPi
increases the activity of this nucleus, which is
exactly the mechanism proposed as the origin of
the symptoms in the classical model of PD. 32-35 In
this context, new hypothesis have been made about
the functioning of the normal and diseased BG.
Current models include dimensionality reduction
models,
action-selection
devices,
learning
reinforcement and some neural network models. 36-
40 These models capture some properties of the
dynamics, but up to now none of them has proven
successful enough to accurately predict outcome
and allow a better programming of the devices. It
has also been proposed that non-linear features
might be the basis of the functioning of the BG-
Th-Ctx loop and that some of these properties
might be altered in PD. 41-45 However, currently
there is not a mathematical model that allows
representing and analyzing these properties in a
quantitative fashion.
In previous works we have found some
interesting mathematical similarities between the
ISIS time series of parkinsonian GPi neurons and
velocity fields of fluids undergoing turbulence. 5 ,8,9
In particular, an exponential decay of the power
spectrum makes it possible to talk about critical
scale length, a property that is not shared by many
systems in the nature. This form of decay of the
spectrum is typical from turbulent fluids, and
follows an analytic solution of Navier-Stokes
equations when energy remains bounded. 46,47
Therefore it could be thought that energy behaves
in a similar manner in both systems. Other
similarities that we have reported refer to the
existence of positive Lyapunov exponents 48 ,49 and
a scaling of the structure function that is typical
from turbulent fluids too 50 and that lead to the
description of a multifractal behavior in the signals
studied. In light of the evidence found up to this
point it seems reasonable to speculate that the
equations underlying the dynamics of the neural
system under discussion might be similar in form
to those that better describe the behavior of fluids:
Navier-Stokes equations. A further step is what we
have tried in the present work, and it consists of
looking for a physical analogy that explains the
sense in which each term of the equations might
apply for neural systems.
Another type of neurophysiological model
are neural fields. These are concerned with firing
rates more than with velocity, 51-55 although
sometimes they have been used to describe the
changes of velocity present
in
the nervous
system16. However they are mainly based on
reaction-diffusion kinds of systems and work
under the assumption of density heterogeneities, a
description that lacks of a strong physiological
basis. In turn these heterogeneities are usually
modeled as continuous functions of space, a
condition that does not seem very physiological
either. So we can say confidently that up to now
there is not a model that handles the behavior of
the velocity of spikes in the nervous system in a
quantitative and physiological enough manner.
Currently there are several experimental
approaches to the study of PD and DBS. Under
some experimental settings, neurons from the basal
ganglia from patients with PD have shown an
oscillatory behavior.56-59 Under this evidence it has
been proposed that a pathologic increment of
oscillatory activity or excessive synchronization
between neurons might be
the pathogenic
mechanism of PD.60 However, other works have
shown the mentioned neurons to behave in a
complex fashion, with an irregular output.8,43 The
fundamental problem is to what extent these
different works are comparable. Probably at least
some of the mathematical properties of the signals
studied depend on the experimental settings used,
such as in vitro vs. in vivo experiments, the level
of external input of the neuron, the type of
anesthesia used, etc. 61-65 In the previous works by
our group that we have mentioned before, we have
worked with data from GPi neurons form patients
with PD. This data were
recorded during
functional neurosurgery, with the patients awake
and under local anesthesia only. Because of that
reason the findings observed might be due at least
to some extent to the state of consciousness of the
patients, which defines the high dimensional input
from the neurons studied at the times of the
recordings. Since
the mathematical properties
5
analyzed were described under these particular
experimental conditions, some questions follow
that will need to be considered in future works. A
control group of a healthy population of neurons
will be needed. For
the reasons
that were
discussed, the best control group would probably
consist of recordings performed on animals awake
with microelectrode implanting techniques. This
experiment
together with
the
analysis of
anesthetized animals could help elucidate if the
observations under discussion are due to the
presence of PD
itself or
to
the state of
consciousness of the animal and to what extent.
Of course, the hypothesis presented here
need to be tested. Many directions of research
follow from the idea that NS-based models might
be useful for the modeling of the nervous system.
First of all, an analytic work might be useful to test
the relation between ISIS and velocity fields, i.e. to
test if the results already known for velocity fields
and the properties observed for ISIS time series
hold under an x-1 transformation. Secondly, to
perform simulations for simple conditions might
be useful, obtaining a comparable measure from
them (times between events). For example, tissue
infarctations from different sizes and shapes might
be a good starting point to run simulations
(different geometric boundary conditions). To
model the effect of DBS it would be probably
necessary to work with the pressure term of the
equations. Since DBS stimulates the nervous tissue
with a pulse of electric
tension,
it can be
interpreted that it produces a change in the level of
energy of the tissue (by affecting ionic gradients
for example), and an analogy could be easily built
with
the pressure of
fluids. Finally, new
experimental approaches could be design to study
the movement of impulses in nervous tissue from
the perspective of velocity. However, and although
new experiments might be desirable, it needs to be
considered that experimental complexity needs to
be kept to a minimum when working with human
patients. For that reasons new experiments might
be necessarily conducted in laboratory animals, or
even more brain slices or cell cultures and, again,
comparison with data from human brains will
become a problem to be considered. Finally, if the
presented approach turns out to be useful for the
modeling of the BG, the universality of the model
will become a question, and other neural systems
under different conditions (awake, anesthetized, in
vivo vs. in vitro) will need to be tested.
8. Andres DS, Cerquetti D, Merello M. Finite dimensional
structure of the GPi discharge in patients with Parkinson’s
Conclusion
In the present work we do not present
simulations or computational results. We have
explained the fundaments of a possible application
of Navier-Stokes-based equations for the modeling
of the movement of nervous impulse in the
nervous system. Velocity and momentum of spikes
are not usually considered when modeling
information transmission in the nervous system,
but
recently we have
found evidence
that
interspike
time series share some
intervals
mathematical properties with the behavior of
velocity fields of fluids under turbulence. In
particular these properties are related to the
behavior of energy, and therefore fluid dynamics’
concepts might be useful in the future for the
understanding of the nervous system, both from a
mathematical perspective and helping us to find
new physical analogies.
References
1. Shadlen MN, NewsomeWT. Noise, neural codes and
cortical organization. Curr Opin Neurobiol 1994; 4(4):569–
79.
2.
Stein RB, Roderich Gossen ER, Jones KE. Neuronal
variability: noise or part of the signal? Nat Rev 2005;
6:389–97.
disease. Int J Neural Syst 2011; in press.
9. Andres DS, Bocaccio H, Cerquetti DF, Merello M. On the
multiple scale of the neural code. 2011; under review.
10. Hashimoto T, Elder CM, Okun MS, Patrick SK, Vitek JL.
Stimulation of the subthalamic nucleus changes the firing
pattern of pallidal neurons. J of Neurosc 2003; 23(5): 1916
–23.
11. Lanotte MM, Rizzone M, Bergamasco B, Faccani G,
Melcarne A, Lopiano L. Deep brain stimulation of the
subthalamic nucleus: anatomical, neurophysiological and
outcome correlations with the effects of stimulation. J
Neurol Neurosurg Psychiatry 2002; 72: 53-58.
12. Pereira EAC, Green AL, Nandi D, Aziz TZ. Deep Brain
Stimulation: indications and evidence. Expert Reviews
2007; 4(5): 591-603.
13. Montgomery Jr. EB, Gale JT. Mechanisms of action of
Deep Brain Stimulation (DBS).
Neuroscience and
Biobehavioral Reviews 2008; 32: 388–407.
14. Nambu A. Seven problems on the basal ganglia. Curr Opin
Neurob 2008; 18(6): 595–604.
15. Rodriguez-Oroz MC, Obeso JA, Lang AE, Houeto JL,
Pollak P, et al. Bilateral deep brain stimulation in
Parkinson’s disease: a multicentre study with 4 years
follow-up. Brain 2005; 128: 2240-49.
16. Bressloff PC. Traveling fronts and wave propagation
failure in an inhomogeneous neural network. Physica D
(2001); 155: 83-100.
17. Newman EA. New roles for astrocytes: regulation of
3. Bialek W, Rieke F, de Ruyter van Steveninck RR, Warland
synaptic transmission. Trends in Neurosci 2003; 26(10):
D. Reading a neural code. Science 1991; 252(5014):1854-
57.
4. Sejnowski TJ. Time for a new neural code? Nature 1995;
376:21.
5. Andres DS, Cerquetti D, Merello M. Turbulence in Globus
Pallidum neurons in patients with Parkinson’s disease:
Exponential decay of the power spectrum. J Neurosci Meth
2011; 197(1): 14–20.
6. Mehta MR, Lee AK, Wilson MA. Role of experience and
oscillations in transforming a rate code in a temporal code.
Nature 2002; 417:741-46.
536-42.
18. Araque A, Carmignoto G, Haydon PG. Dynamic signaling
between astrocytes and neurons. Ann Rev Phys 2001; 63:
795-813.
19. Fellin T. Communication between neurons and astrocytes:
relevance to the modulation of synaptic and network
activity. Jof Neuroch 2009; 108(3): 533-44.
20. Gloor SM. Relevance of NA,K-ATPase
to
local
extracellular potassium homeostasis and modulation of
synaptic transmission. FEBS Letters 1997; 412(1):1-4.
21. Fields RD, Stevens B. ATP: an extracellular signaling
molecule between neurons and glia. Trends in Neurosci
7. Fairhall AL, Lewen GD, Bialek W, de Ruyter van
2000; 23(12):625-33.
Steveninck RR. Eficiency and ambiguity in an adaptive
neural code. Nature 2001; 412:787-92.
6
22. Wichmann T, DeLong MR. Deep brain stimulation for
34. Albin RL, Young A, Penny JB. The functions anatomy of
neurologic and neuropsychiatric disorders. Neuron 2006;
basal ganglia disorders. Trends in Neurosci 1989; 12: 366-
52: 197–204.
75.
23. Sestini S, Ramat S, Formiconi AR, Ammannati F, Sorbi S,
35. DeLong MR. Primate models of movement disorders of
Pupi A. Brain Networks Underlying the Clinical Effects of
basal ganglia origin. Trends in Neurosci 1990; 13: 281–85.
Long-Term Subthalamic Stimulation
for Parkinson’s
Disease: A 4-Year Follow-up Study with rCBF SPECT. J
Nucl Med 2005; 46:1444–54.
24. Bar-Gad I, Bergman H. Stepping out of
the box:
information processing in the neural networks of the basal
ganglia. Current Opinion in Neurobiology 2001; 11: 689–
95.
25. Montgomery
Jr. EB. Deep brain
stimulation
for
hyperkinetic disorders. Neurosurgical Focus 2004; 17: E1.
36. Berns GS, Sejnowski TJ. A Computational Model of How
the Basal Ganglia Produce Sequences. Journal of Cognitive
Neuroscience 1998; 10(1): 108-121.
37. Gurney KN, Prescott TJ, Redgrave P. The Basal Ganglia
viewed as an Action Selection Device. Proceed of the
Eighth Int Conf on Artif Neural Net, September 2-4 1998,
Skövde, Sweden, 1033–38.
38. Terman D, Rubin JE,. Yew AC, Wilson CJ. Activity
Patterns in a Model for the Subthalamopallidal Network of
26. Montgomery Jr. EB. Dynamically coupled, high-frequency
the Basal Ganglia. J of Neurosci 2002; 22(7): 2963–76.
reentrant, non-linear oscillators embedded in scale-free
basal
ganglia-thalamic-cortical
networks mediating
function and deep brain stimulation effects. Nonlinear
Studies 2004; 11: 385–421.
27. Deep Brain Stimulation in Parkinson’s Disease Group.
Deep-brain stimulation of the subthalamic nucleus or the
pars interna of the globus pallidus in Parkinson’s disease.
39. Bar-Gad I, Morris G, Bergman H. Information processing,
dimensionality reduction and reinforcement learning in the
basal ganglia. Prog in Neurobiol 2003; 71: 439–473.
40. Doya K. What are the computations of the cerebellum, the
basal ganglia and the cerebreal cortex? Neural Networks
1999; 12: 961-9.
New England and Journal of Medicine 2001; 345: 956–63.
41. Darbin O, Soares J, Wichmann T. Nonlinear analysis of
28. Koller W, Pahwa R, Busenbark K, Hubble J, Wilkinson S,
et al. High-frequency unilateral thalamic stimulation in the
discharge patterns in monkey basal ganglia. Brain Res
2006; 1118: 84–93.
treatment of essential and parkinsonian tremor. Ann of
42. Lim J, Sanghera MK, Darbin O, Stewart RM, Jankovic J,
Neurol 1997; 42: 292–99.
29. Vitek JL, Hashimoto T, Peoples J, DeLong MR, Bakay
AE. Acute stimulation in the external segment of the
Simpson R. Nonlinear temporal organization of neuronal
discharge in the basal ganglia of Parkinson’s disease
patients. Exp Neurol 2010; 224(2): 542–44.
globus pallidus
improves Parkinsonian motor signs.
43. Obeso JA, Rodríguez-Oroz MC, Rodríguez M, Lanciego
Movement Disorders 2004; 19: 907–15.
JL, et al. Pathophysiology of the basal ganglia in
30. Canavero S, Bonicalzi V, Paolotti R, Castellano G, Greco-
Crasto S, Rizzo L, et al. Therapeutic extradural cortical
Parkinson’s disease. Trends Neurosci 2000; 23(Suppl): S8–
S19.
stimulation
for movement
disorders:
a
review.
44. Dorval AD, Russo GS, Hashimoto T, Xu W, et al. Deep
Neurological Research 2003; 25: 118–22.
brain stimulation reduces neuronal entropy in the MPTP
31. Plaha P, Khan S, Gill SS. Bilateral stimulation of the
caudal zona incerta nucleus for tremor control. J of Neurol
primate model of Parkinson’s disease. J Neurophysiol
2008; 100(5): 2807–18.
Neurosurg and Psychi 2008; 79: 504-513.
45. Lafreniere-Roula M, Darbin O, Hutchison WD, Wichmann
32. Anderson ME, Postupna N, Ruffo M. Effects of high-
frequency stimulation in the internal globus pallidus on the
activity of thalamic neurons in the awake monkey. J of
Neurophys 2003; 89: 1150–60.
33. Montgomery Jr. EB. Effects of globus pallidus interna
stimulation on human thalamic neuronal activity. Clinical
Neurophys 2006; 117: 2691–2702.
T, et al. Apomorphine reduces subthalamic neuronal
entropy in parkinsonian patients. Exp Neurol 2010; 225:
455–58.
46. Bercovici H, Constantin P, Foias C, Manley OP.
Exponential decay of the power spectrum of turbulence. J
Stat Phys 1995; 80:579–602.
47. Doering CR, Titi ES. Exponential decay rate of the power
spectrum or solutions of the Navier–Stokes equations. Phys
Fluids 1995; 76:1384–90.
7
48. Aurell E, Boffetta G, Crisanti A, Paladin G, Vulpiani A.
Growth of noninfinitesimal perturbations in turbulence.
PRL 1996;77(7):1262–5.
49. Lapeyre G. Characterization of finite-time Lyapunov
exponents and vectors in twodimensional turbulence.
Chaos 2002;12(3):688–98.
50. Lin DC, Hughson RL. Modeling Heart Rate Variability in
Healthy Humans: A Turbulence Analogy. Phys Rev Lett
2001; 86: 1650–53.
51. Amari S. Dynamics of pattern formation in
lateral
inhibition type neural fields. Biol Cyb 1977; 27: 77–87.
52. Amari S. Homogeneous nets of neuron-like elements. Biol
Cyb 1975; 17: 211–20.
53. Ermentrout GB, Kleinfeld D. Traveling electrical waves in
cortex: Insights from phase dynamics and speculation on a
computational role. Neuron 2001; 29: 33–44.
54. Nunez PL. The brain wave equation: a model for the EEG.
Math Biosci 1974; 21: 279–97.
55. Pinto DJ, ErmentroutGB. Spatially structured activity in
synaptically coupled neuronal networks: I. Travelling
fronts and pulses. SIAM J on Appl Math2001; 62: 206–25.
56. Gatev P, Darbin O, Wichmann T. Oscillations in the basal
ganglia under normal conditions and
in movement
disorders, Movement Disord 2006; 21(10): 1566–77.
57. Tseng KY, Kasanetz F, Kargieman L, Riquelme LA, Murer
MG. Cortical slow oscillatory activity is reflected in the
membrane potential and spike trains of striatal neurons in
rats with chronic nigrostriatal lesions, J Neurosci 2001;
21(16): 6430–39.
58. Stanford IM. Independent neuronal oscillators of the rat
Globus Pallidus. J Neurophysiol 2003; 89: 1713–17.
59. Levy R, Dostrovsky JO, Lang AE, Sime E, Hutchison WD,
Lozano AM. Effects of Apomorphine on subthalamic
nucleus and Globus Pallidus Internus neurons in patients
with Parkinson’s disease. J Neurophysiol 2001; 86: 249–
60.
60. Brown P. Bad oscillations in Parkinson’s disease. J Neural
Transm 2006; 70(Suppl.): 27–30.
61. Baddeley R, Abbott LF, Booth MCA, Sengpiel F, Freeman
T, Wakeman EA, et al. Responses of neurons in primary
and inferior temporal visual cortices to natural scenes. Proc
R Soc Lond B 1997; 264: 1775–83.
62. Bair W, Koch C, NewsomeW, Britten K. Power spectrum
analysis of bursting cells in area MT in the behaving
monkey. J Neurosci 1994; 74(5): 2870–92.
8
|
1906.00511 | 2 | 1906 | 2019-06-07T03:13:24 | Deep learning from wristband sensor data: towards wearable, non-invasive seizure forecasting | [
"q-bio.NC",
"eess.SP"
] | Seizure forecasting may provide patients with timely warnings to adapt their daily activities and help clinicians deliver more objective, personalized treatments. While recent work has convincingly demonstrated that seizure risk assessment is possible, these early approaches relied largely on complex, often invasive setups including intracranial electrocorticography, implanted devices and multi-channel EEG, which limits translation of these methods to broad clinical application. To facilitate broader adaptation of seizure forecasting in clinical practice, non-invasive, easily applicable techniques that reliably assess seizure risk, in combination with clinical information, are crucial. Wristbands that continuously record physiological parameters, including electrodermal activity, body temperature, blood volume pressure and actigraphy, may afford monitoring of autonomous nervous system function and movement relevant for such a task, hence minimizing potential complications associated with invasive monitoring, and avoiding stigma associated with bulky external monitoring devices on the head. Here, we use deep learning to analyze long-term, multi-modal wristband sensor data from 50 patients with epilepsy (total duration $>$1400 hours) to assess its capability to distinguish preictal from interictal states. Prediction performance is assessed using area under the receiver operating charateristic (AUC) and improvement over chance (IoC) based on F1 scores. Using one- and two-dimensional convolutional neural networks, we identified better-than-chance predictability in out-of-sample test data in 60\% of the patients in leave-one-out and 43\% of patients in pseudo-prospective approaches. These results provide a step towards developing easier to apply, non-invasive methods for seizure risk assessments in patients with epilepsy. | q-bio.NC | q-bio | Deep learning from wristband sensor data: towards
wearable, non-invasive seizure forecasting
Christian Meisel1,2∗
Rima El Atrache 1
Michele Jackson 1
Sarah Schubach 1
Claire Ufongene 1
Tobias Loddenkemper1
1 Boston Children's Hospital, Boston, USA
2 Department of Neurology, University Clinic Carl Gustav Carus, Dresden, Germany
∗ corresponding author email address: [email protected]
9
1
0
2
n
u
J
7
]
.
C
N
o
i
b
-
q
[
2
v
1
1
5
0
0
.
6
0
9
1
:
v
i
X
r
a
1
Seizure forecasting may provide patients with timely warnings to adapt their daily
activities and help clinicians deliver more objective, personalized treatments. While
recent work has convincingly demonstrated that seizure risk assessment is possi-
ble, these early approaches relied largely on complex, often invasive setups includ-
ing intracranial electrocorticography, implanted devices and multi-channel EEG,
which limits translation of these methods to broad clinical application. To facilitate
broader adaptation of seizure forecasting in clinical practice, non-invasive, easily
applicable techniques that reliably assess seizure risk, in combination with clinical
information, are crucial. Wristbands that continuously record physiological param-
eters, including electrodermal activity, body temperature, blood volume pressure
and actigraphy, may afford monitoring of autonomous nervous system function
and movement relevant for such a task, hence minimizing potential complications
associated with invasive monitoring, and avoiding stigma associated with bulky
external monitoring devices on the head. Here, we use deep learning to analyze
long-term, multi-modal wristband sensor data from 50 patients with epilepsy (to-
tal duration >1400 hours) to assess its capability to distinguish pre- from interictal
states. Prediction performance is assessed using area under the receiver operating
charateristic (AUC) and improvement over chance (IoC) based on F1 scores. Using
one- and two-dimensional convolutional neural networks, we identified better-than-
chance predictability in out-of-sample test data in 60% of the patients in leave-one-
out and 43% of patients in pseudo-prospective approaches. These results provide
a step towards developing easier to apply, non-invasive methods for seizure risk
assessments in patients with epilepsy.
2
Introduction
Reliable methods to assess seizure risk could alleviate a major burden for epilepsy patients by
providing timely warning or relief when seizure risk is high or low. From a clinician perspec-
tive, robust seizure risk assessments are desirable because of their ability to improve treatment
by optimizing dosing and timing of antiepileptic drug regimen by objective, personalized stan-
dards, as well as by potentially enabling timely interventions to avert impending seizures.1
Following initial attempts,2 there has been a recent surge of studies demonstrating the pos-
sibility of accurate seizure forecasting.3 -- 6 To this end, most studies have utilized either elec-
trocorticography (ECoG) or scalp electroencephalography (EEG) as well as, to a lesser extent,
electrocardiography (ECG), and have demonstrated that robust differentiation between preictal
and interictal states is possible with a performance better than chance.7 -- 10 In order to make
seizure risk assessments available for broader clinical use, however, methods that build on
non-invasive, easily recordable data streams are desirable.11 Peripheral signals recorded us-
ing wearable devices, such as wristbands, are particularly interesting in this respect since these
signals permit continuous, non-invasive recording of several physiological parameters. At the
same time, the compact design may limit the risk of stigmatization, are easy to apply, and may
altogether increase patient adherence relevant for long-term ambulatory use.
Continuous and simultaneous monitoring of a range of physiological parameters, such as
electrodermal activity, body temperature, blood volume pressure and actigraphy, using wrist-
bands is becoming increasingly available and permits close monitoring of autonomous nervous
system function and movement,12 and may assist in the detection of generalized tonic-clonic
seizures.13
Similar autonomous system measures may also provide information on detection of preictal
patterns or states. Deep learning has been shown to exhibit strong classification performance
3
from complex feature sets.14 It therefore constitutes a promising technique to differentiate pre-
from interictal states from complex, multi-modal wristband data. While more traditional ma-
chine learning approaches rely on hand-designed feature sets, deep learning uses multiple layers
of connections to perform classification tasks without the need of feature designing, which may
be an advantage in relatively under-explored, multi-modal datasets, such as data from wristworn
devices.
In this work, we use convolutional neural networks (CNNs) on a unique dataset comprised
of multi-modal wristband sensor data recorded from patients with epilepsy during multi-day in-
hospital monitoring. Our aim is to evaluate the utility of this novel data and methods approach
in its ability to differentiate pre- from interictal states as a prerequisite for future, non-invasive
seizure forecasting methods.
Materials and Methods
Data recording and preprocessing
We recruited patients with epilepsy admitted to the long-term video-EEG monitoring (LTM)
unit and placed a biosensor wristband (E4, Empatica12) on either left or right wrist or ankle for
long-term recording. For the purpose of this study we considered all patients with wristband
recordings from 02/2015 until 11/2018. For the purpose of this study, we considered data
from one wristband per patient only. When patient recording involved multiple wristbands (e.g.
from wrist and ankle), we selected the data from the biosensor wristband with the longest total
recording time for further analysis.
A prerequisite for seizure risk assessment is the reliable distinction between pre- and inter-
ictal states. We aimed to differentiate pre- from interictal states based on 30-second wristband
recordings composed of six sensor data streams (electrodermal activity (EDA), accelerometer
data in three dimensions, blood volume pressure (BVP), and temperature (TEMP); Fig. 1 A).
4
We considered a 30-second data segment as preictal if it occurred between 61 minutes and one
minute prior to a seizure, thus leaving a one minute buffer prior to seizure onset (Fig. 1 B, red).
Electrographic seizure onset was determined using video and EEG recordings. We analyzed all
epileptic seizure types occurring in a patient, which included subclinical seizures, focal, primary
and secondary tonic-clonic, myoclonic, clonic, tonic, atonic seizures and epileptic spasms (Fig-
ure 4). Only seizures that occurred two hours or more from a preceding seizure were included
to limit our analysis on lead seizures. 30-second data segments were classified as interictal if
they occurred two hours or more from any seizure (Fig. 1 B, green). To allow stable recording
conditions, we removed data from the first and last hour of each recording.
Separation in training, validation and test data
To evaluate classification performance on out of-sample test data, we used two separate ap-
proaches, a leave-one-out approach and a pseudo-prospective approach.
Leave-one-out approach
A leave-one-out cross-validation approach has the advantage that all seizures can be used for
training and validation. Here, we used a leave-one-out approach for each seizure. Similar to,15
if a subject had N seizures, (N−1) seizures were used for training/validation (75%/25% split in
temporal order), and the remaining seizure was used for testing. This was done N times, so that
all seizures were used for testing once. Interictal segments were randomly split into N parts,
where (N − 1) parts were used for training/validation (75%/25% split in temporal order) and
the remaining part was used for testing. Fig. 1 C (top) illustrates the leave-one-out separation
using the first preictal (seizure) segment. The leave-one-out approach consequently required
patients with at least two seizures which reduced the analysis to 50 patients (Figure 4).
5
Pseudo-prospective approach
For seizure prediction under real-world conditions, the risk for seizures at a given time needs to
be assessed using algorithms trained and validated on past data. A pseudo-prospective evalua-
tion method that only evaluates an algorithm using out-of-sample data recorded at a later point
in time than the data used for algorithm training and validation matches these conditions most
closely. A pseudo-prospective evaluation setup benefits from large data sets with sufficient data
for training and pseudo-prospective validation and testing. For this purpose, we separated data
into non-overlapping, consecutive training, validation and test data, where validation and test
data contained preictal segments from one seizure each and training data the remaining, preced-
ing preictal and interictal periods (Fig. 1 C, bottom). The exact cut-off between validation and
test data was chosen so that both contained an equal amount of interictal data. This procedure
guaranteed that training and validation data always chronologically preceded test data and that
training, validation and test data were chronologically separated. This constraint on data, which
required patients to have had at least three seizures, reduced the analysis to 21 patients whose
data could be separated for such a pseudo-prospective approach.
Neural networks and training
We primarily used one-dimensional convolutional neural networks (CNNs), as they have been
shown to provide robust classification performance based on multi-dimensional timeseries data.14
To use the wristband sensor data in a one-dimensional CNN (1DConv), data was down-sampled
to 4 Hz for all sensors in order to provide the same vector length for each 30-second segment.
Figure 5 shows a summary of the CNN parameters used.
We also compared our results to a two-dimensional convolutional neural network (2DConv)
that used the power spectral density (PSD, FFT routine; 30-second, non-overlapping Hanning
windows) of signals sampled at 64 Hz (if originally sampled below 64 Hz, signals were upsam-
6
pled by repeating values). PSD data were then reshaped in 31 by 31 images of depth 6 for use
in neural networks. The use of 2DConv networks using PSD was motivated by encouraging
results of similar approaches using EEG and ECoG data.15, 16
We used balanced learning to handle imbalanced training sets, repeating either pre- or in-
terictal segments in the training set until a 50-50 preictal-interictal balance was achieved. All
networks were trained for 1000 epochs. To limit CNNs from overfitting, we kept the CNN
architecture simple and shallow.
Performance metrics and statistical tests
Performance was assessed on out-of-sample test data using the area under the receiver operating
curve (AUC) and improvement over chance based on F1 scores (IoC). AUC is commonly used
to assess performance in classification problems. AUC scores were obtained for ten network
executions (pseudo-prospective approach) or all folds (leave-one-out approach, five network
executions each). A one-sample t-test was used to compare AUC-values against chance (i.e.,
AUC=0.5).
The F1 score is the harmonic mean between precision and recall and, as a single metric,
evaluates how well a classification algorithm performs on the minority class. The use of F1
scores as an additional method was motivated by the observation that F1 scores may provide
more reliable performance estimates in skewed, imbalanced data sets,17, 18 such as in classifica-
tion problems in epilepsy where interictal data typically outnumber preictal data. F 1T est scores
were obtained for ten network executions (pseudo-prospective approach) or all folds (leave-one-
out approach, five network executions each). To compare our algorithms to a chance predictor,
the F1 score was obtained for random predictions (i.e., randomly shuffled test set lablels, which
maintains the numbers of pre- and interictal classifications but unlinks any correlation to the
data) averaged across 1000 randomizations (pseudo-prospective approach; 1000 randomization
7
per fold for the leave-one-out approach) resulting in a mean F 1Chance score. Improvement over
chance (IoC) was then defined as IoC = F 1T est−F 1Chance
metric in this article. A one-sample t-test was used to compare IoC-values against chance (i.e.,
and used as the main performance
1−F 1Chance
IoC=0). IoC has been repeatedly used in epilepsy forecasting research as a meaningful metric
to characterize algorithm performance.2, 7, 16
Results
We assessed the utility of CNNs in distinguishing pre- from interictal states using long-term,
multi-modal wristband sensor data12 obtained during epilepsy monitoring. Neural network
based assessments were done for each patient individually and for two approaches, a leave-
one-out cross-validation approach and a pseudo-prospective approach (Fig. 1).
Leave-one-out approach
The Leave-one-out approach required patient datasets to contain at least two seizures. With this
constraint, we analyzed a total of 50 patients with a total of 157 seizures and a recording time
of 1425.92 hours (1270.18 hours interictal, 155.74 hours preictal; Figure 4).
Each fold of the leave-one-out approach was executed five times, and average results with
standard deviations were reported for each patient. Area under the receiver operating curve
(AUC) values and improvement over chance (IoC) values using the 1DConv network are de-
picted along with seizure number and data duration for all patients in figure 2. Overall, a
prediction better than chance was obtained for more than half of the patients (Fig. 2 A, B, blue
markers; AUC significant for 30/50 patients, AU C = 0.714 ± 0.121; IoC significant for 29/50
patients, IoC = 0.377 ± 0.229). In comparison, the 2DConv network did not exhibit an overall
better performance (AUC significant for 26/50 patients, AU C = 0.595± 0.069; IoC significant
for 29/50 patients, IoC = 0.164 ± 0.131).
8
Pseudo-prospective approach
Although our data here is comparably short, we attempted a pseudo-prospective evaluation to
assess our approach under more realistic conditions. Requiring validation and test data to be
chronologically following training data with at least one seizure each reduced our analysis to 21
patients. Networks were executed ten times, and average results with standard deviations were
reported (Fig. 3 A, B, blue markers; AUC significant for 9/21 patients, AU C = 0.627 ± 0.090;
IoC significant for 8/21 patients, IoC = 0.221 ± 0.144). The 2DConv network again did not
exhibit an overall better performance (AUC significant for 9/21 patients, AU C = 0.572±0.062;
IoC significant for 13/21 patients, IoC = 0.094 ± 0.061).
Discussion
Here we assessed the utility of physiological sensor data recorded from a wristband to estimate
seizure risk. The ability to robustly differentiate preictal from an interictal states is a neces-
sary prerequisite for any seizure forecasting or seizure risk assessment.2 For this classification
process, we used convolutional neural networks which have demonstrated outstanding perfor-
mance in classification tasks in several domains,14 more recently also including epilepsy.15, 16
Our work is motivated by the potential benefits for patients and clinicians from a robust seizure
gauge. Forecasting seizures would provide patients with timely warning to adapt daily activ-
ities and allow clinicians to titrate therapies and develop novel interventions that potentially
could prevent impending seizures.1, 19 Peripheral sensor data that can be recorded easily and
non-invasively with a wristband would be desirable for such a purpose since approaches based
on ECoG7 or a large number of scalp EEG channels11 limit broad clinical application.
Of the patients included in our analysis, about half displayed a significant improvement over
chance (IoC) for both evaluation schemes, leave-one-out and pseudo-prospective approaches.
9
On the one hand, these performance values may not appear as strong as what is reported in
other recent studies where the majority of patients exhibited predictability levels better than
chance.15, 16 On the other hand, these results suggest that seizure forecasting might also be fea-
sible with relatively short, noisy, multi-modal signals recorded from wristbands far away from
the brain. The better-than-chance classification performance in about half of the patients was
obtained despite the comparably brief duration of data, where training sometimes only involved
one or two seizures, and the variability in seizure types, data duration, age and wristband lo-
cation. Apart from the criteria to label pre- and interictal data, we used no other preselection
or preprocessing, but instead included the data in raw format "as is" in an attempt to maximize
the transferability of our approach to real-world, noisy conditions. Predictability performance
across patients did not depend on wristband location, overall duration of data, or seizure type.
Seizure forecasting builds on the notion that a preictal state, during which a seizure is more
likely to occur soon, can be reliably distinguished from interictal states. To this end, most stud-
ies have focused on data recorded either from ECoG and EEG or from ECG. ECG has thus been
a long-standing example that peri- and preictal changes can not only be detected within the cen-
tral nervous system but are also reflected in a variety of cardiac effects.20 -- 22 Cardiac activity is
controlled by parasympathetic and sympathetic branches of the autonomic nervous system, with
the former producing an inhibitory response and the latter producing an excitatory response on
heart rate.23 Preictal changes in brain activity that occur in or propagate to autonomic control
centers may affect this autonomic balance and, consequently, affect cardiac activity during the
leadup to a seizure. A recent study that compared the information content in ECoG, EEG and
ECG in terms of identifying preictal states found that single-channel ECG contains a compara-
ble amount of information to multi-channel EEG,24 which highlights the relevance of peripheral
sensors for seizure forecasting. Autonomous nervous system changes are captured by the wrist-
band sensors used in this study in several ways. Electrodermal activity is known to be sensitive
10
to sympathetic innervation. Blood volume pressure curves contain information about heart rate
which is controlled by the parasympathetic and sympathetic interplay. Similarly, body temper-
ature is known to be maintained by the autonomic nervous system. The approach proposed in
this study builds on monitoring these autonomous nervous system functions along with actig-
raphy, which indirectly also monitors resting periods and sleep, and therefore pioneers seizure
forecasting capabilities based on such multi-modal sensor data, going beyond more traditional
ECoG/EEG and ECG approaches.
In our approach, we used the same model for all patients, albeit models were trained for each
patient individually. While it is possible that model hyperparameters individualized for each
patient might bring about better performances, we chose to have the same model architecture
across patients that could potentially be implemented "out-of-the-box" in future prospective
settings. Previous studies using CNNs for seizure forecasting have relied on power spectral
densities converted to two-dimensional images used in neural networks.15, 16 In contrast, our
approach relied on one-dimensional convolutional networks where we used raw sensor data
for inputs directly, thus requiring less preprocessing. Although we did not do an extensive
comparison in terms of model hyperparameters, we found that, at least for the two models
chosen here, the 1DConv networks using raw data performed slightly better than the 2DConv
network using the data's power spectrum.
Our evaluation used two different approaches, each of which has certain advantages and
disadvantages. Leave-one-out approaches have the advantage that all seizures can be used for
training and validation to obtain an estimate of algorithm performance. This can be desirable
particularly for relatively short data durations, such as ours. For this reason, leave-one-out cross-
validation approaches have been used to assess seizure prediction and detection performance in
epilepsy research.13, 15 We here applied a leave-one-out approach to preictal intervals belonging
to different seizures in order to assess the algorithmic ability to differentiate pre- from interictal
11
states. For a forecasting algorithm to be used in real-world conditions to assess future seizure
risk, classification of incoming data into pre- and interictal classes has to be performed in a
prospective manner using data recorded after the training and validation periods. A pseudo-
prospective assessment benefits particularly from large data sets with sufficient data for training
and pseudo-prospective validation and testing.16
Results need to be interpreted in the setting of data acquisition. One limitation of our study
is the relative short duration of recordings which covered only a few days of continuous data
per patient. Training data benefits from long periods of data where algorithms can better learn
to generalize and which gives a more realistic account of seizure forecasting capabilities. How-
ever, the current dataset is unique in the sense that it contains multi-modal sensor data over
several days from a relatively large number of epilepsy patients. The better-than-chance pre-
dictability in about half of the patients in this study is therefore encouraging for future, longer
trials using these sensors. Another limitation is the absence of benchmarks to compare our ap-
proach to. While we compared our results to chance predictors in out-of-sample test data, the
uniqueness and novelty of the current dataset limits more comprehensive comparison to other
approaches. There is growing awareness of the benefits of creating data warehouse ecosystems
that allow rigorous and continuous reevaluation and benchmarking by making data and algo-
rithms available to many researchers.25 We expect that these open-science efforts will increase
the reproducibility and help benchmark and improve algorithms, such as the ones proposed in
the current study, in the future. Finally, other clinically meaningful metrics, such as false alarm
rate, or time between seizure warning and actual seizure, that characterize an algorithm's fore-
casting performance may be desirable. These metrics, however, build on top of classifications
and require further post-processing. Due to the brevity of the data and the absence of other
relevant benchmarks to compare our approach to, we did not consider this useful in the current
context. Instead our main goal in this study was more basic, to assess whether the combination
12
of peripheral wristband sensor data and deep learning might be able to differentiate pre- from
interictal states with a better-than-chance performance. For this purpose, AUC and IoC, which
have both been used in several other forecasting studies,2, 7, 16, 25, 26 are valid metric which also
allow future comparison and benchmarking.
Seizure forecasting is likely to bring about notable benefits for epilepsy patients and clini-
cians. In order to make seizure forecasting available for broad use, non-invasive, easily appli-
cable techniques are greatly desirable. We here assessed the capability of multi-modal wrist-
band sensor data in combination with deep learning to reliably distinguish pre- from interictal
states. Our results demonstrate a better-than-random predictability in about half the patients,
even when a pseudo-prospective approach on out-of-sample test data was taken. Future, more
long-term studies should help to validate the utility of this approach.
References
1 S. B. Dumanis, J. A. French, C. Bernard, G. A. Worrell, and B. E. Fureman. Seizure Forecast-
ing from Idea to Reality. Outcomes of the My Seizure Gauge Epilepsy Innovation Institute
Workshop. eNeuro, 4(6), 2017.
2 F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehnertz. Seizure prediction: the long and
winding road. Brain, 130:314 -- 333, 2007.
3 D. R. Freestone, P. J. Karoly, and M. J. Cook. A forward-looking review of seizure prediction.
Curr. Opin. Neurol., 30(2):167 -- 173, 04 2017.
4 Melbourne
University
AES/MathWorks/NIH
Seizure
Prediction.
https://www.kaggle.com/c/melbourne-university-seizure-prediction. Accessed: 2016-10-28.
13
5 F. Mormann and R. G. Andrzejak. Seizure prediction: making mileage on the long and
winding road. Brain, 139(Pt 6):1625 -- 1627, Jun 2016.
6 L. Kuhlmann, K. Lehnertz, M. P. Richardson, B. Schelter, and H. P. Zaveri. Seizure prediction
- ready for a new era. Nat Rev Neurol, 14(10):618 -- 630, Oct 2018.
7 M. J. Cook, T. J. O'Brien, S. F. Berkovic, M. Murphy, A. Morokoff, G. Fabinyi, W. D'Souza,
R. Yerra, J. Archer, L. Litewka, S. Hosking, P. Lightfoot, V. Ruedebusch, W. D. Sheffield,
D. Snyder, K. Leyde, and D. Himes. Prediction of seizure likelihood with a long-term, im-
planted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study.
Lancet Neurol, 12(6):563 -- 571, Jun 2013.
8 K. Lehnertz, H. Dickten, S. Porz, C. Helmstaedter, and C. E. Elger. Predictability of uncon-
trollable multifocal seizures - towards new treatment options. Sci Rep, 6:24584, Apr 2016.
9 K. Fujiwara, M. Miyajima, T. Yamakawa, E. Abe, Y. Suzuki, Y. Sawada, M. Kano, T. Mae-
hara, K. Ohta, T. Sasai-Sakuma, T. Sasano, M. Matsuura, and E. Matsushima. Epileptic
Seizure Prediction Based on Multivariate Statistical Process Control of Heart Rate Variability
Features. IEEE Trans Biomed Eng, 63(6):1321 -- 1332, 06 2016.
10 D. H. Kerem and A. B. Geva. Forecasting epilepsy from the heart rate signal. Med Biol Eng
Comput, 43(2):230 -- 239, Mar 2005.
11 A. Schulze-Bonhage, F. Sales, K. Wagner, R. Teotonio, A. Carius, A. Schelle, and M. Ihle.
Views of patients with epilepsy on seizure prediction devices. Epilepsy Behav, 18(4):388 --
396, Aug 2010.
12 M. Z. Poh, T. Loddenkemper, N. C. Swenson, S. Goyal, J. R. Madsen, and R. W. Picard.
Continuous monitoring of electrodermal activity during epileptic seizures using a wearable
sensor. Conf Proc IEEE Eng Med Biol Soc, 2010:4415 -- 4418, 2010.
14
13 M. Z. Poh, T. Loddenkemper, C. Reinsberger, N. C. Swenson, S. Goyal, M. C. Sabtala, J. R.
Madsen, and R. W. Picard. Convulsive seizure detection using a wrist-worn electrodermal
activity and accelerometry biosensor. Epilepsia, 53(5):e93 -- 97, May 2012.
14 J. Schmidhuber. Deep learning in neural networks: an overview. Neural Netw, 61:85 -- 117,
Jan 2015.
15 N. D. Truong, A. D. Nguyen, L. Kuhlmann, M. R. Bonyadi, J. Yang, S. Ippolito, and O. Kave-
hei. Convolutional neural networks for seizure prediction using intracranial and scalp elec-
troencephalogram. Neural Netw, 105:104 -- 111, Sep 2018.
16 I. Kiral-Kornek, S. Roy, E. Nurse, B. Mashford, P. Karoly, T. Carroll, D. Payne, S. Saha,
S. Baldassano, T. O'Brien, D. Grayden, M. Cook, D. Freestone, and S. Harrer. Epilep-
tic Seizure Prediction Using Big Data and Deep Learning: Toward a Mobile System.
EBioMedicine, 27:103 -- 111, Jan 2018.
17 Nitesh V. Chawla. Data Mining for Imbalanced Datasets: An Overview, pages 853 -- 867.
Springer US, Boston, MA, 2005.
18 L. A. Jeni, J. F. Cohn, and F. De La Torre. Facing Imbalanced Data Recommendations for the
Use of Performance Metrics. Int Conf Affect Comput Intell Interact Workshops, 2013:245 --
251, 2013.
19 M. O. Baud and V. R. Rao. Gauging seizure risk. Neurology, 91(21):967 -- 973, Nov 2018.
20 J. M. VAN BUREN. Some autonomic concomitants of ictal automatism; a study of temporal
lobe attacks. Brain, 81(4):505 -- 528, Dec 1958.
21 D. W. Marshall, B. F. Westmoreland, and F. W. Sharbrough. Ictal tachycardia during temporal
lobe seizures. Mayo Clin. Proc., 58(7):443 -- 446, Jul 1983.
15
22 P. E. Smith, S. J. Howell, L. Owen, and L. D. Blumhardt. Profiles of instant heart rate during
partial seizures. Electroencephalogr Clin Neurophysiol, 72(3):207 -- 217, Mar 1989.
23 B. F. Robinson, S. E. Epstein, G. D. Beiser, and E. Braunwald. Control of heart rate by the
autonomic nervous system. Studies in man on the interrelation between baroreceptor mecha-
nisms and exercise. Circ. Res., 19(2):400 -- 411, Aug 1966.
24 C. Meisel and K. Bailey. Deep learning identifies signal-dependent information about the
preictal state: a comprehensive assessment across ECoG, EEG and EKG. submitted, 2019.
25 B. H. Brinkmann, J. Wagenaar, D. Abbot, P. Adkins, S. C. Bosshard, M. Chen, Q. M. Tieng,
J. He, F. J. Munoz-Almaraz, P. Botella-Rocamora, J. Pardo, F. Zamora-Martinez, M. Hills,
W. Wu, I. Korshunova, W. Cukierski, C. Vite, E. E. Patterson, B. Litt, and G. A. Worrell.
Crowdsourcing reproducible seizure forecasting in human and canine epilepsy. Brain, 139(Pt
6):1713 -- 1722, 06 2016.
26 L. Kuhlmann, P. Karoly, D. R. Freestone, B. H. Brinkmann, A. Temko, A. Barachant, F. Li,
G. Titericz, B. W. Lang, D. Lavery, K. Roman, D. Broadhead, S. Dobson, G. Jones, Q. Tang,
I. Ivanenko, O. Panichev, T. Proix, M. Nahlik, D. B. Grunberg, C. Reuben, G. Worrell, B. Litt,
D. T. J. Liley, D. B. Grayden, and M. J. Cook. Epilepsyecosystem.org: crowd-sourcing
reproducible seizure prediction with long-term human intracranial EEG. Brain, 141(9):2619 --
2630, Sep 2018.
Acknowledgements
This work was supported in part by the Epilepsy Research Fund (ERF). CM acknowledges sup-
port by a NARSAD Young Investigator Grant from the Brain & Behavior Research Foundation.
16
Ethical Publication Statement
We confirm that we have read the Journals position on issues involved in ethical publication and
affirm that this report is consistent with those guidelines.
Disclosure
None of the authors has any conflict of interest to disclose. TL received unrelated research
support and device donations from Empatica, Inc.
17
Figure 1: Schematic outline of data processing in order to predict pre- and interictal states. A,
continuous recording of timeseries data. B, data separation into preictal and interictal segments.
C, separation into training, validation and test data for each approach, leave-one-out (example
shown for first seizure) and pseudo-prospective. D, training, validation and out-of-sample test-
ing using convolutional neural networks.
18
AUTOCORRpredict seizures with theory + deep learningCRITICAL SLOWVARIANCEextract predictive EEG biomarkers no seizureseizure soonBVP blood volume pressureAUTOCORRpredict seizures with theory + deep learningCRITICAL SLOWVARIANCEextract predictive EEG biomarkers no seizureseizure soonDeep LearningTimeseries DataEDA electrodermal activityTEMP temperatureACC-X accelerometer in x-dimensionACC-Y accelerometer in y-dimensionACC-Z accelerometer in z-dimensionValidateTrainAUTOCORRpredict seizures with theory + deep learningCRITICAL SLOWVARIANCEextract predictive EEG biomarkers no seizureseizure soonLeave-One-Out ApproachTime [Hours]240PreictalInterictalData Separation In Pre-/InterictalTime [Hours]240TrainVal.PreictalInterictalPseudo-Prospective ApproachTestTime [Hours]240TestAUTOCORRpredict seizures with theory + deep learningCRITICAL SLOWVARIANCEextract predictive EEG biomarkers no seizureseizure soonAUTOCORRpredict seizures with theory + deep learningCRITICAL SLOWVARIANCEextract predictive EEG biomarkers no seizureseizure soonDeep LearningAUTOCORRpredict seizures with theory + deep learningCRITICAL SLOWVARIANCEextract predictive EEG biomarkers no seizureseizure soonABCD Figure 2:
Prediction performance for the leave-one-out approach using a one-dimensional
convolutional network (1DConv, n=50 patients). A, area under the receiver operating curve
(AUC; blue markers indicate a performance significantly better than chance). B, improvement
over chance (IoC). C, seizure count. D, total duration of data analyzed.
19
IoC (Leave-One-Out Approach)Patient IDChanceMaxNumber Of Seizures AnalysedHours AnalysedBCDAAUC (Leave-One-Out Approach)p≤0.01p≤0.05* Figure 3: Prediction performance for the pseudo-prospective approach using a one-dimensional
convolutional network (1DConv, n=21 patients). A, area under the receiver operating curve
(AUC; blue markers indicate a performance significantly better than chance). B, improvement
over chance (IoC). C, seizure count. D, total duration of data analyzed.
20
IoC (Pseudo-Prospective Approach)Patient IDChanceMaxNumber Of Seizures AnalysedHours AnalysedBCDp≤0.01p≤0.05*AAUC (Pseudo-Prospective Approach) Figure 4: Summary of patient characteristics.
21
Patient dataPatient ID101413130.812138151311542291581313121522103388 390 392 399 410 411 417 418 419 421423 425 427 428 432 440 443 444 446 466 467 475 482 493 496 506Age [Years]LocationWristband1483211653671090.669270.4421479111911117No. of SeizuresPreictal[Minutes]Interictal[Minutes]180.0 180.0 120.0 240.0 120.0 120.0 275.5 180.0 180.0 180.0 120.0 120.0 120.0 240.0 120.0 300.0 181.5 64.5 120.0 240.0 180.0 180.0 120.0 240.0 180.0 480.0597.5 914.5 2243.5 2356.0 204.0 2228.5 2714.0 545.5 1983.5 1753.5 2270.5 807.0 1370.0 1427.5 185.5 1113.5 1764.5 2255.5 4536.5 2986.5 1890.0 2116.5 3806.5 3215.0 187.0 832.0right wristleft wrist right ankle left wrist right ankle right wrist left ankle right ankle right wrist left wrist left wrist left wrist left ankle left ankle right ankle left ankle right ankle right wrist left wrist right wrist right wrist left ankle left ankle left ankle right wrist right wristPatient IDAge [Years]LocationWristbandNo. of SeizuresPreictal[Minutes]Interictal[Minutes]192 198 210 212 221 224 236 243 263 284 296 302 303 309 329 330353 356 365 370 372 378 380 3873 3 2 4 2 2 5 3 3 3 2 2 2 4 2 5 4 2 2 4 3 3 2 4 3 8103 2 3 4 2 2 3 2 2 6 4 5 2 2 2 2 4 3 2 2 2 2 3657.0138.5 120.0 180.0 214.5 120.0 120.0 180.0 120.0 120.0 360.0 240.0 300.0 119.5 120.0 120.0 120.0 193.0 180.0 120.0 120.0 120.0 120.0 180.0 3398.5 1136.0 864.5 395.0 1160.5 2645.5 98.0 570.5 391.0 753.0 1134.5 1478.0 2541.0 617.0231.5 16.5 1660.5 526.5 1616.5 702.5 474.0 22.5 3009.0 4464.0 right wrist left ankle right ankle right ankle left wristleft wristleft ankle left wrist left wrist left ankle right ankle right ankle left ankle right wrist right ankle left ankle left ankle left ankle right ankle right wrist right ankle left ankle right wrist right wrist Figure 5: One-dimensional convolutional (1DConv) network topology used. Learning rate was
set to 0.0001.
22
Network architectureLayer numberLayer typeFilters / Kernel size or NodesLayer parametersConv1DDropoutConv1DMaxPool1DConv1DDropoutConv1DAvgPool1DDropoutDense1234567891050/10NA50/10380/102080/10NANA1activation = ReLUdropout rate = 0.7activation = ReLUNAactivation = ReLUdropout rate = 0.7activation = ReLUNAdropout rate = 0.7activation = sigmoid |
1805.03601 | 3 | 1805 | 2018-11-30T19:57:40 | Finite size effects for spiking neural networks with spatially dependent coupling | [
"q-bio.NC"
] | We study finite-size fluctuations in a network of spiking deterministic neurons coupled with non-uniform synaptic coupling. We generalize a previously developed theory of finite size effects for uniform globally coupled neurons. In the uniform case, mean field theory is well defined by averaging over the network as the number of neurons in the network goes to infinity. However, for nonuniform coupling it is no longer possible to average over the entire network if we are interested in fluctuations at a particular location within the network. We show that if the coupling function approaches a continuous function in the infinite system size limit then an average over a local neighborhood can be defined such that mean field theory is well defined for a spatially dependent field. We then derive a perturbation expansion in the inverse system size around the mean field limit for the covariance of the input to a neuron (synaptic drive) and firing rate fluctuations due to dynamical deterministic finite-size effects. | q-bio.NC | q-bio | Finite size effects for spiking neural networks with spatially
dependent coupling
Si-Wei Qiu1 and Carson C. Chow1
1Mathematical Biology Section, LBM, NIDDK, NIH
(Dated: December 4, 2018)
8
1
0
2
v
o
N
0
3
]
.
C
N
o
i
b
-
q
[
3
v
1
0
6
3
0
.
5
0
8
1
:
v
i
X
r
a
1
Abstract
We study finite-size fluctuations in a network of spiking deterministic neurons coupled with non-
uniform synaptic coupling. We generalize a previously developed theory of finite size effects for
globally coupled neurons with a uniform coupling function. In the uniform coupling case, mean
field theory is well defined by averaging over the network as the number of neurons in the network
goes to infinity. However, for nonuniform coupling it is no longer possible to average over the entire
network if we are interested in fluctuations at a particular location within the network. We show
that if the coupling function approaches a continuous function in the infinite system size limit then
an average over a local neighborhood can be defined such that mean field theory is well defined
for a spatially dependent field. We then use a path integral formalism to derive a perturbation
expansion in the inverse system size around the mean field limit for the covariance of the input
to a neuron (synaptic drive) and firing rate fluctuations due to dynamical deterministic finite-size
effects.
I.
INTRODUCTION
The dynamics of neural networks have traditionally been studied in the limit of very large
numbers of neurons, where mean field theory can be applied, e.g. [1, 2, 4, 11, 16, 19, 20, 25,
26, 30] or for a small number of neurons, where traditional dynamical systems approaches
can be used, e.g. [15, 22, 24]. The intermediate regime of large but finite numbers of
neurons can have interesting properties that are independent of the small and infinite system
limits [6, 10, 12, 14, 17, 18, 23, 29]. However, these previous works have not fully explored
fluctuations due to finite-size effects at specific locations within the network when all the
neurons receive nonhomogeneous input from other neurons because of nonuniform coupling.
Here, we consider finite-size effects in a network of spiking neurons with nonuniform synaptic
coupling. Previously [6, 10, 18], a perturbation expansion in the inverse network neuron
number had been developed for networks with global spatially uniform coupling and we
generalize that theory to include nonuniform coupling. We first show that mean field theory
in the infinite nonuniform system limit can be realized in a single network if a spatial metric
can be imposed on the network and the coupling function is a continuous function of this
distance measure. We then analyze finite-size fluctuations around such mean field solutions
using a path integral formalism to derive a perturbation expansion in the inverse network
2
neuron number for the spatially dependent covariance function for the synaptic drive and
spatially dependent neuron firing rate.
II. COUPLED NEURON MODEL
Consider a network of N theta neurons (phase reduction of quadratic integrate-and-fire
neurons [15]) on a one dimensional periodic domain of size L although the theory can be
applied to any domain. The network obeys the following deterministic microscopic equations:
θi = 1 − cos θi + (Ii + ui(t))(1 + cos θi)
N(cid:88)
j=1
ui =
L
N
wijsj
(cid:88)
sj = −βsj + β
δ(t − tl
j)
(2.1)
(2.2)
(2.3)
l
where θi is the phase of neuron i, ui is the synaptic drive to neuron i, Ii is the external input
to neuron i, β is the decay constant of the synaptic drive, sj is the time dependent synaptic
input from neuron j, and tl
j represents the spike times when the phase of neuron j crosses
π. sj rises instantaneously when neuron j spikes and relaxes to zero with a time constant
of 1/β. The synaptic drive represents the total time dependent synaptic input where the
contribution from each neuron is weighted by the synaptic coupling function wij (a real
N × N matrix). When Ii + ui > 0, the neuron receives suprathreshold input and θi will
progress in time. When it passes π, the neuron is said to spike. When Ii + ui < 0 the neuron
receives subthreshold input and the phase will approach a fixed point. The theta neuron
is the normal form of a Type I spiking neuron near the bifurcation point to firing [15]. By
linearity, the synaptic drive obeys the more convenient form of
ui = −βui + β
L
N
We define an empirical density
N(cid:88)
(cid:88)
δ(t − tl
j)
wij
j=1
l
ηj(θ, t) = δ(θ − θj(t))
(2.4)
(2.5)
the sum of a spike train as(cid:80)
that assigns a point mass to the phase of each neuron in the network. Hence, we can write
θjθj =π = 2
j) = ηj(π, t) θjθj =π. For the theta model,
l δ(t − tl
3
and thus we can then rewrite (2.4) as
ui = −βui + 2β
N(cid:88)
j=1
L
N
wijηj(π, t)
(2.6)
Neuron number is conserved so the neuron density formally obeys a conservation (Klimon-
tovich) equation [10]:
∂tηi(θ, t) + ∂θFi(θ, ui)ηi(θ, t) = 0
(2.7)
where Fi(θ, ui) = 1 − cos θ + (1 + cos θ)(Ii + ui). The Klimontovich equation together with
(2.6) fully describes the system. However, it is only a formal definition since η is not in
general differentiable. In the following, we develop a method to regularize the Klimontovich
equation so that desired quantities can be calculated.
III. MEAN FIELD THEORY
The Klimontovich equation (2.7) only exists in a weak sense. We can regularize it by
taking a suitable average over an ensemble of initial conditions:
∂t(cid:104)ηi(θ, t)(cid:105) + ∂θ(cid:104)Fi(θ, ui)ηi(θ, t)(cid:105) = 0
(3.1)
This equation is not closed because it involves covariances such as (cid:104)ηη(cid:105), which in turn depend
on higher order cumulants in a BBGKY hierarchy [6, 10, 18]. This hierarchy can be rendered
tractable if we can truncate it. Mean field theory truncates the hierarchy at first order by
assuming that all cumulants beyond the first are zero so we can write
∂tρi(θ, t) + ∂θFi(θ, ai)ρi(θ, t) = 0
(3.2)
where ai = (cid:104)ui(cid:105) and ρi = (cid:104)ηi(cid:105). The full set of closed mean field equations are given by
∂tρi(θ, t) + ∂θ [1 − cos θ + (Ii + ai)(1 + cos θ)] ρi(θ, t) = 0
ai = −βai + 2β
L
N
wijρj(π, t)
(3.3)
(cid:88)
j
Although we can always write down the mean field equations (3.3), it is not clear that a
given network would obey it in the infinite N limit. In previous work [9, 10, 25], it was shown
that mean field theory applies to a network of coupled oscillators with uniform coupling in
4
the infinite N limit. However, it is not known when or if mean field theory applies for
nonuniform coupling.
To see this, consider first the stationary system
∂θ [1 − cos θ + (Ii + ui)(1 + cos θ)] ηi(θ) = 0
(cid:88)
j
ui = 2
L
N
wijηj(π)
(3.4)
(3.5)
with uniform coupling, wij = w, and uniform external input, Ii = I. If the neurons are
initialized with random phases and remain asynchronous then we can suppose that in the
limit of N → ∞ the quantity
(cid:88)
j
ρ(π) =
L
N
ηj(π)
(3.6)
converges to an invariant quantity [9, 10, 25]. This then implies that ui = 2wρ ≡ a is also a
constant. Thus each neuron will have identical inputs so if we apply the network averaging
(cid:80)N
operator L
N
i=1 to (3.4) we obtain
∂θ [1 − cos θ + (I + a)(1 + cos θ)] ρ(θ) = 0
a = 2wρ(π)
(3.7)
(3.8)
Covariances vanish and mean field theory is realized in the infinite network limit. Given
that the drive equation (2.6) is linear, the time dependent mean field theory will similarly
hold in the large N limit.
In the case where wij is not uniform, covariances are not guaranteed to vanish and an
infinite network need not obey mean field theory. Our goal is to find conditions such that
mean field theory applies. Again, consider the stationary equations (3.4) and (3.5). Now,
instead of averaging over the entire domain, take a local interval around j, [j − cN/2, j +
cN/2], where c < 1 is a constant that can depend on N and we map j − cN/2 < 1 to
N + j − cN/2 and j + cN/2 > N to j + cN/2 − N . We want to express our mean field
equation in terms of the locally averaged empirical density
2(cid:88)
j+ cN
k=j− cN
2
ρj =
1
cN
ηk
(3.9)
If cN → ∞ for N → ∞ then it is feasible that the local empirical density can be invariant (to
random initial conditions) and correlations can vanish; we seek conditions on the coupling
5
for which this is true. Inserting (3.5) into (3.4), and taking the local average yields
(cid:34)
2(cid:88)
k+ cN
i=k− cN
2
1
cN
1 − cos θ +
∂θ
(cid:32)
(cid:33)
L
N
Ii + 2
N(cid:88)
j=1(cN )−1(cid:80)j+ cN
j=1 wijηj(π) = (cid:80)N
wijηj(π)
j=1
2
Consider the identity (cid:80)N
(cid:35)
We immediately see that correlations can arise from the sums over the product of ηj(π)ηi(θ).
(1 + cos θ)
ηi(θ) = 0
(3.10)
periodic boundary conditions. For nonperiodic boundary conditions there will be an edge
contribution but this should be negligible in the large network limit. Using this summation
l=j− cN
2
wilηl(π), which is exact for
identity, we can rewrite the sum as
N(cid:88)
1
cN
2(cid:88)
k+ cN
i=k− cN
2
wijηj(π)ηi(θ) =
Rjk =
1
(cN )2
+
1
(cN )2
j+ cN
2(cid:88)
2(cid:88)
2
j+ cN
l=j− cN
k+ cN
2(cid:88)
2(cid:88)
i=k− cN
2
k+ cN
l=j− cN
2
i=k− cN
2
N(cid:88)
j=1
(wkjρj(π)ρk(θ) + Rjk)
(3.11)
(wij − wkj)ρj(π)ηi(θ)
(wil − wij)ηl(π)ηi(θ)
(3.12)
j=1
where the remainder
carries the correlations. Mean field theory is valid in the N → ∞ limit if Rjk vanishes. Its
magnitude obeys
(wij − wkj)ρj(π)ηi(θ)
(wil − wij)ηl(π)ηi(θ)
(cid:12)(cid:12)(cid:12)(cid:12) +
(cid:12)(cid:12)(cid:12)(cid:12)
1
(cN )2
2(cid:88)
j+ cN
2(cid:88)
k+ cN
l=j− cN
2
i=k− cN
2
(cid:12)(cid:12)(cid:12)(cid:12)
Rjk ≤
(cN )2
(cid:12)(cid:12)(cid:12)(cid:12)
+
i=k− cN
2
k+ cN
1
j+ cN
l=j− cN
2(cid:88)
(cid:18) 1
(cid:18) 1
2(cid:88)
2(cid:88)
2(cid:88)
2
j+ cN
2
k+ cN
i=k− cN
cN
(cid:19)
2(cid:88)
k+ cN
ηi(θ)
(cN )2
l=j− cN
2
i=k− cN
2
≤ρj(π)
since the density is nonnegative.
Applying (3.9) then leads to
(cid:32)
i∈(k− cN
sup
2 ,k+ cN
2 )
(cid:19)
(ηl(π)ηi(θ))
wij − wkj
i∈(k− cN
2 ,k+ cN
sup
2 ),l∈(j− cN
2 ,j+ cN
2 )
wil − wij
(3.13)
(cid:33)
wil − wij
(3.14)
R ≤ρj(π)ρk(θ)
i∈(k− cN
sup
2 ,k+ cN
2 )
i∈(k− cN
2 ,k+ cN
sup
2 ),l∈(j− cN
2 ,j+ cN
2 )
wij − wkj +
6
We introduce a distance measure z = iL/N , z(cid:48) = jL/N ,z(cid:48)(cid:48) = kL/N , z(cid:48)(cid:48)(cid:48) = lL/N and write
ρi(θ)i=zN/L = ρ(z, θ) and wiji=zN/L,j=z(cid:48)N/L = w(z, z(cid:48)). Then
(cid:32)
R ≤ ρ(z(cid:48), π)ρ(z(cid:48)(cid:48), θ)
sup
z∈[z(cid:48)(cid:48)−cL/2,z(cid:48)(cid:48)+cL/2]
w(z, z(cid:48)) − w(z(cid:48)(cid:48), z(cid:48))
+
z∈[z(cid:48)(cid:48)−cL/2,z(cid:48)(cid:48)+cL/2],z(cid:48)(cid:48)(cid:48)∈[z(cid:48)−cL/2,z(cid:48)+cL/2]
sup
w(z, z(cid:48)(cid:48)(cid:48)) − w(z, z(cid:48))
(cid:33)
(3.15)
(3.16)
Hence, if we set c = N−α, 0 < α < 1, then as N → ∞, the number of neurons in the local
neighborhood cN approaches infinity as N 1−α while c → 0. Then, R → 0 as N → ∞ if
limz→z(cid:48)(cid:48) w(z, z(cid:48)) − w(z(cid:48)(cid:48), z(cid:48)) = 0 and limz(cid:48)(cid:48)(cid:48)→z(cid:48) w(z, z(cid:48)(cid:48)(cid:48)) − w(z, z(cid:48)) = 0, i.e. wij is a continuous
function in both indices. A similar argument shows that
2(cid:88)
k+ cN
i=k− cN
2
1
cN
(Ii − Ik)ηi(θ) → 0
(3.17)
if Ii approaches a continuous function in index i in the infinite N limit. Then (3.4) and (3.5)
can be written as
∂θ [1 − cos θ + (Ik + ak) (1 + cos θ)] ρk(θ) = 0
(cid:88)
j
ak = 2
L
N
wkjρj(π)
(3.18)
(3.19)
Equations (3.18) and (3.19) form a mean field theory that is realized in a nonuniform coupled
network in the infinite size limit as long as the input and coupling function are continuous
functions. By linearity, the time dependent mean field theory should equally apply if the
external input and the coupling are continuous functions of the indices.
continuous, (cid:80)
In the N → ∞ limit, setting i → zN/L, ai(t) → a(z, t), ρi(θ, t) → ρ(z, θ, t), Ii → I(z) is
Ω dz, and wij → w(z, z(cid:48)) is continuous, we can write mean field
i → (N/L)(cid:82)
theory in continuum form as
∂tρ(z, θ, t) + ∂θ [1 − cos θ + (I(z) + a(z, t))(1 + cos θ)] ρ(z, θ, t) = 0
w(z, z(cid:48))ρ(z(cid:48), π, t)dz(cid:48)
∂ta(z, t) = −βa(z, t) + 2β
(cid:90)
The stationary solutions obey
∂θ [1 − cos θ + (I(z) + a(z))(1 + cos θ)] ρ(z, θ) = 0
w(z, z(cid:48))ρ(z(cid:48), π)dz(cid:48)
a(z) = 2
(cid:90)
7
(3.20)
(3.21)
(3.22)
FIG. 1. a) Mean field theory synaptic drive for b) connectivity weight w(z) = −J0 +J2 cos(2π/Lz),
J0 = 0.2 and J2 = 0.8, and c) external input I(z) = I0 + sin(2π/L(z − z0)), I0 = 1, z0 = 0.25.
The stationary solutions will be qualitatively different depending on the sign of I + a.
Consider first the suprathreshold regime where I + a > 0. We can then solve (3.21) to
obtain
(cid:112)I(z) + a(z)
w(z, z(cid:48))(cid:112)I(z) + a(z)dz(cid:48)
(cid:90)
a(z) =
1
π
which has been normalized such that(cid:82) ρ(z, θ)dθ = 1. Inserting this back into (3.22) gives
π [1 − cos θ + (I(z) + a(z))(1 + cos θ)]
ρ(z, θ) =
(3.23)
(3.24)
In the subthreshold regime, I + a < 0, (3.23) has a singularity at 1 − cos θ + (I(z) +
a(z))(1 + cos θ) = 0, for which there are two solutions θ± that coalesce in a saddle node
bifurcation at I + a = 0. Although ρ is no longer differentiable at equilibrium in the
subthreshold regime there is still a weak solution. It has been shown previously [15] that
θ− is stable and θ+ is unstable for a single theta neuron. This implies that the density
is given by ρ(z, θ) = δ(θ − θ−) and that ρ(z, π) = 0 (i.e. no firing) as expected in the
subthreshold regime. Figure 1, shows an example of a stationary "bump" solution for the
periodic coupling function, w(z) = −J0 + J2 cos(2π/Lz), which has been used in models of
orientation tuning of visual cortex [3] and the rodent head direction system [32]).
IV. BEYOND MEAN FIELD THEORY
In the infinite N limit when mean field theory applies, the fields η and u are completely
described by their means. The time trajectories of these fields are independent of the
initial conditions of the individual neurons. For finite N , the trajectories can differ for
8
0.00.20.40.60.81.0z0.20.40.60.81.01.21.4ua)20015010050050100150200i-j6420246810wijb)0.00.20.40.60.81.0z0.40.60.81.01.21.41.6Ic) different initial conditions and going beyond mean field theory involves understanding these
fluctuations. Implicit in going beyond mean field theory is that these fields are themselves
random variables that are drawn from a distribution functional. In this section, we will derive
this distribution functional formally and then use it to compute perturbative expressions for
the covariances of η and u.
Recall, that the microscopic system is fully described by
∂tηi(θ, t) + ∂θFi(θ, ui)ηi(θ, t) − δ(t − t0)η0
i (θ) = 0
ui(t) + βui(t) − 2β
L
N
wijηj(π, t) − δ(t − t0)u0
i = 0
N(cid:88)
j=1
(4.1)
(4.2)
where we have expressed the initial conditions as forcing terms. The probability density
functional for the fields is then comprised of point masses constrained to the dynamical
system marginalized over the distribution of the initial data densities:
(cid:90)
Dη0 (cid:89)
i
P[η, u] =
(cid:34)
×δ
ui + βui − 2β
L
N
j
δ(cid:2)∂tηi + ∂θFi(θ, ui)ηi − δ(t − t0)η0
i (θ)(cid:3)
(cid:35)
(cid:88)
wijηj(π, t) − δ(t − t0)u0
P[η0]
i
where P [η0] is the probability density functional of the initial neuron densities for all neurons
and Dη0 is the functional integration measure. We consider the initial condition of u to be
fixed to u0. Using the functional Fourier transform for the Dirac delta functionals, we then
obtain
P[η, u] =
DηDη0Du e
(cid:90)
(cid:82) dtdθ ηi[∂tηi+∂θFi(θ,ui)ηi−δ(t−t0)η0
−(cid:80)
(cid:82) dt ui[ ui+βui−2β L
−(cid:80)
× e
(cid:80)
i
i (θ)]
j wij ηj (π,t)−δ(t−t0)u0
i ]P[η0]
i
(4.4)
where ηi and ui are response fields for neuron i with functional integration measures Dη and
i (θ) = δ(θ − θi(t = 0)), the distribution over initial densities
Du over all neurons. If we set η0
is given by the distribution over the initial phase, ρ0
N
i (θ). Thus we can write(cid:82) Dη0P [η0] =
(cid:80)
i dθρ0
i (θ). The initial condition contribution is given by the integral
(cid:82)(cid:81)
(4.3)
(4.5)
(4.6)
(4.7)
eW0[η] =
=
i (θi)e
i ηi(θi,t0)
i (θ)e ηi(θ,t0)
= e
i (θ)(e ηi(θ,t0)−1))
dθiρ0
(cid:90) (cid:89)
(cid:90)
(cid:89)
i ln(1−(cid:82) dθρ0
(cid:80)
dθρ0
i
i
9
Hence, the system given by (2.6) and (2.7) can be mapped to the distribution functional
P[η, u] =(cid:82) DηDη0Due−S with action S = Sη + Su given by
(cid:90) π
−π
(cid:90) t1
(cid:90) t1
t0
(cid:88)
(cid:88)
i
i
t0
Sη =
Su =
(cid:32)
ui + βui − 2β
(cid:88)
j
L
N
dt ui(t)
(cid:88)
i
1 −
(cid:18)
(cid:33)
(cid:90)
(cid:19)
i (θ)(e ηi(θ,t0) − 1)
dθρ0
(4.8)
(4.9)
wijηj(π, t) − δ(t − t0)u0
i
dt
dθ ηi(θ, t)[∂tηi(θ, t) + ∂θF (θ, ui)ηi(θ, t)] +
ln
The exponential in the initial data contribution to the action (which corresponds to a
generating function for a Poisson distribution) can be bilinearized via the Doi-Peliti-Janssen
transformation [5 -- 8, 10, 18]: ψi = ηi exp(−ηi), ψi = exp( ηi) − 1, resulting in
dθdt ψi(θ, t)[∂tψi(θ, t) + ∂θF (θ, ui)ψi(θ, t)] +
(cid:32)
ui(t) + βui − δ(t − t0)u0
i − 2β
L
N
dtui(t)
(cid:18)
(cid:90)
1 −
ln
dθρ0
i (θ) ψi(θ, t0)
(cid:19)
(cid:33)
(4.10)
(4.11)
wij( ψj(π, t) + 1)ψj(π, t)
(cid:88)
(cid:88)
i
j
where we have not included the noncontributing terms that arise after integration by parts.
We now make the coarse graining transformation i → zN/L, ui(t) → u(z, t), ψi(θ, t) →
Ω dz, and wij → w(z − z(cid:48)), which
ψ(z, θ, t), ρi(θ, t) → ρ(z, θ, t), Ii → I(z), (cid:80)
i → (N/L)(cid:82)
(cid:90)
(cid:90)
(cid:88)
(cid:88)
i
i
Sψ =
Su =
yields
Sψ =
Su =
(cid:90)
(cid:90)
N
L
N
L
dzdθdt ψ(z, θ, t)[∂tψ(z, θ, t) + ∂θF (θ, u)ψ(z, θ, t)]
(cid:90)
+
N
L
dz ln
(cid:18)
(cid:90)
(cid:18)
1 −
dθρ0(z, θ) ψ(z, θ, t0)
(cid:19)
(cid:90)
(4.12)
(cid:19)
dzdt u(z, t)
u + βu − δ(t − t0)u0(z) − 2β
dz(cid:48)w(z − z(cid:48))( ψ(z(cid:48), π, t) + 1)ψ(z(cid:48), π, t)
We examine perturbations around the mean field solutions a(z, t) and ρ(z, θ, t) of (3.20)
with u → a(z, t)H(t − t0) + v(z, t), u → v, ψ → ρ(z, θ, t)H(t − t0) + ϕ(z, θ, t), and ψ → ϕ,
10
(4.13)
where ρ(z, θ, t = t0) = ρ0(z, θ) and H(t − t0) is the Heaviside function. We then obtain
dθdtdz ϕ[∂tϕ + ∂θ [1 − cos θ + (I + (a + v))(1 + cos θ)] ϕ + ∂θv(1 + cos θ)ρ
dθρ0(z, θ) ϕ(z, θ, t0)
dz ln
dθdtdz ϕ[∂tϕ + ∂θ [1 − cos θ + (I + (a + v))(1 + cos θ)] ϕ
dz
+
dθ(cid:48) ϕ(z, θ(cid:48), t0)ρ0(z, θ)
(cid:19)
(cid:90)
N
L
(cid:90)
(cid:90)
dz
dθ(cid:48) ϕ(z, θ(cid:48), t0)ρ0(z, θ)
dθ ϕ(z, θ, t0)ρ0(z, θ)
(4.14)
Sϕ =
+
(cid:90)
(cid:18)
1 −
=
+ ∂θv(1 + cos θ)ρ] − N
2L
(cid:90)
(cid:90)
(cid:90)
(cid:90)
N
L
N
L
N
L
(cid:90)
(cid:20)
(cid:90)
(
d
dt
−2β
(cid:90)
(cid:90)
Ω
(cid:90)
Sϕ =
N
L
(cid:90)
Sv =
N
L
Sv =
N
L
dtdz v
+ β)v − 2β
dz(cid:48) w(z − z(cid:48))( ϕ(z(cid:48), π, t) + 1)ϕ(z(cid:48), π, t)
dz(cid:48) w(z − z(cid:48)) ϕ(z(cid:48), π, t)ρ(z(cid:48), π) − δ(t − t0)(u0(z) − a(z, t0))
(4.15)
Ω
We have only included the quadratic term of the initial condition since it is the only one
that plays a role at first order perturbation theory (tree level). Finally, if we set the mean
(cid:21)
field solutions to the stationary solutions ρ(z, θ) and a(z) we obtain
dθdtdz ϕ[∂tϕ + ∂θ [1 − cos θ + (I + (a + v))(1 + cos θ)] ϕ
+ ∂θv(1 + cos θ)ρ] +
dz
dθ(cid:48) ϕ(z, θ(cid:48), t0)ρ0(z, θ)
dθ ϕ(z, θ, t0)ρ0(z, θ)
(4.16)
(cid:90)
(cid:90)
N
2L
(cid:20)
(cid:90)
(
d
dt
−2β
(cid:90)
Ω
(cid:90)
(cid:21)
dtdz v
+ β)v − 2β
dz(cid:48) w(z − z(cid:48))( ϕ(z(cid:48), π, t) + 1)ϕ(z(cid:48), π, t)
dz(cid:48) w(z − z(cid:48)) ϕ(z(cid:48), π, t)ρ(z(cid:48), π)
(4.17)
Ω
Without loss of generality, we set L = 1. In the limit of N → ∞, the dominant term in the
probability density functional for the fields will be the extrema of the action, which defines
mean field theory. Moments of the fields can be computed perturbatively as an expansion in
1/N by using Laplace's method around mean field (i.e. a loop expansion). The bilinear terms
in the action (comprising of a product of a field and a response field) are the linear response
functions or propagators. All the other terms are vertices. Each vertex contributes a factor
of N while each propagator contributes 1/N . To make the scaling more transparent, we
make the rescaling transformation where v → v/N and ϕ → ϕ/N . This change will rescale
the propagators to order unity and the vertices to order one or higher depending on how
11
FIG. 2. a) Propagators, ∆ν
ν(z, t; z(cid:48), t(cid:48)) (upper left), ∆ϕ
ϕ(z, θ, t; z(cid:48), t(cid:48))
ϕ(z, θ, t; z(cid:48), θ(cid:48), t(cid:48)) (lower right) b) Vertices for action in (4.18). From left to
ν (z, t; z(cid:48), θ(cid:48), t(cid:48)) (lower left), ∆ν
(upper right), and ∆ϕ
right, they are ∂θ(1 + cos θ), 0, ρ0(z, θ)ρ0(z, θ), − 2β
N w(z − z(cid:48))ρ(z(cid:48), π), − 2β
N w(z − z(cid:48))
many response fields they possess. The resulting action is
(cid:90)
(cid:90)
+ ∂θv(1 + cos θ)ρ] +
dz
(cid:20)(cid:18) d
(cid:90)
1
2N
(cid:19)
v − 2β
(cid:90)
(cid:90)
Ω
Sv =
dtdz v
+ β
dt
(cid:90)
Ω
−2β
N
dz(cid:48) w(z − z(cid:48)) ϕ(z(cid:48), π, t)ρ(z(cid:48), π)
(cid:21)
Sϕ =
dθdtdz ϕ[∂tϕ + ∂θ [1 − cos θ + (I(z) + (a(z) + v))(1 + cos θ)] ϕ
(cid:90)
dθ(cid:48) ϕ(z, θ(cid:48), t0)ρ0(z, θ)
dθ ϕ(z, θ, t0)ρ0(z, θ)
dz(cid:48) w(z − z(cid:48))( ϕ(z(cid:48), π, t)/N + 1)ϕ(z(cid:48), π, t)
(4.18)
The propagators and vertices can be represented by Feynman graphs or diagrams (see Fig. 2).
Each response field corresponds to an outgoing branch (branch on the left) and each field
corresponds to an incoming branch (branch on the right). Time flows from right to left
and causality is respected by the propagators. To each branch is attached a corresponding
propagator.
12
a)b) ∆v
The propagators are defined by
v (x; y(cid:48))
ϕ(y; y(cid:48))
v(x; x(cid:48)) ∆ϕ
ϕ(y; x(cid:48)) ∆ϕ
∆v
G−1 ≡
−1
=
δ2S
δ2S
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)v,ϕ,v, ϕ=0
δv(x)δv(x(cid:48))
δ2S
δv(x)δϕ(y(cid:48))
δ2S
δ ϕ(y)δϕ(y(cid:48))
δ ϕ(y)δv(x(cid:48))
−2βw(z − z(cid:48))δ(π − θ(cid:48))δ(t − t(cid:48))
(d/dt + β)δ(x − x(cid:48))
=
∂θ(1 + cos θ)ρ(z, θ)δ(x − x(cid:48)) (∂t + ∂θ[1 − cos θ + (I + a)(1 + cos θ)])δ(y − y(cid:48))
(4.20)
a(x; x(cid:48)) is the response of field a at the
nonprimed location to field b at the primed location. The propagator satisfies the condition
where x = (z, t), and y = (z, θ, t). The propagator ∆b
dq(cid:48)(cid:48)G−1(q, q(cid:48)(cid:48))G(q(cid:48)(cid:48), q(cid:48)) =
δ(x − x(cid:48))
0
0
δ(y − y(cid:48))
(4.19)
(4.21)
(4.22)
(4.23)
(cid:90)
(cid:90)
(cid:90)
(d/dt + β)∆v
where q is x or y as appropriate. Inserting (4.20) into (4.21) yields
ϕ(z(cid:48)(cid:48), π, t; x(cid:48)) = δ(x − x(cid:48))
ϕ(z(cid:48)(cid:48), π, t; y(cid:48)) = 0
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))∆v
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))∆ϕ
v(x; x(cid:48)) − 2β
v (x; y(cid:48)) − 2β
(d/dt + β)∆ϕ
(∂t + ∂θ[1 − cos θ + (I(z) + a(z))(1 + cos θ)])∆v
ϕ(y; x(cid:48)) + ∂θ(1 + cos θ)ρ(z, θ)∆v
v(x; x(cid:48)) = 0
(∂t + ∂θ[1 − cos θ + (I(z) + a(z))(1 + cos θ)])∆ϕ
ϕ(y; y(cid:48)) + ∂θ(1 + cos θ)ρ(z, θ)∆ϕ
v (x; y(cid:48)) = δ(y − y(cid:48))
(4.24)
(4.25)
A. Computation of Propagators
In order to perform perturbation theory we must compute the Green's functions or prop-
agators. There are four types of propagators at each spatial location. The propagator
equations are comprised of two sets of 2N coupled integro-partial-differential equations.
They can be simplified to ordinary differential equations, which greatly reduces the com-
putational complexity. The solutions of the equations change qualitatively depending on
whether I + a > 0, suprathreshold regime, and I + a ≤ 0, subthreshold regime. Given that
the propagators depend on two coordinates, there are four separate cases. However, the
subthreshold neurons are by definition silent so propagators with the second variable in the
subthreshold regime are zero, which leaves two cases for the first variable being supra- or
sub- threshold.
13
1. Suprathreshold regime
In the suprathreshold regime, z ∈ {ζ : I+a(ζ) > 0}, we make the following transformation
φ = ϑ>(θ) = 2 tan−1
(cid:112)I(z) + a(z)
tan θ
2
(4.26)
ϕ> : θ → φ where
which obeys
dφ
dθ
=
dϑ>(θ)
dθ
=
√
2
I + a
(1 − cos θ) + (I + a)(1 + cos θ)
= 2πρ(z, θ)
(4.27)
where the last equality comes from (3.23). This transformation has the nice property that
ϑ>(π) = π.
Equations (4.22) and (4.23) transform to
(d/dt + β)∆v
v(z, t; z(cid:48), t(cid:48)) − 2β
(cid:90)
>
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))(cid:112)I(z(cid:48)(cid:48)) + a(z(cid:48)(cid:48))∆v
(cid:90)
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))(cid:112)I(z(cid:48)(cid:48)) + a(z(cid:48)(cid:48))∆ϕ
ϕ(z(cid:48)(cid:48), π, t; z(cid:48), t(cid:48)) = δ(z − z(cid:48))δ(t − t(cid:48))
(4.28)
ϕ(z(cid:48)(cid:48), π, t; z(cid:48), θ(cid:48), t(cid:48)) = 0
(d/dt + β)∆ϕ
v (z, t; z(cid:48), θ(cid:48), t(cid:48)) − 2β
>
ϕ(z, φ, t;·) = ∆·
where we set ∆·
Jacobian of the transformation.
ϕ(z, ϑ−1
> (φ), t;·)(dθ/dφ) ≡ ∆·
(4.29)
ϕ(z, φ, t, ;·), where dθ/dφ is the
Equation (4.25) transforms to
(∂t + ∂φ(dφ/dθ)[1 − cos θ + (I + a)(1 + cos θ)])∆ϕ
ϕ(z, φ, t; z(cid:48), θ(cid:48), t(cid:48))(dφ/dθ)
+ ∂φ(dφ/dθ)(1 + cos θ)ρ(z, θ)∆ϕ
v (z, t; z(cid:48), θ(cid:48), t(cid:48)) = δ(z − z(cid:48))δ(t − t(cid:48))δ(φ − ϑ>(θ(cid:48)))(dφ/dθ)
(4.30)
Now consider
(1 + cos θ)ρ(z, θ) =
=
=
=
1
π
1
π
1
π
1
2π
(cid:112)I + a(z)(1 + cos θ)
(1 − cos θ) + (I + a(z))(1 + cos θ)
√
tan2(θ/2) + (I + a)
I + a
√
I + a
(I + a) tan2(φ/2) + (I + a)
(cid:112)I + a(z)
1 + cos φ
14
(4.31)
where we have used (4.26) and the tangent half-angle formula
tan2 θ
2
=
1 − cos θ
1 + cos θ
Inserting (4.31) back into (4.30) gives
√
(∂t + 2
I + a)∂φ∆ϕ
ϕ(z, φ, t; z(cid:48), θ(cid:48), t(cid:48))
√
− sin φ
2π
I + a
v (z, t; z(cid:48), θ(cid:48), t(cid:48)) = δ(x − x(cid:48))δ(φ − ϑ>(θ(cid:48)))δ(t − t(cid:48))
∆ϕ
(4.32)
(4.33)
Similarly, we obtain
√
I + a∂φ)∆v
(∂t + 2
ϕ(z, φ, t; z(cid:48), t(cid:48)) − sin φ
√
v(z, t; z(cid:48), t(cid:48)) = 0
∆v
(4.34)
2π
I + a
The transformed propagator equations are given by equations (4.28), (4.29), (4.33), and
(4.34). Eqs. (4.33) and (4.34) are advection equations in φ, which can be integrated to
(cid:90) t
(cid:90) t
t(cid:48)
t(cid:48)
ϕ(z, φ, t; z(cid:48), t(cid:48)) = C(z)
∆v
ϕ(z, φ, t; z(cid:48), θ(cid:48), t(cid:48)) = C(z)
∆ϕ
where
dτ sin(φ − ν>(z)(t − τ ))∆v
v(z, τ ; z(cid:48), t(cid:48))
dτ sin(φ − ν>(z)(t − τ ))∆ϕ
v (z, τ ; z(cid:48), θ(cid:48), t(cid:48))
+ δ(φ − ϑ>(θ(cid:48)) − ν>(z)(t − t(cid:48)))δ(z − z(cid:48))
(4.35)
C(z) ≡
2π(cid:112)I(z) + a(z)
ν>(z) ≡ 2(cid:112)I(z) + a(z)
1
(cid:90) t
We then define the following variables:
rv(z, t; z(cid:48), t(cid:48)) = ∆v
rϕ(z, t; z(cid:48), θ(cid:48), t(cid:48)) = ∆ϕ
dτ sin(ν>(z)(t − τ ))∆v
v(z, τ ; z(cid:48), t(cid:48))
ϕ(z, π, t; z(cid:48), t(cid:48)) = C(z)
(cid:90) t
ϕ(z, π, t; z(cid:48), θ(cid:48), t(cid:48)) − δ(π − ϑ>(θ(cid:48)) − ν>(z)(t − t(cid:48)))δ(z − z(cid:48))
t(cid:48)
dτ sin(ν>(z)(t − τ ))∆ϕ
v (z, τ ; z(cid:48), θ(cid:48), t(cid:48))
= C(z)
t(cid:48)
We thus obtain after repeated derivatives and using the propagator equations (4.28), (4.29),
15
(4.36)
(4.37)
(4.38)
(4.33), and (4.34):
(
d
dt
d2
d2
dt2 rv(z, t; z(cid:48), t(cid:48)) =
1
π
+ β)∆v
v(z, t; z(cid:48), t(cid:48)) − β
(cid:90)
v(z, t; z(cid:48), t(cid:48)) − ν2
∆v
>(z)rv(z, t; z(cid:48), t(cid:48))
(4.39)
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))rv(z(cid:48)(cid:48), t(cid:48)(cid:48); z(cid:48), t(cid:48)) = δ(z − z(cid:48))δ(t − t(cid:48)) (4.40)
1
π
dt2 rϕ(z, t; z(cid:48), θ(cid:48), t(cid:48)) =
(cid:90)
v (z, t; z(cid:48), θ(cid:48), t(cid:48)) − ν2
∆ϕ
v (z, t; z(cid:48), θ(cid:48), t(cid:48)) − β
d
(
dt
= βw(z − z(cid:48))ν>(z(cid:48))δ(π − ϑ>(θ(cid:48)) − ν>(z(cid:48))(t − t(cid:48)))
+ β)∆ϕ
>(z)rϕ(z, t; z(cid:48), θ(cid:48), t(cid:48))
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))rϕ(z(cid:48)(cid:48), t(cid:48)(cid:48); z(cid:48), θ(cid:48), t(cid:48))
(4.41)
(4.42)
The covariance function (4.72) involves the integral quantity
(cid:90)
U (z, t; z(cid:48), t0) =
dθ(cid:48)∆ϕ
v (z, t; z(cid:48), t0, θ(cid:48))ρ0(z(cid:48), θ(cid:48))
(4.43)
by our choice of transformation convention. However, instead of computing the propagator
at all values of θ(cid:48), we create another pair of ODEs for U . Applying the integral operator
(cid:82) dθ(cid:48)ρ0(z(cid:48), θ(cid:48)) to (4.41) and (4.42) gives
d2
1
dt2 r(z, t; z(cid:48), t(cid:48)) =
π
+ β)U (z, t; z(cid:48), , t(cid:48)) − β
(
d
dt
(cid:90)
>
U (z, t; z(cid:48), t(cid:48)) − ν2
>(z)r(z, t; z(cid:48), t(cid:48))
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))r(z(cid:48)(cid:48), t(cid:48)(cid:48); z(cid:48), t(cid:48))
(cid:90)
dθ(cid:48)ρ0(z(cid:48), θ(cid:48))δ(π − ϑ(θ(cid:48)) − ν>(z(cid:48))(t − t(cid:48)))
= βw(z − z(cid:48))ν>(z(cid:48))
= βw(z − z(cid:48))ν>(z(cid:48))ρ0(z(cid:48), θ0)
dθ
dφ
ρ0(z(cid:48), θ0)
ρ(z(cid:48), θ0)
w(z − z(cid:48))ν>(z(cid:48))
β
2π
=
(cid:12)(cid:12)(cid:12)(cid:12)θ(cid:48)=θ0
(4.44)
(4.45)
where r(z, t; z(cid:48), t(cid:48)) =(cid:82) rϕ(z, t; z(cid:48), θ(cid:48), t(cid:48))ρ0(z(cid:48), θ(cid:48)) dθ(cid:48) and θ0 = −ϑ−1(ν>(z(cid:48))(t− t(cid:48)))). Hence, we
16
need to numerically integrate the following equations
>(z)r(z, t; z(cid:48), t(cid:48))
U (z, t; z(cid:48), t(cid:48)) − ν2
dt2 r(z, t; z(cid:48), t(cid:48)) =
1
π
d2
(cid:90)
+ β)U (z, t; z(cid:48), , t(cid:48)) − β
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))r(z(cid:48)(cid:48), t(cid:48)(cid:48); z(cid:48), t(cid:48)) =
βw(z − z(cid:48))ν>(z(cid:48))
1
2π
(4.47)
>
(4.46)
d2
dt2 rv(z, t; z(cid:48), t(cid:48)) =
1
π
+ β)∆v
v(z, t; z(cid:48), t(cid:48)) − β
(cid:90)
v(z, t; z(cid:48), t(cid:48)) − ν2
∆v
>(z)rv(z, t; z(cid:48), t(cid:48))
(4.48)
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))rv(z(cid:48)(cid:48), t(cid:48)(cid:48); z(cid:48), t(cid:48)) = δ(z − z(cid:48))δ(t − t(cid:48)) (4.49)
dt2 rϕ(z, t; z(cid:48), π, t(cid:48)) =
+ β)∆ϕ
>
1
π
(cid:90)
v (z, t; z(cid:48), π, t(cid:48)) − ν2
∆ϕ
v (z, t; z(cid:48), π, t(cid:48)) − β
∞(cid:88)
>
>(z)rϕ(z, t; z(cid:48), π, t(cid:48))
(4.50)
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))rϕ(z(cid:48)(cid:48), t(cid:48)(cid:48); z(cid:48), π, t(cid:48))
β
2
w(z − z(cid:48))δ(t − t(cid:48)) + β
w(z − z(cid:48))δ(t − t(cid:48) − Tl(z(cid:48)))
l=1
=
(4.51)
where Tl(z(cid:48)) = {sν(z)s = 2π} marks the time intervals from t(cid:48) such that 2πl−ν>(z(cid:48))Tl(z(cid:48)) =
0. The source at t = t(cid:48) in (4.51) has a factor of one half because because it comes form the θ
delta function, which is symmetric about θ = θ(cid:48) since the propagator is symmetric at θ = θ(cid:48),
unlike the contribution from the time delta function, which is one sided due to causality.
(
d
dt
(
d
dt
d2
(
d
dt
are
2. Subthreshold regime
In the subthreshold regime, namely I + a ≤ 0, the mean field solution for the density ρ is
a point mass, and this will change the form of the propagators. The propagator equations
(d/dt + β)∆v
v(x; x(cid:48)) − 2β
v (x; y(cid:48)) − 2β
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))∆v
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))∆ϕ
ϕ(z(cid:48)(cid:48), π, t; x(cid:48)) = δ(x − x(cid:48))
ϕ(z(cid:48)(cid:48), π, t; y(cid:48)) = 0
(d/dt + β)∆ϕ
(∂t + ∂θ[1 − cos θ + (I(z) + a(z))(1 + cos θ)])∆v
+ ∂θ(1 + cos θ)δ(θ − θ−(z))∆v
(∂t + ∂θ[1 − cos θ + (I(z) + a(z))(1 + cos θ)])∆ϕ
v (x; y(cid:48)) = δ(y − y(cid:48))
+ ∂θ(1 + cos θ)δ(θ − θ−(z))∆ϕ
v(x; x(cid:48)) = 0
ϕ(y; x(cid:48))
ϕ(y; y(cid:48))
(4.52)
(4.53)
(4.54)
(4.55)
where the equations are defined on I(z) + a(z) < 0, and θ± are the mean field fixed points,
where sin θ± = ±2(cid:112)I + a/(1+I +a). However, note that the primed variables are defined
(cid:90)
(cid:90)
17
over the entire z domain since subthreshold neurons can receive input from suprathreshold
neurons.
We simplify these equations by breaking the domain of θ into two pieces: D1 = (θ+, θ−)
and D2 = (θ−, θ+). In the two advection equations, there will be a clockwise advection of the
propagators towards θ− in D1 and in a counterclockwise advection towards θ− in D2. π is in
D1 but not D2 so neurons starting in D2 will never fire. In D1, we make the transformation
ϑ< : θ → χ:
χ = ϑ<(θ) = ln
(cid:33)
χ = 2 coth−1
(cid:32)
tan θ
2
sin θ −(cid:112)I + a(1 + cos θ)
sin θ +(cid:112)I + a(1 + cos θ)
(cid:112)I(z) + a(z)
2(cid:112)I + a
(cid:112)I + a
2(cid:112)I + a
(I + a2 + 1) cosh2(χ/2) − 1
(1 − cos θ) − I + a(1 + cos θ)
(I + a2 + 1)(cosh(χ) + 1) − 2
1 + cos θ =
2
1 + I + a coth2 χ/2
dχ
dθ
=
dθ
dχ
=
dθ
dχ
=
(cid:90)
>
(4.56)
(4.57)
(4.58)
(4.59)
(4.60)
(4.61)
(4.62)
(4.63)
which maps D1 to the real line where −∞ corresponds to θ+ and ∞ corresponds to θ−.
We then have the following propagator equations in the χ representation:
(d/dt + β)∆v
v(z, t; z(cid:48), t(cid:48)) − β
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))∆v
ϕ(z(cid:48)(cid:48), π, t; z(cid:48), t(cid:48)) = δ(z − z(cid:48))δ(t − t(cid:48))
(∂t + ν<∂χ)∆v
(d/dt + β)∆ϕ
(∂t + ν<∂χ)∆ϕ
(cid:90)
ϕ(z, χ, t; z(cid:48), t(cid:48)) = Q(z, χ)∆v
v (z, t; z(cid:48), θ(cid:48), t(cid:48)) − β
ϕ(z, χ, t; z(cid:48), θ(cid:48), t(cid:48)) = Q(z, χ)∆ϕ
>
v(z, t; z(cid:48), t(cid:48))
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))∆ϕ
ϕ(z(cid:48)(cid:48), π, t; z, θ(cid:48), t(cid:48)) = 0
(4.64)
v (z, t; z(cid:48), θ(cid:48), t(cid:48)) + δ(z − z(cid:48))δ(χ − ϑ<(θ(cid:48)))δ(t − t(cid:48))
(4.65)
18
where
Q(z, χ) = −∂χ
and ν<(z) = 2(cid:112)I(z) + a(z). Integrating yields
1 + I + a coth2 χ/2
2
δ(cid:0)ϑ−1(χ) − θ−(z)(cid:1)
ϕ(z, χ, t; w(cid:48)) =
∆v
ϕ(z, χ, t; w(cid:48)) =
∆ϕ
Q(z, χ − ν<(z)(t − τ ))∆v
v(z, τ ; z(cid:48), t(cid:48))dτ
Q(z, χ − ν<(z)(t − τ ))∆ϕ
v (z, τ ; z(cid:48), θ(cid:48), t(cid:48))dτ
+ δ(z − z(cid:48))δ(χ − ϑs(θ(cid:48)) − ν<(z)(t − t(cid:48)))
Hence, the only contribution from the subthreshold neurons are from any neuron that is
initially in D1, which for uniformly distributed phases the probability will be (1 − (θ+ −
θ−)/2π. The subthreshold propagators are thus passively driven by the superthreshold
propagators. Hence, for z in the subthreshold regime, the relevant propagator equations
are:
(d/dt + β)∆v
v(z, t; z(cid:48), t(cid:48)) − β
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))rv(z(cid:48)(cid:48), t; z(cid:48), t(cid:48))
= δ(z − z(cid:48))δ(t − t(cid:48))
(d/dt + β)∆ϕ
v (z, t; z(cid:48), π, t(cid:48)) − β
=
β
2
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))rϕ(z(cid:48)(cid:48), t; z(cid:48), π, t(cid:48))
w(z − z(cid:48))δ(t − t(cid:48)) + β
w(z − z(cid:48))δ(t − t(cid:48) − Tl(z(cid:48))) (4.70)
∞(cid:88)
l=1
(cid:90) t
(cid:90) t
t(cid:48)
t(cid:48)
(cid:90)
>
(cid:90)
>
(cid:90)
(4.66)
(4.67)
(4.68)
(4.69)
>
dz(cid:48)(cid:48)w(z − z(cid:48)(cid:48))ν>(z(cid:48)(cid:48))r(z(cid:48)(cid:48), t; z(cid:48), t(cid:48))
w(z − z(cid:48))ν>(z(cid:48))
(4.71)
(d/dt + β)U (z, t; z(cid:48), t(cid:48)) − β
=
β
2π
B. Covariance functions
1. Drive covariance
As described previously [10], the covariances between the fields to order 1/N are com-
prised of vertices with two outgoing branches. Using the diagrams in Figures 2 and 3, we
19
obtain
N(cid:104)δv(z, t)δv(z(cid:48), t(cid:48))(cid:105) = 2β
(cid:90)
(cid:90)
(cid:26)(cid:90)
dz1dz2dτ ∆v
v(z, t; z1, τ )∆ϕ
+ (x ↔ x(cid:48)) −
dz1
dθ∆ϕ
v (z(cid:48), t(cid:48); z2, π, τ )w(z1 − z2)ρ(z2, π)
(cid:27)
v (z, t; z1, t0, θ)ρ(z1, θ, t0)
(cid:90)
×
dθ(cid:48)∆ϕ
v (z(cid:48), t(cid:48); z1, t0, θ(cid:48))ρ(z1, θ(cid:48), t0)
(4.72)
Evaluating the covariance function in (4.72) requires computing the propagators using
the equations derived in the previous section. Our numerical methods for integrating these
equations are in the Appendix. We compared the theory to microscopic simulations of (2.4)
with fixed initial condition of u(z) set to the mean field solution a(z), and the initial condition
of θ(z) is sampled from the probability distribution obeying the mean field solution ρ(z, θ).
For the supra-threshold region, the cumulative distribution function for ρ(z, θ) is
(cid:32)
(cid:33)
P (z, θ) =
tan−1
1
π
(cid:112)I(z) + a(z)(1 + cos θ)
sin θ
+
1
2
(4.73)
from which we can sample θ by applying the inverse of (4.73) to a uniform random number.
For the sub-threshold region, all the samples are taken to be at the stable solution θ−(z) =
−2 tan−1((cid:112)(I(z) + a(z))).
A comparison between the variance of synaptic drive fluctuations for the microscopic
simulation as a function of space at a fixed time for two values of N and the theory is
shown in figure 4a) for external input and synaptic coupling weight as in figure 1. This is
a case where all neurons are in the supra-threshold region. We see that the theory starts
to break down for smaller system sizes at the local maxima and minima of the variance.
This is expected since the theory is valid to order N−1 in perturbation theory and the
maxima and minima are where the effective local population is smallest. Figure 4 b) shows
the variance near a maximum as a function of N , showing an accurate prediction after
N = 800. The sample size for these microscopic simulations is 5 × 105, and we estimate the
error of the variance using bootstrap. The error is of order 10−2. A segment of the spatio-
temporal dynamics is shown in figure 5. The theory matches the simulation quite well with
the greatest deviation near the maxima and minima.
Figure 6 shows the two-time and
two-space covariances of the synaptic drive for the same network parameters. The spatial
covariance mirrors the coupling function as expected.
20
FIG. 3. Tree level diagrams for a) drive covariance (cid:104)vv(cid:105) and b) rate covariance (cid:104)ϕϕ(cid:105). The lower
two diagrams are zero for a) and b). For the upper 3 diagrams in a) and b), the first diagram
corresponds to the third term, while the second and third diagrams correspond to the first and
second terms of eq. 4.72 and eq. 4.80 respectively.
21
a)b) FIG. 4. Variance times N at time t = 10 for parameters in figure 1. a) Comparison between
microscopic simulation and theory calculation for N = 200 and N = 800. b) The N dependence
of N(cid:104)δu(z)u(z)(cid:105) at z = 0.2. Standard errors for the microscopic simulation are estimated by
bootstrap.
Figure 7 shows a comparison between the theory and the microscopic simulation when
subthreshold neurons are included. There is a good match when N is large. As N decreases
the theory starts to fail at the edges of the bump first. This is likely due to the fact
that the location of the edge could move and this is not captured by the theory since it
assumes fluctuations around a stationary mean field solution. However, the spontaneous
firing of sub-threshold neurons due to either the initial conditions or from the fluctuating
inputs of supra-threshold neurons can cause the edge of the bump to move and this is a
22
1.4 1.5 1.6 1.7 1.8 1.9 0 0.5 1N<δuδu>za) 1.78 1.8 1.82 1.84 0 200 400 600 800 1000 1200 1400 1600N<δuδu>Nb) FIG. 5. Spatial-temporal dynamics of the synaptic drive variance for the microscopic simulation
for N = 800 in a) and c), and the theory in b) and d). Parameters are as in Figure 1.
nonperturbative effect.
2. Rate Covariance
The firing rate is defined as ν = 2η(z, π, t) with mean
(cid:104)ν(z, π, t)(cid:105) = 2ρ(z, π, t)
(4.74)
23
FIG. 6.
Spatiotemporal plot of covariance (cid:104)δu(z, 20)δu(z, 20 − τ )(cid:105) for a) theory and b) mi-
croscopic simulation using parameters from Figure 1. c) Covariance at a single spatial location,
(cid:104)δu(0.5, 20)δu(0.5, 20 − τ )(cid:105). d) Covariance at a single time, (cid:104)δu(z = .005, 20)δu(z(cid:48), 20)(cid:105). Standard
errors are estimated by jackknife.
and covariance
(cid:104)δν(z, t)δν(z(cid:48), t(cid:48))(cid:105) = (cid:104)(ν(z, π, t) − (cid:104)ν(z, π, t)(cid:105))(ν(z(cid:48), π, t(cid:48)) − (cid:104)ν(z(cid:48), π, t(cid:48))(cid:105))(cid:105)
= (cid:104)ν(z, π, t)ν(z(cid:48), π, t(cid:48))(cid:105) − (cid:104)ν(z, π, t)(cid:105)(cid:104)ν(z(cid:48), π, t(cid:48))(cid:105)
= 4(cid:104)η(z, π, t)η(z(cid:48), π, t(cid:48)) − 4ρ(z, π, t)ρ(z(cid:48), π, t(cid:48))(cid:105)
= 4(cid:104)( ϕ(z, t)ϕ(z, t) + ϕ(z, t))( ϕ(z(cid:48), t(cid:48))ϕ(z(cid:48), t(cid:48)) + ϕ(z(cid:48), t(cid:48)))(cid:105)
= 4(cid:104)ϕ(z, t)ϕ(z(cid:48), t(cid:48))(cid:105) + 4(cid:104)ϕ(z, t) ϕ(z(cid:48), t(cid:48))(cid:105)(cid:104)ϕ(z(cid:48), t(cid:48))(cid:105)
= 4(cid:104)ϕ(z, t)ϕ(z(cid:48), t(cid:48))(cid:105) +
ϕ(z, π, t; z(cid:48), π, t(cid:48))ρ(z(cid:48), π, t(cid:48))
∆ϕ
4
N
(4.75)
(4.76)
(4.77)
(4.78)
(4.79)
At tree level, from the diagrams in Figure 3 b),
N(cid:104)ϕ(z, t)ϕ(z(cid:48), t(cid:48))(cid:105) = 2β
dz1dz2dτ ∆v
ϕ(z, t; z1, τ )∆ϕ
(cid:90)
(cid:26)(cid:90)
+ (x ↔ x(cid:48)) −
dz1
dθ∆ϕ
×
dθ(cid:48)∆ϕ
ϕ(z(cid:48), t(cid:48); z1, t0, θ(cid:48))ρ(z1, θ(cid:48), t0)
ϕ(z(cid:48), π, t(cid:48); z2, π, τ )w(z1 − z2)ρ(z2, π, τ )
(cid:27)
ϕ(z, t; z1, t0, θ)ρ(z1, θ, t0)
(4.80)
(cid:90)
(cid:90)
24
-1.6-1.2-0.8-0.4 0 0.4 0.8 1.2 1.6 2 0 0.5 1N<δu(z)δu(z')>z'd) FIG. 7. a) Variance multiplied by N and b) mean of the synaptic drive with subthreshold neurons
for constant stimulus I = −1 and c) coupling weight w(z) = A exp(−az)−exp(−bz)+A exp(−a(L−
z)) − exp(−b(L − z)) with A = 150, a = 30, and b = 20. The supra-threshold edge of bump is at
u = 1. a) is evaluated at time 10. Standard errors are estimated by bootstrap.
25
0 10 20 30 0 0.5 1N<δuδu>zedge of bumpa)-2 0 2 4 6 8 10 0 0.5 1uzb)-10 0 10 20 30 40 50-200-150-100-50 0 50 100 150 200wiji-jc) We rewrite as
N(cid:104)ϕ(z, t)ϕ(z(cid:48), t(cid:48))(cid:105) = 2β
(cid:90)
dz1dz2dτ {rv(z, t; z1, τ )(rϕ(z(cid:48), t(cid:48); z2, π, τ )
(cid:90)
(cid:90)
(cid:90)
(cid:26)(cid:90)
(cid:90)
(cid:90)
(cid:88)
(cid:90)
l
+
2β
ν>(z(cid:48))
(cid:90)
(cid:90)
+ (x ↔ x(cid:48)) −
− 1
2π
− 1
4π2
+δ(π − ϑ>(π) + ν>(z(cid:48))(t(cid:48) − τ ))δ(z(cid:48) − z2))w(z1 − z2)ρ(z2, π, τ )} + (x ↔ x(cid:48))
−
dθ(rϕ(z, t; z1, θ, t0) + δ(π − ϑ>(θ) + ν>(z)(t − t0))δ(z − z1))ρ(z1, θ, t0)
dz1
×
dθ(cid:48)(rϕ(z(cid:48), t(cid:48); z1, θ(cid:48), t0) + δ(π − ϑ>(θ(cid:48)) + ν>(z)(t(cid:48) − t0))δ(z(cid:48) − z1))ρ(z1, θ(cid:48), t0)
(cid:27)
= 2β
dτ
dz1dz2rv(z, t; z1, τ )rϕ(z(cid:48), t(cid:48); z2, π, τ )w(z1 − z2)ρ(z2, π, τ )
dz1rv(z, t; z1, t(cid:48) − 2πl/ν)w(z1 − z(cid:48))ρ(z(cid:48), π, τ )
(cid:90)
(cid:90)
dz1dθrϕ(z, t; z1, θ, t0)ρ(z1, θ, t0)
dθ(cid:48)rϕ(z(cid:48), t(cid:48); z1, θ(cid:48), t0)ρ(z1, θ(cid:48), t0)
dθrϕ(z, t; z(cid:48), θ, t0)ρ(z(cid:48), θ, t0) − 1
2π
dz1δ(z − z1)δ(z(cid:48) − z1)
dθ(cid:48)rϕ(z(cid:48), t(cid:48); z, θ(cid:48), t0)ρ(z, θ(cid:48), t0)
where
r(z, t; z(cid:48), t(cid:48)) =
Hence
(cid:104)δν(z, t)δν(z(cid:48), t(cid:48))(cid:105) = ν>(z)ν>(z(cid:48))
(cid:90)
(cid:18)8β
(cid:90)
(cid:88)
N
l
ϕ(z, π, t; z(cid:48), t(cid:48)) = rv(z, t; z(cid:48), t(cid:48))
∆v
ϕ(z, π, t; z(cid:48), θ(cid:48), t(cid:48)) = rϕ(z, t; z(cid:48), θ(cid:48), t(cid:48)) + δ(π − ϑ>(θ(cid:48)) + ν>(z)(t − t(cid:48)))δ(z − z(cid:48))
∆ϕ
rϕ(z, t; z(cid:48), θ(cid:48), t(cid:48))ρ0(z(cid:48), θ(cid:48)) dθ(cid:48)
(cid:90)
(cid:90)
dτ
dz1dz2rv(z, t; z1, τ )rϕ(z(cid:48), t(cid:48); z2, π, τ )w(z1 − z2)ρ(z2, π, τ )
dz1rv(z, t; z1, t(cid:48) − 2πl/ν>(z(cid:48)))w(z1 − z(cid:48))ρ(z(cid:48), π, τ ) + (x ↔ x(cid:48))
+
ν>(z(cid:48))N
8β
(cid:90)
− 4
N
− 2
πN
4
N
+
dz1r(z, t; z1, t0)r(z(cid:48), t(cid:48); z1, t0)
r(z, t; z(cid:48), t0) − 2
πN
ϕ(z, π, t; z(cid:48), π, t(cid:48))ρ(z(cid:48), π, t(cid:48))
∆ϕ
r(z(cid:48), t(cid:48); z, t0) − 1
π2N
(cid:90)
(cid:19)
dz1δ(z − z1)δ(z(cid:48) − z1)
This quantity is well behaved for t (cid:54)= t(cid:48) and z (cid:54)= z(cid:48). However, in the limit of t(cid:48) → t−, the
26
rate covariance is singular since
ϕ(z, π, t; z(cid:48), π, t(cid:48))ρ(z(cid:48), π, t(cid:48)) = δ(z − z(cid:48))δ(ν>(z)(t − t(cid:48)))
t(cid:48)→t− ∆ϕ
lim
(cid:12)(cid:12)(cid:12)(cid:12)θ=π
dφ
dθ
ρ(z(cid:48), π, t(cid:48))
=
ν>(z)
2ν>(z)
δ(z − z(cid:48))δ(0)ρ(z(cid:48), π, t(cid:48))
(4.81)
(4.82)
We regularize the singularity at t = t(cid:48) by considering the time integral over a small interval:
∆ν(z, t) =
δν(z, s)ds
(cid:90) t+∆t/2
(cid:90)
t−∆t/2
giving
(cid:104)∆ν(z, t)∆ν(z(cid:48), t(cid:48))(cid:105)
∆t2
(cid:90)
(cid:18)8β
(cid:90)
(cid:88)
(cid:90)
N
l
= ν>(z)ν>(z(cid:48))
+
8β
ν>(z(cid:48))N
+ (x ↔ x(cid:48)) − 4
N
dτ
dz1dz2rv(z, t; z1, τ )rϕ(z(cid:48), t(cid:48); z2, π, τ )w(z1 − z2)ρ(z2, π, τ )
dz1rv(z, t; z1, t(cid:48) − 2πl/ν>(z(cid:48)))w(z1 − z(cid:48))ρ(z(cid:48), π, t(cid:48) − 2πl/ν>(z(cid:48)))
dz1r(z, t; z1, t0)r(z(cid:48), t(cid:48); z1, t0)
(cid:90)
(cid:19)
dz1δ(z − z1)δ(z(cid:48) − z1)
− 2
πN
2
+
r(z, t; z(cid:48), t0) − 2
πN
ρ(z(cid:48), π, t(cid:48))δ(z − z(cid:48))
N ∆t
r(z(cid:48), t(cid:48); z, t0) − 1
π2N
(4.83)
We regularize the singularity at z = z(cid:48) by taking a local spatial average over [−cN/2 +
z, cN/2 + z]. We make the approximation that within this local region, the propagator is
constant on space, which is valid under the large N limit. This results in
(cid:104)∆ν(z, t)∆ν(z(cid:48), t(cid:48))(cid:105)
∆t2
dτ
dz1dz2rv(z, t; z1, τ )rϕ(z(cid:48), t(cid:48); z2, π, τ )w(z1 − z2)ρ(z2, π, τ )
dz1rv(z, t; z1, t(cid:48) − 2πl/ν>(z(cid:48)))w(z1 − z(cid:48))ρ(z(cid:48), π, t(cid:48) − 2πl/ν>(z(cid:48)))
(cid:90)
(cid:90)
(cid:18)8β
(cid:90)
(cid:88)
(cid:90)
N
l
= ν>(z)ν>(z(cid:48))
+
8β
ν>(z(cid:48))N
+ (x ↔ x(cid:48)) − 4
N
dz1r(z, t; z1, t0)r(z(cid:48), t(cid:48); z1, t0)
(cid:19)
− 2
πN
r(z, t; z(cid:48), t0) − 2
πN
r(z(cid:48), t(cid:48); z, t0) − 1
π2N c
+
2
∆tcN
ρ(z(cid:48), π, t(cid:48))
(4.84)
Figure 8 shows a comparison of the theory in (4.84) to the microscopic simulations. As
shown in Fig. 8 a), at N = 1200, the theory predicts the mean firing rate well. In Fig. 8 c),
we show the variance of the firing rate at fixed location. In Fig. 8d), we show the spatial
structure of the variance. Again, the theory captures the simulations.
27
FIG. 8. a) Comparison between theory and microscopic simulations of time dependence of mean
firing rate at one spatial location. b) Spatial dependence of mean firing rate at time 3. c) and d)
show the same comparisons for the variance given in equation (4.84). Parameters are from Fig 1
and N=1200. Standard errors are estimated by bootstrap.
V. DISCUSSION
Our goal was to understand the dynamics of a large but finite network of determinis-
tic synaptically coupled neurons with nonuniform coupling.
In particular, we wanted to
quantify the dynamics of individual neurons within the network. We first showed that a
28
0.32 0.33 0.34 0.35 0.36 0 1 2 3<υ(z=0.15)>ta) 0.2 0.3 0.4 0.5 0.6 0 0.25 0.5 0.75 1<υ(z)>zb) 0.28 0.29 0.3 0.31 0.32 0 1 2 3<∆υ-(z)∆υ-(z)>/∆t2tc) 0.1 0.2 0.3 0.4 0.5 0.6 0 0.25 0.5 0.75 1<∆υ-(z)∆υ-(z)>/∆t2zd) self-consistent local mean field theory can describe the dynamics of a single network if the
external input and coupling weight are continuous functions. This imposes a spatial metric
on the network where neurons within a local neighborhood experience similar inputs and
can thus be averaged over locally. This local continuity does not impose any conditions on
long range interactions, which can still be random. We thus propose a new kind of network
to study - continuous randomly coupled spiking networks, where the coupling is continuous
but irregular at longer scales.
We show that corrections to mean field theory can be computed as an expansion in the
number of neurons in a local neighborhood.
In this paper, we have chosen to scale the
local neighborhood to the total number of neurons but this is not necessary. We do this by
first writing down a formal and complete statistical description of the theory, mirroring the
Klimontovich approach used in the kinetic theory of plasmas [13, 18, 21]. This formal theory
is regularized by averaging, which leads to a BBGKY moment hierarchy. As in previous
works [5 -- 8, 10, 18], we showed that the Klimontovich description can be mapped to an
equivalent Doi-Peliti-Jansen path integral description from which a perturbation expansion
in terms of Feynman diagrams can be derived. The path integral formalism is a convenient
tool for calculations. Although we only computed covariances to first order (tree level) it is
straight forward (although computationally intensive) to continue to higher order as well as
compute higher order moments. We only considered a deterministic network for clarity but
our method can easily incorporate stochastic effects, which would just add a new vertex to
the action.
We showed that the theory works quite well for large enough network size, which can
be quite small if all neurons receive suprathreshold input. However, the expansion works
less well for neurons with critical input such as neurons at the edge of a bump where
infinitesimally small perturbations can produce qualitatively different behavior. Quanti-
tatively capturing the dynamics at the edge may require renormalization. The formalism
could be a systematic means to understanding randomly connected networks [28] and the
so-called balanced network [27, 31], where the mean inputs from excitatory and inhibitory
synapses are attracted to a fixed point near zero and the neuron dynamics is dominated by
the fluctuations.
29
VI. ACKNOWLEDGMENTS
This research was supported by the Intramural Research Program of the NIH, NIDDK.
VII. APPENDIX: NUMERICAL METHODS
1. Discretization schemes
We use full backward's Euler for green function calculation for propagators.
= sij
drij
dt
Uij − ν2
dsij
dt
∂tUij = −βUij +
1
π
=
i rij
β
N
(cid:88)
j
wijνjrjk +
β
2π
wijνj
ij + hst
ij
1
π
ij = rt−1
rt
ij = st−1
st
ij = U t−1
U t
ij + h(
ij + h(−βU t
ij − ν2
i rt
U t
ij)
β
N
ij +
(cid:88)
j
wijνjrt
jk +
β
2π
wijνj)
ij
ij − hst
ij = rt−1
rt
(cid:88)
ij − h
1
i rt
st
ij + hν2
π
ij − h
β
U t
N
j
ij = st−1
U t
ij
wijνjrt
jk + hβU t
ij = U t−1
ij + h
β
2π
wijνj
I
hν2. ∗ I
−hβ/N w. ∗ ν(cid:48)
0
−hI
I −h/πI
I + hβI
0
rt·j
st·j
U t·j
=
rt−1·j
st−1·j
U t−1
·j + hβ/2πw.jνj
(7.1)
(7.2)
(7.3)
(7.4)
(7.5)
(7.6)
(7.7)
(7.8)
(7.9)
(7.10)
We add the spike terms in equation (4.51) directly to the propagator ∆ϕ
ν (z, t; z(cid:48), π, t(cid:48)) for
all possible l when t− Tl(z(cid:48)) = t(cid:48). These spike terms add stiffness to the differential equation
30
and explicit differential equation solvers like Runge-Kutta have poor stability properties.
[1] L. F. Abbott and Carl van Vreeswijk. Asynchronous states in networks of pulse-coupled
oscillators. Phys. Rev. E, 48:1483 -- 1490, Aug 1993.
[2] Shun-ichi Amari. Dynamics of pattern formation in lateral-inhibition type neural fields. Bio-
logical Cybernetics, 27(2):77 -- 87, Jun 1977.
[3] Sompolinsky H. Ben-Yishai R, Bar-Or RL. Theory of orientation tuning in visual cortex. Proc
Natl Acad Sci U S A., 1995.
[4] Nicolas Brunel and Vincent Hakim. Fast global oscillations in networks of integrate-and-fire
neurons with low firing rates. Neural Computation, 11(7):1621 -- 1671, 1999.
[5] Michael Buice and Carson Chow. Generalized activity equations for spiking neural network
dynamics. Frontiers in Computational Neuroscience, 7:162, 2013.
[6] Michael A. Buice and Carson C. Chow. Correlations, fluctuations, and stability of a finite-size
network of coupled oscillators. Phys. Rev. E, 76:031118, Sep 2007.
[7] Michael A Buice and Carson C Chow. Beyond mean field theory: statistical field theory for
neural networks. Journal of Statistical Mechanics: Theory and Experiment, 2013(03):P03003,
2013.
[8] Michael A. Buice, Jack D. Cowan, and Carson C. Chow. Systematic fluctuation expansion
for neural network activity equations. Neural Computation, 22(2):377 -- 426, 2010. PMID:
19852585.
[9] Rashmi C. Desai and Robert Zwanzig. Statistical mechanics of a nonlinear stochastic model,
volume 19. 07 1978.
[10] Michael A.Buice Carson C.Chow. Dynamics finite size effects in spiking neural network. PLOS,
January 2013.
[11] Michael A. Cohen and Stephen Grossberg. Absolute stability of global pattern formation and
parallel memory storage by competitive neural networks. In Stephen Grossberg, editor, The
Adaptive Brain I, volume 42 of Advances in Psychology, pages 288 -- 308. North-Holland, 1987.
[12] David Dahmen, Hannah Bos, and Moritz Helias. Correlated fluctuations in strongly coupled
binary networks beyond equilibrium. Phys. Rev. X, 6:031024, Aug 2016.
[13] D.R.Nicholson. Introduction to Plasma Theory, volume 2. Cambridge University Press, 1984.
31
[14] Grgory Dumont, Alexandre Payeur, and Andr Longtin. A stochastic-field description of finite-
size spiking neural networks. PLOS Computational Biology, 13(8):1 -- 34, 08 2017.
[15] B. Ermentrout. Type i membranes, phase resetting curves, and synchrony. Neural-Comput,
8, 1996.
[16] Nicolas Fourcaud and Nicolas Brunel. Dynamics of the firing probability of noisy integrate-
and-fire neurons. Neural Computation, 14(9):2057 -- 2110, 2002.
[17] M Helias, T Tetzlaff, and M Diesmann. Echoes in correlated neural systems. New Journal of
Physics, 15(2):023002, 2013.
[18] Eric J. Hildebrand, Michael A. Buice, and Carson C. Chow. Kinetic theory of coupled oscil-
lators. Phys. Rev. Lett., 98:054101, Jan 2007.
[19] J J Hopfield. Neural networks and physical systems with emergent collective computational
abilities. Proceedings of the National Academy of Sciences, 79(8):2554 -- 2558, 1982.
[20] Wilson HR and Cowan JD. Excitatory and inhibitory interactons in localized populations of
model neurons. Biophysical Journal, 1:1 -- 24, Jan 1972.
[21] S. Ichimaru. Basic Principles of Plasma Physics: A Statistical Approach. A lecture note and
reprint series. W. A. Benjamin, 1973.
[22] S.R. Jones and N. Kopell. Local network parameters can affect inter-network phase lags in
central pattern generators. Journal of Mathematical Biology, 52(1):115 -- 140, Jan 2006.
[23] Eva Lang and Wilhelm Stannat. Finite-size effects on traveling wave solutions to neural field
equations. Journal of Mathematical Neuroscience, 7:5, 2017.
[24] Selva K. Maran and Carmen C. Canavier. Using phase resetting to predict 1:1 and 2:2 locking
in two neuron networks in which firing order is not always preserved. Journal of Computational
Neuroscience, 24(1):37 -- 55, Feb 2008.
[25] Renato E. Mirollo and Steven H. Strogatz. Synchronization of pulse-coupled biological oscil-
lators. SIAM J. Appl. Math., 50(6):1645 -- 1662, November 1990.
[26] Duane Q. Nykamp and Daniel Tranchina. A population density approach that facilitates large-
scale modeling of neural networks: Extension to slow inhibitory synapses. Neural Computation,
13(3):511 -- 546, 2001.
[27] Srdjan Ostojic. Two types of asynchronous activity in networks of excitatory and inhibitory
spiking neurons. Nature Neuroscience, 17:594 EP -- , 02 2014.
[28] H. Sompolinsky, A. Crisanti, and H. J. Sommers. Chaos in random neural networks. Phys.
32
Rev. Lett., 61:259 -- 262, Jul 1988.
[29] Jonathan D. Touboul and G. Bard Ermentrout. Finite-size and correlation-induced effects in
mean-field dynamics. Journal of Computational Neuroscience, 31(3):453 -- 484, Nov 2011.
[30] Alessandro Treves. Mean-field analysis of neuronal spike dynamics. Network: Computation
in Neural Systems, 4(3):259 -- 284, 1993.
[31] C van Vreeswijk and H Sompolinsky. Chaos in neuronal networks with balanced excitatory
and inhibitory activity. Science, 274:1724 -- 1726, 1996.
[32] K Zhang. Representation of spatial orientation by the intrinsic dynamics of the head-direction
cell ensemble: a theory. Journal of Neuroscience, 16(6):2112 -- 2126, 1996.
33
|
1202.2491 | 1 | 1202 | 2012-02-12T06:22:12 | Analysis of inverse stochastic resonance and the long-term firing of Hodgkin-Huxley neurons with Gaussian white noise | [
"q-bio.NC"
] | In previous articles we have investigated the firing properties of the standard Hodgkin-Huxley (HH) systems of ordinary and partial differential equations in response to input currents composed of a drift (mean) and additive Gaussian white noise. For certain values of the mean current, as the noise amplitude increased from zero, the firing rate exhibited a minimum and this phenomenon was called inverse stochastic resonance (ISR). Here we analyse the underlying transitions from a stable equilibrium point to the limit cycle and vice-versa. Focusing on the case of a mean input current density $\mu=6.8$ at which repetitive firing occurs and ISR had been found to be pronounced, some of the properties of the corresponding stable equilibrium point are found. A linearized approximation around this point has oscillatory solutions from whose maxima spikes tend to occur. A one dimensional diffusion is also constructed for small noise based on the correlations between the pairs of HH variables and the small magnitudes of the fluctuations in two of them.
Properties of the basin of attraction of the limit cycle (spike) are investigated heuristically and also the nature of distribution of spikes at very small noise corresponding to trajectories which do not ever enter the basin of attraction of the equilibrium point. Long term trials of duration 500000 ms are carried out for values of the noise parameter $\sigma$ from 0 to 2.0, with results appearing in Section 3. The graph of mean spike count versus $\sigma$ is divided into 4 regions $R_1,...,R_4,$ where $R_3$ contains the minimum associated with ISR. | q-bio.NC | q-bio | Analysis of inverse stochastic resonance
and the long-term firing of
Hodgkin-Huxley neurons with Gaussian
white noise
Henry C. Tuckwell1†, Jurgen Jost 2
1 Max Planck Institute for Mathematics in the Sciences
Inselstr. 22, 04103 Leipzig, Germany
† Corresponding author: [email protected]
August 17, 2018
2
1
0
2
b
e
F
2
1
]
.
C
N
o
i
b
-
q
[
1
v
1
9
4
2
.
2
0
2
1
:
v
i
X
r
a
1
Abstract
In previous articles we have investigated the firing properties of the
standard Hodgkin-Huxley (HH) systems of ordinary and partial differen-
tial equations in response to input currents composed of a drift (mean)
and additive Gaussian white noise. For certain values of the mean cur-
rent, as the noise amplitude increased from zero, the firing rate exhibited
a minimum and this phenomenon was called inverse stochastic resonance
(ISR). Here we analyse the underlying transitions from a stable equilib-
rium point to the limit cycle and vice-versa. Focusing on the case of a
mean input current density µ = 6.8 at which repetitive firing occurs and
ISR had been found to be pronounced, some of the properties of the corre-
sponding stable equilibrium point are found. A linearized approximation
around this point has oscillatory solutions from whose maxima spikes tend
to occur. A one dimensional diffusion is also constructed for small noise
based on the correlations between the pairs of HH variables and the small
magnitudes of the fluctuations in two of them. Properties of the basin
of attraction of the limit cycle (spike) are investigated heuristically and
also the nature of distribution of spikes at very small noise correspond-
ing to trajectories which do not ever enter the basin of attraction of the
equilibrium point. Long term trials of duration 500000 ms are carried out
for values of the noise parameter σ from 0 to 2.0, with results appearing
in Section 3. The graph of mean spike count versus σ is divided into 4
regions R1, ..., R4, where R3 contains the minimum associated with ISR.
In R1 noise has practically no effect until a critical value of σ = σc1 is
reached. At a larger critical value σ = σc2 , the probability of transitions
from the basin of attraction of the equilibrium point to that of the limit
cycle becomes greater than zero and the spike rate thereafter increases
with increasing σ. The quantitative scheme underlying the ISR curve
is outlined in terms of exit time random variables and illustrated dia-
grammatically. In the final subsection 3.4, several statistical properties of
the main random variables associated with long term spiking activity are
given, including distributions of exit times from the two relevant basins
of attraction and the interspike interval.
Short Title: Hodgkin-Huxley
Keywords and Phrases: Hodgkin-Huxley equations, noise
1
Introduction
The Hodgkin-Huxley [1] systems of ordinary and partial differential equations,
based on the electrophysiology of the squid giant axon, are the cornerstone of
mathematical models of single neurons as well as several types of cardiac cells.
Recent such studies include those of Komendantov et al. [2] for hypothathalamic
magnocellular neuroendocrine cells, Saarinen et al.
[3] for cerebellar granule
cells, Williams et al. [4] for ventricular myocytes, Kameneva et al. [5] for retinal
ganglion cells and Drion et al.
[6] for dopaminergic neurons. Many of these
computational cell models contain 10 or more components as the important
roles of many different ion channels have been discovered since the appearance
2
of the HH model. Analysis of such higher-dimensional models is very complex
as there may be 50 or more parameters in distinction to the relatively few in
the 4-component Hodgkin-Huxley system. Since the latter does in fact embrace
some of the basic firing properties of neurons in general, there has naturally been
a large number of analyses and computational studies of the HH systems. These
include both deterministic (for example, [7, 8, 9, 10, 11, 12]) and stochastic (for
example, [13, 14, 15, 16, 17]) modeling.
In recent articles [19, 20, 21] we have explored the effects of both additive
Gaussian white noise and conductance noise on repetitive firing in the Hodgkin-
Huxley system. In the additive noise case for the ordinary differential equation
(ODE) model, when the mean input current density µ is not far above the
threshold of 6.4 µA/cm2 for repetitive firing, the number of spikes in the first
500 or 1000 ms was found to undergo a pronounced minimum (ISR) as the noise
level σ increased from zero [19]. The minimum occurred around σ = 0.35. Guo
[22] has recently found similar results for the HH ODE system with colored
(Ornstein-Uhlenbeck process) noise. Similar results were found for the partial
differential equation (PDE) system [20, 21] where the spatial distribution of
the noise was also an important factor, which led to the disinction between the
effects of noise on the instigation and propagation of spikes.
1.1 Model description
In this article we restrict attention to the HH ODE system, which corresponds
to a uniformly polarized or "space-clamped" neuron. The system of stochastic
differential equations was given in our previous articles but are repeated here
for completeness and notation:
dV =
1
C
[µ + gKn4(VK − V ) + gN am3h(VN a − V ) + gL(VL − V )]dt + σdW (1)
and for the auxiliary variables
dn = [αn(1 − n) − βnn]dt
dm = [αm(1 − m) − βmm]dt
dh = [αh(1 − h) − βhh]dt
(2)
(3)
(4)
where C is the membrane capacitance per unit area, µ, which may depend on t,
is the mean input current density, gK, gN a and gL are the maximal (constant)
potassium, sodium and leak conductances per unit area with corresponding
equilibrium potentials VK, VN a, and Vl, respectively. The noise enters as the
derivative of a standard Wiener process W and has amplitude σ. The auxiliary
variables are n(t), the potassium activation, m(t), the sodium activation and
h(t), the sodium inactivation. The coefficients in the differential equations for
the auxiliary variables as functions of depolarization are
αn(V ) =
10 − V
100[e(10−V )/10 − 1]
(5)
3
βn(V ) =
e−V /80
1
8
25 − V
αm(V ) =
10[e(25−V )/10 − 1]
βm(V ) = 4e−V /18
e−V /20
αh(V ) =
7
100
βh(V ) =
1
e(30−V )/10 + 1
(6)
(7)
(8)
(9)
(10)
2 Stable equilibrium point and limit cycle
When µ is above the critical value µc1 for repetitive firing (saddle-node bifur-
cation) and smaller than the value µc2 at which there is a subcritical Hopf
bifurcation, there are two attractors consisting of a limit-cycle (action potential
trajectory) and a stable equilibrium point.
2.1 Stable equilibrium point x∗
Let us denote the random vector (V, n, m, h) by X = (X1, X2, X3, X4) and
rewrite the system of equations as
dX1 = F1(X)dt + σdW
dXk = Fk(X)dt
(11)
(12)
where k = 2, 3, 4. Because a mean input current density of µ =6.8 µA/cm2 was
found to exhibit a pronounced minimum in firing as σ increased, we will focus
on this value of µ unless stated otherwise.
Setting, with zero noise,
F (x∗) = 0
(13)
yields equilibrium points for the deterministic Hodgkin-Huxley system.
The Jacobian matrix at the equilibrium point is defined as
∂xj
x∗
4
(cid:26) ∂Fi
(cid:27)
J(x∗) =
, i, j = 1, ..., 4.
(14)
Figure 1: Eigenvalues in the complex λ-plane for three values of µ. The values
of the real eigenvalue near λ = −0.13 are indistinguishable in this diagram.
Figure 1 shows the eigenvalues in the complex lambda plane for values of
µ = 5, below the critical value µc1, µ = 7.5, between the critical values µc1,
and µc2 , and µ = 10 which is above µc2. The nature of the critical point at
the various values of µ can be readily seen. For the chosen value of µ = 6.8,
numerical evaluation gives an equilibrium point at
x∗ = (4.0536, 0.38107, 0.084327, 0.45129)
(15)
at which the components of F are
(−0.00000003084, 0.000024246,−0.0000013509, 0.0000048941)
(16)
Numerical evaluation of the eigenvalues of J(x∗) gives λ1 = −4.641, the complex
conjugate pair λ2,3 = −0.630 ± 0.548i and λ4 = −0.1323. Hence x∗ is an
asymptotically stable spiral point.
2.2 The linear approximation
The system of stochastic ordinary differential equations for the process X ob-
tained by linearizing about x∗ is
4(cid:88)
k=1
d X1 =
Xkdt + σdW
(17)
∂F1
∂xk
5
4(cid:88)
d Xj =
k=1
∂Fj
∂xk
Xkdt, j = 2, 3, 4,
(18)
where the partial derivatives are evaluated at x∗. The Jacobian is found nu-
merically to be
−0.010891
−1.2764
0.000030551 −0.0019202
0.00032794
−0.000044773
0
0
102 ∗
(19)
1.2710
0.079467
0
−0.034876
0
0
0
−0.0012664
Note that the system (17)-(18) is linear and so X is a Gaussian process.
Using the theory in Rodriguez and Tuckwell [23], in the absence of an imposed
spiking threshold, the exact distribution of X can be found at any t.
Between spikes the linearized system (17)-(18) is expected to provide a rea-
sonable approximation to the fully nonlinear system (1)-(4). This is clearly
demonstrated by the two sets of sample paths shown in Figure 2. With input
parameters µ = 6.8, σ = 0.6, a time segment of length about 60 ms gave the
sample path for V (X1) shown in the top part of the Figure. With the same
path for the Wiener process (or the white noise), the sample path for the voltage
X1 in the linearized system, with the same initial values, is seen in the lower
part of Figure 2, to mimic closely that in the upper part. However, there is one
very striking difference betwen the two paths as at about t = 59 ms the voltage
in the linear system attains a local maximum and then decreases quite rapidly,
continuing in an oscillatory fashion. In contrast, at about the same t and volt-
age values, a spike arises in the nonlinear system. However, these paths indicate
that the time of spiking in the nonlinear system can be well approximated as
the time at which the voltage in the linear sytem first attains a threshold value.
Such a first passage time can be determined from the usual first exit time theory
for diffusion processes.
6
Figure 2: In the top part, voltage is plotted versus time for a sample path in
the nonlinear full Hodgkin.Huxley system with the parameters shown. A spike
forms near the end of the record. In the lower record is shown the voltage path
for the system linearized about the stable equilibrium point. The Wiener path
is the same in both records to enable a comparison to be made.
2.3 Numerical examples
Without an imposed threshold condition, no spikes are possible in the linearized
stochastic system described by (17) and (18) because there is only a stable equi-
librium point about which trajectories fluctuate. This is of course in distinction
to the Hodgkin-Huxley system (with suitable input parameters) where trajec-
tories may, if the fluctuations are large enough, give rise to spikes around the
limit cycle. It is of interest to examine some statistical properties of the original
process X in nonspiking periods. An example of paths with µ = 6.8 and σ = 0.1
over a 50 ms nonspiking period is shown in Figure 2. The basic statistical prop-
erties of the components over a 500 ms time period are given in Table 1. In
Table 2 are given the corresponding correlation coefficients.
7
Table 1: Statistics of variables during non-spiking
period. µ = 6.8, σ = 0.1
Mean Max
4.446
0.3828
0.0874
0.4542
V 4.049
n
0.3811
m 0.0843
h
0.4515
Min
3.658
0.3795
0.0811
0.4488
St Dev Coef Var
0.1337
6.61e-4
0.0012
0.001
0.0330
0.0017
0.0143
0.0023
Table 2: Correlation coefficients during non-spiking
period. µ = 6.8, σ = 0.1
V
V 1.0000
n
0.3068
m 0.9562
h
-0.2137
n
0.3068
1.0000
0.4649
-0.9894
m
0.9562
0.4649
1.0000
-0.3699
h
-0.2137
-0.9894
-0.3699
1.0000
Figure 3: Sample paths for the 4 components of X over a 50 ms period during
which there were no spikes. Input parameters µ = 6.8 and σ = 0.1.
Tables 3 and 4 give the statistical properties during a 500 ms nonspiking
time period when the noise level is σ = 0.6.
8
Table 3: Statistics of variables during non-spiking
period. µ = 6.8, σ = 0.6
Mean Max
8.962
0.419
0.1393
0.4668
V 4.149
n
0.383
m 0.0859
h
0.4471
Min
-0.206
0.368
0.0538
0.3883
St Dev Coef Var
1.301
0.0077
0.0126
0.0122
0.314
0.020
0.147
0.0274
Table 4: Correlation coefficients, µ = 6.8, σ = 0.6
V
V 1.0000
n
0.3505
m 0.9690
h
-0.2408
n
0.3505
1.0000
0.5284
-0.9854
m
0.9690
0.5284
1.0000
-0.4272
h
-0.2408
-0.9854
-0.4272
1.0000
There are two very noticeable features of the sample paths in the nonspiking
periods examined. Firstly, the oscillatory character of the paths, which is trace-
able to the eigenvalues of the Jacobian at the equilibrium point. Secondly, the
strong positive correlation between V and m and the strong negative correlation
between n and h.
Figure 4: Showing how single, double and multiple spikes arise from oscillations
in the nonlinear full Hodgkin-Huxley system with noise. µ = 6.8
9
The oscillatory nature of the paths can never be captured in the standard
integrate and fire models nor the leaky integrate and fire models [24]. When
the noise is sufficiently large to make for fairly frequent spiking, the times of
spiking must tend to arise at the maxima in the oscillations of V . This is seen
dramatically in Figure 3 and in Figure 2 of the previous subsection.
2.4 A one-dimensional diffusion
Examination of the statistical properties of the process as given in Tables 1
and 2 shows that for small noise the coefficients of variation of n and h are an
order of magnitude smaller than those for V and m and that the correlation
coefficient of V and m is close to unity. These observations make it reasonable
to consider a 1-dimensional approximation V to the HH system in which a
constant n ≈ n and a constant h ≈ h, with m = kV , where k is another
constant. These approximations lead, on putting C = 1 in (1), to the following
stochastic differential equation for V ,
d V = (µ + c1 − c2 V + c3 V 3 − c4 V 4)dt + σdW,
where
c1 = gKn4VK + gLVL
c2 = gKn4 + gL
c3 = gN ak3hVN a
c4 = gN ak3h.
(20)
(21)
(22)
(23)
(24)
Standard theory gives for such a diffusion that the mean exit time from a value
x ∈ (a, b) to outside this interval satisfies the ordinary differential equation
(σ2/2)M(cid:48)(cid:48) + (µ + c1 − cx + c3x3 − c4
x)M(cid:48) = −1, x ∈ (a, b),
(25)
where primes denote differentiation, along with suitable boundary conditions.
Preliminary investigations of the validity of this approximation were made using
values for µ = 6.8 and σ = 0.1 (see Tables 1 and 2) but a detailed study will be
reported in a subsequent article.
2.5 The limit cycle and its basin of attraction
With constant input current density in the interval [µc1 , µc2), repetitive spiking
may occur with a fixed period. For µ = 6.8 and no noise the period is about
17.65 ms. The limit cycle is in 4-space but we here show in the upper part of
Figure 2 the projection of the limit cycle obtained by plotting n versus V . In
the lower part of the figure is shown the position of the stable equilibrium point,
here designated R, for the same value of µ. It can be seen that the limit cycle
approaches quite close to R. In a previous article in which the moment method
was used to explore the effects of noise on HH spiking (Tuckwell and Jost, 2009),
we heuristically estimated the basin of attraction of the stable equilibrium point.
10
Part of the basin of attraction of the limit cycle can be numerically estimated
by taking the union of all stochastic paths which do not collapse to the stable
equilibrium point. As explained later, there are ranges of values of σ where the
stochastic paths have this property and Figure 6 shows a sample of such paths
for various such σ. Here paths were taken over 1000 msec involving about 50
spikes for values of σ ranging from 0.05 to 0.25. This picture gives an idea of
how far off the deterministic limit cycle a path may wander without entering
the basin of attraction of the stable equilibrium point. The variability of paths
is relatively small.
Figure 5: In the top part, voltage is plotted versus potassium activation vari-
ablefor the deterministic path of repetitive spikes, depicting the limit cycle for
the HH ODE system with µ = 6.8 > µc1. In the lower part the limit cycle is
magnified in the vicinity of the stable rest point.
11
Figure 6: The union of many stochastic paths for values of the noise parameter
σ for which paths did not collapse into the stable point in a time interval of at
least 1000 ms.
Another way to see the limited variability of the times to complete a spike
orbit (limit cycle) is displayed in Figure 7. Here spikes were observed during
repetitive firing during periods in which no transitions to the basin of attraction
of the rest point occurred, with noise levels from σ = 0 to σ = 0.4. The statistical
properties of the interspike interval (ISI) are shown in the Figure. The most
salient feature is that the mean ISI is practically constant (blue triangles) as the
noise varies, staying in the interval [17.58, 17.82] for these values of σ. Naturally
the standard deviation of the ISI increases (green squares) as σ increases, the
maximum ISI increases /red circles) and the minimum ISI decreases (black
diamonds).
12
Figure 7: Some statistical measures of the ISI during repetitive spiking at various
noise levels with µ = 6.8. For σ = 0, there is no variability. The mean ISI is
shown with (blue) triangles, the (green) squares denote the mean + 3 standard
deviations, the (red) circles denote the maxima and the (black) diamonds denote
the minima. Until σ is about 0.07, there are no cessations of spiking up to 500000
ms.
Examples of distributions of the ISI during repetitive spiking at small noise
levels are shown in Figure 8. In the top panel is an ISI histogram for a noise level
(σ = 0.07) at which there is apparently no cessation of spiking over extremely
long (infinite?) time periods with 28429 ISIs in 500000 ms. The distribution is
roughly Gaussian and has a mean of 17.59 ms and a standard deviation of 0.221
ms. For the second histogram shown, σ = 0.085, which is just greater than the
critical value at which spiking may stop in a finite time. The number of spikes
is only 4397 and the mean and standard deviation of the ISI are 17.60 ms and
0.276 ms, respectively. The distribution of the ISI is practically the same in
both cases.
13
Figure 8: Histograms of the ISI during repetitive spiking at two small noise
levels. Top, σ = 0.070; bottom, σ = 0.085. For σ = 0.07 there are no cessations
of spiking up to 500000 ms whereas for σ = 0.085 firing stops after about 4400
spikes.
3 Results on spiking for the HH system and in-
verse stochastic resonance
In the following it is assumed that µ is such that repetitive spiking does in
fact occur. In relation to the stochastic paths for the HH system of stochastic
equations we define the following two random variables.
Firstly, the exit time of the process to escape from the basin of attraction
BL of the limit cycle L to that BR of the rest point R is by
TL→R(x0), x0 ∈ BL
(26)
where (x0) = (V0, n0, m0, h0) is an initial point. Secondly the exit time for the
process to escape from the basin of attraction BR of R to that of the limit cycle
is
TR→L(x0), x0 ∈ BR.
(27)
14
Using the standard theory for diffusion processes, Kolmogorov second order par-
tial differential equations for the moments and distributions of these quantities
as a function of initial values were described in our previous work (Tuckwell et
al., 2009). However, any attempt to solve these equations analytically or even
by numerical methods for PDEs, apart from being a formidable task, requires
an exact or even approximate knowledge of the two basins of attraction which is
unfortunately not presently available. In fact the probabilistic nature of BL and
BR is not completely understood because it seems that for small enough noise,
at least for some values of µ, escape from BL may be impossible, not being
observed in extremely long time periods and similarly, for particular ranges of
σ, for escape from BR.
3.1 Long term trials and data collected
We are interested in obtaining samples of meaningful sizes for the above random
variables and determining their properties, mainly as a function of σ. Since some
of the exits from one basin of attraction to the other are, for some values of σ,
very rare events the simulations of solutions of the full stochastic HH system
have to be performed over very long time periods. These were chosen to be
500000 ms or about 8 minutes and 20 seconds with a timestep of 0.065 ms.
Results for the 500000 ms interval were obtained as the union of 100 trials of
length 5000 ms, where the final values of V, n, m, h from any trial were used
as the initial values for the next trial. Then the data on t, V, n, m, h for the
100 trials were concatenated to give 5 single vectors each with over 7.5 million
elements. The records for V were analyzed to determine the times of occurrence
of spikes and from these the interspike intervals were determined. In total, 50
trials of length 500000 were simulated.
3.2 All spikes
The numbers of spikes were recorded in each of 50 trials of length 500000 ms
for 40 values of σ. The mean number of spikes, denoted by E[NSP], is plotted
against σ in Figure 9. In the top panel all results are included, whereas in the
lower left panel detail is shown for 0 ≤ σ ≤ 0.1 and in the lower right panel, for
0.1 ≤ σ ≤ 0.5.
The general form of the plot of E[NSP] versus σ is similar to that in Figure
5 of Tuckwell et al. (2009) where the total time period was much less at 1000
ms. Thus, Figure 9 exhibits the phenomenon of inverse stochastic resonance,
which refers to a firing rate which, as noise level increases, at first declines to
a minimum and then becomes greater. However, Figure 9 shows more detail
and reveals 4 distinct regimes marked R1,...,R4. In region R1, which extends
from the deterministic setting of σ = 0 to very close to σ = 0.07, the noise
is almost without effect and the number of spikes is always close to that for
the zero noise, 28431. Region R2 is characterized by a rapidly falling number
of spikes as σ increases from 0.07 to 0.14. At the latter value of σ the mean
number of spikes is 104.8 which is about 0.4% of the maximum value. In region
15
R3, from σ = 0.15 to σ = 0.35, E[NSP] stays below 100 with a broad minimum
of about 9.5 spikes (0.03% of the maximum) around σ = 0.295 to 0.300. By
σ = 0.375, the beginning of region R4, the mean number of spikes has reached
over 120 and then increases sharply and eventually at a slower rate to reach
25883 around σ = 2.
Figure 9: The dependence on the noise parameter σ of the mean of the total
number of spikes E[NSP] in a time interval of length 500000 ms in the nonlinear
HH system. Here µ = 6.8 with 50 trials at each point. The values of σ are
divided into 4 regimes designated R1 to R4. In the top panel, results are given
for all values of σ. In the lower left panel the detail of R1 and the start of R2
are shown. In the lower right panel, the detail of R3 and the start of R4 are
shown.
3.3 Underlying scheme for ISR
The following observations constitute a basis for ISR which occurs at some values
of µ in the HH system described by Equations (1)-(4). We define two critical
values σc1 and σc2 of the noise parameter σ, with 0 < σc1 < σc2 < ∞. σc1 and
σc2 depend on µ, and are only relevant above the critical value for repetitive
firing and possibly for selected initial values of the process (see Tuckwell et al.,
2009). With reference to the two random variables defined by (26) and (27),
16
but without any specific values of x0, we have the following, as schematized
diagrammatically in Figure 10.
0 < σ < σc1.
Repetitive spiking continues, presumably indefinitely.
The probability of a transition from BL to BR at any time is zero.
The expectation of TL→R(x0), x0 ∈ BL, is infinite.
σc1 < σ < σc2.
Repetitive spiking ceases at some time t1 < ∞.
The probability of a transition from BL to BR is greater than zero and increases
with noise level.
The expectation of TL→R(x0), x0 ∈ BL, is finite and and tends to decrease with
increasing σ.
After a transition from BL to BR, the probability of the reverse transition from
BR to BL is zero.
The expectation of TR→L(x0), x0 ∈ BR, is infinite.
After the cessation of repetitive spiking, there is no more spiking.
σ > σc2.
After a transition from BL to BR, the probability of a transition from BR to
BL is greater than zero and tends to increase as σ incresases.
The expectation of TR→L(x0), x0 ∈ BR, is finite and tends to decrease as σ
increases.
After a transition back from BR to BL, the reverse transition occurs with non-
zero pobability, followed again by a transition from BL to BR and so forth,
indefinitely.
Epochs of spiking and nonspiking alternate with variable durations which tend
to become shorter as σ increases.
Figure 10 illustrates some of these aspects. The bottom panel shows part of
the actual E[NSP] curve (normalized) from Figure 9 and the approximate values
of σc1 and σc2 are obtained from this curve. In the top panel, the probabilities
of transitions from BL to BR and BR to BL are indicated as commencing to
increase from zero at the two critical values of σ and rising to unity, although
this final value depends on the initial value x0 which is not fixed throughout.
In the middle panel, descriptions of the expectations of the exit times from BL
and BR are given.
17
Figure 10: A schematic representation for the transitions underlying ISR. There
are shown two critical values of the noise parameter σ and the probabilities of
transitions between the basins of attraction BL and BR of the limit cycle and
rest point are sketched in the top panel for the various ranges of values of σ. In
the bottom panel is shown the normalized expected number of spikes obtained
by simulation over 500000 ms for µ = 6.8, a value of µ at which ISR had been
found to be pronounced.
The scheme given in Figure 10 is in accordance with the results for sample
paths of V for simulated spike trains shown in Figure 11. Here the time periods
are all 500000 ms but in the top two records time is only shown to 2000 ms. The
very small noise case gives apparently indefinite repetitive spiking. Somewhat
larger noise leads to an initial isolated burst followed by silence, possibly for
an infinite time. Increasing the noise further leads to occasional bursting and
eventually the frequent bursting case occurs.
18
Figure 11: Illustrating the 4 basic patterns of spiking activity. Top left: σ =
0.05. Incessant spiking, at least to 500000 ms. Top right: σ = 0.25. One initial
burst with no further spikes, at least to 500000 ms. Bottom left: σ = 0.34.
Occasional bursts with separations of order 100000 ms. Bottom right: σ = 0.45.
Frequent bursts.
3.4 Some statistics of TL→R and TR→L
In this subsection results on certain statistical aspects of the underlying random
variables TL→R and TR→L, defined by (26) and (27), obtained from the long-
term simulation of the HH system (1)-(4) are given.
3.4.1 Escape time from BL
During repetitive spiking the trajectories of the process do not deviate much
from the deterministic limit cycle as illustrated in Figure 6. It is assumed that
there is a well-defined but unkown set in (V, n, m, h)-space, denoted by BL,
containing the deterministic repetitive spiking trajectory, which is called the
basin of attraction of the limit cycle. This implies that if the process started in
BL, then, with no noise, the path would approach the limit cycle. With noise,
trajectories may escape from BL whereupon they enter the basin of attraction
19
BR of the stable equilibrium point. The nature of BL is unknown and it is
not clear that it is a regular open set because if it was it is likely that escape
from it would occur eventually with probability one no matter how small the
noise level. Thus, even though it seems that for σ < σc1 , P (L → R) = 0 and
E[TL→R = ∞], it may be the case that P (L → R) is so small for small enough σ
that the event of this escape is unlikely to be observed in the course of feasible
simulations. It is a remaining mathematical challenge to ascertain the veracity
of these remarks.
Notwithstanding these uncertainties, the mean value of E[TL→R] of the exit
time from BL was estimated from sample paths and the results are shown for
σ ≥ 0.2 in Figure 12. For these noise levels the mean exit time declines rapidly
from very large values to a minimum of about 57 ms at σ = 1.25. Thereafter,
E[TL→R] seems to increase slightly to about 72 ms at σ = 2.
Figure 12: The dependence on the noise parameter σ of the expectation E[TL→R]
of the random variable which is the time of exit of the process from the basin
of attraction of the limit cycle to that of the stable rest point. 50 trials at each
point with µ = 6.8. For values of σ less than a critical value σc1 ≈ 0.07 the
value of this expectation is apparently infinite. Extremely large values which
occurred for σc1 < σ < 0.2 are not shown.
20
3.4.2 Escape time from BR
The basin of attraction of the stable rest point is also unknown exactly. Figures
13 and 14 show estimates of the mean and standard deviation of TR→L over
various ranges of values of σ. Again, it is not known with certainty, but it
appears that the probability of escape from BR to BL is zero (or extremely
close to zero) until σ ≥ σc2. In Figure 13, the mean is shown dropping from
values of order 300000 ms at σ = 0.25 to eventually reach values about 30 ms
at σ = 2. Figure 14 shows corresponding results for the standard deviation, but
only for σ ≥ 0.35 because sample sizes were too small for lesser values of the
noise parameter. In the lower panel of Figure 14 is shown the dependence of the
coefficient of variation of TR→L on noise level. It can be seen that there is an
initial increase in this quantity until a maximum is attained at about σ = 0.5,
whereupon it declines monotonically.
Figure 13: The dependence on the noise parameter σ of the expectation E[TR→L]
of the random variable which is the time of exit of the process from the basin
of attraction of the stable rest point to that of the limit cycle. Here µ = 6.8. 50
trials at each point. For values of σ less than some value just less than 0.25 the
value of this expectation is apparently infinite.
21
Figure 14: Top two panels. The dependence on the noise parameter σ of the
standard deviation σTR→L,of the random variable which is the time of exit of
the process from the basin of attraction of the stable rest point to that of the
limit cycle. Here µ = 6.8. 50 trials at each point. Bottom panel. The coefficient
of variation of the random variable TR→L as a function of σ.
3.4.3 Distributions
There are three main random variables of interest in relation to the empirical
spike trains obtained in this study. These are the previously defined exit times
TL→R and TR→L and in addition the general ISI for the whole train. Samples for
the latter random variable can be viewed as the union of those of the previous
two. Examples of histograms of these random variables are shown in Figure 15.
In all cases the results for 50 trials of duration 500000 ms are combined. In
Figure 15A is given a histogram for all ISIs for σ = 0.2. This occurs in region
R3 of Figure 9, where there are very few spikes, all with short ISIs as there are
no returns to BL fron BR. The histogram is of the same nature as those in
Figure 8. See also Figure 7.
In Figure 15C the raw histogram is shown for all ISIs with σ = 0.350. There
is a preponderance of small intervals and then an exponential-type distribution
of longer (> 21.5 ms) intervals as depicted in Figure 15E. The tail of the expo-
22
nential distribution is very long and extends out to about 500000 ms which is
the limit in these simulations. The distribution depicted in Figure 15E is close
to the actual distribution of TR→L for this value of σ. Similar remarks apply to
the pairs of Figures 15B and 15D (σ = 0.375) and 15F and 15G (σ = 0.5). In
the final plot of Figure 15H, the number of spikes has become enormous and the
majority of ISIs are less than 50 ms as expected for σ = 1 which is in region R4
of Figure 9. In the histograms of Figure 15 C, B, F and H, there are small bin
counts out to very large times and these are not visible compared to the large
bin count at short intervals. However, they are visible at relatively small values
in the truncated histograms.
Figure 15: Examples of distributions (histograms) of ISIs and exit times from
BR based on 50 trials (data pooled) of length 500000 ms. Red histograms, raw
data. Blue histograms, ISIs truncated at 21.5 ms. A. Raw data for σ = 0.2.
B. Raw data for σ = 0.375. C. Raw data for σ = 0.350. D. Truncated data
for σ = 0.375. E. Truncated data for σ = 0.350. F. Raw data for σ = 0.5. G.
Truncated data for σ = 0.5. H. Raw data for σ = 1.0.
To complete the picture for the distributions of the three key random vari-
ables, Figure 16 shows histograms of mean numbers of spikes per burst for two
23
values, 0.4 and 1.0, of σ. If these are multiplied by the mean ISI within bursts,
which is about 17.6 ms, the mean times spent in the basin of attraction of
the limit cycle (BL) and hence of TL→R are estimated. For smaller σ some
very large means were obtained (not shown). Noticeable in Figure 16 is the
smaller magnitude and small variability when σ = 1.0 compared to σ = 0.4.
The distributions in both cases shown are approximately Gaussian which may
be compared with the exponential-types shown in Figures 15 D, E and G.
Figure 16: Histograms of mean numbers of spikes per burst, which indicate
magnitudes of exit times from BL to BR. based on 50 trials of length 500000.
4 Discussion
The inhibitory effects of noise on repetitive spiking in squid axon, on which the
HH system is based, were well documented experimentally by Paydarfar et al.
[25]. The first theoretical evidence of ISR in the full HH sytem was obtained for
µ = 5 in the non-repetitive spiking mode [26] (Figure 2), as it was found that
there was a maximum in the mean ISI at small values of σ. Subsequently ISR has
been demonstrated for repetitive spiking in both the HH system of ODEs and
PDEs [18, 20, 21, 22]. In addition to using the standard initial conditions and
additive noise, in [18] the phenomenon was confirmed with respect to random
24
initial conditions and conductance-based input. Some theory of ISI was also
outlined [18] in terms of bifurcations in the HH sytem [7, 10, 27] and of the exit
times [28] from the basins of attraction of a stable equilibrium point and a limit
cycle. However the simulations were over relatively short time periods and the
statistical details of the exit times were not explored in detail.
In this article we have first explored the properties of the stable equilibrium
point in depth. Linearizing the stochastic HH system about that point gave an
approximate sytem of stochastic differential equations whose oscillatory solu-
tions were found to mimic those of the full system in nonspiking periods. The
oscillations are evidently an integral part of firing in the HH system with the
considered parameters, because spikes did tend to emerge at the maxima. The
distribution of ISIs with trajectories not departing greatly from the limit cycle
for very small noise was estimated and is, as perhaps expected, Gaussian-like.
Long-term simulations, to 500000 ms were performed with µ = 6.8 for many
values of the noise parameter σ. Four ranges of values of σ were distinguished,
based on mean total numbers of spikes. These regions were denoted R1, ..., R4
in Figure 9. Two critical values of σ were also apparent, denoted by 0 < σc1 <
σc2 < ∞. A detailed discussion of the mechanisms underlying ISR in terms of
these critical values was given in Section 3.3. It is not certain whether escape
from the basin of attraction of the limit cycle can ever occur when σ < σc1 or
whether escape from the basin of attraction from the basin of attraction of the
stable equilibrium point can ever occur when σc1 < σ < σc2, but these events
were never observed in 50 trials of length 500000 ms with several values of σ.
References
[1] A.L. Hodgkin, A.F. Huxley, A quantitative description of membrane cur-
rent and its application to conduction and excitation in nerve, J. Physiol.
117 (1952) 500-544.
[2] A.O. Komendantov, N.A.Trayanova, J.G. Tasker, Somato-dendritic
mechanisms underlying the electrophysiological properties of hypotha-
lamic magnocellular neuroendocrine cells: A multicompartmental model
study. J. Comput. Neurosci. 23 (2007) 143-168.
[3] A. Saarinen, M-L.Linne, O. Yli-Harja, Stochastic differential equation
model for cerebellar granule cell excitability. PLoS Comp. Biol. 4 (2008)
e1000004.
[4] G.S.B. Williams, G.D. Smith, E.A. Sobie, M.S. Jafri MS, Models of
cardiac excitation-contraction coupling in ventricular myocytes. Math.
Biosci. 226 (2010) 1-15.
[5] T. Kameneva, H. Meffin, A.N. Burkitt, Modelling intrinsic electrophysi-
ological properties of ON and OFF retinal ganglion cells. J. Comp. Neu-
rosci. 31 (2011) 547-561.
25
[6] G. Drion, L. Massotte, R. Sepulchre, V. Seutin V, How modeling can rec-
oncile apparently discrepant experimental results: the case of pacemaking
in dopaminergic neurons. PLoS Comput. Biol. 7 (2011) e1002050.
[7] B. Hassard, Bifurcation of periodic solutions of the Hodgkin- Huxley
model for the squid giant axon. J. Theor. Biol. 71 (1978) 401-420.
[8] E.N. Best, Null space in the Hodgkin-Huxley equations a critical test.
Biophys J 27 (1979) 87-104.
[9] K. Aihara, G. Matsumoto G, Two stable steady states in the Hodgkin-
Huxley axons. Biophys J 41 (1983) 87-89.
[10] H. Fukai, S. Doi, T. Nomura, S. Sato, Hopf bifurcations in multiple-
parameter space of the Hodgkin-Huxley equations I. Global organization
of bistable periodic solutions. Biol. Cybern. 82 (2000) 215-222.
[11] J.Guckenheimer, R.A. Oliva, Chaos in the HodgkinHuxley model. SIAM
J. App. Dyn. Sys. 1 (2002)105-114.
[12] D. Calitoiu, B.J. Oommen, D. Nussbaum, Spikes annihilation in the
Hodgkin-Huxley neuron. Biol. Cybern. 98 (2008) 239-257.
[13] Y. Horikawa, Noise effects on spike propagation in the stochastic
Hodgkin- Huxley models. Biol. Cybern. 66 (1991)19-25.
[14] D. Brown, J-F. Feng, S. Feerick, Variability of firing of Hodgkin- Hux-
ley and FitzHugh-Nagumo neurons with stochastic synaptic input. Phys.
Rev. Lett. 82 (1999) 4731-4734.
[15] P.H.E. Tiesinga, J.V. Jos´e, T.J. Sejnowski, Comparison of current-driven
and conductance-driven neocortical model neurons with HodgkinHuxley
voltage-gated channels. Phys. Rev. E 62 (2000) 8413-8419.
[16] T.D. Austin, The emergence of the deterministic Hodgkin-Huxley equa-
tions as a limit from the underlying stochastic ion-channel mechanism.
Ann. Appl. Prob. 18 (2006) 1279-1325.
[17] M. Ozer, L.J. Graham LJ, Impact of network activity on noise delayed
spiking for a Hodgkin-Huxley model. Eur. Phys. J. B 61 (2008) 499-503.
[18] H.C. Tuckwell, J.Jost, B.S. Gutkin, Inhibition and modulation of rhyth-
mic neuronal spiking by noise. Phys. Rev. E 80 (2009) 031907.
[19] H.C. Tuckwell, J. Jost J (2009) Moment analysis of the Hodgkin-Huxley
system with additive noise. Physica A 388 (2009) 4115-4125.
[20] H.C.Tuckwell, J. Jost, Weak noise powerfully inhibits rhythmic spiking
but not its propagation. PLoS Comp. Biol. 6 (2010) e1000794.
26
[21] H.C. Tuckwell, J. Jost, The effects of various spatial distributions of weak
noise on rhythmic spiking. J. Comp. Neurosci. 30 (2011) 361-371.
[22] D. Guo, Inhibition of rhythmic spiking by colored noise in neural systems.
Cogn Neurodyn. 5 (2011) 293-300.
[23] R. Rodriguez, H.C. Tuckwell, Statistical properties of stochastic nonlinear
dynamical models of single spiking neurons and neural networks. Phys.
Rev. E 54(1996) 5585-5590.
[24] H.C. Tuckwell. Introduction to Theoretical Neurobiology, Volume 2.
Cambridge University Press, Cambridge UK, 1988.
[25] D. Paydarfar, D.B. Forger, J.R. Clay , Noisy inputs and the induction of
on-off switching behavior in a neuronal pacemaker. J. Neurophysiol. 96
(2006) 3338-3348.
[26] H.C. Tuckwell, Spike trains in a stochastic Hodgkin-Huxley system.
BioSystems 80 (2005) 25-36.
[27] J. Jost, Dynamical Systems, Springer, Berlin, 2005.
[28] H. C. Tuckwell, Stochastic Processes in the Neurosciences, SIAM,
Philadelphia, 1989.
27
|
1807.04691 | 1 | 1807 | 2018-07-12T15:57:20 | Spatial Embedding Imposes Constraints on the Network Architectures of Neural Systems | [
"q-bio.NC"
] | A fundamental understanding of the network architecture of the brain is necessary for the further development of theories explicating circuit function. Perhaps as a derivative of its initial application to abstract informational systems, network science provides many methods and summary statistics that address the network's topological characteristics with little or no thought to its physical instantiation. Recent progress has capitalized on quantitative tools from network science to parsimoniously describe and predict neural activity and connectivity across multiple spatial and temporal scales. Yet, for embedded systems, physical laws can directly constrain processes of network growth, development, and function, and an appreciation of those physical laws is therefore critical to an understanding of the system. Recent evidence demonstrates that the constraints imposed by the physical shape of the brain, and by the mechanical forces at play in its development, have marked effects on the observed network topology and function. Here, we review the rules imposed by space on the development of neural networks and show that these rules give rise to a specific set of complex topologies. We present evidence that these fundamental wiring rules affect the repertoire of neural dynamics that can emerge from the system, and thereby inform our understanding of network dysfunction in disease. We also discuss several computational tools, mathematical models, and algorithms that have proven useful in delineating the effects of spatial embedding on a given networked system and are important considerations for addressing future problems in network neuroscience. Finally, we outline several open questions regarding the network architectures that support circuit function, the answers to which will require a thorough and honest appraisal of the role of physical space in brain network anatomy and physiology. | q-bio.NC | q-bio |
Spatial Embedding Imposes Constraints on the Network Architectures
of Neural Systems
Jennifer Stiso1,2 and Danielle S. Bassett1,3,4,5,6
1Department of Bioengineering, School of Engineering and Applied Sciences,
University of Pennsylvania, Philadelphia, PA, 19104
5Department of Physics & Astronomy, College of Arts and Sciences,
University of Pennsylvania, Philadelphia, PA, 19104
3Department of Electrical and Systems Engineering, School of Engineering and Applied Sciences,
University of Pennsylvania, Philadelphia, PA, 19104
4Department of Neurology, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA, 19104
2Neuroscience Graduate Group, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA, 19104 and
6To whom correspondence should be addressed: [email protected]
A fundamental understanding of the network architecture of the brain is necessary for the further
development of theories explicating circuit function. Recent progress has capitalized on quantitative
tools from network science to parsimoniously describe and predict neural activity and connectiv-
ity across multiple spatial and temporal scales. Perhaps as a historical artifact of its origins in
mathematics or perhaps as a derivative of its initial application to abstract informational systems,
network science provides many methods and summary statistics that address the networks topologi-
cal characteristics with little or no thought to its physical instantiation. Yet, for embedded systems,
physical laws can directly constrain processes of network growth, development, and function, and
an appreciation of those physical laws is therefore critical to an understanding of the system. Re-
cent evidence demonstrates that the constraints imposed by the physical shape and volume of the
brain, and by the mechanical forces at play in its development, have marked effects on the observed
network topology and function. Here, we review the rules imposed by space on the development
of neural networks and show that these rules give rise to a specific set of complex topologies. We
present evidence that these fundamental wiring rules affect the repertoire of neural dynamics that
can emerge from the system, and thereby inform our understanding of network dysfunction in dis-
ease. We also discuss several computational tools, mathematical models, and algorithms that have
proven useful in delineating the effects of spatial embedding on a given networked system and are
important considerations for addressing future problems in network neuroscience. Finally, we out-
line several open questions regarding the network architectures that support circuit function, the
answers to which will require a thorough and honest appraisal of the role of physical space in brain
network anatomy and physiology.
NETWORK TOPOLOGY VERSUS GEOMETRY
by studies of regional activation (Box 1).
IN NEURAL SYSTEMS
In contemporary neuroscience, increasing volumes of
data are being brought to bear on the question of how
heterogeneous and distributed interactions between neu-
ral units might give rise to complex behaviors. Such
interactions form characteristic patterns across multiple
spatial scales, spanning from the relatively small scales of
molecules and cells, to the relatively large scales of areas
and lobes [1]. A natural language in which to describe
such interactions is network science, which elegantly rep-
resents interconnected systems as sets of nodes linked by
edges [2]. Intuitively, nodes often represent proteins, neu-
rons, subcortical nuclei, or large cortical areas, and edges
often represent either (i) structural links in the form of
chemical bonds, synapses, or white-matter tracts, or (ii)
functional links in the form of statistical relations be-
tween nodal activity time series. Generally, the resultant
network architecture can be fruitfully studied using tools
from graph theory to obtain mechanistic insights perti-
nent to cognition [3, 4], above and beyond those provided
In particular, several fundamental questions in neuro-
science are quintessentially network questions concern-
ing the physical relationships between functional units.
How does the physical structure of a circuit affect its
function? How does coordinated activity at small spa-
tial scales give rise to emergent phenomena at large spa-
tial scales? How might alterations in neurodevelopmental
processes lead to circuit malfunction in psychiatric disor-
ders? How might pathology spread through cortical and
subcortical tissue giving rise to the well-known clinical
presentations of neurological disease? These questions
collectively highlight the fact that the brain - and its
multiple networks of interacting units - is physically em-
bedded into a fixed three-dimensional enclosure. Natural
consequences of this embedding include diverse physical
drivers of early connection formation and physical con-
straints on the resultant adult network architecture. An
understanding of the system's constitution and basal dy-
namics therefore require not only approaches to quantify
and predict network topology, but also tools, theories,
and methods to quantify and predict network geometry
and its role in both enabling and constraining system
function.
In this review, we provide evidence to support the no-
tion that a consideration of the brain's physical embed-
ding will prove critical for a holistic understanding of neu-
ral circuit function. We focus our comments on the utility
of informing this consideration with emerging computa-
tional tools developed for the characterization of spatial
networks. Indeed, while network science was originally
developed in the context of systems devoid of clear spatial
characteristics [5], the field has steadily developed tools
and intuitions for spatially embedded network systems
[6]. In the light of these recent technical developments,
we begin by recounting observations from empirical stud-
ies addressing the question of how brain networks are
embedded into physical space. Next, we discuss the rele-
vance of this spatial embedding for an understanding of
network function and dysfunction. We complement these
empirical discussions with a more technical exposition on
the relevant tools, methods, and statistical approaches to
be considered when analyzing brain networks. Lastly, we
close by outlining a few important future directions in
methodological development and neuroscientific investi-
gation that would benefit from an honest appraisal of the
role of space in neural network architecture and dynam-
ics.
PHYSICAL CONSTRAINTS ON NETWORK
TOPOLOGY AND GEOMETRY
The processes and influences that guide the formation
of structural connections in neural systems are disparate
and varied [7, 8]. Evidence from genetics suggests that
neurons with similar functions as operationalized by sim-
ilar gene expression tend to have more similar connection
profiles than neurons with less similar functions [7, 9, 10].
Of course, it is important to note that some spatial sim-
ilarity of expression profiles is expected due to the in-
fluence of small scale spatial gradients in growth factors
over periods of development [7]. However, evidence from
the Allen Brain Atlas suggests that interareal connec-
tivity profiles in rodent brains are even more correlated
with gene co-expression than expected simply based on
such spatial relationships [9]. This heightened correlation
could be partially explained by observations in mathe-
matical modeling studies that neurons with similar in-
puts (and therefore potentially performing similar func-
tions) tend to have more similar connection profiles than
neurons with dissimilar inputs [11].
Yet, while genetic coding and functional utility each
play important roles, a key challenge lies in summariz-
ing the various constraints on connection formation in
a simple and intuitive theory that can guide future pre-
dictions. One particularly acclaimed candidate mecha-
nism for such a theory is that of physical constraints on
the development, maintenance, and use of connections.
Metabolism related to neural architecture and function
2
is costly, utilizing 20% of the body's energy, despite com-
prising only 2% of its volume [12, 13]. Even the devel-
opment of axons alone, comprising only a small portion
of cortical tissue, extorts a large material cost [7]. The
existence of these pervasive costs motivated early work
to postulate that wiring minimization is a fundamental
driver of connection formation. Consistent with this hy-
pothesis, the connectomes of multiple species are pre-
dominantly comprised of wires extending over markedly
short distances [13–19], and this observation holds across
different methods of data collection [16, 18, 19].
However, mounting evidence suggests that pressures
for wiring minimization may compete against pressures
for efficient communication [20–22]. Early evidence sup-
porting the role of efficient communication came from the
observation that one can fix the network architecture of
inter-areal projections in the macaque cortex and then
rearrange the location of areas in space to obtain a con-
figuration with significantly (32%) lower wiring cost than
that present in the real system [21]. A similar method
can be used to obtain a configuration of the C. elegans
neuronal connectome with 48% lower wiring cost than
that present in the real system [21].
Interestingly, the
connections whose length is decreased most also tend to
be those that shorten the characteristic path length –
one of many ways to quantify how efficiently a network
can communicate [21, 23]. Consistent observations have
been made in human white matter tractography [20], the
mouse inter-areal connectome [10], and dendritic arbors
[7, 14]. Notably, computational models that instantiate
both constraints on wiring and efficient communication
produce topologies more similar to the true topologies
than models that instantiate a constraint on wiring min-
imization alone [11, 24]. Moreover, models that allow for
changes in this tradeoff over developmental time peri-
ods better fit observed connectome growth patterns than
prior models [25].
It is precisely this balance between wiring minimiza-
tion and communication efficiency that is thought to pro-
duce the complex network topologies observed in neural
systems, along with markedly precise spatial embedding
[24, 26]. To better understand this spatially embedded
topology, it is useful to consider methods that can si-
multaneously (rather than independently) assess topol-
ogy and geometry. One such method that has proven
particularly useful in the study of neural systems from
mice to humans is Rentian scaling, which assesses the
efficiency of a network's spatial embedding [20, 27–30].
Originally developed in the context of computer circuits,
Rentian scaling describes a power-law scaling relation-
ship between the number of nodes in a volume and the
number of connections crossing the boundary of the vol-
ume [7, 20]. The existence of such a power law relation-
ship with an exponent known as Rent's exponent is con-
sistent with an efficient spatial embedding of a complex
topology [31, 32]. In turn that efficient spatial embedding
is thought to support a broad repertoire of functional
dynamics. For example, tracts that bridge disparate ar-
eas of cortex to increase communication efficiency despite
greater wiring cost, also critically add to the functional
diversity of the brain in a manner that is distinct from
that predicted by path length alone [33].
REFLECTIONS OF PHYSICAL CONSTRAINTS
IN LOCAL, MESOSCALE, AND GLOBAL
NETWORK TOPOLOGY
Across species, the brain consistently exhibits a set
of topological features at local, meso-, and global scales
that can be relatively simply explained by spatial wiring
rules [34–36]. At the local scale, multiple modalities have
been used to demonstrate that a key conserved topologi-
cal feature is the existence of hubs, or nodes of unexpect-
edly high degree [37, 38]. Such hubs emerge naturally
in computational models in which the location of nodes
are fixed in space, and edges between nodes are rewired
to miminize average wiring length and to maximize topo-
logical efficiency by minimizing the average shortest path
length (Box 1). Interestingly, the number and degree of
hubs varies systematically with the relative importance of
the two constraints [14, 24] (Fig. 1). When wiring min-
imization is not enforced, networks become star graphs
with a single giant hub [14, 24]; when both constraints
are balanced, networks contain several hubs of varying
degrees, consistent with the topology observed in brain
networks [24].
It is notable that such constraints can
be implemented within the natural processes of devel-
opment; for example, in adult C. elegans, hub neurons
have been tracked back to the earliest born neurons in
the embryo, which accumulate a large number of connec-
tions along the normative growth trajectory [26, 39].
At the mesoscale, a key conserved topological feature is
modularity, or the existence of internally dense and exter-
nally sparse communities of nodes [34, 40]. The strength
of modularity in a network is commonly quantified using
a modularity quality index (Box 1). In computational
models, this index obtained under pressures of wiring
minimization and communication efficiency (quantified
with path length) was more similar to that empirically
measured in the connectomes of the macaque and C. el-
egans than to that obtained under either constraint sep-
arately [8, 24]. Again it is notable that such constraints
can be implemented within the natural processes of de-
velopment; for example, in C. elegans, communities form
when many neurons are born in a similar temporal win-
dow, and therefore typically share a common progenitor
type, spatial location, and genetic profile [26, 41]. Spaces
between modules can form cavities or cycles, or intu-
itively holes in the network, that can be identified with
emerging tools from applied algebraic topology (Box 2)
[42]. The locations, prevalence, and weight structure of
these cycles differs markedly between geometric and ran-
dom networks [43, 44], with patterns of functional con-
nectivity among neurons exhibiting characteristics simi-
lar to those observed in spatially constrained geometric
3
networks [45].
It will be interesting in future to gain
a deeper understanding of the relations between cycles
and modules, and their emergence through the spatially
constrained processes of development.
At the global scale, a key conserved topological feature
is small-worldness, or the confluence of unexpectedly high
clustering and short path length (Box 1) [46]. Such an
architecture is thought to be particularly conducive to a
balance between local information processing within the
clusters, and global information transmission across the
topologically long distance connections [47]. Similar to
the existence of hubs, modules, and cavities, small-world
architecture in a network can naturally arise from spa-
tial constraints on wiring [48]. Intuitively, clusters tend
to form in spatially nearby regions in order to minimize
wiring cost, while long distance connections facilitating
efficient communication tend to form only occasionally
due to their elevated wiring cost [49].
In concert with
these empirical observations, computational models that
account for wiring economy produce networks with small-
world architecture reminiscent of that observed in real
neural systems [50]. Collectively, these studies demon-
strate the influence of parsimonious wiring rules on com-
plex network topology. Future work could be directed to
better understand the aspects of connectome topology
that remain unexplained and thus may arise from more
subtle rules [8].
RELEVANCE OF NETWORK GEOMETRY FOR
DYNAMICS AND COGNITION
Pressures for wiring minimization and communica-
tion efficiency can exist alongside developmental pro-
cesses that produce non-isotropically structured organs.
Such processes include migration, elongation, segrega-
tion, folding, and closure that accompany neurulation
resulting in the bilaterally symmetric nervous system
composed of proencephalon, mesencephalon, rhomben-
cephalon, and spinal cord as well as patterning across
multiple overlapping signaling gradients. It is intuitively
possible that such processes could also explain the ob-
served differences in the network topologies of different
sectors of the brain [51, 52], which can impinge on the
functions that those sectors are optimized to perform
(Box 3). Indeed, prior work has noted the co-existence
of complex structural topologies and spatial gradients of
specific function [53], although it has been difficult to
achieve a mechanistic understanding of exactly how the
two relate to one another. One particularly promising
recent line of investigation has proposed the existence of
a set of primary spatial gradients that explain variance
in large-scale connectivity [54, 55]. In both humans and
macaques, the primary axis of variance is bounded on
one end by the transmodal default mode system, and on
the other end by the unimodal sensory systems [55]. No-
tably, this gradient is tightly linked to the geometry of
the network, with the regions located at one end having
4
FIG. 1. Effect of wiring minimization and communication efficiency on network topology. Networks were generated
by modulating the balance between a constraint on wiring (in the figure referred to as a spatial cost) and a constraint on
information routing efficiency (in the figure referred to as a temporal cost). The parameter β, which ranges between 0 and 1,
tunes this balance by weighting spatial cost against temporal cost. When β = 0 only the spatial cost is considered, while when
β = 1 only the temporal cost is considered. (A) Examples of networks at different values of β when only the spatial constraint
exists (left), when only the temporal constraint exists (right), and when the two constraints are balanced (middle). Root nodes
are shown in green and all other nodes are shown in yellow. (B) Spatial costs (blue) and temporal costs (red) vary as a function
of β. This figure was adapted with permission from [14].
maximal spatial distances from the regions located at the
other end [55]. Additionally, the regions located at the
peaks of the transmodal gradient have substantial overlap
with structural hubs in human connectomes [4, 56]. As
discussed earlier, these hub could serve to optimally bal-
ance wiring cost and communication efficiency across the
connectome, in addition to explaining patterns of func-
tional connectivity. Put simply, such evidence supports
the notion that the cortex is fundamentally organized
along a dimension of function from concrete to abstract,
and that dimension manifests clearly in the network's
spatial embedding.
The diverse spatial location of cortical and subcortical
areas has important implications for the patterns of neu-
ral dynamics that one would expect to observe. Consider,
for example, the patterns of intrinsic activity noted con-
sistently across species, individuals, and imaging modal-
ities [55, 57–59]. While prior work has reported a con-
sistent architecture of correlations between regional time
series [59], the manner in which the across-region activity
pattern at one time point is related to the across-region
activity pattern at other time points is not well under-
stood. Recent work addressing this gap has posited the
existence of so-called lag threads, or spatial similarities
between whole-brain activity patterns at non-zero time
lags [60]. Unexplained by vasculature, the orthogonal
threads are thought to reflect slow, subthreshold changes
in the membrane potential of large neuronal populations
[61], a notion that is supported by subtle changes in
lag thread structure across sleep and during cognitively
demanding tasks [60, 62]. Notably, regions of the de-
fault mode (which coincide with peaks of the transmodal
gradient discussed above) participate in lag thread mo-
tifs, where changes in the activity in one region lead to
changes in the activity of another region more frequently
than expected in an appropriate statistical null model
AB β = 0 β = .8 β = 1Trade-o(cid:30) Parameter (β)Relative Temporal Cost (%)Relative Spatial Cost (%) [58, 59]. The existence of these dependencies is consistent
with time-invariant, possibly structural features of brain
organization producing highly reproducible patterns of
activity.
RELEVANCE OF NETWORK GEOMETRY FOR
DISEASE
The spatial architecture of brain networks not only
impacts our understanding of dynamics and cognition,
but also our understanding of neurological disease and
psychiatric disorders. Mounting evidence suggests that
many diseases and disorders of mental health can be
thought of fruitfully as network disorders, where the
anatomy and physiology of cross-regional communication
can go awry [63, 64]. Intuitively, spatial anisotropies of
developmental processes, spatial specificity of pathology,
and spatial inhomogeneities of drug targets that either
lead to or accompany these disorders can also explain al-
terations in the spatial characteristics of brain networks
[65]. In this section we briefly discuss this correspondence
in epilepsy, a particularly common neurological disease,
and in schizophrenia, a particularly devastating psychi-
atric disorder.
Despite a diverse pathophysiology and a renitent uni-
fying biological mechanism, epilepsy is characterized by
altered network dynamics in the form of seizures that
display spatially consistent patterns. For example, an
ictal period often begins with a marked spatial decor-
relation between distributed brain regions followed by a
period in which abnormally synchronized activity prop-
agates in consistent spatial patterns [66, 67].
In addi-
tion to broad patterns of spatial decorrelation, individ-
ual siezures also show stereotyped patterns of both spi-
ral waves and travelling waves of activity [66, 68–73]. In
silico studies have demonstrated that a simple adaptive
model of synaptically coupled and spatially embedded
excitatory neurons can reproduce many basic features of
these waveforms, including their speed and the size of
the wavefront [70]. Yet, it is worth noting that travel-
ling waves are not unique to epilepsy, but also occur in
healthy human and non-human primates where they are
thought to play a role in transporting task-relevant infor-
mation [71, 74–79]. However, marked differences in wave
propagation in healthy and epileptic cortical tissue sug-
gests that the precise spatial progression is important,
potentially supported by distinct underlying microstruc-
tures [80]. Finally, even interictal dynamics are altered
in epilepsy, as manifest by marked decreases in average
functional connectivity across the brain combined with
local increases in functional connectivity and efficiency
in default mode areas [81–83]. These connectivity pat-
terns have some utility in predicting seizure spread, but
the guiding principles leading to these changes and how
they relate to fine scale patterns of activity remains un-
clear [84].
While its pathophysiology is quite distinct from that
5
implicated in epilepsy, schizophrenia is also a condition
marked by severe network disturbances that have broad
ramifications for cognitive function [22, 85, 86]. Some
of these network alterations appear to selectively affect
connections of certain physical lengths, reflecting an al-
teration in the network's spatial embedding [87]. Specif-
ically, evidence suggests a reduced hierarchical struc-
ture and increased connection distance in the anatom-
ical connectivity of multimodal cortex in patients with
schizophrenia compared to healthy controls, indicative of
less efficient spatial wiring [85]. Moreover, in functional
brain networks, patients display longer high-weight con-
nections, decreased clustering, and increased topological
efficiency in comparison to healthy controls [87]. The
lack of strong, short distance functional connections is
in line with evidence from animal studies suggesting an
over-pruning of synapses in childhood onset schizophre-
nia [87]. Here, the intuitions gained from a consideration
of the network's spatial embedding offer important di-
rections for future work in linking non-invasive imaging
phenotypes with invasive biomarkers of neural dysfunc-
tion in disease.
STATISTICS, NULL MODELS, AND
GENERATIVE MODELS
Collectively,
In the previous sections, we outlined developmental
rules for efficient wiring and we discussed the reflections
of these rules in spatial patterns of healthy and diseased
brain dynamics.
the studies that we
have reviewed motivate the broader use and further
development of sophisticated and easily-implementable
tools for the analysis of a network's spatial embedding
[88]. Here we outline the current state of the field in
developing effective network statistics, network null
models, and generative network models that account for
spatial embedding.
Network Statistics. A simple way to examine net-
work architecture in the context of spatial embedding
is to incorporate the Euclidean distance of connections
into local, meso-scale, and global statistics [87, 89]. Ar-
guably the simplest local statistics that remain spatially
sensitive are moments of the distribution of edge lengths
in the network, including the mean, variance, skewness,
and kurtosis. One can also compute graph metrics that
have been extended to consider space, such as the physi-
cal network efficiency and the physical edge betweenness
[90]. Intuitively, both begin by defining the length of the
shortest physical (as opposed to topological) path along
network edges between any two nodes. The physical net-
work efficiency then takes the inverse of the harmonic
mean of this length, while the physical edge betweenness
provides the fraction of shortest physical paths between
all node pairs that traverse a given edge [91]. One could
also define a physical clustering coefficient in a similar
manner. Finally, one can assess the system for Rentian
6
FIG. 2. Spatial distribution of intrinsic neural activity. Principal gradients of functional connectivity calculated in the
structural connections of both humans and macaques. The first two principal gradients explained approximately 40% of the
observed variance. (A) (Left) A scatter plot of the first two principal gradients, with transmodal regions shown in red, visual
regions shown in blue, and sensorimotor regions shown in green. (Right) The same colors are used to show the distribution of
points visualized on a cortical surface. The pattern suggests the existence of a macroscale gradient of connectivity that reflects
the systematic integration of information across different sensory modalities. (B, Left) The minimum geodesic distance (mm)
between each point on the cortical surface and the positive peaks of the first principal gradient. The peaks are shown as white
circles. (B, Right) A scatter plot depicting the relationship between distance and location on the trans- to uni-modal gradient.
Put differently, transmodal regions with high values in the principal gradient are maximally distant from unimodal regions with
low values in the principal gradient. This figure was adapted with permission from [55].
scaling as described earlier, providing information on how
efficiently the complex network topology has been embed-
ded into the physical space [20, 27–30]. In the context
of neural systems, these spatially informed graph statis-
tics can be used to account for the physical nature of
information processing, propagation, and transmission.
Complementing local and global graph statistics is
an assessment of a network's community structure, a
mesoscale property frequently assessed by considering
the existence and strength of network modules [92].
From that community structure, one can determine the
spatial embedding of communities, for example by as-
sessing their laterality in bilaterally symmetric systems
such as the brain [93–95]. One of the most com-
mon ways to assess community structure is to maxi-
mize a modularity quality function, which identifies as-
sortative modules with dense within-module connectivity
and sparse between-module connectivity [96] (although
see [97] for methods to identify non-assortative commu-
nities). Mechanistically, this algorithm compares the
strength of observed connections between two nodes in a
community to that expected under a given a null model.
The most commonly used null model in this context is
the Newman-Girvan or configuration model, which pre-
serves the strength distribution of the network [96]. How-
ever, this null model operationalizes a purely topological
constraint – the strength distribution – and does not ac-
knowledge any spatial constraints that may exist in the
system. For this reason, many investigators across sci-
entific domains have begun developing alternative null
models that account for physical laws [98] or physical
contraints [99–101] on their system of interest.
it is worth con-
In the context of brain networks,
sidering three distinct null models
for modularity
maximization that incorporate information about the
physical space of the network's embedding.
First,
one can directly incorporate the wiring minimization
constraint observed in brain networks by defining a null
model with a probability of connection between two
nodes that decays exponentially as a function of distance
[99]. Using this model, one can detect different, and
more spatially distributed modules than those obtained
when one uses the configuration model [99]. Second, one
can employ gravity models [100], which account for the
number of connections expected given a certain distance
(typically a power law or inverse of distance), weighted
by the relative importance of each location (typically
a quantification of the population or size of a given
location) [100, 101]. Third, one can employ radiation
models designed to capture flow of information between
regions, by weighting distance functions by the flux
or flow of each location [101]. Of course, there exists
no single correct null model for community detection
that will suit every question in neuroscience. However,
we propose that many studies could test tighter, more
targeted hypotheses about community structure in brain
networks by using a null model that accounts for the
brain's spatial nature.
Network Null Models. When considering a network
representation of a neural system, one often computes
a statistical quantitity of interest and then compares
that quantity to that expected in a random network null
model. If the observed quantity is significantly greater
than or less than that expected, one concludes that the
network under study shows meaningful architecture of
potential relevance to the biology. Perhaps the most com-
mon random network null model is that which randomly
permutes the locations of edges in the network while pre-
serving the number of nodes, number of edges, and edge
7
ical network are compared to the statistics of each of
the generative models with the goal of inferring which
wiring rule was most likely to have produced the ob-
served architecture [11, 28, 111]. Evidence from such
studies suggests that spatially embedded models tend to
more accurately reproduce network measures of large-
scale neural systems than models that do not account
for space [28]. One particularly influential study consid-
ered 13 generative models that all incorporated a wiring
probability that increased with distance [111]. Consis-
tent with other work, the authors found that the model
that only included the wiring minimization constraint
was unable to recreate long distance connections of in-
dividual connectomes in humans [8, 11, 111]. Successive
generative models were then added that attempted to
recreate certain aspects of topology in addition to these
geometric constraints [111]. The models that performed
the best were those that preserved homophilic attrac-
tion such that connections preferentially formed between
nodes that had similar connection profiles [111]. Contin-
ued advancement of generative network models, and in-
clusion of additional biological features such as bilateral
symmetry, serves as an exciting approach to test mech-
anistic predictions about how network topology forms in
spatially embedded neural systems.
FUTURE DIRECTIONS
The spatial embedding of the brain is an impor-
tant driver of its connectivity, which in turn directly
constrains neural function and by extension behavior.
Emerging tools from network science can be used to as-
sess this spatial architecture, thereby allowing investiga-
tors to test more specific hypotheses about brain network
structure and dynamics. While we envision that the use
of these tools will significantly expand our understand-
ing, it is also important to acknowledge their limitations.
In particular, the majority of currently available network
tools make the simplifying assumption that all of the re-
lations of interests are strictly dyadic in nature, and ex-
ist between inherently separable components [112].
In
truth, however, features that arise from spatial embed-
ding can also manifest as continuous or overlapping maps
and gradients [53], motivating the use of tools from ap-
plied algebraic topology that can account for non-dyadic
interactions (Box 2). As the field moves forward, we
envision existing and yet-to-be-developed tools for char-
acterizing the spatial embedding of brain networks will
prove critical for our understanding of network processes
underlying cognition, and alterations to those processes
accompanying disease.
weight distribution. However, one may also be inter-
ested to determine whether observed statistics are dif-
ferent from what one would expect simply from the spa-
tial embedding or wiring rules of the network [102–104].
To address these questions, one can rewire the observed
network by conditionally swapping two links if the swap
preserves the mean wiring length of the network [103].
By pairing this model with a reduced null model in which
connections are only swapped if they reduce connnection
length, one can assess the role of long distance connec-
tions in the network, which will be preserved in the spa-
tial null but not preserved in the reduced null [103]. In
addition to preserving the mean wiring length, one might
also wish to preserve the full edge length distribution
by, for example, (1) fitting a function to the relation-
ship between the mean and variance of edge weights and
their distances, (2) removing the effect of that relation-
ship from the data, (3) randomly rewiring the network,
and (4) adding the effect back into the rewired network
[102].
To complement insights obtained from edge swapping
algorithms, one can also construct null model networks
by stipulating a wiring rule a priori while fixing the
locations of nodes within the embedded system.
In
this vein, studies have fruitfully used null models based
on minimum spanning tree and greedy triangulation
methods [105–107]. A minimum spanning tree is a graph
that connects all of the nodes in a network such that
the sum of the total edge weights is minimal. To extend
this notion to spatial networks, one can preserve the
true geographic locations of all nodes in the empirical
network and compute the minimum spanning tree on
the matrix of Euclidean distances between all node pairs
[91]. Representing the opposite extreme is the greedy
triangulation model, which is particularly relevant for
the study of empirical networks that are planar (lying
along a surface) as opposed to non-planar (lying within
a volume).
In the context of neural systems, planar
or planar-like networks are observed in vasculature,
and in thinned models of cortex that either consider a
single lamina or a coarse-grained model collapsing across
laminae [108, 109]. To construct a greedy triangula-
tion null model, one can preserve the true geographic
locations of all nodes in the empirical network and
iteratively connect pairs of nodes in ascending order
of their distance while ensuring that no edges cross.
After constructing such minimally and maximally wired
null models, one can calculate relative measures of
wiring length, physical efficiency, physical betweenness
centrality, and community structure by normalizing the
empirical values by those expected in the two extremes
[91].
Generative Network Models. Generative network
models can be used to test hypotheses about the rules
guiding network growth, development, and evolution
[110]. Often, an ensemble of generative models are con-
structed, and summary graph statistics from the empir-
8
FIG. 3. Community structure obtained with spatially embedded and non-embedded null models. (A) A schematic
of a spatially-informed null model. The model expects fewer long distance connections than short distance connections. (B)
A schematic of the anticipated difference between the spatial null model and the Newman-Girvan (NG) null model; spatial
communities will have longer distance connections and not capture clustering of spatially nearby regions. (C) Differences in
the association matrices between the two models. Positive (negative) numbers indicate when two nodes were more likely to
be co-assigned to the same module under the spatial (NG) model. (D) The difference in the participation coefficient between
the spatial and NG models. The participation coefficient quantifies how diverse a node's connections are across modules. (E)
The difference in spatial spread of modules in both models; the spatially embedded model tends to produce modules that cover
larger distances. This figure was adapted with permission from [99].
BOX 1: SIMPLE NETWORK STATISTICS
In a network representation of the brain, units rang-
ing from neurons or neuronal ensembles to nuclei and
areas are represented as network nodes and unit-to-unit
interactions ranging from physical connections to statis-
tical similarities in activity time series are represented as
network edges. The architecture of the network can be
quantitatively characterized using statistics from graph
theory [113]. Here, we mathematically define some of the
topological statistics mentioned elsewhere in this paper.
• Degree and Strength. The degree of a node is the
number of connections it has.
In a binary graph
encoded in the adjacency matrix A, where two re-
gions i and j are connected if Aij = 1, and not
connected if Aij = 0, then the degree ki is defined
i,j∈N Aij, where N is the set of all nodes.
In a weighted graph, where Aij is the strength of
the connection between nodes i and j, then the
as ki =(cid:80)
strength si is defined as si =(cid:80)
i,j∈N Aij.
• Path Length and Network Efficiency. The term
path length frequently refers to the average length
of the shortest path in a network. The short-
est path between any two nodes is given by the
path requiring the fewest hops. The network ef-
ficiency is given by the inverse of the harmonic
mean of the shortest path length. To be pre-
cise, we can write the path length of node i as
Li = 1
, where di,j is the short-
n
est path length between two nodes and n is the
number of nodes.
(cid:80)
i∈N,j(cid:54)=i di,j
(cid:80)
n−1
i∈N
2ti
i∈N
(cid:80)
• Clustering Coefficient. The clustering coefficient
can be used to quantify the fraction of a node's
neighbors that are also neighbors with each other.
Specifically, the clustering coefficient of node i is
given by Ci = 1
ki(ki−1) , where ti is the
n
number of triangles around node i [46]. The clus-
tering coefficient of the network is the average clus-
tering coefficient of all of its nodes.
tions exist, the most common is Q = (cid:80)
• Modularity. While several modularity quality func-
ij[Aij −
γPij]δ(cicj), where Q is the modularity quality in-
dex, Pij is the expected number of connections be-
tween node i and node j under a specified null
model, δ() is the Kroenecker delta, and ci indicates
the community assignment of node i. The tuning
parameter γ ranges from (0,∞) and can be used to
tune the average community size.
BOX 2: APPLIED ALGEBRAIC TOPOLOGY
While graph theory is a powerful and accessible frame-
work for analyzing complex networks, complementary in-
formation can be gained by using different mathematical
9
formalisms. Here, we describe an alternative approach to
studying structure in networks that relies on tools devel-
oped in the field of applied algebraic topology, specifically
persistent homology [115–117]. Persistent homology can
be used to study intrinsically mesoscale structures called
cycles and cliques [118]. Cliques are all-to-all connected
subsets of nodes in a network. The presence of many,
large cliques indicates many highly connected units are
present in the network [119]. Cycles are looped patterns
of cliques which may enclose a cavity, or topological void,
within the network. Cliques and cavities by definition
reside within a binary graph, however one can expand a
weighted network into a sequence of binary graphs via
iterative thresholding [45, 120]. Then using persistent
homology one can track the birth, persistence, and death
of cavities along this sequence which gives a wholistic
insight into the global network (Fig. Box 2, panel A).
In random graphs, the number of births and deaths
across thresholds follows a characteristic pattern [44]. At
high thresholds and low edge density, a few low dimen-
sional cavities exist, while at low thresholds and high
edge density, more high-dimensional cavities exist (Fig.
Box 2, panel B) [44, 121, 122]. Interestingly, geometric
graphs – which can be used to instantiate spatial con-
straints on the topology – show a markedly different dis-
tribution. There are many low dimensional cavities, and
fewer cavities with increasing dimension[43, 44, 121, 123]
(Fig. Box 2, panel C). This general pattern has been
recapitulated in functional networks constructed from
firing of hippocampal neurons,
indicating a geometric
rather than random nature to neuronal co-firing [45].
Furthermore, the persistent homology of human connec-
tomes [42] and rat microcircuits [124] is distinct from that
expected in a minimally wired null model. In humans,
the presence of widespread subcortical connections leads
to more cavities being born at high densities [42], while
rat microcircuits display more high dimensional cavities
in general [124]. Further investigation into how wiring
rules shape the topology of neural systems may shed light
on how the brain's spatial embedding shapes connectivity
across scales and species.
BOX 3: CONTROL THEORY
Network control theory provides a potentially power-
ful approach for modeling neural dynamics [125]. Hail-
ing from physics and engineering, network control theory
characterizes a complex system as composed of nodes
interconnected by edges, and then specifies a model of
network dynamics to determine how external input af-
fects the nodes' time-varying activity [126]. Most studies
of network control in neural systems stipulate a linear,
time-invariant model of dynamics:
x(t) = Ax(t) + Bu(t),
where x is some measure of brain state, A is a struc-
tural connectivity matrix, u is the input into the system
(exogenous stimulation, or endogenous input from other
brain regions), and B selects the control set, or regions
10
FIG. 4. Schematic of network measures. (A) An illustration of network space (topology) and physical space (geometry).
The network is embedded into a physical space, indicated by the x- and y-axes. The topological and physical distances between
the nodes are not necessarily related. (B) The network representation enables the calculation of local, mesoscale, and global
features to describe the pattern of connections in topological space (as shown here) as well as the pattern of connections in
physical space (as we describe in the main text). This figure was adapted with permission from [1] and from [114].
to provide input to [127, 128]. Assuming this model of
dynamics, one can calculate the control energy required
to reach specific brain states, which can be used a state
dependent measure of the efficiency of control [129, 130].
Control theory can also posit control metrics that quan-
tify how efficiently a node would drive the brain to vari-
ous states. Two commonly used metrics are average con-
trollability and modal controllability [131]. When every
node is included in the control set, average controllabil-
ity is proportional to the average energy required to drive
the node to any state [132]. Conversely, modal controlla-
bility is high in nodes where a small input will result in
large perturbations to all eigenmodes of the system, and
is interpreted to be high in nodes that can easily drive
the brain to hard-to-reach states [133–135].
If these properties are important for helping the brain
transition between states, one would expect them not
to be randomly distributed across the cortex, but to
be clustered into spatially constrained, functionally rele-
vant systems. More specifically, one might expect func-
tional systems that drive the brain to many accessible
states, such as the default mode system, to have high
average controllability, while regions that drive the brain
to hard-to-reach, cognitively demanding states (executive
control areas) to have high modal controllability. Data
from healthy human adults supports these two hypothe-
ses [133]. Moreover, both average and modal control-
lability increase across development and are correlated
with cognitive performance generally [134] and executive
function specifically [136]. The manner in which network
control tracks individual differences reflects the fact that
the capacity for a network to enact control is dependent
upon its topology [137, 138]. Further efforts are needed
to distill exactly how spatial embedding and wiring con-
straints impinge on that control capacity, and how it is
altered in psychiatric disorders [132] and neurological dis-
ease [139].
BOX 4: OUTSTANDING QUESTIONS
• How do spatially constrained developmental pro-
cesses constrain the formation of cycles in brain
networks?
• What are the aspects of connectome topology that
remain unexplained by wiring minimization or com-
munication efficiency and thus may arise from more
subtle rules?
• How do the structural topologies that arise from
physical growth rules support functional gradients?
Local Topology (degree, clustering)Path, distancesModule detection, centralityBA 11
• To what extent can we link macro-scale structural
topology with small scale developmental rules?
• Can a deeper understanding of connectome devel-
opment be used to help identify new biomarkers for
network diseases?
• What additional rules can be incorporated into gen-
erative models of the brain to recapitulate its topol-
ogy?
• How can frontiers in network science help charac-
terize non-dyadic relationships in the brain?
GLOSSARY
• adjacency matrix: The adjacency matrix of a
graph is an N × N matrix, where N is the number
of nodes. Each element Aij of the matrix gives the
strength of the edge between nodes i and j.
• cycle:
In applied algebraic topology, a cycle is
an empty space (or lack of edges) in a graph sur-
rounded by all-to-all connected subgraphs of the
same dimension. The dimension here refers to the
number of nodes included in each all-to-all con-
nected subgraph.
• edge: From the perspective of graph theory, an
edge is a connection between nodes. From the per-
spective of neuroscience, an edge is a statistical de-
pendency (functional) or estimated physical con-
nection (structural) between nodes.
• geometry: The geometry of a network reflects fea-
tures of a graph in the context of physical space.
• hub: A central node in the network, typically hav-
ing many connections.
• node: From the perspective of graph theory, a
node is the unit where edges connect in the graph.
From the perspective of neuroscience, a node is a
brain region, neuron, or protein whose interactions
one wishes to understand.
• topology: The quantification of
features of a
graph in the context of space defined by the graph
itself, without respect to any physical embedding.
HIGHLIGHTS
FIG. 5. Applied algebraic topology. (A) An illustration
of thresholding a weighted network across different densities
(ρ). At ρ1, a cavity of dimension 1 is born (shown in yellow),
which then dies at ρ3. (B) The characteristic pattern of births
and deaths (called Betti curves) for cycles of dimension 1
(yellow), dimension 2 (red), and dimension 3 (blue) from a
random network. The cliques of each dimension are shown
near each corresponding Betti curve for reference. (C) The
same pattern, but for geometric networks. Different lines of
the same color indicate different dimensions of embedding.
This figure was adapted with permission from [45].
• What is the precise relationship between invari-
ant features of brain activity and the underlying
anatomical structure?
• How does the development of the connectome de-
termine spatial progression of activity through the
cortex in health and disease?
• The physical embedding of neural systems imposes
constraints on the possible patterns of connections
that form, and in turn on the possible repertoire
of functional motifs. Utilizing new tools from net-
work science that account for these constraints can
help researchers understand fundamentals of brain
dynamics in health and disease.
ABC 12
acknowledge support from the John D. and Catherine T.
MacArthur Foundation, the Alfred P. Sloan Foundation,
the ISI Foundation, the Paul Allen Foundation, the Army
Research Laboratory (W911NF-10-2-0022), the Army
Research Office (Bassett-W911NF-14-1-0679, Grafton-
W911NF-16-1-0474, DCIST-W911NF-17-2-0181),
the
the National Institute of
Office of Naval Research,
Mental Health (2-R01-DC-009209-11, R01-MH112847,
R01-MH107235, R21-M MH-106799),
the National
Institute of Child Health and Human Development
(1R01HD086888-01), National Institute of Neurological
Disorders and Stroke (R01 NS099348), and the National
Science Foundation (BCS-1441502, BCS-1430087, NSF
PHY-1554488 and BCS-1631550). The content is solely
the responsibility of the authors and does not necessarily
represent the official views of any of the funding agencies.
• Prominent, competing rules guiding the forma-
tion of brain networks include the minimization of
wiring cost, and the maximization of communica-
tion efficiency. These competing mechanisms lead
to networks with high local clustering with sparse
long distance connections.
• Recent work suggests that intrinsic functional con-
nectivity varies along dimensions that are tightly
linked to the spatial embedding of the brain and
the topological properties that arise in the pres-
ence of spatial constraints. Similarly, these topo-
logical properties show widespread changes in the
context of the network diseases such as epilepsy and
schizophrenia.
• A rich, and continually growing repertoire of statis-
tics, null models, and generative models exist to aid
researchers in testing focused hypotheses about the
role of physical embedding in observed neural phe-
nomena.
ACKNOWLEDGMENTS
We would like to thank Ann Sizemore for her help with
Box 2: Applied Algebraic Topology. D.S.B. and J.S.
[1] D. S. Bassett and O. Sporns, Nature Neuroscience 20,
[15] C. Cherniak, The Journal of Neuroscience 14, 2418
353 (2017), arXiv:0106096v1 [arXiv:cond-mat].
(1994).
[2] D. S. Bassett, P. Zurn, and J. I. Gold, Nature Reviews
Neuroscience In Press (2018).
[3] J. D. Medaglia, M. E. Lynall, and D. S. Bassett, J Cogn
Neurosci 27, 1471 (2015).
[4] S. E. Petersen and O. Sporns, Neuron 88, 207 (2015).
[5] C. Ducruet and L. Beauguitte, Networks and Spatial
Economics 14, 297 (2014).
[6] M. Barth´elemy, Physics Reports 499, 1 (2011).
[7] E. Bullmore and O. Sporns, Nature Reviews Neuro-
science 13, 336 (2012), arXiv:NIHMS150003.
[8] Y. Chen, S. Wang, C. C. Hilgetag, and C. Zhou, PLoS
computational biology 13, e1005776 (2017).
[16] C. Cherniak, Z. Mokhtarzada,
and R. Rodriguez-
Esteban, Evolution of Nervous Systems 1, 269 (2010).
[17] A. Raj and Y.-H. Chen, PLoS ONE 6 (2011), 10.1371/.
[18] M. Ria Ercsey-Ravasz, N. T. Markov, C. Lamy, D. C.
and
Van Essen, K. Knoblauch, Z. N. Toroczkai,
H. Kennedy, Neuron 80, 184 (2013).
[19] H. F. Song, H. Kennedy, and X.-J. Wang, Proceedings
of the National Academy of Sciences 111, 16580 (2014).
[20] D. S. Bassett, D. L. Greenfield, A. Meyer-Lindenberg,
D. R. Weinberger, S. W. Moore,
and E. T. Bull-
more, PLoS Comput Biol 6 (2010), 10.1371/jour-
nal.pcbi.1000748.
[9] L. French and P. Pavlidis, PLoS Comput Biol 7 (2011),
[21] M. Kaiser and C. C. Hilgetag, PLoS Computational Bi-
10.1371/journal.pcbi.1001049.
[10] M. Rubinov, R. J. F. Ypma, C. Watson, and E. T. Bull-
more, Proceedings of the National Academy of Sciences
112, 10032 (2015).
[11] P. E. Vertes, A. F. Alexander-Bloch, N. Gogtay, J. N.
Giedd, J. L. Rapoport, and E. T. Bullmore, Proceed-
ings of the National Academy of Sciences 109, 5868
(2012).
[12] S. B. Laughlin and T. J. Sejnowski, Science 108, 22
(2002).
[13] J. E. Niven and S. B. Laughlin, Journal of Experimental
Biology (2008), 10.1242/jeb.017574.
[14] J. M. L. Budd and Z. F. Kisv´arday, Frontiers in Neu-
roanatomy 6 (2012), 10.3389/fnana.2012.00042.
ology 2, 0805 (2006), arXiv:0607034 [q-bio].
[22] A. Zalesky, A. Fornito, G. F. Egan, C. Pantelis, and
E. T. Bullmore, Human Brain Mapping 33, 2535 (2012),
arXiv:arXiv:1011.1669v3.
[23] A.
B. Misic,
Avena-Koenigsberger,
and
O. Sporns, Nature Publishing Group 19 (2018),
10.1038/nrn.2017.149.
[24] Y. Chen, S. Wang, C. C. Hilgetag, and C. Zhou, PLoS
Comput Biol 9 (2013), 10.1371/journal.pcbi.1002937.
[25] V. Nicosia, P. E. Vertes, W. R. Schafer, V. Latora, and
E. T. Bullmore, Proceedings of the National Academy
of Sciences 110, 7880 (2013), arXiv:1304.7364.
[26] M. Kaiser, Trends in Cognitive Sciences 21, 703 (2017).
[27] M. M. Sperry, Q. K. Telesford, F. Klimm, and D. S.
Bassett, Journal of Complex Networks 5, 199 (2017).
13
[28] F. Klimm, D. S. Bassett, J. M. Carlson, P. J. Mucha,
and C. C. Hilgetag, PLoS Comput Biol 10 (2014),
10.1371/journal.pcbi.1003491.
[29] J. A. Pineda-Pardo, K. Martinez, A. B. Solana, J. A.
Hernandez-Tamames, R. Colom, and F. del Pozo, Brain
Topogr 28, 187 (2015).
[30] A. J. Sadovsky and J. N. MacLean, J Neurosci 34, 7769
(2014).
[31] P. Christie and D. Stroobandt, IEEE Trans. on VLSI
Systems 8, 639 (2000).
[53] S. Jbabdi, S. N. Sotiropoulos,
and T. E. Behrens,
Current Opinion in Neurobiology 23, 207 (2013),
arXiv:NIHMS150003.
[54] J. M. Huntenburg, P. L. Bazin, and D. S. Margulies,
Trends in Cognitive Sciences 22, 21 (2018).
[55] D. S. Margulies, S. S. Ghosh, A. Goulas, M. Falkiewicz,
J. M. Huntenburg, G. Langs, G. Bezgin, S. B. Eick-
hoff, F. X. Castellanos, M. Petrides, E. Jefferies, and
J. Smallwood, Proceedings of the National Academy of
Sciences 113, 12574 (2016), arXiv:arXiv:1408.1149.
[32] F. Alcalde Cuesta, P. Gonzalez Sequeiros,
and
[56] M. P. van den Heuvel and O. Sporns, Trends in Cogni-
A. Lozano Rojo, Sci Rep 7, 5378 (2017).
tive Sciences 17, 683 (2013).
[33] R.
and D.
S.
Betzel
of
F.
ceedings
ences
http://www.pnas.org/content/early/2018/05/07/1720186115.full.pdf.
Pro-
Sci-
10.1073/pnas.1720186115,
the National Academy
(2018),
Bassett,
of
[57] H. Lu, Q. Zou, H. Gu, M. E. Raichle, E. A. Stein, and
Y. Yang, Proceedings of the National Academy of Sci-
ences 109, 3979 (2012).
[58] A. Mitra, A. Z. Snyder, T. Blazey, and M. E. Raichle,
[34] J. A. Henderson and P. A. Robinson, Physical Review
Letters 107 (2011), 10.1103/PhysRevLett.107.018102.
[35] H. J. Park and K. Friston, Science 342 (2013),
10.1126/science.1238411.
[36] O. Sporns, D. R. Chialvo, M. Kaiser, and C. C. Hilge-
tag, Trends in Cognitive Sciences 8, 418 (2004).
[37] M. P. van den Heuvel, R. S. Kahn, J. Goni,
and
O. Sporns, Proceedings of the National Academy of Sci-
ences 109, 11372 (2012).
[38] J. Seidlitz, F. V´asa, M. Shinn, R. Romero-Garcia,
K. J. Whitaker, P. E. V´ertes, K. Wagstyl, P. Kirk-
patrick Reardon, L. Clasen, S. Liu, A. Messinger, D. A.
Leopold, P. Fonagy, R. J. Dolan, P. B. Jones, I. M.
Goodyer, A. Raznahan, and E. T. Bullmore, Neuron
97, 231 (2018).
[39] S. Varier and M. Kaiser, PLoS Comput Biol 7 (2011),
10.1371/journal.pcbi.1001044.
[40] C. C. Hilgetag and A. Goulas, Brain Struct Funct 221,
2361 (2016).
[41] R. Perin, T. K. Berger, and H. Markram, Proceedings
of the National Academy of Sciences 108, 5419 (2011),
arXiv:arXiv:1408.1149.
[42] A. E. Sizemore, C. Giusti, A. Kahn, J. M. Vettel, R. F.
Betzel, and D. S. Bassett, Journal of Computational
Neuroscience 44, 1 (2017), arXiv:1608.03520.
[43] M. Kahle, "Random Geometric Complexes," (2018).
[44] M. Kahle, Discrete Mathematics 309, 1658 (2009),
arXiv:0605536 [math].
[45] C. Giusti, E. Pastalkova, C. Curto,
and V. It-
skov, Proc Natl Acad Sci U S A 112, 13455 (2015),
arXiv:1502.06172.
[46] D. J. Watts and S. H. Strogatz, Nature 393, 440 (1998),
arXiv:0803.0939v1.
[47] C. T. Shih, O. Sporns, S. L. Yuan, T. S. Su, Y. J. Lin,
C. C. Chuang, T. Y. Wang, C. C. Lo, R. J. Greenspan,
and A. S. Chiang, Curr Biol 25, 1249 (2015).
[48] M. Kaiser and C. C. Hilgetag, Phys Rev E Stat Nonlin
Soft Matter Phys 69, 036103 (2004).
Proc Natl Acad Sci 17112, 2235 (2015).
[59] M. E. Raichle, Annual Review of Neuroscience 38, 433
(2015), arXiv:0402594v3 [arXiv:cond-mat].
[60] A. Mitra, A. Z. Snyder, C. D. Hacker,
and M. E.
Raichle, Journal of Neurophysiology 111, 2374 (2014).
[61] A. Mitra, A. Kraft, P. Wright, B. Acland, A. Z. Sny-
der, Z. Rosenthal, L. Czerniewski, A. Bauer, L. Snyder,
J. Culver, J.-M. Lee, and M. E. Raichle, Neuron (2018),
https://doi.org/10.1016/j.neuron.2018.03.015.
[62] A. Mitra, A. Z. Snyder, E. Tagliazucchi, H. Laufs, and
M. E. Raichle, eLife 4, 1 (2015).
[63] U. Braun, A. Schaefer, R. F. Betzel, H. Tost, A. Meyer-
Lindenberg, and D. S. Bassett, Neuron 97, 14 (2018).
[64] C. J. Stam, Nat Rev Neurosci 15, 683 (2014).
[65] D. S. Bassett, C. H. Xia, and T. D. Satterthwaite, Biol
Psychiatry Cogn Neurosci Neuroimaging S2451-9022,
30079 (2018).
[66] V. K. Jirsa, W. C. Stacey, P. P. Quilichini, A. I.
and C. Bernard, Brain 137, 2210 (2014),
Ivanov,
arXiv:arXiv:1011.1669v3.
[67] F. Wendling, F. Bartolomei, J. J. Bellanger, J. Bourien,
and P. Chauvel, Brain 126, 1449 (2003).
[68] A. C. Chamberlain, J. Viventi, J. A. Blanco, D. H. Kim,
J. A. Rogers, and B. Litt, Proceedings of the Annual
International Conference of the IEEE Engineering in
Medicine and Biology Society, EMBS , 761 (2011).
[69] J. Viventi, D. H. Kim, L. Vigeland, E. S. Frechette, J. A.
Blanco, Y. S. Kim, A. E. Avrin, V. R. Tiruvadi, S. W.
Hwang, A. C. Vanleer, D. F. Wulsin, K. Davis, C. E.
Gelber, L. Palmer, J. Van Der Spiegel, J. Wu, J. Xiao,
Y. Huang, D. Contreras, J. A. Rogers, and B. Litt,
Nature Neuroscience 14, 1599 (2011).
[70] L. R. Gonz´alez-Ram´ırez, O. J. Ahmed, S. S. Cash, C. E.
Wayne, and M. A. Kramer, PLoS Computational Biol-
ogy 11 (2015), 10.1371/journal.pcbi.1004065.
[71] K. A. Richardson, S. J. Schiff,
and B. J. Gluck-
man, Physical Review Letters 94 (2005), 10.1103/Phys-
RevLett.94.028103.
[49] D. S. Bassett and E. T. Bullmore, Neuroscientist 23,
[72] M. Ursino and G. E. La Cara, Journal of Theoretical
499 (2017), arXiv:1608.05665.
Biology 242, 171 (2006).
[50] A. Avena-Koenigsberger, J. Goni, R. Sole,
and
O. Sporns, Journal of The Royal Society Interface 12,
20140881 (2014).
[51] L. H. Scholtens, R. Schmidt, M. A. de Reus, and M. P.
van den Heuvel, J Neurosci 34, 12192 (2014).
[52] M. P. van den Heuvel, L. H. Scholtens, L. Feldman Bar-
rett, C. C. Hilgetag, and M. A. de Reus, J Neurosci 35,
13943 (2015).
[73] L. E. Martinet, G. Fiddyment, J. R. Madsen, E. N.
Eskandar, W. Truccolo, U. T. Eden, S. S. Cash,
and M. A. Kramer, Nature Communications 8 (2017),
10.1038/ncomms14896.
[74] J. M. Beggs and N. Timme, Frontiers in Physiology 3
JUN, 1 (2012).
[75] M. Massimini, Journal of Neuroscience 24, 6862 (2004).
[76] D. Rubino, K. A. Robbins, and N. G. Hatsopoulos,
Nature Neuroscience 9, 1549 (2006).
[77] K. Takahashi, S. Kim, T. P. Coleman, K. A. Brown,
A. J. Suminski, M. D. Best, and N. G. Hatsopoulos, Na-
ture Communications 6 (2015), 10.1038/ncomms8169.
[78] K. Takahashi, M. Saleh, R. D. Penn, and N. G. Hat-
sopoulos, Frontiers Human Neuroscience 5, 40 (2011).
[79] H. Zhang, A. J. Watrous, A. Patel, and J. Jacobs, Neu-
ron 98, 1269 (2018).
[80] A. Benucci, R. A. Frazor, and M. Carandini, Neuron
55, 103 (2007).
[81] L. Douw, M. N. DeSalvo, N. Tanaka, A. J. Cole, H. Liu,
C. Reinsberger, and S. M. Stufflebeam, Annals of Clin-
ical and Translational Neurology 2, 338 (2015).
[82] L. Bonilha, T. Nesland, G. U. Martz, J. E. Joseph, M. V.
Spampinato, J. C. Edwards, and A. Tabesh, J Neurol
Neurosurg Psychiatry 83, 903 (2012).
[83] M. N. DeSalvo, L. Douw, N. Tanaka, C. Reinsberger,
and S. M. Stufflebeam, Radiology (2014), 10.1148/ra-
diol.13131044.
[84] V. K. Jirsa, T. Proix, D. Perdikis, M. M. Woodman,
H. Wang, C. Bernard, C. Bınar, P. Chauvel, F. Bar-
tolomei, F. Bartolomei, M. Guye, J. Gonzalez-Martinez,
and P. Chauvel, NeuroImage 145, 377 (2017).
[85] D. S. Bassett, E. Bullmore, B. A. Verchinski, V. S. Mat-
tay, D. R. Weinberger, and A. Meyer-Lindenberg, Jour-
nal of Neuroscience 28, 9239 (2008).
[86] A. Griffa, P. S. Baumann, J. P. Thiran, and P. Hag-
mann, NeuroImage 80, 515 (2013).
[87] A. F. Alexander-Bloch, P. E. V´ertes, R. Stidd,
F. Lalonde, L. Clasen, J. Rapoport, J. Giedd, E. T. Bull-
more, and N. Gogtay, Cerebral Cortex 23, 127 (2013).
[88] M. Kaiser, NeuroImage 57, 892 (2011).
[89] J. M. Duarte-Carvajalino, N. Jahanshad, C. Lenglet,
K. L. McMahon, G. I. De Zubicaray, N. G. Martin, M. J.
Wright, P. M. Thompson, and G. Sapiro, NeuroImage
59, 3784 (2012), arXiv:NIHMS150003.
[90] J. Buhl, J. Gautrais, R. V. Sole, P. Kuntz, S. Valverde,
J. L. Deneubourg, and G. Theraulaz, The European
Physical Journal B 42, 123 (2004).
[91] L. Papadopoulos, P. Blinder, H. Ronellenfitsch,
F. Klimm, E. Katifori, D. Kleinfeld, and D. S. Bas-
sett, arxiv 1612, 08058 (2018).
14
[100] P. Expert, T. Evans, V. D. Blondel,
and R. Lam-
biotte, Proceedings of the National Academy of Sciences
(2010), 10.1073/pnas.1018962108, arXiv:1012.3409.
[101] M. Sarzynska, E. A. Leicht, G. Chowell, M. A. Porter,
D. Bassett, and M. Sarzynska, Journal of Complex Net-
works (2015), 10.1093/comnet/cnv027.
[102] J. A. Roberts, A. Perry, A. R. Lord, G. Roberts, P. B.
Mitchell, R. E. Smith, F. Calamante, and M. Breaks-
pear, NeuroImage 124, 379 (2016).
[103] D. Samu, A. K. Seth, and T. Nowotny, PLoS Computa-
tional Biology 10 (2014), 10.1371/journal.pcbi.1003557.
[104] M. Wiedermann, J. F. Donges, J. Kurths, and R. V.
Donner, Physical Review E 93 (2016), 10.1103/Phys-
RevE.93.042308, arXiv:1509.09293.
[105] X. Cui, J. Xiang, H. Guo, G. Yin, H. Zhang, F. Lan,
and J. Chen, Front Comput Neurosci 12, 31 (2018).
[106] T. W. P. Janssen, A. Hillebrand, A. Gouw, K. Gelade,
R. Van Mourik, A. Maras, and J. Oosterlaan, Clin Neu-
rophysiol 128, 2258 (2017).
[107] D. J. Smit, E. J. de Geus, M. Boersma, D. I. Boomsma,
and C. J. Stam, Brain Connect 6, 312 (2016).
[108] P. Blinder, P. S. Tsai, J. P. Kaufhold, P. M. Knutsen,
H. Suhl, and D. Kleinfeld, Nat Neurosci 16, 889 (2013).
and
[109] F. Schmid, P. S. Tsai, D. Kleinfeld, P. Jenny,
B. Weber, PLoS Comput Biol 13, e1005392 (2017).
[110] R. F. Betzel and D. S. Bassett, J R Soc Interface 14,
20170623 (2017).
[111] R. F. Betzel, A. Avena-Koenigsberger, J. Goni, Y. He,
M. A. de Reus, A. Griffa, P. E. V´ertes, B. Misic, J. P.
Thiran, P. Hagmann, M. van den Heuvel, X. N. Zuo,
E. T. Bullmore, and O. Sporns, NeuroImage 124, 1054
(2016), arXiv:1506.06795.
[112] C.
T.
Butts,
Science
325,
414
(2009),
arXiv:arXiv:1010.0725v1.
[113] M. Rubinov and O. Sporns, NeuroImage 52, 1059
(2010).
[114] J. O. Garcia, A. Ashourvan, S. F. Muldoon, J. M. Vettel,
and D. S. Bassett, Proceedings of the IEEE 106, 846
(2018).
[115] C. Giusti, R. Ghrist,
Neurosci 41, 1 (2016).
and D. S. Bassett, J Comput
[116] A. Zomorodian and G. Carlsson, Discrete and Compu-
tational Geometry 33, 249 (2005).
[92] S. Fortunato and D. Hric, Physics Reports 659, 1
[117] G. Carlsson, Bulletin of the American Mathematical So-
(2016).
[93] K. W. Doron, D. S. Bassett,
and M. S. Gazzaniga,
Proceedings of the National Academy of Sciences 109,
18661 (2012).
[94] L. R. Chai, M. G. Mattar, I. A. Blank, E. Fedorenko,
and D. S. Bassett, Cereb Cortex 26, 4148 (2016).
[95] X. He, D. S. Bassett, G. Chaitanya, M. R. Sperling,
ciety 46, 255 (2009), arXiv:arXiv:1312.6184v5.
[118] A. E. Sizemore, J. Phillips-Cremins, R. Ghrist, and
D. S. Bassett, arXiv 1806, 05167 (2018).
[119] P. Dotko, K. Hess, R. Levi, M. Nolte, M. Reimann,
M. Scolamiero, K. Turner, E. Muller, and H. Markram,
Frontiers in Computational Neuroscience 11, 48 (2016),
arXiv:1601.01580.
L. Kozlowski, and J. I. Tracy, Brain , 141 (2018).
[120] G. Petri, M. Scolamiero, I. Donato, F. Vaccarino, and
[96] M. E. J. Newman and M. Girvan, Physical Review
E (2003), 10.1103/PhysRevE.69.026113, arXiv:0308217
[cond-mat].
R. Lambiotte, PLoS ONE 8 (2013), 10.1371/.
[121] A. Sizemore, C. Giusti, and D. S. Bassett, Journal of
Complex Networks 5, 245 (2017), arXiv:1512.06457.
[97] R. F. Betzel, J. D. Medaglia, and D. S. Bassett, Nat
[122] D. Horak, S. Maleti,
and M. Rajkovi, Stat. Mech
Commun 9, 346 (2018).
[98] L. Papadopoulos, J. G. Puckett, K. E. Daniels, and
D. S. Bassett, Phys Rev E 94, 032908 (2016).
[99] R. F. Betzel, J. D. Medaglia, L. Papadopoulos,
G. Baum, R. Gur, R. Gur, D. Roalf, T. D. Satterth-
waite, and D. S. Bassett, Network Neuroscience (2016),
arXiv:1608.01161.
(2009), 10.1088/1742-5468/2009/03/P03034.
[123] O. Bobrowski and M. Kahle, Journal of Applied and
Computational Topology 1, 10.1007/s41468-017-0010-0.
[124] P. Dotko, K. Hess, R. Levi, M. Nolte, M. Reimann,
M. Scolamiero, K. Turner, E. Muller, and H. Markram,
Frontiers in Computational Neuroscience 1, 1 (2016),
1601.01580.
[125] S. J. Schiff, Neural Control Enginerring. The emerg-
ing intersection between control theory and neuroscience
(MIT Press, 2011).
J.
[126] Y. Y.
and A.
Liu,
J.
Slotine,
473,
L.
(2011),
15
[133] S. Gu, F. Pasqualetti, M. Cieslak, Q. K. Telesford, A. B.
Yu, A. E. Kahn, J. D. Medaglia, J. M. Vettel, M. B.
Miller, S. T. Grafton, and D. S. Bassett, Nature Com-
munications 6, 1 (2015), arXiv:1406.5197.
[134] E. Tang, C. Giusti, G. L. Baum, S. Gu, E. Pollock, A. E.
Kahn, D. R. Roalf, T. M. Moore, K. Ruparel, R. C. Gur,
R. E. Gur, T. D. Satterthwaite,
and D. S. Bassett,
Nature Communications 8 (2017), 10.1038/s41467-017-
01254-4, arXiv:1607.01010.
[135] E. Wu-Yan, R. F. Betzel, E. Tang, S. Gu, F. Pasqualetti,
and D. S. Bassett, Journal of Nonlinear Science (2017),
arXiv:1706.05117.
[136] E. J. Cornblath, E. Tang, G. L. Baum, T. M. Moore,
D. R. Roalf, R. C. Gur, R. E. Gur, F. Pasqualetti, T. D.
Satterthwaite, and D. S. Bassett, arXiv 1801, 04623
(2018).
[137] T. Menara, V. Katewa, D. S. Bassett,
and
Nature
Barab´asi,
arXiv:/www.nature.com/nature/journal/v473/n7346/abs/10.1038-
nature10011-unlocked.html#supplementary-
information [http:].
167
[127] A. E. Bryson, IEEE Control Systems 16, 26 (1996).
[128] G. Yan, P. E. Vertes, E. K. Towlson, Y. L. Chew, D. S.
Walker, W. R. Schafer, and A. L. Barabasi, Nature
550, 519 (2017).
[129] S. Gu, R. F. Betzel, M. G. Mattar, M. Cieslak, P. R.
Delio, S. T. Grafton, F. Pasqualetti, and D. S. Bassett,
Neuroimage 148, 305 (2017).
[130] R. F. Betzel, S. Gu, J. D. Medaglia, F. Pasqualetti, and
D. S. Bassett, Sci Rep 6, 30770 (2016).
[131] F. Pasqualetti, S. Zampieri, and F. Bullo, IEEE Trans-
F. Pasqualetti, arXiv 1706, 05120 (2018).
actions on Control of Network Systems 1, 40 (2014).
[132] J. Jeganathan, A. Perry, D. S. Bassett, G. Roberts, P. B.
Mitchell, and M. Breakspear, NeuroImage Clinical In
Press (2018).
[138] J. Z. Kim, J. M. Soffer, A. E. Kahn, J. M. Vettel,
F. Pasqualetti, and D. S. Bassett, Nat Phys 14, 91
(2018).
[139] P. N. Taylor, J. Thomas, N. Sinha, J. Dauwels,
M. Kaiser, T. Thesen, and J. Ruths, Front Neurosci
9, 202 (2015).
|
1803.03304 | 2 | 1803 | 2019-03-01T23:25:16 | A model of reward-modulated motor learning with parallelcortical and basal ganglia pathways | [
"q-bio.NC",
"cs.NE"
] | Many recent studies of the motor system are divided into two distinct approaches: Those that investigate how motor responses are encoded in cortical neurons' firing rate dynamics and those that study the learning rules by which mammals and songbirds develop reliable motor responses. Computationally, the first approach is encapsulated by reservoir computing models, which can learn intricate motor tasks and produce internal dynamics strikingly similar to those of motor cortical neurons, but rely on biologically unrealistic learning rules. The more realistic learning rules developed by the second approach are often derived for simplified, discrete tasks in contrast to the intricate dynamics that characterize real motor responses. We bridge these two approaches to develop a biologically realistic learning rule for reservoir computing. Our algorithm learns simulated motor tasks on which previous reservoir computing algorithms fail, and reproduces experimental findings including those that relate motor learning to Parkinson's disease and its treatment. | q-bio.NC | q-bio | A model of reward-modulated motor learning with parallel
cortical and basal ganglia pathways
Ryan Pyle1,∗ and Robert Rosenbaum1,2,∗
1: Department of Applied and Computational Mathematics and Statistics, University of Notre Dame, Notre Dame,
Indiana, USA.
2:Interdisciplinary Center for Network Science and Applications, University of Notre Dame, Notre Dame, Indiana,
USA.
* [email protected], [email protected]
Acknowledgments
This work was supported by National Science Foundation grants DMS-1517828, DMS-1654268,
and DBI-1707400. The authors thank Jonathan Rubin, Robert Turner, and Robert Mendez for
helpful discussions.
Abstract
Reservoir computing is a biologically inspired class of learning algorithms in which the intrinsic
dynamics of a recurrent neural network are mined to produce target time series. Most existing
reservoir computing algorithms rely on fully supervised learning rules, which require access to an
exact copy of the target response, greatly reducing the utility of the system. Reinforcement learning
rules have been developed for reservoir computing, but we find that they fail to converge on complex
motor tasks. Current theories of biological motor learning pose that early learning is controlled by
dopamine modulated plasticity in the basal ganglia that trains parallel cortical pathways through
unsupervised plasticity as a motor task becomes well-learned. We developed a novel learning algo-
rithm for reservoir computing that models the interaction between reinforcement and unsupervised
learning observed in experiments. This novel learning algorithm converges on simulated motor tasks
on which previous reservoir computing algorithms fail, and reproduces experimental findings that
relate Parkinson's disease and its treatments to motor learning. Hence, incorporating biological
theories of motor learning improves the effectiveness and biological relevance of reservoir computing
models.
9
1
0
2
r
a
M
1
]
.
C
N
o
i
b
-
q
[
2
v
4
0
3
3
0
.
3
0
8
1
:
v
i
X
r
a
1
Introduction
Even simple motor tasks require intricate, dynamical patterns of muscle activations. Understanding
how the brain generates this intricate motor output is a central problem in neuroscience that can
inform the development of brain-machine interfaces, treatments for motor diseases, and control
algorithms for robotics. Recent work is largely divided into addressing two distinct questions: How
are motor responses encoded and how are they learned?
From the coding perspective, it has been shown that the firing rates of cortical neurons ex-
hibit intricate dynamics that do not always code for specific stimulus or movement parame-
ters (Churchland et al., 2012; Russo et al., 2018). A prevailing theory poses that these firing rate
patterns are part of an underlying dynamical system that serves as a high-dimensional "reservoir"
of dynamics from which motor output signals are distilled (Shenoy, Sahani, & Churchland, 2013;
Sussillo, 2014). This notion can be formalized by reservoir computing models, in which a chaotic
or near chaotic, recurrent neural network serves as a reservoir of firing rate dynamics and synaptic
readout weights are trained to produce target time series (Maass, Natschlager, & Markram, 2002;
Jaeger & Haas, 2004; Sussillo & Abbott, 2009; Lukosevicius, Jaeger, & Schrauwen, 2012; Sussillo,
2014).
Reservoir computing models can learn to generate intricate dynamical responses and natu-
rally produce firing rate dynamics that are strikingly similar to those of cortical neurons (Sussillo,
Churchland, Kaufman, & Shenoy, 2013; Mante, Sussillo, Shenoy, & Newsome, 2013; Laje & Buono-
mano, 2013; Hennequin, Vogels, & Gerstner, 2014). However, most reservoir computing models
rely on biologically unrealistic, fully supervised learning rules. Specifically, they must learn from a
teacher signal that can already generate the target output. Many motor tasks are not learned in an
environment in which such a teacher signal is available. Instead, motor learning is at least partly
realized through reward-modulated, reinforcement learning rules (Izawa & Shadmehr, 2011).
A large body of studies are committed to understanding how reinforcement learning is im-
plemented in the motor systems of mammals and songbirds (Brainard & Doupe, 2002; Olveczky,
Andalman, & Fee, 2005; Kao, Doupe, & Brainard, 2005; Ashby, Turner, & Horvitz, 2010; Izawa &
Shadmehr, 2011; Fee, 2014). The basal ganglia and their homolog in songbirds play a critical role
in reinforcement learning of motor tasks through dopamine-modulated plasticity at corticostriatal
synapses. This notion inspired the development of a reward-modulated learning rule for reservoir
computing (Hoerzer, Legenstein, & Maass, 2014). However, we found that this learning rule fails
to converge on many simulated motor tasks.
We propose that the shortcomings of previous reservoir computing models can be resolved by
a closer inspection of the literature on biological motor learning. A large body of evidence across
multiple species supports a theory of learning in which dopamine-modulated plasticity in the basal
ganglia or its homologs is responsible for early learning and this pathway gradually trains a parallel
cortical pathway that takes over as tasks become well learned or "automatized" (Bottjer, Miesner,
& Arnold, 1984; Carelli, Wolske, & West, 1997; Brainard & Doupe, 2000; Pasupathy & Miller,
2005; Ashby, Ennis, & Spiering, 2007; Obeso et al., 2009; Andalman & Fee, 2009; Ashby et al.,
2010; Turner & Desmurget, 2010; Fee & Goldberg, 2011; Olveczky, Otchy, Goldberg, Aronov, &
Fee, 2011), although the biology is not settled (Kawai et al., 2015). This model of motor learning
has only been tested computationally in discrete choice tasks that do not capture the intricate,
dynamical nature of motor responses (Ashby et al., 2007).
Inspired by this theory of automaticity from parallel pathways, we derived a new architecture
2
Figure 1: Network diagrams for three reservoir computing algorithms. A) In FORCE
learning, readout weights are trained to match a target using a fully supervised error signal. Output
is fed back to the reservoir. B) RMHL is similar to FORCE, but learning is driven by a reward-
modulated plasticity and exploratory noise. C) SUPERTREX combines elements of FORCE and
RMHL. The exploratory pathway (green) is driven by noise, and acts similarly to RMHL. The
mastery pathway (purple) is analogous to FORCE, but uses the output of the exploratory pathway
in place of the fully supervised signal used by FORCE. The sum of both pathways provides the
total output.
and learning rule for reservoir computing. In this model, a reward-modulated pathway is respon-
sible for early learning and serves as a teacher signal for a parallel pathway that takes over the
production of motor output as the task becomes well-learned. This algorithm is applicable to a
large class of motor learning tasks to which fully supervised learning models cannot be applied and
it outperforms previous reward-modulated models. We additionally show that our model naturally
produces experimental and clinical findings that relate Parkinson's disease and its treatment to
motor learning (Ashby et al., 2007, 2010; Turner & Desmurget, 2010).
Results
We first review two previous learning rules for reservoir computing, then introduce a new, biologi-
cally inspired learning rule that combines their strengths.
FORCE learning
One of the most powerful and widely used reservoir computing algorithms is first-order reduced and
controlled error or FORCE (Sussillo & Abbott, 2009), which is able to rapidly and accurately learn
to generate complex, dynamical outputs. The standard architecture for FORCE is schematized
in Fig. 1A (FORCE variants exist, although the underlying principle is the same). The reservoir
is composed of a recurrently connected population of "rate-model" neurons. The output of the
reservoir is trained to produce a target time series by modifying a set of readout weights, and the
output affects the reservoir through a feedback loop.
The reservoir dynamics are defined by
= −x + Jr + Qz
τ
dx
dt
3
(1)
outputrecurrent netFORCESUPERTREXReward-modulated plasticityUnsupervised plasticityFully supervised plasticityrecurrent netRMHLexploratorynoiserecurrent netoutputoutputexploratorynoiseABC Here,
r = tanh(x) +
is a time-dependent vector representing the activity of units within the reservoir, τ is a time
constant, is a small noise term,
z = W r
is the reservoir output, J is the recurrent connectivity matrix, Q the feedback weights and W
the readout weights. A feedforward input term is also commonly included (Sussillo & Abbott,
2009), but we do not use one here. When the spectral radius of J is sufficiently large, the intrinsic
dynamics of r(t) become rich and chaotic (Sompolinsky, Crisanti, & Sommers, 1988). The goal
of FORCE is to utilize these rich dynamics by modifying readout weights, W , in such a way that
the output, z(t), matches a desired target function, f (t). A powerful and widely used learning
algorithm for FORCE, recursive least squares (RLS), is defined by (Sussillo & Abbott, 2009)
where
τw
dW
dt
= −erT P
e(t) = z(t) − f(t)
(2)
(3)
is the error vector and τw is the learning time scale. The matrix, P , is a running estimate of the
inverse of the correlation matrix of rates, r (see Materials and Methods).
FORCE excels at generating a target time series by harvesting reservoir dynamics, but is incom-
plete as a model of motor learning. As a fully supervised learning rule, FORCE must have access
to the correct output to determine its error (see the presence of f in Eq. (3)). Since the correct
output must already be generated to compute the error, FORCE can only learn target functions
that are already known explicitly and can be generated. Many motor learning tasks require the
generation of an unknown target using a lower-dimensional error signal (Izawa & Shadmehr, 2011).
We consider examples of such tasks below. A potential solution for these issues is provided by
appealing to biological motor learning, which is controlled at least in part by dopamine-modulated
reinforcement learning in the basal ganglia (Turner & Desmurget, 2010; Ashby et al., 2010; Izawa
& Shadmehr, 2011).
Reward modulated Hebbian learning
Reward-modulated Hebbian learning (RMHL) (Hoerzer et al., 2014) is a reinforcement learning
rule for reservoir computing in which reward is indicated by a one-dimensional error signal using
a plasticity rule inspired by dopamine-dependent Hebbian plasticity observed in the basal ganglia.
RMHL uses the same reservoir dynamics (Eq. (1)) and the same basic architecture as FORCE
(Fig. 1B), but the learning rule is fundamentally different.
The original RMHL algorithm (Hoerzer et al., 2014) used a binary error signal, despite poten-
tially poorer performance than other options, to demonstrate that the algorithm could learn with
minimal information. We implement a modified version of RMHL with an error signal,
e(z(t), t),
that can be any time-dependent, non-negative function of the output, z, which the algorithm will
seek to minimize. Equivalently, e is proportional to the negative of reward.
4
In contrast to the fully supervised error signal, e, used in FORCE learning, e is scalar (one-
dimensional) even when z is a higher-dimensional vector. Moreover, e can quantify any notion of
"error" or "cost" associated with the output, z, and does not assume that a target output is known
or even that there exists a unique target output. This allows RMHL to be applied to a large class
of learning tasks to which FORCE cannot be applied, as we demonstrate below.
To decrease error, RMHL makes random perturbations to the reservoir output as a form of
exploration. Specifically, the output, z, is given by
z = W r + Ψ(e)η
where η(t) is exploratory noise and Ψ is a sublinear function that serves to damp runaway oscilla-
tions during learning. The learning rule is then given by
τw
dW
dt
= Φ(e)zrT
(4)
where x denotes a high-pass filtered version of x, which represents recent changes in x and Φ is a
sublinear function that controls when to update the weights. We assume that Ψ is an increasing
function and Φ an odd function with Ψ(0) = Φ(0) = 0. This assures that exploration and learning
are effectively quenched when the error is consistently near zero. Intuitively, the learning rule can be
understood as follows: If a random perturbation from η has recently decreased e, this perturbation
is then incorporated into W .
SUPERTREX: A new learning algorithm for reservoir computing
Unfortunately, on many tasks, the weights trained by RMHL fail to converge to an accurate solution,
as we show below. RMHL models dopamine-modulated learning in the basal ganglia, but does
not account for experimental evidence for the eventual independence of well-learned tasks on the
activity of the basal ganglia: It has been proposed that the basal ganglia are responsible for early
learning, but train a parallel cortical pathway that gradually takes over the generation of output as
tasks become well-learned and "automatized" (Pasupathy & Miller, 2005; Ashby et al., 2007, 2010;
Turner & Desmurget, 2010; H´elie, Paul, & Ashby, 2012). This could explain why some neurons
in the basal ganglia are active during early learning and exploration, but inactive as the task
becomes well-learned (Carelli et al., 1997; Miyachi, Hikosaka, & Lu, 2002; Pasupathy & Miller,
2005; Poldrack et al., 2005; Ashby et al., 2007; Tang et al., 2009; Ashby et al., 2010; H´elie et
al., 2012). It could also explain why patients or animals with basal ganglia lesions can perform
previously learned tasks well, but suffer impairments at learning new tasks (Miyachi, Hikosaka,
Miyashita, K´ar´adi, & Rand, 1997; Obeso et al., 2009; Turner & Desmurget, 2010). This idea
is also consistent with many findings suggesting that the basal ganglia homolog in song birds is
responsible for early learning and exploration of novel song production, but not for the vocalization
of well-learned songs (M. S. Brainard, 2004; Kao et al., 2005; Aronov, Andalman, & Fee, 2008;
Andalman & Fee, 2009; Fee & Goldberg, 2011).
The FORCE and RMHL algorithms could be seen as analogous to the individual pathways in
this theory of motor learning: RMHL learns through reward-modulated exploration analogous to
the basal ganglia, while FORCE models cortical pathways that learn from the output produced by
the basal ganglia pathway. Inspired by this analogy, we next introduce a new algorithm, Supervised
5
Learning Trained by Rewarded Exploration (SUPERTREX), that combines the strengths RMHL
and FORCE to overcome the limitations of each.
The architecture of SUPERTREX (Fig. 1C) is different than the architectures of FORCE and
RMHL: There are now two distinct sets of weights from the reservoir to the outputs, and each
is trained with a separate learning rule. The "exploratory pathway" learns via an RMHL-like,
reinforcement learning algorithm, requiring only a one-dimensional metric of performance rather
than an explicit error signal. The exploratory pathway is roughly based off of the biological basal
ganglia pathway. The "mastery pathway" learns through a FORCE-like algorithm. The key idea is
that the activity of the exploratory pathway can act as a target for the mastery pathway to learn,
replacing the supervised error signal required by FORCE. Hence, SUPERTREX does not need the
explicit supervisory error signal that FORCE does. The mastery pathway is roughly based off of
the biological cortical pathway.
Importantly, the convergence issues we have found with RMHL are not problematic for SU-
PERTREX because weights in the RMHL-like exploratory pathway do not need to converge to a
correct solution because weights in the mastery pathway converge instead.
SUPERTREX uses the same reservoir dynamics (Eqs. (1)), but the outputs are determined by
z1 = W1r + Ψ(e)η
z2 = W2r
z = z1 + z2.
Here, z1 is the output from the exploratory pathway, z2 from the mastery pathway, and z is the
total output. The learning rules are defined by
τw1
τw2
dW1
dt
dW2
dt
= Φ(e)z rT
= (z − z2)rT P.
(5)
Intuitively, the first learning rule works like RMHL to quickly minimize the total error, as it uses
the error of the total output, z. However, it only controls the z1 component of z, resulting in
z1 + z2 ≈ f. Error between the z2 component and f is therefore just z1, which replaces the error in
the second learning rule since z− z2 = z1. As z2 approaches f, learning in the exploratory pathway
causes z1 to approach 0 in order to keep the total z correct.
Additionally, we added one extra component to the SUPERTREX algorithm. Learning transfer
from the exploratory pathway is soft thresholded based on total error - if error grows above this
point, the transfer rate is gradually reduced to 0. This means that transfer can only occur if
the total combined output of both pathways is correct.
In practice, this is true for the entire
learning period except for a small initial period while the exploratory pathway is finding a solution.
Performance without this addition was similar overall, but slightly slower.
Note that the learning rule for W1 is local in the sense that it only involves values of the
presynaptic and postsynaptic variables in addition to the error signal, e. The learning rule for W2
would be local if it were not for the computation of P , which is biologically unrealistic. However,
P can be replaced by the identity matrix to make the learning rules for SUPERTREX purely
local. This slows down learning, but the network can still learn to produce target outputs from a
one-dimensional error signal (see (Hoerzer et al., 2014) and the disrupted learning example below).
6
In summary, RMHL-like learning in the exploratory pathway uses a one-dimensional error signal,
e, to track the target while FORCE-like learning in the mastery pathway uses the exploratory
pathway as a teacher signal until it learns the output and takes over. This models current theories
of biological motor learning in which early learning is dominated by dopamine-dependent plasticity
in the basal ganglia, which gradually trains parallel cortical pathways as the task becomes well-
learned.
We next test SUPERTREX on three increasingly difficult motor tasks, comparing the perfor-
mance of SUPERTREX to that of FORCE and RMHL.
Task 1: Generating a known target output
We first consider a task in which the goal is to draw a parameterized curve of a butterfly by directly
controlling the coordinates of a pen (Fig. 2A). Specifically, the target is given by f (t) = (x(t), y(t))
where x(t) and y(t) parameterize the x- and y-coordinates of a pen that successfully traces out the
butterfly. The reservoir output, z(t), controls the coordinates of the pen, so the goal is to train the
weights so that z(t) closely matches the target, f (t).
The learning algorithms are first allowed to learn for ten repetitions of the task. As a diagnostic,
the error signals are not computed and the weights are frozen for a further five repetitions. This
provides a way to check the accuracy of the final solution, demonstrating whether or not the
algorithm has converged to an accurate solution. During this testing phase, feedback to the system
comes from the true solution (Sussillo & Abbott, 2009). Specifically, Qz is replaced by Qf in
Eq. (1). This avoids a drift in the phase of the solution that otherwise occurs when weights are
frozen (see below). Additionally, for SUPERTREX, the exploratory pathway was shut off during
these last five repetitions (z1 set to zero) to test how well the mastery pathway converged.
This simple task is well suited to FORCE, which requires a known target, f , in order to compute
the fully supervised error signal,
e = z − f .
FORCE was able to quickly find the correct solution to the task and maintained the correct result
even after weights were frozen (Fig. 2B,C). Another measure of convergence is the activity of weight
matrix, W , which quickly converged then stabilized (Fig. 2Cii, bottom). Error also remained low
after learning was disabled (Fig. 2B). In summary, as expected, FORCE learned this task quickly
and accurately.
To apply RMHL and SUPERTREX to this task, we set
e = (cid:107)z − f(cid:107)2
where (cid:107) · (cid:107) is the Euclidean norm, i.e. the distance of the pen from its target. This error contains
strictly less information than the fully supervised error used by FORCE. RMHL performed well
during learning, but the performance after weights were frozen (Fig. 2B,D) along with the cyclical
changes in (cid:107)W(cid:107) during learning (Fig. 2Dii) demonstrate that the RMHL algorithm never actually
converged. Instead, RMHL relied on rapid changes in W to mimic the correct output at each time
point without truly learning it. Even when the number of learning trials was dramatically increased
(to 100, not shown), RMHL's W continually oscillated rather than converging.
SUPERTREX performed well on this task. During learning SUPERTREX performed slightly
worse than FORCE and similar to RMHL (Fig. 2B,E). Unlike RMHL, though, SUPERTREX
7
Figure 2: Performance of three learning algorithms on Task 1. A) Task 1 is to draw a
butterfly curve by directly controlling the x- and y-coordinates of a pen. Specifically, the outputs
of the reservoir is z(t) = (x(t), y(t)) where x(t) and y(t) are the Cartesian coordinates of the pen.
B) Euclidean distance of pen from target for FORCE (orange), RMHL (green), and SUPERTREX
(purple). Learning was halted by freezing weights and exploration after ten periods, so the remain-
ing five periods represent a testing phase. Ci) Target butterfly curve (red) versus the butterfly
drawn by FORCE (blue) during the testing phase. Cii) Target (red) and actual (blue) outputs,
x(t) and y(t), and the norm of the weight matrix, (cid:107)(W W T )
2(cid:107)2, produced by FORCE. Di-ii) Same
as C, but for RMHL. Di-ii) Same as C, but for SUPERTREX and the norm of each matrix, W1
(exploratory; green) and W2 (mastery; purple) are plotted separately. Vertical gray bar indicates
the time at which weights were frozen. Note exploratory weight do change, but primarily at the
start and is hard to see over the full trial timescale. See figure 3 for more details.
1
8
CiATask 1LearningTestingCii1sFORCEDiDiiRMHLEiEiiSUPERTREXDistance from Targetx(t)y(t)Wx=z1y=z2OutputTarget0.0010.010.1x(t)y(t)Wx(t)y(t)WRMHLFORCESUPERTREXExploratoryMasteryB Figure 3: The dynamics of SUPERTREX with an abruptly changed target. A) Target
(red) and actual (blue) output. Same as Task 1, but the target was changed from a butterfly to
a circle after ten periods. B) Detail around the time of change. Same colors as A (red for target
and blue for total output), with addition of exploratory (green) and mastery (purple) components
of the total output. Note that exploratory + mastery = total output.
continued to track the target after learning was disabled, and performed similarly to FORCE
during that phase (Fig. 2B). This, combined with the apparent convergence of (cid:107)W(cid:107) during learning
(Fig. 2Eii, purple and green curves converge), indicates that the SUPERTREX algorithm did
converge - albeit more slowly than FORCE.
Interestingly, SUPERTREX produced less error during the testing phase than during learning
(Fig. 2B). This is because exploration introduces random errors during learning, but exploration was
turned off during testing so that output was produced only by the well-trained mastery pathway.
This is comparable to findings in song birds in which natural or artificial suppression of neural
activity in brain areas homologous to the basal ganglia reduce exploratory song variability and
vocal errors (Kao et al., 2005).
Looking at Fig. 2, it can be hard to tell whether the exploratory pathway is active, since the
weights do not seem to change. This is due to the large timescale of the trial compared to the
exploratory dominated phase, which only occurs as the algorithm is first adjusting to the task.
An interesting illustration of the exploration / mastery handoff in SUPERTREX is provided by
suddenly changing the target from a butterfly to a circle during learning (Fig. 3). The relative
contributions from the exploratory pathway and the mastery pathway show that the exploratory
pathway initially tracks the new target (Fig. 3B). Since the exploratory pathway is equivalent to
RMHL, we know from above that the pathway is only mimicking the output through rapid weight
changes. Over time, the mastery pathway learns from the activity of the exploratory pathway
and begins taking over the generation of the output. This "handoff" from the exploratory to the
mastery pathway produces a damped oscillation around the target (Fig. 3B).
Task 2: Generating an unknown target from a scalar error signal
Task 1 is a simple introductory task to compare the three learning algorithms, but it is also
unrealistic in some ways that play towards FORCE's strengths. Specifically, the task involves
producing an output, z, to match a known target, f , and error is computed in terms of the difference
between z and f . In many tasks, motor output has indirect effects on the environment and the
target and error are given in terms of these indirect effects. For example, consider a human or
robot performing a drawing task. Motor output does not control the position of the pen directly,
9
x(t)y(t)A1s1mstotal outputtargetchangetargetexploratorymasteryBx(t)y(t)changetarget Figure 4: Performance of RMHL and SUPERTREX on Task 2. A) Task 2 is to draw
the same butterfly (red) as in Task 1, but the reservoir output now controls the arm joint angles,
z1(t) = θ1(t) and z2(t) = θ2(t). Error is still computed in terms of pen coordinates. FORCE
is not applicable to this task. B) Euclidean distance of pen from target for RMHL (green) and
SUPERTREX (purple). Learning was halted by freezing weights after ten periods, so the remaining
five periods represent a testing phase. C-D) Same as Fig. 2D-E except that angles, θ1(t) and θ2(t)
are plotted in place of pen coordinates.
but instead controls the angles of the arm joints, which are nonlinearly related to pen position. On
the other hand, error might be evaluated in terms of the distance of the pen from its target. Task
2 models this scenario.
The goal in Task 2 is to draw the same butterfly from Task 1, parameterized by the same target
coordinates f (t) = (x(t), y(t)). However, the reservoir output controls the angles of two arm joints
(Fig. 4A),
z(t) = (θ1(t), θ2(t)).
We assume that the subject does not have access to the target angles that draw the butterfly.
Instead, they only have access to the target pen coordinates, f , and its distance to the actual pen
coordinates, which are related to the angles through a nonlinear function,
h(θ1, θ2) = (x, y).
FORCE cannot be applied directly to this task since the fully supervised error required for FORCE
would need to be computed in terms of target angles instead of target pen position.
RMHL and SUPERTREX can be applied to this task since they only require a signal that pro-
vides enough information to determine whether error recently increased or decreased. In particular,
10
ATask 1LearningTesting1sCiCiiRMHLDiDiiSUPERTREXDistance from TargetOutputTarget10.010.1WRMHLSUPERTREXExploratoryMasteryBz =θ11z =θ22θ1(t)θ2(t)Wθ1(t)θ2(t) this is accomplished by setting
e = (cid:107)h(z) − f(cid:107)2
where h(z) = (x, y) is the pen position.
Once again, the task is divided into 10 learning cycles and 5 test cycles, with learning algorithms
and the exploratory pathway of SUPERTREX disabled during the test cycles. Since the target
angles, z(t), are unknown, feedback during testing cannot be replaced by the target as was done
for Task 1. Instead, it is provided by the output from five previous periods.
RMHL performed poorly on this task. It eventually mimicked the target (Fig. 4B), but once
again failed to converge (Fig. 4B,C). SUPERTREX was able to track the target, and continue to
produce it even after weight changes ceased (Fig. 4B,D). Hence, the combination of FORCE-like
learning and RMHL-like learning implemented by SUPERTREX is able to learn a task that neither
FORCE nor RMHL can learn on their own.
Task 3: Learning and optimizing a task with multiple candidate solutions
While FORCE cannot be applied to Task 2 as it is currently defined, it could be applied if the
inverse of h were explicitly computed off-line to provide the target angles, (θ1, θ2) = h−1(f ), from
which to compute a fully supervised error signal. This approach assumes that the subject knows
the inverse of h and therefore does not easily extend to learning tasks in which h is difficult or
impossible to invert. We now consider a task in which the error is not an invertible function of the
motor output.
Specifically, we consider an arm with three joints (Fig. 5A) and a cost function, C(θ(cid:48)
2, θ(cid:48)
3),
that penalizes the movement of some joints more than others. Here, θ(cid:48)
j is the time-derivative of θj.
SUPERTREX can work with any penalty structure, making the choice arbitrary. Given that, we
decided to loosely model our arm on a real human arm, with the joints corresponding to shoulder,
elbow, and wrist. The penalties are larger for the angles controlling larger arm lengths, so the cost
is lowest for the wrist joint, θ3, and largest for the shoulder joint, θ1, based on the fact that you
are more likely to move your wrist than your entire shoulder and arm for a small reaching task -
this can also be seen as an energy conservation principle, with larger costs associated to the more
costly shoulder joint compared to the wrist joint.
1, θ(cid:48)
For this task, there are infinitely many candidate solutions that successfully draw the butterfly,
differing by the cost of joint movement. This turns our learning task into an optimization problem.
Using FORCE for this task does not make sense, as it would require prior, explicit knowledge of
the desired time series of joint angles. Essentially, it would require that the optimization problem
had already been solved offline. RMHL and SUPERTREX can be applied to this problem by setting
e = (cid:107)h(z) − f(cid:107)2 + C(θ(cid:48)
1, θ(cid:48)
2, θ(cid:48)
3).
In this context, RMHL and SUPERTREX work as greedy search algorithms that make local changes
to the angular output to reduce error and cost. Note, however, that the solution they find may not
be globally optimal.
We applied RMHL and SUPERTREX to this task using the same protocol for the learning and
testing phases that we used for Task 2. RMHL performed poorly on this task (Fig. 5B,C), which is
not surprising considering its poor performance on Task 2. SUPERTREX performed much better
11
Figure 5: Performance of RMHL and SUPERTREX on Task 3. A) Task 3 is to draw the
same butterfly (red) as in Tasks 1 and 2, but the reservoir output now controls three arm joint
angles, z1(t) = θ1(t), z2(t) = θ2(t), and z3(t) = θ3(t) with a different cost function associated
to moving each joint. Error is computed in terms of pen coordinates and cost of moving joints.
FORCE is not applicable to this task. B) Euclidean distance of pen from target (top) and cost
(bottom; C(θ(cid:48)
3)) for RMHL (green) and SUPERTREX (purple). C,D) Same as Di and Ei in
Fig. 2. E) Angular outputs and distance from target across two different SUPERTREX
trials. The overall solution found was similar, with a mirrored rotation in one joint
angle.
2, θ(cid:48)
1, θ(cid:48)
12
ATask 3LearningTesting1sCRMHLDSUPERTREXOutputTarget0.01RMHLSUPERTREXBz =θ11z =θ22z =θ3310-50.110.00010.0010.01DistanceCostθ10.11Distance1s0.01θ2θ3ETrial 1Trial 2 than RMHL: It was able to track the target, and continue to produce it even after weights were
frozen (Fig. 5B,D). Over multiple runs SUPERTREX will find different solutions, as seen in Fig
5 E). The solution found will primarily depend on the initial condition, but the randomness in
searching will also play a role. In this task, SUPERTREX tended to find similar solutions, except
for random mirroring of certain angles. In summary, SUPERTREX can solve motor learning tasks
in which there are multiple "correct" solutions with different costs.
Disrupted learning as a model of Parkinson's disease
The design of SUPERTREX was motivated in part by observations about the role of the basal
ganglia in motor learning and Parkinson's disease (PD). PD is caused by the death of dopamine
producing neurons in the basal ganglia, resulting in motor impairment. A common treatment for
PD is a lesion of basal ganglia output afferents. Such lesions alleviate PD symptoms and impair
performance on new learning tasks more than well-learned tasks (Obeso et al., 2009; Turner &
Desmurget, 2010). These and other findings have inspired a theory of motor learning in which the
basal ganglia are responsible for early learning, but not the performance of well-learned tasks and
associations (Turner & Desmurget, 2010; H´elie et al., 2012). SUPERTREX is consistent with this
theory if the exploratory pathway is interpreted as a basal ganglia pathway and the mastery pathway
the cortical pathway. To test this model, we next performed an experiment in SUPERTREX that
mimics the effects of PD and its treatment with basal ganglia lesion.
The hand-off of learning from the exploratory to mastery pathway occurs extremely quickly in
SUPERTREX due to the powerful, but biologically unrealistic RLS learning rule used in the mastery
pathway (see, e.g., Fig. 3). To make SUPERTREX more biologically plausible for this experiment,
we replaced the RLS learning rule with a least-mean-squares (LMS) rule by replacing P in Eq. (2)
with the identity matrix (Hoerzer et al., 2014). This modified rule is more realistic because it avoids
the complicated computation of the matrix, P , it makes the learning rules local, and it causes the
mastery pathway to learn more slowly, which slows the hand-off from the exploratory pathway.
We applied this modified SUPERTREX algorithm to Task 1. For 100 trials, learning proceeded
normally. SUPERTREX learned the target more slowly than in Fig. 2 and with a slight degradation
in performance due to the use of LMS instead of RLS learning in the mastery pathway (Fig. 6,
early and late learning). This phase models normal learning before the onset of PD. By the end
of this phase (Fig. 6, late learning), the task has become "well-learned" in the sense that output
is generated by the mastery pathway instead of the exploratory pathway. The system output
depending primarily on the mastery rather than exploratory pathway can be seen in Fig 6 C).
For the next five trials, we corrupted the error signal to model the effects of PD. Since e
models the error or cost of motor output, it is negatively related to dopamine release. Specifically,
e ∝ Dmax − D where D quantifies dopamine release and Dmax is the maximum possible value of D.
Hence, PD-induced dopamine depletion is modeled by artificially increasing e, which we achieve by
setting
e = (cid:107)f − z(cid:107)2
e = e + p
where p(t) increases over time. Here, p = 0 corresponds to a healthy subject and, as p increases,
SUPERTREX falsely evaluates more of its actions as being in error or costly. For our Parkinsonian
task, we chose a p(t) that linearly increased to .1 over the duration of the corrupted learning phase.
13
Figure 6: SUPERTREX with a corrupted error signal models Parkinson's disease and its
treatment. A) Euclidean distance of pen from its target for a modified version of SUPERTREX
on Task 1. Learning proceeded normally for 100 trials. First five trials (early learning) and last
five trials (late learning) are shown. The error signal was corrupted over the following five trials
(corrupted learning) and the exploratory pathway was lesioned for the last five trials (post-lesion).
B) Target (red) and actual (blue) outputs. C) Normed outputs from mastery pathway (purple)
and exploratory pathway (green). D) Target butterfly (red) versus drawn curve during each of the
plotted groups of five trials.
14
Distance from TargetABD0.010.10.6early learninglate learningcorrupted learningpost-lesiony(t)x(t)1sOutput Norm01C Even though the mastery pathway had taken over motor output before the error signal of the
exploratory pathway was corrupted, the perceived increase in error caused the exploratory path-
way to take over during the corrupted learning phase because the contribution of the exploratory
pathway increases with error. Although the actual disruption may seem small (See Fig 6 C), where
the exploratory activity is similar to that of early learning), the mismatch between actual error
and perceived error during the corrupted learning phase results in highly inaccurate motor output
(Fig. 6, corrupted learning phase) as activity leaves the learned manifold and is unable to recover.
These results model the motor impairments associated with PD. Indeed, PD symptoms are believed
to be caused, at least in part, by aberrant learning in the basal ganglia (Turner & Desmurget, 2010;
Ashby et al., 2010).
In the last five trials, we disabled the exploratory pathway, modeling basal ganglia lesion, and
the feedback term, Qz, was replaced by Qf in Eq. (1) (see below and Discussion). SUPERTREX
recovered nearly correct output during this last stage (Fig. 6, post-lesion phase) because the output
had been stored in the mastery pathway before learning in the exploratory pathway was corrupted.
As shown in Fig. 6, immediately before corruption began the mastery pathway was essentially
solely responsible for generating the correct output. After the Parkinsonian effect, the final output
is given solely by the mastery pathway as the malfunctioning exploratory pathway is lesioned. Thus,
any degradation in the drawn butterfly is due to harmful changes made to the mastery pathway
during the Parkinsonian effect. There are two main reasons why these harmful changes should be
small. One is that the exploratory pathway changes are only kept if they result in a decrease in error
even after taking into account the additional Parkinsonian error, or that the Parkinsonian error
term makes changes due to exploration less likely to be accepted. Additionally, for sufficiently large
errors the SUPERTREX component that controls transfer from exploration to mastery pathways
shuts down, limiting the degree to which harmful perturbations can be assimilated. Thus, post-
lesion performance will depend on the specific p(t) Parkinsonian effect used, along with the overall
duration of the Parkinsonian effect.
State Information Promotes Stability of Learned Output
During our previous examples comparing FORCE, RMHL, and SUPERTREX, the comparison
was made by allowing 10 trials of training the algorithm, and then with 5 trials of the learning
algorithm shut off (weights frozen) to see if the method had converged. During this testing phase,
feedback was modified.
In task 1 it was replaced with the correct output (the target), and for
tasks 2 and 3 it was replaced with the output from previous trials during learning, which nearly
matched the target due to the learning algorithm being active. This allowed us to check whether
an algorithm had converged, in the sense that there would be no further feedback and weight
changes. However, providing the correct answer as feedback, also known as teacher forcing, could
be considered cheating here. Teacher forcing essentially ignores stability of the solution and instead
only checks whether the system can correctly produce the next time step of the solution given a
perfect fit to the current time step.
In order to address this, we repeated task 1, but without
teacher forcing.
FORCE has previously been shown to perform well in the absence of teacher forcing (Sussillo
& Abbott, 2009; Abbott, Depasquale, & Memmesheimer, 2016), but it failed in our simulations
(Fig. 7A, solid orange). We suspected that this was due to the extra additive noise, , added
15
Figure 7: Including target information in feedback promotes stability without teacher
forcing. A) Euclidean distance of pen from its target for FORCE (orange), RHML (green), and
SUPERTREX (purple) on Task 1. Same as Fig. 2B except feedback during the testing phase was
not replaced with the true solution (teacher forcing), but is instead given by Qz exactly. B) Same as
A, but for Task 2 and without FORCE (since it cannot be applied to Task 2). C) Butterfly drawn
by SUPERTREX from the simulation in B. D-F) Same as A-C, except feedback was augmented
by the target, Q[x y f ].
during learning. Noise is not typically included in applications of FORCE, but reservoir learning
is known to be sensitive to noise and other perturbations (Vincent-Lamarre, Lajoie, & Thivierge,
2016; Sussillo, 2014; Miconi, 2017), which are ubiquitous in biological neuronal networks. Indeed,
FORCE performed better when this noise was removed (Fig. 7A, dashed orange). Noise is an
inherent part of RMHL and SUPERTREX, so they cannot be tested without it. Unsurprisingly,
RMHL and SUPERTREX also perform poorly without teacher forcing (Fig. 7A,B,C). In summary,
learning a noisy version of the target prevents all three algorithms from reproducing the target
post-learning in the absence of teacher forcing.
We resolve this issue by augmenting the feedback to include full information about the state
of the system, allowing the system to self-correct. Specifically, we concatenated the x- and y-
coordinates of the target pen position onto the feedback signal, replacing the Qz term in Eq. (1)
with Q[z f ], during both training and testing. Under this modified framework, we again tested
all three algorithms on Task 1 and tested RMHL and SUPERTREX on Task 2. For Task 1, this
16
Distance from Target10.010.1LearningTestingRMHLFORCESUPERTREX10.010.10.001LearningTestingTask 1Distance from Target10.1Task 20.0110.10.011sFeed Back Output OnlyFeed Back Output and TargetABDECF change is analogous to teacher forcing (since the target coordinates are the same as the target
reservoir output). For Task 2, it is distinct from teacher forcing because the feedback is in terms
of the Cartesian coordinates of the target, whereas the output must be in terms of arms' angles.
Hence, for Task 2, the system must learn to self correct: If z and f differ then the networks need
to learn how to generate the correct θ1 and θ2 to correct the error. This change greatly improved
accuracy of FORCE and SUPERTREX, but not RMHL (Fig. 7D,E,F).
Note that this change is not the same as just providing the correct answer as teacher forcing
does. Teacher forcing essentially "resets" the system to be correct after every time step by replacing
Qz with Qf , preventing drift. The augmented feedback instead provides sufficient information for
the system to be autonomously self-correcting and the feedback is provided as-is with no context. In
task 2, the algorithm does not have access to the solution it must produce (in terms of arm angles),
but only has access to the target pen coordinates, which are non-linearly related to arm angles. This
is akin to including a sensory feedback term, where the algorithms have sensory information about
the actual and target positions, but do not have explicit information on necessary joint movements
to make them overlap. Note that simply replacing the feedback from position to target will not
result in convergence, e.g. replacing Qz with Qf throughout training and testing does not work.
Both pieces of information together are required to build a stable system.
The extra feedback term can be simplified further by changing f into a simple phase variable,
which gives similar results as those shown in Fig. 7D,E,F (data not shown). Similar approaches
have been proposed previously (Vincent-Lamarre et al., 2016). These approaches can model the
presence of time-keeping neural populations. For example, in songbird, motor learning is believed
to be supported by a timekeeping signal from HVC, which is extensively used in models of songbird
learning (Doya & Sejnowski, 1995; Fiete, Fee, & Seung, 2007; Fee & Goldberg, 2011).
Reward modulated learning with velocity control
In all examples considered so far, the output of the reservoir controlled the position of a pen or
the angle of arm joints. In control problems, motor output controls velocity or acceleration (e.g.
applied force) of limbs or joints. From a naive perspective, SUPERTREX should still be able to
complete such a task - random perturbations still change error, and SUPERTREX can learn to
produce perturbations associated with lower error.
However, a more careful consideration reveals that SUPERTREX and RMHL applied directly
to control velocity would not be effective. To understand why, we first review and schematicize
how SUPERTREX and RMHL successfully learn Task 1 where the output controls the position of
the pen, then consider why they would not work when the reservoir output control the velocity of
the pen.
In Task 1, suppose the pen is displaced from its target (Fig. 8 top left) and an exploratory
perturbation is made to the reservoir output that successfully moves the pen closer to its target
(Fig. 8 bottom left). In this case, the change in error is negative (∆e ≈ e < 0), so the perturbation
is correctly rewarded (see Eqs. (4) and (5)).
Now consider Task 1 except that the reservoir output controls the velocity of the pen instead
of the position. Again, suppose the pen is displaced from its target and also suppose that it is
moving away from the target (Fig. 8 top middle). A beneficial exploratory perturbation changes
the velocity of the pen in the direction of the target (Fig. 8 bottom middle). However, if the
17
Figure 8: Velocity Control in SUPERTREX. A) When the position of a pen is controlled
and error is the distance of the pen from its target, a beneficial perturbation correctly results in
a decreased error. B) When the velocity of the pen is controlled and error is computed in the
same way, there are situations where a beneficial perturbation results in increased error. C) With
velocity control, replacing the error by the derivative of the distance causes beneficial perturbations
to correctly produce decreased error.
perturbation was not strong enough to change the direction of the pen, then the error (which is
measured as the distance of the pen from its target) will still have increased after the perturbation
(as in Fig. 8 bottom middle), so that ∆e ≈ e > 0 and this perturbation will be penalized instead
of rewarded (as again indicated by Eqs. (4) and (5)).
This problem is overcome by taking the derivative of the error, specifically defining e to be
the derivative of the distance between the pen and its target. When this change is made, then
a reservoir controlling pen velocity will be correctly rewarded for beneficial perturbations Fig. 8
right) and penalized for harmful perturbations.
To test these conclusions, we repeated Task 1 with SUPERTREX, except with output now
corresponding to velocity rather than position,
d[x, y]
dt
= z1 + z2
and we set [x(0) y(0)] = f(0). During the course of training this model, we discovered two other
adjustment were required. As our goal was to track a signal, rather than reach a target, adding a
penalty term based on velocity was helpful in order to prevent oscillations around our target, e.g.
e = ∆((cid:107)f − z(cid:107)2 + γdtz)
Unfortunately, we did not find a systematic way to determine γ. Instead, γ is chosen via iteration
in order to prevent over- or under-damped oscillatory behavior.
18
Δe<0e = P-TΔe>0PerturbationΔe<0VelocityPositionTargetPerturbationPositionTargetPerturbationtime=ttime=t+dtPositional ControlRegular ErrorVelocity ControlRegular ErrorVelocity ControlDerivative Errore = P-Te = ΔP-TVelocityPositionTarget222ABC Figure 9: Velocity Control in SUPERTREX. A) A schematic of the velocity control task,
which is identical to Task 1, except the velocity of the pen is controlled by the reservoir instead of
its position. B) Using the regular error (distance of pen from target) produces large errors, but
using the derivative of the distance produces smaller errors, especially during testing. C) Butterfly
drawn during the testing phase using regular error and D) derivative error.
19
CATask 1LearningTesting1sRegular ErrorDistance from TargetxyOutputTarget0Regular ErrorDerivative ErrorB0.10.2z 1(t)=x'(t)z 2(t)=y'(t)DDerivative Error Additionally, standard feedback Qz clearly does not provide enough information - if we don't
explicitly know our starting position, only knowing velocity does not help. Instead, we provided
full-state information Q[x y f] since we care more about our position than our velocity in terms of
feedback. This is also more realistic - it makes sense to modify z based on our position, rather than
velocity, and position is more likely to be available as sensory feedback. Making these changes, we
can compare SUPERTREX with error computed as the distance between the pen and its target
(Fig. 9, "regular error") and with error computed as the derivative of the distance between the
pen and its target (Fig. 9, "derivative error"). As predicted, SUPERTREX with velocity control
performs better when using the derivative of the distance as the error signal (Fig. 9, compare red
to purple in panel B, and compare panel C to D).
Discussion
We presented a novel, reward-modulated method of reservoir computing, SUPERTREX, that per-
forms nearly as well as fully supervised methods. This is desirable as there are a broad class or
problems where traditional supervised methods are not applicable, such as Tasks 2 and 3 that we
considered. Moreover, humans can learn motor tasks from reinforcement signals alone (Izawa &
Shadmehr, 2011). In place of a supervised error signal, SUPERTREX bootstraps from a dopamine-
like, scalar error signal to a full error signal using rewarded exploration. This serves as an approx-
imate target solution which is then transferred to a more traditional reservoir learning algorithm.
This transfer of learned behavior to a mastery pathway, along with continued rewarded exploration,
automatically creates a balanced system where the total output is correct, but the composition
shifts over time from exploration to mastery. SUPERTREX performed similarly to FORCE on
tasks where both were applicable, but also worked well on tasks where FORCE was not applicable.
SUPERTREX also outperformed RMHL, a previously developed reward-modulated algorithm, on
all tasks we considered.
Unlike RMHL and other reinforcement learning models, SUPERTREX models the complemen-
tary roles of cortical and basal ganglia pathways in motor learning. Under this interpretation,
dopamine concentrations play the role of the reward signal, and the basal ganglia is the site of the
RMHL-like, exploratory learning. Direct intra-cortical connections would then learn from Hebbian
plasticity in the mastery pathway. Consistent with this interpretation, SUPERTREX produces
inaccurate motor output when the reward signal is corrupted, modeling dopamine depletion in PD,
but recovers the generation of well-learned output when the exploratory pathway is removed, model-
ing basal ganglia lesions used to treat PD. Hence SUPERTREX provides a model for understanding
the role of motor learning in PD and its treatments.
As models of motor learning, reward-modulated algorithms like SUPERTREX and RMHL as-
sume no knowledge of the relationship between motor output and error. In contrast, fully supervised
algorithms like FORCE require perfect knowledge of this relationship. In reality, we learn through
some combination of supervisory and reward-modulated error signals (Izawa & Shadmehr, 2011). To
account for this, SUPERTREX could potentially be extended to incorporate both one-dimensional
reward and higher-dimensional sensory feedback.
The FORCE-like learning algorithm used for the mastery pathway of SUPERTREX is biolo-
gistically unrealstic in some ways. The presence of the matrix, P , causes the rule to be non-local.
However, we showed that SUPERTREX still works when P is removed to implement a local, LMS
20
learning rule (Fig. 6). Indeed, one can replace the mastery pathway with any supervised learning
rule. This could open the way for an implementation of SUPERTREX with spiking neural net-
works using existing supervised learning rules (Maass et al., 2002; Bourdoukan & Deneve, 2015;
Abbott et al., 2016; Pyle & Rosenbaum, 2017). In order to have a fully spiking-based version of
SUPERTREX, this would also require a spike-based reinforcement learning rule, most likely an
eligibility-trace based rule (Seung, 2003; Xie & Seung, 2004; Fiete & Seung, 2006; Miconi, 2017).
As with most other reservoir computing algorithms, SUPERTREX implements online learning
in which a local error signal is provided and used at every time step. This is partly by design
- SUPERTREX learns extremely (even unrealistically) quickly as weights are updated at a high
frequency. This learning is slowed by some extent by switching to the more realistic LMS learning
rule (as in Fig. 6). For some biological learning tasks, however, error signals are temporally sparse or
reflect temporally non-local information. Trial-based learning rules for reservoir computing (Fiete
& Seung, 2006; Miconi, 2017) are applicable in the presence of sparse or non-local rewards. At least
one of these algorithms learns very slowly, requiring thousands of trials (Miconi, 2017), which may
be an inevitable consequence of learning from sparse rewards. In reality, biological motor learning
likely makes use of both online and sparse feedback. An extension of SUPERTREX that accounts
for both types of feedback could be more versatile and realistic.
SUPERTREX is conceptually an extension of SPEED (Ashby et al., 2007), which has a similar
framework for categorization and other discrete tasks. SPEED learns to map arbitrary discrete
inputs to discrete outputs, such as in categorization tasks. While the architecture and learning rule
are similar to SUPERTREX, SPEED cannot produce continuous, dynamical output and requires
a separate pathway for each possible input-output pairing.
SUPERTREX could also be compared to a class of RNN algorithms that use a teacher network
to train the final output network. However, many of these networks use the activity of the teacher
network as a way to train the recurrence J of the output network; in SUPERTREX, there is only
one recurrent network (used for both outputs). These methods are often even more biologically
implausible - for example, the recent FULL-FORCE extension of FORCE (DePasquale, Cueva,
Rajan, Abbott, et al., 2018) feeds the target signal info the first, chaotic reservoir, and then uses the
activities of each reservoir unit in the teacher network as a target for training the second network,
drastically increasing the amount of supervision required.
SUPERTREX loses accuracy when learning is halted when feedback consists solely of the sys-
tem's output (Fig. 7A-C) due to the fact that it learns from a noisy estimate of the target. This
shortcoming can be overcome by augmenting the feedback with the target, allowing the system to
learn to self-correct noise-induced errors (Fig. 7D-F). FORCE is susceptible to the same instabili-
ties as SUPERTREX under the biologically realistic assumption of noise during learning (Fig. 7A),
but SUPERTREX can solve tasks that FORCE cannot (Figs. 4 and 5). RMHL is also susceptible
to the same instabilities and is applicable to the same tasks as SUPERTREX, but the instabil-
ities in RMHL are not resolved by including target information in the feedback as they are for
SUPERTREX (Fig. 7C,D). Hence, SUPERTREX is the only one of the three algorithms that can
be applied to reward-modulated learning tasks and achieves stability with target information in
the feedback. Stability in reward-modulated reservoir computing without target information in the
feedback term remains an open problem. This problem could potentially be solved by providing
external input in-phase with the target output. This could help the reservoir "keep time" by re-
aligning the reservoirs' state on each trial, allowing the system to self-correct its phase. A similar
21
approach was shown to improve robustness of FORCE to perturbations in previous work (Vincent-
Lamarre et al., 2016).
Interestingly, biology may have already solved this problem. Research by Toledo-Suarez, Duarte,
and Morrison (Toledo-Su´arez, Duarte, & Morrison, 2014) has found that the striatum may act as
a reservoir computer that processes state information. Rather than rely on raw inputs, the motor
learning system instead has access to pre-processed state information that is both simpler and
more relevant. In SUPERTREX, this could correspond to replacing our simple feedback of raw
state information Qz or Q[zf ] with Qs, where s is a pre-processed state information vector. s could
even come from another reservoir, designed to ensure s contains maximally relevant information to
the task at hand. This would be an interesting extension to SUPERTREX.
In summary, SUPERTREX is a new biologically inspired framework for reservoir computing
that is more realistic and more effective than its predecessors. Using a general error signal allows for
SUPERTREX to be used in places where a more powerful algorithm like FORCE cannot. The hand
off from exploration to mastery allows SUPERTREX to perform nearly as well as FORCE with
the generality of reward-modulated algorithms. Moreover, SUPERTREX offers a computational
formalization of widely supported theories of motor learning and reproduces several experimental
and clinical findings. Hence, this new framework opens the way for a truly two-way communication
between biological and computational theories of motor learning.
Materials and methods
Simulation and Reservoir Parameters
All simulations were performed using a forward Euler method, with dt = 0.2ms. Each task period
or "trial" was 104 ms long and all simulations except those in Fig. 6 had 15 trials. Fig. 6 had 110
trials.
The reservoir equation used in all algorithms was
= −x + Jr + Qz
τ
dx
dt
where r = tanh(x) + αη, η was uniformly drawn from [−1, 1] on every time step, τ =10, and
α = 2.5 × 10−2 during training and α = 0 during testing. Reservoir size was set to N = 1000
neurons, with connection probability p = 0.1. Connection strengths in J were normally distributed
with mean 0 and variance λ2/(pN ) with λ = 1.5. Feedback Q was dense, with weights uniformly
between −1 and 1. Initial readout weights for RMHL and SUPERTREX exploratory pathway, as
well as weights for FORCE and the SUPERTREX mastery pathway were initialized at 0. Initial
voltages were set uniformly between −0.5 and 0.5, while initial rates were the hyperbolic tangent
of initial voltages. Displayed outputs and errors were low pass filtered according to
τM SE
dM SE(t)
dt
dz(t)
dt
τbar
= −M SE(t) + M SE(t)
= −z(t) + z(t)
where τM SE = 1000, τbar = 10, and x represents a low pass filtered version of the variable x. The
22
plotted "distance from target" was computed as
the pen from its target.
√
M SE where M SE(t) is the squared distance of
FORCE
Reservoir output was z = W r and the learning rule is
τw
dW
dt
= −[z − f ]rT P
with τw = 0.02. The matrix P is a running estimate of the inverse of the correlation matrix of rates
r, initialized to
and updated according to
P (0) =
1
γ
I
where τP = dt, γ = 10 is a constant and I is the identity matrix. The matrix P is only updated
every 10 time steps in order to save on computing time.
τp
dP
dt
= − P r rT P
1 + rT P r
RMHL
For RMHL, outputs were given by
and the learning rule was
z = W r + Ψ(e)η
τw
dW
dt
= Φ(e)(z)rT
where τw = 0.02, η is uniformly distributed noise between [−1, 1], and the high-pass filtered version,
x, of variable x was computed as
τ
with τ = 1 used for all tasks and trials.
SUPERTREX
= −x + x
dx
dt
x = x − x.
Updates to P were identical to the method used in FORCE above. Relevant other changes are
z1 = W1r + Ψ(e)η
z2 = W2r
z = z1 + z2
for η uniformly drawn from [−1, 1]. For the learning algorithm,
τw
τw
dW1
dt
dW2
dt
= Φ(e)zrT
= −kz1rTP
23
with τw remaining 0.02 and constant learning rate k which varies per task. Finally, an extra
condition was imposed on updates to P , W2. Both updates were multiplied by (−0.5 ∗ tanh(5 ×
105 ∗ (e − (1.5 × 10−3))) + 0.5), which acts as a soft threshold around e = 1.5 × 10−3. Effectively,
for errors larger than this the mastery pathway would not activate. Performance was similar, but
slightly slower, without this thresholding.
Tasks
In all tasks, the target was to draw a butterfly, given by a polar curve x(t) = r(t) cos(t) and
y(t) = r(t) sin(t) where
r(t) =c[9 − sin(qt) + 2 sin(3qt)+
2 sin(5qt) − sin(7qt) + 3 cos(2qt) − 2 cos(4qt)]
and c = 1/ maxt[r(t)] is a normalizing constant. For a single repetition, t went from 0 to 104ms,
and q = 2π
104 scales the system such that qt goes from 0 to 2π over the duration.
In task 2, the task is instead to draw a butterfly by controlling two angles, representing radians
from y axis, and radians from the first joint. The arm is positioned at (0,-2), and each arm segment
has fixed length of 1.8. h(z) is therefore
(cid:20)
h(z) =
1.8 sin(z1π) + 1.8 sin((z1 + z2)π)
−2 + 1.8 cos(z1π) + 1.8 cos((z1 + z2)π)
(cid:21)
(cid:20)
In task 3, now there are three angles to control. The arm is positioned at (0,-2), and each arm
segment has fixed length. The first segment has length 1.8, the next 1.2, and the final .6. h(z) is
therefore
h(z) =
1.8 sin(z1π) + 1.2 sin((z1 + z2)π) + .6 sin((z1 + z2 + z3)π)
−2 + 1.8 cos(z1π) + 1.2 cos((z1 + z2)π) + .6 cos((z1 + z2 + z3)π)
√
In the first task, Ψ(x) = 0.025 × 4
√
10x and Φ(x) = −5 4
√
When testing for swapping targets, Ψ(x) = 0.1×(−5x)0.3 and Φ(x) = 2.5× 4
learning rate k was 0.5.
√
For the second task, SUPERTREX learning rate k was still 0.5, Ψ(x) = 0.01 × 5
x for both RMHL and SUPERTREX.
x. For SUPERTREX,
10x, and
(cid:21)
√
Φ(x) = 5 × 4
x.
√
For the third task, SUPERTREX learning rate was k = 0.9, Ψ(x) = .025 × 4
√
5 × 4
x. The error metric was changed slightly, to
10x, and Φ(x) =
e = h(z) − f2 + α z1 + β z2 + γ z3
for α = 0.1, β = 0.05, γ = 0. This implemented an additional cost for moving joints; highest for
the longest arm segment, and 0 for the smallest arm segment.
For the corrupted learning example, LMS learning was used, which is obtained by setting P = I.
The learning rate was changed to k = 0.003. Note that LMS learning rather than RMS learning
generally requires a much lower learning rate. Other parameter values were the same as in the first
task. The perturbation, p(t), increased linearly from 0 to 0.1 over the corrupted learning timeframe.
24
For the velocity controlled example in Fig. 8, more significant changes were needed. As detailed,
and
[x, y](0) = f (0)
d[x, y]
dt
= z1 + z2
as well as using full state feedback Q[x y f]. Learning rate k was .025, as smaller velocities were
needed relative to direct control of output. Velocity penalty γ = .3. Ψ and Φ were the same as in
task 1. Finally, we changed how we we calculated e. Rather than use a high pass filter as a crude
derivative estimator, we instead used a finite difference approximation e = e(t) − e(t − dt). Note
that, as described above, e(t) now refers to ∆d(t) = d(t)− d(t− dt) where d(t) = (cid:107)f − z(cid:107)2 + γdtz,
e.g. the squared euclidean distance between the position and target plus a penalty term. Thus,
our total update metric e = d(t) − 2d(t − dt) + d(t − 2dt), or the finite difference approximation to
the second derivative of our error metric, which is euclidean distance plus penalty.
Acknowledgments
This work was supported by National Science Foundation grants DMS-1517828, DMS-1654268, and
DBI-1707400. The authors thank Jonathan Rubin, Robert Turner, and Robert Mendez for helpful
comments.
References
Abbott, L. F., Depasquale, B., & Memmesheimer, R.-m. (2016). Building Functional Networks of
Spiking Model Neurons. Nat. Neurosci., 19 (3), 1 -- 16.
Andalman, A. S., & Fee, M. S. (2009). A basal ganglia-forebrain circuit in the songbird biases
motor output to avoid vocal errors. Proc. Natl. Acad. Sci. U. S. A., 106 (30), 12518 -- 23.
Aronov, D., Andalman, A., & Fee, M. (2008). A specialized forebrain circuit for vocal babbling in
the juvenile songbird. Science (80-. )., 320 , 630 -- 635.
Ashby, F. G., Ennis, J. M., & Spiering, B. J. (2007). A neurobiological theory of automaticity in
perceptual categorization. Psychol. Rev., 114 (3), 632 -- 56.
Ashby, F. G., Turner, B. O., & Horvitz, J. C. (2010). Cortical and basal ganglia contributions to
habit learning and automaticity. Trends Cog. Sci., 14 (5), 208 -- 215.
Bottjer, S. W., Miesner, E. a., & Arnold, a. P. (1984). Forebrain lesions disrupt development but
not maintenance of song in passerine birds. Science (80-. )., 224 (4651), 901 -- 903.
Bourdoukan, R., & Deneve, S. (2015). Enforcing balance allows local supervised learning in spiking
recurrent networks. In Adv. neur. in. (pp. 982 -- 990).
Brainard, & Doupe, A. (2000). Interruption of a basal gangliaforebrain circuit prevents plasticity
of learned vocalizations. Nature, 404 .
Brainard, & Doupe, A. (2002). What songbirds teach us about learning. Nature, 417 , 351 -- 358.
Brainard, M. S. (2004). Contributions of the anterior forebrain pathway to vocal plasticity. Ann.
NY Acad. Sci., 1016 (1), 377 -- 394.
25
Carelli, R. M., Wolske, M., & West, M. O. (1997). Loss of lever press-related firing of rat striatal
forelimb neurons after repeated sessions in a lever pressing task. J. Neurosci., 17 (5), 1804 --
1814.
Churchland, M. M., Cunningham, J. P., Kaufman, M. T., Foster, J. D., Nuyujukian, P., Ryu, S. I.,
& Shenoy, K. V. (2012). Neural population dynamics during reaching. Nature, 487 (7405),
51 -- 6.
DePasquale, B., Cueva, C. J., Rajan, K., Abbott, L., et al. (2018).
full-force: A target-based
method for training recurrent networks. PloS one, 13 (2), e0191527.
Doya, K., & Sejnowski, T. J. (1995). A novel reinforcement model of birdsong vocalization learning.
In Adv. neur. in. (pp. 101 -- 108).
Fee, M. S. (2014). The role of efference copy in striatal learning. Curr. Opin. Neurobiol., 25 ,
194 -- 200.
Fee, M. S., & Goldberg, J. H. (2011). A hypothesis for basal ganglia-dependent reinforcement
learning in the songbird. Neuroscience, 198 , 152 -- 170.
Fiete, I. R., Fee, M. S., & Seung, H. S. (2007). Model of birdsong learning based on gradient
estimation by dynamic perturbation of neural conductances. J. Neuropysiol., 98 (4), 2038 --
2057.
Fiete, I. R., & Seung, H. S. (2006). Gradient learning in spiking neural networks by dynamic
perturbation of conductances. Phys. Rev. Lett., 97 (4), 048104.
H´elie, S., Paul, E. J., & Ashby, F. G. (2012). A neurocomputational account of cognitive deficits
in Parkinson's disease. Neuropsychologia, 50 (9), 2290 -- 302.
Hennequin, G., Vogels, T. P., & Gerstner, W. (2014). Optimal control of transient dynamics in
balanced networks supports generation of complex movements. Neuron, 82 (6), 1394 -- 406.
Hoerzer, G. M., Legenstein, R., & Maass, W.
(2014). Emergence of complex computational
structures from chaotic neural networks through reward-modulated hebbian learning. Cereb.
Cort., 24 (3), 677 -- 690.
Izawa, J., & Shadmehr, R. (2011). Learning from sensory and reward prediction errors during
motor adaptation. PLoS Comput. Biol., 7 (3), 1 -- 11.
Jaeger, H., & Haas, H. (2004). Harnessing nonlinearity: predicting chaotic systems and saving
energy in wireless communication. Science, 304 (5667), 78 -- 80.
Kao, M. H., Doupe, A. J., & Brainard, M. S. (2005). Contributions of an avian basal ganglia --
forebrain circuit to real-time modulation of song. Nature, 433 (7026), 638 -- 643.
Kawai, R., Markman, T., Poddar, R., Ko, R., Fantana, A. L., Dhawale, A. K., . . . Olveczky, B. P.
(2015). Motor cortex is required for learning but not for executing a motor skill. Neuron,
86 (3), 800 -- 812.
Laje, R., & Buonomano, D. V. (2013). Robust timing and motor patterns by taming chaos in
recurrent neural networks. Nat. Neurosci., 16 (7), 925 -- 33.
Lukosevicius, M., Jaeger, H., & Schrauwen, B. (2012). Reservoir Computing Trends. KI - Kunstliche
Intelligenz , 26 (4), 365 -- 371.
Maass, W., Natschlager, T., & Markram, H. (2002). Real-time computing without stable states:
a new framework for neural computation based on perturbations. Neural Comput., 14 (11),
2531 -- 60.
Mante, V., Sussillo, D., Shenoy, K. V., & Newsome, W. T. (2013). Context-dependent computation
by recurrent dynamics in prefrontal cortex. Nature, 503 (7474), 78 -- 84.
26
Miconi, T. (2017). Biologically plausible learning in recurrent neural networks reproduces neural
dynamics observed during cognitive tasks. E-Life, 6 .
Miyachi, S., Hikosaka, O., & Lu, X. (2002). Differential activation of monkey striatal neurons in
the early and late stages of procedural learning. Exp Brain Res, 146 (1), 122 -- 126.
Miyachi, S., Hikosaka, O., Miyashita, K., K´ar´adi, Z., & Rand, M. K. (1997). Differential roles of
monkey striatum in learning of sequential hand movement. Exp. brain res., 115 (1), 1 -- 5.
Obeso, J. a., Jahanshahi, M., Alvarez, L., Macias, R., Pedroso, I., Wilkinson, L., . . . Rothwell,
J. C. (2009). What can man do without basal ganglia motor output? The effect of com-
bined unilateral subthalamotomy and pallidotomy in a patient with Parkinson's disease. Exp.
Neurol., 220 (2), 283 -- 92.
Olveczky, B. P., Andalman, A. S., & Fee, M. S. (2005). Vocal experimentation in the juvenile
songbird requires a basal ganglia circuit. PLoS Biol., 3 (5), e153.
Olveczky, B. P., Otchy, T. M., Goldberg, J. H., Aronov, D., & Fee, M. S. (2011). Changes in
the neural control of a complex motor sequence during learning. J. Neurophysiol., 106 (1),
386 -- 97.
Pasupathy, A., & Miller, E. K. (2005). Different time courses of learning-related activity in the
prefrontal cortex and striatum. Nature, 433 (7028), 873 -- 876.
Poldrack, R. A., Sabb, F. W., Foerde, K., Tom, S. M., Asarnow, R. F., Bookheimer, S. Y., &
Knowlton, B. J. (2005). The neural correlates of motor skill automaticity. J. Neurosci.,
25 (22), 5356 -- 5364.
Pyle, R., & Rosenbaum, R. (2017). Spatiotemporal dynamics and reliable computations in recurrent
spiking neural networks. Phys. Rev. Lett., 118 (1), 018103.
Russo, A. A., Bittner, S. R., Perkins, S. M., Seely, J. S., London, B. M., Lara, A. H., . . . others
(2018). Motor cortex embeds muscle-like commands in an untangled population response.
Neuron, 97 (4), 953 -- 966.
Seung, H. S. (2003). Learning in spiking neural networks by reinforcement of stochastic synaptic
transmission. Neuron, 40 (6), 1063 -- 1073.
Shenoy, K. V., Sahani, M., & Churchland, M. M. (2013). Cortical control of arm movements: a
dynamical systems perspective. Annu. Rev. Neurosci., 36 , 337 -- 59.
Sompolinsky, H., Crisanti, a., & Sommers, H. J. (1988). Chaos in random neural networks. Phys.
Rev. Lett., 61 (3), 259 -- 262.
Sussillo, D. (2014). Neural circuits as computational dynamical systems. Curr. Opin. Neurobiol.,
25 , 156 -- 163.
Sussillo, D., & Abbott, L. F. (2009). Generating coherent patterns of activity from chaotic neural
networks. Neuron, 63 (4), 544 -- 557.
Sussillo, D., Churchland, M. M., Kaufman, M. T., & Shenoy, K. V. (2013). A neural network that
finds naturalistic solutions for the production of muscle activity. Nat. Neurosci., 18 (7).
Tang, C. C., Root, D. H., Duke, D. C., Zhu, Y., Teixeria, K., Ma, S., . . . West, M. O. (2009).
Decreased Firing of Striatal Neurons Related to Licking during Acquisition and Overtraining
of a Licking Task. J. Neurosci., 29 (44), 13952 -- 13961.
Toledo-Su´arez, C., Duarte, R., & Morrison, A. (2014). Liquid computing on and off the edge of
chaos with a striatal microcircuit. Frontiers in computational neuroscience, 8 , 130.
Turner, R. S., & Desmurget, M. (2010). Basal ganglia contributions to motor control: a vigorous
tutor. Curr. Opin. Neurobiol., 20 (6), 704 -- 16.
27
Vincent-Lamarre, P., Lajoie, G., & Thivierge, J.-P.
(2016). Driving reservoir models with os-
cillations: a solution to the extreme structural sensitivity of chaotic networks. J. Comput.
Neurosci., 41 (3), 305 -- 322.
Xie, X., & Seung, H. S. (2004). Learning in neural networks by reinforcement of irregular spiking.
Physical Review E , 69 (4), 041909.
28
|
1809.01051 | 1 | 1809 | 2018-09-04T15:42:00 | Scaling Spike Detection and Sorting for Next Generation Electrophysiology | [
"q-bio.NC"
] | Reliable spike detection and sorting, the process of assigning each detected spike to its originating neuron, is an essential step in the analysis of extracellular electrical recordings from neurons. The volume and complexity of the data from recently developed large scale, high density microelectrode arrays and probes, which allow recording from thousands of channels simultaneously, substantially complicate this task conceptually and computationally. This chapter provides a summary and discussion of recently developed methods to tackle these challenges, and discuss the important aspect of algorithm validation, and assessment of detection and sorting quality. | q-bio.NC | q-bio |
Scaling Spike Detection and Sorting for Next Generation
Electrophysiology
Ma(cid:138)hias H. Henniga, Cole Hurwitza and Martino Sorbaroa,b
September 5, 2018
a Institute for Adaptive Neural Computation, School of Informatics, University of Edinburgh, Informatics Forum, 10
Crichton Street, Edinburgh EH8 9AB, Scotland, United Kingdom b School of Computer Science and Communication,
KTH Royal Institute of Technology, Lindstedtsvagen 5, 114 28 Stockholm, Sweden
Abstract Reliable spike detection and sorting, the process of assigning each detected spike to its
originating neuron, is an essential step in the analysis of extracellular electrical recordings from
neurons. (cid:140)e volume and complexity of the data from recently developed large scale, high density
microelectrode arrays and probes, which allow recording from thousands of channels simultane-
ously, substantially complicate this task conceptually and computationally. (cid:140)is chapter provides
a summary and discussion of recently developed methods to tackle these challenges, and discuss
the important aspect of algorithm validation, and assessment of detection and sorting quality.
1 Introduction
Extracellular electrical recording of neural activity is an essential tool in neuroscience. If an elec-
trode is placed su(cid:129)ciently close to a spiking neuron, the extracellular potential recorded o(cid:137)en
contains a clear, readily detectable signature of the action potential. As extracellular electrodes do
not interfere with neural function, such recordings provide an unbiased and precise record of the
functioning of intact neural circuits.
Recent progress in CMOS technology (Complementary metal-oxide semiconductor technology
for low power integrated circuits) has provided systems that allow recording from thousands of
closely spaced channels simultaneously with ever increasing density and sampling rates. With this
technology, it becomes possible to reliably monitor several thousand neurons simultaneously both
in vitro and in vivo (Eversmann et al., 2003; Berdondini et al., 2005; Frey et al., 2010; Ballini et al.,
2014; Muller et al., 2015; Yuan et al., 2016; Lopez et al., 2016; Jun et al., 2017b; Dimitriadis et al.,
2018). (cid:140)is is a signi(cid:128)cant advancement as it enables, for the (cid:128)rst time, the systematic investigation
of interactions between neurons in large circuits. Understanding these interactions will contribute
to learning more about how neural circuitry is altered by cellular changes in diseases, injury, and
during pharmacological interventions.
To appreciate the advantages of recording the activity of many neurons, it is important to em-
phasize that neural circuits are usually highly diverse and heterogeneous (Hrom´adka et al., 2008;
Buzs´aki and Mizuseki, 2014; Panas et al., 2015). Not only do they consist of di(cid:130)erent neuron types,
1
but even within groups of neurons of the same type, the (cid:128)ring rates may di(cid:130)er by orders of magni-
tude. (cid:140)is observation has been made consistently in vitro and in vivo, and it stands to reason that
this has biological relevance. Conventional technologies, which allow simultaneous recording of
a handful (rarely more than a hundred) of neurons, severely under-sample highly heterogeneous
populations. If the recorded neurons are not representative of the whole population, both exper-
imental accuracy and reproducibility between experiments will be negatively a(cid:130)ected. Moreover,
dense recording systems increase the fraction of neurons isolated in a local population, to a level
that was, so far, only accessible with calcium imaging.
A further advantage of recording many neurons at once is that it can be an e(cid:130)ective way
of probing neural excitability and connectivity, using functional interactions as a proxy measure
for the e(cid:130)ects of synaptic interactions.
In vitro assays are particularly suited for investigation
of functional interactions, as they can be augmented with stimulation, (cid:131)uorescent labeling and
targeted optogenetic stimulation (Zhang et al., 2009; Obien et al., 2015). A combination of dense
multielectrode arrays and imaging technologies could allow phenotyping at the level of single
cells, potentially in combination with further modalities such as gene expression pro(cid:128)ling. (cid:140)e
high yield of such approaches thus provides entirely new possibilities for systematic assessment
of the roles of di(cid:130)erent genotypes and of drug e(cid:130)ects.
(cid:140)e analysis of single neuron activity requires the correct assignment of each detected spike to
the originating neuron, a process called spike sorting. In this chapter, we will provide an overview
of the most frequently employed methods for the spike sorting for large-scale, dense multielec-
trode arrays. While many of the issues discussed will also apply to dense in vivo probes, the
focus is on in vitro arrays, because they typically provide a large surface area evenly covered with
recording channels, which is advantageous for spike sorting. A major additional challenge in in
vivo recordings is tissue movement, which causes the signals of neurons to dri(cid:137) over time. For
an excellent review of the challenges encountered in vivo, and of methodology for conventional
recording devices with fewer channels, the reader may consult Rey et al. (2015).
In the (cid:128)rst section, we will discuss in more detail the technical and practical issues that are
introduced when moving from conventional devices with tens of channels to larger, more dense
systems. Next, we will introduce the main components of modern spike sorting pipelines, and then
discuss each component and existing algorithms in detail. Finally, we will provide an overview of
approaches for validation of the quality of these algorithms.
2 Challenges for large-scale spike sorting
On both conventional and high density recording devices, electrodes will usually pick up the ac-
tivity of multiple neurons. While it is possible to directly analyse the multi-unit activity (MUA)
from each channel, spike sorting is required to resolve single-unit activity (SUA). Spike sorting
resembles the classic "cocktail party" problem: to isolate the voice of a single speaker in a crowd
of people. Since the recorded spike waveforms di(cid:130)er in shape and amplitude among neurons,
the resulting signal can be de-mixed using either dimensionality reduction paired with cluster-
ing or spike templates along with template matching. (cid:140)ese approaches have been successfully
employed on conventional devices with few, spatially well separated channels. On large-scale,
dense arrays, however, these traditional methods become more di(cid:129)cult both computationally and
algorithmically. Instead of (cid:128)nding a single voice in a crowd, the challenge is to isolate the voices
of thousands of speakers in a room equipped with thousands of microphones. Overcoming this
2
challenge is imperative as wrong assignments can severely bias subsequent analysis of neuronal
populations (Ventura and Gerkin, 2012).
Spike sorting is a tractable problem for conventional extracellular recordings as it is commonly
done for each recording channel separately. In this case, only a small number of neurons are ex-
pected to contribute to the signal on each channel, which allows the use of precise, but computa-
tionally more costly algorithms. Also, most existing algorithms for spike sorting still include an
element of manual intervention to adjust or improve sorting results. (cid:140)ese traditional algorithms
struggle when faced with large-scale, dense arrays.
On dense arrays, a single action-potential from a neuron is visible on multiple, nearby chan-
nels. As a result, spike sorting on single channels is no longer appropriate. Removing duplicate
events is feasible in principle, but becomes challenging when nearby neurons are (cid:128)ring with high
synchrony. Poor treatment of duplicate removal can lead to false exclusions of action potentials
or retention of multiple spikes from the same action potential.
Conventional spike sorting algorithms also struggle with the sheer volume of data large-scale
arrays produce. For instance, a recording from 4,096 channels with 18kHz sampling rate yields
about 140 megabytes per second, or over 8 gigabytes per minute. Simply reading this data volume
from hard disk into memory for analysis can be a severe bo(cid:138)leneck in any spike sorting pipeline.
In addition, the massive data volume prevents extensive manual curation of spike sorting results.
Highly automated pipelines with minimal need for intervention are needed to overcome these
challenges and to fully exploit the capabilities of dense arrays.
3 From raw data to single neuron activity
A typical spike sorting pipeline begins with the detection of candidate events followed by some
method of assigning these events to speci(cid:128)c neurons. (Lewicki, 1998; Rey et al., 2015). On large-
scale arrays, two approaches have emerged as particularly suitable. One method is based on cre-
ating spike templates and then performing template matching. (cid:140)e other method relies on feature
extraction and clustering, using both the spike shape and estimated location of the event. A sum-
mary of the steps required to obtain sorted spikes from raw data is shown in Figure 1. Each of
these steps is discussed in more detail below.
3.1 Spike detection
Spikes in the raw signal take the form of biphasic de(cid:131)ections from a baseline level. (cid:140)ey can be
found through detection of threshold crossings and by using additional shape parameters such
as the presence of a biphasic shape as acceptance criteria. As the noise levels may vary among
channels and over time, the threshold is usually de(cid:128)ned relative to the noise level, which is esti-
mated from portions of the raw signal that do not contain spikes. It is worth noting that signal
(cid:131)uctuations in extracellular data are typically highly non-Gaussian. As a result, a noise estimate
based on percentiles is more accurate and also easier to obtain, as opposed to computing the signal
variance (Fee et al., 1996; Muthmann et al., 2015).
(cid:140)e choice of the detection threshold determines which events are retained for further analysis.
Spikes from well-detected neurons are easily identi(cid:128)able, but de(cid:131)ections with amplitudes closer
to the background noise level are harder to isolate. Since there is typically no clear-cut separation
between spikes and noise, events detected close to the threshold may originate from neurons for
3
Figure 1: Schematic overview of existing spike sorting pipelines for high density microelectrode arrays. Following
detection, either neuron templates are formed based on the spatio-temporal event footprints and then used to detect
these units in a second pass, or current sources are estimated and clustered together with waveform features, or a mask
is created to restrict clustering to channels with a detectable signal. (cid:140)e output consists of a list of spike time stamps
for each identi(cid:128)ed neuron, which o(cid:137)en has to be corrected in a (cid:128)nal manual curation step.
which only an incomplete activity record can be obtained. (cid:140)e magnitude of electrical noise,
which can be estimated when recording from an empty array, is usually much smaller than the
magnitude of the (cid:131)uctuations recorded in the absence of clearly visible spiking activity (Muthmann
et al., 2015). (cid:140)is indicates that a large component of the recorded signal (cid:131)uctuations are due to
neural activity, such as neurons located further away from the electrode, or smaller events such as
currents during synaptic transmission. (cid:140)e analysis of recordings from the retina shows that even
very small detected signals may re(cid:131)ect activity that is typical of stimulus-evoked responses from
retinal ganglion cells, hence carries signatures of neural activity rather than noise (Figure 2).
As a result, the detection step signi(cid:128)cantly a(cid:130)ects the subsequent isolation of single neuron
activity. Choosing a high detection threshold is not an ideal solution as this will potentially leave
valid spikes undetected. In contrast, a low threshold guarantees reliable detection of neurons with
stronger signals, but also increases the fraction of false positives. As a good compromise, a strategy
can be adopted to detect events with a low threshold, and to subsequently discard unreliable units.
(cid:140)is can either be done a(cid:137)er detection, for instance by using a classi(cid:128)er trained on true spikes
and noise events obtained from channels not reporting neural activity (Hilgen et al., 2017), or by
removing sorted units with a small number of spikes or poor clustering metric scores post spike
sorting (Hill et al., 2011).
Recently, a new method for spike detection using a pre-trained neural network was introduced
by Lee et al. (2017). (cid:140)is method was shown to outperform conventional threshold-based methods
on simulated ground truth data, in particular by achieving a lower false positive rate. When run
on a modern GPU, a neural-network based method is also much faster. (cid:140)is is a very promising
avenue, although the considerations regarding detection thresholds outlined above still remain
4
Feature extractionPC1PC2ConstructTemplatesby ClusteringEstimatespike locationsMethod #2Clusterlow dim.embeddingsMethod #1TemplatematchFlood filland maskSpike Detection Manualcuration Figure 2: Even small events detected on a high density array contain signatures of neural activity. A, Density plot
of spatially binned spike counts, estimated from detected and spatially localised spikes using the method described by
Hilgen et al. (2017). Spikes were recorded from a light-stimulated mouse retina. Spike detection was performed with a
low threshold, hence false positives were registered in areas where no neural activity was recorded, such as the optic
disk on the centre le(cid:137). B, No clear separation between spikes and noise is seen for recorded amplitudes, or for average
amplitudes of units following spike sorting. C, Individual, randomly selected units with small (le(cid:137)) and large amplitudes
(right) both show signatures of light stimulation during presentation of full (cid:128)eld (cid:131)ashes. For each unit, the light-evoked
peri-stimulus time histogram (le(cid:137)), and examples and the average of spike waveforms (right) are shown. (cid:140)e recording
was contributed by Gerrit Hilgen and Evelyne Sernagor, University of Newcastle.
5
02550Channels0204060123log10(Spike count/bin)0510Spike amplitude (z-score)0.00.51.01.5DensitySpikesUnits024Time (s)050Rate (Hz)100V1ms024Time (s)ABClight onlight offlight onlight offSmall spikesLarge spikes relevant.
3.2 Dealing with duplicate spikes
Unlike in conventional recordings, on dense arrays, spikes are detectable on multiple channels.
(cid:140)ese duplicate spikes pose two signi(cid:128)cant problems for traditional spike sorting algorithms. First,
the amount of computation and memory used for processing each detected event increases with
the number of duplicates. Second, the rate of misclassi(cid:128)cation in spike sorting potentially increases
since each duplicate spike must be sorted into the same event.
To avoid the pitfalls associated with duplicate spikes, it is suggested to identify and remove du-
plicates during detection. One naive method for duplicate removal is to remove all but the largest
amplitude spike in a radius that encompasses the spatial footprint of the event. (cid:140)is method will
remove almost every duplicate event, but as the radius of duplicate removal increases, so does the
number of spikes removed that are not associated with the original event. A more rigorous method
for duplicate removal involves keeping the largest amplitude spikes and removing all spikes in a
radius that have decayed in amplitude. (cid:140)is method allows for the separating of near-synchronous
events that are in the same spatial area of the array. Its success, however, relies on the assump-
tion that the timing of spikes from the same event on nearby electrodes is almost identical and
only weakly in(cid:131)uenced by noise and that the signal spatially decays away from its current source
(Hagen et al., 2015).
3.3 Feature Extraction
(cid:140)e relevant signal a spike causes in extracellular recordings lasts around 3 ms, which, depending
on the acquisition rate, may correspond to up to 90 data points per event. However, spike shapes
are highly redundant and can be e(cid:129)ciently represented in a low dimensional space. (cid:140)us, an
appropriate projection method can be used to compute a small number of features for each event,
which can be more e(cid:129)ciently clustered than raw waveforms.
(cid:140)e most common feature extraction method for extracellular spikes is Principal Component
Analysis (PCA), performed on whitened and peak-aligned spike waveforms. PCA (cid:128)nds princi-
pal components, or orthogonal basis vectors, whose directions maximize the variance in the data.
Extracellular spikes can be summarized well by just 3-4 principal components, a manageable di-
mensionality for most clustering algorithms. (Adamos et al., 2008). Other less frequently used
methods include independent component analysis (ICA) (Hermle et al., 2004) and wavelet decom-
position ((cid:139)iroga et al., 2004). A comparison of these methods showed that the performance of
sorting algorithms depends not only on the feature extraction method employed, but also on the
clustering algorithm ((cid:139)iroga et al., 2004). In practice, the comparably low computational cost
and relative e(cid:130)ectiveness of PCA in discriminating between di(cid:130)erent neurons and neuron types
makes it particularly suitable for large scale recordings (Adamos et al., 2008). To reduce memory
load, the PCA decomposition can be evaluated for a subset of events from a large recording and
all events can be projected along the chosen dimensions e(cid:129)ciently in batches (Hilgen et al., 2017).
3.4 Clustering spatio-temporal event footprints
(cid:140)ere are (cid:128)ve fundamental problems with the clustering phase. (cid:140)e (cid:128)rst problem is that the ex-
tracellular waveform of neurons are known to change amplitude and shape during bursting (Fee
6
et al., 1996). (cid:140)e second problem is that some recorded waveforms are distorted by overlapping
action potentials from synchronous, spatially-local events. (cid:140)is occurs frequently in dense arrays
and usually exist as outliers in the chosen feature space. (cid:140)e third problem is that electrodes can
dri(cid:137) in the extracellular medium, changing the relative position of each neuron to the electrodes.
Dri(cid:137) distorts waveform shapes over the duration of the recording. (cid:140)e fourth problem is that the
duplication of spikes over neighboring channels can lead to refractory period violations or mis-
classi(cid:128)cations. (cid:140)e (cid:128)(cid:137)h and (cid:128)nal problem is that the number of observed neurons is unknown,
which requires the use of non-parametric clustering algorithms or requires the user to estimate
the number of neurons for a parametric clustering algorithm.
(cid:140)e choice of the clustering algorithm will be determined by the speed and scalability con-
siderations, by hypotheses over the typical shape of a cluster in this space, and by how well the
algorithm can deal with the previously listed problems. Many spike sorting methods cluster by
(cid:128)(cid:138)ing Gaussian Mixture Models (GMMs), modelling the feature density pro(cid:128)les as a sum of Gaus-
sians (Harris et al., 2000; Rossant et al., 2016), or by (cid:128)(cid:138)ing a mixture of t-distributions (Shoham
et al., 2003). (cid:140)e unknown number of actual neurons can be introduced as a latent variable, and
the inference problem be solved with the expectation-maximisation (EM) algorithm. Bayesian ap-
proaches, which also quantify parameter uncertainty, have also been introduced (Wood and Black,
2008). (cid:140)ese approaches, however, only perform well for single channels and are conceptually and
computationally hard to scale up to large, full-chip datasets.
More recent clustering algorithms for spike sorting are density-based. Density-based algo-
rithms generally detect peaks or high-density regions in the feature space that are separated by
low-density regions. (cid:140)ese algorithms are non-parametric, allowing the classi(cid:128)cation pipeline to
be fully automatic, however, the number of clusters found can depend heavily on both hyper-
parameters and the chosen feature space. Density-based clustering algorithms have been imple-
mented for spike sorting with promising results (Hilgen et al., 2017; Chung et al., 2017).
For dense arrays, an added complication arises since the information contained in event foot-
prints cannot be used directly for sorting spikes, since it is unknown which channels contain
responses of a single neuron and how many neurons cause the observed responses. (cid:140)e resulting
combinatorial explosion can be dealt with in three ways:
Masked clustering A straight-forward way to reduce the dimensionality of the clustering prob-
lem is to include only channels with detectable responses for each event. Classical expectation
maximisation on a mixture model is then possible when the irrelevant parts of the data are masked
out and replaced with a tractable noise model (Kadir et al., 2014). (cid:140)is strategy produces excellent
results with the help of a semi-automated re(cid:128)nement step. (Rossant et al., 2016). A main limitation
is, however, a super-linear scaling with the number of recordings channels, which makes it less
suitable for the latest generations of large-scale arrays.
Template matching Since the raw recorded signal can be linearly decomposed into a mixture
of footprints from di(cid:130)erent neurons (Segev et al., 2004), template matching has been a successful
strategy for spike sorting, and implementations are available that scale up to thousands of channels
(Pachitariu et al., 2016; Lee et al., 2017; Yger et al., 2018). (cid:140)is approach has two steps. First a col-
lection of spatio-temporal footprints is obtained in a single pass over the data and dimensionality
reduction and clustering is used to build templates for single neurons. In a second pass, all events
7
are assigned to the most likely template or combination of templates in the case of temporally
overlapping events.
A major advantage of template matching that temporally overlapping spikes are naturally ac-
counted for through addition of two relevant templates. (cid:140)is makes it very suitable for recordings
with high (cid:128)ring rates and correlations between nearby neurons. A potential limitation is that
neurons spiking at very low rates may remain undiscovered as no reliable template can be built
through averaging. Moreover, current implementations require a (cid:128)nal manual curation step. (cid:140)is
is, however, simpli(cid:128)ed by correcting the assignment based on templates, which can be merged or
split, rather than based on single events.
Spike localisation As explained above, the spatial spike footprint allows event localisation
through an estimation of the barycentre from the peak event amplitudes in nearby channels. (cid:140)is
produces density maps with clear, isolated peaks in event density, which represent spikes from
single or multiple, nearby neurons (see Figure 2 for an example). A two-dimensional density map
can be clustered very e(cid:129)ciently, and the combination of locations and waveform features obtained
through dimensionality reduction allows successful separation of nearby neurons. Density-based
clustering algorithms have been successfully employed to solve this task: DPClus, based on the
identi(cid:128)cation of density peaks (Jun et al., 2017a), ISO-SPLIT to grow uni-modal clusters from small
seeds (Chung et al., 2017) and Mean Shi(cid:137), which herds data points towards high-density areas
(Hilgen et al., 2017).
Of all methods discussed here, spike localisation and clustering potentially has the best com-
putational performance, since the actual computation is performed on a data set with much lower
dimensionality than the original data (Hilgen et al., 2017; Jun et al., 2017a). Because the number of
dimensions in the clustering step has to be kept small, it also discards useful information. However,
usually locations and spatio-temporal waveform features exhibit substantial redundancy (Hilgen
et al., 2017), making this approach the most suitable for very large arrays.
4 Evaluation
(cid:140)e evaluation of spike detection and sorting quality is complicated by data volume and com-
plexity, which makes both manual and automated curation challenging. It is however possible to
assess the quality of an algorithm using data with ground truth annotation. Moreover, methods for
post-hoc quality assessment of desirable properties of single units can be used to accept or reject
units found through spike sorting.
Speci(cid:128)cally, the desired result of a spike sorting pipeline to minimise the false detection of
noise as spikes (false positives in detection), and the number of real spikes le(cid:137) undetected (false
negatives in detection). Moreover, it should not assign spikes to the wrong neuron, hence it should
minimise false positives and negatives in a cluster assignment.
When ground truth annotations are available, false positives and false negatives can be easily
counted. A direct, but technically challenging method to obtain ground truth information, is the
simultaneous recording of a single neuron, together with an array recording, which will then be
analysed using the spike sorting algorithm in question. (cid:140)ree such data sets recorded with dense
arrays are currently available, two from the rat cortex recorded in vivo (Neto et al., 2016; Marques-
Smith et al., 2018), and one from the mouse retina recorded in vitro (Yger et al., 2018). In both cases,
a single juxtacellular electrode placed very closely to the array reliably recorded all spikes from
8
a single neuron. A systematic analysis of spike sorting has shown a clear relationship between
measured spike amplitude and classi(cid:128)cation accuracy with errors strongly increasing for events
smaller than 50 mV (Yger et al., 2018). (cid:140)is important result can help motivate exclusion of units
with weaker signals.
Ground truth for spike sorting can also be produced by simulations. Recently, it has become
possible to simulate a complete biophysical forward model for recorded extracellular potentials
in neural tissue (Hagen et al., 2015). (cid:140)is has produced several data sets that are now used to
benchmark spike sorting algorithms (see e.g. Lee et al. (2017)). In another study, ground truth
data was generated by superimposing synthetic spikes onto a recording from an empty array.
(cid:140)is data was used to evaluate the e(cid:130)ect of noise on event localisation accuracy and to discover
that localisation is inevitably a trade-o(cid:130) between position uncertainty and bias (Muthmann et al.,
2015). It is an open question, however, how well results collected from simulated data generalise,
since the precise noise model, which may di(cid:130)er between recording systems, impacts spike sorting
algorithm performance (Muthmann et al., 2015).
Finally, for cases where no ground truth data is available, Hill et al. (2011) proposed a set
of metrics that should accompany all spike sorting methods as an evaluation of their reliability.
(cid:140)eir metrics, applied a posteriori, are based on di(cid:130)erent features of the sorted dataset, which can
be summarised as follows:
• (cid:140)e waveforms in each cluster. (cid:140)e average waveform can present non-biological features,
hinting that the cluster may be a collection of wrongly detected events. Additionally, if
properties of the typical waveform change over time, this may be a sign of neurons dri(cid:137)ing
away from their initial position on the chip. Finally, anomalous variability of each feature
above the noise level may be a sign that multiple neurons contributed to the same cluster.
• (cid:140)e times of all spikes in a cluster. Violations of the refractory period show the cluster
contains false positives: these can be studied via the autocorrelation function or inter-spike
histogram of each cluster.
• (cid:140)e amplitudes of action potentials. A sharp drop in the amplitude distribution, caused by
the detection threshold, signi(cid:128)es that the la(cid:138)er has introduced an arti(cid:128)cial bias.
• (cid:140)e separation between pairs of clusters. Ample, sharp interfaces between clusters mean
the properties of each neuron's spikes overlap in the selected feature space. If this occurs,
there will be a theoretical minimum of false positives and negatives due to the incorrect
assignment of events to the wrong cluster.
(cid:140)e last point can be evaluated by re-examining a group of clustered neurons with a mixture model
(usually Gaussian), which can be (cid:128)t using more features than the original algorithm. Assuming
that this (cid:128)t is at least as reliable as the original sorting algorithm, a comparison of the two assign-
ments is informative regarding the reliability of each unit. A statistic summarizing all these tests
can then be used to exclude events and units post hoc. Using this method, detection and clustering
parameters do not have to be adjusted carefully prior to each analysis.
9
Name and reference
Kilosort (Pachitariu et al. (2016))
github.com/cortex-lab/KiloSort
YASS (Lee et al. (2017))
yass.readthedocs.io
Herding Spikes (Hilgen et al. (2017))
github.com/mhhennig/HS2
MountainSort (Chung et al. (2017))
github.com/(cid:131)atironinstitute/mountainsort
JRCLUST (Jun et al. (2017a))
jrclust.org
SpyKING CIRCUS (Yger et al. (2018))
spyking-circus.rtfd.org
Method Notes
TM
TM
SL+D
D
SL+D
TM
detection
template
GPU support; MATLAB based; semi-
automated (cid:128)nal curation.
network-based
Neural
(GPU); outlier
triaging;
matching; clustering.
Fast and scalable; tested on multiple
array geometries
Fully automatic;
graphi-
cal user interface; unique clustering
method
Probe dri(cid:137) correction; GPU support.
scalable;
GPU support; tested on many datasets;
robust to overlapping spikes; graphical
user interface.
Table 1: Summary of the most recent spike sorting methods developed for large, dense arrays. For a summary of
older algorithms -- mostly for smaller, sparser arrays -- see Bestel et al. (2012). TM = Template Matching; SL = Spike
Localisation; D = Density-based clustering (see section 3.4)
5 Outlook
In this chapter, we discussed the existing methodology for recovering single neuron activity from
high density recordings and the challenges and problems that each approach faces. Six freely
available spike sorting pipelines for large-scale extracellular arrays and the methods they use are
summarised in Table 1. For more information on their unique advantages and disadvantages,
please review their associated references.
Since inaccurate detection and sorting can in(cid:131)uence subsequent analysis of neural populations
(Ventura and Gerkin, 2012), manual curation steps are o(cid:137)en still required to guarantee good data
quality. However, the recent methods we summarised in this chapter take signi(cid:128)cant steps in
increasing the speed, automatisation, and accuracy of the spike sorting pipeline. Looking forward,
it may be possible to apply novel machine learning techniques to improving spike sorting. (cid:140)is
has already been put into practice with a recent spike sorting algorithm where a neural network
is used to improve detection of neural events (Lee et al., 2017). Although neural networks are
showing promising results in detection, it may be possible to (cid:128)nd new breakthroughs in both
feature extraction and in classi(cid:128)cation using these methods. Moreover, a neural network approach
may have the potential of encompassing all of the spike sorting steps within a single model. A
challenge when using these machine learning algorithms, however, is the di(cid:129)culty of obtaining
ground truth data, which is poorly available and usually under speci(cid:128)c experimental conditions
that may not generalize to other data sets.
Increased automation also means more work is needed in developing reliable methods for
validation and quality control of spike sorting results. (cid:140)e introduction of synthetic (Hagen et al.,
10
2015) and experimetal ground truth datasets (Neto et al., 2016; Yger et al., 2018) is an important
step forward in this direction. A standardisation, both of the sorting pipeline and of its evaluation,
should be considered among the next objectives of the spike sorting community. A joint e(cid:130)ort
should be made in order to guarantee that methods are intuitive to use and results are easy to
compare.
References
Adamos, D. A., Kosmidis, E. K., and (cid:140)eophilidis, G. (2008). Performance evaluation of pca-based
spike sorting algorithms. Computer Methods and Programs in Biomedicine, 91(3):232 -- 244.
Ballini, M., Muller, J., Livi, P., Chen, Y., Frey, U., Ste(cid:138)ler, A., Shadmani, A., Viswam, V., Jones, I. L.,
Jackel, D., Radivojevic, M., Lewandowska, M. K., Gong, W., Fiscella, M., Bakkum, D. J., Heer, F.,
and Hierlemann, A. (2014). A 1024-channel CMOS microelectrode array with 26,400 electrodes
for recording and stimulation of electrogenic cells in vitro. IEEE Journal of Solid-State Circuits,
49(11):2705 -- 2719.
Berdondini, L., van der Wal, P. D., Guenat, O., de Rooij, N. F., Koudelka-Hep, M., Seitz, P., Kauf-
mann, R., Metzler, P., Blanc, N., and Rohr, S. (2005). High-density electrode array for imaging in
vitro electrophysiological activity. Biosensors & Bioelectronics, 21(1):167 -- 74.
Bestel, R., Daus, A. W., and (cid:140)ielemann, C. (2012). A novel automated spike sorting algorithm with
adaptable feature extraction. Journal of Neuroscience Methods, 211(1):168 -- 178.
Buzs´aki, G. and Mizuseki, K. (2014). (cid:140)e log-dynamic brain: how skewed distributions a(cid:130)ect
network operations. Nature Reviews Neuroscience, 15(4):264.
Chung, J. E., Magland, J. F., Barne(cid:138), A. H., Tolosa, V. M., Tooker, A. C., Lee, K. Y., Shah, K. G., Felix,
S. H., Frank, L. M., and Greengard, L. F. (2017). A fully automated approach to spike sorting.
Neuron, 95(6):1381 -- 1394.
Dimitriadis, G., Neto, J. P., Aarts, A., Alexandru, A., Ballini, M., Ba(cid:138)aglia, F., Calcaterra, L., David,
F., Fiath, R., Frazao, J., et al. (2018). Why not record from every channel with a cmos scanning
probe? bioRxiv, page 275818.
Eversmann, B., Jenkner, M., Hofmann, F., Paulus, C., Brederlow, R., Holzap(cid:131), B., Fromherz, P.,
Merz, M., Brenner, M., Schreiter, M., Gabl, R., Plehnert, K., Steinhauser, M., Eckstein, G., Schmi(cid:138)-
landsiedel, D., and (cid:140)ewes, R. (2003). A 128 128 CMOS Biosensor Array for Extracellular Record-
ing of Neural Activity. IEEE Journal of Solid-State Circuits, 38(12):2306 -- 2317.
Fee, M. S., Mitra, P. P., and Kleinfeld, D. (1996). Variability of extracellular spike waveforms of
cortical neurons. Journal of Neurophysiology, 76(6):3823 -- 3833.
Frey, U., Sedivy, J., Heer, F., Pedron, R., Ballini, M., Mueller, J., Bakkum, D., Ha(cid:128)zovic, S., Faraci,
F. D., Greve, F., Kirstein, K. U., and Hierlemann, A. (2010). Switch-matrix-based high-density
microelectrode array in CMOS technology. IEEE Journal of Solid-State Circuits, 45(2):467 -- 482.
11
Hagen, E., Ness, T. V., Khosrowshahi, A., Sørensen, C., Fyhn, M., Ha(cid:137)ing, T., Franke, F., and
Einevoll, G. T. (2015). ViSAPy: A Python tool for biophysics-based generation of virtual spiking
activity for evaluation of spike-sorting algorithms. Journal of Neuroscience Methods, 245:182 --
204.
Harris, K. D., Henze, D. A., Csicsvari, J., Hirase, H., and Buzs´aki, G. (2000). Accuracy of tetrode
spike separation as determined by simultaneous intracellular and extracellular measurements.
Journal of Neurophysiololgy, 84(1):401 -- 414.
Hermle, T., Schwarz, C., and Bogdan, M. (2004). Employing ica and som for spike sorting of
multielectrode recordings from cns. Journal of Physiology-Paris, 98(4-6):349 -- 356.
Hilgen, G., Sorbaro, M., Pirmoradian, S., Muthmann, J.-O., Kepiro, I. E., Ullo, S., Ramirez, C. J.,
Encinas, A. P., Maccione, A., Berdondini, L., et al. (2017). Unsupervised spike sorting for large-
scale, high-density multielectrode arrays. Cell Reports, 18(10):2521 -- 2532.
Hill, D. N., Mehta, S. B., and Kleinfeld, D. (2011). (cid:139)ality metrics to accompany spike sorting of
extracellular signals. Journal of Neuroscience, 31(24):8699 -- 705.
Hrom´adka, T., Deweese, M. R., and Zador, A. M. (2008). Sparse representation of sounds in the
unanesthetized auditory cortex. PLoS Biology, 6(1):e16.
Jun, J. J., Mitelut, C., Lai, C., Gratiy, S., Anastassiou, C., and Harris, T. D. (2017a). Real-time spike
sorting platform for high-density extracellular probes with ground-truth validation and dri(cid:137)
correction. bioRxiv, page 101030.
Jun, J. J., Steinmetz, N. A., Siegle, J. H., Denman, D. J., Bauza, M., Barbarits, B., Lee, A. K., Anastas-
siou, C. A., Andrei, A., Aydın, C¸., et al. (2017b). Fully integrated silicon probes for high-density
recording of neural activity. Nature, 551(7679):232.
Kadir, S. N., Goodman, D. F., and Harris, K. D. (2014). High-dimensional cluster analysis with the
masked em algorithm. Neural Computation, 26(11):2379 -- 2394.
Lee, J. H., Carlson, D. E., Razaghi, H. S., Yao, W., Goetz, G. A., Hagen, E., Ba(cid:138)y, E., Chichilnisky,
E., Einevoll, G. T., and Paninski, L. (2017). Yass: Yet another spike sorter. In Advances in Neural
Information Processing Systems, pages 4005 -- 4015.
Lewicki, M. S. (1998). A review of methods for spike sorting: the detection and classi(cid:128)cation of
neural action potentials. Network, 9(January):R53 -- R78.
Lopez, C. M., Mitra, S., Putzeys, J., Raducanu, B., Ballini, M., Andrei, A., Severi, S., Welkenhuysen,
M., Van Hoof, C., Musa, S., et al. (2016). 22.7 a 966-electrode neural probe with 384 con(cid:128)gurable
channels in 0.13 µm soi cmos. In Solid-State Circuits Conference (ISSCC), 2016 IEEE International,
pages 392 -- 393. IEEE.
Marques-Smith, A., Neto, J. P., Lopes, G., Nogueira, J., Calcaterra, L., Frazo, J., Kim, D., Phillips,
M. G., Dimitriadis, G., and Kamp(cid:130), A. R. (2018). Recording from the same neuron with high-
density cmos probes and patch-clamp: a ground-truth dataset and an experiment in collabora-
tion. bioRxiv, page 370080.
12
Muller, J., Ballini, M., Livi, P., Chen, Y., Radivojevic, M., Shadmani, A., Viswam, V., Jones, I. L.,
Fiscella, M., Diggelmann, R., Ste(cid:138)ler, A., Frey, U., Bakkum, D. J., Hierlemann, A., Muller, J.,
Ballini, M., Livi, P., Chen, Y., Radivojevic, M., Shadmani, A., Viswam, V., Jones, I. L., Fiscella, M.,
Diggelmann, R., Ste(cid:138)ler, A., Frey, U., Bakkum, D. J., and Hierlemann, A. (2015). High-resolution
CMOS MEA platform to study neurons at subcellular, cellular, and network levels. Lab on a
Chip, 15(13):2767 -- 2780.
Muthmann, J.-O., Amin, H., Sernagor, E., Maccione, A., Panas, D., Berdondini, L., Bhalla, U. S.,
and Hennig, M. H. (2015). Spike Detection for Large Neural Populations Using High Density
Multielectrode Arrays. Frontiers in Neuroinformatics, 9(December):1 -- 21.
Neto, J. P., Lopes, G., Frazao, J., Nogueira, J., Lacerda, P., Baiao, P., Aarts, A., Andrei, A., Musa, S.,
Fortunato, E., et al. (2016). Validating silicon polytrodes with paired juxtacellular recordings:
method and dataset. Journal of Neurophysiology, 116(2):892 -- 903.
Obien, M. E. J., Deligkaris, K., Bullmann, T., Bakkum, D. J., and Frey, U. (2015). Revealing neuronal
function through microelectrode array recordings. Frontiers in Neuroscience, 9(JAN):423.
Pachitariu, M., Steinmetz, N. A., Kadir, S. N., Carandini, M., and Harris, K. D. (2016). Fast and accu-
rate spike sorting of high-channel count probes with kilosort. In Advances in Neural Information
Processing Systems, pages 4448 -- 4456.
Panas, D., Amin, H., Maccione, A., Muthmann, O., van Rossum, M., Berdondini, L., and Hennig,
M. H. (2015). Sloppiness in Spontaneously Active Neuronal Networks. Journal of Neuroscience,
35(22):8480 -- 8492.
(cid:139)iroga, R. Q., Nadasdy, Z., and Ben-Shaul, Y. (2004). Unsupervised spike detection and sorting
with wavelets and superparamagnetic clustering. Neural Computation, 16(8):1661 -- 87.
Rey, H. G., Pedreira, C., and (cid:139)ian (cid:139)iroga, R. (2015). Past, present and future of spike sorting
techniques. Brain Research Bulletin, 119:106 -- 117.
Rossant, C., Kadir, S. N., Goodman, D. F. M., Schulman, J., Hunter, M. L. D., Saleem, A. B., Grosmark,
A., Belluscio, M., Den(cid:128)eld, G. H., Ecker, A. S., Tolias, A. S., Solomon, S., Buzs´aki, G., Carandini,
M., and Harris, K. D. (2016). Spike sorting for large, dense electrode arrays. Nature Neuroscience,
19(4):634 -- 641.
Segev, R., Goodhouse, J., Puchalla, J., and Berry II, M. J. (2004). Recording spikes from a large
fraction of the ganglion cells in a retinal patch. Nature Neuroscience, 7(10):1155.
Shoham, S., Fellows, M. R., and Normann, R. A. (2003). Robust, automatic spike sorting using
mixtures of multivariate t-distributions. Journal of Neuroscience Methods, 127(2):111 -- 122.
Ventura, V. and Gerkin, R. C. (2012). Accurately estimating neuronal correlation requires a new
spike-sorting paradigm. Proceedings of the National Academy of Sciences, 109(19):7230 -- 7235.
Wood, F. and Black, M. J. (2008). A nonparametric bayesian alternative to spike sorting. Journal
of Neuroscience Methods, 173(1):1 -- 12.
13
Yger, P., Spampinato, G. L., Esposito, E., Lefebvre, B., Deny, S., Gardella, C., Stimberg, M., Je(cid:138)er,
F., Zeck, G., Picaud, S., et al. (2018). A spike sorting toolbox for up to thousands of electrodes
validated with ground truth recordings in vitro and in vivo. eLife, 7:e34518.
Yuan, X., Kim, S., Juyon, J., D'Urbino, M., Bullmann, T., Chen, Y., Ste(cid:138)ler, A., Hierlemann, A.,
and Frey, U. (2016). A microelectrode array with 8,640 electrodes enabling simultaneous full-
frame readout at 6.5 kfps and 112-channel switch-matrix readout at 20 ks/s.
In VLSI Circuits
(VLSI-Circuits), 2016 IEEE Symposium on, pages 1 -- 2. IEEE.
Zhang, J., Laiwalla, F., Kim, J. A., Urabe, H., Van Wagenen, R., Song, Y.-K., Connors, B. W., Zhang,
Integrated device for optical stimulation and
F., Deisseroth, K., and Nurmikko, A. V. (2009).
spatiotemporal electrical recording of neural activity in light-sensitized brain tissue. Journal of
neural engineering, 6(5):055007.
14
|
1412.1713 | 2 | 1412 | 2015-07-02T13:28:29 | Networks that learn the precise timing of event sequences | [
"q-bio.NC"
] | Neuronal circuits can learn and replay firing patterns evoked by sequences of sensory stimuli. After training, a brief cue can trigger a spatiotemporal pattern of neural activity similar to that evoked by a learned stimulus sequence. Network models show that such sequence learning can occur through the shaping of feedforward excitatory connectivity via long term plasticity. Previous models describe how event order can be learned, but they typically do not explain how precise timing can be recalled. We propose a mechanism for learning both the order and precise timing of event sequences. In our recurrent network model, long term plasticity leads to the learning of the sequence, while short term facilitation enables temporally precise replay of events. Learned synaptic weights between populations determine the time necessary for one population to activate another. Long term plasticity adjusts these weights so that the trained event times are matched during playback. While we chose short term facilitation as a time-tracking process, we also demonstrate that other mechanisms, such as spike rate adaptation, can fulfill this role. We also analyze the impact of trial-to-trial variability, showing how observational errors as well as neuronal noise result in variability in learned event times. The dynamics of the playback process determine how stochasticity is inherited in learned sequence timings. Future experiments that characterize such variability can therefore shed light on the neural mechanisms of sequence learning. | q-bio.NC | q-bio | Noname manuscript No.
(will be inserted by the editor)
Networks that learn the precise timing of event
sequences
Alan Veliz-Cuba · Harel Z. Shouval · Kresimir Josi´c* · Zachary
P. Kilpatrick*
the date of receipt and acceptance should be inserted later
Abstract Neuronal circuits can learn and replay
firing patterns evoked by sequences of sensory stim-
uli. After training, a brief cue can trigger a spa-
tiotemporal pattern of neural activity similar to
that evoked by a learned stimulus sequence. Net-
work models show that such sequence learning can
occur through the shaping of feedforward excita-
tory connectivity via long term plasticity. Previous
models describe how event order can be learned,
but they typically do not explain how precise tim-
ing can be recalled. We propose a mechanism for
learning both the order and precise timing of event
A. Veliz-Cuba
Department of Mathematics, University of Houston,
Houston TX 77204 USA
E-mail: [email protected]
H.Z. Shouval
Department of Neurobiology and Anatomy, University
of Texas Medical School, Houston TX 77030 USA
E-mail: [email protected]
K. Josi´c
Department of Mathematics and Department of Biology,
University of Houston, Houston TX 77204 USA
E-mail: [email protected]
Z.P. Kilpatrick
Department of Mathematics, University of Houston,
Houston TX 77204 USA
E-mail: [email protected]
*equal contribution
sequences. In our recurrent network model, long
term plasticity leads to the learning of the sequence,
while short term facilitation enables temporally pre-
cise replay of events. Learned synaptic weights be-
tween populations determine the time necessary
for one population to activate another. Long term
plasticity adjusts these weights so that the trained
event times are matched during playback. While we
chose short term facilitation as a time-tracking pro-
cess, we also demonstrate that other mechanisms,
such as spike rate adaptation, can fulfill this role.
We also analyze the impact of trial-to-trial vari-
ability, showing how observational errors as well
as neuronal noise result in variability in learned
event times. The dynamics of the playback process
determine how stochasticity is inherited in learned
sequence timings. Future experiments that charac-
terize such variability can therefore shed light on
the neural mechanisms of sequence learning.
Keywords serial recall · short term facilitation ·
long term plasticity
1 Introduction
Networks of the brain are capable of precisely learn-
ing and replaying sequences, accurately represent-
ing the timing and order of the constituent events
5
1
0
2
l
u
J
2
]
.
C
N
o
i
b
-
q
[
2
v
3
1
7
1
.
2
1
4
1
:
v
i
X
r
a
2
Alan Veliz-Cuba et al.
(Conway and Christiansen, 2001; Buhusi and Meck,
2005). Recordings in awake monkeys and rats re-
veal neural mechanisms that underlie such sequence
representation. After a training period consisting
of the repeated presentation of a cue followed by
a fixed sequence of stimuli, the cue alone can trig-
ger a pattern of neural activity correlated with the
activity pattern evoked by the stimulus sequence
(Eagleman and Dragoi, 2012; Xu et al, 2012). Im-
portantly, the temporal patterns of the stimulus-
driven and cue-evoked activity are closely matched
(Shuler and Bear, 2006; Gavornik and Bear, 2014).
Sequence learning and replay has been iden-
tified in a number of different brain areas. Re-
cent electrophysiological recordings have located
patterned neural activity in V1 (Xu et al, 2012;
Gavornik and Bear, 2014) and V4 (Eagleman and
Dragoi, 2012), corresponding to learned visual se-
quences. Experiments on motor sequence learning
found the underlying activity was coordinated by
a combination of prefrontal, associative, and mo-
tor cortical areas (Jenkins et al, 1994; Sakai et al,
1998). Training networks of the brain to replay mo-
tor sequences is important since it allows quick mo-
tor skill execution, faster than deliberate muscle
control allows (Hikosaka et al, 2002). In addition,
learning visual sequences can aid in experience-
based prediction, so animals can react quickly to
an unexpected chain of events (Meyer and Olson,
2011; Kok et al, 2012). Learning serial order is also
an essential component of language and speech pro-
duction in humans (Burgess and Hitch, 1999). In
a similar way, music perception and production re-
quires that humans learn to recognize and generate
auditory-motor sequences (Zatorre et al, 2007). In
total, sequence learning plays a large role in the
daily cognitive tasks of a wide variety of animals.
Various neural mechanisms have been proposed
for learning the duration of a single event (Buono-
mano, 2000; Rao and Sejnowski, 2001; Durstewitz,
2003; Reutimann et al, 2004; Karmarkar and Buono-
mano, 2007; Gavornik et al, 2009), as well as the
order of events in a sequence (Amari, 1972; Kle-
infeld, 1986; Wang and Arbib, 1990; Abbott and
Blum, 1996; Jun and Jin, 2007; Fiete et al, 2010;
Brea et al, 2013). However, mechanisms for learn-
ing the precise timing of multiple events in a se-
quence remain largely unexplored. The activity of
single neurons evolves on the timescale of tens of
milliseconds. It is therefore likely that sequences
on the timescale of seconds are represented in the
activity of populations of cells. Recurrent network
architecture could determine activity patterns that
arise in the absence of input, but how this architec-
ture can be reshaped by training to support pre-
cisely timed sequence replay is not understood.
Long term potentiation (LTP) and long term
depression (LTD) are fundamental neural mecha-
nisms that change the weight of connections be-
tween neurons (Kandel, 2001). Learning in a wide
variety of species, neuron types, and parts of the
nervous system has been shown to occur through
LTP and LTD (Alberini, 2009; Takeuchi et al, 2014;
Nabavi et al, 2014). It is therefore natural to ask
whether LTP and LTD can play a role in the learn-
ing of sequence timing (Karmarkar and Buono-
mano, 2007; Ivry and Schlerf, 2008; Bueti and Buono-
mano, 2014), in addition to their proposed role in
learning sequence order (Abbott and Blum, 1996;
Fiete et al, 2010).
We introduce a neural network model capa-
ble of learning the timing of events in a sequence.
The connectivity and dynamics in the network are
shaped by two mechanisms: long term plasticity
and short term facilitation. Long term synaptic
plasticity allows the network to encode sequence
and timing information in the synaptic weights,
while slowly evolving short term facilitation can
mark time during event playback. These ideas are
quite general, and we show that they do not depend
on the particulars of the time-tracking mechanisms
we implemented. The impact of stimulus variabil-
ity and neural noise is largely determined by the
trajectory of the time-tracking process. Thus, we
predict that errors in event sequence recall may be
indicative of the mechanism that encodes them.
Precise timing of event sequences
2 Material and Methods
2.1 Population rate model with short term
facilitation
Pyramidal cells in cortex form highly connected
clusters (Song et al, 2005; Perin et al, 2011), which
can correspond to neurons with similar stimulus
tuning (Ko et al, 2011). We therefore considered
a rate model describing the activity of N excita-
tory populations (clusters) uj (j = 1, . . . , N ), and a
single inhibitory population v. The excitatory and
inhibitory populations were coupled via long range
connections. Each population j received an exter-
nal input Ij(t). Our model took the form:
τ
τf
duj
dt
dpj
dt
τ
dv
dt
= −uj + ϕ (Ij(t) + Ij,syn − θ) ,
= 1 − pj + (pmax − 1)uj,
= −v + ϕ
Zkuk − θv
,
(cid:33)
(1)
(cid:32) N(cid:88)
k=1
N(cid:88)
k(cid:54)=j
where synaptic inputs to the jth population are
given by
Ij,syn = wjjuj +
wjkpkuk − Lv.
A complete description of the model functions and
parameters is given in Table 1. Population firing
rates ranged between a small positive value (the
background firing rate), and a maximal value (the
rate of a driven population), normalized to be 0
and 1, respectively. The baseline weight of the con-
nection from population k to j was denoted by wjk.
Connections within a population were denoted wjj,
and these were not subject to short-term facilita-
tion. Furthermore, the global inhibitory popula-
tion, modeled by v, was typically active for very
short epochs, so the effects of short term plasticity
were not considered. These assumptions did not
change our results, but made the analysis more
transparent.
Synapses between populations were subject to
short term facilitation, and the facilitated connec-
tion had "effective synaptic strength" wjkpk (Tsodyks
3
et al, 1998). Without loss of generality we assumed
that short term facilitation varied between 1 and 2
so that the effective synaptic strength varied from
wjk to 2wjk. Note that rescaling the maximal level
of short term facilitation will simply rescale the
relationship we will derive between the baseline
synaptic weight and activation time of single pop-
ulations. We assumed τf (cid:29) τ , in keeping with the
observation that synaptic facilitation dynamics are
much slower than changes in firing rates (Markram
et al, 1998).
In the absence of external stimulus or input
from other populations, the dynamics of each pop-
dt = −uj + ϕ(wjjuj − θ),
ulation is described by τ duj
and the stationary firing rate is given by the solu-
tions of ϕ(wjjuj − θ) = uj. We typically took ϕ
to be a Heaviside step function. Thus, if wjj > θ,
there are two equilibrium firing rates: uj = 0 (in-
active population) and uj = 1 (active population).
This assumption simplified the analysis, however
we show show in Section 4.2 that they are not es-
sential.
2.2 Rate-based long term plasticity
Connectivity between the populations in the net-
work was subject to long term potentiation (LTP)
and long term depression (LTD). Following experi-
mental evidence (Bliss and Lømo, 1973; Dudek and
Bear, 1992; Markram and Tsodyks, 1996; Sjostrom
et al, 2001), connections were modulated using a
rule based on pre- and post-synaptic activity with
'soft' bounds (Gerstner and Kistler, 2002).
We made three main assumptions about the
long term evolution of synaptic weight, w = wpre→post:
(a) If the presynaptic population activity was low
(upre ≈ 0), the change in synaptic weight was neg-
ligible ( w(t) = 0); (b) If the presynaptic population
was highly active (upre ≈ 1) and the postsynaptic
population responded weakly (upost ≈ 0), then the
synaptic weight decayed toward zero ( w(t) ∝ −w);
and (c) If both populations had a high level of
activity (upost ≈ 1 and upre ≈ 1), then synaptic
weight increased towards an upper bound ( w(t) ∝
wmax − w).
4
Alan Veliz-Cuba et al.
Table 1 Variables and parameters with their default values. The default values were used in all simulations, unless
otherwise noted. References indicate work where similar parameter values were used or estimated: 1Graupner and
Brunel (2012), 2Xu et al (2012), 3Gavornik and Bear (2014), 4Markram and Tsodyks (1996), 5Lundstrom (2015);
6Hausser and Roth (1997).
Variables
symbol
Ij
uj
v
pj
wjk, w
Tj, T
symbol
τ
τf
τw
τa
τs
Tcue
D
D(cid:48)
description
external stimulus for excitatory population j
non-dimensional firing rate of excitatory population j (maximum uj = 1)
non-dimensional firing rate of global inhibitory population (maximum v = 1)
level of facilitation of synapses from population j (baseline pj = 1)
strength of excitation from population k to excitatory population j
duration of stimulus
Time parameters (default values in parenthesis)
description
timescale of neuronal firing (10ms 6)
timescale of short term facilitation (1s 4)
timescale of learning rule (150s 1)
time scale of adaptation (400ms 5)
time scale of synaptic inputs from other populations (50ms 6)
duration of stimulus to trigger replay (50ms 2,3)
delay in presynaptic firing affecting connections between populations (30ms 1)
delay in presynaptic firing affecting connections within populations (20ms 1)
Other parameters (default values in parenthesis)
description
firing rate response function (Heaviside step function)
threshold for activation of excitatory population (0.5)
threshold for activation of inhibitory population (0.5)
maximum level of short term facilitation (2)
strength of excitation from population k to inhibitory population (0.3)
weight of global inhibition (0.6)
strength of adaptation (1)
learning rule threshold (1)
maximum synaptic weight between populations (0.4852)
maximum synaptic weight within populations (4.1312)
minimum synaptic weight within populations (1.3488)
strength of LTD between populations (150 1)
strength of LTP between populations (3614.5 1)
strength of LTD within populations (7500)
strength of LTP within populations (267.86 1)
symbol
ϕ
θ
θv
pmax
Zk
L
b
M
wmax
w(cid:48)
max
wmin
γd
γp
γ(cid:48)
γ(cid:48)
d
p
Precise timing of event sequences
5
Similar assumptions have been used in previ-
ous rate-based models of LTP/LTD (von der Mals-
burg, 1973; Bienenstock et al, 1982; Oja, 1982;
Miller, 1994), and it has been shown that calcium-
based (Graupner and Brunel, 2012) and spike-time
dependent (Clopath et al, 2010; Gjorgjieva et al,
2011) plasticity rules can be reduced to such rate-
based rules (Pfister and Gerstner, 2006). Further-
more, the fact that pre-synaptic activity is neces-
sary to initiate either LTP or LTD is supported
by experimental observations that plasticity de-
pends on calcium influx through NMDA receptors
(Malenka and Bear, 2004). A simple differential
equation that implements these assumptions is
τw
dw
dt
= − γd w upre(t − D)(M − upost(t))
(2)
+ γp (wmax − w) upre(t − D) upost(t),
where τw is the time scale, γd (γp) represents the
strength of LTD (LTP), D is a delay accounting
for the time it takes for the presynaptic firing rate
to trigger plasticity processes, and M is a param-
eter that determines the threshold and magnitude
of LTD. We note that we initially model only the
molecular processes that detect correlations in fir-
ing rates, and thus set τw = 150s. We will later ex-
tend this model to account for the longer timescales
of synaptic weight changes (Section 4.4).
Eq. (2) describes a Hebbian rate-based plas-
ticity rule with soft bounds involving only linear
and quadratic dependences of the pre- and post-
synaptic rates (Gerstner and Kistler, 2002). Tem-
poral asymmetry that accounts for the causal link
between pre- and post-synaptic activity is incorpo-
rated with a small delay in the dependence of pre-
synaptic activity upre(t − D) (Gutig et al, 2003).
This learning rule is a firing-rate version of the
calcium-based plasticity model proposed by Graup-
ner and Brunel (2012).
2.3 Encoding timing of event sequences
Our training protocol was based on several recent
experiments that explored cortical learning in re-
sponse to sequences of visual stimuli (Xu et al,
2012; Eagleman and Dragoi, 2012; Gavornik and
Bear, 2014). During a training trial, an external
stimulus Ij(t) activated one population at a time.
Each individual stimulus could have a different du-
ration (Fig. 1B). We stimulated n populations, and
enumerated them by order of stimulation; that is,
population 1 was stimulated first, then population
2, and so on. This numbering is arbitrary, and the
initial recurrent connections have no relation to
this order. We denote the duration of input j by
Tj. All inputs stop at Ttot = T1 + T2 + . . . + Tn. A
sequence was presented m times.
Repeated training of the network described by
Eq. (1) with a fixed sequence drove the synap-
tic weights wij to equilibrium values. We assumed
that during sequence presentation, the amplitude
of external stimuli Ij(t) was sufficiently strong to
dominate the dynamics of the population rates, uj.
Then, the activity of the populations during train-
ing evolved according to:
τ
duj
dt
dwjk
dt
τw
j = 1, . . . , n,
= −uj + ϕ (Ij(t) − θ) ,
= −γdwjkuk(t − D)(M − uj(t))
(3)
+ γp(wmax − wjk)uk(t − D)uj(t), j (cid:54)= k.
Thus, the timing of population activations mim-
icked the timing of the input sequence, i.e. the
training stimulus.
2.3.1 Synaptic weights for consecutive activations
For simplicity, we begin by describing the case of
two populations, N = 2, and we consider the thresh-
old that determines the level of LTD equal to 1,
M = 1, so that LTD is absent when the postsy-
naptic population is active (upost = 1). Suppose
that I1(t) = IS on [0, T1] and I2(t) = −IH , and
I2(t) = IS on [T1, T1 + T2] and I1(t) = −IH , where
IS and IH are large enough so that Eq. (3) is valid.
The positive inputs with weight IS model feedfor-
ward excitation to cells tuned to the cue from up-
stream visual processing regions in thalamus. Neg-
ative inputs with weight −IH model strong effec-
tive feedforward inhibition to cortical cells that are
6
not tuned to the present cue (Wang et al, 2007;
Haider et al, 2013). Representing the effect of feed-
forward inhibition as static inputs simplified the
model, and did not affect our results.
We assumed that Ti > D, Ti (cid:29) τ , and τw (cid:29) τ ,
so that the stimulus was longer than the plastic-
ity delay, and plasticity slower than the firing rate
dynamics. Separation of timescales in Eq. (3) im-
plies that uj ≈ ϕ(Ij(t) − θ), so the firing rate of
populations 1 and 2 is approximated by u1(t) ≈ 1
on [0, T1] and zero elsewhere, and u2(t) ≈ 1 on
[T1, T1 + T2] and zero elsewhere (Fig. 2A). Hence,
during a training trial on a time interval [0, Ttot],
we obtain from Eq. (3) the following piecewise equa-
tion for the synaptic weight, w21, in terms of the
duration of the first stimulus, T1,
0,
− γd
w21,
τw
γp(wmax − w21)
τw
t (cid:54)∈ [D, T1 + D],
t ∈ [D, T1],
, t ∈ [T1, T1 + D].
(4)
dw21
dt
=
Note that we assumed that u1(t) = 0 for t < 0.
Eq. (4) allows the network to encode T1 using
the weight w21. Namely, solving Eq. (4) we obtain
w21(Ttot) =w21(0)e−T1γd/τw e−(γp−γd)D/τw
+ (1 − e−Dγp/τw )wmax,
which relates the synaptic weight at the end of a
presentation, w21(Ttot), to the synaptic weight at
the beginning of the presentation, w21(0) (Fig. 2A).
Thus, there is a recursive relation that relates the
weight w21 at the end of the i + 1st stimulus to the
weight at the end of the ith stimulus:
wi+1
21 =wi
21e−T1γd/τw e−(γp−γd)D/τw
+ (1 − e−Dγp/τw )wmax.
(5)
As long as γp > γd, the sequence (wi
to
w∞
21 :=
(1 − e−Dγp/τw )wmax
1 − e−T1γd/τw e−(γp−γd)D/τw
,
21) converges
(6)
as seen in Fig. 2B. An equivalent expression also
holds in the case of an arbitrary number of popu-
lations.
Alan Veliz-Cuba et al.
The relative distance to the fixed point w∞
21 is
computed by noting that (for T1 fixed)
wi+1
21 −w∞
21/wi
21−w∞
21 = e−T1γd/τw e−(γp−γd)D/τw ,
from which we calculate
21 ∝(cid:16)
wi
21 − w∞
e−T1γd/τw e−(γp−γd)D/τw
(cid:17)i
.
Thus, the sequence converges exponentially with
the number of training trials, i. The relative dis-
tance to the fixed point is proportional to e−mT1γd/τw ,
so the convergence is faster for larger values of T1,
as shown in Fig. 2C.
2.3.2 Synaptic weights of populations that are not
co-activated
To compute the dynamics of w12, we note that dur-
ing a training trial on the time interval [0, Ttot], the
following piecewise equation governs the change in
synaptic weight,
dw12
dt
=
t (cid:54)∈ [T1 + D, T1 + T2 + D],
0,
w12, t ∈ [T1 + D, T1 + T2 + D],
− γd
τw
(cid:40)
which can be solved explicitly to find
w12(Ttot) = w12(0)e−T2γd/τw .
(7)
We can therefore write a recursive equation for
the weight after the i + 1st stimulus in terms of the
weight after the ith stimulus
wi+1
12 = 0. Thus, w∞
12e−T2γd/τw ,
12 = wi
which converges to w∞
jk = 0 for all
pairs of populations (j, k) for which population j
was not activated immediately after population k
during training. In total, sequential activation of
the populations leads to the strengthening only of
the weights wj+1,j, while other weights are weak-
ened.
Precise timing of event sequences
7
2.4 Reactivation of trained networks
To examine how a sequence of event timings could
be encoded by our network, the first neural popu-
lation in the sequence was activated with a short
cue. Typically, this cue was of the form I1(t) = 1
for t ∈ [0, Tcue], I1(t) = 0 for t ∈ [Tcue,∞), and
Ij(t) = 0 for j (cid:54)= 1 (Fig. 1C). During replay, aside
from the initial cue, the activity in the network was
generated through recurrent connectivity.
We describe the case of two populations where
the first population is cued, and remains active
due to self-excitation (u1(t) ≈ 1), Fig. 3A. Since
u2(0) = 0 and ϕ is the Heaviside step function,
the equations governing the dynamics of the sec-
ond population are
= −u2 + ϕ(w21p1 − θ),
du2
τ
dt
p1(t) = pmax + (1 − pmax)e−t/τf .
(8)
Thus, for population 2 to become active, w21p1(t)
must have reached θ (Fig. 3A). The time T be-
tween when u1 becomes active and u2 becomes
active ("replay time") could be controlled by the
synaptic weight w21. Fig. 3B shows the effect of
changing the synaptic weight: For very small val-
ues of the baseline weight w21, the effective synap-
tic strength w21p1(t) never reaches θ and activation
does not occur. Increasing the baseline weight w21
causes more rapid activation of the second popula-
tion, and for very large weights w21 the activation
is immediate. The weight required for a presynap-
tic population to activate a postsynaptic popula-
tion after T units of time is given in closed form
by
W(T ) :=
pmax + (1 − pmax)e−T /τf
(9)
θ
.
Similarly, the inverse of this function,
T (w) := τf ln
(cid:18) pmax − 1
(cid:19)
,
pmax − θ/w
(10)
gives the activation time as a function of the synap-
tic weight (Fig. 3C). Note that Eq. (10) is valid for
θ/pmax < w < θ. If w ≤ θ/pmax, then activation
of the next population does not occur. If w ≥ θ,
activation is immediate.
To ensure the first population becomes inactive
when population 2 becomes active, we assumed
that global inhibition overcame the self excitation
in the first population, w11 + w12p2(T )− L− θ < 0.
Also, for the second population to remain active,
we needed the self excitation plus the input re-
ceived from population 1 to be stronger than the
global inhibition; namely, w22 +w21p1(T )−L−θ =
w22 − L > 0. These two inequalities are satisfied
whenever w12 is small enough and L < wjj < L+θ.
2.5 Matching training parameters to reactivation
parameters
To guarantee that long term plasticity leads to a
proper encoding of event times, it is necessary that
the learned weight, w∞
21 given by Eq. (6), matches
the desired weight W(T ) given by Eq. (9). This
can be achieved by equating the right hand sides
of Eq. (6) and Eq. (9), so that
(1 − e−Dγp/τw )wmax
1 − e−T γd/τw e−(γp−γd)D/τw
=
=
θ/pmax
1 + e−T /τf (1 − pmax)/pmax
.
(11)
Eq. (11) can be satisfied for all values of T by
choosing parameters that satisfy
τw/γd = τf ,
(1 − e−Dγp/τw )wmax = θ/pmax,
e−(γp−γd)D/τw = (pmax − 1)/pmax.
(12)
Since there are fewer equations than model param-
eters, there is a multi-dimensional manifold of pa-
rameters for which Eq. (11) holds for all T . For
instance, for fixed short-term facilitation parame-
ters θ, pmax, and τf and restricting specific plastic-
ity parameters τw and D, the appropriate γd, γp,
and wmax can be determined using Eq. (12). This
is how we determined the parameters in Figs. 4
and 5.
The first relationship in Eq. (12) states τw/γd =
τf , relating the timescale of short term facilitation
8
Alan Veliz-Cuba et al.
to the timescale of long term plasticity through the
depression amplitude parameter γd. It is important
to note that this does not mean that the timescales
of the two processes need to match. As stated in
Table 1, following experimental data (Markram and
Tsodyks, 1996; Alberini, 2009; Graupner and Brunel,
2012; Nabavi et al, 2014), we chose τf = 1s and
τw = 150s for our simulations. This implies that
γd = 150, for training to yield the correct weights.
As we demonstrate in Supplementary Fig. 1,
perturbing parameters of the long term plasticity
process away from the optimal relationships deter-
mined by Eq. (12) does alter the learned time. The
relative size of errors depends on the parameters we
perturb, and we find the model is most sensitive to
perturbation of wmax. Perturbations of other pa-
rameters such as γp and γd lead to to errors roughly
equal in to the magnitude of the parameter pertur-
bation (e.g., a 5% perturbation of γd leads to a 5%
change in Treplay).
For more detailed models, the analog of Eq. (12)
is more cumbersome or impossible to obtain explic-
itly. Specifically, when we incorporated noise into
our models in the Section 4.1 and considered spike
rate adaptation in the Section 4.3, we had to use
an alternative approach. We found it was always
possible to use numerical means to approximate
parameter sets that allowed a correspondence be-
tween the trained and desired weight for all possi-
ble event times. A simple way of finding such pa-
rameters was to use the method of least squares:
We selected a range of stimulus durations, e.g. [.1s, 3s],
and sampled timings from it, e.g. S = {.1s, .2s, .3s,
. . . , 3s}. For each T ∈ S we computed the learned
weight, wlearned(T, pars), where "pars" denotes the
list of parameters to be determined. Then, we com-
puted the replay time, Treplay(wlearned(T, pars)).
We defined the "cost" function
(cid:88)
T∈S
J(pars) =
(Treplay(wlearned(T, pars)) − T )2 ,
and the desired parameters were given by
parsbest := argminpars{J(pars)}.
This approach was successful for different models
and training protocols, and allowed us to find a
(13)
working set of parameters for models that included
noise or different slow processes for tracking time.
Eq. (13) can also be interpreted as a learning
rule for the network parameters. Starting with arbi-
trary network parameter values, any update mech-
anism that decreases the cost function J will result
in a network that can accurately replay learned se-
quence times.
2.6 Training and replay simulations
To test our model, we trained the network with a
sequence of four events. Each event in the sequence
corresponded to the activation of a single neuronal
population (Fig. 4A). Since each population was
inactivated (received strong negative input) when
the subsequent populations became active, we also
assumed that an additional, final population inac-
tivated the population responding to the last event
(additional population not shown in the figure). In-
put during training was strong enough so that acti-
vation of the different populations was only deter-
mined by the external stimulus overriding global
inhibition and recurrent excitation.
During reactivation, the recurrent connections
were assumed fixed. This assumption can be re-
laxed if we assume that LTP/LTD are not imme-
diate, but occur on long timescales, as in the Sec-
tion 4.4.
We used m = 10 training trials, the default pa-
rameter values in Table 1, and estimated γd, γp,
and wmax using Eq. (12). After the training tri-
als were finished, we cued the first population in
the sequence, using I1(t) = 1 for t ∈ [0, Tcue] and
Ij(t) = 0 otherwise. We also started with this set
of weights, and retrained the network with a novel
sequence of stimuli.
3 Results
3.1 Training
We explore sequence learning in a network model
of neural populations, where each population is ac-
tivated by a distinct stimulus or event. The initial
Precise timing of event sequences
9
Fig. 1 The precise timing of training sequences is learned via long term plasticity. Each stimulus in the sequence
is represented by a different color. A. Before training, network connectivity is random, and a cue does not trigger a
sequential pattern of activity. B. During training, a sequence of events is presented repeatedly. Each event activates
a corresponding neural population for some amount of time, which is fixed across presentations. Long term plasticity
reshapes network architecture to encode the duration and order of these activations. C. After sufficient training, a
cue triggers the pattern of activity evoked during the training period. Learned synaptic connectivity along with short
term facilitation steer activity along the path carved by the training sequence (arrow width and contrast correspond
to synaptic weight).
connectivity between the populations is random
(Fig. 1A). To make our analysis more transpar-
ent, we initially consider a deterministic firing rate
model, Eq. (1). Each individual neural population
is bistable, having both a low activity state and a
high activity state that is maintained through re-
current excitation. Our results also hold for more
biologically plausible firing rate response functions
and are robust to noise (Sections 4.1 and 4.2).
To train the network, we stimulated popula-
tions in a fixed order, similar to the training paradigm
used by Xu et al (2012); Eagleman and Dragoi
(2012); Gavornik and Bear (2014). The duration
of each event in the training sequence was arbi-
trary (Fig. 1B), and each stimulus in the sequence
drove a single neural population. Synaptic connec-
tions between populations were plastic. To keep the
model tractable, population activity was assumed
to immediately impact the weight of synaptic con-
nections. Our results also extend to a model with
synaptic weights changing on longer timescales (Sec-
tion 4.4).
Changes in the network's synaptic weights de-
pended on the firing rates of the pre- and post-
synaptic populations (Bliss and Lømo, 1973; Bi-
enenstock et al, 1982; Dudek and Bear, 1992; Markram
and Tsodyks, 1996; Sjostrom et al, 2001). When
a presynaptic population was active, either: (a)
synapses were potentiated (LTP) if the postsynap-
tic population was subsequently active or (b) syna-
pses were depressed (LTD) if the postsynaptic pop-
ulation was not activated soon after (Materials and
Methods). Such rate-based plasticity rules can be
derived from spike time dependent plasticity rules
(Kempter et al, 1999; Pfister and Gerstner, 2006;
Clopath et al, 2010).
To demonstrate how the timing of events can be
encoded in the network architecture, we start with
timesequence trainingLTP LTDtime-codingrecurrent networkBCAtime (s)firingrate01timecuebefore trainingtime (s)firingrate01timecueafter trainingtime (s)firingrate011234arbitrary connectivitysequence presentationI1I2I3I412341234 10
Alan Veliz-Cuba et al.
ods). While the first stimulus was present, popu-
lation 1 was active and LTD dominated, decreas-
ing the synaptic weight, w21, from population 1 to
population 2. After T1 seconds, the first stimulus
ended, and the second population was activated.
However, population 1 did not become inactive in-
stantaneously, and for some time both population
1 and 2 were active. During this overlap window,
LTP dominated leading to an increase in synap-
tic weight w21. Shortly after population 1 became
inactive, changes in the weight w21 ceased, as plas-
ticity only occurs when the presynaptic population
is active. The initial and final synaptic weights
(w0
21, respectively) can be computed in
closed form (Materials and Methods). Repeated
presentations of the training sequence leads to ex-
ponential convergence of the synaptic weights, wi
21
(weight after ith training trial), to a fixed value
(Fig. 2B). On the other hand, the synaptic weight
w12 is weakened during each trial because the presy-
naptic population 2 is always active after the post-
synaptic population 1 (Materials and Methods). In
the case of N populations, each weight wk+1,k will
converge to a nonzero value associated with Tk,
whereas all other weights will become negligible
during replay. Thus, the network's structure even-
tually encodes the order of the sequence.
21 and w1
The duration of activation in population 1, T1,
determines the equilibrium value of the synaptic
weight from population 1 to population 2, w∞
21 (Ma-
terials and Methods). For larger values of T1, LTD
lasts longer, weakening w21 (Fig. 2C). Hence, weaker
synapses are associated with longer event times.
Reciprocally, weaker synapses lead to longer acti-
vation times during replay (Section 3.2).
As the stimulus duration, T1, determines the
asymptotic synaptic weight, w∞
21(T1), there is a map-
ping T1 → w∞
21(T1) from stimulus times to the re-
sulting weights. Event timing is thus encoded in
the asymptotic values of the synaptic weights.
Fig. 2 Encoding timing in synaptic weights. A. Synap-
tic connections evolve during training. When a presy-
naptic population (1) is active and a postsynaptic pop-
ulation (2) is inactive, LTD reduces the synaptic weight
w21. When the populations (1 and 2) are co-active (over-
lap window between dashed lines), LTP increases w21.
Shortly after, global inhibition inactivates the presynap-
tic population (1), so long term plasticity ceases (see
Materials and Methods). As in the text, wi
21 denotes
the weight at the end of the i-th trial. Arrow width and
contrast correspond to synaptic weights. B. After sev-
eral training trials, the synaptic weight wi
21 converges to
a fixed point, w∞
21, whose amplitude depends on the ac-
tivation time of the presynaptic population. C. Starting
from the same initial value, w0
21(T1) = 0.25, the weight
21(T1) converges to different values, w∞
wi
21(T1), depend-
ing on the the training time, T1. Pink bar at T1=300ms
corresponds to the value used in A and B.
two populations (Fig. 2). During training, popula-
tion 1 was stimulated for T1 seconds followed by
stimulation of population 2 (Fig. 2A). The stimu-
lus was strong enough to dominate the dynamics
of the population responses (Materials and Meth-
3.2 A slow process allows precise temporal replay
We next describe how the trained network replays
sequences. The presence of a slow process, which
12001T1Aw012T10T1LTDLTPBT1firingratesynapticstrength w21timetrial 1trial 2trial 3trial 4sequencepresentationbeforeafterCduration of first stimulus (s)020.30.4timesynapticstrength w21synaptic strength w21T1timew1w0w1w∞w1w2w4w3w1(T1)w2(T1)w3(T1)w4(T1)w∞(T1)2121212121212121212121212121 Precise timing of event sequences
11
Such ramping models have previously been pro-
posed as mechanisms for time-keeping(Buonomano,
2000; Durstewitz, 2003; Reutimann et al, 2004; Kar-
markar and Buonomano, 2007; Gavornik et al, 2009).
Without such a slow process, cued activity would
result in a sequence replayed in the proper order,
but information about event timing would be lost.
For simplicity we focus on two populations, whe-
re activity of the first population represents a timed
event (Fig. 3). To simplify the analysis, we also
assumed that synaptic weights are fixed during re-
play. This assumption is not essential (Section 4.4).
After population 1 is activated with a brief cue, it
remains active due to recurrent excitation (Mate-
rials and Methods). Meanwhile, short term facili-
tation leads to an increase in the effective synaptic
strength from population 1 to population 2. Popu-
lation 2 becomes active when the input from pop-
ulation 1 crosses an activation threshold (Fig. 3A).
When both populations are simultaneously active,
a sufficient amount of global inhibition is recruited
to shut off the first population, which receives only
weak input from population 2. The second popula-
tion then remains active, as the strong excitatory
input from the first population and recurrent exci-
tation exceed the global inhibition.
The weight of the connection from population
1 to population 2 determines how long it takes
to extinguish the activity in the first population
(Fig. 3B). This synaptic weight therefore encodes
the time of this first and only event. We demon-
strate how this principle extends to multiple event
sequences in the Section 3.3. The time until the ac-
tivation of the second population decreases as the
initial synaptic weight increases, since a shorter
time is needed for facilitation to drive the input
from population 1 to the activation threshold (Fig.
3C). Note that when the baseline synaptic weight is
too weak, synaptic facilitation saturates before the
effective weight reaches the activation threshold,
and the subsequent population is never activated.
When the baseline synaptic weight is above the
activation threshold, the subsequent population is
activated instantaneously.
Fig. 3 Replay timing. A. Once population 1 is acti-
vated, the weight of the connection to population 2
slowly increases, eventually becoming strong enough to
activate this next population in the sequence. B. Ac-
tivation time T := T1 decreases monotonically with
the weight of the connection between the populations,
w := w21. For weak connections (wA = 0.33) the
synapses must be strongly facilitated to reach the thresh-
old. If the weight is larger (wB = 0.42) threshold is
reached more quickly. For weights above the threshold
(wC = 0.58 > θ), population 2 is activated immediately.
Activation will not occur when the synaptic connection
is smaller than θ/pmax. C. Activation time T1 plotted
against the initial synaptic weight, w21. For intermedi-
ate values of w21, the relationship is given by Eq. (10).
Here wA, wB, wC and T A, T B are the same as in panel B.
we assumed here to be short term facilitation, is
critical. This slow process tracks time by ramping
up until reaching a pre-determined threshold. An
event's duration corresponds to the amount of time
it takes the slow variable to reach this threshold.
12001synapseTT0Aθw21p1(t)w21p1(t)B12θimmediateactivationtimedactivationnoactivationfacilitatedsynapseθpmaxeffectivestrengthw21p1firingratetimetimeactivationtime0Cno activationwAwBimmediate activationTAactivation time (s)TBsynaptic strength1activation time (s)synaptic strengthwCθpmaxθwAwBwCTBTAvglobalinhibition 12
Alan Veliz-Cuba et al.
Therefore, long term plasticity allows for en-
coding an event time in the weight of the connec-
tions between populations in the network, while
short term facilitation is crucial for replaying the
events with the correct timing. The time of acti-
vation during cued replay will match the timing in
the training sequence as long as training drives the
synaptic weights to the value that corresponds to
the appropriate event time (Fig. 2C):
W(T ) :=
θ
pmax + (1 − pmax)e−T /τf
.
(14)
Here pmax and τf are the strength and timescale
of facilitation (Materials and Methods). An exact
match can be obtained by tuning parameters of the
long term plasticity process so the learned weight
matches Eq. (14). There is a wide range of param-
eters for which the match occurs (Materials and
Methods). We next show that the timing and or-
der of sequences containing multiple events can be
learned in a similar way.
3.3 Repeated presentation of the same sequence
produces a time-coding feedforward network
To demonstrate that the mechanism we discussed
extends easily to arbitrary sequences, we consider
a concrete sequence of four stimuli. We set the pa-
rameters of the model so that the training parame-
ters match the reactivation parameters (Eqs. 11-13
in Materials and Methods).
We trained the network using the event sequence
1-2-3-4 (Fig. 4A), repeatedly stimulating the corre-
sponding populations in succession. The duration
of each population activation was fixed across tri-
als. After each training trial, we cued the network
by stimulating the first population for a short pe-
riod of time to trigger replay (Materials and Meth-
ods). Thus, after population 1 is activated the sub-
sequent activity is governed by the network's ar-
chitecture. As our theory suggests, the cue-evoked
network activity pattern converged with training
to the stimulus-driven activity pattern (Fig. 4A).
Fig. 4 Learning and replaying event sequences. A. As
the number of training trials increases, a cue results in
an activation pattern that approaches that evoked by
the training sequence. Network architecture is reshaped
to encode the precise duration and order of events in the
sequence, with stronger feedforward connections corre-
sponding to shorter events (thicker and darker arrows
correspond to larger synaptic weights). All weights are
learned independently and training most strongly affects
the weights wi→i+1. The event times were T1 = 0.6,
T2 = 0.4, T3 = 1, and T4 = 0.5 seconds, with indices
denoting the population stimulated. The initial weights
ij = 0.025 for i (cid:54)= j and wii = 1. B. The network
were w0
in the last row of A was retrained with the sequence
T1 = 0.4, T4 = 1, T3 = 0.6, and T2 = 0.8, presented
in the order 1-4-3-2. After 10 training trials, the cued
network replays the new training sequence.
We further tested whether the same network
can be retrained to encode a sequence with a dif-
ferent order of activation (1-4-3-2) with different
event times. Fig. 4B shows that after training, the
network encodes and replays the new training se-
quence. Thus, the network architecture can be shaped
by long term plasticity to encode an arbitrary se-
training sequencetime (s)firing rateA01120101010time (s)01210Bcued activitynew training sequencetrial 1trial 2trial 3trial 101234cued activity (trial 10)12341234123412341432trained networkfiring rateretrained networkrearranged bystimulus order Precise timing of event sequences
13
Fig. 5 Effect of noise on learning and replay. The ef-
fects of adding normally distributed noise to A the ob-
served training time T := T1 (duration of first stimulus)
and B neural activity uj. Starting from a uniform dis-
tribution (p(w0
21)) and considering either a noise level
of C σ = 0.1(cid:104)T(cid:105) in the training time or D σ = 0.03 in
neural activity, the probability density function p(wi
21)
converges to the steady state distribution p(w∞
21) (this is
nearly identical to w10
21). Increasing noise in the training
times (E) or neural activity (F) widens the steady state
distribution p(w∞
21). The mean (dashed lines) and stan-
dard deviation (shaded region) of w∞
21(T ) are pictured in
panel G when noise with standard deviation σ = 0.20(cid:104)T(cid:105)
is added to the training times and H when noise with
standard deviation σ = 0.05 is added to neural activity.
As in Fig. 2C, the mean learned weight E [w∞
21] decreases
with (cid:104)T(cid:105). The mean of the replay time (dashed blue line)
and its standard deviation (shaded cyan region) are plot-
ted against (cid:104)T(cid:105) for the case of I noise in training times
and J noise in neural populations. As the mean train-
ing time (cid:104)T(cid:105) increases, so does the effect of noise on
replay time, Treplay (grey lines show the diagonal line
Treplay = T ). For a suitable choice of "corrected-for-
noise" parameters, the effect of noise on the mean re-
play time can be removed (dashed green line and green
shaded region). Insets in I and J show the root-mean-
square error in replay time as a function of training time
for the parameters obtained from the deterministic case
(blue) and corrected-for-noise parameters (green).
we explored how noise impacts the training and
recall of event sequences in our model. We exam-
ined the effects of stochasticity in event times as
well as noise in the network activity and the im-
pact such variability has on the training and recall
of event sequences. For simplicity, we focus on the
case of two populations, but our results extend to
sequences of arbitrary length.
To examine the impact of variability in stimu-
lus durations, we sampled T := T1 from a normal
distribution (Fig. 5A) with mean (cid:104)T(cid:105) = 0.5 and
variance σ2 = (cv (cid:104)T(cid:105))2, where cv = 0.1 is the co-
efficient of variation. Randomness in the observed
event time may be due to variability in the exter-
nal world, temporal limitations on sensation (Butts
et al, 2007), or other observational errors (Ma et al,
2006). We selected w0 := w0
21 from a uniform dis-
tribution on [θ/pmax, θ]. Fig. 5C shows the evolu-
tion of the probability density function of wm as
m increases (using 20,000 initial w0's). The synap-
Fig. 5 Effect of noise on learning and replay (caption
continued on next page).
quence of event times, and a brief cue evokes the
replay of the learned sequence.
4 Extensions
4.1 Effects of noise
To examine the impact of the many sources of vari-
ability in the nervous system (Faisal et al, 2008),
w10w2w0σ=0.15⟨T⟩σ=0.10⟨T⟩σ=0.20⟨T⟩θanalyticssimulations000200ww∞Tsynapticstrengthsynaptic strengthθpmaxθθpmaxθ02trained time (s)T02trainedtime (s)percentagepercentageθpmaxw∞synaptic strengthreplay time (s)02θθpmaxw∞synaptic strengthσ=0.05σ=0.03σ=0.06replay time (s)Treplay10101010percentagepercentagew10w2w0TreplaysimulationsCEGIDFHJ0202ww∞synapticstrengthsynaptic strengthθpmaxθθpmaxθ020.5RMSE⟨T⟩⟨T⟩trained time (s)⟨T⟩trained time (s)0201RMSETN(T,σ2)trial 1trial 2trial 3001TfiringratetimeABnoise in training timenoise in firing activity~T1T2T30202.1.2.2.4⟨T⟩TCVCV 14
Alan Veliz-Cuba et al.
tic weight after the ith training, wi
21, is described
by a probability density function that converges in
the limit of many training trials. The peak (mode)
of this distribution is the most likely value of the
learned synaptic weight after repeated presenta-
tion of the sequence (Fig. 5C). The variance of
the learned synaptic weight, w∞
21, increases mono-
tonically with the variance in the training time, σ2
(Fig. 5E). We explored the effects of different levels
of noise, σ = 0.1(cid:104)T(cid:105), 0.15(cid:104)T(cid:105), 0.2(cid:104)T(cid:105) (20,000 initial
conditions for each), and estimated the probability
density function of w∞ numerically (Fig. 5E). To
see the effect of noise for different mean training
durations, we estimated the mean and standard
deviation of w∞ for (cid:104)T(cid:105) ranging from 0.1s to 2s
(step size of 0.05s) using σ = 0.2(cid:104)T(cid:105) (5,000 initial
conditions for each) (Fig. 5G). For a distribution
of training times with mean (cid:104)T1(cid:105) we obtain a uni-
modal probability density for the weights, p(w21).
As in the noise-free case, the mode of this weight
distribution decreases with (cid:104)T1(cid:105). Note that the pa-
rameters given by Eq. (12), which guarantee that
training and replay time coincide in the determin-
istic case, may not be the same as the parameters
needed when noise is present. We numerically esti-
mated these parameters using Eq. (13) so the mean
training time and mean replay time coincided.
To determine how noise affects activation tim-
ing during sequence replay, we compared the mean
event time with the mean replayed time. Since the
network parameters used here are those obtained
from the noise-free case, we expect that replay times
are biased. Indeed, Fig. 5I shows that activity dur-
ing replay is slightly longer on average than the
corresponding training event. Also, the variance in
activation during replay increases with the mean
duration of the trained event. We can search nu-
merically and find a family of parameters for which
the mean activation time during replay and train-
ing coincide (Fig. 5I). Error and the coefficient of
variation (CV) in replay time increases with the
duration of the trained time (Fig. 5I, shaded re-
gion and inset).
We also estimated the mean learned synaptic
weight and its variance analytically: Since the synap-
tic weight, wi
21, evolves according to the rule
21A(T i) + C,
wi+1
21 = wi
where A(T ) := e−T γd/τw e−(γp−γd)D/τw and C :=
(1 − e−Dγp/τw )wmax, we obtain
E[wi+1
21A(T i)] + C.
21 ] = E[wi
Since wi
lows that wi
variables; then,
21 only depends on T j for j < i, it fol-
21 and A(T i) are independent random
21 ] = E[wi
21]E[A(T i)] + C,
E[wi+1
and by taking the limit i → ∞ and solving for
limi→∞ E[wi
21] we find
µw := lim
i→∞ E[wi
21] =
C
1 − µA
,
where µA := E[A(T i)]. Squaring wi+1
C gives
21 = wi
21A(T i)+
(wi+1)2 = (wi)2A2(T i) + C 2 + 2wi
21A(T i)C,
and then it similarly follows that
(cid:0)E[(wi
21)2] − E[wi
C 2σ2
A
21]2(cid:1)
σ2
w := lim
i→∞
,
=
A − µ2
A)
(1 − µA)2(1 − σ2
A := E[A(T i)2] − E[A(T i)]2. To quantify
where σ2
the average error, we computed numerically the
mean and the standard deviation of the replay time
(Fig. 5I). The root-mean-square error (RMSE) was
computed by
RMSE((cid:104)T(cid:105)) :=
E[((cid:104)T(cid:105) − Treplay)2],
(cid:113)
where the expected value is taken over the replay
time, Treplay. The coefficient of variation (CV) was
computed by
(cid:112)V ar(Treplay)
(cid:104)Treplay(cid:105)
.
CV ((cid:104)T(cid:105)) :=
To introduce neural noise (Fig. 5B), we added
white noise to the rate equations of the populations
during training and replay so that
Precise timing of event sequences
15
duj =
1
τ
[−uj + ϕ (Ij(t) + Ij,syn − θ)] dt + σdξj,
where dξj is a standard white noise process with
variance σ2. The analysis of the effect of noise in
population activity is similar to the analysis per-
formed on stimulus duration noise, the only dif-
ference being that the noise level was σ = 0.03 in
panel Fig. 5D, σ = 0.03, 0.05, 0.06 in panel Fig. 5F,
and σ = 0.05 in panels Fig. 5H and 5J. After re-
peated presentation of a sequence, the distribution
of the learned synaptic weights converged (Fig. 5D).
The variance of the synaptic weight increased mono-
tonically with the variance of the noise (Fig. 5F),
and the mean weight decreased monotonically with
the event time (Fig. 5H). Since we used the pa-
rameters found from the noise-free case, we ex-
pect some bias in replay time. After training, the
replayed event times are shorter than the corre-
sponding events in the training sequence (Fig. 5J),
and the effect is much more significant than the
lengthening of times due to observation noise. This
systematic bias in the replayed time error is due
to the saturating nature of the time-tracking pro-
cess, short term facilitation (Markram et al, 1998).
Input to the second population remains close to
threshold for longer periods of time for longer trained
times, leading to more frequent noise-induced thresh-
old crossings (Gardiner, 2004). However, parame-
ters for which mean event time and mean replay
time coincide can be found numerically (Materi-
als and Methods). Interestingly, the CV attains a
minimum in the vicinity of the timescale of short
term facilitation (τf = 1s), suggesting the network
best encodes events on the timescale of the slow
process (Fig. 5J, lower inset). This principle holds
across a range of short term facilitation timescales
τf (Supplementary Material, Fig. 2).
4.2 Alternative firing rate response function
ϕγ(x) =
We next show that the mechanism for learning the
precise timing of an event sequence does not de-
pend on the particulars of the model. In previous
where the parameter γ determines the steepness.
Note that in the absence of other population in-
Fig. 6 Alternative firing rate function. The mechanism
for sequence learning and replay also works for other
firing rate functions (see Materials and Methods). A.
Using a nonlinear and nonsaturating response function
ϕγ (u) =
γu, long term plasticity still results in the
coding of a training sequence as synaptic weights. B.
After training, the cued network replays the training
sequence similarly to the replay seen in Fig. 4A.
√
sections, we used a Heaviside step function as the
firing rate function and chose short term facilita-
tion as the slow, time-tracking process. However,
the principles we have identified do not depend on
these specific choices. More general circuit models
of slowly ramping units can learn and replay timed
event sequences. The elements needed to learn and
replay precise time durations are a chain of slow
accumulators and a learning rule which modifies
weights to precisely time a threshold-crossing event
for each accumulator (Supplementary Material, Fig.
3).
Precise replay relies on a threshold-crossing pro-
cess which occurs as long as each population is
bistable, having only low and high activity states
rather than graded activity. Indeed, detailed spik-
ing models (Litwin-Kumar and Doiron, 2012) and
experimental recordings (Major and Tank, 2004)
suggest that cell assemblies can exhibit multiple
stable states. To test whether our conclusions hold
with different firing rate response functions, we re-
place the Heaviside function with the nonsaturat-
ing function (Fourcaud-Trocm´e et al, 2003) (Fig.
6A)
(cid:40)
0
√
if x < 0,
if x ≥ 0,
γx
time (s)01210training sequencecued activity (trial 10)firing rate0nonlinear responsefunctionpre.post.φAB 16
puts,
τ
duj
dt
= −uj + ϕγ (wjjuj − θ) ,
which has steady states determined by the equa-
tion u = ϕγ (wjju − θ). One of the stable steady
states is u = 0 and there is a positive stable steady
state, u∗, which is the largest root of the quadratic
equation u2 − γwjju + γθ = 0. For simplicity, we
normalize γ and self-excitation so that the stable
states are u = 0 and u = 1; namely, we consider
γ(cid:48) := γ/u2∗ and w(cid:48)
jj := wjju∗.
Since a population is activated when its input
reaches the threshold due to short term facilita-
tion, the derivations that led to Eqs. 9 and 10 are
still valid for this model. However, the activation
of a neuronal population (uj → 1) was delayed
since ϕγ has finite slope. This delay was negligible
when firing rate response was modeled by a Heavi-
side function, and activation was instantaneous. To
take this delay into account, we can modify Eqs. 9
and 10 to obtain
W(T ) :=
pmax + (1 − pmax)e−(T−d)/τf
θ
,
and
T (w) := d + τf ln
(cid:18) pmax − 1
pmax − θ/w
(cid:19)
,
where d is a heuristic correction parameter to ac-
count for the time it takes for uj to approach 1.
Following the arguments that led to Eq. (12),
we were able to derive constraints on parameters
to ensure the correct timings are learned:
1 − e
(M−1)γd+γp
−D(
1 + (M−1)γd
τw
γp
= τf ,
τw
M γd
wmax = θ/pmax,
)
(15)
e(γd−γp)D/τw =
pmax − 1
pmax
ed/τf .
For simulations we used the parameters M = 1.5,
d = 30ms, γ = 3, and estimated γd, γp, and wmax
using Eq. (15) (we then normalized ϕγ and wjj to
Alan Veliz-Cuba et al.
make 0 and 1 the stable firing rates). As in the pre-
vious simulations, the number of presentations was
m = 10; the durations of the events were 0.6s, 0.4s,
1s, and 0.5s for events 1, 2, 3 and 4, respectively.
Network architecture converges, and the replayed
activity matches the order and timing of the train-
ing sequence (Fig. 6B).
4.3 An alternative slow process: spike rate
adaptation
We also examined whether spike frequency adapta-
tion, i.e. a slow decrease in firing rates in response
to a fixed input to a neural population, can play
the role of a slow, time tracking process (Benda
and Herz, 2003), instead of short term facilitation.
In contrast to the case of short term facilitation,
adaptation causes the effective input from one pop-
ulation to decrease over time.
In this case population activity was modeled by
τ
τa
τs
duj
dt
daj
dt
dsj
dt
τ
dv
dt
= −uj + ϕ(wjjuj + sj − θ − Lv − aj),
= −aj + buj,
= −sj +
wjkuk,
= −v + ϕ
Zkuk − θv
N(cid:88)
(cid:32) N(cid:88)
k(cid:54)=j
k=1
(cid:33)
,
where aj denotes the adaptation level of popula-
tion j, τa is the time scale of adaptation, and b
is the adaptation strength. Feedback between pop-
ulations was assumed to be slower than feedback
within a population; thus, the total input for pop-
ulation j was split into self-excitation (wjjuj), and
synaptic inputs from other populations (sj) which
evolved on the time scale τs. Note that in the limit
τs → 0, synapses are instantaneous.
For a suitable choice of parameters, global in-
hibition tracks activity faster than excitation be-
tween populations. Then, when a population be-
comes inactive due to adaptation, the level of global
Precise timing of event sequences
17
Fig. 7 Alternative slow process based on spike fre-
quency adaptation. A. When a population is active,
LTP increases the synaptic weight w11. After becoming
inactive, LTD decreases w11 (see Materials and Meth-
ods). B. Synaptic weight wi
11 is updated after the ith
training trial. After several trials, wi
11 converges to a
fixed point w∞
11 that depends on the activation time
of the first population. C. Once population 1 is ac-
tive, adaptation builds up until overcoming self excita-
tion. This will occur when the effective strength, w11 −
a1(t) ([excitation]−[adaptation]), crosses below θ + L
([threshold]+[inhibition]). D. Activation time T := T1
increases with synaptic weight w := w11. For strong self
excitation (wA) adaptation takes longer to shut off the
first population, so the next population in the sequence
is activated later. Weaker self excitation (wB) will result
in quicker extinction of activity in the first population,
result in the next population activating sooner. For self
excitation below θ + L ([threshold] + [inhibition]), the
first population will inactivate immediately, resulting in
immediate activation of the next population. If self ex-
citation of the first population is greater than θ + L + b
([threshold] + [inhibition] + [maximum adaptation]), it
will remain active indefinitely, and the subsequent pop-
ulation is never activated. E. When the parameters of
the long term plasticity process and the replay process
match, the network can learn the precise timing of se-
quences.
analogous to wjk with the additional assumption
that since wjj represented the synaptic weight within
a population, it could not decrease below a cer-
tain value wmin. Also, the parameters for long term
plasticity within a population are allowed to be dif-
ferent from the parameters for long term plasticity
between populations.
The learning rule was then
τw
dwjj
dt
= − γ(cid:48)
− γ(cid:48)
d(wjj − wmin)uj(t − D(cid:48))(1 − uj(t))
p(wjj − w(cid:48)
max)uj(t − D(cid:48))uj(t).
When the population was activated (u1(t) ≈
1) for t ∈ [0, T1] (Fig. 7A), the changes in the
weight w11 were governed by the piecewise differ-
ential equation
Fig. 7 Alternative slow process based on spike fre-
quency adaptation (caption continued on next page).
inhibition decreases, allowing subsequent popula-
tions to become active. This means the weight of
self excitation can encode timing. Thus, in this
setup we modeled long term plasticity within a
population as well. The learning rule for wjj was
dw11
dt
=
t (cid:54)∈ [D(cid:48), T1 + D(cid:48)]
0,
max − w11) t ∈ [D(cid:48), T1]
(w(cid:48)
(w11 − wmin) t ∈ [T1, T1 + D(cid:48)].
γ(cid:48)
p
τw
− γ(cid:48)
d
τw
time (s)01210training sequencecued activity (trial 10)firing rateθ+Lw11-a1(t)12001T1A12T10T1LTDLTPBT1firingratesynapticstrength w11timetrial 1trial 2trial 3trial 4sequencepresentationbeforeaftertimesynapticstrength w11timew011w111w111w011w∞11w111w211w311w411w11-a1(t)C12001synapseTT0Deffectivestrengthfiringratetimetimeactivationtime0no activationimmediate activationTAactivation time (s)TBsynaptic strength1wAwBwCvw11-a1(t)Eθ+Lθ+L+b 18
Alan Veliz-Cuba et al.
The following equation relates the synaptic weight
at the end of a presentation, w11(Ttot), to the synap-
tic weight at the beginning of the presentation,
w11(0):
w11(Ttot) =w11(0)e−T1γ(cid:48)
maxe−D(cid:48)γ(cid:48)
p/τw e(γ(cid:48)
d/τw (1 − e−(T1−D(cid:48))γ(cid:48)
d)D(cid:48)/τw
p/τw )
p−γ(cid:48)
+ w(cid:48)
+ wmin(1 − e−D(cid:48)γ(cid:48)
d/τw ).
This recurrence relation between the weight at the
i + 1st stimulus, wi+1
11 , and the weight at the ith
stimulus, wi
11 converges to the
limit (Fig. 7B)
11, implies that wi
w∞
11 =
w(cid:48)
maxe−D(cid:48)γ(cid:48)
d/τw (1 − e−(T1−D(cid:48))γ(cid:48)
d)D(cid:48)/τw
1 − e−T1γ(cid:48)
p/τw e(γ(cid:48)
wmin(1 − e−D(cid:48)γ(cid:48)
p−γ(cid:48)
p−γ(cid:48)
d/τw )
d)D(cid:48)/τw
p/τw e(γ(cid:48)
1 − e−T1γ(cid:48)
+
p/τw )
.
Thus, for each stimulus duration a unique synaptic
weight is learned. Also, as shown in previous sec-
tions, w21 will converge to a fixed value and w12
is weakened. Note that in this case timing will be
encoded as the weight of self-excitation, and the
order will be encoded as the weights between pop-
ulations.
da1
dt
du1
dt
1, and we obtain τ
During replay a cue activates population 1, u1 =
= −u1 + ϕ(w11− θ− a− L),
= −a + b. Population 1 will become inac-
τa
tive (u1 ≈ 0) when w11 − a1(t) decreases to θ + L.
Then, the next population will become active due
to the decrease in global inhibition and the remain-
ing feedback from the first population due to the
slower dynamics of feedback between populations
(Fig. 7C).
The precise time of activation can be controlled
by tuning the synaptic weight w11. Furthermore,
since the activation time satisfies w11 − a1(T ) =
θ + L, we have a formula that relates the synaptic
weight to the activation time (Fig. 7D)
w11 = W(T ) := θ + L + b(1 − e−T /τa ).
When self excitation is too strong, adaptation will
not affect the activity of the first population and
deactivation will never occur. On the other hand,
when self excitation is too weak, activation is not
sustained and the population will be shut off imme-
diately. To guarantee correct time coding and de-
11(T ) and W(T ) had to be approximately
coding, w∞
equal for all T . The appropriate parameters could
not be found in closed form, so we again resorted
to finding them numerically using Eq. (13).
For simulations we used the parameters Zk =
0.6, L = 0.8, γp = 3750, γd = 100, γ(cid:48)
p = 267.86,
γ(cid:48)
d = 7500, wmax = 1.5, w(cid:48)
max = 4.1312, and
wmin = 1.3488. As in the previous simulations, the
number of presentations was m = 10; the duration
of the events were 0.6s, 0.4s, 1s, and 0.5s for events
1, 2, 3 and 4, respectively.
This idea generalizes to any number of events
and populations. Timing is encoded in the weight
of the excitatory self-connections within a popula-
tion, while sequence order is encoded in the weight
of the connections between populations. Moreover,
for a range of network parameters, the duration
of the sequences during training and reactivation
coincide (Materials and Methods). Presenting the
event sequence used in Fig. 4A, the network can
learn the precise timing and order of the events
(Fig. 7E).
4.4 Incorporating long timescale plasticity
Thus far, we have assumed that during sequence
replay, synaptic connections remained unchanged
(Section 3.3). However, if synaptic changes occur
on the same timescale as the network's dynamics,
and are allowed to act during replay, the network's
architecture can become unstable. This problem
can be solved by assuming that synaptic weights
change slowly compared to network dynamics (Al-
berini, 2009).
We therefore extended our model so that long
term plasticity occurs on more realistic timescales.
The impact of rate covariation on the network's
synaptic weights was modeled by a two step pro-
cess: (a) rate correlation detection, which occurs on
the timescale of seconds and (b) translation of this
information into an actual weight change, which
Precise timing of event sequences
19
occurs on the timescale of minutes or hours. The
initial and immediate signal shaped by the firing
rates of pre- and postsynaptic neural populations
was modeled by intermediate variables we refer to
as proto-weights (Gavornik et al, 2009) (Fig. 8A).
Changes to the actual synaptic weights occur on
a much longer timescale, and slowly converged to
values determined by the proto-weights (Fig. 8B).
Introducing proto-weights leads to repeated reen-
forcement of learned activity patterns making them
robust to spontaneous network activations. The
model takes the form
= −uj + ϕ (Ij(t) + Ij,syn − θ) ,
= 1 − pj + (pmax − 1)uj,
(cid:32) N(cid:88)
= −v + ϕ
(cid:33)
Zkuk − θv
,
k=1
= −γdwjkuk(t − D)(M − uj(t))
+ γp(wmax − wjk)uk(t − D)uj(t), j (cid:54)= k,
= wjk(t − Dp) − Wjk(t),
(16)
τ
τf
duj
dt
dpj
dt
τ
dv
dt
τw
dwjk
dt
τI
dWjk
dt
where
Ij,syn = Wjjuj +
N(cid:88)
k(cid:54)=j
Wjkpkuk − Lv,
τI is the time scale of the actual weights, and Dp is
a time delay in the process of transforming changes
in the proto-weights to changes in the actual weights.
Since τw represents the timescale of changes in
proto-weights, the timescale τI represents the timescale
of actual synaptic weight changes. The process of
translating a coincidence in firing rates into a weight
change can be much longer than detecting the coin-
cidence, and we can thus take τI much larger than
τw (Markram and Tsodyks, 1996).
It was still possible to analyze the model given
in Eqs. (16), but the results were less transpar-
ent. During training, the proto-weights satisfied
Eq. (5). As the number of training trials increased,
the proto-weights converged to the limit given in
Fig. 8 A two-stage learning rule. A. Proto-weights track
the signaled change to synaptic connectivity brought
about by firing rate covariance. Actual weights evolve on
a slower time scale and hence do not track changes in the
proto-weights immediately. B. Actual synaptic weights
approach proto-weight values on a longer timescale, so
both eventually converge to the same value.
Eq. (6). The actual weights also slowly approached
the same limit; that is, Wjk converged to the limit
given in Eq. (6). During replay, the proto-weights
evolved according to Eqs. 16, but due to the delay
Dp and the time scale τI , the actual weights re-
mained unchanged and replay occurred accurately
as in the Section 2.6.
Modeling long term plasticity as a two-stage
process allowed us to incorporate more realistic de-
tails: We were able to assume that synaptic con-
nections are plastic during replay, as well as during
training. Replaying the sequence of neural popula-
tion activations evokes long term plasticity signals
through the proto-weights, and alterations to ac-
tual synaptic weights do not take place until af-
ter the epochs of neural activity. During reacti-
vation, the actual weights already equal the val-
ues necessary to elicit the timed event sequence.
Then, the weakening of the connection due to LTD
will precisely equal the strengthening due to LTP,
resulting in no net change in the actual connec-
tion weight. As long as the time scale of synaptic
weight consolidation is much larger than the se-
quence timescale, long term plasticity during re-
firingratesynapticstrength 01proto-weight05time (s)actual weightproto-weightactual weight30time (min)synapticstrength AB 20
Alan Veliz-Cuba et al.
play reinforces the learned network of weights that
is already present.
5 Discussion
Sequences of sensory and motor events can be en-
coded in the architecture of neuronal networks.
These sequences can then be replayed with cor-
rect order and timing when the first element of the
sequence is presented, even in the absence of any
other sensory input. Experimental evidence shows
that after repeated presentations of a cued training
sequence, the presentation of the cue alone trig-
gers a temporal pattern of activity similar to that
evoked by the training stimulus (Eagleman and
Dragoi, 2012; Xu et al, 2012; Shuler and Bear,
2006; Gavornik and Bear, 2014). Our goal here
was to provide a biologically plausible mechanism
that could govern the learning of precisely timed
sequences.
5.1 Learning both the precise timing and order of
sequences
We demonstrated how a complex learning task can
be accomplished by combining two simple mech-
anisms. First, the timing of a single event can be
represented by a slowly accumulating positive feed-
back process (Buonomano, 2000; Durstewitz, 2003;
Reutimann et al, 2004; Shea-Brown et al, 2006;
Karmarkar and Buonomano, 2007; Gavornik et al,
2009; Simen et al, 2011). Second, rate dependent
long term plasticity can reshape synaptic weights
so that the order and precise timing of events in
a sequence is encoded by the network's architec-
ture (Amari, 1972; Abbott and Blum, 1996).
To make the problem analytically tractable, we
considered an idealized model of neural population
firing rates, long term plasticity, and short term fa-
cilitation. This allowed us to obtain clear relation-
ships between parameters of the time-tracking pro-
cess (short term facilitation) and the learning pro-
cess (long term plasticity). The assumptions about
model structure and parameters that were essential
for sequence learning could be explicitly described
in this model. Similar conditions were required for
learning in more realistic models, which incorpo-
rated the long timescale of LTP/LTD.
A novel feature of our network model is that
long term plasticity influences the length of time a
neural population is active. Typical computational
models of sequence learning employ networks of
neurons (Jun and Jin, 2007; Fiete et al, 2010; Brea
et al, 2013) or populations (Abbott and Blum, 1996)
that are each active for equal amounts of time dur-
ing replay. However, sensory and motor processes
can be governed by networks whose neurons have
a fixed stimulus tuning (Xu et al, 2012; Gavornik
and Bear, 2014). Therefore, a sequence of events of
varying time lengths should be represented by neu-
ral populations that are each active for precisely
the length of time of the corresponding event. Our
model demonstrates that this can be achieved us-
ing rate-based long term plasticity.
5.2 Experimental predictions
The general mechanisms we described here imply
a number of experimentally testable features of
the neural substrates of the learning and recall of
event sequences. Our analysis of the impact of noise
on time encoding demonstrates a relationship be-
tween the dynamical mechanism for encoding and
error statistics. If time interval estimation is ac-
complished through a slow process that saturates
toward a threshold, then the relative error of an
interval in the sequence should increase with inter-
val length, and average interval estimate should be
shortened during replay (Fig. 5J). This provides an
innate mechanistic explanation for underestimate
of time duration, contributing to existing literature
that has found environmental conditions that can
lead to such systematic errors (Morrone et al, 2005;
Terao et al, 2008). When time is marked by a slow
process that scales linearly in time, average dura-
tion estimates will be close to the true estimate
and relative error will be invariant to trained du-
rations (Supplementary Material, Fig. 3). We sug-
gest a way in which average interval estimates may
Precise timing of event sequences
21
be shorter than the trained interval, if the trained
interval is longer than the timescale of the slow
process encoding it. If the slow process that reads
out the stored time grows linearly or exponentially,
average interval estimates may be nearly equal or
longer than the true duration (Supplementary Ma-
terial, Fig. 3).
The mechanics of sequence learning could be
understood further by examining the development
of sequence replay accuracy with the number of
trainings. Errors in sequence recall will tend to be
greatest after very few trainings (Fig 4A). Corti-
cal recordings reveal that, indeed, the correlation
between replayed activity and training sequence
evoked activity increases with the number of se-
quence exposures (Eagleman and Dragoi, 2012).
Additionally, our model suggests that errors in the
replayed time of each event's beginning will build
serially, if they rely upon a sequence of popula-
tion activations. This means that errors in the total
run time of a sequence will increase with sequence
duration, as in Hass et al (2008). However, errors
in the individual estimates of each event duration
will not depend on their placement in the sequence.
Such errors should decrease at a similar rate across
all individual events, as suggested by our analysis
in Materials and Methods. This prediction could
be tested experimentally by examining how sub-
jects' individual event duration estimates depend
on the event's position in a learned sequence.
Lastly, we predict that event sequences can be
learned through rate correlation based synaptic plas-
ticity acting on connection between stimulus-tuned
populations. This mechanism could be probed ex-
perimentally in a number of ways. First, if neu-
ral activity underlying sequence learning were be-
ing recorded electrophysiologically (Xu et al, 2012;
Gavornik and Bear, 2014), subsequent experiments
could be performed to see if electrically stimulating
neurons out of sequence could disrupt learned se-
quence memory. This would provide evidence that
plasticity mechanisms that result from neural ac-
tivity are involved in the consolidation of sequence
memory. Inactivating populations in the sequence
could also disrupt the replay of the remainder of
the sequence, supporting our network chain model
of sequence learning. For example, optogenetic meth-
ods could be used to inactivate a large fraction of
cells that respond to one of the events in sequence.
If such a disruption were to terminate sequence re-
play, this would be strong evidence for the events
being represented by a chain of active populations.
Furthermore, long term plasticity processes could
be disrupted through the local injection of transla-
tional inhibitors (Alberini, 2009). If this leads to a
reduction of sequence memory robustness, it would
constitute strong evidence for the importance of
long term plasticity in the local circuit for sequence
memory formation.
5.3 Comparison to previous models of interval
timing
Several previous theoretical studies have proposed
neural mechanisms for the learning and recall of
timed events. Models capable of representing serial
event order have utilized individual units that are
oscillators (Brown et al, 2000) or bistable popula-
tions (Grossberg and Merrill, 1992). Recent studies
have found that continuous temporal trajectories
can be learned in networks of chaotic elements by
training weights to downstream neurons that con-
stitute a linear readout (Buonomano and Maass,
2009; Hennequin et al, 2014). A complementary
approach has been used to infer time by fitting a
maximum likelihood model to the rates and phases
of spiking neurons in hippocampal networks (It-
skov et al, 2011). Our approach is most similar
to previous studies that utilize discrete popula-
tions or neurons to represent serial order (Gross-
berg and Merrill, 1992; Abbott and Blum, 1996;
Fiete et al, 2010; Brea et al, 2013). Namely, we
assume that the memory of each individual event
duration is learned in parallel with the others as in
Fiete et al (2010), in contrast to the serial build-
ing of chains demonstrated in the model of Jun
and Jin (2007). Reset models of sequence replay
are supported by comparisons of human behavioral
data to models that mark event durations using a
clock that is reset after each event (McAuley and
22
Alan Veliz-Cuba et al.
Jones, 2003), suggesting errors are made locally
in time, rather than accumulated event-to-event.
We have extended previous work by developing a
mechanism for altering the activation time of each
unit in the sequence. This learning process is dis-
tinct from the approach outlined in Buonomano
and Maass (2009); Hennequin et al (2014), since
it solely trains the recurrent architecture between
populations encoding time; tuning of a downstream
readout is unnecessary.
5.4 Internal tuning of long term plasticity
parameters
There is a large set of parameters for which the
network can be trained to accurately replay train-
ing sequences. While some parameter tuning is re-
quired, in simple cases we could find these param-
eters explicitly. In all cases appropriate parame-
ters could be obtained computationally using gra-
dient descent. In biological systems plasticity pro-
cesses could be shaped across many generations by
evolution, or within an organism during develop-
ment. Indeed, recent experimental evidence sug-
gests that networks are capable of internally tuning
long term plasticity responses through metaplastic-
ity (Abraham, 2008). For instance, NMDA receptor
expression can attenuate LTP (Huang et al, 1992;
Philpot et al, 2001), while metabotropic glutamate
receptor activation can prime a network for future
LTP (Oh et al, 2006). We note that such mech-
anisms would affect the timescale and features of
LTP/LTD, not the synaptic weights themselves.
5.5 Models that utilize ramping processes with
different timescales
Our proposed mechanism relies on a ramping pro-
cess that evolves on the same timescale as the train-
ing sequence. Short term facilitation (Markram et al,
1998) as well as rate adaptation (Benda and Herz,
2003) can fulfill this role. However, other ramp-
ing processes that occur at the cellular or network
level are also capable of marking time (Supplemen-
tary Material, Fig. 3). For instance, slow synap-
tic receptor types such as NMDA can slowly in-
tegrate sensory input (Wang, 2002), resulting in
population firing rate ramping similar to experi-
mental observations in interval timing tasks (Xu
et al, 2014). Were we to incorporate slow recur-
rent excitatory synapses in this way, the duration
of represented events would be determined by the
decay timescale of NMDA synapses. Alternatively,
we could have also employed short term depres-
sion as the slow process in our model. Mutual in-
hibitory networks with short term depression can
represent dominance time durations that depend
on the network's inputs, characteristic of percep-
tual rivalry statistics (Laing and Chow, 2002). This
relationship between population inputs and popu-
lation activity durations could be leveraged to rep-
resent event times in sequences.
Events that occur on much shorter or longer
timescales than those we explored here could be
marked by processes matched to those timescales.
For instance, fast events may be represented sim-
ply using synaptic receptors with rapid kinetics,
such as AMPA receptors (Clements, 1996). AMPA
receptor states evolve on the scale tens of millisec-
onds, which would allow representation of several
fast successive events. However, we would expect
a lower bound on the duration of an event repre-
sented by this mechanism, given by the neuronal
membrane time constant (Dayan and Abbott, 2001).
Slow events could also be represented by a long
chain of sub-populations, each of which is activated
for a shorter amount of time than the event. In the
context of our model, this would mean each pop-
ulation would contain sub-populations connected
as a feedforward chain (Goldman, 2009). Networks
of cortical neurons can have different subpopula-
tions with distinct sets of timescales, due to the
variety of ion channel and synaptic receptor kinet-
ics (Ulanovsky et al, 2004; Bernacchia et al, 2011;
Pozzorini et al, 2013; Costa et al, 2013). This reser-
voir of timescales could be utilized to learn events
whose timings span several orders of magnitude.
Precise timing of event sequences
23
5.6 Learning the repeated appearance of an event
We only considered training sequences in which no
event appeared more than once (e.g. 1-2-3-4). If
events appear multiple times (e.g. 1-2-1-4), then a
learned synaptic weight (e.g., w21) would be weak-
ened when the repeated event appears again. This
can be resolved by representing each event repe-
tition by the activation of a different subpopula-
tion of cells. There is evidence that this occurs in
hippocampal networks responding to spatial nav-
igation sequences on a figure eight track (Griffin
et al, 2007). Even for networks where each stimu-
lus activates a specific population, sequences with
repeated stimuli could be encoded in a deeper layer
of the underlying sensory or motor system. The
same idea can be used to create networks that
can store several different event sequences contain-
ing the same events (e.g. 1-2-3-4; 2-4-3-1; 4-3-2-1).
If multiple sequences begin with the same event
(e.g. 1-2-3-4; 1-3-2-4), evoking the correct sequence
would require partial stimulation of the sequence
(e.g. 1-2 or 1-3). Networks would then be less likely
to misinterpret one learned sequence for another
sequence with overlapping events (Abbott and Blum,
1996).
5.7 Feedback correction in learned sequences
We emphasize that we did not incorporate any
mechanisms for correcting errors in timing during
replay. However, this could easily be implemented
by considering feedback control via a stimulus that
activates the population that is supposed to be
active, if any slippage in event timing begins to
occur. This assumes there is some external sig-
nal indicating how accurately the sequence is be-
ing replayed. For instance, human performance of
a piece of music relies on auditory feedback sig-
nals that are used by the cerebellum to correct
motor errors (Zatorre et al, 2007; Kraus and Chan-
drasekaran, 2010). If feedback is absent or is ma-
nipulated, performance deteriorates (Finney and
Palmer, 2003; Pfordresher, 2003). Similar princi-
ples seem to hold in the replay of visual sequences.
Gavornik and Bear (2014) showed that portions
of learned sequence are replayed more accurately
when preceded by the correct initial portion of the
learned sequence. We could incorporate feedback
into our model by providing external input to the
network at several points in time, not just the ini-
tial cue stimulus.
5.8 Conclusions
Overall, our results suggest that a precisely timed
sequence of events can be learned by a network
with long term synaptic plasticity. Sequence play-
back can be accomplished by a ramping process
whose timescale is similar to the event timescales.
Trial-to-trial variability in training and neural ac-
tivity will be inherited by the sequence representa-
tion in a way that depends on the learning process
and the playback process. Therefore, errors in se-
quence representation provide a window into the
neural processes that represent them. Future ex-
perimental studies of sequence recall that statis-
tically characterize these errors will help to shed
light on the neural mechanisms responsible for se-
quence learning.
Acknowledgements We thank Jeffrey Gavornik helpful
comments. Funding was provided by NSF-DMS-1311755
(Z.P.K.); NSF/NIGMS-R01GM104974 (A.V-C. and K.J.);
and NSF-DMS-1122094 (A.V-C. and K.J.).
References
Abbott LF, Blum KI (1996) Functional signif-
icance of long-term potentiation for sequence
learning and prediction. Cereb Cortex 6(3):406 --
416
Abraham WC (2008) Metaplasticity:
tuning
synapses and networks for plasticity. Nat Rev
Neurosci 9(5):387
Alberini C (2009) Transcription factors in long-
term memory and synaptic plasticity. Physiol
Rev 89(1):121 -- 145
24
Alan Veliz-Cuba et al.
Amari SI (1972) Learning patterns and pattern se-
quences by self-organizing nets of threshold ele-
ments. IEEE Trans Comput 21(11):1197 -- 1206
Benda J, Herz AV (2003) A universal model
for spike-frequency adaptation. Neural Comput
15(11):2523 -- 2564
Bernacchia A, Seo H, Lee D, Wang XJ (2011) A
reservoir of time constants for memory traces in
cortical neurons. Nat Neurosci 14(3):366 -- 72
Bienenstock EL, Cooper LN, Munro PW (1982)
Theory for the development of neuron selectiv-
ity: orientation specificity and binocular interac-
tion in visual cortex. J Neurosci 2(1):32 -- 48
Bliss TVP, Lømo T (1973) Long-lasting potentia-
tion of synaptic transmission in the dentate area
of the anaesthetized rabbit following stimulation
of the perforant path. J Physiol 232(2):331 -- 356
Brea J, Senn W, Pfister JP (2013) Matching recall
and storage in sequence learning with spiking
neural networks. J Neurosci 33(23):9565 -- 9575
Brown GD, Preece T, Hulme C (2000) Oscillator-
based memory for serial order. Psychological re-
view 107(1):127
Bueti D, Buonomano DV (2014) Temporal per-
learning. Timing & Time Perception
ceptual
2(3):261 -- 289
Buhusi CV, Meck WH (2005) What makes us tick?
functional and neural mechanisms of interval
timing. Nat Rev Neurosci 6(10):755 -- 65
Buonomano DV (2000) Decoding temporal infor-
mation: A model based on short-term synaptic
plasticity. J Neurosci 20(3):1129 -- 1141
Buonomano DV, Maass W (2009) State-dependent
computations: spatiotemporal processing in cor-
tical networks. Nature Reviews Neuroscience
10(2):113 -- 125
Burgess N, Hitch GJ (1999) Memory for serial or-
der: A network model of the phonological loop
and its timing. Psychological Review 106(3):551
Butts DA, Weng C, Jin J, Yeh CI, Lesica NA,
Alonso JM, Stanley GB (2007) Temporal pre-
cision in the neural code and the timescales of
natural vision. Nature 449(7158):92 -- 95
Clements J (1996) Transmitter timecourse in the
synaptic cleft: its role in central synaptic func-
tion. Trends Neurosci 19(5):163 -- 171
Clopath C, Busing L, Vasilaki E, Gerstner W
(2010) Connectivity reflects coding: a model of
voltage-based stdp with homeostasis. Nat Neu-
rosci 13(3):344 -- 352
Conway CM, Christiansen MH (2001) Sequential
learning in non-human primates. Trends Cogn
Sci 5(12):539 -- 546
Costa RP, Sjostrom PJ, Van Rossum MC (2013)
Probabilistic inference of short-term synaptic
plasticity in neocortical microcircuits. Frontiers
in computational neuroscience 7
Dayan P, Abbott LF (2001) Theoretical neuro-
science. Cambridge, MA: MIT Press
Dudek SM, Bear MF (1992) Homosynaptic long-
term depression in area ca1 of hippocampus and
effects of n-methyl-d-aspartate receptor block-
ade. Proc Natl Acad Sci USA 89(10):4363 -- 4367
Durstewitz D (2003) Self-organizing neural inte-
grator predicts interval times through climbing
activity. J Neurosci 23(12):5342 -- 5353
Eagleman SL, Dragoi V (2012) Image sequence re-
activation in awake v4 networks. Proc Natl Acad
Sci USA 109(47):19,450 -- 19,455
Faisal AA, Selen LP, Wolpert DM (2008) Noise in
the nervous system. Nat Rev Neurosci 9(4):292 --
303
Fiete IR, Senn W, Wang CZ, Hahnloser RH (2010)
Spike-time-dependent plasticity and heterosy-
naptic competition organize networks to produce
long scale-free sequences of neural activity. Neu-
ron 65:563 -- 576
Finney S, Palmer C (2003) Auditory feedback and
memory for music performance: Sound evidence
for an encoding effect. Mem Cognition 31(1):51 --
64
Fourcaud-Trocm´e N, Hansel D, Van Vreeswijk C,
Brunel N (2003) How spike generation mecha-
nisms determine the neuronal response to fluc-
tuating inputs. J Neurosci 23(37):11,628 -- 11,640
Gardiner CW (2004) Handbook of Stochastic
Methods. Springer
Gavornik JP, Bear MF (2014) Learned spatiotem-
poral sequence recognition and prediction in pri-
mary visual cortex. Nat Neurosci 17(5):732 -- 737
Precise timing of event sequences
25
Gavornik JP, Shuler MGH, Loewenstein Y, Bear
MF, Shouval HZ (2009) Learning reward timing
in cortex through reward dependent expression
of synaptic plasticity. Proc Natl Acad Sci USA
106(16):6826 -- 6831
Gerstner W, Kistler WM (2002) Mathematical for-
mulations of hebbian learning. Biol Cybern 87(5-
6):404 -- 15
Gjorgjieva J, Clopath C, Audet J, Pfister
JP (2011) A triplet spike-timing -- dependent
plasticity model generalizes the Bienenstock --
Cooper -- Munro rule to higher-order spatiotem-
poral correlations. Proc Natl Acad Sci USA
108(48):19,383 -- 19,388
Goldman MS (2009) Memory without feedback in
a neural network. Neuron 61(4):621 -- 634
Graupner M, Brunel N (2012) Calcium-based plas-
ticity model explains sensitivity of synaptic
changes to spike pattern, rate, and dendritic lo-
cation. Proc Natl Acad Sci USA 109(10):3991 --
3996
Griffin AL, Eichenbaum H, Hasselmo ME (2007)
Spatial representations of hippocampal ca1 neu-
rons are modulated by behavioral context in a
hippocampus-dependent memory task. J Neu-
rosci 27(9):2416 -- 2423
Grossberg S, Merrill JW (1992) A neural network
model of adaptively timed reinforcement learn-
ing and hippocampal dynamics. Cognitive brain
research 1(1):3 -- 38
Gutig R, Aharonov R, Rotter S, Sompolinsky H
(2003) Learning input correlations through non-
linear temporally asymmetric hebbian plasticity.
The Journal of neuroscience 23(9):3697 -- 3714
Haider B, Hausser M, Carandini M (2013) Inhibi-
tion dominates sensory responses in the awake
cortex. Nature 493(7430):97 -- 100
Hass J, Blaschke S, Rammsayer T, Herrmann JM
(2008) A neurocomputational model for optimal
temporal processing. Journal of computational
neuroscience 25(3):449 -- 464
Hausser M, Roth A (1997) Estimating the time
course of the excitatory synaptic conductance in
neocortical pyramidal cells using a novel volt-
age jump method. The Journal of neuroscience
17(20):7606 -- 7625
Hennequin G, Vogels TP, Gerstner W (2014) Op-
timal control of transient dynamics in balanced
networks supports generation of complex move-
ments. Neuron 82(6):1394 -- 1406
Hikosaka O, Nakamura K, Sakai K, Nakahara H
(2002) Central mechanisms of motor skill learn-
ing. Current opinion in neurobiology 12(2):217 --
222
Huang YY, Colino A, Selig DK, Malenka RC
(1992) The influence of prior synaptic activity
on the induction of long-term potentiation. Sci-
ence 255(5045):730 -- 3
Itskov V, Curto C, Pastalkova E, Buzs´aki G
(2011) Cell assembly sequences arising from
spike threshold adaptation keep track of time in
the hippocampus. The Journal of Neuroscience
31(8):2828 -- 2834
Ivry RB, Schlerf JE (2008) Dedicated and intrin-
sic models of time perception. Trends Cogn Sci
12(7):273 -- 280
Jenkins I, Brooks D, Nixon P, Frackowiak R, Pass-
ingham R (1994) Motor sequence learning: a
study with positron emission tomography. The
Journal of Neuroscience 14(6):3775 -- 3790
Jun JK, Jin DZ (2007) Development of neu-
ral circuitry for precise temporal sequences
through spontaneous activity, axon remodeling,
and synaptic plasticity. PLoS ONE 2(8):e723
Kandel ER (2001) The molecular biology of mem-
ory storage: a dialogue between genes and
synapses. Science 294(5544):1030 -- 1038
Karmarkar UR, Buonomano DV (2007) Timing in
the absence of clocks: encoding time in neural
network states. Neuron 53(3):427 -- 438
Kempter R, Gerstner W, van Hemmen JL (1999)
Hebbian learning and spiking neurons. Phys Rev
E 59:4498 -- 4514
Kleinfeld D (1986) Sequential state generation by
model neural networks. Proc Natl Acad Sci USA
83(24):9469 -- 9473
Ko H, Hofer SB, Pichler B, Buchanan KA,
Sjostrom PJ, Mrsic-Flogel TD (2011) Functional
specificity of local synaptic connections in neo-
cortical networks. Nature 473(7345):87 -- 91
26
Alan Veliz-Cuba et al.
Kok P, Jehee JF, de Lange FP (2012) Less is more:
expectation sharpens representations in the pri-
mary visual cortex. Neuron 75(2):265 -- 270
Kraus N, Chandrasekaran B (2010) Music training
for the development of auditory skills. Nat Rev
Neurosci 11(8):599 -- 605
Laing CR, Chow CC (2002) A spiking neuron
model for binocular rivalry. J Comput Neurosci
12(1):39 -- 53
Litwin-Kumar A, Doiron B (2012) Slow dynam-
ics and high variability in balanced cortical net-
works with clustered connections. Nat Neurosci
15(11):1498 -- 505
Lundstrom B (2015) Modeling multiple time scale
firing rate adaptation in a neural network of local
field potentials. J Comput Neurosci 38(1):189 --
202
Ma WJ, Beck JM, Latham PE, Pouget A (2006)
Bayesian inference with probabilistic population
codes. Nat Neurosci 9(11):1432 -- 1438
Major G, Tank D (2004) Persistent neural activity:
prevalence and mechanisms. Curr Opin Neuro-
biol 14(6):675 -- 84
Malenka RC, Bear MF (2004) Ltp and ltd: an em-
barrassment of riches. Neuron 44(1):5 -- 21
von der Malsburg C (1973) Self-organization of ori-
entation sensitive cells in the striate cortex. Ky-
bernetik 14(2):85 -- 100
Markram H, Tsodyks M (1996) Redistribution of
synaptic efficacy between neocortical pyramidal
neurons. Nature 382(6594):807 -- 810
Markram H, Wang Y, Tsodyks M (1998) Differ-
ential signaling via the same axon of neocorti-
cal pyramidal neurons. Proc Natl Acad Sci USA
95(9):5323 -- 5328
McAuley JD, Jones MR (2003) Modeling effects of
rhythmic context on perceived duration: a com-
parison of interval and entrainment approaches
to short-interval timing. Journal of Experimen-
tal Psychology: Human Perception and Perfor-
mance 29(6):1102
Meyer T, Olson CR (2011) Statistical learning of
visual transitions in monkey inferotemporal cor-
tex. Proceedings of the National Academy of Sci-
ences 108(48):19,401 -- 19,406
Miller KD (1994) A model for the development
of simple cell receptive fields and the ordered
arrangement of orientation columns through
activity-dependent competition between on-and
off-center inputs. J Neurosci 14:409 -- 441
Morrone MC, Ross J, Burr D (2005) Saccadic eye
movements cause compression of time as well as
space. Nat Neurosci 8(7):950 -- 4
Nabavi S, Fox R, Proulx CD, Lin JY, Tsien RY,
Malinow R (2014) Engineering a memory with
ltd and ltp. Nature 511(7509):348 -- 52
Oh MC, Derkach VA, Guire ES, Soderling TR
(2006) Extrasynaptic membrane trafficking reg-
ulated by glur1 serine 845 phosphorylation
primes ampa receptors for long-term potentia-
tion. J Biol Chem 281(2):752 -- 8
Oja E (1982) Simplified neuron model as a princi-
pal component analyzer. J Math Biol 15(3):267 --
273
Perin R, Berger TK, Markram H (2011) A synaptic
organizing principle for cortical neuronal groups.
Proc Natl Acad Sci USA 108(13):5419 -- 5424
Pfister JP, Gerstner W (2006) Triplets of spikes in
a model of spike timing-dependent plasticity. J
Neurosci 26(38):9673 -- 9682
Pfordresher PQ (2003) Auditory feedback in mu-
sic performance: Evidence for a dissociation of
sequencing and timing. Journal of Experimen-
tal Psychology: Human Perception and Perfor-
mance 29(5):949
Philpot BD, Sekhar AK, Shouval HZ, Bear MF
(2001) Visual experience and deprivation bidi-
rectionally modify the composition and func-
tion of NMDA receptors in visual cortex. Neuron
29(1):157 -- 169
Pozzorini C, Naud R, Mensi S, Gerstner W
(2013) Temporal whitening by power-law adap-
tation in neocortical neurons. Nature neuro-
science 16(7):942 -- 948
Rao RP, Sejnowski TJ (2001) Spike-timing-
dependent hebbian plasticity as temporal differ-
ence learning. Neural Comput 13(10):2221 -- 2237
Reutimann J, Yakovlev V, Fusi S, Senn W
(2004) Climbing neuronal activity as an event-
based cortical representation of time. J Neurosci
Precise timing of event sequences
27
ralistic viewing. Neuron 55(3):465 -- 78
Wang XJ (2002) Probabilistic decision making by
slow reverberation in cortical circuits. Neuron
36(5):955 -- 68
Xu M, Zhang SY, Dan Y, Poo Mm (2014) Repre-
sentation of interval timing by temporally scal-
able firing patterns in rat prefrontal cortex. Proc
Natl Acad Sci USA 111(1):480 -- 485
Xu S, Jiang W, Poo Mm, Dan Y (2012) Activity
recall in a visual cortical ensemble. Nat Neurosci
15:449 -- 455
Zatorre RJ, Chen JL, Penhune VB (2007) When
the brain plays music: auditory -- motor interac-
tions in music perception and production. Nat
Rev Neurosci 8(7):547 -- 558
24(13):3295 -- 3303
Sakai K, Hikosaka O, Miyauchi S, Takino R, Sasaki
Y, Putz B (1998) Transition of brain activation
from frontal to parietal areas in visuomotor se-
quence learning. The Journal of Neuroscience
18(5):1827 -- 1840
Shea-Brown E, Rinzel J, Rakitin BC, Malapani
C (2006) A firing rate model of parkinsonian
deficits in interval timing. Brain Res 1070(1):189
-- 201
Shuler MG, Bear MF (2006) Reward timing in the
primary visual cortex. Science 311(5767):1606 --
1609
Simen P, Balci F, Cohen JD, Holmes P, et al (2011)
A model of interval timing by neural integration.
J Neurosci 31(25):9238 -- 9253
Sjostrom PJ, Turrigiano GG, Nelson SB (2001)
timing, and cooperativity jointly de-
synaptic plasticity. Neuron
Rate,
termine cortical
32(6):1149 -- 1164
Song S, Sjostrom PJ, Reigl M, Nelson S, Chklovskii
DB (2005) Highly nonrandom features of synap-
tic connectivity in local cortical circuits. PLoS
Biol 3(3):e68
Takeuchi T, Duszkiewicz AJ, Morris RG (2014)
The synaptic plasticity and memory hypothesis:
encoding, storage and persistence. Philos T Roy
Soc Lond B 369(1633):20130,288
Terao M, Watanabe J, Yagi A, Nishida S (2008)
Reduction of stimulus visibility compresses ap-
parent time intervals. Nat Neurosci 11(5):541 -- 2
Tsodyks M, Pawelzik K, Markram H (1998) Neural
networks with dynamic synapses. Neural Com-
put 10(4):821 -- 835
Ulanovsky N, Las L, Farkas D, Nelken I (2004)
Multiple time scales of adaptation in auditory
cortex neurons. The Journal of Neuroscience
24(46):10,440 -- 10,453
Wang D, Arbib M (1990) Complex temporal se-
quence learning based on short-term memory. P
IEEE 78(9):1536 -- 1543
Wang X, Wei Y, Vaingankar V, Wang Q, Koepsell
K, Sommer FT, Hirsch JA (2007) Feedforward
excitation and inhibition evoke dual modes of
firing in the cat's visual thalamus during natu-
|
1507.00235 | 1 | 1507 | 2015-07-01T13:52:07 | Energy-efficient neuromorphic classifiers | [
"q-bio.NC",
"cs.NE"
] | Neuromorphic engineering combines the architectural and computational principles of systems neuroscience with semiconductor electronics, with the aim of building efficient and compact devices that mimic the synaptic and neural machinery of the brain. Neuromorphic engineering promises extremely low energy consumptions, comparable to those of the nervous system. However, until now the neuromorphic approach has been restricted to relatively simple circuits and specialized functions, rendering elusive a direct comparison of their energy consumption to that used by conventional von Neumann digital machines solving real-world tasks. Here we show that a recent technology developed by IBM can be leveraged to realize neuromorphic circuits that operate as classifiers of complex real-world stimuli. These circuits emulate enough neurons to compete with state-of-the-art classifiers. We also show that the energy consumption of the IBM chip is typically 2 or more orders of magnitude lower than that of conventional digital machines when implementing classifiers with comparable performance. Moreover, the spike-based dynamics display a trade-off between integration time and accuracy, which naturally translates into algorithms that can be flexibly deployed for either fast and approximate classifications, or more accurate classifications at the mere expense of longer running times and higher energy costs. This work finally proves that the neuromorphic approach can be efficiently used in real-world applications and it has significant advantages over conventional digital devices when energy consumption is considered. | q-bio.NC | q-bio |
Energy-efficient neuromorphic classifiers
Daniel Mart´ı ∗ †, Mattia Rigotti ‡ † Mingoo Seok § and Stefano Fusi †
∗ D´epartement d’´Etudes Cognitives, ´Ecole Normale Sup´erieure - PSL Research University, Paris, France.
Institut nationale de la sant´e et de la recherche m´edicale,
France.,‡Physical Sciences Department, IBM T. J. Watson Research Center, Yorktown Heights, NY 10598,†Center for Theoretical Neuroscience, Columbia University, New
York, USA, and §Department of Electrical Engineering, Columbia University, New York, USA
Neuromorphic engineering combines the architectural and computational
principles of systems neuroscience with semiconductor electronics, with the
aim of building efficient and compact devices that mimic the synaptic and
neural machinery of the brain. Neuromorphic engineering promises ex-
tremely low energy consumptions, comparable to those of the nervous sys-
tem. However, until now the neuromorphic approach has been restricted to
relatively simple circuits and specialized functions, rendering elusive a direct
comparison of their energy consumption to that used by conventional von
Neumann digital machines solving real-world tasks. Here we show that a re-
cent technology developed by IBM can be leveraged to realize neuromorphic
circuits that operate as classifiers of complex real-world stimuli. These cir-
cuits emulate enough neurons to compete with state-of-the-art classifiers.
We also show that the energy consumption of the IBM chip is typically 2 or
more orders of magnitude lower than that of conventional digital machines
when implementing classifiers with comparable performance. Moreover,
the spike-based dynamics display a trade-off between integration time and
accuracy, which naturally translates into algorithms that can be flexibly
deployed for either fast and approximate classifications, or more accurate
classifications at the mere expense of longer running times and higher en-
ergy costs. This work finally proves that the neuromorphic approach can be
efficiently used in real-world applications and it has significant advantages
over conventional digital devices when energy consumption is considered.
neuromorphic electronic hardware VLSI technology neural networks classifica-
tion
Abbreviations: SVM: support vector machine — SV: support vector — RCN: ran-
domly connected neuron
Introduction Recent developments in digital technology and
machine learning are enabling computers to perform an in-
creasing number of tasks that were once solely the domain
of human expertise, such as recognizing a face in a picture
or driving a car in city traffic. These are impressive achieve-
ments, but we should keep in mind that the human brain
carries out tasks of such complexity using only a small frac-
tion of the energy needed by conventional computers, the dif-
ference in energy consumption being often of several orders
of magnitude. This suggests that one way to reduce energy
consumption is to design machines whose architecture takes
inspiration from the biological brain, an approach that was
proposed by Carver Mead in the late 1980s [1] and that is
now known as “neuromorphic engineering”. Mead’s idea was
to use very-large-scale integration (VLSI) technology to build
electronic circuits that mimic the architecture of the nervous
system. The first electronic devices inspired by this concept
were analog circuits that exploited the subthreshold properties
of transistors to emulate the biophysics of real neurons. Nowa-
days the term “neuromorphic” refers to any analog, digital,
or hybrid VLSI system whose design principles are inspired
by those of biological neural systems [2].
Neuromorphic hardware has convincingly demonstrated
its potential for energy efficiency, as proven by devices that
consume as little as a few picojoules per neural event (spike)
[3, 4, 5]. These devices contain however a relatively small
number of elements (neurons and synapses) and they can typ-
ically perform only simple and specialized tasks, making it
1–12
difficult to directly compare their energy consumption to that
of conventional digital machines.
The situation has changed recently with the development
by IBM of the TrueNorth processor, a neuromorphic device
that implements enough artificial neurons to perform com-
plex real-world tasks, like large-scale pattern classification [6].
Here we show that a pattern classifier implemented on the
IBM chip can achieve performances comparable to those of
state-of-the-art conventional devices based on the von Neu-
mann architecture. More importantly, our chip-implemented
classifier uses 2 or more orders of magnitude less energy than
current digital machines performing the same classification
tasks. These results show for the first time the deployment
of a neuromorphic device able to solve a complex task, while
meeting the claims of energy efficiency contented by the neu-
romorphic engineering community for the last few decades.
Results
We chose pattern classification as an example of a complex
task because of the availability of well-established bench-
marks. A classifier takes an input, like the image of a hand-
written character, and assigns it to one among a set of discrete
classes, like the set of digits. To train and evaluate our clas-
sifiers we used three different datasets consisting of images of
different complexity (see Fig. 1a).
We start by describing the architecture of the classifier
that we plan to implement on the neuromorphic chip. The
classifier is a feed-forward neural network with three layers of
neurons, and it can be simulated on a traditional digital com-
puters. We will call this network the ‘neural classifier’ to dis-
tinguish it from its final chip implementation, which requires
adapting the architecture to the connectivity constraints im-
posed by the hardware. The neural classifier also differs from
the final hardware implementation in that it employs neu-
rons with a continuous activation function, whereas the IBM
neuromorphic chip emulates spiking neurons. Despite the dif-
ferences, the functionality of the neural classifier and its final
chip implementation is approximately the same, as we show
below. We list the procedure for adapting the architecture of
the neural classifier into its chip implementation as a contribu-
tion in its own right, since it can be directly extended for the
implementation of generic neural systems on other hardware
substrates.
Architecture of the neural classifier Figure 1b illustrates the
three-layer neural classifier. The first layer encodes the pre-
processed input and projects to the neurons in the interme-
diate layer through connections with random weights. Each
of these Randomly Connected Neurons (RCNs) receives there-
fore a synaptic current given by a randomly weighted sum of
the inputs, which the RCNs transform into activation levels
in a non-linear way—in our case, through a linear rectifica-
tion function: f (x) = x if x > 0, and 0 otherwise. The
combination of a random mixing of the inputs together with
a non-linear input-output transformation efficiently expands
the dimensionality of the resulting signal (see e.g. [7, 8, 9]),
thereby increasing the chances that downstream neurons can
Figure 1. Datasets, architecture of the classifier, architecture of a single core, and chip implementation. a Samples of the three
datasets used to evaluate the performance of our classifier. MNIST contains handwritten digits (10 classes); MNIST-back-image contains the digits of the MNIST dataset
on a background patch extracted randomly from a set of 20 images downloaded from the Internet; LATEX contains distorted versions of 293 characters used in the LATEX
document preparation system. For more details about the datasets, see Methods. b Architecture of the neural network classifier. The images to classify are preprocessed (see
Methods) and represented as patterns of activity of a population of Nin input neurons (left, black dots). These input neurons send random projections to a second layer of
N Randomly Connected Neurons (RCNs) (green circles), which transform nonlinearly their synaptic inputs into firing activity. The activity of the RCNs is then read
out by a third layer of neurons, each of which is trained to respond to only one class (red circles). c Architecture of a single core in the chip. Horizontal lines represent inputs,
provided by the axons of neurons that project to the core. Vertical lines represent the dendrites of the neurons in the core (one dendrite per neuron). Active synapses are shown
as dots in a particular axon-dendrite junction. The synaptic input collected by the dendrites is integrated and transduced into spike activity at the soma (filled squares on top).
The spikes emitted by the neuron are sent via its axon to a particular input line, not necessarily on the same core. Blue lines represent the flow of input and output signals.
The panel includes an example of internal connection: the upmost axon carries the output activity of the leftmost neuron in the core (other connections are left unspecified).
d Implementation of the neural network classifier in a chip with connectivity constraints. The input is fed into all the cores in the RCN layer (shaded blue), whose neurons
project to the input lines of readout cores (shaded yellow) in a one-to-one manner (green curves). The outputs of the readout units are combined together off-line to generate
the response of the output neuron (shaded red). See the main text for the description of the different modules.
discriminate signals belonging to distinct classes. This dis-
crimination is carried out by a set of output units in the last
layer, which compute a weighted sum of the RCNs activity.
The weights are trained so that each output unit responds to
one separate class (one-vs-all code). Details are given in the
Methods. Once the network is trained, a class is assigned to
each input patterns according to which output unit exhibits
the highest activation.
Chip implementation of the neural classifier We implemented
the neural classifier on the IBM neuromorphic chip described
in [10, 6]. The first step of the conversion of the abstract neu-
ral classifier to an explicit chip implementation is the transfor-
mation of the input patterns into a format that is compatible
with the spike-based coding of the TrueNorth system. For this
we simply employ a firing rate coding and convert the integer
value of every input component to a spike train with a propor-
tional number of spikes, a prescription that is commonly used
in neurocomputational models such as the Neural Engineering
framework [11]. Specifically, input patterns are preprocessed
and formatted into 256-dimensional vectors representing the
firing activity of the input layer (the same preprocessing step
was applied in the neural classifier, see Methods). This vector
of activities is then used to generate 256 regular-firing spike
trains that are fed into a set of cores with random and sparse
connectivity. This set of cores constitutes the RCN layer. Like
in the neural classifier, the neurons in the RCN layer receive
synaptic inputs that consist of randomly weighted combina-
tions of the input, and transform their synaptic inputs into
firing activity according to a nonlinear function. On the chip
this function is given by the neuronal current-to-rate transduc-
tion, which approximates a linear-rectification function [12].
Discriminating the inputs coming from the RCN layer re-
quires each output unit to read from the whole layer of RCNs,
which in our implementation contains a number of neurons N
that can be as large as 214. Moreover, all the readout connec-
titions have to be set at the weights computed by the training
procedure. These requirements exceed the constraints set by
the chip design, in terms of the maximal number of both in-
coming and outgoing connections per neuron, as well as the
resolution and the freedom with which synaptic weights can
In this paragraph we will present a set of prescrip-
be set.
2
tions that will allow us to circumvent these limitations, and
successfully instantiate our neural classifier on the IBM sys-
tem. The prescriptions we are presenting are specific to the
TrueNorth architecture, but the types of constraints that they
solve are shared by any physical implementation of neural
systems, whether it is biological or electronic. It is therefore
instructive to discuss the constraints and the prescriptions to
obviate them in detail, as they can be easily extended to other
more generic settings.
1. Constraints on connectivity. The IBM chip is organized in
cores, each of which contains 256 integrate-and-fire neurons
and 256 input lines that intersect with one another forming
a crossbar matrix of programmable synapses (Fig. 1c). Each
neuron can connect to other neurons by projecting its axon
(output) to a single input line, either on the same core or
on a different core. With this hardware design the maximum
number of incoming connections per neuron, or fan in, is 256.
Likewise, the maximal number of outgoing connections per
neuron, or fan out, is 256, each of which are restricted to
target neurons within a single core.
2. Constraints on synaptic weight precision. Synapses can be
either inactive or active. The weight of an active synapse can
be selected from a set of four values given by signed integers
of 9-bit precision. These values can differ from neuron to neu-
ron. Which of the four values is assigned to an active synapse
depends on the input line: all synapses on the same input line
are assigned an index that determines which of the four val-
ues is taken by each synapse (e.g. if the index assigned to the
input line is 2, all synapses on the input line take the second
value of the set of four available synaptic weights, which may
differ from neuron to neuron).
The design constraints that we just described can be over-
come with the following set of architectural prescriptions.
P1. Overcoming the constraints on connectivity. We intro-
duced an intermediate layer of neurons, each of which inte-
grates the inputs from 256 out of the total N RCNs. Accord-
ingly, the firing rates of these intermediate neurons represent
a 256/N portion of the total input to an output unit. These
partial inputs can then be combined by a downstream neu-
ron, which will have the same activity as the original output
9876543210inputoutput2randomconnectionsreadoutconnectionsnonlinearunits(RCNlayer)RCNlayerreadoutoutputinput+++adbcnaxons(inputs)nneurons(outputs)MNISTMNIST-back-imageLATEX unit. If the total number of the partial inputs is larger than
the total number of incoming connections of the neurons that
represent the output units (in our case 256), the procedure
can be iterated by introducing additional intermediate layers.
The final tree will contain a number of layers that scales only
logarithmically with the total number of RCNs. For simplic-
ity we did not implement this tree on chip and we summed
off-chip the partial inputs represented by the firing activity
of the readout neurons. Notice also that this configuration
requires readout neurons to respond approximately linearly
to their inputs, which can be easily achieved by tuning read-
out neurons to operate in the linear regime of their current-
to-rate transduction function (i.e., the regime in which their
average input current is positive). This procedure strongly
relies on the assumption that information is encoded in the
firing rates of neurons;
if the spiking inputs happen to be
highly synchronized and synchronization encodes important
information, this approach would not work.
P2. Overcoming the constraints on synaptic weight preci-
sion. Reducing the weight precision after learning usually
only causes moderate drops in classification performance. For
example, in the case of random uncorrelated inputs, the scal-
ing properties of the capacity of the classifier (i.e., number
of classes that can be correctly classified) remain unchanged,
even when the number of states of the synaptic weights is
reduced to two [13]. Instead, the performance drop is catas-
trophically larger when the weight precision is limited also
during learning [14, 15] and in some situations the learning
problem becomes NP-complete [16] In our case the readout
weights are determined off-chip, using digital conventional
computers that operate on 64 bit numbers, and then quan-
tized in the chip implementation. The performance drop is
almost negligible for a sufficient number of synaptic levels.
In our case we quantized the readout weights of the original
classifier on an integer scale between −28 and 28. Each quan-
tized weight was then implemented as the sum of four groups
of 6 synaptic contacts, where each contact in the group can
either be inactive (value 0) or activated at one of the 6 values:
±1,±2,±4. The multiplicity of this decomposition (19 can be
for instance decomposed as (4) + (1 + 4) + (1 + 4) + (1 + 4) or
(2)+(2+4)+(2+4)+(1+4)) is resolved by choosing the decom-
position that is closest to a balanced assignment of the weights
across the 4 groups (e.g. 19 = (4) + (1 + 4) + (1 + 4) + (1 + 4)).
This strategy requires that each original synapse be repre-
sented by 24 synapses. We implemented this strategy by
replicating each readout neuron 24 times and by distributing
each original weight across 24 different dendritic trees. These
synaptic inputs are then summed together by the off-line sum-
mation of all readout neuron activities that correspond to the
partial inputs to a specific output unit (see Methods for de-
tails). A similar strategy can be used to implement networks
with synaptic weights that have even a larger number of lev-
els and the number of additional synapses would scale only
logarithmically with the total number of synaptic levels that
is required. However, it is crucial to limit individual synapses
to low values, in order to avoid synchronization between neu-
rons. This is why we limited to 4 the maximum synaptic value
of individual synapses of the chip.
Classification performance and speed-accuracy trade-off Our
neuromorphic classifier implemented on the TrueNorth chip
was emulated on a simulator developed by IBM. As the
TrueNorth chip is entirely digital, the simulator reproduces
exactly the behavior of the chip [10].
In Fig. 2a we show
the dynamics of two typical runs of the simulator classifying
images from the MNIST-back-image dataset. Upon image
presentation, the RCNs in the intermediate layer start inte-
Figure 2. The neuromorphic classifier in action. a Spikes emit-
ted by readout neurons during an easy (top) and a difficult (bottom) classification,
after removing the trend caused by the intrinsic constant currents. Each curve corre-
sponds to the readout output associated with the digit indicated by the color code.
Samples are drawn from the MNIST-back-image dataset. b Test error as a function
of classification time (i.e., the time over which spikes are integrated) and energy. The
error is averaged over the first 1000 test samples of the MNIST (red) and MNIST-
back-image (blue) datasets. Each dashed horizontal line indicates the best test error
achieved with support vector classifiers for a given dataset, based on the evaluation
of the whole test set. c Classification times for different thresholds in spike differ-
ence (as indicated in the legend), for the MNIST and MNIST-back-image datasets.
For each threshold we plot all classification times (thin lines) as well as the sample
mean (shorter ticks on top). The performances associated with each threshold are
indicated in the y-axis. When the threshold in spike difference is infinite (black), the
classification is assessed at t = 500 ms (i.e., there is no stopping criterion). In all
panels the chip uses N = 16384 RCNs.
grating the input signal (not shown) and, a few tens of mil-
liseconds later, they start emitting spikes, which are passed
to the readout neurons. The figure shows the total number of
spikes emitted by the readout neurons since input activation,
after subtracting the overall activity trend caused by baseline
activity.
For simple classifications, in which the input is easily rec-
ognizable, the readout neuron associated with the correct class
is activated in less than 100 ms (Fig. 2a, top). More diffi-
cult cases require the integration of spikes over longer time
intervals, as the average synaptic inputs to different readout
neurons can be very similar (Fig. 2a, bottom). This sug-
gests that the performance of the classifier, as measured by
the classification error rate on the test set, should improve
with longer integration intervals. This trade-off between speed
and performance is illustrated in Fig. 2b, which shows the
classification performance versus elapsed time for the MNIST
and MNIST-back-image datasets. The performance increases
monotonically with time until it saturates in about half a
second, with a highest performance of 97.27% for MNIST
(98.2% with 10-fold bagging), and 77.30% for MNIST-back-
image. These performances are not too far from the best
classification results achieved so far: 99.06% for MNIST (us-
ing maxout networks on the permutation invariant version of
the MNIST dataset, which does not exploit any prior knowl-
edge about the two-dimensional structure of the patterns [17])
and 77.39% for MNIST-back-image (with support vector clas-
sifiers [18], although methods combining deep nets, feature
3
05010000.51performance(%)energyconsumption(mJ)MNISTMNIST-back-imageb0250500time(ms)69.977.987.997.20200400performance(%)classificationtime∞80402010spikedifferenceMNISTc59.667.675.078.40200400classificationtimeMNIST-back-image05000123456789−25002500500time(ms)emittedspikes,detrendeda learning, and feature selection can achieve performances as
high as 87.75% [19]).
Energy-speed-accuracy trade-off As just discussed, accuracy
has a cost in term of energy because longer integration times
entail more emitted spikes per classification and a larger base-
line energy costs, which in our case is the dominant contri-
bution to the total energy consumption. We estimated the
energy consumption as described in section 5 and we found
that the energy per classification never exceeds 1 mJ for our
network configuration. With the energy needed to keep lit a
100 W light bulb for a second, one could perform 105 classi-
fications, which is equivalent to around one classification per
second uninterruptedly for almost one day. Notice that this
estimate is based on a classification that lasts 0.5 s and, there-
fore, does not take into account the fact that most patterns are
correctly classified in a significantly shorter time (see Fig. 2a,
top). If the integration and emission of spikes is stopped as
soon as one of the output units is significantly more active
than the others, then the average energy consumption can be
strongly reduced. The criterion we used to decide when to
stop the integration of spikes (and thus the classification) was
based on the spikes emitted by the readout units. Specifically,
we monitored the cumulative activity of each output unit by
counting all the spikes emitted by the corresponding readout
neurons. We stopped the classification when the accumulated
activity of the leading unit exceeded that of the second unit
by some threshold. The decision was the class associated with
the leading output unit.
In Fig. 2c we show the performances and the correspond-
ing classification times for several thresholds. Low thresholds
allow for faster yet less accurate classifications. In both the
MNIST and MNIST-back-image datasets, the patterns that
require long classifications times are rare. While the perfor-
mance barely changes for large enough thresholds, the average
classification time can be substantially reduced by lowering
the threshold. For example, for the MNIST dataset the clas-
sification time drops by a factor of 5 (from 500 ms to 100 ms)
and, accordingly, so does the energy consumption (from 1 mJ
to 0.2 mJ). Faster classifications are also possible by increas-
ing either the average firing rate or the total number of RCNs,
both of which entail an increase in energy consumption, which
might be partially or entirely compensated by the decrease in
the classification time. These expedients will speed up the
integration of spike-counts and, as a result, the output class
will be determined faster.
In all cases both the energy cost and the classification per-
formance increase with the total number of emitted spikes or,
equivalently, with integration time, if the average firing rate
is fixed. This is a simple form of a more general energy-speed-
accuracy trade-off, a phenomenon that has been described in
several biological information-processing systems (e.g. [20]),
and that can confer great functional flexibility to our classi-
fier. One advantage of basing the computation on a temporal
accumulation of spikes is that the classifier can be interrupted
at any time at the cost of reduced performance, but with-
out compromising its function. This is in stark contrast to
some conventional clock-based centralized architectures whose
mode of computation crucially relies on the completion of en-
tire monolithic sets of instructions. We can then envisage uti-
lization scenarios where a spiking-based chip implementation
of our classifier is required to flexibly switch between precise
long-latency classifications (like, e.g., those involving the iden-
tification of targets of interest) and rapid responses of limited
accuracy (like the quick avoidance of imminent danger).
Notice that both the simulated and implemented net-
works, although entirely feed-forward, exhibit complex dy-
4
namics leading to classification times that depend on the dif-
ficulty associated with the input. This is because neurons are
spiking and the final decision requires some sort of accumu-
lation of evidence. When a stimulus is ambiguous, the units
representing the different decisions receive similar inputs and
the competition becomes harder and longer. This type of be-
havior is also observed in human brains [21].
We will now focus on the comparison of energy consump-
tion and performance between the neuromorphic classifier and
more conventional digital machines.
Energy consumption and performance: comparison with con-
ventional digital machines We compared both the classifica-
tion performance and the energy consumption of our neuro-
morphic classifier to those obtained with conventional digi-
tal machines implementing Support Vector Machines (SVMs).
SVMs offer a reasonable comparison because they are among
the most successful and widespread techniques for solving
machine-learning problems involving classification and regres-
sion [22, 23, 24], and because they can be efficiently imple-
mented on digital machines.
To better understand how the energy consumption scales
with the complexity of the classification problem, it is useful
to summarize how SVMs work. After training, SVMs classify
an input pattern according to its similarity to a set of tem-
plates, called the support vectors, which are determined by the
learning algorithm to define the boundaries between classes.
The similarity is expressed in terms of the scalar product be-
tween the input vectors and the support vectors. As argued
above, we can improve classification performance by embed-
ding the input vectors in a higher-dimensional space before
classifying them. In this case SVMs evaluate similarities by
computing classical scalar products in the higher-dimensional
space. One of the appealing properties of SVMs is that there
is no need to compute explicitly the transformation of inputs
into high-dimensional representations. Indeed, one can skip
this step and compute directly the scalar product between the
transformed vectors and templates, provided that one knows
how the distances are distorted by the transformation. This is
known as the “kernel trick” because the similarities in a high-
dimensional space can be computed and optimized over with
a kernel function applied to the inputs. Interestingly, the ker-
nel associated with the transformation induced by the RCNs
of our neural classifier can be computed explicitly in the limit
of a large number of RCNs [18]. This is also the kernel that
we used to compare the performance of SVMs against that of
our neural classifier.
Unfortunately, classifying a test input by computing its
similarity to all support vectors becomes unwieldy and com-
putationally inefficient for large datasets, as the number of
support vectors typically scales linearly with the size of the
training set in many estimation problems [25]. This means
that the number of operations to perform, and hence the en-
ergy consumption per classification, also scales with the size of
the training set. This makes SVMs and kernel methods com-
putationally and energetically expensive in many large-scale
tasks. In contrast, our neural network algorithm evaluates a
test sample by means of the transformation carried out by
the RCNs.
If the RCN layer comprises N neurons and the
input dimension is Nin, evaluating the output of a test sam-
ple requires O(Nin · N ) synaptic events. Thus for large sam-
ple sizes, evaluating a test sample in the network requires far
fewer operations than when using the “kernel trick”, because
the number is effectively independent of the size of training
set (cfr. [26, 27]). Systems such as ours may therefore display
considerable energy advantages over SVMs when datasets are
large.
Figure 3. Energy-accuracy trade-off. a,c Dependence of the classification accuracy on the number of Randomly Connected Neurons (RCNs) in the neural
classifier and on the number of support vectors (SVs) in the SVC. Panel a shows this dependence for the MNIST dataset, and panel b for the MNIST-back-image. As the
number of RCNs increases, the classifier becomes more accurate at the cost of higher energy consumptions (b,d). The energy consumption is based on the average time it
takes to the neural classifier to perform the classification (see Fig. 2c). We also show the performance achieved by three different implementations of support vector classifiers
(legend code: SVC, libsvm; rSVC, reduced primal; SVC; SVCperf, cutting plane subspace pursuit). The algorithms rSVC, SVCperf minimize the number of support vectors
(SVs) with respect to the optimal value and reduce, therefore, the energy consumption levels at test time. The number of SVs used by the standard algorithm (libsvm), on
the other hand, can go beyond the optimal value by reducing sufficiently the soft-margin parameter and pushing the classifier to overfit the data.
In all cases, the energy
consumption increases linearly with the number of SVs, as the number of operations per classification at test time scales linearly with the number of SVs. The vertical thin
lines indicate the abscissa at best performance for the IBM chip (red) and SVM implementations (black). For reference we indicate the best performance achieved by the
chip with a horizontal dashed line. The horizontal arrow indicates the reduction in energy consumption that would be attained if the efficiency of digital machines reached the
theoretical lower bound estimated by [28]. The relation between number of SVs and energy consumption was determined by simulating the i7 Intel chip running a program that
implements an SVM at test time. b Same as a, but on the MNIST-back-image dataset. In both cases our neuromorphic classifier exhibits an energy cost per classification
that is orders of magnitude smaller.
In Fig. 3 we compare the energy consumption and per-
formance of the neuromorphic classifier to those of an SVM
implemented on a conventional digital machine. More specif-
ically, we estimated the energy expenditure of a digital SVM
using a simulator of the Intel i7 processor, which was the ma-
chine with the best energy performance among those that we
simulated (see Methods section 5 and Discussion). The en-
ergy cost per support vector per pattern was estimated to be
around 5.2 µJ, a quantity that is not far above what is con-
sidered as a lower bound on energy consumption for digital
machines [28]. For both the neuromorphic classifier and the
digital SVM we progressively increased the performance of
the classifiers by increasing the number of RCNs (in the case
of the neuromorphic classifier), and by varying the number of
support vectors (in the case of the SVM), see Figures 4a,c. For
the SVM we tried three different algorithms to minimize the
number of support vectors and hence the energy consump-
tion (for more details, see caption of Fig. 3 and Methods).
For the IBM chip we estimated the energy consumption both
in the case in which we stopped the classifications with the
criterion described in the previous section and in the case in
which the classification time was fixed at 500 ms (see Fig. 7 in
Suppl. Info.). In both cases the energy consumption is signif-
icantly lower for the neuromorphic classifier, being in the for-
mer case approximately 2 orders of magnitude smaller for both
the MNIST and the MNIST-back-image datasets, while still
achieving comparable maximal performances (Fig. 4b,d).
Scalability The MNIST dataset only has 10 output classes. We
wondered whether the advantage of the neuromorphic clas-
sifier in terms of energy consumption is preserved when the
number of classes increases and the classification task becomes
more complex. To study how the energy consumption scales
with the number of classes we used the LATEX dataset, which
contains 293 classes of distorted characters. We progressively
increased the number of classes to be learned and classified
and we studied the performance and the energy consumption
of both the digital implementation of the SVM and the neuro-
morphic classifier. Specifically, given a number of classes that
was varied between 2 and 293, we selected a random subset of
all the available classes, and we trained both the SVM and the
neural classifier on the same subset. The results are averaged
over 10 repetitions, each one with a different sample of output
classes.
To make a meaningful comparison between the the en-
ergy consumed by a SVM and the neuromorphic classifier, we
equalized all the classification accuracies, as follows. For each
classification problem we varied the margin penalty param-
eter of the standard SVC using grid search and picked the
best performance achieved. We then varied the relevant pa-
rameters of the other two classifiers so that their classification
accuracy matched or exceeded the accuracy of the standard
SVC. Specifically, we progressively increased the number of
basis functions (in the primalSVC method) and the number
of RCNs (in the neural classifier) until both reached the tar-
get performance. For each classification problem we averaged
over 10 realizations of the random projections of the neural
classifier.
The results are summarized in Fig. 4. The energy con-
sumption is about two orders of magnitude larger for the
SVM throughout the entire range of variation of the num-
ber of classes that we considered, although for a very small
(2–3) number of classes the advantage of the neuromorphic
classifier strongly reduces, most likely because the algorithms
to minimize the number of SVs work best when the number
5
1011021031046080100numberofRCNsorsupportvectorsperformance(%)networkchipaMNISTlibsvmSVCperfprimalSVC255075101102103104performance(%)numberofRCNsorsupportvectorsnetworkchipcMNIST-back-imagelibsvmSVCperfprimalSVCb60801000.010.1110100averageenergyconsumption(mJ)networkchiplibsvmSVCperfprimalSVC2550750.010.1110averageenergyconsumption(mJ)networkchipdlibsvmSVCperfprimalSVC of classes is low. This plot indicates that the energy advan-
tage of the neuromorphic classifier over SVMs implemented
on conventional digital machines is maintained also for more
complex tasks involving a larger number of classes.
It is interesting to discuss the expected scaling for grow-
ing number of classes. Consider the case of generic C classes
multi-class problems solved through reduction with multiple
combined binary SVMs.
In a one-vs-all reduction scheme,
each binary classifier is trained to respond to exactly one of
the C classes, and hence C SVMs are required. For each
SVM, one needs to compute the scalar products between the
test sample to be classified and the NSV support vectors. Each
scalar product requires Nin multiplications and sums. In the
favorable case in which all binary classifiers happen to share
the same support vectors, the scalar products can be com-
puted only once and would require Nin·NSV operations. These
NSV scalar products then need to be multiplied by the cor-
responding coefficients, which are different for the different
SVMs. This requires additional CNSV operations.
If NSV
scales linearly with C, as in the cases we analyzed, then the
total energy E will scale as
E ∼ NinC + C 2.
When C is small compared to Nin, the first term dominates,
and the expected scaling is linear. However, for C > Nin
the scaling is expected to be at least quadratic. It can grow
more rapidly if the support vectors are different for different
classifiers.
Interestingly the expected scaling for the neural network
classifiers that we considered is the same. The energy con-
sumption mostly depends on the number of needed cores. This
number will be proportional to the number of RCNs, N , mul-
tiplied by the number of classes. Indeed, each core can receive
up to 256 inputs, so the total number of needed cores will be
proportional to (cid:100)N/256(cid:101), with (cid:100)·(cid:101) denoting the ceiling func-
tion. Moreover, the number of readout units, which are the
output lines of these cores, will be proportional to the num-
ber of classes. Hence the N C dependence. In the cases we
analyzed N depends linearly on the number of classes, and
hence the energy depends quadratically on C, as in the case
of the SVMs when C is large enough. Notice that the there
is a second term which also scales quadratically with C that
contributes to the energy. The second term comes from the
necessity of replicating the RCNs C times, due to the limited
fan out of the RCNs. Again, under the assumption of N ∼ C,
also this term will scale quadratically with C.
Given that the scaling with the number of classes is basi-
cally the same for the neuromorhic classifier as for the SVMs,
it is not unreasonable to hypothesize that the energy consump-
tion advantage of the neuromorphic implementation would be
preserved also for a much larger number of classes.
Discussion
Our results indicate that neuromorphic devices are mature
enough to achieve performances on a real-world machine-
learning task that are comparable to those of state-of-the-art
conventional devices with von Neumann architecture, all just
by using a tiny fraction of their energy. Our conclusions are
based on a few significant tests, based on a comparison limited
to our neuromorphic classifier and a few digital implementa-
tions of SVMs. This clearly restricts the generality of our
results and does not preclude situations in which the advan-
tage of the neuromorphic approach might be less prominent.
In any case, the merit of our study is to offer a solid com-
parison with implementations on current conventional digital
6
platforms that are energy-efficient themselves. In particular,
the algorithm we used on conventional digital machines in-
volves only multiplications between matrices and vectors, the
efficiency of which has been dramatically increased in the last
decades thanks to optimized parallelization. Furthermore, not
only we tried to match the classification performance of the
competitors, but we also considered two additional SVM al-
gorithms that minimize the number of support vectors, and
hence the final number of operations. Other choices for SVM
algorithms would certainly lead to different estimates for en-
ergy consumption, but it is rather unlikely that they would
change across 2 orders of magnitude. It is possible that full
custom unconventional digital machines based, e.g., on field
programmable gate arrays (FPGAs) would be more energy-
efficient, but it is hard to imagine that they would break the
predicted energy wall discussed in [28]. If this assumption is
right, neuromorphic hardware would always be more efficient
when performing the type of tasks that we considered. More-
over, analog neuromorphic VLSI or unreliable digital tech-
nologies might allow for a further reduction of energy con-
sumption, probably by another order of magnitude [5, 29, 30].
The current energy consumption levels achieved by analog sys-
tems are very close to those of biological brains in terms of en-
ergy per spike, although many of these systems are relatively
small and it is unclear whether they can ever be extended to
brain-scale architectures.
Other custom chips that can solve real-world tasks have
been designed. An example is the FPGA chip NeuFlow, de-
signed to implement convolutional networks for visual recog-
nition. The chip is digital and uses as little as 4.9 ×
1011 operations/W or, equivalently, 2 pJ/operation.
It is also interesting to discuss the performance of other
conventional digital processors in the benchmarks we exam-
ined. Let us consider for example the implementations of
SVMs classifying the MNIST digits with about 104 support
vectors, which is roughly the number of vectors we need to
achieve the best classification accuracy. As we have shown,
the Intel i7 takes about 10 ms to perform a classification, at
an approximate cost of 50 mJ. The IBM chip, in contrast, re-
quired 1 mJ for the longest classification times (500 ms), and
0.2 mJ for the average classification time (100 ms). We also
quantified the energy cost of the ARMv7, which is a more
energy-efficient yet slower microprocessor often used in mo-
bile technologies.
Its energy consumption per classification
was substantially higher, around 700 mJ. The main reason for
this high consumption is that it takes more than 0.6 seconds to
perform a single classification. And the baseline consumption,
which increases linearly with the classification time, is a large
portion of the total energy needed for a classification. Finally,
we considered the recent Xeon Phi, which has a massively par-
allel architecture and is employed in high performance com-
puting applications. As we do not have a simulator for the
Phi, we could only indirectly estimate a lower bound for the
energy consumption (see Methods for more details). Accord-
ing to our estimate, a single classification requires only 0.2 µs
and uses about 16 mJ, which would be significantly lower than
the energy cost of the i7 and very close to the estimated lower
limit of energy consumption [28], but still larger than the con-
sumption of the IBM chip. Notice however that both the clas-
sification time and the energy consumption of the Xeon Phi
processor are very likely to be grossly underestimated, as they
are simply derived from the peak performance of 100 Tflop/s.
The estimates for the i7 and the ARMv7 are significantly more
reliable, because we derived them by simulating the proces-
sors.
To summarize, our results compellingly suggest that the
neuromorphic approach is finally competitive in terms of en-
Figure 4. Dependence of the energy consumption on the number of classes a Classification accuracy for the neural classifier and for two SVM
algorithms, as a function of the number C of classes for the LATEX dataset. The parameters of the different classifiers are tuned to have approximately the same classification
accuracy. b Energy consumption as a function of the number of classes, for the LATEX dataset. Given a number C of classes, every point in the plot is obtained by training
a given classifier on C randomly sampled classes among the 293 available ones. This procedure is repeated 10 times for every value of C and every type of classifier. Each
datapoint associated with the neural classifier (’RCN’) was in turn estimated from a sample of 10 realizations of the random connections (squares indicate sample means,
errorbars indicate the 0.1 and 0.9 fractile of the sample). c As in b, but number of support vectors and RCNs as a function of the number of classes.
Dataset
MNIST
MNIST-back-image
LATEX
size
image
28 × 28
28 × 28
32 × 32
num.
classes
size trai-
ning set
10
10
293
60000
12000
14650
size
test set
10000
50000
9376
ergy consumption in useful real-world machine learning tasks
and constitutes a promising direction for future scalable tech-
nologies. The recent success of deep networks for large-scale
machine learning [31, 32] makes neuromorphic approaches
particularly relevant and valuable. This will be certainly true
for neuromorphic systems with synaptic plasticity, which will
enable these devices to learn autonomously from experience.
Learning is now available only in small neuromorphic systems
[33, 34, 35], but hopefully new VLSI technologies will allow
us to implement it also in large-scale neural systems.
ACKNOWLEDGMENTS. This work was supported by DARPA SyNAPSE, Gatsby
Charitable Foundation, Swartz Foundation and Kavli Foundation. We are grateful to
the IBM team led by Dr. D. Modha for their assistance with the IBM chip simulator.
In particular we thank John Arthur and Paul Merolla for their help with the estimate
of the chip energy consumption. DM acknowledges the support from the FP7 Marie
Curie Actions of the European Commission and the ANR-10-LABX-0087 IEC and
ANR-10-IDEX-0001-02 PSL grants.
Materials and Methods
Images sets for classification benchmarksWe used three
datasets in our study: MNIST, MNIST-back-image, and
LATEX. The MNIST dataset consists of images of handrwit-
ten digits (10 classes) [36]. The MNIST-back-image dataset
contains the same digits of MNIST, but in this case the back-
ground of each pattern is a random patch extracted from a
set of 20 black and white images downloaded from the In-
ternet [37]. Patches with low pixel variance (i.e. containing
little texture) are discarded. The LATEX dataset consists of
distorted versions of 293 characters used in the LATEX docu-
ment preparation system [38, 39]. All datasets consist of l × l
pixel gray-scale images, and each of such pixel images is asso-
ciated with one out of C possible classes. The size of the pixel
images, the number of classes, and the sizes of the training
and test sets depend on the data set (see table below).
Preprocessing Every sample image was reshaped as a l2-
dimensional vector, and the average gray level of each compo-
nent was subtracted from the data. The dimensionality of the
resulting image vector was then reduced to 256 using PCA. To
guarantee that all the selected components contributed uni-
formly to the patterns, we applied a random rotation to the
principal subspace (see, e.g., [40]). We denote by Nin = 256
the dimension of that subspace.
The architecture of the network and the training algorithm
We map the preprocessed Nin-dimensional vector image, s,
into a higher dimensional space through the transformation
xi = f (wi · s),
i = 1, . . . , N,
where wi is an Nin-dimensional sparse random vector and f (·)
is a nonlinear function. This is the transformation induced by
a neural network with Nin input units and N output units
with activation function f (·). More succinctly,
x = f (WT s),
[ 1 ]
where W is a weight matrix of dimensions Nin × N formed
by adjoining all the column weight vectors wi, and where
f (·) acts componentwise, i.e., f (x) ≡ (f (x1), . . . , f (xNin ))T .
The output of the random nonlinear transformation, x, is
used as the input to a linear NC -class discriminant, consist-
k=1 Jjkxk, with
ing of NC linear functions of the type yj =(cid:80)N
j = 1, . . . , C. More compactly,
y = Jx,
[ 2 ]
where y = (y1, . . . , yC )T , J is a C×N matrix, and x is given by
Eq. [ 1 ]. A pattern x is assigned to class Cj if yj(x) > yk(x)
for all j (cid:54)= k. The elements of J are learned offline by impos-
ing a 1-of-NC coding scheme on the output: if the target class
is j then the target output t is a vector of length NC where
all components are zero except component tj, which is 1. For
the offline training of weights we use the pseudoinverse, which
minimizes the mean squared error of the outputs. This tech-
nique has been shown to be a good replacement for empirical
minimization problems when the dataset is embedded in a ran-
dom high-dimensional space, which is our case [41, 26, 42, 27].
Neuromorphic chip implementation
The chip is composed of multiple identical cores, each of which
consists of a neuromorphic circuit that comprises n = 256 ax-
ons, n neurons, and n2 adjustable synapses ([43, 10, 6], see
also Fig. 1c). Each axon provides the inputs by feeding the
spiking activity of one given neuron that may or not reside in
the core. The incoming spiking activity to all n axons in a core
is represented by a vector of activity bits (A1(t), . . . , An(t))
whose elements indicate whether or not the neurons associ-
ated with the incoming axons emitted a spike in the previ-
ous time step. The intersection of the the n axons with the
n neurons forms a matrix of programmable synapses. The
weight of active synapses is determined by the type of axon
and the type of neuron the synapse lies on. Specifically, each
core can contain up to four different types of axon, labeled
Gj = {1, 2, 3, 4}, whereas it can accommodate an unlimited
7
1010000.51numberofclassesperformanceRCNlibsvmprimalSVC101000.991a0.010.111010100energyconsumption(mJ)numberofclassesRCNprimalSVClibsvmb10100101102103104numberofclassesnumberSVs,RCNsRCNlibsvmprimalSVCc i, . . . , S4
i
number of neuron types, each of which having four associated
synaptic weights Si = (S1
i ). The strength of an active
synapse connecting axon j with neuron i is SGj
, that is, the
axon type determines which weight to pick among the weights
associated with neuron i. The net input received by neuron i
i BijAj(t), where
Bij is 1 or 0 depending on whether the synapse between axon
j and dendrite i is active or inactive.
at time step t is therefore hi(t) =(cid:80)n
j=1 SGj
At each time step the membrane potential Vi(t) of neu-
ron i receiving input h(t) is updated according to Vi(t + 1) =
Vi(t) − β + hi(t), where β is a constant leak. If Vi(t) becomes
negative after an update, it is clipped to 0. Conversely, when
Vi(t) reaches the threshold Vthr, the potential is reset to Vreset
and the neuron emits a spike, which is sent through the neu-
ron’s axon to the target core and neuron. This design implies
that each neuron can connect to at most n neurons, which
are necessarily in the same core. The initial voltage of each
neuron was initialized by drawing randomly and with equal
probability from a set of 4 evenly spaced values from Vreset to
Vthr.
Signal-to-rate transduction The input to the neuromor-
phic chip consists of a set of spike trains fed to the neurons
of the input layer. To transform the vector signal s into spike
trains, we first shifted the signal by ¯s = 3σ, where σ is the
standard deviation across all signal components of all pat-
terns. The shifted signal was then scaled by a factor νsc cho-
sen to ensure moderate output rates in the RCN layer, and
the result was linear-rectified to positive values. In short, the
input rate νi associated with signal si is νi = νsc[si + ¯s]+,
i = 1, . . . , Nin, where [x]+ is x if x > 0, or 0 otherwise. The
values νi were then used to generate regular spike trains with
fixed inter-spike-interval 1/νi.
Basic architecture The circuit is divided in two functional
groups, or layers, each of which comprises several cores. The
first functional group is the RCN layer, which computes the
random nonlinear expansion in Eq.
[ 1 ]. The second func-
tional group computes the C-class discriminant y = Jx. The
output of the classifier is just argmaxj yj, where j runs over
the C possible categories. The argmax operation was not com-
puted by the chip, but was determined off-line by comparing
the accumulated spike counts across all outputs. In the fol-
lowing, we describe the implementation of the two layers in
more detail.
RCN layer We first set the dimensionality of the input to
the number of available axons per core, i.e., Nin = n = 256. A
convenient choice for W is a n× N matrix where each column
is vector of zeros except for exactly m < n nonzero entries,
which are randomly placed and take a fixed integer value w.
We took m = 26, which corresponds to a connectivity level
of around 0.1. Lowering the connectivity has the advantage
of decreasing energy costs by reducing the number of total
spikes and active synapses, without impacting the classifica-
tion performance. The random expansion was mapped in the
chip by splitting the matrix WT into (cid:100)N/n(cid:101) submatrices of
size n× n, and using each submatrix as the (boolean) connec-
tivity matrix Bij of a core.
With this arrangement, each of the N neurons distributed
among the (cid:100)N/n(cid:101) cores receives a sparse and random linear
combination of signals. Specifically, the average current re-
ceived by each RCN is
n(cid:88)
hi =
Wjiνj,
i = 1, . . . , N.
j=1
8
A zero-th order approximation of the firing rate of a gen-
eral vlsi neuron receiving a current hi is
ri =
[hi − β]+
Vthr − Vreset
,
[ 3 ]
where Vthr is the threshold for spike emission and Vreset is the
reset potential [12].
We chose the parameters w and β to meet two criteria.
First, we required the fraction of RCNs showing any firing
activity (i.e., the coding level f ) to be around 0.25. This
coding level is a good compromise between the need for dis-
crimination and generalization, and it keeps finite-size effects
at bay [9]. Second, we required the distribution of activities
across active RCNs to be sufficiently wide. Otherwise the in-
formation carried by the spiking activity of the RCNs is too
imprecise to discriminate among patterns.
All the cores in the RCN layer receive exactly the same
n-dimensional input signal.
Readout
The readout matrix J was trained offline and mapped to the
chip architecture as follows.
Weight quantization Because the chip can hold only
integer-valued synapses, we need to map the set of all compo-
nents of J into an appropriate finite set of integers. We started
clipping the synaptic weights within the bounds (−4σ, 4σ),
where σ is the standard deviation of the sample composed of
all the components of J. We then rescaled the weights to a
convenient magnitude Jmax = 28 (see below), and rounded
the weight values to the nearest integer.
Weight assignment The TrueNorth connectivity con-
straints dictate that each RCN can project to only one axon,
meaning that there are at most n = 256 synaptic contacts
available to encode the C = 10 weights, J0i, . . . , J9i associ-
ated with the i-th RCN. We allocated 24 contacts per class
and per axon (see Fig. 5). Each of these 24 contacts were
divided in four groups comprising 6 weights each, with val-
ues 1, 2, 4,−1,−2,−4. This allowed us to represent any inte-
ger weight from −28 to 28 (each of the 4 groups encodes a
maximum weight of 7, sign aside). To distribute any weight
value w across the available synaptic contacts, we decom-
posed w in a sum of four terms, given by the integer divi-
sion of w by 4 with the remainder spread evenly across terms
(Ex: 19 = 4 + 5 + 5 + 5). Each of such values was assigned
to one group, represented in base 2, and mapped to a pat-
tern of active-inactive synapses according to the weight as-
sociated with each axon-dendrite intersection. Positive and
negative weights, as well as strong and week weights, were
balanced along a dendrite by changing the sign and order of
the weights in the crossbar (see alternating colors and satura-
tions in Fig. 5).
Negative threshold For the readout to work properly, the
firing activity of readout neurons must be proportional to the
linear sum of the inputs from the RCNs. This requires neu-
rons to operate in the linear regime of their dynamic range,
a regime that can be enforced by lowering the threshold βout
of readout neurons. We set βout < 0, which is equivalent
to adding a constant positive current to each neuron.
If
the current-to-rate transduction function were the threshold-
linear function of Eq.
[ 3 ], the baseline activity induced by
this constant current would be βout/(Vthr − Vreset) per read-
out neuron. The contribution of this background signal should
Figure 5. Implementation of the readout matrix in a core. The diagram represents the first 8 input lines and first 48 dendrites (two output units)
of a typical readout core. Under each axon-dendrite contact is a square that indicates the potential synaptic strength at the site: color indicates whether the connection is
excitatory (red) or inhibitory (blue), while the saturation level represents the absolute value of the synaptic strength, which can be 1, 2, or 4 (low, medium, and high saturation,
respectively). Only the sites marked with a dot are active. The green frame highlights all the synaptic contacts allocated for an arbitrary weight of the readout matrix, in this
case J13 = −9, which is decomposed as the 4-term sum −9 = −2 − 3 − 2 − 2 = −0102 − 0112 − 0102 − 0102. Note that in this particular axon the ordering of
weights is 20, 21, 22 (rightmost bit is the most significant).
be subtracted from the readout outputs if one wants to get the
equivalent to Eq. [ 2 ], although the step is unnecessary if one
only wishes to compare output magnitudes (as we implicitly
do in order to find the maximal output).
Support Vector Machines. We trained SVMs to perform mul-
ticlass classifications based on a one-vs-all scheme, so that
the number of output units coincides with the number of
classes (as in the neural classifier). SVMs were evaluated us-
ing arc-cosine kernels, whic mimic the computation of large
feedforward networks with one or more layers of hidden non-
linear units [18]. For our particular architecture, based on
one hidden layer built with threshold-linear units, the kernel
is k(x, y) = (cid:107)x(cid:107)(cid:107)y(cid:107)J1(θ), where J1(θ) = sin θ + (π − θ) cos θ
and θ is the angle between the inputs x and y.
We considered three types of SVM. For the standard SVM
we used the open library libsvm [44], which we patched to in-
clude the arccos kernel. The other two SVMs reduce the num-
ber of support vectors without sacrificing performance sub-
stantially. One of such algorithms is primalSVC, which selects
greedily the basis functions by optimizing the primal objec-
tive function [45]. The other method is based on the so-called
Cutting-Plane Subspace Pursuit algorithm, which reduces the
number of support vectors by using basis functions that, un-
like standard SVMs, are not necessarily training vectors [46].
Such method is implemented in the library SVMperf. Unlike
the other two classifiers, SVMperf used RBF kernels instead of
arccos kernels.
Estimation of the IBM chip energy consumptionThe en-
ergy consumption of the IBM chip was estimated from the
TrueNorth specifications [6]. The total energy consumption
comprises the baseline energy (15.9 µW per core), the energy
to emit spikes (109 pJ per spike), the energy needed to read
active synapses (10.7 pJ per active synapse), and the energy
necessary to update membrane potentials (1.2 pJ per neuron).
We ignored the input-output energy needed to transmit spikes
off chip and receive spikes on chip. These numbers provide
a reasonable estimate of the energy consumption of systems
with a conservative supply voltage of 0.775 V; most chips op-
erate near or below this estimate. For a setup with 214 RCNs,
26 dendrites per class, and 10 classes, the power was about
2.08 mW, 95% of which corresponds to the baseline power.
Scaling of the energy with the number of classes The
estimation was based on the energy cost of the simulated clas-
sifications of the MNIST dataset, and extrapolated to the de-
signs required by an increasing number of classes. As the
number of classes C increases, so does the number of readout
neurons necessary to perform a classification and, therefore,
so does the required number of readout cores. Specifically, if
we assign sc synaptic contacts per axon and per class, we will
need a total of scC output lines. These output lines need to
be connected to all the N neurons through the input lines of
the readout cores . Because each readout core can accomodate
256 output lines, connected to 256 input lines, the total num-
ber of readout cores will be (cid:100)N/256(cid:101)(cid:100)scC/256(cid:101) ((cid:100)·(cid:101) indicates
the ceiling function). In principle the number of RCN cores
will be simply (cid:100)N/256(cid:101). However, each RCN should project to
(cid:100)scC/256(cid:101) cores, which implies that each RCN core must be
cloned (cid:100)scC/256(cid:101) times due to the fan-out constraint—each
RCN can project to only one core. The total number of cores
is therefore Ncores = 2(cid:100)N/256(cid:101)(cid:100)scC/256(cid:101), where the factor 2
accounts for the contributions of both the readout and the
RCN cores.
The total number of spikes emitted was estimated from
the reference value we got from the chip simulation (for 10
classes, N = 214, sc = 24, and 500 ms of classification time),
scaled appropriately for the new Ncores. More concretely, if
we denote by n0
sp the number of spikes emitted during our
reference simulation, the number of emitted spikes in a gen-
sp(cid:100)scC/256(cid:101)(T /500)(N/214), where T is
eral case is nsp = n0
the duration of the simulation in milliseconds. We chose this
duration to be T = 108 ms, which is the average classification
time of the chip implementation the MNIST dataset, when
the spike difference is 80 spikes and which yields only 0.1%
less in performance than in the fixed-duration case (97.2%
vs 97.3%). With T and the estimated values of Ncores and
Figure 6. Simulation of a digital support vector machine. a
Number of operations (black circles, left ordinate) and runtime (blue dots, right ordi-
nate) required by a digital SVM to classify 10 test patterns from the MNIST dataset,
as a function of the number of support vectors. The SVM performance was estimated
with a simulator of the Intel i7 processor. b Energy consumption associated to the
datapoints shown in a (squares). The straight line is a least-square fit.
9
fromRCN12345678J13=−9+0+1025005000750002.5·1085·10800.050.1numberofSVsnumberofoperationsrunningtime(s)ab025005000750000.20.4numberofSVsenergyconsumption(J)0.053mJperSV(10classifications) Figure 7. Performance versus energy consumption at a fixed classification time. Panels are like in Figs. 3b,d, classification time is now
fixed at 500 ms, rather than determined by a stopping criterion.
nsp, it is straightforward to compute the energy consumption
according to the values given in the previous paragraph.
Energy consumption in von Neumann digital machines
Configuration The runtime and power of microprocessors
with von Neumann architectures were estimated with the re-
cently developed simulators GEM5 (gem5.opt 2.0) [47] and
McPAT (ver. 1.2) [48]. For the estimation we used an architec-
ture configuration similar to that of the recent Intel CoreTM
i7 processors [49], which incorporate state-of-the-art CMOS
technology. Specifically, we used an x86 64, O3, single core
architecture at 2.66 GHz clock frequency, with 32KB 8-way
L1-i and 32KB 8-way L1-d caches, 256KB 8-way L2 cache,
64B cache line size, and 8GB DDR3 1600 DRAM. Channel
length was 22 nm, HP type, using long channel if appropriate.
VDD was 0.9V, so slightly higher than the 0.775 V used for
the IBM chip. However, could we use the same voltage in
Intel i7 simulator, the energy consumption would be lower by
a factor (0.775/0.9)2 = 0.74. This 26% reduction would not
change the main conclusions about the energy consumption
gap between the IBM chip and the conventional von Neumann
digital machines, which is 2–3 orders of magnitude.
Simulations The benchmark was the test phase of the
SVMs, already trained. Simulations showed that a modern
microprocessor based on a von-Neumann architecture takes
115.5 ms to evaluate the test set with 8087 SVs, while con-
suming 424.6 mJ (DRAM energy consumption not included).
When we varied the number of support vectors from 9 to
8087, both the runtime and energy consumption grew propor-
tionally to the number of SVs, while the power was roughly
constant due to the fixed hardware configuration (see Fig 6).
To estimate how the energy used by von Neumann digital
SVMs scales with the number of classes, we ran another set
of simulations with Intel i7 simulator, this time varying both
the number of support vectors and the number of classes in
the classification problem. This step was necessary to deter-
mine the overhead incurred when we increase the number of
output units. For a given number of classes, the energy cost
per support vector was estimated from the least-square fit of
the energies against the number of support vectors.
Mobile processor We also investigated the runtime and
energy consumption of a more energy-efficient but slower
mobile microprocessor performing the same target workload.
The architecture configuration was: ARMv7, O3, single core,
1GHz CPU clock frequency, 32kB 4-way L1i and 32kB 4-way
L1d caches, and 128kB 8-way L2 cache, which is similar to the
architecture of ARM Cortex-A9 [50]. The technology node
(22 nm) and simulators were the same as in the experiment
with the microprocessor mimicking Intel Core i7. For the
benchmark code with the largest number of SVs, the task re-
quired 1.2· 1010 operations that took 6.35 s at a cost of 7.34 J.
Discussion on Intel Xeon Phi Massively parallel archi-
tectures have gained a significant amount of attention to
improve the throughput and power efficiency of the high-
performance computing (HPC) technology, in response to the
relatively stagnated improvement in clock frequency. The
Xeon Phi coprocessor, recently developed by Intel, is one of
such efforts [51]. It integrates more than 50 CPU cores to-
gether with L1/L2 caches, network-on-chips, GDDR memory
controller, and PCIe interface. Each core supports up-to 4-
thread in-order operation and the 512b SIMD VPU (Vector
processing unit). While the runtime and energy-consumption
of the coprocessor are highly dependent on the target work-
loads, several recent investigations quantified the performance
and energy-efficiency. In the high-performance configuration,
the system integrating Xeon and Xeon Phi shows the through-
put of 100 Tera floating-point operations (flop) per second, the
power consumption of 72.9 kW, marking the energy efficiency
of 0.74 nJ/flop [51]. The classification benchmark codes (with
the largest number of SVs) require 0.02235 Gigaflop on the
desktop processor configuration similar to Intel Core i7. At
a first order approximation, therefore, the Xeon and Xeon
Phi-based system takes 0.2235 µs and uses 16.5 mJ per clas-
sification. This energy consumption seems significantly lower
than the one of the Intel Core i7, and very close to its lower
bound, which is approximately 3 mJ. However, one should
keep in mind that the energy is grossly underestimated, as
not only we ignored the energy needed for the RAM, but we
also neglected the cost of the non floating point operations,
which are approximately twice as many as the floating point
operations. For all these reasons it is difficult to compare the
energy consumption for the Xean Phi to the Intel Core i7.
In any case, even for our very conservative energy consump-
tion estimate, the IBM chip remains significantly more energy
efficient.
10
a60801000.1110100performance(%)averageenergyconsumption(mJ)networkchiplibsvmSVCperfprimalSVC2550750.1110performance(%)averageenergyconsumption(mJ)networkchipblibsvmSVCperfprimalSVCMNISTMNIST-back-image 1. Mead C (1989) Analog VLSI implementation of neural systems (Addison Wesley Publishing
28. Hasler J, Marr B (2013) Finding a roadmap to achieve large neuromorphic hardware systems.
Company).
Indiveri G, et al. (2011) Neuromorphic silicon neuron circuits. Front. Neurosci. 5.
2.
3. Livi P, Indiveri G (2009) A current-mode conductance-based silicon neuron for address-event
neuromorphic systems pp 2898–2901.
Front. Neurosci. 7.
29. Arthur JV, Boahen K (2011) Silicon-neuron design: A dynamical systems approach. Circuits
and Systems I: Regular Papers, IEEE Transactions on 58:1034–1043.
30. Han J, Orshansky M (2013) Approximate computing: An emerging paradigm for energy-
4. Rangan V, Ghosh A, Aparin V, Cauwenberghs G (2010) A subthreshold aVLSI implementation
efficient design (IEEE), pp 1–6.
of the Izhikevich simple neuron model pp 4164–4167.
31. Krizhevsky A, Sutskever I, Hinton GE (2012) ImageNet classification with deep convolutional
5. Chicca E, Stefanini F, Indiveri G (2014) Neuromorphic electronic circuits for building au-
neural networks pp 1097–1105.
tonomous cognitive systems. Proceedings of the IEEE PP:1–22.
6. Merolla PA, et al. (2014) A million spiking-neuron integrated circuit with a scalable commu-
nication network and interface. Science 345:668–673.
7. Jaeger H, Haas H (2004) Harnessing nonlinearity: predicting chaotic systems and saving energy
in wireless communication. Science 304:78–80.
8. Buonomano DV, Maass W (2009) State-dependent computations: spatiotemporal processing
in cortical networks. Nat Rev Neurosci 10:113–125.
9. Barak O, Rigotti M, Fusi S (2013) The sparseness of mixed selectivity neurons controls the
generalization–discrimination trade-off. J. Neurosci. 33:3844–3856.
10. Arthur JV, et al. (2012) Building block of a programmable neuromorphic substrate: A digital
neurosynaptic core (IEEE), pp 1–8.
11. Eliasmith C, et al. (2012) A large-scale model of the functioning brain. Science 338:1202–1205.
12. Fusi S, Mattia M (1999) Collective behavior of networks with linear (VLSI) integrate-and-fire
neurons. Neural Comput. 11:633–652.
13. Sompolinsky H (1986) Neural networks with non-linear synapses and static noise. Phys. Rev.
A 34:2571.
14. Amit DJ, Fusi S (1994) Learning in neural networks with material synapses. Neural Compu-
tation 6:957–982.
15. Fusi S (2002) Hebbian spike-driven synaptic plasticity for learning patterns of mean firing rates.
Biological cybernetics 87:459–470.
16. Garey MR, Johnson DS (1979) Computers and intractability: a guide to NP-completeness
(WH Freeman New York).
17. Goodfellow I, Warde-farley D, Mirza M, Courville A, Bengio Y (2013) Maxout Networks pp
1319–1327.
32. Deng L, Hinton G, Kingsbury B (2013) New types of deep neural network learning for speech
recognition and related applications: An overview (IEEE), pp 8599–8603.
33. Mitra S, Fusi S, Indiveri G (2009) Real-time classification of complex patterns using spike-
based learning in neuromorphic vlsi. Biomedical Circuits and Systems, IEEE Transactions on
3:32–42.
34. Giulioni M, et al. (2011) Robust working memory in an asynchronously spiking neural network
realized with neuromorphic vlsi. Frontiers in neuroscience 5.
35. Arthur J, Boahen K (2006) Learning in silicon: timing is everything. Advances in neural
information processing systems 18:75.
36. LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document
recognition. Proc. IEEE 86:2278–2324.
37. Larochelle H, Erhan D, Courville A, Bergstra J, Bengio Y (2007) An empirical evaluation of
deep architectures on problems with many factors of variation, ICML ’07 (ACM, New York,
NY, USA), pp 473–480.
38. Amit Y, Geman D (1997) Shape quantization and recognition with randomized trees. Neural
Comput. 9:1545–1588.
39. Amit Y (2002) 2D Object Detection and Recognition: Models, Algorithms, and Networks
(MIT Press).
40. Raiko T, Valpola H, LeCun Y (2012) Deep Learning Made Easier by Linear Transformations
in Perceptrons Vol. 22, pp 924–932.
41. Huang GB, Zhu QY, Siew CK (2006) Extreme learning machine: theory and applications.
Neurocomputing 70:489–501.
42. Tapson J, van Schaik A (2013) Learning the pseudoinverse solution to network weights. Neural
Netw. 45:94–100.
43. Merolla P, et al. (2011) A digital neurosynaptic core using embedded crossbar memory with
18. Cho Y, Saul LK (2010) Large-margin classification in infinite neural networks. Neural Comput.
45pJ per spike in 45nm (IEEE), pp 1–4.
22:2678–2697.
19. Sohn K, Zhou G, Lee C, Lee H (2013) Learning and Selecting Features Jointly with Point-wise
Gated Boltzmann Machines pp 217–225.
44. Chang CC, Lin CJ (2011) LIBSVM: A library for support vector machines. ACM T. Intel. Sys.
Techn. 2:27:1–27:27 Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.
45. Keerthi SS, Chapelle O, DeCoste D (2006) Building support vector machines with reduced
20. Lan G, Sartori P, Neumann S, Sourjik V, Tu Y (2012) The energy-speed-accuracy trade-off in
classifier complexity. J. Mach. Learn. Res. 7:1493–1515.
sensory adaptation. Nature physics 8:422–428.
46. Joachims T, Yu CN (2009) Sparse kernel SVMs via cutting-plane training. Mach. Learn.
21. Tang H, et al. (2014) Spatiotemporal dynamics underlying object completion in human ventral
76:179–193.
visual cortex. Neuron 83:736–748.
47. Binkert N, et al. (2011) The GEM5 simulator. ACM SIGARCH Computer Architecture News
22. Boser BE, Guyon IM, Vapnik VN (1992) A training algorithm for optimal margin classifiers
39:1–7.
(ACM), pp 144–152.
23. Cortes C, Vapnik V (1995) Support-vector networks. Mach. Learn. 20:273–297.
24. Vapnik V, Golowich SE, Smola A (1997) Support vector method for function approximation,
48. Li S, et al. (2009) McPAT: an integrated power, area, and timing modeling framework for
multicore and manycore architectures (IEEE), pp 469–480.
49. (2014) Intel core i7 processor., Technical report
http://www.intel.com/content/www/us/en/
regression estimation, and signal processing pp 281–287.
processors/core- i7- processor.html.
25. Steinwart I, Christmann A (2008) Support vector machines (Springer).
26. Rahimi A, Recht B (2008) Weighted sums of random kitchen sinks: Replacing minimization
with randomization in learning pp 1313–1320.
27. Le Q, Sarl´os T, Smola A (2013) Fastfoodapproximating kernel expansions in loglinear time.
50. (2014) Arm cortex-a9., Technical report http://www.arm.com/products/processors/cortex- a/
cortex- a9.php.
51. Chrysos G, Engineer SP (2012) Intel Xeon Phi coprocessor (codename knights corner).
11
|
1701.07775 | 1 | 1701 | 2017-01-26T16:59:20 | A Forward Model at Purkinje Cell Synapses Facilitates Cerebellar Anticipatory Control | [
"q-bio.NC",
"eess.SY",
"math.OC"
] | How does our motor system solve the problem of anticipatory control in spite of a wide spectrum of response dynamics from different musculo-skeletal systems, transport delays as well as response latencies throughout the central nervous system? To a great extent, our highly-skilled motor responses are a result of a reactive feedback system, originating in the brain-stem and spinal cord, combined with a feed-forward anticipatory system, that is adaptively fine-tuned by sensory experience and originates in the cerebellum. Based on that interaction we design the counterfactual predictive control (CFPC) architecture, an anticipatory adaptive motor control scheme in which a feed-forward module, based on the cerebellum, steers an error feedback controller with counterfactual error signals. Those are signals that trigger reactions as actual errors would, but that do not code for any current or forthcoming errors. In order to determine the optimal learning strategy, we derive a novel learning rule for the feed-forward module that involves an eligibility trace and operates at the synaptic level. In particular, our eligibility trace provides a mechanism beyond co-incidence detection in that it convolves a history of prior synaptic inputs with error signals. In the context of cerebellar physiology, this solution implies that Purkinje cell synapses should generate eligibility traces using a forward model of the system being controlled. From an engineering perspective, CFPC provides a general-purpose anticipatory control architecture equipped with a learning rule that exploits the full dynamics of the closed-loop system. | q-bio.NC | q-bio |
A Forward Model at Purkinje Cell Synapses
Facilitates Cerebellar Anticipatory Control
Ivan Herreros-Alonso
SPECS lab
Xerxes D. Arsiwalla
SPECS lab
Universitat Pompeu Fabra
Universitat Pompeu Fabra
Barcelona, Spain
[email protected]
Barcelona, Spain
Paul F.M.J. Verschure
SPECS, UPF
Catalan Institution of Research
and Advanced Studies (ICREA)
Barcelona, Spain
Abstract
How does our motor system solve the problem of anticipatory control in spite
of a wide spectrum of response dynamics from different musculo-skeletal sys-
tems, transport delays as well as response latencies throughout the central nervous
system? To a great extent, our highly-skilled motor responses are a result of a
reactive feedback system, originating in the brain-stem and spinal cord, combined
with a feed-forward anticipatory system, that is adaptively fine-tuned by sensory
experience and originates in the cerebellum. Based on that interaction we design
the counterfactual predictive control (CFPC) architecture, an anticipatory adaptive
motor control scheme in which a feed-forward module, based on the cerebellum,
steers an error feedback controller with counterfactual error signals. Those are
signals that trigger reactions as actual errors would, but that do not code for any cur-
rent or forthcoming errors. In order to determine the optimal learning strategy, we
derive a novel learning rule for the feed-forward module that involves an eligibility
trace and operates at the synaptic level. In particular, our eligibility trace provides
a mechanism beyond co-incidence detection in that it convolves a history of prior
synaptic inputs with error signals. In the context of cerebellar physiology, this
solution implies that Purkinje cell synapses should generate eligibility traces using
a forward model of the system being controlled. From an engineering perspective,
CFPC provides a general-purpose anticipatory control architecture equipped with a
learning rule that exploits the full dynamics of the closed-loop system.
1
Introduction
Learning and anticipation are central features of cerebellar computation and function (Bastian, 2006):
the cerebellum learns from experience and is able to anticipate events, thereby complementing a
reactive feedback control by an anticipatory feed-forward one (Hofstoetter et al., 2002; Herreros
and Verschure, 2013). This interpretation is based on a series of anticipatory motor behaviors that
originate in the cerebellum. For instance, anticipation is a crucial component of acquired behavior in
eye-blink conditioning (Gormezano et al., 1983), a trial by trial learning protocol where an initially
neutral stimulus such as a tone or a light (the conditioning stimulus, CS) is followed, after a fixed
delay, by a noxious one, such as an air puff to the eye (the unconditioned stimulus, US). During early
trials, a protective unconditioned response (UR), a blink, occurs reflexively in a feedback manner
following the US. After training though, a well-timed anticipatory blink (the conditioned response,
CR) precedes the US. Thus, learning results in the (partial) transference from an initial feedback
action to an anticipatory (or predictive) feed-forward one. Similar responses occur during anticipatory
postural adjustments, which are postural changes that precede voluntary motor movements, such
as raising an arm while standing (Massion, 1992). The goal of these anticipatory adjustments is to
counteract the postural and equilibrium disturbances that voluntary movements introduce. These
30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
behaviors can be seen as feedback reactions to events that after learning have been transferred to
feed-forward actions anticipating the predicted events.
Anticipatory feed-forward control can yield high performance gains over feedback control whenever
the feedback loop exhibits transmission (or transport) delays (Jordan, 1996). However, even if a
plant has negligible transmission delays, it may still have sizable inertial latencies. For example,
if we apply a force to a visco-elastic plant, its peak velocity will be achieved after a certain delay;
i.e. the velocity itself will lag the force. An efficient way to counteract this lag will be to apply
forces anticipating changes in the desired velocity. That is, anticipation can be beneficial even when
one can act instantaneously on the plant. Given that, here we address two questions: what is the
optimal strategy to learn anticipatory actions in a cerebellar-based architecture? and how could it be
implemented in the cerebellum?
To answer that we design the counterfactual predictive control (CFPC) scheme, a cerebellar-based
adaptive-anticipatory control architecture that learns to anticipate performance errors from experience.
The CFPC scheme is motivated from neuro-anatomy and physiology of eye-blink conditioning.
It includes a reactive controller, which is an output-error feedback controller that models brain
stem reflexes actuating on eyelid muscles, and a feed-forward adaptive component that models the
cerebellum and learns to associate its inputs with the error signals driving the reactive controller.
With CFPC we propose a generic scheme in which a feed-forward module enhances the performance
of a reactive error feedback controller steering it with signals that facilitate anticipation, namely,
with counterfactual errors. However, within CFPC, even if these counterfactual errors that enable
predictive control are learned based on past errors in behavior, they do not reflect any current or
forthcoming error in the ongoing behavior.
In addition to eye-blink conditioning and postural adjustments, the interaction between reactive
and cerebellar-dependent acquired anticipatory behavior has also been studied in paradigms such
as visually-guided smooth pursuit eye movements (Lisberger, 1987). All these paradigms can be
abstracted as tasks in which the same predictive stimuli and disturbance or reference signal are
repeatedly experienced. In accordance to that, we operate our control scheme in trial-by-trial (batch)
mode. With that, we derive a learning rule for anticipatory control that modifies the well-known
least-mean-squares/Widrow-Hoff rule with an eligibility trace. More specifically, our model predicts
that to facilitate learning, parallel fibers to Purkinje cell synapses implement a forward model that
generates an eligibility trace. Finally, to stress that CFPC is not specific to eye-blink conditioning, we
demonstrate its application with a smooth pursuit task.
2 Methods
2.1 Cerebellar Model
Figure 1: Anatomical scheme of a Cerebellar Purkinje cell. The xj denote parallel fiber inputs to
Purkinje synapses (in red) with weights wj. o denotes the output of the Purkinje cell. The error signal
e, through the climbing fibers (in green), modulates synaptic weights.
We follow the simplifying approach of modeling the cerebellum as a linear adaptive filter, while
focusing on computations at the level of the Purkinje cells, which are the main output cells of the
cerebellar cortex (Fujita, 1982; Dean et al., 2010). Over the mossy fibers, the cerebellum receives
a wide range of inputs. Those inputs reach Purkinke cells via parallel fibers (Fig. 1), that cross
2
xj x1 xN oew1wjwN dendritic trees of Purkinje cells in a ratio of up to 1.5 × 106 parallel fiber synapses per cell (Eccles
et al., 1967). We denote the signal carried by a particular fiber as xj, j ∈ [1, G], with G equal to the
total number of inputs fibers. These inputs from the mossy/parallel fiber pathway carry contextual
information (interoceptive or exteroceptive) that allows the Purkinje cell to generate a functional
output. We refer to these inputs as cortical bases, indicating that they are localized at the cerebellar
cortex and that they provide a repertoire of states and inputs that the cerebellum combines to generate
its output o. As we will develop a discrete time analysis of the system, we use n to indicate time (or
time-step). The output of the cerebellum at any time point n results from a weighted sum of those
cortical bases. wj indicates the weight or synaptic efficacy associated with the fiber j. Thus, we
(cid:124)
(where the transpose, (cid:124), indicates
have x[n] = [x1[n], . . . , xG[n]]
that x[n] and w[n] are column vectors) containing the set of inputs and synaptic weights at time n,
respectively, which determine the output of the cerebellum according to
(cid:124)
and w[n] = [w1[n], . . . , wG[n]]
o[n] = x[n]
(cid:124)
w[n]
(1)
The adaptive feed-forward control of the cerebellum stems from updating the weights according to a
rule of the form
∆wj[n + 1] = f (xj[n], . . . , xj[1], e[n], Θ)
(2)
where Θ denotes global parameters of the learning rule; xj[n], . . . , xj[1], the history of its pre-
synaptic inputs of synapse j; and e[n], an error signal that is the same for all synapses, corresponding
to the difference between the desired, r, and the actual output, y, of the controlled plant. Note that in
drawing an analogy with the eye-blink conditioning paradigm, we use the simplifying convention
of considering the noxious stimulus (the air-puff) as a reference, r, that indicates that the eyelids
should close; the closure of the eyelid as the output of the plant, y; and the sensory response to the
noxious stimulus as an error, e, that encodes the difference between the desired, r, and the actual
eyelid closures, y. Given this, we advance a new learning rule, f, that achieves optimal performance
in the context of eye-blink conditioning and other cerebellar learning paradigms.
2.2 Cerebellar Control Architecture
Figure 2: Neuroanatomy of eye-blink conditioning and the CFPC architecture. Left: Mapping of
signals to anatomical structures in eye-blink conditioning (De Zeeuw and Yeo, 2005); regular arrows
indicate external inputs and outputs, arrows with inverted heads indicate neural pathways. Right:
CFPC architecture. Note that the feedback controller, C, and the feed-forward module, F F , belong
to the control architecture, while the plant, P , denotes an object controlled. Other abbreviations: r,
reference signal; y, plant's output; e, output error; x, basis signals; o, feed-forward signal; and u,
motor command.
We embed the adaptive filter cerebellar module in a layered control architecture, namely the CFPC
architecture, based on the interaction between brain stem motor nuclei driving motor reflexes and
the cerebellum, such as the one established between the cerebellar microcircuit responsible for
conditioned responses and the brain stem reflex circuitry that produces unconditioned eye-blinks
(Hesslow and Yeo, 2002) (Fig. 2 left). Note that in our interpretation of this anatomy we assume
that cerebellar output, o, feeds the lower reflex controller (Fig. 2 right). Put in control theory terms,
within the CFPC scheme an adaptive feed-forward layer supplements a negative feedback controller
steering it with feed-forward signals.
3
+-US(airpu(cid:31))[r]Eyelids(Blink)[P][y]Facialnucleus[C]Trigeminalnucleus[e][e]CS(Context, e.g.: sound, light)[u]Cerebellum (cortex and nuclei) and Inferior olive [FF][x]Pons[o]FFxoreCuPy+-+ADAPTIVE(cid:31)ANTICIPATORY(cid:30)FEED(cid:31)FORWARD(cid:29) LAYER REACTIVE (cid:30)FEEDBACK(cid:29) LAYERFEEDBACK CLOSED(cid:31)LOOP SYSTEM Our architecture uses a single-input single-output negative-feedback controller. The controller
receives as input the output error e = r − y. For the derivation of the learning algorithm, we assume
that both plant and controller are linear and time-invariant (LTI) systems. Importantly, the feedback
controller and the plant form a reactive closed-loop system, that mathematically can be seen as a
system that maps the reference, r, into the plant's output, y. A feed-forward layer that contains the
above-mentioned cerebellar model provides the negative feedback controller with an additional input
signal, o. We refer to o as a counter-factual error signal, since although it mechanistically drives the
negative feedback controller analogously to an error signal it is not an actual error. The counterfactual
error is generated by the feed-forward module that receives an output error, e, as its teaching signal.
Notably, from the point of view of the reactive layer closed-loop system, o can also be interpreted as
a signal that offsets r. In other words, even if r remains the reference that sets the target of behavior,
r + o functions as the effective reference that drives the closed-loop system.
3 Results
3.1 Derivation of the gradient descent update rule for the cerebellar control architecture
We apply the CFPC architecture defined in the previous section to a task that consists in following
a finite reference signal r ∈ RN that is repeated trial-by-trial. To analyze this system, we use the
discrete time formalism and assume that all components are linear time-invariant (LTI). Given this,
both reactive controller and plant can be lumped together into a closed-loop dynamical system, that
can be described with the dynamics A, input B, measurement C and feed-through D matrices. In
general, these matrices describe how the state of a dynamical system autonomously evolves with
time, A; how inputs affect system states, B; how states are mapped into outputs, C; and how inputs
instantaneously affect the system's output D (Astrom and Murray, 2012). As we consider a reference
of a finite length N, we can construct the N-by-N transfer matrix T as follows (Boyd, 2008)
T =
D
CB
CAB
0
D
CB
0
0
D
...
...
CAN−2B CAN−3B CAN−4B . . . D
...
...
0
0
0
. . .
. . .
. . .
...
With this transfer matrix we can map any given reference r into an output yr using yr = T r, obtaining
what would have been the complete output trajectory of the plant on an entirely feedback-driven trial.
Note that the first column of T contains the impulse response curve of the closed-loop system, while
the rest of the columns are obtained shifting that impulse response down. Therefore, we can build
the transfer matrix T either in a model-based manner, deriving the state-space characterization of
the closed-loop system, or in measurement-based manner, measuring the impulse response curve.
Additionally, note that (I − T )r yields the error of the feedback control in following the reference, a
signal which we denote with e0.
Let o ∈ RN be the entire feed-forward signal for a given trial. Given commutativity, we can consider
that from the point of view of the closed-loop system o is added directly to the reference r, (Fig. 2
right). In that case, we can use y = T (r + o) to obtain the output of the closed-loop system
when it is driven by both the reference and the feed-forward signal. The feed-forward module only
outputs linear combinations of a set of bases. Let X ∈ RN×G be a matrix with the content of the
G bases during all the N time steps of a trial. The feed-forward signal becomes o = Xw, where
w ∈ RG contains the mixing weights. Hence, the output of the plant given a particular w becomes
y = T (r + Xw).
We implement learning as the process of adjusting the weights w of the feed-forward module in a
trial-by-trial manner. At each trial the same reference signal, r, and bases, X, are repeated. Through
learning we want to converge to the optimal weight vector w∗ defined as
(r − T (r + Xw))
(r − T (r + Xw))
w∗ = arg min
c(w) = arg min
e = arg min
(3)
(cid:124)
(cid:124)
e
1
2
1
2
w
w
w
where c indicates the objective function to minimize, namely the L2 norm or sum of squared errors.
With the substitution X = T X and using e0 = (I − T )r, the minimization problem can be cast as a
4
canonical linear least-squares problem:
w∗ = arg min
w
1
2
(cid:124)
(e0 − Xw)
(e0 − Xw)
(4)
One the one hand, this allows to directly find the least squares solution for w∗, that is, w∗ = X†e0,
where † denotes the Moore-Penrose pseudo-inverse. On the other hand, and more interestingly, with
w[k] being the weights at trial k and having e[k] = e0 − Xw[k], we can obtain the gradient of the
error function at trial k with relation to w as follows:
∇wc = − X
(cid:124)
e[k] = −X
(cid:124)T (cid:124)
e[k]
Thus, setting η as a properly scaled learning rate (the only global parameter Θ of the rule), we can
derive the following gradient descent strategy for the update of the weights between trials:
w[k + 1] = w[k] + ηX
(cid:124)T (cid:124)
e[k]
(5)
This solves for the learning rule f in eq. 2. Note that f is consistent with both the cerebellar anatomy
(Fig. 2left) and the control architecture (Fig. 2right) in that the feed-forward module/cerebellum only
requires two signals to update its weights/synaptic efficacies: the basis inputs, X, and error signal, e.
3.2 T (cid:124) facilitates a synaptic eligibility trace
The standard least mean squares (LMS) rule (also known as Widrow-Hoff or decorrelation learning
rule) can be represented in its batch version as w[k + 1] = w[k] + ηX
e[k]. Hence, the only
difference between the batch LMS rule and the one we have derived is the insertion of the matrix
factor T (cid:124). Now we will show how this factor acts as a filter that computes an eligibility trace at each
weight/synapse. Note that the update of a single weight, according Eq. 5 becomes
(cid:124)
(6)
where xj contains the sequence of values of the cortical basis j during the entire trial. This can be
rewritten as
e[k]
(cid:124)
jT (cid:124)
wj[k + 1] = wj[k] + ηx
with hj ≡ T xj. The above inner product can be expressed as a sum of scalar products
(cid:124)
wj[k + 1] = wj[k] + ηh
j e[k]
(7)
N(cid:88)
wj[k + 1] = wj[k] + η
hj[n]e[k, n]
(8)
n=1
where n indexes the within trial time-step. Note that e[k] in Eq. 7 refers to the whole error signal
at trial k whereas e[k, n] in Eq. 8 refers to the error value in the n-th time-step of the trial k. It is
now clear that each hj[n] weighs how much an error arriving at time n should modify the weight
wj, which is precisely the role of an eligibility trace. Note that since T contains in its columns/rows
shifted repetitions of the impulse response curve of the closed-loop system, the eligibility trace codes
at any time n, the convolution of the sequence of previous inputs with the impulse-response curve of
the reactive layer closed-loop. Indeed, in each synapse, the eligibility trace is generated by a forward
model of the closed-loop system that is exclusively driven by the basis signal.
Consequently, our main result is that by deriving a gradient descent algorithm for the CFPC cerebellar
control architecture we have obtained an exact definition of the suitable eligibility trace. That
definition guarantees that the set of weights/synaptic efficacies are updated in a locally optimal
manner in the weights' space.
3.3 On-line gradient descent algorithm
The trial-by-trial formulation above allowed for a straightforward derivation of the (batch) gradient
descent algorithm. As it lumped together all computations occurring in a same trial, it accounted for
time within the trial implicitly rather than explicitly: one-dimensional time-signals were mapped onto
points in a high-dimensional space. However, after having established the gradient descent algorithm,
we can implement the same rule in an on-line manner, dropping the repetitiveness assumption inherent
to trial-by-trial learning and performing all computations locally in time. Each weight/synapse must
5
have a process associated to it that outputs the eligibility trace. That process passes the incoming
(unweighted) basis signal through a (forward) model of the closed-loop as follows:
sj[n + 1] = Asj[n] + Bxj[n]
hj[n] = Csj[n] + Dxj[n]
where matrices A, B, C and D refer to the closed-loop system (they are the same matrices that we
used to define the transfer matrix T ), and sj[n] is the state vector of the forward model of the synapse
j at time-step n. In practice, each "synaptic" forward model computes what would have been the
effect of having driven the closed-loop system with each basis signal alone. Given the superposition
principle, the outcome of that computation can also be interpreted as saying that hj[n] indicates what
would have been the displacement over the current output of the plant, y[n], achieved feeding the
closed-loop system with the basis signal xj. The process of weight update is completed as follows:
(9)
At each time step n, the error signal e[n] is multiplied by the current value of the eligibility trace
hj[n], scaled by the learning rate η, and subtracted to the current weight wj[n]. Therefore whereas
the contribution of each basis to the output of the adaptive filter depends only on its current value and
weight, the change in weight depends on the current and past values passed through a forward model
of the closed-loop dynamics.
wj[n + 1] = wj[n] + ηhj[n]e[n]
3.4 Simulation of a visually-guided smooth pursuit task
We demonstrate the CFPC approach in an example of a visual smooth pursuit task in which the
eyes have to track a target moving on a screen. Even though the simulation does not capture all the
complexity of a smooth pursuit task, it illustrates our anticipatory control strategy. We model the
plant (eye and ocular muscles) with a two-dimensional linear filter that maps motor commands into
angular positions. Our model is an extension of the model in (Porrill and Dean, 2007), even though
in that work the plant was considered in the context of the vestibulo-ocular reflex. In particular, we
use a chain of two leaky integrators: a slow integrator with a relaxation constant of 100 ms drives the
eyes back to the rest position; the second integrator, with a fast time constant of 3 ms ensures that
the change in position does not occur instantaneously. To this basic plant, we add a reactive control
layer modeled as a proportional-integral (PI) error-feedback controller, with proportional gain kp and
integral gain ki. The control loop includes a 50 ms delay in the error feedback, to account for both
the actuation and the sensing latency. We choose gains such that reactive tracking lags the target by
approximately 100 ms. This gives kp = 20 and ki = 100. To complete the anticipatory and adaptive
control architecture, the closed-loop system is supplemented by the feed-forward module.
Figure 3: Behavior of the system. Left: Reference (r) and output of the system before (y[1]) and
after learning (y[50]). Right: Error before e[1] and after learning e[50] and output acquired by
cerebellar/feed-forward component (o[50])
The architecture implementing the forward model-based gradient descent algorithm is applied to a
task structured in trials of 2.5 sec duration. Within each trial, a target remains still at the center of
the visual scene for a duration 0.5 sec, next it moves rightwards for 0.5 sec with constant velocity,
remains still for 0.5 sec and repeats the sequence of movements in reverse, returning to the center.
The cerebellar component receives 20 Gaussian basis signals (X) whose receptive fields are defined
in the temporal domain, relative to trial onset, with a width (standard-deviation) of 50 ms and spaced
by 100 ms. The whole system is simulated using a 1 ms time-step. To construct the matrix T we
computed closed-loop system impulse response.
6
00.511.522.500.20.40.60.81time (s)angular position (a.u.) ry[1]y[50]00.511.522.5−0.100.10.2time (s)angular position (a.u.) e[1]e[50]o[50] At the first trial, before any learning, the output of the plant lags the reference signal by approximately
100 ms converging to the position only when the target remains still for about 300 ms (Fig. 3 left). As
a result of learning, the plant's behavior shifts from a reactive to an anticipatory mode, being able to
track the reference without any delay. Indeed, the error that is sizable during the target displacement
before learning, almost completely disappears by the 50th trial (Fig. 3 right). That cancellation
results from learning the weights that generate a feed-forward predictive signal that leads the changes
in the reference signal (onsets and offsets of target movements) by approximately 100 ms (Fig. 3
right). Indeed, convergence of the algorithm is remarkably fast and by trial 7 it has almost converged
to the optimal solution (Fig. 4).
Figure 4: Performance achieved with different learning rules. Representative learning curves of the
forward model-based eligibility trace gradient descent (FM-ET), the simple Widrow-Hoff (WH) and
the Widrow-Hoff algorithm with a delta-eligibility trace matched to error feedback delay (WH+50
ms) or with an eligibility trace exceeding that delay by 20 ms (WH+70 ms). Error is quantified as the
relative root mean-squared error (rRMSE), scaled proportionally to the error in the first trial. Error of
the optimal solution, obtained with w∗ = (T X)†e0, is indicated with a dashed line.
To assess how much our forward-model-based eligibility trace contributes to performance, we test
three alternative algorithms. In both cases we employ the same control architecture, changing the
plasticity rule such that we either use no eligibility trace, thus implementing the basic Widrow-Hoff
learning rule, or use the Widrow-Hoff rule extended with a delta-function eligibility trace that matches
the latency of the error feedback (50 ms) or slightly exceeds it (70 ms). Performance with the basic
WH model worsens rapidly whereas performance with the WH learning rule using a "pure delay"
eligibility trace matched to the transport delay improves but not as fast as with the forward-model-
based eligibility trace (Fig. 4). Indeed, in this case, the best strategy for implementing a delayed
delta eligibility trace is setting a delay exceeding the transport delay by around 20 ms, thus matching
the peak of the impulse response. In that case, the system performs almost as good as with the
forward-model eligibility trace (70 ms). This last result implies that, even though the literature
usually emphasizes the role of transport delays, eligibility traces also account for response lags due
to intrinsic dynamics of the plant.
To summarize our results, we have shown with a basic simulation of a visual smooth pursuit task
that generating the eligibility trace by means of a forward model ensures convergence to the optimal
solution and accelerates learning by guaranteeing that it follows a gradient descent.
4 Discussion
In this paper we have introduced a novel formulation of cerebellar anticipatory control, consistent
with experimental evidence, in which a forward model has emerged naturally at the level of Purkinje
cell synapses. From a machine learning perspective, we have also provided an optimality argument
for the derivation of an eligibility trace, a construct that was often thought of in more heuristic terms
as a mechanism to bridge time-delays (Barto et al., 1983; Shibata and Schaal, 2001; McKinstry et al.,
2006).
The first seminal works of cerebellar computational models emphasized its role as an associative
memory (Marr, 1969; Albus, 1971). Later, the cerebellum was investigates as a device processing
correlated time signals(Fujita, 1982; Kawato et al., 1987; Dean et al., 2010). In this latter framework,
7
0102030405000.20.40.60.81#trialrRMSEWHWH+50msWH+70msFM−ET the use of the computational concept of an eligibility trace emerged as a heuristic construct that
allowed to compensate for transmission delays in the circuit(Kettner et al., 1997; Shibata and Schaal,
2001; Porrill and Dean, 2007), which introduced lags in the cross-correlation between signals.
Concretely, that was referred to as the problem of delayed error feedback, due to which, by the time
an error signal reaches a cell, the synapses accountable for that error are no longer the ones currently
active, but those that were active at the time when the motor signals that caused the actual error were
generated. This view has however neglected the fact that beyond transport delays, response dynamics
of physical plants also influence how past pre-synaptic signals could have related to the current
output of the plant. Indeed, for a linear plant, the impulse-response function of the plant provides the
complete description of how inputs will drive the system, and as such, integrates transmission delays
as well as the dynamics of the plant. Recently,
Even though cerebellar microcircuits have been used as models for building control architectures,
e.g., the feedback-error learning model (Kawato et al., 1987), our CFPC is novel in that it links
the cerebellum to the input of the feedback controller, ensuring that the computational features of
the feedback controller are exploited at all times. Within the domain of adaptive control, there are
remarkable similarities at the functional level between CFPC and iterative learning control (ILC)
(Amann et al., 1996), which is an input design technique for learning optimal control signals in
repetitive tasks. The difference between our CFPC and ILC lies in the fact that ILC controllers
directly learn a control signal, whereas, the CFPC learns a conterfactual error signal that steers a
feedback controller. However the similarity between the two approaches can help for extending
CFPC to more complex control tasks.
With our CFPC framework, we have modeled the cerebellar system at a very high level of abstraction:
we have not included bio-physical constraints underlying neural computations, obviated known
anatomical connections such as the cerebellar nucleo-olivary inhibition (Bengtsson and Hesslow,
2006; Herreros and Verschure, 2013) and made simplifications such as collapsing cerebellar cortex and
nuclei into the same computational unit. On the one hand, such a choice of high-level abstraction may
indeed be beneficial for deriving general-purpose machine learning or adaptive control algorithms.
On the other hand, it is remarkable that in spite of this abstraction our framework makes fine-grained
predictions at the micro-level of biological processes. Namely, that in a cerebellar microcircuit (Apps
and Garwicz, 2005), the response dynamics of secondary messengers (Wang et al., 2000) regulating
plasticity of Purkinje cell synapses to parallel fibers must mimic the dynamics of the motor system
being controlled by that cerebellar microcircuit. Notably, the logical consequence of this prediction,
that different Purkinje cells should display different plasticity rules according to the system that they
control, has been validated recording single Purkinje cells in vivo (Suvrathan et al., 2016).
In conclusion, we find that a normative interpretation of plasticity rules in Purkinje cell synapses
emerges from our systems level CFPC computational architecture. That is, in order to generate
optimal eligibility traces, synapses must include a forward model of the controlled subsystem. This
conclusion, in the broader picture, suggests that synapses are not merely components of multiplicative
gains, but rather the loci of complex dynamic computations that are relevant from a functional
perspective, both, in terms of optimizing storage capacity (Benna and Fusi, 2016; Lahiri and Ganguli,
2013) and fine-tuning learning rules to behavioral requirements.
Acknowledgments
The research leading to these results has received funding from the European Commission's Horizon
2020 socSMC project (socSMC-641321H2020-FETPROACT-2014) and by the European Research
Council's CDAC project (ERC-2013-ADG 341196).
References
Albus, J. S. (1971). A theory of cerebellar function. Mathematical Biosciences, 10(1):25–61.
Amann, N., Owens, D. H., and Rogers, E. (1996). Iterative learning control for discrete-time systems with
exponential rate of convergence. IEE Proceedings-Control Theory and Applications, 143(2):217–224.
Apps, R. and Garwicz, M. (2005). Anatomical and physiological foundations of cerebellar information
processing. Nature reviews. Neuroscience, 6(4):297–311.
Astrom, K. J. and Murray, R. M. (2012). Feedback Systems: An Introduction for Scientists and Engineers.
Princeton university press.
8
Barto, A. G., Sutton, R. S., and Anderson, C. W. (1983). Neuronlike adaptive elements that can solve difficult
learning control problems. IEEE transactions on systems, man, and cybernetics, SMC-13(5):834–846.
Bastian, A. J. (2006). Learning to predict the future: the cerebellum adapts feedforward movement control.
Current Opinion in Neurobiology, 16(6):645–649.
Bengtsson, F. and Hesslow, G. (2006). Cerebellar control of the inferior olive. Cerebellum (London, England),
5(1):7–14.
Benna, M. K. and Fusi, S. (2016). Computational principles of synaptic memory consolidation. Nature
neuroscience.
Boyd, S. (2008). Introduction to linear dynamical systems. Online Lecture Notes.
De Zeeuw, C. I. and Yeo, C. H. (2005). Time and tide in cerebellar memory formation. Current opinion in
neurobiology, 15(6):667–74.
Dean, P., Porrill, J., Ekerot, C.-F., and Jörntell, H. (2010). The cerebellar microcircuit as an adaptive filter:
experimental and computational evidence. Nature reviews. Neuroscience, 11(1):30–43.
Eccles, J., Ito, M., and Szentágothai, J. (1967). The cerebellum as a neuronal machine. Springer Berlin.
Fujita, M. (1982). Adaptive filter model of the cerebellum. Biological cybernetics, 45(3):195–206.
Gormezano, I., Kehoe, E. J., and Marshall, B. S. (1983). Twenty years of classical conditioning with the rabbit.
Herreros, I. and Verschure, P. F. M. J. (2013). Nucleo-olivary inhibition balances the interaction between the
reactive and adaptive layers in motor control. Neural Networks, 47:64–71.
Hesslow, G. and Yeo, C. H. (2002). The functional anatomy of skeletal conditioning. In A neuroscientist's guide
to classical conditioning, pages 86–146. Springer.
Hofstoetter, C., Mintz, M., and Verschure, P. F. (2002). The cerebellum in action: a simulation and robotics
study. European Journal of Neuroscience, 16(7):1361–1376.
Jordan, M. I. (1996). Computational aspects of motor control and motor learning. In Handbook of perception
and action, volume 2, pages 71–120. Academic Press.
Kawato, M., Furukawa, K., and Suzuki, R. (1987). A hierarchical neural-network model for control and learning
of voluntary movement. Biological Cybernetics, 57(3):169–185.
Kettner, R. E., Mahamud, S., Leung, H. C., Sitkoff, N., Houk, J. C., Peterson, B. W., and Barto, a. G. (1997).
Prediction of complex two-dimensional trajectories by a cerebellar model of smooth pursuit eye movement.
Journal of neurophysiology, 77:2115–2130.
Lahiri, S. and Ganguli, S. (2013). A memory frontier for complex synapses. In Advances in neural information
processing systems, pages 1034–1042.
Lisberger, S. (1987). Visual Motion Processing And Sensory-Motor Integration For Smooth Pursuit Eye
Movements. Annual Review of Neuroscience, 10(1):97–129.
Marr, D. (1969). A theory of cerebellar cortex. The Journal of physiology, 202(2):437–470.
Massion, J. (1992). Movement, posture and equilibrium: Interaction and coordination. Progress in Neurobiology,
38(1):35–56.
McKinstry, J. L., Edelman, G. M., and Krichmar, J. L. (2006). A cerebellar model for predictive motor control
tested in a brain-based device. Proceedings of the National Academy of Sciences of the United States of
America, 103(9):3387–3392.
Porrill, J. and Dean, P. (2007). Recurrent cerebellar loops simplify adaptive control of redundant and nonlinear
motor systems. Neural computation, 19(1):170–193.
Shibata, T. and Schaal, S. (2001). Biomimetic smooth pursuit based on fast learning of the target dynamics. In
Intelligent Robots and Systems, 2001. Proceedings. 2001 IEEE/RSJ International Conference on, volume 1,
pages 278–285. IEEE.
Suvrathan, A., Payne, H. L., and Raymond, J. L. (2016). Timing rules for synaptic plasticity matched to
behavioral function. Neuron, 92(5):959–967.
Wang, S. S.-H., Denk, W., and Häusser, M. (2000). Coincidence detection in single dendritic spines mediated by
calcium release. Nature neuroscience, 3(12):1266–1273.
9
|
1902.05845 | 1 | 1902 | 2019-02-15T15:30:38 | Assessing the Level of Autonomic Nervous Activity for Effective Biofeedback Training | [
"q-bio.NC",
"physics.med-ph"
] | This paper proposes a prototype of a new biofeedback training based on mathematical models of cardiovascular control. For this purpose we develop a low-cost device that is able to record and process arterial pulse wave via photoplethysmograph and skin temperature on the peripheral part of the arm. A benchmark analysis of our device against a registered cardiovascular measurement system (Biopac MP35 from Biopac Inc., USA) shows that heart rate and skin temperature values delivered by our device are acceptable with a very low residual error (+/-1 beat/min, +/-1 oC).
Both measured signals are feed into a mathematical model which estimates the level of activity of both sympathetic and parasympathetic branches of the autonomic nervous system. That information is used jointly with other biological parameters for investigating the stress score of the subject. The data is processed in real-time and continuously displayed to the user for effective biofeedback training.
The complete solution was preliminary tested on three volunteers who used the displayed biofeedback information in order to regulate their emotional state successfully during Biofeedback training. They exhibited a significant reduction in stress score compared to three control subjects who did not used our solution during biofeedback training. This supports the benefits of biofeedback training with autonomic nervous tone assessment as effective holistic healing method. | q-bio.NC | q-bio | 1
Assessing the Level of Autonomic Nervous Activity for
Effective Biofeedback Training
Michel KANA
Dept. of Biomedical Informatics, Czech Technical University in Prague, Nám. Sítná 3105, 272 01 Kladno, Czech Republic
Abstract. This paper proposes a prototype of a new
biofeedback training based on mathematical models of
cardiovascular control. For this purpose we develop a low-
cost device that is able to record and process arterial pulse
wave via photoplethysmograph and skin temperature on
the peripheral part of the arm. A benchmark analysis of
our
cardiovascular
measurement system (Biopac MP35 from Biopac Inc.,
USA) shows that heart rate and skin temperature values
delivered by our device are acceptable with a very low
residual error (± 1 beat/min, ± 1ºC).
registered
against
device
a
Both measured signals are feed into a mathematical model
which estimates the level of activity of both sympathetic
and parasympathetic branches of the autonomic nervous
system. That
jointly with other
biological parameters for investigating the stress score of
the subject. The data is processed in real-time and
continuously displayed to the user for effective biofeedback
training.
information
is used
The complete solution was preliminary tested on three
volunteers who used the displayed biofeedback information
in order to regulate their emotional state successfully
during Biofeedback training. They exhibited a significant
reduction in stress score compared to three control
subjects who did not used our solution during biofeedback
training. This supports the benefits of biofeedback training
with autonomic nervous tone assessment as effective
holistic healing method.
Keywords
Biofeedback, pulse photoplethysmograph, autonomic
nervous tone, cardiovascular control, mathematical
modeling.
1. Background
to
is a
Biofeedback
influence physiological activities
learning process for gaining
capability
that
normally happen unconsciously, such as heart function,
breathing and skin temperature. Biofeedback can be
effective in combination with other therapies for diverse
disorders of psychical and physical function such as
anxiety, psychological stress, hypertension, attention
disorders, urinary incontinence or depressive disorders [1].
for
sensors
therapists use
Biofeedback
recording
physiological signals such as skin temperature, muscle
activity, heart rate, respiration, skin conductance, or
brainwave activity. This stream of information is processed
and presented in a simplified graphical or audible form that
allows patients to sense and influence changes in their
physiological activity in real time. The main limitation of
traditional biofeedback is that subjects receive very simple
audio and video feedback information [2].
they
important
We believe that assessing the level of autonomic
nervous activity and feeding it back to subject's awareness
will improve the effectiveness of biofeedback therapy.
Autonomic control centers in the brain monitor and
regulate
functions. They maintain body
temperature and body osmolarity; they control reproductive
intake; and
function, food
influence behavior,
emotions and
the cardiovascular control center [3].
Assessing autonomic function is usually indirectly done by
studying plasma or urinary levels of neurotransmitters,
baroreflex sensitivity or heart rate variability [4]. Results
are however not always reliable [5]. An attempt for
improving
the quantitative assessment of autonomic
cardiovascular control is to combine the information
obtained from the recorded physiological signal with the
information derived from mathematical models [6]. Such a
method is developed in this paper together with a complete
hardware and software solution for effective biofeedback
training.
Fig. 1. Biofeedback solution based on mathematical models
of cardiovascular neural control
2. Materials and Methods
As depicted in Fig. 1, our biofeedback method is
based on measurement of skin temperature and pulse wave.
The temperature is recorded using a thermocouple. The
pulse wave is recorded using a photoplethysmograph
(PPG). The data from both sensors is processed in a
2
mathematical model which provides information about the
balanced state of autonomic nervous system to the user.
2.1 Hardware Development
signal is 1000 times amplified in order to acquire a signal
in order of volts using the instrumentation amplifier
INA128 by Texas Instruments. Power supply is obtained
from ±9 volts batteries.
Fig. 4. Thermocouple circuit used for temperature
measurement
(R1=56Ω, R2=10kΩ, R3=1kΩ, R4=500Ω, R5=10kΩ,
R6=500Ω, R7=10kΩ, P1=10kΩ; C1=47nF, C2=10pF;
U1 INA128, U2 TL074)
The Object Oriented Programmable Integrated Circuit
(OOPic by Savage Innovations) microcontroller was used
for the purpose of 5 volts A/D conversion and RS-232
serial communication with the computer.
Fig. 2. Schematic of the hardware prototype
the
We use the photoplethysmographic method for heart
rate measurement [7]. Light can be transmitted through
tissues, which are full of small capillaries. They are filled
via arterial pulsations during each heart beat. The resulting
changes in volume of those vessels modify the absorption,
reflection, and scattering of the light. By measuring the
level of modification of
light source using a
photoreceptor, it is possible to measure heart rate. As light
source we used an AlGaAs
(Aluminum, Gallium,
Arsenide) LED, which produces a narrow-band source with
a peak spectral emission at a wavelength of 850 nm. As
photosensor we used a BPW42 phototransistor which is
sensible to the infrared band, from 560 to 980 nm, with
peak wavelength sensitivity at 830 nm. As depicted in Fig.
3, the output of a light-emitting diode is altered by tissue
absorption and modulates the phototransistor. The DC level
is blocked by the capacitor. A noninverting amplifier drives
low impedance loads, and provides a gain of 100. Power
supply of photoplethysmograph circuit is obtained from 4.5
volts batteries.
Fig. 5. Box with boards, power supply and supply switches
The final hardware prototype is shown in Fig. 5. We
validated our hardware solution against the well-known
Biopac MP35 system, developed by Biopac Systems Inc.,
which is a registered system for physiological data
acquisition and analysis. Simultaneous data streams from
both Biopac and our systems were statistically analyzed
and compared. The mean heart rate obtained by Biopac was
almost the same, with maximum one beat per minute error,
as the mean heart rate obtained by our system. The skin
temperature delivered by Biopac corresponded to the skin
temperature delivered by our system.
2.2 Software Development
A schematic representation of our software prototype
is shown in Fig. 6. When a biofeedback training session is
started by a user, the COM Reader opens the COM port.
Raw data stream comes from the OOPic microcontroller in
the format: Data Separator -- Pulse Value -- Temperature
Value. As a separator we used the number '2' that is not
engaged in both signals. COM Reader recognizes the Data
Fig. 3. Photoplethysmograph circuit used for heart rate
measurement
(R1=100Ω, R2= 100kΩ, R3=1.6mΩ, R4=1.6mΩ, R5= 100kΩ,
R6= 1kΩ, C=2µF; Diode LED IR850, 5nm, 160 nW,
λ=850nm; Phototransistor BPW42, 3mm, λp=830nm)
We use the Seebeck effect for skin temperature
measurement. It describes the electromotive force that
exists across a junction of two dissimilar metals.
We choose the Iron -- Constantan thermocouple type J
with 0.25 mm diameter. It has a linear characteristic in the
area around 37°C and sufficient range of values. The raw
Separator and sends the next two bytes to Buffer Manager
that stores pulse and temperature values. Buffer Manager
also immediately sends the data to Plot Manager for
display in real-time. After a time window of 10 seconds is
elapsed, the content of the buffer is sent to Pulse Processor
for processing. The Pulse Processor detects peaks in the
arterial pulse waveform and calculates the heart rate (HR).
Mean arterial pressure (MAP) is estimated from the pulse
amplitude using the equation (1) which was proposed in
[8]. The temperature values (T) are converted from volts
values to °C according the table for thermocouple type J.
HR, MAP and T are sent to Plot Manager for display.
MAP a e
= ⋅
b A
⋅
+ ⋅
c e
d A
⋅
(1)
where the unknown parameters a, b, c and d can be
determined by fitting correct blood pressure values to pulse
amplitudes during a calibration phase.
Fig. 6. Structure of a software solution for biofeedback training
After a time window of 30 seconds is elapsed, the HR
values corresponding to the buffer content are sent to the
HRV Processor that performs a heart rate variability
analysis (HRV) in the frequency band [0.015 -- 0.4] Hz by
calculating the power of LF, HF components and their ratio
LF/HF. The obtained HRV parameters are normalized over
the total power, and then sent to Plot Manager for display.
The HR, MAP, T values corresponding to the buffer
content are also sent to the Model Simulator, that simulates
a mathematical model of cardiovascular neural control (see
section 2.3 of this paper) and estimates the level of activity
of sympathetic and parasympathetic branches as well as the
so-called stress score. All three values are sent to Plot
Manager for display.
A graphical user interface was developed in the
Matlab environment according
to user requirements
gathered during on-site research with 30 respondents from
a private relaxation center and a university (see Fig. 7). It
displays measured and processed physiological data as
feedback for the user. This included a percentage of stress
score, with an appropriate color (red for high and green for
low stress score) and a smiley icon (smiling for low,
satisfied for normal and angry for high stress score).
3
Fig. 7. User interface for biofeedback training with autonomic
nervous activity assessment
2.3 Mathematical Model for Biofeedback
The level of activity of the autonomic nervous system
can be estimated using mathematical models of
cardiovascular control. The main component of such a
model would be the baroreflex or baroreceptor loop, which
stabilizes blood pressure by the mean of neural negative
feedback with three components: an afferent or sensory
component, an efferent component corresponding
to
autonomic nervous control and an effector component for
modulating heart rate and vascular resistance [9]. We
selected model equations from Ursino [10] because of their
simplicity and
the scientific
community. We adapted the equations for our biofeedback
purpose as presented below.
their acceptability
in
The Sensory Component consists of mechanoreceptor
cells located in the carotid sinus and aorta arch walls
responding to pressure by stretching. Their activity is
modeled using a linear derivative first-order dynamic block
and a sigmoidal static block as follows.
%
P P
n
−
k
ab
f
ab
=
f
ab,min
+
f
ab,max
⋅
e
%
P P
n
−
k
ab
1
+
e
τ
pb
⋅
%
dP
dt
=
P
as
+
τ
zb
⋅
dP
as
dt
−
%
P
(2)
where
abf
is the firing rate of the carotid baroreceptors in
response to the pressure sensed P% which depends on the
are the upper and lower
arterial pressure
, f
f
asP ;
ab,max
ab,min
saturation levels of firing rate with values 2.52 Hz and
47.78 Hz;
τ τ are the time constants for the real pole
,pb
zb
2.07 sec and 6.37 sec;
and the real zero in the linear dynamic block with values
nP is a model parameters, related to
the value of baroreceptor pressure at the central point of the
abk is a model parameter, with the
dimension of pressure, related to the slope of the static
function at the central point.
sigmoidal function;
4
The Efferent Component consists of sympathetic and
vagal nerves which project to blood vessels and heart.
Their activity is modeled as exponential trends using
weighted sum of afferent inputs from baroreceptors as
follows.
The changes in the heart rate depend of both
sympathetic and vagal drive as follows.
HR
=
∆
HR
S
60
+ ∆
HR HR
0
+
V
f
s
=
f
es
,
∞
+
(
f
es
,0
−
f
f
k
es
)
⋅
e
es
,
∞
es,max
f
ev
,0
+
f
ev
,
∞
⋅
e
f
ab
−
f
ab
,0
k
ev
f
ab
−
f
ab
,0
k
ev
1
+
e
−
θ
v
f
s
f
f
es,max
f
v
ev
,
∞
f
v
=
f
s
=
f
v
=
(
⋅ −
W f
⋅
,
b s
−
θ
s
)
ab
if
if
f
f
s
s
<
≥
f
f
es,max
es,max
∆
d HR
S
dt
∆
d HR
V
dt
=
=
1
τ
HR
S
1
τ
HR
V
(
⋅ −∆
HR
S
+
σ
HR
S
)
(
⋅ −∆
HR
V
+
σ
HR
V
)
(5)
(3)
σ
HR
S
σ
HR
V
G
=
G
=
HR
v
⋅
ln f
[
s
(
t D
−
HR
S
HR
S
)
−
f
es,min
+
1]
if f
s
if
f
s
≥
<
f
f
es,min
es,min
0
)
⋅
f
v
(
t D
−
HR
V
1
HR
0
=
1.97 9.5 10
−
⋅
−
3
⋅
age
where
sf is the firing rate of efferent sympathetic fibers to
vessels and heart;
vf is the firing rate of efferent vagal
es
,
,
,
f
f
k
k
es,max
fibers to heart;
∞ are constants with
values 0.06, 60 Hz, 2.1 Hz, 7 Hz, 2.1 Hz and 3.15 Hz;
vθ θ are offset
,b spW is a synaptic weight with value 0.3;
terms for sympathetic and vagal neural activation with
,s
f
,
∞
ev
,
,
ev
es
values -49 Hz and -0.68 Hz.
f
es
,0
,
f
ev
,0
are model
to
the
related
parameters,
rate of
sympathetic and vagal nerves respectively at the absence of
sf is the so-called sympathetic tone, i.e.
a quantification of sympathetic activity in the normalized
any sensory input.
intrinsic
firing
range [0 .. 1]. Similarly,
vf is the parasympathetic tone.
The Effectors Component consists of
the heart
sinoatrial node and peripheral blood vessels. The changes
in skin temperature as response to sympathetic activity are
modeled to include a pure latency, a monotonic logarithmic
static function, and low-pass first order dynamics.
( )
T t
s
∆
d T
s
dt
= ∆
( )
T t
s
+
T
s
,0
=
1
τ
T
S
(
⋅ −∆ +
T
s
σ
T
s
)
(4)
σ
T
s
=
G ln f
[
⋅
T
S
−
(
t D
T
S
s
)
−
f
es,min
+
1]
0
if f
if
f
s
s
≥
<
f
f
es,min
es,min
where
sT is the skin temperature in ºC, related to the
value as model parameter;
resistance of peripheral vessels with
,0sT being its baseline
STτ is a time constant with value
STD is the
time delay for sympathetic response to take effect with
STG is a constant gain factor with value 1.94;
6;
value 5 sec;
esf
, min
is a threshold for sympathetic stimulation
with value 2.66 Hz.
where HR is the heart rate in beats/min with
intrinsic duration of 1 cardiac cycle, depending on the age ;
HR being the
0
τ τ are a time constants with values 2 and 1.5;
,
HR
S
HR
V
G
HR
S
,
G
HR
V
are constant gain factors with values -0.13 and
0.09;
D
,
D
HR
V
HR
S
are the time delays for sympathetic and
parasympathetic responses to take effect with values 2 sec
and 0.2 sec.
2.4 Experiments
Three healthy men and three healthy women (age
37±13 years) with different employments and stress
environments were divided into two randomly selected
groups. The first one undergoes whole biofeedback training
and the second one was a non-training control group. The
control and trained group were managed according to
training plans described in the Fig. 8 and 9. All subjects
were instrumented with the hardware device we have
developed
temperature
measurement (see section 2.1).
for arterial pulse and skin
The goal of the training for both groups was to
improve the ability to calm down and relax. The methods
for relaxation were deep diaphragmatic breathing and
enhanced peripheral capillary return. Deep diaphragmatic
breathing was performed at a rate of 6 cycles per minute in
a comfortable sitting position with one hand placed on the
upper chest and the other just below your rib cage in order
to feel the diaphragm moving. The subjects were instructed
to tighten the stomach muscles, letting them fall inward as
they exhale through pursed lips, while the hand on the
upper chest remains as still as possible. Enhanced
peripheral capillary return was upheld by repeating
autogenic phrases such as "I have warm stones in my hands
which heat them". Relaxation was further enhanced in the
training group by gentle massage and few minutes resting
in yoga asana Savasana which has the same sedative effect
as some kind of antidepressants. The control group did not
undergo massage and yoga.
5
The principle of training was based on scoring the
overall stress state in order to provide the learner with
combined information about his or her heart rate, skin
temperature, heart rate variability, level of sympathetic or
parasympathetic activity and stress score in real time. The
control group did not have access to that combined
information. Instead control subjects had to achieve the
biofeedback goal without being presented with any
information about their physiological state.
Each biofeedback session had 4 phases. The pure
measurement phase is for the subject to get familiarized
with the experimental setup. The deep diaphragmatic
breathing phase is preparation phase where subjects get
relaxed by activating their parasympathetic system via deep
breathing and fictive image of warm hands. The resting
phase is a pause prior to the final phase. The Self-calming
phase is the active biofeedback session, where the subject
consciously tries to take control of its physiological
variables while reducing his stress score, sympathetic tone
and low frequency components in heart rate variability, and
increasing his parasympathetic tone and high frequency
components in heart rate variability at the same time.
Fig. 8. Experimental schedule of the training group.
Fig. 9. Experimental schedule of the control group.
2.5 Parameters estimation
Beat-to-beat heart rate (
HR
meas
) and skin temperature
( measT
) were directly
extracted
from
the pulse
plethysmograph signal which was recorded during the all
four phases of the biofeedback training. The obtained HR
time series was analyzed in the frequency domain using
Fourier transformation in order to calculate the normalized
power of low and high frequency components of heart rate
variability (LF, HF). Mean arterial pressure (
MAP
meas
) was
estimated from the amplitude of peaks in the pulse
plethysmograph signal using the mathematical equation
(1).
In order to compute the sympathetic tone
sf and the
parasympathetic
tone
vf ,
the mathematical model
described in section 2.3 was implemented in the Simulink
software environment and model parameters were
estimated by formulating and resolving a least-squares
the errors between
optimization problem minimizing
measured
skin
temperature
heart
rate,
and
HR
smea
,
MAP
,
T
meas meas
and simulated heart rate, and skin
temperature
HR P T . The corresponding cost function,
,
,as
s
i.e. the discrepancy between measured and simulated
values is defined as optimization criteria as follows, where
n is the number of data points.
J =
n
∑
[
i=1
HR
meas
]
(i) - HR(i)
2
+
[
MAP
meas
]
(i) - P (i)
2
as
+
[
T
meas
]
(i) - T (i)
2
s
(5)
The cost function was minimized in the Matlab
software environment with parameters values shown in
table 1 for a representative subject.
Model Parameter
Value
nP
abk
,0esf
,0evf
,0sT
SHRG
VHRG
92 mmHg
11.7 mmHg
16.1 Hz
3.14 Hz
25 ºC
-0.13
0.09
Tab. 1. Typical values of model parameters
The simulated sympathetic and parasympathetic tones
were combined with variables extracted from experimental
data in order to calculate the stress score using the formula
(6) below. Sympathetic tone, high fluctuations in heart rate,
average skin temperature close to 37 ºC and average heart
rate less than 70% of the intrinsic heart rate contribute to
lower stress score, meaning that the subject is fully relaxed
under predominant activity of the parasympathetic branch
of the autonomic nervous system, compare to minimal
activity of the sympathetic branch.
6
SC =
100
⋅
−
1
f HF f
v
+
−
3. Results
HR
0
−
60
HR
meas
0.7
⋅
HR
0
+
T
,0
s
37
−
LF
+
s
3
(6)
In order
to evaluate
two consecutive phases of
the effectiveness of our
biofeedback method which is based on autonomic nervous
assessment, we compared the change in different variables
between
the biofeedback
training for the control group, the trained group before
relaxation and the trained group after relaxation. Especially
we are interested in the capability of the trained group to
more effectively reduce its stress score and increase its
parasympathetic tone during self calming (phase 4).
Heart rate decreased during deep diaphragmatic
breathing (phase 2) for all groups. This trend was also
observed in the trained group during self calming. The
control group displayed an increasing heart rate instead, as
shown in Fig. 10.
Fig. 10. Change in the average heart rate.
in
The changes
the normalized power of
low
frequency components of heart rate variability, which
reflect sympathetic activity, are the most expressive for the
trained group. Subjects in the trained group had visual
information about their autonomic nervous activity and
could reduce sympathetic activity in phase 4 (see Fig. 11)
while enhancing vagal activity. In Fig. 12, the power of
high frequency components of heart rate variability, which
are known to reflect vagal activity and a state of relaxation,
significantly increases. The control group did not exhibit
any change.
Fig. 12. Changes in the normalized power of high frequency
components (HF) of heart rate variability.
As shown in Fig. 13, the control group had the highest
stress score. Stress decreased during biofeedback training
with our software solution. Relaxation prior to biofeedback
training provided additional reduction of stress score.
Fig. 13. Stress score achieved during self calming.
Differences between the control and trained group are
obvious as explained in previous paragraphs. The trained
groups show better results (reduced stress score, reduced
sympathetic tone, increased parasympathetic activity). It
was also observed that passive relaxation techniques are
valuable for effective biofeedback training. We believe that
our complete and functional solution for biofeedback
training can be used in alternative medicine, i.e. therapy
with limited adverse effects.
Future work could be in the design of a better
photosensor gripping on the finger in order to reduce noise
in the pulse signal. Moreover we should improve the
mathematical model of autonomic nervous control and
include additional neural pathways that are involved in
sympathetic and parasympathetic function. We also plan to
test the solution on more volunteers in order to bring a
solid statistical proof of its significant effectiveness in
biofeedback training. A further room of improvement
would be the inclusion of environmental parameters to our
solution such as ambient temperature, humidity and
atmospheric pressure which can be very influential on the
autonomic nervous system of human subjects.
Fig. 11. Changes in the normalized power of low frequency
components of heart rate variability.
References
[1] Schwartz, M and Andrasik, F., Biofeedback: A
Practitioner's Guide. 3rd Edition. s.l. : The Guilford Press,
2005. ISBN 1593852339.
7
[2] Arns, M, et al., "Efficacy of neurofeedback
treatment in ADHD: the effects on inattention, impulsivity
and hyperactivity: a meta-analysis." Clin EEG Neurosci.,
2009, Issue 3, Vol. 40, pp. 180-9.
[3] Unglaub, D., Human Physiology: An Integrated
[ed.] 4th Edition. s.l. : Pearson, 2009.
Approach.
0805368493.
[4] Malik, M., "Heart Rate Variability. Standards of
Measurement, Physiological Interpretation and Clinical
Use." Circulation, 1996, Issue 5, Vol. 93, pp. 1043-1065.
[5] Goldberger, J J, et al., "Dissociation of heart rate
variability from parasympathetic tone." Am J Physiol Heart
Circ Physiol, 1994, Issue 5, Vol. 266, pp. H2152-H2157.
[6] Batzel, J, et al., Cardiovascular and Respiratory
Systems -- Modeling, Analysis and Control. Philadelphia :
Society for Industrial and Applied Mathematics, 2007. pp.
6-16. ISBN 0-89871-617-9.
[7] Allen,
its
application in clinical physiological measurement." Physiol
Meas, 2007, Issue 3, Vol. 28, pp. R1-39.
J., "Photoplethysmography and
[8] Shaltis, P, Reisner, A and Asada, H., "Calibration
of the photoplethysmogram to arterial blood pressure:
capabilities and
for continuous pressure
monitoring." 2005. Conf Proc IEEE Eng Med Biol Soc.
Vol. 4, pp. 3970-3.
limitations
[9] Ottesen, J T., "Modelling of the baroreflex-
Journal of
feedback mechanism with
Mathematical Biology, 1997, Issue 1, Vol. 36, pp. 41-63.
time-delay."
[10] Ursino, M., "Interaction between carotid
baxoregulation and the pulsating heart: a mathematical
model." Am J Physiol Heart Circ Physiol, 1998, Vol. 275,
pp. 1733-1747.
|
1602.01889 | 2 | 1602 | 2016-06-11T01:45:12 | Discovering Neuronal Cell Types and Their Gene Expression Profiles Using a Spatial Point Process Mixture Model | [
"q-bio.NC",
"stat.ML"
] | Cataloging the neuronal cell types that comprise circuitry of individual brain regions is a major goal of modern neuroscience and the BRAIN initiative. Single-cell RNA sequencing can now be used to measure the gene expression profiles of individual neurons and to categorize neurons based on their gene expression profiles. While the single-cell techniques are extremely powerful and hold great promise, they are currently still labor intensive, have a high cost per cell, and, most importantly, do not provide information on spatial distribution of cell types in specific regions of the brain. We propose a complementary approach that uses computational methods to infer the cell types and their gene expression profiles through analysis of brain-wide single-cell resolution in situ hybridization (ISH) imagery contained in the Allen Brain Atlas (ABA). We measure the spatial distribution of neurons labeled in the ISH image for each gene and model it as a spatial point process mixture, whose mixture weights are given by the cell types which express that gene. By fitting a point process mixture model jointly to the ISH images, we infer both the spatial point process distribution for each cell type and their gene expression profile. We validate our predictions of cell type-specific gene expression profiles using single cell RNA sequencing data, recently published for the mouse somatosensory cortex. Jointly with the gene expression profiles, cell features such as cell size, orientation, intensity and local density level are inferred per cell type. | q-bio.NC | q-bio |
Discovering Neuronal Cell Types and Their Gene Expression
Profiles Using a Spatial Point Process Mixture Model
Furong Huang
Animashree Anandkumar
UC Irvine
UC Irvine
Christian Borgs
Microsoft Research
[email protected]
[email protected]
[email protected]
Jennifer Chayes
Ernest Fraenkel
Michael Hawrylycz
Microsoft Research
MIT
Allen Institute
Ed Lein
Allen Institute
[email protected]
[email protected]
[email protected]
[email protected]
Alessandro Ingrosso
Srinivas Turaga
Politecnico di Torino
HHMI Janelia Research Campus
[email protected]
[email protected]
Abstract
Cataloging the neuronal cell types that comprise circuitry of individual brain regions is a major goal of modern
neuroscience and the BRAIN initiative. Single-cell RNA sequencing can now be used to measure the gene expression
profiles of individual neurons and to categorize neurons based on their gene expression profiles. While the single-cell
techniques are extremely powerful and hold great promise, they are currently still labor intensive, have a high cost per
cell, and, most importantly, do not provide information on spatial distribution of cell types in specific regions of the
brain. We propose a complementary approach that uses computational methods to infer the cell types and their gene
expression profiles through analysis of brain-wide single-cell resolution in situ hybridization (ISH) imagery contained
in the Allen Brain Atlas (ABA). We measure the spatial distribution of neurons labeled in the ISH image for each
gene and model it as a spatial point process mixture, whose mixture weights are given by the cell types which express
that gene. By fitting a point process mixture model jointly to the ISH images, we infer both the spatial point process
distribution for each cell type and their gene expression profile. We validate our predictions of cell type-specific gene
expression profiles using single cell RNA sequencing data, recently published for the mouse somatosensory cortex.
Jointly with the gene expression profiles, cell features such as cell size, orientation, intensity and local density level
are inferred per cell type.
1 Introduction
1.1 Motivations and Goals
The human brain comprises about one hundred billion neurons and one trillion supporting glial cells. These cells
are specialized into a surprising diversity of cell types. The retina alone boasts well over 50 cell types, and it is an
active area of research to perform a census of the various neuronal cell types that comprise the central nervous system.
Many criteria have been used to categorize neuronal cell types, from neuronal morphology and connectivity to their
functional response properties. Neurons can also be categorized based on the proteins they make. Immunohistochem-
istry has been used with great success for many decades to differentiate excitatory neurons from inhibitory neurons by
labeling for known proteins involved in the synthesis and regulation of glutamate and GABA, the primary excitatory
and inhibitory neurotransmitters respectively.
More recently, there has been an effort to systematically measure the complete transcriptome of single neurons.
Single-cell RNA sequencing (RNA-Seq) is an extremely powerful technique that can quantitatively determine the ex-
pression level of every gene that is expressed in individual neurons. This so-called transcriptome or gene expression
1
/ transcription profile can then be used to define cell types by clustering. A recent study produced the most compre-
hensive census of cell types to date in the mouse somatosensory cortex and hippocampus by performing single-cell
RNA-Seq on over 3000 neurons [12]. While this study is quite exciting, tyring to replicate it for all brain regions
might well require the equivalent of a thousand such experiments. Thus, it is likely that the unprecedented insights
that RNA-Seq can provide will be slow to arrive. More importantly, single cell sequencing methods are not currently
able to capture the precise three-dimensional location of the individual neurons.
Here we propose a complementary approach that uses computational strategies to identify cell types and their
spatial distribution by re-analysing data published by the Allen Institute for Brain Research. The Allen Brain Atlas
(ABA) contains cellular resolution brain-wide in-situ hybridization (ISH) images for 20,000 genes1. ISH is a histo-
logical technique that labels the mRNA in all cells expressing the corresponding gene in a manner roughly proportion
to the gene expression level. An example of an ISH image can be seen in figure 1(a).
The ABA contains genome-wide and brain-wide ISH images of the adult mouse brain. These images were gener-
ated by slicing the brain into a series of 25 µm thin sections and performing ISH. Image series of ISH performed for
different genes come from different mouse brains, since ISH can only be performed for one gene at a time. The ISH
image series for different genes were then computational aligned into a common reference brain coordinate system.
Such data have been productively used to infer the average transcriptomes corresponding to different brain regions.
It is commonly thought that the ABA cannot be used to infer the transcriptomes of individual cells in a given
brain region since mouse brains cannot be aligned to the precision of a single cell. This is because there is individual
variation in the precise number and location of neurons from brain to brain. However, we expect that the average
number and spatial distribution of neurons from each cell type to be conserved from brain to brain, for a given brain
area. More concretely, we might expect that parvalbumin-expressing (PV) inhibitory interneurons in layer 2/3 of
the mouse somatosensory cortex comprise approximately 7% of all neurons and have a conserved spatial and size
distribution from brain to brain. We use this fact to derive a method for simultaneously inferring the cell types in a
given brain region and their gene expression profiles from the ABA.
We propose to model the spatial distribution of neurons in a brain as being generated by sampling from an unknown
but consistent brain-region and cell-type dependent spatial point process distribution. And since each gene might only
be expressed in a subset of cell types, an ISH image for a single gene can be thought of as a mixture of spatial point
processes where the mixture weights represent the individual cell types expressing that gene. We infer cell types, their
gene expression profiles and their spatial distribution by unmixing the spatial point processes corresponding to the ISH
images for 1743 genes. This is in notable contrast to the information provided by single-cell RNA sequencing which
can only measure the gene expression profile of individual cells to high accuracy but where, due to the destructive
measurement process, all information about the spatial position and distribution of cell types is lost.
1.2 Previous Work
Allen Brain Atlas (ABA) [9] is a landmark study which mapped the gene expression of about 20,000 genes across the
entire mouse brain. The ABA dataset consists of cellular high-resolution 2d imagery of in-situ hybridized series of
brain sections, digitally aligned to a common reference atlas. However, since the in-situ images for each gene come
from different mouse brains and since there is significant variability in the individual locations of labeled cells, it is
not possible to register brain-wide gene expression at a resolution higher than about 250µm. Therefore, the cellular
resolution detail was down-sampled to construct a coarser 3d representation of the average gene expression level in
250µm × 250µm × 250µm voxels.
The coarse-resolution averaged gene expression representation has been widely used and analyzed to understand
differences in gene expression at the level of brain region. Hawrylycz et al [5] analyzed the correlational structure
of gene expression at this scale, across the entire mouse brain. However, due to the poor resolution of the average
gene expression representation, it has proven challenging to use the ABA to discover the microstructure of gene
expression within a brain region. To address this issue from a complementary perspective, Grange et al [3] used the
gene expression profiles of 64 known cell-types, combined with linear unmixing to determine the spatial distribution of
1 Although the Atlas contains ISH data for approximately 20,000 distinct mouse genes, we focus on the top 1743 reliable genes whose sagittal
and coronal experiments are highly correlated.
2
these known cell-types. However, such an approach can be confounded by the presence of cell-types whose expression
profiles have yet to be characterized, and limited by the resolution of the averaged gene expression representation.
In contrast to previous approaches, we aim to solve the difficult problem of automatically discovering the gene
expression profiles of cell-types within a brain region by analyzing the original cellular resolution ISH imagery. We
propose to use the spatial distributions of labeled cells, and their shapes and sizes, which are a far richer representation
than simply the average expression level in 250µm × 250µm × 250µm voxels. This spatial point process is then
un-mixed to determine the gene expression profile of cell types.
Most previous work on unmixing point process mixtures adopted parametric generative models where the point
process is limited to some distribution family such as Poisson or Gaussian [6, 7]. However, since we are not interested
in building a generative model of a point process, but rather care more about inferring the mixing proportions (gene
expression profile), we take a simpler parameter-free approach. This approach models only the statistics of the point
process, but is not a generative model, and so cannot be use to model individual points/cells.
2 Modeling the Spatial Distribution of Cell-types Using Spatial Point Pro-
cess Features
Most analyses of the ABA in situ hybridization dataset have utilized a simple measure of average expression level in
relatively large 250µm × 250µm × 250µm voxels of brain tissue. Due to the large volume over which the expression
level is averaged, such a representation cannot distinguish between large numbers of cells expressing small amounts
of RNA vs. small numbers of cells expressing large amounts of RNA. All information about the spatial organization
of labeled cells, their shapes, sizes and spatial density are lost and summarized by a single scalar number. Here,
we describe a more sophisticated representation of the labeled cells in an ISH image based on marked spatial point
processes.
2.1 The Marked Spatial Point Process Representation of ISH Images
Our approach requires processing the high-resolution ISH images to detect individual labeled cells and their visual
characteristics. We developed a cell detection algorithm described in the Supplementary section. Our algorithm
additionally also estimates the expression level of each detected cell, its shape, size and orientation. Figure 1(a) and
Figure 1(b) illustrate the results of our cell detection algorithm.
Since cell-types differ not only in terms of gene expression pattern, but also display a diversity of shapes, sizes
and spatial densities, we sought to characterize these properties. We measured: (1) cell size s = [r1, r2]: the radius in
two principal directions of an ellipse fit to each cell; (2) cell orientation o: the orientation of the first principle axis
of the ellipse; (3) gene intensity level p: intensity of labeling of a cell relative to the image background; (4) spatial
distribution c: the number of cells within a local area centered around the cell, which can be regarded as a measure
of the local cell density.
The collection of detected cells within an atlas-defined brain region, along with their features, constitutes a marked
spatial point process. This point process is considered "marked", because each point is characterized by the shape,
size, expression level and local density features, in addition to just their location in space.
2.2 A Model-free Approach to Representing Spatial Point Processes Using Joint Feature
Histograms
The statistical modeling of repulsive spatial point processes such as those that arise in biology is non-trivial, and many
generative models such as determinantal point processes [8]and Matern point processes have high computational
complexity. But since we are not interested in directly modeling the individual labeled cells, but instead in modeling
only their aggregate spatial statistics, and in inferring their gene expression profiles, we can take a simpler approach.
We use a joint histogram simple statistics of the collection of detected cells to characterize the underlying point
process from which they are drawn. This is an empirical moment approach which side-steps the need to carefully
define a generative point process distribution.
3
Extract Point Process:
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
PSfrag replacements
0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
(a) Patch from gene Pvalb slice
(b) Cell detection and extraction of
spatial point process features
Joint Histogram:
Pvalb
Rasgrf2
PSfrag replacements
cell number
PSfrag replacements
cell number
(c1) Size
π/6 2π/6 3π/6 4π/6 5π/6
(c2) Orientation
PSfrag replacements
cell number
PSfrag replacements
cell number
(c3) Expression level
(c4) Cell counts in 100 µm radius
Discover Cell Types:
(d) Point process histogram
representation: [xm
n ] ∈ R+
NG×NF
(f) LDA model for inferring cell types
Figure 1: Overview of the proposed framework - Discovering Neuronal cell Types via Un-mixing of Spatial Point Process
Mixtures. (a) & (b) An in situ hybridization image for gene Pvalb along with detected cells. (c) Marginalized point process feature
histograms for genes Pvalb and Rasgrf2. Note that size denotes the principal axis diameter. We have NG genes and 4d joint
histogram with NF bins.
As we describe in the next section, we propose to model the point process measured from the ISH image for
each gene as a mixture of point processes belonging to individual cell-types. For this, we use a linear mixing model,
the Latent Dirichlet Allocation model. The use of this model is greatly simplified if we carefully choose our feature
representation such that the linear mixture of point processes results in a linear mixture of histogram statistics. This is
clearly the case for the features we have chosen. For instance, if we sample equally from two point process distributions
P1 and P2 with average densities of d1 and d2, the addition of these two point processes P = P1 + P2 results in the
addition of the two densities d = d1 + d2. This is not the case for second order features, such as the distances to the
nearest neighbors, which would have a more nonlinear relationship.
In figure 1(c), we display marginal histograms corresponding to the joint histogram for two genes, Pvalb and
4
Rasgrf2, which are well-known markers for a specific class of inhibitory and excitatory cortical neuronal cell-types
respectively.
3 Un-mixing Spatial Point Processes to Discover Cell-types
3.1 Generative Model: A Variation of Latent Dirichlet Allocation
The spatial point process histogram representation of the ABA ISH dataset results, for each brain region, is an NF ×NG
matrix [xm
n ], where NF is the total number of histogram bins (henceforward called the number of histogram features)
2, NG is the number of genes, and xm
n is the number of cells expressing gene n in histogram bin m.
We model the gene-spatial histogram matrix [xm
n ] by assuming it is generated by a Variation of Latent Dirichlet
Allocation (vLDA) [2] model of cell types. This matrix factorization based latent variable model assumes that the
ISH histograms are generated from a small number of cell-types, K, and each cell-type i is associated with a type-
dependent spatial point process histogram hi and a gene expression profile βi.
n xm
spatial distribution) is as follows: Let Lm =PNG
Our generative model for each histogram bin m (characterizing a particular bin in the size/ orientation/ gene profile/
n be the detected number of cells in the joint histogram bin m.
For each cell l in this bin, its cell-type t is sampled from the multinomial distribution hm. And given the cell-type t
of cell l, the genes n expressed by this cell are sampled from a multinomial distribution given by the type-dependent
gene expression profile/distribution βt. For a given gene n and histogram bin m, this generative process determines
n .
the number of cells that would be detected xm
We further place a Dirichlet prior over hm ∼ Dir(α), with the concentration parameter α which determines the
prior probability over the number of cell-types present in a given histogram bin m. This prior represents our prior
knowledge of how many cell-types express each gene, and also how well our feature representation separates cells of
different types into different histogram bins. In principle, we could generalize this to be a gene-specific prior, if we
had such information available. We could also use α to incorporate information about our prior knowledge over the
distribution of cells from each cell-type, for instance that excitatory neurons greatly outnumber inhibitory neurons in
a roughly 5 : 1 ratio.
We now describe how we estimate the model parameters -- the cell-type specific multinomial gene expression
profile β and the cell-type specific spatial point process histogram h from the gene-specific spatial point process
histograms measured from the ISH images.
3.2 Estimating the Cell-type Dependent Gene Expression Profile β
After testing several estimation methods for the parameters of our model, we found that non-negative matrix factor-
ization (NMF) performed well in estimating the cell-type specific gene expression profiles β, see Figure 2a. We solve
the following optimization problem:
min
β,h
NFXm
NGXn
(xm
n −
KXt
hm
t βt
nLm)2,
s.t. βt
n ≥ 0,
βt
n = 1, hm
t ≥ 0,
NGXn
KXt
hm
t = 1
(1)
Here, the non-negativity and sum-to-one constraints on hm
n ensure that h and β result in properly normalized
multinomial distributions. While this estimation procedure results in joint estimates for h and β, it does not enforce
the Dirichlet prior over h. So we refine our NMF-derived estimates for h using variational inference [2].
t and βt
2Note that there are two types of features -- the features characterizing each detected cell, and the features characterizing the collection of
detected cells that constitute a single sample from a spatial point process
5
3.3 Estimating the Cell-type Dependent Spatial Point Process Histogram h
We use a standard maximum likelihood estimation procedure for h [2]. Iteratively, we refine the inference of the cell
type membership hm ∈ ∆k under each joint histogram feature m. We update hm
i until convergence [11].
hm
i ←
1
Lm +PK
t αt
xm
n
NGXn=1
KPl=1
hm
i βi
n
hm
l βl
n
+ αi, ∀i ∈ [K], m ∈ [NF ]
(2)
Recall that the Dirichlet prior α encodes the number of cell-types that we expect on average to express each gene. We
set α to be a symmetric Dirichlet with α1 = α2 = . . . = αK, andPt αt = 0.01 for all cell-types t. In practice, we
observe that our estimates of h are fairly insensitive to the specific choice for α as long asPt αt is small enough. The
smaller α is, the fewer cell-types expressing a given gene we expect to observe in a single histogram bin.
4 Results and Evaluation
4.1 Implementation Details
We tested our proposed cell-type discovery algorithm using the high-resolution in situ hybridization image series
for 1743 of the most reliably imaged and annotated genes in the ABA. Individual cells were detected in the cellular
resolution ISH images using custom algorithms (detailed in Supplementary Information). For each detected cell, we
fit ellipses and extract several local features: (a) size and shape represented as the diameters along the principle axes of
the ellipse, (b) orientation of the first principle axis, (c) gene intensity level as measured by the intensity of labeling of
the cell body, and (d) the number of cells detected with-in a 100 µm radius around the cell, which is a measure of the
local cell density. We aligned the ISH images to the ABA reference atlas and, for this paper, focused our attention on
cells in the somatosensory cortex, since independent RNA-Seq data exist for this region the can be used to evaluate our
approach. We computed joint histograms for the collection of cells found with-in the somatosensory cortex, resulting
in a spatial point process feature vector of NF = 10010 histogram bins per gene.
Synthetic experiment:
The vLDA model we proposed is then fit to NG × NF gene point process histogram
matrix to estimate the cell-type gene expression profile matrix β using the non-negative matrix factorization (NNMF)
algorithm. The reason why we choose NNMF over Variational Inference (which is a popular approach for LDA) for β
estimation is that NNMF produces more accurate β estimation in simulated data, illustrated in Fig 2a. In the synthetic
experiment, we simulate point process data ( with some predefined golden standard β) and use the data to estimate
bβ. The errors were computed after pairing the estimated columns of β with a closest golden standard β column via
hypothesis testing. Note that the columns of β are normalized to 1, so the errors are bounded.
4.2 Evaluating Cell-type Gene Expression Profile Predictions
A recent study performed single-cell RNA sequencing on 1691 neurons isolated from mouse somatosensory cortex.
We use this dataset to evaluate the quality of the cell-types we discover.
The single cell RNA-seq data, G := [g1g2 . . . gNC ] ∈ RNG×NC , contains the gene expression profiles for
NC = 1691 cells. We infer the cell types hi for these cells using equation (2), and then compute the likelihood
Li of observing each for each cell under our estimated cell-type dependent gene expression profile matrix β using
equation (4). We can then evaluate the perplexity, a commonly used measure of goodness of fit under the vLDA
model, of single cell RNA-seq data on the model we learned from our spatial point process data.
The perplexity score is a standard metric, which is defined as the geometric mean per-cell likelihood. It is a
monotonically decreasing function of the log-likelihood L(G) of test data G.
perplexity(G) = exp(−PNC
PNC
i=1 Li
i=1 log p(gi)
6
)
(3)
PSfrag replacements
0.2
Permute β
VI estimated β
NNMF estimated β
NNMF robust
k
2.5
Spatial point process (ours)
(
permuted
)
Average expression level (baseline)
(
permuted
)
PSfrag replacements
2.0
e
p
y
T
r
e
P
r
o
r
r
E
0.15
0.1
0.05
0
3 10
20
30
40
50
Number of Cell Types
Spatial point process (ours)
Average expression level (baseline)
Spatial point process (ours, permuted)
Average expression level (baseline, permuted)
60
80
70
90 100
Perplexity Score
1.5
1.0
0.5
(a) Validate NNMF Method
Number of Cell Types
(b) Validate Point Process Data
Figure 2: (a) Synthetic Experiment : comparison of Non-negative Matrix Factorization (NNMF) with Variational
Inference (VI) on simulated point process cell data using known gene expression profile β. An additional robustness
test of NNMF is done to see how good the algorithm is when a wrong number of cell types K is input. A permutation
test (shuffling the gene expression levels between cell) is done to access statistical significance. Comparing with
permute test shows that our cell-types are significantly different from chance. Error per type is computed by pairing the
columns of estimatedbβ with the columns of the ground-truth β. (b) Comparison of gene expression profiles recovered
for cell-types in the somatosensory cortex by fitting an LDA model using spatial point process features (ours) vs the
standard average gene expression level feature (baseline). Our features provide a significantly better match, with lower
perplexity, to ground truth single-cell RNA sequencing derived transcriptomes. A permutation test is done to access
statistical significance. Perplexity is computed by matching to surrogate single-cell RNA transcriptomes by shuffling
the gene expression levels between cells. Comparing with permute test shows that our cell-types are significantly
different from chance.
where the likelihood is evaluated as
p(gmhm, α, β) =
Γ (Pi αi)
Qi Γ (αi)
(hm
i )αi−1
kYi=1
LmYj=1 kXi=1
NGXn=1
δgm
j ,en hm
i βi
n! .
(4)
where δi,j is the Kronecker delta, δi,j = 1 when i = j and 0 otherwise. en is the nth basis vector.
4.3 Comparison to Standard Average Gene Expression Features Baseline and a Permuta-
tion Test for Significance
Here we demonstrate the superiority of our method and its statistical significance in two ways. First we compared
the perplexity of the single-cell RNA seq dataset G under our model (figure 2b, solid blue) against the perplexity
of a surrogate dataset with the same marginal statistics, but whose gene-cell correlations were destroyed (figure 2b,
dashed blue). We generated this surrogate dataset by randomly permuting the gene expression levels for each gene
across cells. This permuted dataset had a significantly higher (worse) perplexity than the true single-cell dataset. This
demonstrates that our model trained to un-mix the ISH-derived spatial point processes discovered cell-types whose
gene expression profiles are significantly better match to single-cells than by chance.
We also compared the predictions of cell-type gene expression profiles derived by un-mixing our spatial point
process features against gene expression profiles derived by un-mixing the more standard 250µm × 250µm × 250µm
averaged gene expression level features. We see a very large improvement in perplexity by switching from the standard
simple averaging of gene expression, to extracting spatial point process features (figure 2b). The single-cell RNA seq
dataset analysis from figure 2b shows that the perplexity of our recovered cell-types rapidly flattens after we recover
approximately 10 clusters (K = 10).
7
[2,4]
[3,3]
[3,4]
[3,5]
[3,6]
[3,7]
[4,4]
[4,5]
[4,6]
[4,7]
[5,7]
[6,6]
[6,7]
[7,7]
[7,8]
[7,9]
[8,8]
[8,9]
[9,9]
Gad1
Sp8
Tox3
Nkx2-1
Lhx6
Pax6
Dlx5
Arx
Dlx2
Dlx1
Elavl2
Sp9
Tbr1
Foxp2
Tshz2
CA1Pyramidal
Stat4
Ascl1
Cux2
Neurod1
Mef2c
Oligodendrocytes
S1Pyramidal
Ependymal
Microglia
Astrocytes
Interneurons
Endothelial
M ural
PSfrag replacements
PSfrag replacements
Ptrf
Cldn5
Maf
Hcls1
Spi1
Myb
Fhl1
Aldoc
Sall3
Sox21
Mbp
CA1 Pyramidal
Etv6
Sox10
St18
Olig2
Oligodendrocytes
S1 Paramidal
Ependymal
Microglia
Astrocytes
Interneurons
Endothelial
M ural
Figure 3: Estimated memberships β on marker genes for 8 cell types. These marker genes are used to label the
columns of the membership matrix.
PSfrag replacements
CA1Pyramidal
Cell Fraction
Cell Fraction
4.4 A Brief Analysis of Recovered Cell Types in Somatosensory Cortex
Cell Fraction
Cell Fraction
π
6
In this section we describe the representative spatial point process statistics and gene expressions for 8 cell-types
we recovered. We attempted to align our 8 clusters to cell-types defined by [12] in the single-cell RNA sequencing
paper. We found high overlap in the gene expression profiles for all 8 clusters with known cell-types defined in [12],
Interneurons, S1 Pyramidal, Mural, Endothelial, Microglia, Ependymal, Astrocytes and Oligodendrocytes, in Figure 3.
4π
6
3π
6
2π
6
5π
6
Axis 1
Axis 2
CA1Pyramidal
150
100
50
9
8
7
6
5
4
3
0.9
0.8
0.7
0.6
0.5
0.4
0.3
CA1Pyramidal
50
40
30
20
10
CA1Pyramidal
2
n s
I n ter n e u r o
y ra
P
o
O li g
1
S
d r o c y te s
d ial
m
M icr o
d e n
E
g lia
o t h elial
A str o c y te s
p e n
d
n
E
al
m
M
y
d
u ral
0
n s
I n ter n e u r o
y ra
P
o
O li g
1
S
d r o c y te s
d ial
m
M icr o
d e n
E
o t h elial
g lia
A str o c y te s
p e n
d
n
E
y
d
m
al
M
u ral
0.2
n s
I n ter n e u r o
y ra
P
o
O li g
1
S
d r o c y te s
d ial
m
M icr o
d e n
E
o t h elial
g lia
A str o c y te s
p e n
d
n
E
al
m
M
y
d
u ral
(a) Cell diameter in principal axes
(b) Orientation
(c) Intensity
0
S
1
al
d r o c y te s
d ial
m
M icr o
d e n
E
g lia
o t h elial
A str o c y te s
p e n
d
n
n s
I n ter n e u r o
y ra
P
o
O li g
(d) Cells in 100 µm radius
u ral
M
m
y
d
E
Figure 4: Figure of 5% and 95% percentile estimated cell features for 8 cell types we detected. Inference is performed
on the Spatial point process histograms data we estimated.
The estimate of β was combined with MLE to infer the cell-type specific spatial point process representation hm
.
l
In examining the spatial point process distributions that we predict for each of these cell types, we discover that while
the distribution of cell body orientations is quite broad and similar across cell types, the cell count distribution, which
is a measure of cell density, varies in a systematic way from one cell type to another. Fig 4d shows that inhibitory
Interneurons are less dense than S1Pyramidal neurons. This is consistent with their known prevalence, roughly 20%
of all neurons are GABAergic interneurons [10], while the remaining 80% are excitatory glutamatergic pyramidal
neurons. As expected, this excitatory neuronal category of S1Pyramidal is the most common and hence most dense
class of neuronal cells. They also have slightly larger cell bodies, compared to interneurons, as can be seen in Fig 4a.
The remaining 6 cell types correspond to various glial sub-types.
8
5 Conclusion
We developed a computational method for discovering cell types in a brain region by analyzing the high-resolution in
situ hybridization image series from the Allen Brain Atlas. Under the assumption that cell types have unique spatial
distributions and gene expression profiles, we used a varied latent Dirichlet allocation (vLDA) based on spatial point
process process mixture model to simultaneously infer the cell feature spatial distribution and gene expression profiles
of cell types. By comparing our gene expression profile predictions to a single-cell RNA sequencing dataset, we
demonstrated that our model improves significantly on state of the art.
The accuracy of our method relies heavily on the assumption that cell-types differ in their spatial distribution, and
that our point process features perform a good job of distinguishing these differences. Thus the performance of our
method can be improved by better estimates of better features. We would expect our method to perform better for
large brain areas, which can be more accurately aligned, and which have more cells to estimate point process features.
There are several modifications to our vLDA model which might improve the faithfulness of our generative model
to the biology. We place a symmetric Dirichlet prior over cell-type multinomial distribution hm for a given histogram
bin m. This assumes that the number of cell-types expressing each gene is the same for all genes. But since some genes
are expressed more commonly and non-specifically than others, we might expect a gene-specific prior to be a better
model. Further, the symmetric Dirichlet assumes that all cell-types have equal proportions of cells. But evidence
suggests that excitatory neurons are more common than inhibitory neurons in cortex [4], and using a non-uniform
Dirichlet prior could account for this.
References
[1] Website:
2014 allen institute for brain science. allen mouse brain atlas [internet].
Available from:
http://mouse.brain-map.org/. Accessed: 2014-11-06.
[2] David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. the Journal of machine Learning
research, 3:993 -- 1022, 2003.
[3] Pascal Grange, Jason W Bohland, Benjamin W Okaty, Ken Sugino, Hemant Bokil, Sacha B Nelson, Lydia Ng,
Michael Hawrylycz, and Partha P Mitra. Cell-type -- based model explaining coexpression patterns of genes in the
brain. Proceedings of the National Academy of Sciences, 111(14):5397 -- 5402, 2014.
[4] Kenneth D Harris and Thomas D Mrsic-Flogel. Cortical connectivity and sensory coding. Nature, 503(7474):51 --
58, 2013.
[5] Mike Hawrylycz, Lydia Ng, Damon Page, John Morris, Chris Lau, Sky Faber, Vance Faber, Susan Sunkin, Vilas
Menon, Ed Lein, et al. Multi-scale correlation structure of gene expression in the brain. Neural Networks,
24(9):933 -- 942, 2011.
[6] Chunlin Ji, Daniel Merl, Thomas B Kepler, and Mike West. Spatial mixture modelling for unobserved point
processes: Examples in immunofluorescence histology. Bayesian analysis (Online), 4(2):297, 2009.
[7] Athanasios Kottas and Bruno Sans´o. Bayesian mixture modeling for spatial poisson process intensities, with
applications to extreme value analysis. Journal of Statistical Planning and Inference, 137(10):3151 -- 3163, 2007.
[8] Alex Kulesza and Ben Taskar. Determinantal point processes for machine learning. Machine Learning, 5(2-
3):123 -- 286, 2012.
[9] Ed S Lein, Michael J Hawrylycz, Nancy Ao, Mikael Ayres, Amy Bensinger, Amy Bernard, Andrew F Boe,
Mark S Boguski, Kevin S Brockway, Emi J Byrnes, et al. Genome-wide atlas of gene expression in the adult
mouse brain. Nature, 445(7124):168 -- 176, 2007.
[10] Henry Markram, Maria Toledo-Rodriguez, Yun Wang, Anirudh Gupta, Gilad Silberberg, and Caizhi Wu. In-
terneurons of the neocortical inhibitory system. Nat Rev Neurosci, 5(10):793 -- 807, October 2004.
9
[11] Alexander Smola and Shravan Narayanamurthy. An architecture for parallel topic models. Proceedings of the
VLDB Endowment, 3(1-2):703 -- 710, 2010.
[12] Amit Zeisel, Ana B Munoz-Manchado, Simone Codeluppi, Peter Lonnerberg, Gioele La Manno, Anna Jur´eus,
Sueli Marques, Hermany Munguba, Liqun He, Christer Betsholtz, et al. Cell types in the mouse cortex and
hippocampus revealed by single-cell rna-seq. Science, 347(6226):1138 -- 1142, 2015.
10
Appendix for Discovering Neuronal Cell Types and Their Gene Expression
Profiles Using a Spatial Point Process Mixture Model
A Morphological Basis Extraction
We aim to characterize the morphological basis for all cells with different size, orientation, expression profiles and
spatial distribution. The traditional sparse coding introduces too many free parameters and is not suitable for compact
morphological basis learning. We instead propose Gaussian prior convolutional sparse coding (GPCSC). The intuition
for using convolution is due to the frequent replication of cells of similar shapes and the translation invariance property.
Traditional sparse coding would learn both the shape of the cell and the location of the cell. But the convolutional
sparse coding would only learn the shape here. We characterize cell spatial distribution via decoding the sparse
activation map.
To formulate the problem formally: let I be the image observed, then the convolutional sparse coding model
generates observed image I using filters (resembling cell shapes)F superposed at locations indicated by the activation
map M (whose sparsity pattern indicates cell spatial distribution and activation amplitude indicates gene expression
profiles. )
Our goals of segmenting cells, extracting cell basis, and estimating gene profiles and cell locations are reduced to
this optimization learning problem:
min
Fm,M n
m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xn
2
F
m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
I n −
kXm=1
Fm ⋆ M n
+Xn Xm
λ kM n
mk0 , s.t. Fm(x, y) ≥ 0, kFmk2
F = 1, M (n)
m (x, y) ≥ 0.
(5)
where I n is the nth image associated with the gene we are interested in with Dx × Dy pixels, i.e., I n ∈ RD×D.
We call the Fm ∈ Rd×d filter, where d is set to capture the local cell morphological information. The spatial
m ∈ R(D−d+1)×(D−d+1) which represents the position of the filter Fm being
m(x, y) = 1, then Fm is active at I n(x : x + d − 1, y : y + d − 1).
coefficient for image I n is denoted as H (n)
active on image I n. More precisely, if H n
A.1 Gaussian Prior Convolutional Sparse Coding
The popular alternating approach between matching pursuit to learn activation map M and k-SVD to learn F is
general applicable to any object detection problem in image processing. However, this approach causes inexact cell
number estimation as filters with multi-modality (i.e., multiple cells) are learnt. We resolve this issue by proposing an
Gaussian probability density function prior on the filters to guarantee single cell detection and achieve accurate cell
number estimation. The support of M is also limited to the local maxima indicating cell centers. Note that our cell are
not donut shaped, and it is reasonable to assume the darkest point being the cell center.
Therefore, we optimize over the objective min kPn I n −Pm Fm ⋆ M n
solving the optimization problem. If we define the residual asPn I n−PnPm bFm ⋆cM n
mk0 such that Fm
are 2 − D Gaussian densities with priori set top 2 principal radius and orientation. Alternating Minimization is used to
m, the gradient of the objective
reduced to an iterative approach of updating filters, compute residual, optimizing activation map based on residual,
(i, j) are convolution of the
compute residual and updating filters again. It is easy to see that both ∂L
∂Fm
residual and the other variable rotated by angle π.
2 +PnPm λ kM n
(i, j) and ∂L
∂Hm
mk2
A.2 Image Registration/Alignment
A structure represents a neuronanatomical region of interest. Structures are grouped into ontologies and organized in
a hierarchy or structure graph. We are interested in the somatosensory cortex area. So we use the affine transform
from Allen Brain Institute [1, 9] to align all the in-situ hybridization images with the Atlas brain to extract the correct
region.
11
|
1805.11704 | 2 | 1805 | 2018-06-29T20:17:16 | Deep Semantic Architecture with discriminative feature visualization for neuroimage analysis | [
"q-bio.NC",
"cs.AI",
"q-bio.QM"
] | Neuroimaging data analysis often involves \emph{a-priori} selection of data features to study the underlying neural activity. Since this could lead to sub-optimal feature selection and thereby prevent the detection of subtle patterns in neural activity, data-driven methods have recently gained popularity for optimizing neuroimaging data analysis pipelines and thereby, improving our understanding of neural mechanisms. In this context, we developed a deep convolutional architecture that can identify discriminating patterns in neuroimaging data and applied it to electroencephalography (EEG) recordings collected from 25 subjects performing a hand motor task before and after a rest period or a bout of exercise. The deep network was trained to classify subjects into exercise and control groups based on differences in their EEG signals. Subsequently, we developed a novel method termed the cue-combination for Class Activation Map (ccCAM), which enabled us to identify discriminating spatio-temporal features within definite frequency bands (23--33 Hz) and assess the effects of exercise on the brain. Additionally, the proposed architecture allowed the visualization of the differences in the propagation of underlying neural activity across the cortex between the two groups, for the first time in our knowledge. Our results demonstrate the feasibility of using deep network architectures for neuroimaging analysis in different contexts such as, for the identification of robust brain biomarkers to better characterize and potentially treat neurological disorders. | q-bio.NC | q-bio |
Deep Semantic Architecture with discriminative
feature visualization for neuroimage analysis
Arna Ghosh
Integrated Program in Neuroscience
McGill University
Montréal, QC H3A 0G4. Canada
[email protected]
Fabien Dal Maso
École de kinésiologie et des
sciences de l'activité physique
Université de Montréal
Montréal, QC H3T 1J4. Canada
[email protected]
Marc Roig
School of Physical and
Occupational Therapy
McGill University
Montréal, QC H3A 0G4. Canada.
[email protected]
Georgios D Mitsis
Department of Bioengineering
McGill University
Montréal, QC H3A 0G4. Canada.
[email protected]
Marie-Hélène Boudrias
School of Physical and
Occupational Therapy
McGill University
Montréal, QC H3A 0G4. Canada.
[email protected]
Abstract
Neuroimaging data analysis often involves a-priori selection of data features to
study the underlying neural activity. Since this could lead to sub-optimal feature
selection and thereby prevent the detection of subtle patterns in neural activity, data-
driven methods have recently gained popularity for optimizing neuroimaging data
analysis pipelines and thereby, improving our understanding of neural mechanisms.
In this context, we developed a deep convolutional architecture that can identify
discriminating patterns in neuroimaging data and applied it to electroencephalogra-
phy (EEG) recordings collected from 25 subjects performing a hand motor task
before and after a rest period or a bout of exercise. The deep network was trained to
classify subjects into exercise and control groups based on differences in their EEG
signals. Subsequently, we developed a novel method termed the cue-combination
for Class Activation Map (ccCAM), which enabled us to identify discriminating
spatio-temporal features within definite frequency bands (23–33 Hz) and assess the
effects of exercise on the brain. Additionally, the proposed architecture allowed
the visualization of the differences in the propagation of underlying neural activity
across the cortex between the two groups, for the first time in our knowledge.
Our results demonstrate the feasibility of using deep network architectures for
neuroimaging analysis in different contexts such as, for the identification of robust
brain biomarkers to better characterize and potentially treat neurological disorders.
Preprint. Work in progress.
1
Introduction
Skilled motor practice facilitates the formation of an internal model of movement, which may be
later used to anticipate task specific requirements. These internal models are more susceptible to
alterations during and immediately following practice and become less susceptible to alterations over
time, a process called consolidation [5, 23]. A single bout of cardiovascular exercise, performed
in close temporal proximity to a session of skill practice, has shown to facilitate motor memory
consolidation [24]. Several potential mechanisms underlying the time-dependent effects induced by
acute exercise on motor memory consolidation have been identified, such as increased availability of
neurochemicals [28] and increased cortico-spinal excitability [19]. However, the distinct contribution
of specific brain areas and the precise neurophysiological mechanisms underlying the positive effects
of acute cardiovascular exercise on motor memory consolidation remain largely unknown.
Electroencephalography (EEG) is a popular technique used to study the electrical activity from
different brain areas. The EEG signal arises from synchronized postsynaptic potentials of neurons
that generate electrophysiological oscillations in different frequency bands. During movement,
the EEG signal power spectrum within the alpha (8–12 Hz) and beta (15–29 Hz) range decreases
in amplitude and this is thought to reflect increased excitability of neurons in sensorimotor areas
[7, 18, 20, 25]. This phenomenon is termed Event-Related Desynchronization (ERD). Alpha- and
beta-band ERD have been shown to be modulated during motor skill learning in various EEG studies
[4, 14, 34]. There is converging evidence suggesting an association of cortical oscillations in the
motor cortex with neuroplasticity events underlying motor memory consolidation [4, 22]. In this
context, our aim was to study the add-on effects of exercise on motor learning in terms of modulation
of EEG-based ERD.
Many neuroimaging studies, including EEG ones, rely on the a-priori selection of features from the
recorded time-series. This could lead to sub-optimal feature selection and could eventually prevent the
detection of subtle discriminative patterns in the data. Alternatively, data-driven approaches such as
deep learning allow discovery of the optimal discriminative features in a given dataset. Convolutional
Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have been applied to computer
vision and speech processing datasets [17, 12, 32, 15] with great success. They have also been used
successfully in the neuroimaging domain to learn feature representations for Magnetic Resonance
Image segmentations [21] and EEG data decoding [2, 26, 30] among others. Most studies using
CNNs for EEG have been restricted to the classification of EEG data segments into known categories.
However, the usefulness of CNNs to improve our understanding of the underlying neural bases is
less straightforward, primarily due to the difficulty into visualizing and interpreting the feature space
learnt by the CNN.
Our work addresses existing caveats in applying deep learning architectures, such as CNNs, to
analyzing EEG data by means of three novel contributions –
1. We used two parallel feature extraction streams to discover informative features from EEG
data before and after an intervention and subsequently characterize the modulatory effect on
these derived features rather than on the raw EEG data itself;
2. We incorporated a subject prediction adversary component in the network architecture to
learn subject-invariant, group-related features instead of subject-specific features;
3. We developed a novel method, termed cue-combination for Class Activation Map (ccCAM),
to visualize the features extracted by the CNN after training
We used this CNN-based deep network architecture to identify exercise-induced changes in neural
activity from EEG signals recorded during an isometric motor task. The training was carried out in a
hierarchical structure – first for time-frequency and then for topographical data maps. Visualizing the
features after each stage of the training allowed us to identify frequency bands and the corresponding
brain regions that were modulated by the add-on effects of acute exercise on motor learning.
The majority of previous related studies have leveraged large-scale datasets consisting of hundreds of
subjects for training purposes, which may be a limiting factor for applying powerful deep learning
methodologies to data of smaller sample size. Therefore, one of our aims was to develop a method
that can be used both for small-scale and large-scale studies. To this end, we added a regularizer
that prevented the feature extraction part of the CNN from learning subject-specific features, thus
promoting the identification of group-specific features only.
2
2 Dataset
The dataset used in this work consisted of EEG recordings from 25 healthy subjects. The experiment
and data collection are detailed elsewhere [8]. Briefly, 25 right-handed healthy subjects were recruited
and assigned to the Control (CON, n=13 subjects) or Exercise (EXE, n=12 subjects) groups.
Figure 1: Illustration of Experiment Protocol.
Each subject reported to the laboratory on four occasions as shown in Figure 1. Visit 1 required
the participants to go through a Graded Exercise Test (GXT), which was used to determine their
cardiorespiratory fitness. Visit 2 was conducted at least 48 hrs after the GXT to avoid potential long-
term effects of exercise on memory [3, 13]. EEG recordings were collected at baseline while subjects
performed isometric handgrips, which corresponded to 50 repetitions of visually cued isometric
handgrips with their dominant right hand using a hand clench dynamometer (Biopac, Goleta, CA,
USA). Each contraction was maintained for 3.5 sec at 15% of each participant's maximum voluntary
contraction (MVC) and followed by a 3 to 5 sec rest period. The baseline assessment was followed by
the practice of a visuo-motor tracking task (skill acquisition), which was used for the calculation of
the motor learning score. Participants were then randomly assigned to two groups. The exercise group
(EXE) performed a bout of high-intensity interval cycling of 15 min, while the control group (CON)
rested on the cycle ergometer for the same amount of time. The same EEG recordings collected at
baseline were repeated 30, 60 and 90 min after the exercise or rest period.EEG activity was recorded
using a 64-channel ActiCap cap (BrainVision, Munich, Germany) with electrode locations arranged
according to the 10–20 international system. Electrical conductive gel was inserted at each electrode
site to keep impedances below 5 kΩ. EEG signals were referenced to the FCz electrode and sampled
at 2500 Hz.
3 Methods
The analysis pipeline was first applied to the time and frequency domain data without incorporating
spatial information. Subsequently, it was applied to the data obtained by creating topographical maps
corresponding to distribution of activity in specific frequency bands across the cortex. The entire
pipeline consisted of 3 segments , i.e.– Preprocessing, CNN training and ccCAM generation.
3.1 Time-Frequency (TF) maps
3.1.1 Preprocessing
EEG data preprocessing was similar to that performed in a previous study [8] and was performed
using the Brainstorm Matlab toolbox [29]. EEG signals were bandpass-filtered between 0.5 Hz and
55 Hz and average-referenced. Continuous data were visually inspected and portions of signals
with muscle or electrical transient artifacts were rejected. Independent component analysis (ICA)
was subsequently applied on each dataset (total number of components: 20) and between one and
two eye-blink related components were rejected based on their topography and time signatures [9].
The resulting dataset was epoched with respect to the period of time (3.5 sec) corresponding to the
appearance of the visual cue that triggered the initiation of the isometric handgrips (n = 50/subject).
Finally, each trial was visually inspected and those containing artifacts were manually removed.
Morlet wavelet (wave number = 7) coefficients between 1 to 55 Hz with 1 Hz resolution were
extracted to obtain time-frequency decompositions of the EEG data. The time-frequency data for
3
each electrode were consequently normalized with respect to their spectral power before the start
of the grip event, as calculated from a window of 0.5 sec. Following this, an average over all trials
was calculated in order to obtain a single time-frequency map for each electrode. Further steps were
applied on the EEG recording segment corresponding to 0.5–3.5 sec after the appearance of the visual
cue, i.e. during the handgrip task, to perform the subsequent analysis.
3.1.2 CNN training
t and As
t and As
The overall CNN architecture that we developed is shown in Figure 2. Following preprocessing of
the data, time-frequency maps for each electrode and session – at baseline and 90 min after exercise
or a rest period (post-intervention session) – were obtained. The data for each session was then
rearranged to form 2D matrices comprising of the frequency spectra for all electrodes at a given time
instant t. Each matrix had a dimension of 64 × 55 (64 electrodes × 55 frequency bands). For training
the network, a pair of matrices was used – the first corresponding to time point t from the baseline
session and the second corresponding to the same time point t from the post-intervention session.
Each pair was labeled as either exercise or control, depending on the group allocation. Structuring the
data in this fashion allowed the network to take into account the inter-subject variability in baseline
measures and therefore did not require the experimenter to adopt techniques for normalizing the EEG
signal from the post-intervention session with respect to the baseline session. Thus, the network was
expected to capture the EEG features that were modulated by the add-on effects of acute exercise.
Dataset Notation:- B and A represent the entire data tensor at baseline and post-intervention respec-
tively. Each data tensor consists of data matrices from all 25 subjects and timepoints. For subject
s, the goal was to classify whether the tuple containing the matrices Bs
t (where t denotes
timepoint) belongs to the EXE or CON groups.
To this end, we used a deep convolutional network that was optimized for the task. The network
architecture is similar to the one used in [1]. Features from matrices Bs
t were extracted
using a network termed the Base CNN. The difference between the obtained feature vectors was
passed to a discriminator network, termed the Top NN, to predict the correct group in which each pair
belongs to. The schematic view of the architecture is shown in Figure 2 and details of each network's
architecture are provided in Tables S1 and S2 in supplementary material respectively. Since the
sampling frequency was 2500 Hz and the time period of interest was 3 sec long, t ∈ [1, 7500] for
each subject s.
The convolutions performed in the Base CNN were with respect to the frequency domain and not
the electrode (sensor) domain. This is because the former was laid out in a semantic order of
increasing frequencies, as opposed to the latter, which was not arranged by the spatial locations of the
electrodes. Consequently, we expected the features extracted by the Base CNN to be the frequency
bands significantly affected by exercise. Therefore, all convolutional filters in the Base CNN were
implemented as 1 × n 2D filters, where n is the extent of the filter in the frequency domain. The
same holds for the Max-Pooling layers.
Initially, a network that did not include an adversary loss component (Figure S1.a from Supplementary
material) was used; however, it was found that this network was able to learn subject-specific
features as opposed to subject-invariant, exercise-related features. This is illustrated in Figure S2
(Supplementary material) and Table 1. In most neuroimaging studies, the number of participants
scanned is limited, which typically restricts deep networks from learning subject-invariant features.
To address this issue, we followed a domain adaptation approach. Specifically, each subject was
considered as a separate domain comprising of subject-specific features along with subject-invariant,
exercise-related features. Since our goal was to learn features mainly related to the effect of exercise
on the consolidation of motor memory, we incorporated the domain confusion strategy [31] to
train the network, thus adding the subject discriminator as an adversary (Figure 2 – bottom right).
Specifically, we added this network in parallel to the Top NN with similar model capacity (see Table
S3 in supplementary material for architecture details).
Network Architecture Notation:- The feature extractor operation and parameters of the Base CNN
are denoted as fθf and θf respectively, the Top NN feature discrimination operator and its parameters
are denoted by hθt and θt respectively, while the subject discrimination operator and its parameters
are denoted by hθs and θs respectively. The input tuple is denoted by x and its corresponding group
and subject labels by yg and ys respectively. We used the Negative Log Likelihood (NLL) loss for
each classifier with the Adam optimizer [16] in Torch [6] for training the network. The Subject
4
Figure 2: Modified deep network architecture with an adversary component (bottom right) to avoid
subject discrimination. The adversary is a novel addition to enable the use of CNNs to smaller-scale
neuroimaging studies with a limited number of subjects. Dimensions corresponding to TF maps are
shown here.
Discriminator was trained to minimize the subject prediction loss given by –
Js(θs, θf ) = −[
m
X
i=1
logh(y(i)
s )
θs
(fθf (x(i)))]
The Top NN was trained to minimize the group prediction loss given by –
Jg(θt, θf ) = −[
m
X
i=1
logh
(y(i)
g )
θt
(fθf (x(i)))]
(1)
(2)
We trained the feature extractor, Base CNN, in a manner such that the features extracted would be
agnostic to the originating subject, therefore, the target distribution for the subject prediction network
was a uniform distribution. Hence, we used the domain confusion loss [31] over the gradient reversal
layer [11] and used the Kullback-Leibler (KL) divergence from the uniform distribution over 25
classes (25 subjects) as our loss metric. Conclusively, the Base CNN was trained to minimize the loss
given by –
Jf (θf , θt, θs) = −[
m
X
i=1
logh
(y(i)
g )
θt
(fθf (x(i)))] + λ[
m
X
i=1
KL(U, hθs (fθf (x(i)))]
(3)
where KL(P, Q) denotes the KL divergence between distributions P & Q, U denotes the uniform
distribution, m denotes the total number of training examples and λ is a hyperparamater that deter-
mines the weight for the subject discrimination regularizer. Here, we used a 80-20 split of the data
set, whereby 80% was used for training and 20% was used for validation.
ccCAM
3.1.3
A major contribution of the present work is the development of a novel method for the visualization
of the features that guide the proposed network's decision. Although well-known techniques used in
the computer vision literature include the use of Global Average Pooling (GAP) [33] and grad-CAM
[27], these methods are not suited for the neuroscience paradigm considered here. For instance, GAP
requires averaging the activations of each filter map, i.e. each channel of the extracted feature tensor.
This leads to loss of information related to electrode positions, as convolutions were performed only
in the frequency domain. Specifically, we applied GAP and grad-CAM to our data and we were
unable to obtain adequate classification accuracy (≈ 56%) with a GAP layer in the network. Also,
grad-CAM is sensitive to absolute scale of the features in the input data and hence yielded results that
were biased towards frequency bands with higher power-values, namely the lower frequency bands
(<10 Hz).
5
Given these limitations in existing analytic methods, we used the linear cue-combination theory used
in human perception studies [10] to develop a method that explains the network's decisions. Let
us consider for example, a CNN with only 2 channels , i.e. filter maps, in the final feature tensor
extracted after convolutions. Each of these filter maps preserve the spatial and/or semantic structure
of the input data. Each of these filter maps acts as a "cue" to the network's classifier layers, denoted
as c1 and c2. If we denote the desired class label as y1 and assuming c1 and c2 to be independent to
each other, we can use Bayes' Theorem to write –
P (y1c1, c2) =
P (c1, c2y1)P (y1)
P (c1, c2)
=
P (c1y1)P (c2y1)P (y1)
P (c1)P (c2)
=
P (y1c1)P (y1c2)
P (y1)
(4)
If the likelihood for predicting y1 due to cue ci is Gaussian with mean µi and variance σ2
i , the
maximum likelihood estimate (MLE) yields the combined cue, denoted by c∗, that summarizes the
important features on which the network bases its decisions. Therefore, the combined cue, c∗, is the
desired Class Activation Map (CAM).
c∗ =
2
X
i=1
wiµi where wi =
1/σ2
i
2
1/σ2
i
P
i=1
Since the network is trained, µi = ci. To calculate the values of σi, we used the NLL loss values.
The NLL loss with a cue removed provided an estimate of the σ associated with that cue.
(5)
(6)
(7)
ǫ = −logP (y1c1, c2)
= −logP (y1c1) − logP (y1c2) + logP (y1) [From eq 4]
ǫ1 = ǫc1=0 = −logP (y1c1 = 0) − logP (y1c2) + logP (y1)
ǫ1 − ǫ = logP (y1c1) − logP (y1c1 = 0)
µ2
1
2σ2
1
2(ǫ1 − ǫ)
µ2
1
=
=
21
1 σ
Therefore,
Using the estimated σi, the CAM corresponding to the correct class for each input was generated.
Since in the present case µi corresponds to a 2D matrix, the denominator in Equation 7 was replaced
by the mean-squared value of the corresponding matrix. The obtained CAMs were subsequently
group-averaged to extract frequency bands that contain features characteristic to each group (CON
and EXE).
3.2 Topographical maps
Topographical maps were created using the frequency bands obtained from the ccCAM corresponding
to the TF-maps. The average power within each frequency band for all electrodes at time point t
was used to construct a 64 × 64 matrix by projecting the average power value of each electrode to a
point corresponding to the its spatial position. Since this procedure yielded a sparse matrix, cubic
interpolation was used to obtain a continuous image depicting the distribution of activity within each
frequency band over the entire head. A total of three such matrices were packed together to form a
3 × 64 × 64 tensor corresponding to activity maps at times t, t + 1 and t + 2 respectively. The entire
data tensor for a given subject was created by taking non-overlapping time windows. Hence, the total
number of tensors for each subject was equal to 2500.
Similar to the analysis of TF-maps, we trained a CNN-based network to classify each data tensor into
the CON and EXE groups. Since the inputs were 2D image tensors here, we used 2D convolutional
6
[µ2
j ](i)
m
X
i=1
2j
1 σ
σi is estimated over the entire dataset as shown in Equation 7.
2[(ǫj − ǫ)](i)
=
filters in the Base CNN (see Tables S4, S5 and S6 from supplementary material for more details).
Following training, ccCAM was applied to obtain CAMs for each subject at each time instant during
the task execution.
4 Results and Discussion
The results presented here illustrate the differences between the Baseline and 90 min post exercise/rest
datasets. The network architecture details for each type of data (TF and Topographical) map are
presented in the supplementary material, along with details regarding the chosen hyperparameters.
4.1 Time-Frequency maps
We observed that the features extracted by Base CNN, without any subject prediction regularizer, could
be used to identify the subject corresponding to any given data tensor. As the subject discriminator
regularization was given more weight by increasing λ, the Base CNN learned to extract features that
were agnostic to the originating subject. However for very high λ values, the extracted features could
not be used to discriminate the EXE and CON groups, suggesting that the Base CNN was unable to
learn discriminative features. The loss values obtained post-training for four different values of λ are
shown in Table 1. The choice of an optimal value for λ depends on two factors – group prediction
accuracy and subject prediction accuracy. To identify subject-invariant features, we aimed to obtain
an optimal value of λ that achieved good group prediction accuracy but poor subject prediction
accuracy. Consequently, this required a good tradeoff between the two prediction accuracies.
Table 1: Variation of Loss values with λ after training network on TF maps.
Group prediction loss
Subject prediction loss
(NLL)
≈ 0
≈ 1.5
≈ 2.6
≈ 3.2
KL divergence loss from
Uniform distribtion
≈ 0.3
≈ 0.07
≈ 0.004
≈ 0.0002
λ
0
10
13
15
(NLL)
≈ 0
≈ 0.1
≈ 0.4
≈ 0.68
According to this procedure, the model corresponding to λ = 13 was used for ccCAM generation.
The average loss over a batch for subject prediction was around 2.6, which roughly predicted the
13. The group prediction accuracy was 99.984% (99.969% for
correct subject with a confidence of 1
CON and 100% for EXE). Hence the extracted features achieved excellent group prediction, while
all subjects in the group were predicted with roughly equal probability (CON and EXE consisted of
13 and 12 subjects respectively). The ccCAM obtained is shown in Figure 3.
Figure 3: Time-averaged TF map ccCAM averaged over electrodes & subjects showing "discrimina-
tive" frequencies.
As one of the main goals of this study was to identify the frequency bands that contained significant
information, we calculated the ccCAMs for all timepoints and then group-averaged (averaging across
all timepoints and subjects in a group) the maps to get two 2D maps – for the CON and EXE groups.
We plotted the average activation within each frequency band in each of these 2D maps to obtain
the plots in Figure 3. The bold lines denote the group-mean and the shaded regions span 1 standard
7
error over all subjects in the group. The two plots are significantly different within the band 23–33
Hz. This band lies within the wider beta-band and agrees with findings in [8] where beta-band
desynchronization was found to be significantly modulated by exercise. It is important to note that
the ccCAM highlights the differences between the 90min and baseline EEG recordings. The negative
values in a frequency band indicate that the ERD was smaller after than before the exercise. This
also agrees with findings in [8] and implies that decreased neural activity was required to perform the
hand-grip task after exercise. The p-value calculated from the ccCAM outputs within this frequency
band was equal to 0.021, while the corresponding p-value from the original time-frequency data
tensor was equal to 0.0134. This suggests that had the band of interest in the previous study [8] been
chosen to be 23–33 Hz, instead of the wider beta-band (15–29 Hz), similar, statistically significant
inferences would have been drawn.
Topographical maps were created to understand the distribution of the activity within the 23–33
Hz frequency band over the cortex. After training a network to classify into the CON and EXE
groups from topographical maps, a classification accuracy of 98.70% (98.94% for CON and 98.43%
for EXE) was obtained for λ = 5. Generating ccCAMs for the topographical maps revealed the
propagation of the discriminative activity across the cortex. A video showing this traveling property
of this activity is included in the supplementary material. Some snapshots from the video are shown
in Figure 4.
Figure 4: Topographical map ccCAM averaged over subjects in a group showing regions with
difference in activity before and 90min after rest/exercise. Top row shows CON group. Bottom row
shows EXE group for a brief time window, where a streak of activity can be seen to move across time
from the parietal to motor cortex.
To the best of our knowledge, this traveling pattern of activity across the cortex while performing an
isometric handgrip has not been demonstrated before. These oscillations could allow us to visualize
the neural mechanisms involved in maintaining a constant grip-force output. Further investigation
into the correlation of these activities with the observed error signal while performing the task is
required to understand these mechanisms more precisely. As expected, differences in the ccCAM
of EXE group before and after exercise were higher in magnitude as compared to those in the CON
group (see Figure S4 in supplementary material), thus indicating the modulatory effects of an acute
bout of high-intensity exercise.
5 Conclusion
This work introduces a deep learning architecture for the analysis of EEG data and shows promising
results in terms of discriminating the participants that underwent an acute bout of high-intensity
exercise/rest in close temporal proximity to performing a motor learning task. Importantly, the
proposed novel method enabled us to visualize the features learnt by deep networks such as CNNs,
which may in turn yield better interpretation of their classification basis. The results are in general
agreement with those reported in a previous study using more standard statistical analysis for a-priori
selected features on the same dataset [8], with our analysis revealing a narrower, more-specific
frequency band associated with exercise-induced changes. In addition, our method revealed, for the
first time, the traveling pattern of cortical activity while subjects were performing isometric handgrips.
8
Therefore, our approach demonstrates scope of identifying discriminative features in a completely
data-driven manner. The proposed method is not restricted to the EEG modality and dataset described
here. Hence, it paves the way for applying equivalent deep learning methods to datasets obtained
from neuroimaging studies of differing scales and varying modalities (eg. magnetoencephalography –
MEG). This, in turn, yields great potential to accelerate research oriented towards identification of
neurophysiological changes associated with various neurological disorders and ultimately lead to
design of optimized and individualized intervention strategies.
References
[1] P. Agrawal, J. Carreira, and J. Malik. Learning to see by moving. In Computer Vision (ICCV),
2015 IEEE International Conference on, pages 37–45. IEEE, 2015.
[2] P. Bashivan, I. Rish, M. Yeasin, and N. Codella. Learning representations from eeg with deep
recurrent-convolutional neural networks. arXiv preprint arXiv:1511.06448, 2015.
[3] N. Berchtold, G. Chinn, M. Chou, J. Kesslak, and C. Cotman. Exercise primes a molecu-
lar memory for brain-derived neurotrophic factor protein induction in the rat hippocampus.
Neuroscience, 133(3):853–861, 2005.
[4] T. W. Boonstra, A. Daffertshofer, M. Breakspear, and P. J. Beek. Multivariate time–frequency
analysis of electromagnetic brain activity during bimanual motor learning. Neuroimage,
36(2):370–377, 2007.
[5] T. Brashers-Krug, R. Shadmehr, and E. Bizzi. Consolidation in human motor memory. Nature,
382(6588):252, 1996.
[6] R. Collobert, K. Kavukcuoglu, and C. Farabet. Torch7: A matlab-like environment for machine
learning. In BigLearn, NIPS workshop, number EPFL-CONF-192376, 2011.
[7] N. E. Crone, D. L. Miglioretti, B. Gordon, J. M. Sieracki, M. T. Wilson, S. Uematsu, and R. P.
Lesser. Functional mapping of human sensorimotor cortex with electrocorticographic spectral
analysis. i. alpha and beta event-related desynchronization. Brain: a journal of neurology,
121(12):2271–2299, 1998.
[8] F. Dal Maso, B. Desormeau, M.-H. Boudrias, and M. Roig. Acute cardiovascular exercise
promotes functional changes in cortico-motor networks during the early stages of motor memory
consolidation. NeuroImage, 174:380–392, 2018.
[9] A. Delorme and S. Makeig. Eeglab: an open source toolbox for analysis of single-trial
eeg dynamics including independent component analysis. Journal of neuroscience methods,
134(1):9–21, 2004.
[10] M. O. Ernst and M. S. Banks. Humans integrate visual and haptic information in a statistically
optimal fashion. Nature, 415(6870):429, 2002.
[11] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand,
and V. Lempitsky. Domain-adversarial training of neural networks. The Journal of Machine
Learning Research, 17(1):2096–2030, 2016.
[12] A. Graves, A.-r. Mohamed, and G. Hinton. Speech recognition with deep recurrent neural net-
works. In Acoustics, speech and signal processing (icassp), 2013 ieee international conference
on, pages 6645–6649. IEEE, 2013.
[13] M. E. Hopkins, F. C. Davis, M. R. VanTieghem, P. J. Whalen, and D. J. Bucci. Differential
effects of acute and regular physical exercise on cognition and affect. Neuroscience, 215:59–68,
2012.
[14] S. Houweling, A. Daffertshofer, B. W. van Dijk, and P. J. Beek. Neural changes induced by
learning a challenging perceptual-motor task. Neuroimage, 41(4):1395–1407, 2008.
[15] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video
classification with convolutional neural networks. In Proceedings of the IEEE conference on
Computer Vision and Pattern Recognition, pages 1725–1732, 2014.
9
[16] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
[17] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional
neural networks. In Advances in neural information processing systems, pages 1097–1105,
2012.
[18] C. Neuper and G. Pfurtscheller. Event-related dynamics of cortical rhythms: frequency-specific
features and functional correlates. International journal of psychophysiology, 43(1):41–58,
2001.
[19] F. Ostadan, C. Centeno, J.-F. Daloze, M. Frenn, J. Lundbye-Jensen, and M. Roig. Changes in
corticospinal excitability during consolidation predict acute exercise-induced off-line gains in
procedural memory. Neurobiology of learning and memory, 136:196–203, 2016.
[20] G. Pfurtscheller, B. Graimann, J. E. Huggins, S. P. Levine, and L. A. Schuh. Spatiotemporal
patterns of beta desynchronization and gamma synchronization in corticographic data during
self-paced movement. Clinical neurophysiology, 114(7):1226–1236, 2003.
[21] S. M. Plis, D. R. Hjelm, R. Salakhutdinov, E. A. Allen, H. J. Bockholt, J. D. Long, H. J. Johnson,
J. S. Paulsen, J. A. Turner, and V. D. Calhoun. Deep learning for neuroimaging: a validation
study. Frontiers in neuroscience, 8:229, 2014.
[22] B. Pollok, D. Latz, V. Krause, M. Butz, and A. Schnitzler. Changes of motor-cortical oscillations
associated with motor learning. Neuroscience, 275:47–53, 2014.
[23] E. M. Robertson, A. Pascual-Leone, and R. C. Miall. Current concepts in procedural consolida-
tion. Nature Reviews Neuroscience, 5(7):576, 2004.
[24] M. Roig, S. Nordbrandt, S. S. Geertsen, and J. B. Nielsen. The effects of cardiovascular exercise
on human memory: a review with meta-analysis. Neuroscience & Biobehavioral Reviews,
37(8):1645–1666, 2013.
[25] R. Salmelin, M. Hámáaláinen, M. Kajola, and R. Hari. Functional segregation of movement-
related rhythmic activity in the human brain. Neuroimage, 2(4):237–243, 1995.
[26] R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger,
M. Tangermann, F. Hutter, W. Burgard, and T. Ball. Deep learning with convolutional neural
networks for eeg decoding and visualization. Human brain mapping, 38(11):5391–5420, 2017.
[27] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra. Grad-cam:
Visual explanations from deep networks via gradient-based localization. See https://arxiv.
org/abs/1610.02391 v3, 7(8), 2016.
[28] K. Skriver, M. Roig, J. Lundbye-Jensen, J. Pingel, J. W. Helge, B. Kiens, and J. B. Nielsen.
Acute exercise improves motor memory: exploring potential biomarkers. Neurobiology of
learning and memory, 116:46–58, 2014.
[29] F. Tadel, S. Baillet, J. C. Mosher, D. Pantazis, and R. M. Leahy. Brainstorm: a user-friendly
application for meg/eeg analysis. Computational intelligence and neuroscience, 2011:8, 2011.
[30] P. Thodoroff, J. Pineau, and A. Lim. Learning robust features using deep learning for automatic
seizure detection. In Machine Learning for Healthcare Conference, pages 178–190, 2016.
[31] E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko. Simultaneous deep transfer across domains and
tasks. In Computer Vision (ICCV), 2015 IEEE International Conference on, pages 4068–4076.
IEEE, 2015.
[32] X. Zhang and Y. LeCun. Text understanding from scratch. arXiv preprint arXiv:1502.01710,
2015.
[33] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for
discriminative localization. In Computer Vision and Pattern Recognition (CVPR), 2016 IEEE
Conference on, pages 2921–2929. IEEE, 2016.
10
[34] P. Zhuang, C. Toro, J. Grafman, P. Manganotti, L. Leocani, and M. Hallett. Event-related
desynchronization (erd) in the alpha frequency during development of implicit and explicit
learning. Electroencephalography and clinical neurophysiology, 102(4):374–381, 1997.
11
Deep Semantic Architecture with discriminative
feature visualization for neuroimage analysis
Arna Ghosh
Integrated Program in Neuroscience
McGill University
Montréal, QC H3A 0G4. Canada
[email protected]
Fabien Dal Maso
École de kinésiologie et des
sciences de l'activité physique
Université de Montréal
Montréal, QC H3T 1J4. Canada
[email protected]
Marc Roig
School of Physical and
Occupational Therapy
McGill University
Montréal, QC H3A 0G4. Canada.
[email protected]
Georgios D Mitsis
Department of Bioengineering
McGill University
Montréal, QC H3A 0G4. Canada.
[email protected]
Marie-Hélène Boudrias
School of Physical and
Occupational Therapy
McGill University
Montréal, QC H3A 0G4. Canada.
[email protected]
Supplementary Material
1 Network Architecture
Notation:- Conv denotes the 2D Spatial Convolutional layer. ReLU denotes the Rectified Linear Unit
Layer that adds non-linearity to the network. MaxPool denotes the 2D Spatial Max Pooling layer.
FullyConn denotes a Fully Connected layer, also known as the linear layer of the network.
1.1 TF maps
Layer
0
1
2
3
4
5
6
Type
Input
Conv
ReLU
Conv
ReLU
MaxPool
MaxPool
Maps and Neurons
1M × 64 × 55N
6M × 64 × 28N
6M × 64 × 28N
6M × 64 × 14N
16M × 64 × 14N
16M × 64 × 14N
16M × 64 × 7N
Filter Size
-
1×5
-
1×2
1×5
-
1×2
Table S1: Network architecture used for EEG feature extraction network (Base CNN). The output of
the network is a tensor of dimensions 16 × 64 × 7.
Preprint. Work in progress.
(a) Basic Architecture without adversary
(b) Modified Architecture with adversary to avoid subject discrimination
Figure S1: Deep Network Architecture. The initial choice of architecture (without any adversary)
gives good subject prediction accuracy from features extracted by the Base CNN. Therefore, a subject
discriminator of roughly the same model capacity as the Top NN is added. The subject discrimination
acts as a regularizer while training and avoids the Base CNN from learning subject specific features.
Type
Input
Flatten
Maps and Neurons
16M × 64 × 7N
7168N
Filter Size
Layer
0
1
2
3
4
5
Layer
0
1
2
3
4
5
Dropout (p=0.5)
FullyConn
ReLU
FullyConn
Dropout (p=0.5)
FullyConn
ReLU
FullyConn
-
8N
8N
2N
-
8N
8N
25N
2
-
-
-
1×1
-
1×1
-
-
-
1×1
-
1×1
Table S2: Network architecture used for group discrimination network (Top NN). The output of the
network is a vector of dimension 2, values corresponding to the probability that the data tuple belongs
to particular class.
Type
Input
Flatten
Maps and Neurons
16M × 64 × 7N
7168N
Filter Size
Table S3: Network architecture used for subject discrimination network (adversary). The output of
the network is a vector of dimension 25, values corresponding to the probability that the data tuple
belongs to particular subject.
1.2 Topographical maps
Layer
0
1
2
3
4
5
6
7
8
9
Type
Input
Conv
ReLU
Conv
ReLU
MaxPool
MaxPool
Conv
ReLU
MaxPool
Maps and Neurons
3M × 64 × 64N
16M × 32 × 32N
16M × 32 × 32N
16M × 16 × 16N
32M × 16 × 16N
32M × 16 × 16N
32M × 8 × 8N
64M × 8 × 8N
64M × 8 × 8N
64M × 4 × 4N
Filter Size
-
5×5
-
2×2
5×5
-
2×2
3×3
-
2×2
Table S4: Network architecture used for EEG feature extraction network (Base CNN). The output of
the network is a tensor of dimensions 64 × 4 × 4.
Type
Input
Flatten
Maps and Neurons
64M × 4 × 4N
1024N
Filter Size
Layer
0
1
2
3
4
5
Layer
0
1
2
3
4
5
Dropout (p=0.5)
FullyConn
ReLU
FullyConn
Dropout (p=0.5)
FullyConn
ReLU
FullyConn
-
8N
8N
2N
-
8N
8N
25N
-
-
-
1×1
-
1×1
-
-
-
1×1
-
1×1
Table S5: Network architecture used for group discrimination network (Top NN). The output of the
network is a vector of dimension 2, values corresponding to the probability that the data tuple belongs
to particular class.
Type
Input
Flatten
Maps and Neurons
64M × 4 × 4N
1024N
Filter Size
Table S6: Network architecture used for subject discrimination network (adversary). The output of
the network is a vector of dimension 25, values corresponding to the probability that the data tuple
belongs to particular subject.
3
2 Training curves
2.1 Time-Frequency Maps
Hyperparameter
Learning Rate
Learning Rate Decay
Weight Decay
Value
0.001
0.0001
0.001
Table S7: List of hyperparameters used for training the networks on TF maps.
(a) Group prediction loss (λ = 0)
(b) KL divergence (λ = 0)
(c) Subject prediction loss (λ = 0)
(d) Group prediction loss (λ = 15)
(e) KL divergence (λ = 15)
(f) Subject prediction loss (λ = 15)
(g) Group prediction loss (λ = 13)
(h) KL divergence (λ = 13)
(i) Subject prediction loss (λ = 13)
Figure S2: Time-Frequency Maps Training curves for three different weight values to the subject
predictor regularizer.
4
2.2 Topographical Maps
Hyperparameter
Learning Rate
Learning Rate Decay
Weight Decay
Value
0.001
0.001
0.03
Table S8: List of hyperparameters used for training the networks on Topographical maps.
(a) Group prediction loss (λ = 0)
(b) KL divergence (λ = 0)
(c) Subject prediction loss (λ = 0)
(d) Group prediction loss (λ = 10)
(e) KL divergence (λ = 10)
(f) Subject prediction loss (λ = 10)
(g) Group prediction loss (λ = 5)
(h) KL divergence (λ = 5)
(i) Subject prediction loss (λ = 5)
Figure S3: Topographical Maps Training curves for three different weight values to the subject
predictor regularizer.
5
3
ccCAM on Topographical Maps
Figure S4: ccCAM matrices generated for Topographical Maps of 23–33 Hz activity group-averaged
over time windows of 0.6 sec. The CAMs have higher values for EXE group as compared to the
CON group which indicates that the network learns to look at the differences in EEG activity while
performing the fixed-force handgrip task induced by acute bout of high-intensity exercise. The time
described here is calculated from 0.5 sec after the presentation of visual cue.
6
|
1509.07897 | 1 | 1509 | 2015-09-16T06:02:26 | Quantum Look at two Common Logics: the Logic of Primitive Thinking and the Logic of Everyday Human Reasoning | [
"q-bio.NC",
"cs.AI",
"math.LO"
] | Based on ideas of quantum theory of open systems and psychological dual system theory we propose two novel versions of Non-Boolean logic. The first version can be interpreted in our opinion as simplified description of primitive (mythological) thinking and the second one as the toy model of everyday human reasoning in which aside from logical deduction, heuristic elements and beliefs also play the considerable role. Several arguments in favor of the interpretations proposed are adduced and discussed in the paper as well. | q-bio.NC | q-bio |
Quantum Look at two Common Logics: the Logic of Primitive Thinking and the Logic
of Everyday Human Reasoning
B. Verkin Institute for Low Temperature Physics and Engineering of the National
Academy of Sciences of Ukraine 47, Lenin Ave., Kharkov 61103, Ukraine.
(Dated: October 24, 2018)
E. D. Vol∗
Based on ideas of quantum theory of open systems and psychological dual system theory we
propose two novel versions of Non-Boolean logic. The first version can be interpreted in our opinion
as simplified description of primitive (mythological) thinking and the second one as the toy model of
everyday human reasoning in which aside from logical deduction, heuristic elements and beliefs also
play the considerable role. Several arguments in favor of the interpretations proposed are adduced
and discussed in the paper as well.
PACS numbers: 05.40.-a
I.
INTRODUCTION
It is well known that the Boolean propositional logic
which is consistent mathematical presentation of the clas-
sical Aristotelian logic has applications of two kinds. On
the one hand it correctly takes into account deductive
elements of the everyday human reasoning and on the
other hand this logic can be considered as the relevant
framework of common scientific language both in exact
sciences and humanities. For example referring to the
classical physics it may be argued that the result of any
experiment realized in this area can be described by the
Boolean logic in a consistent manner. However in the sit-
uations when distinct features of the objects under study
strongly correlate between themselves and moreover may
be incompatible the application of standard Boolean logic
may lead to errors. Just such case has place for example
in quantum mechanics. Actually, as firstly G. Birkhoff
and J. von Neumann revealed [1] in quantum mechanics,
in view of existence of certain non-commuting observ-
ables relating to the same particle such as coordinate and
momentum or the different components of the spin, the
laws of the Boolean logic (in particular the distributive
law) have been failed and their generalization i.e. quan-
tum logic is desirable. The similar situation may arise
also in everyday life when various reasons, emotions and
beliefs governing the behavior of concrete person begin
to contradict each other.
In this connection it should
be noted that still in the early days of quantum theory
N. Bohr, when he formulated the Complementarity Prin-
ciple, (which just maintains the presence in physics cer-
tain supplementary properties of the microscopic objects)
underlined more than once essential analogy existing be-
tween atomic processes and such mental phenomena as
thoughts, sentiments and acts of decision making. It is
remarkable that modern cognitive psychology using the
data of a large number of experiments came to the similar
conclusions as well. Really according to the psychologi-
∗Electronic address: [email protected]
cal dual system theory all basic cognitive processes such
as attention, memory, learning and so on are connected
with certain dual systems and dual mechanisms. In the
present paper we will interested in only human reasoning
mechanisms where two primary dual systems of interest
can be specified. The first of them (we will term it below
as deductive reasoning system) is rational, sequential and
consistent but acts relatively slow and its resources are
limited by the capacity of human working memory. The
second one (we will term it as heuristic reasoning system
) is intuitive, rapid and automatic but its activity may be
biased to a large degree by emotions and old unconscious
ideas.
In addition a hidden interaction exists between
these two cognitive systems which as a rule is not aware
by the reasoning person. The main goal of present pa-
per is to demonstrate how by using some ideas, concepts
and technical tools of quantum theory of open systems
(QTOS) one can describe the dual nature of human rea-
soning in the framework of consistent logical theory that
generalizes and modifies the rules of ordinary Boolean
logic.
The paper is organized as follows.
In Section.2 fol-
lowing the previous author paper [2] we briefly outline
main ideas of the approach proposed using the simple
and instructive example of probabilistic Boolean logic
(PBL) that describes the situations when in the absence
of complete information about the surrounding objects
and events all human judgments acquire inevitably prob-
abilistic nature. To take into account this crucial point it
is convenient, instead of usual Boolean functions of dis-
crete variables which take only two values unit and zero,
to associate with every plausible proposition (PP) certain
diagonal representative density matrix (RDM) of some
auxiliary two state quantum system, whose elements de-
fine the plausibility of corresponding proposition. The
essential and novel element of the approach proposed
is the universal and constructive algorithm (based on
QTOS) which allows one to define all logical connec-
tives between plausible propositions using the powerful
and effective method of positive valued (PV) quantum
operations. It is worth noting that this method will be
used continually throughout the paper for the construc-
tion of various kinds of logics. Also we study here an-
other important problem, namely : how the possible log-
ical correlations between various PP may be took into
account in the approach proposed. Further in the Sec-
tion.3 in order to describe the dual nature of human rea-
soning we have extended our approach to the arbitrary
2 × 2 non-diagonal density matrices that will represent
in this case the generalized propositions (GP). We as-
sume that non-diagonal elements of the RDMs of such
propositions relate to heuristic ( believable) elements of
human reasoning and therefore by means of them one
can define the believability of the corresponding propo-
sitions. The main problem of the approach proposed is:
whether one may specify the set of logical connectives
for similar generalized propositions and if the answer is
yes in what way. To answer this question we will use
again the approved method of PV quantum operations.
It turns out however that in general case (beyond the
usual PBL) it is necessary to impose certain restrictions
either on the form of admissible quantum operations or
on the form of input GP. In this connection we proposed
two possible alternatives to construct the consistent Non-
Boolean (NB) logic operating with such GP. The NB
logic of the first kind aside from negation includes only
the single pair of two place connectives which are simi-
lar to the pair of relations: equivalence-non-equivalence
in the standard Boolean logic. At first sight such logic
looks much poorer than standard Boolean logic and we
will term it as 'atomic' logic or prime logic. It turns out
however that, starting from GP and using only these two
connectives, one may construct any prescribed diagonal
PP and after that to handle with these PP according
to the known rules of PBL .Thus PBL appears now as
secondary or 'molecular' logic with respect to prime or
'atomic' NB logicof the first kind.
We argue that described version of Non-Boolean logic
of the first kind (prime logic) can be considered as sim-
plified version of the logic of primitive (or mythological)
thinking. In favor of such interpretation we adduce sev-
eral cogent in our opinion arguments. Finally in the Sec-
tion.4. we study the NB logic of the second kind that
assumes the existence of special primordial correlations
between deductive and heuristic components of general-
ized propositions .In this case using the formalism of PV
quantum operations one can specify all the same logical
connectives between GP including implication as in usual
PBL .It turns out that a number of known from everyday
life curious psychological phenomena such as belief-bias
effect may be explained in the framework of this logic by
natural way. Now let us go to the concrete presentation
of announced results.
II. PRELIMINARIES
In this section we briefly remind some results of our
previous paper [2] in which based on ideas of quantum
theory of open systems the formalism of probabilistic
2
Boolean logic (PBL) was developed. To this end in view,
instead of standard Boolean variables with two values 1
and 0 that represent true and false propositions respec-
tively one has to consider so called plausible propositions
(PP), whose truth or falsity are determined only with
certain probability. We propose to represent such propo-
sitions by means of 2 × 2 diagonal matrices with positive
elements the sum of which is equal to unit. Every such
representative matrix of PP may be associated with the
density matrix of certain two state quantum system .
Therefore, if it does not lead to confusion, we will often
identify the propositions in PBL with their representative
density matrices (RDMs).
In explicit form the RDM of arbitrary PP a looks as
0 1 − pa(cid:19) where pa is a probability for propo-
ρ (a) =(cid:18)pa
0
sition a to be true(index a of the proposition we will
often omit later ).
It is convenient also to define the
plausibility of the proposition a as P (a) = 2pa − 1. Ob-
viously that plausibility takes its values in the interval
[−1, 1]. It turns out the approach proposed allows one to
express all logical connectives between PP by universal
way as certain positive valued (PV) quantum operations
under their RDMs. Referring the reader for the details of
this approach to [2] let us merely demonstrate here how
this approach works using several concrete examples. So,
if someone wants to obtain the RDM of the negation of
PP a =(cid:18)p
with the following 2×2 matrix Gnot =(cid:18)0 1
0 1 − p(cid:19) that is (not a) she(or he) must make
1 0(cid:19) and GT is,
the next quantum operation: ρ (nota) = Gnotρ (a) GT
0
not
as usually, the matrix transposed to matrix G. It turns
out that any two place connective in PBL (aRb) may be
expressed in similar manner as well, namely:
(aRb) = GR [ρ (a) ⊗ ρ (b)] GT
R
(1)
where ρ (a) ⊗ ρ (b) is standard tensor product of RDMs
of propositions a and b and GR is a 2 × 4 matrix (we
will term it below as admissible matrix for connective
R) which has two defining properties:1) every element
of GR is equal 0 or 1 and 2) in each column of matrix
GR the only element is equal to one and all the rest
are equal to zero. Let us now present (without proof)
the concrete form of matrices GR for basic two place
connectives, namely :(a and b), (a or b),(a =⇒ b) . As
shown in [2] they have the next form:
0 1 1 1(cid:19) ,
G(a and b) = (cid:18)1 0 0 0
0 0 0 1(cid:19) ,
G(a or b) = (cid:18)1 1 1 0
0 1 0 0(cid:19) .
G(a=⇒b) = (cid:18)1 0 1 1
(2)
(3)
(4)
For example for the implication (a =⇒ b) one can easily
3
(5)
1) ρ1 (a, b) =
1
2 0 0 0
0 0 0 0
0 0 0 0
0 0 0 1
2
0 0 0 0
0 1
2 0 0
0 0 1
2 0
0 0 0 0
,
and
2) ρ2 (a, b) =
3) ρ3 =
1
4 0 0 0
0 1
4 0 0
0 0 1
4 0
0 0 0 1
4
.
writes down the RDM as:
0
ρ (a =⇒ b) =(cid:18)1 − p + pq
p (1 − q)(cid:19) .
In Eq. (5) we mean the next notation: a = (cid:18)p
and b = (cid:18)q
0 1 − p(cid:19),
0 1 − q(cid:19). Note that this relation coincides
0
0
0
with standard Boolean expression for the implication in
the case when probabilities p and q may take only two
values 0 and 1. Let us make now the generalization of
the approach proposed on the important case when ini-
tial PPs a and b are not independent propositions or,
in other words, they possess some logical correlations.
In such situation the best way to take these correla-
tions into account is to avail of the direct analogy be-
tween our approach and the theory of composite cor-
related quantum systems. To realize this analogy one
must take as a starting point, instead of tensor produc-
tion of two initial PPs a and b the more general positive
diagonal matrix of the form: ρ (a, b) =
4Pi=1
with additional normalization condition:
p1 0
0
0 p2 0
0
0
0
0
0 p3 0
0
0 p4
pi = 1. Fur-
thermore we naturally assume that this joint matrix of
correlated propositions ρ (a, b) is connected with matri-
ces of partial propositions that is RDM of a and RDM
of b by the standard relations: a = (cid:18)p1 + p2
b = (cid:18)p1 + p3
p3 + p4(cid:19),
p2 + p4(cid:19). Now let us suppose again that
0
0
0
0
all logical connectives for correlated PP may be obtained
by the approved method that is by means of the PV quan-
tum operations. Here we restrict ourselves only with the
case of implication between correlated propositions a and
b. Let us introduce the following notation p = p1 + p2 ,
q = p1 + p3 and C = p1p4 − p2p3
0
pq + C
p (1 − q) − C
It is easy to see that ,using this notation, the ma-
trix ρ (a, b) can be rewritten in the form: ρ (a, b) =
0 1 0 0(cid:19) to the ρ (a, b) one may obtain the required
(cid:18)1 0 1 1
(1 − p) (1 − q) + C
Now by applying the quantum operation G(a=⇒b) =
q (1 − p) − C
0
0
0
0
0
0
0
0
0
0
0
.
result for RDM of implication, namely:
ρ (a, b) =(cid:18)1 − p + pq + C
0
p − pq − C(cid:19) .
0
(6)
It is easy to see that in all cases the RDMs for
partial propositions a and b coincide and represent
the same ambiguous proposition: ρi (a) = ρi (b) =
2(cid:19) (i = 1, 2, 3). However the expressions for impli-
(cid:18) 1
2 0
0 1
cation in these situations differ significantly. Really us-
ing the general rule Eq. (1) stated above and its spe-
cial case in the form of Eq. (6) one may easily obtain:
(a =⇒ b)1 = (cid:18)1 0
(a =⇒ b)2 =(cid:18) 1
and (a =⇒ b)3 = (cid:18) 3
0 0(cid:19), that is the true proposition, while
2(cid:19) is the same ambiguous proposition
4(cid:19) .Note that only the last case
4 0
0 1
2 0
0 1
with composite matrix ρ3 corresponds to the case of the
independent partial propositions. This simple but in-
structive example clearly demonstrates the considerable
influence of logical correlations on the results of inference
process. In this point we have finished our brief review
of PBL. It should be noted in conclusion that in the case
of PBL considered in this Section were taken into ac-
count only propositions that may be specified by the sin-
gle defining characteristic , namely plausibility. Now we
pass to the study of Non-Boolean (NB) logics that would
operate with the propositions of more general kind .These
propositions can be determined already by two defining
features which reflect dual nature of the objects (or sub-
jects) connecting with corresponding propositions . We
believe that similar logics may be more appropriate for
the description (at least in simplified form) of essential
peculiarities of such common logics as the logic of prim-
itive (mythological) thinking and the logic of everyday
human reasoning.
III. NON-BOOLEAN LOGIC OF THE FIRST
KIND AS THE SIMPLIFIED DESCRIPTION OF
PRIMITIVE (MYTHOLOGICAL) THINKING
Note that the variable C is just a measure of logical corre-
lations between propositions a and b. In this connection
we will termed it as the context variable.
Let us consider now the simple example with three con-
crete initial RDMs of composite correlated propositions,
namely :
In the beginning of this Section we would like to dis-
cuss the following primary question namely: to what de-
gree the approach stated above may be extended on the
propositions of more general form .To this end in view
let us consider instead of exclusively diagonal RDM rep-
resenting a certain PP the more general class of 2 × 2
non-diagonal positive valued matrices with unit trace and
examine them as certain RDMs of generalized propo-
sitions (GP). Let us suppose that an arbitrary propo-
sition of this class a can be written in the next form
z ∗ 1 − p(cid:19) ,where z is some complex number.
ρ (a) = (cid:18) p
p(cid:19).
is (not a) by natural way as ρ (not a) = (cid:18)1 − p z ∗
Let us define also the negation of such proposition, that
z
z
Furthermore let us require that the RDMs of the oppo-
site propositions a and not a commute with each other.
The sense of this restriction (having in mind the obvious
analogy with quantum theory) is quite clear and does not
need in additional comments. It is easy to see that this
restriction may be reduced to the simple relation z = iα
or, in other words, non-diagonal elements of the RDM in
the approach proposed have to be pure imaginary. Thus
the class of generalized RDMs which we will consider in
the present Section contains all 2 × 2 positive valued ma-
trices of the form:
ρ (a) =(cid:18) p
−iα 1 − p(cid:19) .
iα
4
we have to define all possible connectives related to them.
It turns out however that the direct extension of the ad-
missible quantum operation method in order to obtain
required two place logical operations for GPs is impossi-
ble ( because the automatic satisfaction of normalization
condition, that was guaranteed for PPs in PBL, is not
always holds now).However there are at least two possi-
ble ways to get round this formal obstacle. In this Sec-
tion we consider the first of them. To this end we retain
as admissible operations only two concrete 2 × 4 matri-
ces,namely: G∆ = (cid:18)0 1 1 0
0 1 1 0(cid:19).
1 0 0 1(cid:19) and G∆ = (cid:18)1 0 0 1
It is easy to see directly that these operations transform
the tensor product of two propositions a ⊗ b to the RDM
that belongs to the required class of GP . Really, let
a = (cid:18) p
−iα 1 − p(cid:19) and b = (cid:18) q
−iβ 1 − q(cid:19) are two GP,
iα
iβ
then it is easy to obtain that:
G∆ (a ⊗ b) GT
∆ =(cid:18) A
−iγ 1 − A(cid:19) ≡ (a∆b) .
iγ
(9)
(7)
where A = p + q − 2pq + 2αβ and γ = α (1 − 2q) +
β (1 − 2p). In the same way one can find that
As to formal interpretation of these GPs we assume as
before that their diagonal elements determine the plausi-
bility of corresponding GP according to the above men-
tioned rule: P (a) = 2p − 1 whereas their non-diagonal
elements define its believability, that is the readiness of a
subject to accept the corresponding proposition as valid
by virtue of some irrational and unconscious reasons or,
in other words. on the strength of certain beliefs. Thus
the similar GPs combine together the objective and sub-
jective reasons which stimulate a person to accept the
concrete proposition as valid.
It is naturally to define
the believability of concrete GP a with RDM Eq. (9) as
B (a) = −2α. The justification of such definition became
clear if we write down this GP by standard way using the
Bloch sphere representation, namely:
G∆ (a ⊗ b) GT
∆ = (cid:18)1 − A −iγ
A (cid:19) =
iγ
(10)
= not(a∆b) ≡ (a∆b).
In the case when x-component of the Bloch vector
is identically equal to zero it is convenient to intro-
duce the complex Bloch vector as follows : P = Pz −
iPy.Comparing this definition with the above mentioned
Bloch sphere representation of density matrix one can
state the simple relation connecting the complex vector
R of the proposition (a∆b)with corresponding vectors P
and Q of the GP a and b respectively, namely:
R = −P Q.
(11)
ρ (a) =(cid:18) 1+Pz
Px+iPy
2
2
Px
−iPy
2
2 (cid:19) .
1−Pz
(8)
Note that for GPs which we consider here, the compo-
−→
nent Px of the Bloch vector
P is always equal to zero.
Now if one compares the representation Eq. (8) with ex-
pression Eq. (7) she(he) can see that the plausibility and
believability of the same GP being expressed in terms of
the Bloch vector components look similarly. Furthemore
it should be noted that relation P 2
y 6 1 is obviously
holds for any GP and this fact distinctly reflects the com-
petition existing between two features of the same propo-
sition. It should be noted also that our assumption is in
a full agreement with modern psychological dual system
theory [3] that asserts the presence in brain two com-
plementary cognitive systems: (deductive and heuristic)
, which both are responsible for human reasoning, al-
though often compete with each other. Now in order to
formulate the consistent NB logic with GPs stated above
z + P 2
So, we found that the NB logic of the first kind aside
from negation contains in addition only the single pair of
two place connectives, namely ∆ and ∆. At first sight
such logic is much poorer that standard Boolean logic
and therefore one may call it as 'atomic logic' or prime
logic. Nevertheless it should be noted that, having in
hands only these two connectives and starting from the
set of GPs, one can easily obtain any PP belonging to
PBL with prescribed diagonal elements. After making
that, one may operate with them already according to
the rules of standard Boolean logic. Let us demonstrate
the validity of this statement .To this end let us take two
opposite GP of the special form: a = (cid:18) 1
(not a) =(cid:18) 1
2(cid:19) and
2 (cid:19). If one applies the operation ∆ to
2 −iα
iα
iα
−iα 1
2
1
them the obtained result looks as:
2 − 2α2
(a∆ not a) =(cid:18) 1
0
0
2 + 2α2(cid:19) .
1
(12)
2
Thus we see that any proposition a of PBL with pa 6 1
2
may be expressed by similar way. As for propositions of
PBL with pa ≥ 1
2 they can be expressed by means of op-
eration ∆ in the same manner. Thus the PBL plays now
the role of secondary or 'molecular' logic with respect
to prime or 'atomic' NB logic of the first kind. Note
in addition that the GP of the form : a = (cid:18) 1
iα
−iα 1
2(cid:19)
whose plausibility is equal to zero can be considered as
the simplest example of so-called intuitive judgments. In
this place we refer the reader to the inspiring book [4] in
which distinction between discoursive and intuitive judg-
ments have been deeply analyzed. Now let us pass to
the most intriguing question, relating to the NB logic of
the first kind stated above, namely: whether the logic
proposed above has any actual applications in everyday
life and (or) in science? We believe that the answer is
yes and are going to argue that logic of this kind could
be appropriate for example as simplified description of
primitive ( mythological) thinking. The author is aware
that without being an expert in the area of anthropology
or mythology he is unable to disclose this topic in proper
depth and moreover to prove the above assertion exactly.
Nevertheless let me adduce several arguments in favor of
the hypothesis proposed above. In the beginning let us-
remind some basic facts relating to the primitive thinking
(PT), or the thinking of savages ( all these and the other
concrete facts the reader can derive in detail for exam-
ple from classical book by C. Levi-Strauss) [5].The major
fact on which we base is that there are a set of elemen-
tary and at the same time fundamental units of PT, so-
called binary oppositions, such as : top-bottom, right-
left, birth-death, male-female and so on. Another impor-
tant regularity determines the basic principle governing
the primitive thinking, that is, its goals and mental tools
which are used to achieve these goals . It turns out that
PT while it operates with binary oppositions does not
seek to avoid any contradictions, that is typical for the
"normal" logic, but rather tends to find and then to re-
construct all kinds of possible intermediate links existing
between two antagonistic terms in concrete binary oppo-
sition. To this end in view there are two special person-
ages in myth, namely: the Culture Hero and the Trixter,
(for example in Greek mythology Hermes plays the role of
such Trixter) which realize this important function. Fur-
thermore certain totem animals for example the Raven
and the Coyote in Indian myths sometimes play the part
of Culture hero as well. The basic tool using in PT that
allows one to achieve the desired goal is so-called 'medi-
ation' -- the special mental or workable operation which
is able to reduce the distinction between two antagonis-
tic terms in binary oppositions. For example the Raven
and the Coyote as the animals- scavengers symbolically
realize the mediation function between the world of the
living beings and the world of the dead . Now let us
demonstrate how in the NB logic of the first kind de-
scribed above its two basic operations may be considered
as certain analogues of the mediation operation in myth.
5
To this end, in order to define the degree of difference be-
tween two distinct GPs, we will use well-known definition
of the distance between two mixed quantum states ρ1 and
ρ2, namely D (ρ1, ρ2) = 1
2 Sp ρ1 − ρ2 [6]. It is easy to
see that in the case of two dimensional Hilbert space
and in the Bloch sphere representation the above expres-
sion takes the simple form: D (ρ1, ρ2) = 1
where
the states.
−→
P 2 are corresponding Bloch vectors of
P 2(cid:12)(cid:12)(cid:12)
−→
P 1 and
2(cid:12)(cid:12)(cid:12)
−→
P 1 −
−→
It is obvious that maximum distance is realized be-
tween opposite propositions or , (using the PT language)
between antagonistic members of binary opposition. Now
if one applies the basic two place operation ∆ to the pair
of opposite propositions a with complex vector P and
(nota) with complex vector −P she(he) obtains the new
proposition with the complex vector R = P 2. It is easy
to see that the distance between the opposite proposi-
tions with complex vectors P 2 and −P 2 is less that the
distance between original pair of opposite propositions.
This fact confirms that logical connective ∆ possesses the
characteristic property of operation 'mediation' in PT.
Now we are going to adduce another forcible argument
in favor of the doubtless connection between NB logic
of the first kind considered above and the logic of PT.
Let us remind that according to the known ethnologist
Levy-Bruhl [7] (who was the first European researcher of
primitive and magical thinking PT is governed to a great
extent by the so-called law of participation that claims
the universal links existing between various things and
events in the world.In particular just this law allows the
shaman, who fell into trans, to perform a wide variety of
magical acts and transformations with surrounding ob-
jects or people using their mutual contiguity and (or)
similarity.
It is interesting to note that this feature of
primitive thinking afterwards was adopted by art and
literature (especially poetry) in the form of extensive use
of such specific tropes as metaphor and metonymy. In
the NB logic of the first kind proposed above this fea-
ture of PT may be simulated by additional logic opera-
tions,namely by logic rotations, that have no analogues
in the standard Boolean logic. Really, it is possible to
define one-parameter continuous group of logic rotations
(in the plane Pz − Py) according to the usual mathemati-
cal rule, namely: let GP a is represented by the complex
vector P then GPea obtained by rotation of a at an angle
Φ has complex vector eP with components fPy and fPz so
thatfPy = Py cos Φ+Pz sin Φ andfPz = Pz cos Φ−Py sin Φ
.Clearly that in this groupof logical transformations the
negation of the proposition a coincides with its rotation
at angle π and, furthermore, if one rotates GP a at angle
Φ1 and the other GP b at angle Φ2 then the GP (a∆b)
would be rotated at angle (Φ1 +Φ2).Thus this continuous
group of logical rotations let one to transform any propo-
sition into the arbitrary other (with the same modulus
certainly). In this point we are obliged to finish our brief
discussion of the relationship existing between primitive
thinking and the NB logic of the first kind although it
is clear that this complex and intriguing topic deserves
much more detail study.
IV. NON-BOOLEAN LOGIC OF THE SECOND
KIND AS THE TOY MODEL OF EVERYDAY
HUMAN REASONING
As we see in the previous Section in the case of Non-
Boolean logic of the first kind it is possible to define aside
from negation only the single pair of two- place logical op-
erations ∆ and ∆. The application of the rest admissible
quantum operations to the tensor product of two input
GP results in the disturbance of normalization condition
for output RDM. This obstacle may be formally removed
if one assume the existence of certain primordial correla-
tions between plausible and believable components of the
GPs. Let us explain more detail what we have in mind.
Suppose that instead of tensor product of the two input
GP a⊗ b (that means their logical independence) we take
as input more general composite proposition ρ (a, b) , the
RDM of which has the next form:
ρ (a, b) =
p1
−iα p2
0
−iβ
iα iβ
0
p3
0
iγ
iδ
0 −iγ −iδ p4
.
(13)
Using the obvious analogy with theory of compos-
ite quantum systems one can determine now the
partial RDM of propositions a and b as a =
−i (β + γ) p3 + p4 (cid:19) and b = (cid:18) p1 + p3
(cid:18) p1 + p2
−i (α + δ) p2 + p4(cid:19)
i (β + γ)
i (α + δ)
respectively. It is clear that input composite proposition
ρ (a, b) Eq.
(13) corresponds to the case of two corre-
lated partial GP a and b and the degree and nature of
their correlation is uniquely determined by elements of
composite matrix ρ (a, b). It should be emphasized that
construction of the input composite proposition ρ (a, b)
in the form Eq.
(13) has the essential advantage over
simple tensor product from formal point of view since
it allows one to define all 16 two place logical connec-
tives consistently by approved method of PV quantum
operations. Significantly that all admissible matrices GR
in this NB logic of the second kind have the same form
as in the case of PBL (see Eq. (1) ) and corresponding
two place connectives can be obtained according to the
similar rule:
(aRb) = GRρ (a, b) GT
R.
(14)
Let us present here (without details) the expressions for
the two place connectives (and), (or) in the NB logic of
the second kind, obtained by means of Eq. (14):
p1
(a and b) = (cid:18)
−i (α + β) p2 + p3 + p4(cid:19) ,
(a or b) = (cid:18)p1 + p2 + p3 i (γ + δ)
p4 (cid:19)
−i (γ + δ)
i (α + β)
(15)
(16)
6
Now as in the previous section, the important question
arises about the possibility of concrete realizations of the
NB logic of the second kind.
In this connection we are going to bring several argu-
ments in favor of hypothesis that above version of NB
logic can be considered as simplified model of everyday
human reasoning. To this end in view let us remind some
basic facts relating to the psychological dual system the-
ory of human cognition(see for example [3] for the com-
prehensive exposition). As numerous psychological ex-
periments and observations had unambiguously demon-
strated there are two distinct types of mental processes
that are responsible for human reasoning: 1) cognitive
processes of the first type that are rapid, automatic and
intuitive and 2) the processes of the second type that are
slow, sequent and rational. Cognitive system which is
responsible for the processes of the first type is termed
as heuristic (or intuitive) while the second one is termed
as deductive (or analytical). Clearly these two systems
may sometimes come into conflict and compete with each
other. Therefore human reasoning is often biased and
subjected to various fallacies. In particular the reason-
ing person often overestimates the cogence of the argu-
ments that lead to the believable and expected conclu-
sions and underestimates those arguments which lead to
the conclusions contradictory to her (his) beliefs. This
widespread psychological phenomenon is known as belief-
bias effect. Now the natural question arises: is it possible
to describe this human hybrid intuitive-deductive think-
ing in the framework of any consistent logical theory?
In this Section we propose brief outline of possible the-
ory based on the NB logic of the second kind that was
exposed above. Note that we present here only the lit-
tle fragment of this theory which let one to understand
belief-bias effect in the framework of closed logical point
of view. To this end let us examine the expression for
implication connecting two correlated GP. According to
the general rule Eq. (14) the RDM of the implication
can be represented in the form:
(a =⇒ b) = Gimpρ (a, b) GT
imp =(cid:18)p1 + p3 + p4 i (α − γ)
p2 (cid:19) .
0 1 0 0(cid:19) .
where the matrix Gimp =(cid:18)1 0 1 1
−i (α − γ)
(17)
We see that the believability of the GP (a =⇒ b) as it
follows from Eq. (17) is equal to 2 (γ − α), and it does
not connect by simple way with corresponding believabil-
ities of partial propositions a and b. Let us now impose
another additional restrictionon on the phases α, β, γ, δ of
the matrix ρ (a, b) which looks enough naturally, namely,
we require that in simple version of NB logic of the sec-
ond kind the known De Morgan dual formulas for all
logical propositions, namely: not (a and b) = (not a)
or (not b) and not (a or b) = (not a) and (not b) hold.
This is just the case what we have in mind when we
are speaking about the toy model of everyday human
reasoning. Really, it is easy to see that the restriction
above immediately leads to the next condition imposed
on the phases of input matrix ρ (a, b) of Eq. (13), namely:
α+β = γ+δ = const. Since this constant must be identi-
cal for all propositions in the toy model of NB logic of the
second kind we can without loss of generality set it equal
to zero and then it turns out that all non-diagonal ele-
ments in implication ( and in other connectives as well)
can be expressed by means of the non-diagonal elements
of partial propositions. Comparing the believability of
the implication (a =⇒ b) with the believability of it con-
sequence b that is equal to −2 (α + δ) = 2 (γ − α)we im-
mediately come to the desired conclusion , namely, in
the toy version of NB logic of the second kind the values
of believabilities of the propositions (a =⇒ b) and b com-
pletely coincide (and both are equal to 2 (γ − α)).Strictly
speaking just this conclusion gives us the good reason to
believe that NB logic of the second kind has essential
common features with everyday human logic in which
(as was shown in numerous experiments) the believabil-
ity of conclusions stimulates the subjects more positively
evaluate the correctness of the process of their inference
and to pay less attention to their logical rigor (belief-bias
effect). It is necessary to mention also that as a matter
of fact the whole subjective evaluation of the validity of
7
the given proposition a simultaneously should take into
account both the plausibility of this proposition and its
believability as well. Therefore the appropriate expres-
sion for the integral validity V of the GP a should be
actually look as follows:
V (a) = αP (a) + (1 − α) B (a) .
(18)
(where the coefficient α reflects the mental nature of rea-
soning person.It remains to point out
that by appli-
cation of the results of the approach proposed in this
Section one can evaluate quantitatively both believabil-
ity and plausibility of various GPs in complex logical
situations relating to everyday human reasoning. How-
ever this amusing task is already beyond the scope of the
present paper.
In the conclusion we should like to emphasize only that
all guesses and conjectures about proposed interpreta-
tions of two NB logics expressed in this paper must be
taken with a grain of salt, that is as plausible but prelim-
inary hypotheses . Obviously, the rigorous confirmation
of all results obtained here will require more painstaking
extra work in the future.
[1] G. Birkhoff, J. von Neumann, Ann.Math.,37,823, (1936)
[2] E.D. Vol, Int.J. of Theor. Phys.,52,514,(2013)
[3] Jonathan Evans,Keith Frankish, In Two Minds: Dual Pro-
[6] M.A. Nielsen,
I.L. Chuang, Quantum Computation
and Quantum Information, Cambridge University Press,
(2001)
cesses and Beyond,Oxford University Press,(2009 )
[7] L. Levy-Bruhl, La Mythologie Primitive, Paris, (1935)
[4] E.L. Feinberg, Zwei Kulturen, Springer-Verlag,(1998)
[5] C. Levi-Strauss, La Pensee Sauvage, Paris, (1962)
|
1111.6573 | 1 | 1111 | 2011-11-28T20:23:29 | Neural integrators for decision making: A favorable tradeoff between robustness and sensitivity | [
"q-bio.NC"
] | A key step in many perceptual decision tasks is the integration of sensory inputs over time, but fundamental questions remain about how this is accomplished in neural circuits. One possibility is to balance decay modes of membranes and synapses with recurrent excitation. To allow integration over long timescales, however, this balance must be precise; this is known as the fine tuning problem. The need for fine tuning can be overcome via a ratchet-like mechanism, in which momentary inputs must be above a preset limit to be registered by the circuit. The degree of this ratcheting embodies a tradeoff between sensitivity to the input stream and robustness against parameter mistuning.
The goal of our study is to analyze the consequences of this tradeoff for decision making performance. For concreteness, we focus on the well-studied random dot motion discrimination task. For stimulus parameters constrained by experimental data, we find that loss of sensitivity to inputs has surprisingly little cost for decision performance. This leads robust integrators to performance gains when feedback becomes mistuned. Moreover, we find that substantially robust and mistuned integrator models remain consistent with chronometric and accuracy functions found in experiments. We explain our findings via sequential analysis of the momentary and integrated signals, and discuss their implication: robust integrators may be surprisingly well-suited to subserve the basic function of evidence integration in many cognitive tasks. | q-bio.NC | q-bio | Title:
Neural integrators for decision making: A favorable tradeoff between robustness and sensitivity
Abbreviated Title:
Neural integrators for decision making
Authors:
Nicholas Cain, Department of Applied Mathematics, University of Washington
Andrea K. Barreiro, Department of Applied Mathematics, University of Washington
Michael Shadlen, Department of Physiology and Biophysics, University of Washington
Eric Shea-Brown, Department of Applied Mathematics, University of Washington
Corresponding Author:
Eric Shea-Brown
Department of Applied Mathematics
Box 325420
Seattle, WA 98195-2420
[email protected]
Number of pages: 30
Number of figures: 16
Number of tables: 1
Number of words (Abstract): 226
Number of words (Introduction): 476
Number of words (Discussion): 1467
Conflict of Interest: None
Acknowledgements: This research was supported by a Career Award at the Scientific Interface
from the Burroughs-Wellcome Fund (ESB), the Howard Hughes Medical Institute, the National
Eye Institute Grant EY11378, and the National Center for Research Resources Grant RR00166
(MS), by a seed grant from the Northwest Center for Neural Engineering (ESB and MS), and by
NSF Teragrid allocation TG-IBN090004.
arXiv:1111.6573v1 [q-bio.NC] 28 Nov 2011
1
Abstract
A key step in many perceptual decision tasks is the integration of sensory inputs over time,
but fundamental questions remain about how this is accomplished in neural circuits. One
possibility is to balance decay modes of membranes and synapses with recurrent excitation. To
allow integration over long timescales, however, this balance must be precise; this is known as the
fine tuning problem. The need for fine tuning can be overcome via a ratchet-like mechanism, in
which momentary inputs must be above a preset limit to be registered by the circuit. The degree
of this ratcheting embodies a tradeoff between sensitivity to the input stream and robustness
against parameter mistuning.
The goal of our study is to analyze the consequences of this tradeoff for decision making
performance. For concreteness, we focus on the well-studied random dot motion discrimination
task. For stimulus parameters constrained by experimental data, we find that loss of sensitivity
to inputs has surprisingly little cost for decision performance. This leads robust integrators
to performance gains when feedback becomes mistuned. Moreover, we find that substantially
robust and mistuned integrator models remain consistent with chronometric and accuracy func-
tions found in experiments. We explain our findings via sequential analysis of the momentary
and integrated signals, and discuss their implication: robust integrators may be surprisingly
well-suited to subserve the basic function of evidence integration in many cognitive tasks.
1
Introduction
Many decisions are based on the balance of evidence that arrives at different points in time. This
process is quantified via simple perceptual discrimination tasks, in which the momentary value of a
sensory signal carries negligible evidence but correct responses arise from summation of this signal
over the duration of a trial. At the core of such decision making must lie neural mechanisms that
integrate signals over time (Gold and Shadlen, 2007; Wang, 2008; Bogacz et al., 2006). The function
of these circuits is intriguing, because perceptual decisions develop over hundreds of milliseconds
to seconds, while individual neuronal and synaptic activity often decays on timescales of several
to tens of milliseconds -- a difference of at least an order of magnitude. A mechanism that bridges
this gap is feedback connectivity tuned to balance -- and hence cancel -- inherent voltage leak and
synaptic decay (Cannon et al., 1983; Usher and McClelland, 2001).
The tuning of recurrent connections to achieve this balance presents a challenge (Seung, 1996;
Seung et al., 2000), illustrated in Figure 4(A) via motion of a ball on an energy surface. Here,
the ball position E(t) represents the total activity of a circuit (relative to a baseline marked 0);
momentary sensory input perturbs E(t) to increase or decrease. If decay dominates (upper-right),
then E(t) always has a tendency to "roll back" to baseline values, thus forgetting accumulated
sensory input. Conversely, if feedback connections are in excess, then activity will grow away
from the baseline value (center). If balance is perfectly achieved via fine-tuning, (left) temporal
integration can occur. That is, inputs can then smoothly perturb network activity back and forth,
so that the network state at any given time represents the time-integral of past inputs.
Koulakov et al. (2002) proposed an alternate model: a ratchet-like accumulator, equivalent to
movement along a scalloped energy surface (Figure 4(A), bottom) (Pouget and Latham, 2002).
Importantly, even without finely-tuned connectivity, network states can hold prior values without
decay or growth, allowing integration of inputs over time. Thus, this mechanism is called a robust
integrator. Energy wells can be spaced arbitrarily close together while maintaining their depth,
so that the robust integrator can represent a practically continuous range of values. However, the
energy wells imply a minimum input strength to transition between adjacent states, with inputs
below this limit effectively ignored.
2
(A)
(B)
Figure 1: Schematic of neural integrator models. (A) Visualizing integration via an energy surface
(Pouget and Latham, 2002). A robust integrator can "fixate" at a range of discrete values, indicated
by a sequence of potential wells, despite mistuning of circuit feedback. Without these wells (the
non-robust case), activity in a mistuned integrator would either exponentially grow or decay, as
in the top panels. Perturbing the robust integrator from one well to the next, however, requires
sufficiently strong momentary input. (B) As a consequence, low-amplitude segments in the input
signal ∆I(t), below a robustness limit R, are not accumulated by a robust integrator: only the high-
amplitude segments are. The piecewise-defined differential Equation (5) captures this robustness
behavior, resulting in the accumulated activity shown, and may be related to, e.g., a detailed
bistable-subpopulation model. A decision is expressed when the accumulated value E(t) crosses
the decision threshold θ.
The two models just introduced present a tradeoff between robustness to parameter mistun-
ing and sensitivity to inputs. Here, we ask how this tradeoff impacts behavioral performance in
perceptual decision making. Focusing on the moving dots task (Shadlen and Newsome, 1996;
Roitman and Shadlen, 2002), enables us to constrain model parameters to known physiology and
behavior. Our aim is to establish whether or not the robust integrator model is consistent with
known data, and to assess the performance benefits, if any, that it affords when network parameters
cannot be fine-tuned.
2 Materials and methods
2.1 Model and task overview
To explore the consequences of the robust integrator mechanism for decision performance, we begin
by constructing a two-alternative decision making model similar to that proposed by Mazurek et
al. (2003). For concreteness, we concentrate on the forced choice motion discrimination task (Roit-
man and Shadlen, 2002; Mazurek et al., 2003; Gold and Shadlen, 2007; Churchland et al., 2008;
Shadlen and Newsome, 1996; Shadlen and Newsome, 2001). Here, subjects are presented with a
field of random dots, of which a subset move coherently in one direction; the remainder are relo-
cated randomly in each frame. The task is to correctly choose the direction of coherent motion
from two alternatives (i.e., left vs. right).
As in Mazurek et al. (2003) (see also Smith (2010)), we first simulate a population of neurons
that represent the sensory input to be integrated over time. This population is a rough model of
cells in extrastriate cortex (Area MT) which encode momentary information about motion direc-
tion (Britten et al., 1993; Britten et al., 1992; Salzman et al., 1992). We pool spikes from model
3
Not Robust:
Perfect Over-tuned Under-tuned:
Robust:
Perfect Over-tuned Under-tuned:
Figure 2: Overview of model setup. Simulations of sensory neurons and neural recordings are used
to define the left and right inputs ∆Il(t), ∆Ir(t) to neural integrators (see text). These inputs are
modeled by gaussian (OU) processes, which capture noise in the encoding of the motion strength
by each pool of spiking neurons. See Equations (1)-(3) for definition of input signals. Similar to
Mazurek et al. (2003), the activity levels of the left and right integrators El(t) and Er(t) encode
In the reaction time task, El(t) and Er(t) race to
accumulated evidence for each alternative.
thresholds in order to determine choice on each trial. In the controlled duration task the choice is
made in favor of the integrator with higher activity at the end of the stimulus presentation.
MT cells that are selective for each of the two possible directions into separate streams, labeled
according to their preferred "left" and "right" motion selectivity: see Figure 2.
Two corresponding integrators then accumulate the difference between these streams, left-less-
right or vice-versa. Each integrator therefore accumulates the evidence for one alternative over the
other. Depending on the task paradigm, different criteria may be used to terminate accumulation
and give a decision.
In the reaction time task, accumulation continues until activity crosses a
decision threshold: if the leftward evidence integrator reaches threshold first, a decision that overall
motion favored the leftward alternative is registered.
Accuracy is defined as the fraction of trials that reach a correct decision. Speed is measured
by the time taken to cross threshold starting from stimulus onset. Reaction Time (RT) is then
defined as the time until threshold (decision time) plus 350 ms of non-decision time, accounting
for other delays that add to the time taken to select an alternative (e.g. visual latencies, or motor
preparation time, cf. (Mazurek et al., 2003; Luce, 1986)). The exact value of this parameter was
not critical to our results. Task difficulty is determined by the fraction of coherently moving dots
C (Britten et al., 1992; Mazurek et al., 2003; Roitman and Shadlen, 2002). Accuracy and RT across
multiple levels of task difficulty define the accuracy and chronometric functions in the reaction time
task, and together can be used to assess model performance. When necessary, these two numbers
can be collapsed into a single metric, such as the reward per unit time or reward rate.
In a second task paradigm, the controlled duration task, motion viewing duration is set in
advance by the experimenter. A choice is made in favor of the integrator with greater activity
at the end of the stimulus duration. Here, the only measure of task performance is the accuracy
function.
4
Sensory Neurons:
Input Signals:
Recurrent Integration:
Threshold:
Left:
Right:
Figure 3: Construction of gaussian (OU) processes to represent fluctuating, trial-by-trial firing rate
of a pool of weakly correlated MT neurons (Bair et al., 2001; Zohary et al., 1994). As in Mazurek
and Shadlen (2002), these motion sensitive neurons provide direct input to our model integrator
circuits. Simulated spike trains from weakly-correlated, direction selective pools of neurons are
shown as a rastergram. All spikes prior to time t -- a sum over the jth spike from the ith neuron,
for all i and j -- are convolved with an exponential filter, and then summed to create a continuous
stochastic output (right); here, H(t) is the Heaviside function. We approximated this output by
a simpler gaussian (OU) process in order to simplify numerical and analytical computations that
follow.
2.2 Sensory input
We now describe in detail the signals that are accumulated by the integrators corresponding to
the "left" and "right" alternatives. First, we model the pools of leftward or rightward direction-
selective sensory (MT) neurons as N = 100 weakly correlated spiking cells (Pearson's correlation
ρ = .11 (Zohary et al., 1994; Bair et al., 2001)); see Figure 3. Specifically, as in Mazurek and
Shadlen (2002), each neuron is modeled via an unbiased random walk to a spiking threshold; the
random walks of neurons in the same pool are correlated. Increasing the variance of each step in
the random walk increases the firing rate of each model neuron; it was therefore chosen at each
coherence value to reproduce the linear relationship between coherence C and mean firing rate µl,r
of the left and right selective neurons observed in MT recordings:
µl,r(C) = r0 + bl,rC .
(1)
Here the parameters r0, bl, and br are derived from firing rates observed across a range of coher-
ences (Britten et al., 1993). If evidence favors the left alternative, bl = .4 and br = −.2; if the right
alternative is favored, these values are exchanged.
Next, the output of each spiking pool was aggregated. Each spike emitted from a neuron in
the pool was convolved with an exponential filter with time constant 20 ms. This is intended as
an approximate model of the smoothing effect of synaptic transmission. These smoothed responses
were then summed to form a single stochastic process for each pool (see Figure 3, right).
We then approximated the smoothed output of each spiking pool by a simpler stochastic process
that captures the mean, variance, and temporal correlation of this output as a function of dot
coherence. We used gaussian processes Il(t) and Ir(t) for the rightward- and leftward-selective pools
(See Figure 3). Specifically, we chose Ornstein-Uhlenbeck (OU) processes, which are continuous
5
Cell #
Sensory Neurons
gaussian process generated by the stochastic differential equations
dIl,r =
µl,r(C) − Il,r
τ
dt +r 2νl,r(C)
τ
dWt
(2)
with mean µl,r(C) as dictated by Equation 1. The variance νl,r(C) and timescale τ were chosen to
match the steady-state variance and autocorrelation function of the smoothed spiking process. As
we will see, this timescale plays an important role in determining the decision making performance
of robust integrators.
Our construction so far accounts for variability in output from left vs. right direction selective
neurons. We now incorporate an additional noise source into the output of each pool. These noise
terms (ηl(t) and ηr(t), respectively) could approximate, for example, neurons added to each pool
that are nonselective to direction. Each noise source is modeled as an independent OU process with
mean 0, timescale 20 ms as above, and a strength (variance) νγ/2. This noise strength is a free
parameter that we vary to match behavioral data (see "A robust integrator circuit" and Figure 14).
We note that previous studies also found that performance based on the direction-sensitive cells
alone can be more accurate than behavior, and therefore incorporated variability in addition to the
output of "left" and "right" direction selective MT cells (Shadlen et al., 1996; Mazurek et al., 2003;
Cohen and Newsome, 2009).
Finally, the signals that are accumulated by the left and right neural integrators are constructed
by differencing the outputs of the two neural pools:
∆Il(t) = [Il(t) + ηl(t)] − [Ir(t) + ηr(t)]
∆Ir(t) = −∆Il(t) .
(3)
2.3 Neural integrator circuit and feedback mistuning
A central focus of our paper is variability in the relative tuning of recurrent feedback vs. decay
in an integrator circuit. Below, we will introduce the mistuning parameter β, which determines
the extent to which feedback and decay fail to perfectly balance. We first define the dynamics of
the integrator circuit on which our studies are based. This is described by the firing rates El,r(t)
of integrators that receive outputs from left-selective or right-selective pools ∆Il,r(t) respectively.
The firing rates El,r(t) increase as evidence for the corresponding task alternative is accumulated
over time:
dEl,r
dt
τE
= −El,r + (1 + β)El,r + κ∆Il,r(t).
(4)
The three terms in this equation account for leak, feedback excitation, and the sensory input (scaled
by a weight κ), respectively. When the mistuning parameter β = 0, leak and self-excitation exactly
cancel; we describe such an integrator as perfectly tuned, while an integrator with β 6= 0 is said
to be mistuned. Imprecise feedback tuning is modeled by randomly setting β to different values
from trial to trial (but constant during a given trial), with a mean value ¯β and a precision given
by a standard deviation σβ. We assume that ¯β = 0 for most of the study. Thus the spread of β,
which we take to be gaussian, represents the intrinsic variability in the balance between circuit-level
feedback and decay. Perfect tuning corresponds to σβ = ¯β = 0, while σβ 6= 0 or ¯β 6= 0 corresponds
to a mistuned integrator. Finally, we set initial activity in the integrators to zero (El,r(0) = 0),
and impose reflecting boundaries at Er = 0, El = 0 (as in, e.g., Smith and Ratcliff (2004)) so that
firing rates never become negative.
6
2.4 A robust integrator circuit
A robust integrator can be constructed from a series of bistable subpopulations, which sequentially
activate in order to represent the accumulated evidence (Koulakov et al., 2002; Nikitchenko and
Koulakov, 2008). The many equations that describe the evolution of these systems can be closely
approximated with reduced models, as demonstrated in Goldman et al. (2003). We derived a
single piecewise-defined differential equation model that approximates the dynamics of a robust
integrator constructed from bistable pools.
All subsequent results are based on this simplified model, which captures the essence of the
robust integration computation:
τE
dEl,r
dt
=(cid:26) 0
βEl,r + κ∆Il,r
: βEl,r + κ∆Il,r ≤ R
: otherwise
(5)
The first line represents the series of potential wells discussed in the Introduction (see Figure 4): if
the sum of the mistuned integrator feedback and the input falls below the robustness limit R, the
activity of the integrator remains fixed. If this summed input exceeds R, the activity evolves as for
the non-robust integrator in Equation (4). To interpret the robustness limit R, it is convenient to
normalize by the standard deviation of the input signal:
R =
R
ST D [∆Il,r(t)]
.
(6)
In this way, R can be interpreted in units of standard deviations of input OU process that are
"ignored" by the integrator. We note that Equation (5) is similar to the effective equation derived
for a different implementation of a robust integrator (Goldman et al., 2003).
To summarize, Equation (5) defines a parameterized family of neural integrators, distinguished
by the robustness limit R. As R → 0, the model reduces to Equation (4). When additionally β = 0,
the (perfectly tuned) integrator computes an exact integral of its input: Equation (5) then yields
El,r(t) ∝R t
0 ∆Il,r(t0)dt0.
2.5 Computational methods
Monte Carlo simulations of Equations (1)-(5) were performed with Euler-Maruyama method (Higham,
2001), with dt = 0.1 ms. For a fixed choice of input statistics and threshold θ, a minimum of 10, 000
trials were simulated to estimate accuracy and RT values. During simulations of the reaction time
task, in order to prevent excessively long trials (particularly at low coherence values) a maximum
simulation time was set at 10,000 ms. At this time, if neither integrator had reached threshold, the
indeterminate result was broken by a numerical "coin flip", (this rarely occurred, as indicated by
the RT histograms in Figures 15, 16).
In simulations where σβ > 0, results were generated across a range of β values and then
marginalized by weighting according to a normal distribution. The range of values was chosen with
no less than 19 linearly spaced points, across a range of ± 3 standard deviations around the mean
¯
β.
Reward rate values presented in "Reward rate and the robustness-sensitivity tradeoff" are pre-
sented as maximized by varying the free parameter θ; values were computed by simulating across a
range of θ values. The range and spacing of these values were chosen dependent on the values of R
and β for the simulation; the range was adjusted to capture the relative maximum of reward rate
as a function of θ, while the spacing was adjusted to find the optimal θ value with a resolution of
7
Figure 4: Parameter space view of four integrator models, with different values of the robustness
limit R and feedback mistuning σβ. The impact of transitioning from one model to another by
changing parameters is either to enhance or diminish performance, or to have a neutral effect (see
text).
± 0.1.
Values of θ and νγ in the table included in Figure 14 were chosen to best match accuracy and
chronometric functions to behavioral data reported in Roitman and Shadlen (2002). This was
accomplished by minimizing the sum-squared error in data vs. model accuracy and chronometric
curves across a discrete grid of θ and νγ values, with a resolution of 0.1. When data between
simulated values were needed, linear interpolation was used to approximate the corresponding
accuracy and RT values.
Autocovariance functions of integrator input, presented later, were computed by simulating an
Ornstein-Uhlenbeck process using the exact numerical technique in Gillespie (1996) with dt = 0.1
ms, to obtain a total of 227 sample values. Sample values of the process less than the specified
robustness limit R were set to 0, and the autocovariance function was computed using standard
Fourier transform techniques.
Simulations were performed on NSF Teragrid clusters.
3 Results
3.1 How do robustness and mistuning affect decision speed and accuracy?
In the Methods, we define a general neural integrator model (Equation (5)) that accumulates signals
representing the output of motion sensitive neurons (Equation 3). The integrator model includes
two key parameters. The first is β, which represents the mistuning of feedback from a value that
perfectly balances decay; the extent of this mistuning is measured by σβ, the standard deviation
of β from the ideal value β = 0. The second is the robustness limit R. We emphasize twin
effects of R: as R increases, the integrator becomes able to produce a range of graded persistent
activity for ever-increasing levels of mistuning (see Figure 4 (A), where R corresponds to the
depth of energy wells). This prevents runaway increase or decay of activity when integrators are
mistuned; intuitively, this might lead to better performance on sensory accumulation tasks. At the
same time, as R increases integrator activity remains fixed even for increasingly strong positive or
negative momentary input ∆Il,r (see Figure 4 (B), where R specifies a limit within which inputs
are "ignored"). Such sensitivity loss should lead to worse performance. This implies a fundamental
tradeoff between competing effects: (1) one would prefer to not ignore relevant input stimulus,
8
"Mistuned"
Model
"Recovery"
Model
"Baseline"
Model
"Robust"
Model
Figure 5: Mistuned feedback diminishes decision performance. (Inset) In both figures we consider a
move in parameter space from the "baseline" model to the "mistuned" model by changing σβ = 0 →
0.1 (A) In the controlled duration task, accuracy is lower for the "mistuned" model (dashed line)
than for the "baseline" model (solid line) at every trial duration T , indicating a loss of performance
when σβ increases. (B) In the reaction time task, we plot the curve of all (RT, accuracy) pairs
attained by varying the decision threshold θ (see text). Once again, accuracy is diminished by
mistuning.
favoring small R, and (2) one would prefer an integrator robust to mistuning, favoring large R.
Figure 4 gives a schematic of how the two model parameters, σβ and R, define a plane of possible
integrator models. Here, we explore decision performance in four different cases arranged in this
plane. By contrasting integrators with different values of the robustness limit R, we can assess how
the fundamental tradeoff plays out, to either improve or degrade decision making performance.
In order to assess this performance, we consider relationships between decision speed and accu-
racy in both controlled duration and reaction time tasks. In the controlled duration task, we simply
vary the stimulus presentation duration, and plot accuracy vs. experimenter-controlled stimulus
duration. In the reaction time task, we vary the decision threshold θ -- treated as a free parameter
-- over a range of values, thus tracing out the speed accuracy curve for all possible pairs of speed
and accuracy values. Here, speed is measured by reaction time (RT), the latency between the onset
of stimulus and crossing of the decision threshold. For both cases, we use a single representative
dot coherence (C=12.8 in Equation 1); results are qualitatively similar for other coherence values
(data not shown); slightly (approx. 25%) lower robustness limits are required at the lowest dot
coherence of C = 3.2.
We first study a case we call the "baseline" model, for which there is no mistuning or robustness:
σβ = R = 0. Speed accuracy plots for this model are shown as a solid line in Figs. 5(A) and (B),
for the controlled duration and reaction time tasks respectively. We compare the "baseline" model
with the "mistuned" model, for which the feedback parameter has a standard deviation of σβ = 0.1
(10% of the mean feedback) and robustness R = 0 remains unchanged. In the controlled duration
task (Figure 5(A)) we observe that mistuning diminishes accuracy by as much as 10%, and this
effect is sustained even for arbitrarily long viewing windows (cf. (Usher and McClelland, 2001;
9
Controlled Duration (sec)
Reaction Time (ms)
Accuracy
Accuracy
Figure 6: Increasing the robustness limit R helps recover performance lost due to feedback mis-
tuning. (Inset) We illustrate this by moving in parameter space from the "mistuned" model to
the "recovery" model, by changing R = 0 → .85. The impact on decision performance is shown
for both the controlled duration (A) and reaction time (B) tasks. For each task we plot the rela-
tionship between speed and accuracy as above: solid lines indicate the "baseline" model, dotted
the "mistuned" model, and now dash-dotted the "recovery" model. We find that R > 0 yields a
modest performance gain for the "recovery" model in comparison with the "mistuned" model.
Bogacz et al., 2006)). The reaction time task (Panel B) produces a similar effect: for a fixed RT,
the corresponding accuracy is decreased.
Next we increase the robustness limit to R = 0.85 -- so that almost ± a standard deviation of
the input stream is "ignored" by the integrators -- while maintaining feedback mistuning. We call
this case the "recovery" model because robustness compensates in part for the performance loss due
to feedback mistuning: the speed accuracy plots in Figure 6 for the recovery case lie above those
for the "mistuned" model. For example, at the longer controlled task durations (Panel A) and
reaction times (Panel B) plotted, 20% of the accuracy lost due to integrator mistuning is recovered
via the robustness limit R = 0.85.
Finally, we study the remaining possibility, when the robustness parameter R is increased from
zero in a perfectly tuned integrator (σβ = 0); this is the "robust" case in Figure 4. We expected
performance to be substantially diminished as a consequence of lost sensitivity to inputs. However,
Figure 7 demonstrates that this is not the case: speed accuracy curves for R = 0.85 almost coincide
with those for the "baseline" case of R = 0. We emphasize again that because R measures ignored
input in units of the standard deviation, the integrator circuit is actually not integrating the weakest
60% of the input stimulus. Given this large amount of ignored stimulus, the fact that the robust
integrator produces nearly the same accuracy and speed as the "baseline" case is surprising. This
implies that the "robust" model can protect against feedback mistuning, without substantially
sacrificing performance when feedback is perfectly tuned.
To summarize, the ratchet-like mechanism of the robust integrator appears well-suited to the
decision tasks at hand. This mechanism counteracts some of the performance lost when feedback
fails to be perfectly fine-tuned. Moreover, even when this fine-tuning is achieved, a robust integrator
10
Controlled Duration (sec)
Reaction Time (ms)
Accuracy
Accuracy
Figure 7: Increasing R alone does not compromise performance.
(Inset) We illustrate this by
moving in parameter space from the "baseline" to the "robust" model. For both (A) controlled
duration and (B) reaction time tasks, we plot the relationship between speed and accuracy. Solid
lines give results for the "baseline" model, and dash-dotted for the robust model. The curves are
very similar in the "baseline" and robust cases, indicating little change in decision performance due
to the robustness limit R = 0.85.
still performs as well as the "baseline" case that is perfectly sensitive to the input signal. In the
next section, we begin to explain this observation by constructing several simplified models and
employing results from statistical decision making theory.
3.2 Analysis: Robust integrators and decision performance
3.2.1 Controlled duration task: Discrete time analysis
We can begin to understand the effect of the robustness limit on decision performance by formu-
lating a simplified version of the evidence accumulation process. We focus first on the controlled
duration task, where the analysis is somewhat simpler.
Our first simplification is to consider a single accumulator E which receives evidence for or
against a task alternative in discrete time. The value of E on the ith time step, Ei, is allowed
to be either positive or negative, corresponding to accumulated evidence favoring the leftward or
rightward alternatives, respectively. On each time step, Ei increments by an independent, random
value Zi with a probability density function (PDF) fZ(Z). We first describe an analog of the
"baseline" model above; i.e., in the absence of robustness ( R = 0). Here, we take the increments
Zi to be independent, identical, and normally distributed, with a mean µ > 0 (i.e. biased toward
the leftward alternative; we call this the preferred alternative) and standard deviation σ: that is,
Zi ∼ N(cid:0)µ, σ2(cid:1). After the nth step, we have
En =
Zi .
n
X
i=1
11
Controlled Duration (sec)
Reaction Time (ms)
Accuracy
Accuracy
(A)
(B)
Figure 8: The effect of R on a discrete time increment distribution, and the second real root of
the moment generating function of this distribution. (A) The PDF of the random variable Z R,
with probability mass for values between the robustness limit R re-allocated as a delta function
centered at zero ( R = 1). (B) The second real root h0 of MZ R
(s) remains unchanged as R increases
from 0 → 2. (Lines are uniformly distributed in this range.) This implies that in the reaction time
task, no changes in the accuracy and chronometric functions will be observed until the deviation
in E[ZR] becomes large (Discussed in "Reaction time task: Continuous analysis").
In the controlled duration task, a decision is rendered after a fixed number of time steps N , (i.e.
n = N ) and a correct decision (i.e., in favor of the preferred alternative) occurs when EN > 0. By
construction, En ∼ N(cid:0)nµ, nσ2(cid:1), which implies that accuracy (Acc) can be computed as a function
of the signal-to-noise ratio (SNR) s = µ
σ of a sample:
Acc =Z ∞
0
1
√2πN σ2
e− (x−N µ)2
2N σ2 dx =
1 + Erf(cid:18)q N
2 s(cid:19)
2
.
(7)
Next, we change the distribution of the accumulated increments Zi to construct a discrete
time analog of the robust integrator. Specifically, increasing the robustness parameter to R > 0
affects increments Zi by redefining the PDF fZ(Z) so that weak samples do not add to the total
accumulated "evidence", precisely as in Equation (5). (Models where such a central "region of
uncertainty" of the sampling distribution is ignored have previously been studied in a race-to-
bound model (Smith and Vickers, 1989); see Discussion). This requires reallocating probability
mass below the robustness limit to zero. We plot the resulting PDF in Figure 8(A), where the
reallocated mass gives a weighted delta function at zero. Specifically:
fZR (Z) = δ (0)Z R
−R
fZ(cid:0)Z0(cid:1) dZ0 +(cid:26)
0
fZ(Z)
Z < R
:
: otherwise
(8)
The central limit theorem then allows us to approximate the new cumulative sum EN R
as
a normal distribution (for sufficiently large N ), with µ and σ in Equation (7) replaced by the
mean and standard deviation of the PDF defined by Equation (8). As before, we normalize R
by the standard deviation of the increment, R = R
σ , and then express the fraction correct Acc R
as a function of R and s. One can think of R as perturbing the original Acc function (Equation
(7)), and although this perturbation has a complicated form, we can understand its behavior by
12
Figure 9: Comparison of the decision accuracy predicted by discrete and continuous time approxi-
mate models for the controlled duration task (as above, coherence C = 12.8). (A) Both discrete and
continuous time models predict that accuracy (Acc) will stay roughly constant as R increases up
to ≈ 0.5, followed by a gradual decrease. However, the continuous time approximation provides a
closer match to results for the full model pictured in Figure 7(A) (see text). Additionally, compar-
ing the dashed and dot-dashed lines shows that the approximation in Equation 9 provides a good
description of the discrete time model for R . 1. (B) The (numerically computed) autocovariance
functions of the input signal Z R(t) at various levels of R that were used to construct the continuous
time curve in (A) (Equation (9)). (Inset) Two of these same autocovariance functions (for R = 0
and R = 1.5) are plotted normalized to their peak value. This shows that autocorrelation falls off
faster as R increases.
observing that its Taylor expansion has only one nonzero term up to fifth order in R:
Acc R(N ) ≈ Acc(N )(1 − P (N ))
(9)
P (N ) =
√N s(cid:0)1 + 2s2(cid:1) e− (1+N )s2
6π
2
R3 + O( R5)
Thus, for small values of R (giving very small R3), there will be little impact on accuracy.
Equation (9) can therefore partially explain the key observation in Fig 7(A) that R can be
increased to 0.85 while incurring very little performance loss. For concreteness, we focus on decisions
at T = 500 ms. To make a rough comparison, we first assume that a new sample of evidence arrives
in the discrete time model every 10 ms. We then set the SNR s so that Acc = 0.92 for the discrete
model when N = 50, agreeing with the accuracy obtained from the continuous model at 500
ms (Fig 7(A)) when R = 0. We then increase the robustness limit R. Figure 9(A) shows that
accuracy for both the discrete time model itself (dashed line), and its approximation up to O( R5)
(dot-dashed), barely decrease at all while R is less than ≈ 0.5, and then begin to fall off. This is
consistent with the results for the full model in Fig 7(A). However, the discrete time model does
predict a small decrease in accuracy at R = 0.85 that is not seen in the full model. In the next
section, we explain how this discrepancy can be resolved.
3.2.2 Controlled duration task: Continuous time analysis
We next extend the analysis of the controlled duration task in the previous section to signal inte-
gration in continuous time. In brief, we follow a method developed in Gillespie (1996) to describe
the evolution of the mean and variance of a continuous input signal that has been integrated over
13
Accuracy
time. This is challenging and interesting because, as for the signals used in modeling the random
dots task above (see Methods, Sensory Input), this input signal contains temporal correlations. As
in the previous section, we describe the distribution of the integrated signal at the final time T ,
which determines accuracy in the controlled duration task.
We first replace the discrete input samples Zi from the previous section with a continuous signal
Z(t), which we take to be a (OU) gaussian process with a correlation timescale derived from our
model sensory neurons (see Methods). We define the integrated process
dE
dt
= Z(t) → E(t) =Z t
0
Z(t0)dt0
with initial condition E(0) = 0.
Assuming that Z(t) satisfies certain technical conditions that are easily verified for the OU pro-
cess (wide-sense stationarity, α-stability, and continuity of sample paths (Gardiner, 2002; Billings-
ley, 1986; Gillespie, 1996)), we can construct differential equations for the first and second moments
hE(t)i and(cid:10)E2(t)(cid:11) evolving in time. We start by taking averages on both sides of our definition of
E(t), and, noting that E(0) = 0, compute the time-varying mean:
dhE(t)i
dt
= hZ(t)i =⇒ hE(t)i = thZ(t)i .
(10)
Similarly, we can derive a differential equation for the second moment of E(t):
d(cid:10)E2(t)(cid:11)
dt
= 2hZ(t)E(t)i .
The righthand side of this equation can be related to the area under the autocovariance function
A (τ ) ≡ hZ(t)Z(t + τ )i − hZ(t)i2 of the process Z(t):
hZ(t)E(t)i =(cid:28)Z(t)Z t
0
Z(s)ds(cid:29) = Z t
= Z t
= Z t
0
0 hZ(t)Z(s)i ds
0 hZ(t)Z(t − τ )i dτ
A (τ ) + hZ(t)i2 dτ
We now have an expression for how the second moment evolves in time. We can simplify the result
via integration by parts:
(cid:10)E2(t)(cid:11)
=
2Z t
0 Z s
0
A (τ ) + hZ(t)i2 dτ ds = 2Z t
0
(t − τ )A (τ ) dτ + t2 hZ(t)i2
=⇒ Var[E(t)] = 2Z t
0
(t − τ )A (τ ) dτ.
(11)
Because E(t) is an accumulation of gaussian random samples Z(t), it will also be normally dis-
tributed, and hence fully described by the mean (Equation (10)) and variance (Equation (11)) (Billings-
ley, 1986).
To model a non-robust integrator, as discussed above we take Z(t) to be a OU process with
steady-state mean and variance µ and σ2, and time constant τ . For the robust case, we can
follow Equation 5 and parameterize a family of processes Z R(t) with momentary values below
14
the robustness limit R set to zero. (Here, we again normalize the robustness limit by the standard
deviation of the OU process.) We numerically compute the autocovariance functions A R (τ ) of these
processes, and use the result to compute the required mean and variance, and hence time-dependent
signal-to-noise ratio SNR(t), for the integrated process E(t). This yields
SN R R(t) =
.
(12)
hE(t)i
pV ar[E(t)]
=
q2R t
tE[Z R(t)]
0 (t − τ )A R (τ ) dτ
Under the assumption that E(T ) is approximately gaussian for sufficiently long T (which can be
verified numerically), we use this SNR to compute decision accuracy at T :
1 + Erf(cid:16)q 1
2 SN R R(T )(cid:17)
2
Acc R(T ) ≈
.
(13)
This function is plotted for T = 500 ms as the solid line in Figure 9(A). The plot shows that
accuracy remains relatively constant until the robustness limit R exceeds ≈ 0.85. Interestingly, this
is a longer range of R values than for the discrete time case (compare dotted line in Figure 9(A)),
and is closer to the results for the full model pictured in Figure 7(A).
Why does the robustness limit appear to have a milder effect on degrading decision accuracy
for our continuous vs. discrete time input signals? We can get some insight into the answer by
examining the autocovariance functions A R (τ ), which we present in Figure 9(B). When normalized
by their peak value, the autocovariance for R > 0.5 falls off more quickly vs. the time lag τ (see inset
in Figure 9(B)), indicating that subsequent samples become less correlated in time. Thus, there
are effectively more "independent" samples that are drawn over a given time range T , improving
the fidelity of the signal. This effect is not present in our discrete time model.
Summary: Our analysis of decision performance for the controlled duration task shows that two
factors contribute to the preservation of decision performance for robust integrators. The first is
that, for robustness limits up to R ≈ 0.5, the momentary SNR of the inputs is barely changed by
setting values below robustness limit to zero. The second is that, as R increases, the signal Z R(t)
being integrated becomes less correlated in time. This means that (roughly) more independent
samples will arrive over a given time period.
3.2.3 Reaction time task: Discrete analysis
We begin our analysis of the reaction time task by introducing a discrete time, discrete space
random walk model. In this model, schematized in Figure 10(A) with five intermediate states, a
particle representing the accumulated value E starts at a state balanced between two absorbing
"sink" states. At every time step, the particle moves towards the "correct" (i.e. preferred) sink
with probability p(1−R), and towards the "incorrect" (null) sink with probability (1−p)(1−R) (we
consider p > 0.5, biasing the random walk toward the "correct" sink). There is also the possibility
that the particle might remain in the current state, with probability R.
We now draw an analogy between the states in this random walk model and the ratcheting
dynamics among energy wells in a robust integrator (see Figure 4 and Introduction). Here, the po-
sition of the particle represents the accumulated evidence for the left vs. right alternatives, and the
absorbing states represent crossing of the corresponding decision thresholds. When the robustness
limit R is increased, the wells -- each of which could represent a bistable neural subpopulation (see
Methods) -- act to hold the particle in a given state, with a probability set by R.
15
(A)
(B)
Figure 10: Analysis of the effect of the robustness limit in the reaction time task. (A) Biased
random walk between two absorbing boundaries. The correct and incorrect states act as "sinks"
of the discrete time discrete space Markov chain. The intermediate states are analogous to the
potential wells in the robust integrator model. Here, a particle will remain in the current state
at the next time step with probability R. The final probability of ending up in either sink is
independent of R. (B) Speed accuracy tradeoff curves from Figure 7 are replotted and labeled (c),
while a new line labeled (d) shows the performance predicted by the discrete time, continuous space
model described in "Reaction time task: Continuous analysis". Here the signal-to-noise ratio of
the discrete increments were chosen so that the line generated in the R = 0 case would overlay the
line from the R = 0 continuous model, and is therefore not plotted. We see that the discrete time
model over-predicts the impact of the robustness limit R, just as in the controlled duration case
(Figure 9).
16
Reaction Time (ms)
Accuracy
As R is increased in the random walk model, the probability of transitioning out of a given state
similarly decreases. Standard results on Markov chains (see, for example, Kemeny and Snell (1960))
provide formulas for the probability that the particle will end in one vs. the other sink, as well as
the expected number of time steps until this occurs, based on the transition matrix associated with
the random walk. The probability of ending in the "correct" sink corresponds to decision accuracy,
and is found at the middle entry in the solution vector x of the matrix equation
(I − Q) x = (1 − R)pe1 .
(14)
Here Q is a tridiagonal matrix with R on the main diagonal, p(1 − R) on the lower diagonal, and
(1 − p)(1 − R) on the upper diagonal; e1 is the canonical basis vector with e(1)
1 = 1, and all other
entries equal to 0. After some factoring, we find a common factor of (1 − R) on both sides of the
equation; thus the solution to x is independent of R. This implies that the probability of ending
up in the correct state is unchanged by increasing R from the non-robust case (R = 0). Intuitively
this makes sense:
if one conditions on the fact that one will leave the current state on the next
time step, the probability of moving toward the correct and incorrect states are independent of R.
The same is not true for the expected number of steps necessary to reach a sink (by analogy,
the reaction time). This is because the matrix system that yields reaction times is:
(I − Q) t = 1.
(15)
Here the right-hand side of this equation is the vector of all ones, and therefore no equivalent
cancellation can occur. However, we do notice that the reaction time with R 6= 0 is just a rescaling
of the original reaction time with R = 0. Specifically, if tR is the expected number of steps required
to reach an absorbing state, then
tR = t0
.
(16)
1
1 − R
(17)
(18)
h0(cid:21)
θ 2
Acc =
RT =
1
1 + eθh0
θ
E[Z]
tanh(cid:20)−
17
Summary: We have used a simplified random walk model to gain intuition about the effect of the
robustness limit in the reaction time task, and to show that adding a robustness limit only affects
decision latency, but not accuracy. In the next section, we will derive a similar result for continuous
sample distributions.
3.2.4 Reaction time task: Continuous analysis
We return to the continuous sampling distribution introduced in "Controlled duration task: Dis-
crete time analysis", but now in the context of threshold crossing in the reaction time task. The
accumulation of these increments toward decision thresholds can be understood as the sequen-
tial probability ratio test, where the log-odds for each alternative are summed until a predefined
threshold is reached (Wald, 1945; Gold and Shadlen, 2002; Luce, 1963; Laming, 1968). Wald (1944)
provides an elegant method of computing decision accuracy and speed (RT). The key quantity is
given by the moment generating function (MGF, denoted MZ(s) and defined in Equation 19) for the
samples Z (see Luce (1986) and Doya (2007), Chapter 10). Under the assumption that thresholds
are crossed with minimal overshoot, we have the following expressions:
Thus, the only effect of the robustness limit R is to delay arrival at the sinks.
where h0 is one of the two real roots of the equation MZ(s) = 1 (the other root is precisely 1) and
θ is the decision threshold.
We first consider the case of a non-robust integrator, for which the samples Z are again normally
distributed. In this case, we must solve the following equation to find s = h0:
MZ(s) = EZ [es∗z] =Z ∞
It follows that s = 1 and s = h0 = −2 µ
σ2 provide two real solutions of this equation. (Wald's
Lemma ensures that there are exactly two such real roots, for any sampling distribution meeting
easily satisfied technical criteria.)
fZ (z) es∗zdz = e
s2σ2
2 +sµ = 1.
(19)
−∞
When the robustness limit R > 0, we can again compute the two real roots of the associated
MGF. Here, we use the increment distribution fZR(Z) given by Equation (8), for which all proba-
bility mass within R of 0 is reassigned precisely to 0. Surprisingly, upon plugging this distribution
into the expression MZ(s) = 1, we find that s = 1, h0 continue to provide the two real solutions to
this equation regardless of R, as depicted in Figure 8(B).
This observation implies that (1) accuracies (Equation (17)) are unchanged as R is increased,
and (2) reaction times (Equation (18)) only change when E[ZR] changes. In other words, the inte-
grator can ignore inputs below an arbitrary robustness limit at no cost to accuracy, and a penalty
in terms of reaction time will only be observed when E[ZR] changes appreciably. Generalizing our
result, we note that a sufficient condition for h0 to be unchanged as R changes is that the original
sampling distribution fZ(z) obeys
fZ(z) = fZ(−z)eh0z;
it is straightforward to verify that the Gaussian satisfies this property.
We next determine the magnitude of R necessary to change E[ZR]. When we substitute R = R
σ
and compute the perturbation to E[ZR], we again find only one term up to fifth order in R:
(20)
(21)
E[ZR] =Z ∞
P =r 2
−∞
9π
z fZR (z) dz = E[Z] (1 − P )
2 ( µ
σ )2
e− 1
R3 + O( R5)
This outcome is similar to the controlled duration case: small values of R will have little effect on
E[ZR], and therefore little effect on increasing decision speed (via Equation (18)). Moreover, as we
have already shown, accuracy is unaffected by robustness limits R of any value. As a consequence
we expect speed accuracy curves to change only modestly for small values of R. We illustrate
this via a speed accuracy plot in Figure 10(B). Here, the present discrete time, continuous space
model produces the chain-dotted curve (marked (d)), showing a moderate decrease in performance
at R = 0.85. This decrease is purely due to the increase in RT just discussed.
However, the model at hand does not reproduce the speed accuracy curve for the continuous
time model shown in Figure 10(B). Indeed, the continuous time model produces better performance
(higher accuracy at a given speed). This suggests an additional effect in the continuous time case:
once again, the fact that R reduces autocorrelation of the integrated signal increases the fidelity of
the input, improving performance (see inset in Figure 9(B)). Unlike the simpler controlled duration
task, attempting a mathematical analysis of this effect is beyond the scope of this paper.
Summary of analysis: We pause to summarize our analysis of how the robustness limit R impacts
decision performance. For both the controlled duration and reaction time tasks, we first studied
18
Figure 11: Using the Reward Rate metric to quantify recovery of decision performance as the
robustness limit is increased, C = 12.8. (A) Speed accuracy curves plotted for multiple values
of R; as in previous figures, the greater accuracies found at fixed reaction times indicate that
performance improves as R increases. The heavy line indicates the "baseline" case of a perfectly
tuned, non-robust integrator (repeated from Figure 5(B)). RR isoclines are plotted in background
(dotted lines; see text), and points along speed accuracy curves that maximize RR are shown as
circles. These maximal values of reward rate are plotted in (B), demonstrating the non-monotonic
relationship between R and the best achievable RR.
the effect of this limit on the evidence carried by momentary values of sensory inputs. In each task,
this effect was more favorable than might have been expected: in the controlled duration case, the
signal to noise ratio of momentary inputs was preserved for a fairly broad range of R, while in
the reaction time task, R was shown to affect accuracy but not speed at fixed decision threshold.
Moreover, the robustness mechanism serves to decorrelate input signals in time. This contributes
further to decision performance being preserved as the robustness limit increases.
3.3 Reward rate and the robustness-sensitivity tradeoff
Until now, we have examined performance in the reaction time task by plotting the full range
of attainable speed and accuracy values. The advantage of this approach is that it demonstrates
decision performance in a general way. An alternative, more compact approach, is to assume a
specific method of combining speed and accuracy into a single performance metric. This approach
is useful in quantifying decision performance, and rapidly comparing a wide range of models.
Specifically, we use the reward rate (RR) (Gold and Shadlen, 2002; Bogacz et al., 2006):
RR =
F C
RT + Tdel
.
(22)
Reward rate can be thought of as the number of correct responses made per unit time, with
a delay Tdel imposed between responses to penalize rapid guessing.
Implicitly, this assumes a
motivation on the part of the subject which may not be true; in general, subjects rarely achieve
19
0.0
.6
.85
1.15
Reaction Time (ms)
Accuracy
(A)
(B)
Figure 12: A nonzero robustness limit improves performance across a range of mistuning baises
¯
β.
In both the reaction time (A) and controlled duration (B) tasks, robustness helps improve
performance when β ∼(cid:0) ¯β, .12(cid:1), for all values of ¯β shown. Here, as in previous panels, the coherence
of the sensory input is C = 12.8. In the reaction time task, θ is varied for each value of ¯β to find
the maximal possible reward rate RR (see text); Tdel = 3 seconds. In the controlled duration case,
θ = 15.3 is fixed, in agreement with a value indicated by behavioral data (see below and Figure 14).
optimality under this definition as they tend to favor accuracy over speed in two-alternative forced
choice trials (Zacksenhouse et al., 2010). Here, we simply use this quantity to formulate a scalar
performance metric that provides a clear, compact interpretation of reaction time data.
Plotted in Figure 11(A) are multiple accuracy vs. speed curves. The heavy solid line corresponds
to the "baseline" model with robustness and mistuning set to zero (see Figure 4). The lighter solid
line corresponds to the "mistuned" model with σβ = .1. The remaining dashed lines correspond to
the "recovery" model for three different, nonzero levels of the robustness limit R. Also plotted in
the background as dashed lines are RR isoclines -- that is, lines along which RR takes a constant
value, with Tdel = 3 sec. On each accuracy vs. speed curve, there exists a RR-maximizing (RT,
accuracy) pair. This corresponds to a tangency with one RR isocline, and is plotted as a filled
circle. In general, each model achieves maximal RR via a different threshold θ; values are specified
in the legend of Panel (B). (A general treatment of RR-maximizing thresholds for drift-diffusion
models is given in Bogacz et al. (2006).)
In sum, we see that mistuned integrators with a range of increasing robustness limits R achieve
greater RR, as long as their thresholds are adjusted in concert. The optimal values of RR for a range
of robustness limits R are plotted in Figure 11(B). This figure illustrates the fundamental tradeoff
between robustness and sensitivity discussed above. If there is variability in feedback mistuning
(σβ > 0), increasing R can help recover performance. However, beyond at a certain point increasing
R further starts to diminish performance, as too much of the input signal is ignored.
3.4 Biased mistuning towards leak or excitation
We next consider the possibility that variation in mistuning from trial to trial could occur with a
systematic bias in favor of either leak or excitation, and ask whether the robustness limit has quali-
tatively similar effects on decision performance as for the unbiased case studied above. Specifically,
we draw the mistuning parameter β from a gaussian distribution with standard deviation σβ = 0.1
as above, but with various mean values ¯β (see Methods). In Figure 12(A) we show reward rates as
a function of the bias ¯β, for several different levels of the robustness limit R. At each value of ¯β, the
highest reward rate is achieved for a value of R > 0; that is, regardless of the mistuning bias, there
20
Accuracy
RR
Figure 13: Effect of the robustness limit R on decision performance in a controlled duration task,
under the bounded integration model of Kiani et al (2008) (see text). Dot coherence C = 12.8.
(A) Increasing the robustness limit R helps recover performance lost to mistuning at multiple
reaction times in the controlled duration task. Specifically, moving from the "baseline" model
to the "mistuned" model decreases decision accuracy (solid arrow), but this lost accuracy can be
partially or fully recovered for R > 0 (dotted arrow). (B) When allowing for biased mistuning
( ¯β 6= 0, σ = .1), R still allows for recovery of performance; effects are most pronounced when
¯
β > 0.
exists a R > 0 that will improve performance vs. the non-robust case ( R = 0). We note that this
improvement appears minimal for substantially negative mistuning biases, but is more noticable
for the values of ¯β that yield the highest RR. Finally, the ordering of the curves in Figure 12(A)
shows that, for many values of ¯β, this optimal robustness limit is an intermediate value less than
R = 2.
While Figure 12 only assesses performance via a particular performance metric (RR, Tdel = 3
sec.), the analysis in "Reward rate and the robustness-sensitivity tradeoff" suggests that the result
will hold for other performance metrics as well. Moreover, Figure 12(B) demonstrates the analogous
effect for the controlled duration task: for each mistuning bias ¯β, decision accuracy increases over
the range of robustness limits shown.
3.5 Bounded integration as a model of the fixed duration task
We have demonstrated that increasing the robustness limit R can improve performance for mistuned
integrators, in both the reaction time and controlled duration tasks. In the latter, a decision is made
by examining which integrator had accumulated more evidence at the end of the time interval. In
contrast, Kiani et al. (2008) argue that decisions in the controlled duration task are actually made
with a decision threshold (or bound). That is, evidence accumulates toward a bound as in the
reaction time task; if accumulated evidence crosses the bound before the end of the task duration,
the subject simply waits for the opportunity to report the choice, ignoring any further evidence.
Figure 13 demonstrates that our observations about the how the robustness limit can recover
performance lost to mistuned feedback carry over to this model of decision making as well. Specif-
ically, Panel 13(A) shows how setting R > 0 improves performance in a mistuned integrator. In
fact, more of the lost performance (up to 100%) is recovered than in the previous model of the
controlled duration task (cf. Figure 6(A)). Panel 13(B) extends this result to show that some value
of R > 0 will recovers lost performance over a wide range of mistuning biases ¯β (cf. Figure 12(B)).
21
Accuracy
Accuracy
Controlled Duration (sec)
3.6 Compatibility of the robust integrator model with behavioral data
Given the fact that the robustness property can improve decision performance in our model, we
next ask whether robust limits R > 0 are compatible with known behavioral data. To answer this
question, we fit accuracy and chronometric functions from robust integrator models to behavioral
data of Roitman and Shadlen (2002) in the reaction time task. This fit is via least squares across
the range of coherence values, and requires two free parameters: additive noise variance νγ (see
Methods) and the decision bound θ.
Figure 14 shows the results. Panels (A) and (B) display accuracy and chronometric data (dots)
together with fits for various integrator models. First, the solid line gives the fit for the "baseline"
model. The close match between model and data agrees with findings of prior studies (Mazurek et
al., 2003). Next, the dashed and dotted lines give fits for mistuned models (σβ = 0.1), with three
values of bias in feedback mistuning ( ¯β). To obtain these fits, both νγ and θ are changed from
their values for the baseline case. In particular, the noise variance νγ is lowered when feedback is
mistuned. This makes intuitive sense: we have seen in Figure 5 that mistuned feedback worsens
performance for a given signal, so that matching a fixed dataset with a mistuned integrator requires
improving the fidelity of the incoming signal.
Figures 14(C), (D) show analogous results for robust integrators. For all cases in these panels,
we take the robustness limit R = 1.15. We fix levels of additive noise to values found for the
non-robust case above, on order to demonstrate that by adjusting the decision threshold, one can
obtain approximate fits to the same data. This is expected from our results above: Figure 6
shows that, while accuracies at given reaction times are higher for mistuned robust vs. non-robust
models, the effect is modest on the scale of the full range of values traced over an accuracy curve.
Moreover, for the perfectly tuned case, accuracies at given reaction times are very similar for
robust and non-robust integrators (Figure 7, with a slightly lower value of R). Thus, comparable
pairs of accuracy and RT values are achieved for robust and non-robust models, leading to similar
matches with data.
In sum, the accuracy and chronometric functions in Figure 14 show that
all of the models schematized in Figure 4 -- "baseline", "mistuned", "robust", and "recovery" --
are generally compatible with the chronometric and accuracy functions reported in Roitman and
Shadlen (2002).
In order to further test whether empirical data are consistent with the robust integrator model,
we compared simulated reaction time histograms with those found in Roitman and Shadlen (2002).
First, Figure 15 compares the reaction time histograms resulting from the "baseline" model (Panel
A) and the "recovery" model (Panel B). These are plotted in blue; the red histograms are data
are taken from Subject "B".
In both panels, the histograms have similar means, but differ in
their shape; in particular, the model predicts a broader range of reaction times and a more slowly
decaying tail of the RT distribution. From these data, we conclude that neither the "baseline" nor
the "recovery" model quantitatively reproduce the details of reaction time distributions, when the
free parameters θ and σβ are constrained by fitting the accuracy and chronometric functions.
An urgency signal was introduced in Churchland et al. (2008) to better capture behavioral and
physiological data. We next incorporated such a signal into our model to determine whether it
would better align our predicted reaction time histograms with the empirical data. We chose to
implement urgency by assuming a collapsing decision bound, which decreases monotonically from
a peak value of θ0 to a steady state value θss with a halflife t1/2:
θ(t) = θ0 − (θ0 − θss)
t
t + t 1
2
.
(23)
22
(A)
(C)
(B)
(D)
Perfect Tuning
Mistuning
β
∼ δ(0)
∼ N(cid:0)0, .12(cid:1)
∼ N(cid:0)−.05, .12(cid:1)
∼ N(cid:0).05, .12(cid:1)
--
- -
-.-
...
Not Robust ( R = 0) Robust ( R = 1.15)
(cid:0)θ, √νγ(cid:1)
(15.3, 14.0)
(11.0, 12.0)
(9.2, 12.1)
(13.8, 11.5)
(7.6, 14.0)
(6.8, 12.0)
(5.7, 12.1)
(8.6, 11.5)
Figure 14: Accuracy (A,C) and chronometric (B,D) functions: data and model predictions. Solid
dots are behavioral data for rhesus monkeys performing the dot-motion discrimination task (Roit-
man and Shadlen, 2002). In each panel, the accuracy and chronometric functions are identified
with behavioral data via a least-squares fitting procedure over the free parameters θ and νγ. In
Panels (A,B), the robustness threshold R = 0, and results are shown for "baseline" and exemplar
"mistuned" models (see legend in table). In Panels (C,D), R = 1.15, and results are shown for
"robust" and "recovery" models. The close matches to data points indicate that each model can
be broadly reconciled with known psychophysics. Parameter values for each curve are summarized
in the included table.
23
Dot Coherence (C)
Dot Coherence (C)
Reaction Time (ms)
RT (ms.)
Accuracy
Dot Coherence (C)
Dot Coherence (C)
Reaction Time (ms)
RT (ms.)
Accuracy
Figure 15: Reaction time histograms, with decision bounds held constant in time and dot coherence
C=12.8. Histograms for reaction times for our model with both R = 0 (A) and R = 1.15 (B)
are plotted in blue. Overlayed are the reaction time histograms for one subject (Subject "B")
in (Roitman and Shadlen, 2002), in red (semitransparent). Box-and-whisker plots indicate the
quartiles for each data set. Both histograms have identical means, but clearly differ in basic shape
(i.e., the model produces longer tails). We emphasize that this mismatch is a property of our basic
integration to bound model, regardless of the value of the robustness limit R (see text).
Figure 16: Reaction time histograms, with collapsing decision bounds. This shows that introducing
collapsing bounds produces simulated RT histograms that are a closer match to data, for both non-
robust (Panel A, R = 0) and robust (Panel B, R = 1.15) integrators. This shows that RT histograms
can be similarly well matched to data for both cases. Here t 1
2
= 500, θ0 = 25, and θss = 0.
24
Reaction Time (ms)
Reaction Time (ms)
P[DT]
P[DT]
Reaction Time (ms)
Reaction Time (ms)
P[DT]
P[DT]
Figure 16 compares model reaction time histograms produced with the collapsing bound against the
data, and indeed finds a closer fit: qualitatively, the improvement in fit is similar for both the non-
robust ( R = 0) and robust ( R = 1.15) cases. In sum, this shows that the robust integrator model
is capable of producing roughly similar patterns of reaction times compared with those observed
experimentally.
4 Discussion
A wide range of cognitive functions require the brain to process information over time scales that
are at least an order of magnitude greater than values supported by membrane time constants,
synaptic integration, and the like. Integration of evidence in time, as occurs in simple perceptual
decisions, is one such well studied example, whereby evidence bearing on one or another alternative
is gradually accumulated over time. This is formally modeled as a bounded random walk or drift-
diffusion process in which the state (or decision) variable is the accumulated evidence for one
choice and against the alternative(s). Such formal models explain both the speed and accuracy of
a variety of decision-making tasks studied in both humans and nonhuman primates (Ratcliff, 1978;
Luce, 1986; Gold and Shadlen, 2007; Palmer et al., 2005), and neural correlates have been identified
in the firing rates of neurons in the parietal and prefrontal association cortex (Mazurek et al., 2003;
Gold and Shadlen, 2007; Churchland et al., 2008; Shadlen and Newsome, 1996; Schall, 2001; Shadlen
and Newsome, 2001; Kim et al., 2008). The obvious implication is that neurons must somehow
integrate evidence supplied by the visual cortex, but there is mystery as to how.
The reason this is a challenging problem is that the biological building blocks operate on rel-
atively short time scales. From a broad perspective, the challenge is to assemble neural circuits
that that can sustain a stable level of activity (i.e., firing rate) and yet retain the capability to
increase or decrease firing rate when perturbed with new input (e.g., momentary evidence). A well
known solution is to suppose that recurrent excitation might balance perfectly the decay modes
of membranes and synapses (Cannon et al., 1983; Usher and McClelland, 2001). However, this
balance must be fine tuned (Seung, 1996; Seung et al., 2000), or else the signal will either dis-
sipate or grow exponentially (Figure (A), top). Several investigators have proposed biologically
plausible mechanisms that mitigate somewhat the need for such fine tuning (Lisman et al., 1998;
Goldman et al., 2003; Goldman, 2009; Romo et al., 2003; Miller and Wang, 2006; Koulakov et al.,
2002). These are important theoretical advances because they link basic neural mechanism to an
important element of cognition and thus provide grist for experiment.
Although they differ in important details, many of the proposed mechanisms can be depicted
as if operating on a scalloped energy landscape with relatively stable (low energy) values, which
are robust to noise and mistuning in that they require some activation energy to move the system
to a larger or smaller value (Figure (A), bottom; cf. (Pouget and Latham, 2002)). The energy
landscape is a convenient way to view such mechanisms -- which we refer to as robust integrators
-- because it also draws attention to a potential cost. The very same effect that renders a location
on the landscape stable also implies that the mechanism must ignore information in the incoming
signal (i.e., evidence). Here, we have attempted to quantify the costs inherent in this loss. How
much loss is tolerable before the circuit misses substantial information in the input? How much
loss is consistent with known behavior and physiology?
We focused our analyses on a particular well-studied task because it offers critical benchmarks to
assess both the potential costs of robustness to behavior and a gauge of the degree of robustness that
might be required to mimic neurophysiological recordings with neural network models. Moreover,
we know key statistical properties of the signal and noise to be accumulated over time, based on
25
firing properties in area MT.
Our central finding is that ignoring a surprisingly large part of the motion evidence would have
almost negligible impact on performance. Indeed, we found that speed and accuracy are preserved
even when almost a full standard deviation of the input distribution is ignored. We also found that
a similar degree of robustness provides protection of performance against mistuning of recurrent
excitation. Although in general this protection is only partial (Figure 6), for the controlled duration
task it can be nearly complete (Figure 13(A), T > 2) depending on the presence of a decision bound.
We can appreciate the impact of robust integration intuitively by considering the distribution
of random values that would increment the stochastic process of integrated evidence. Instead of
imagining a scalloped energy surface, we simply replace all the small perturbations in integrated
evidence with zeros. Put simply, if a standard integrator would undergo a small step in the positive
or negative direction, a robust integrator instead stays exactly where it was.
In the setting of
drift-diffusion, this is like removing a portion of the distribution of momentary evidence (the part
that lies symmetrically about zero) and replacing the mass with a delta function at 0. At first
glance this appears to be a dramatic effect -- see the illustration of the distributions in Figure 8 --
and it is surprising that it would not result in strong changes in accuracy or reaction time or both.
Three factors appear to mitigate this loss of momentary evidence. First, we showed that setting
weak values of the input signal to zero can reduce both its mean and its standard deviation by a
similar amount, creating compensatory effects that result in a small change to the input signal-
to-noise ratio. Second, we showed that, surprisingly, the small loss of signal to noise that does
occur would not result in any loss of accuracy if the accumulation were to the same bound as
for a standard integrator. The cost would be to decision time, but mainly in the regime that is
dominated by drift -- that is, the shorter decision times -- hence not a large cost overall. Third, even
this slowing is mitigated by the temporal dynamics of the input. Unlike for idealized drift diffusion
processes, real input streams possess finite temporal correlation. Left unchecked, this would imply
greater variability in the integrated signal. Interestingly, removing the weakest momentary inputs
reduces the temporal correlation of the noise component of the input stream. This can be thought
of as allowing more independent samples in a given time period, thereby improving accuracy at a
given response time.
Our robust integrator framework shares features with existing models in sensory discrimination.
The interval of uncertainty model of Smith and Vickers (1989) and the gating model of Purcell et
al. (2010) ignore part of the incoming evidence stream, yet they can explain both behavioral and
neural data. We suspect that the analyses developed here might also reveal favorable properties
of these models. Notably, some early theories of signal detection also featured a threshold, below
which weaker inputs fail to be registered -- so called high threshold theory (reviewed in (Swets,
1961)). The primary difference in the current work is to consider single decisions made based on
an accumulation of many such thresholded samples (or a continuous stream of them).
Although they are presented at a general level, our analyses make testable predictions. For
example, they predict that pulses of motion evidence added to random dot stimulus would affect
decisions in a nonlinear fashion consistent with a soft threshold. Such pulses are known to affect
decisions in a manner consistent with bounded drift diffusion (Huk and Shadlen, 2005) and its
implementation in a recurrent network (Wong et al., 2007). A robust integration mechanism further
predicts that brief, stronger pulses will have greater impact on decision accuracy than longer, weaker
pulses containing the same total evidence.
However, we believe that the most exciting application of our findings will be to cases in
which the strength of evidence changes over time, as expected in almost any natural setting. One
simple example is for task stimuli that have an unpredictable onset time, and whose onset is
not immediately obvious. For example, in the moving dots task, this would correspond to subtle
26
increases in coherence from a baseline of zero coherence. Our preliminary calculations agree with
intuition that robust integrator mechanism will improve performance:
in the period before the
onset of coherence, less baseline noise would be accumulated; after the onset of coherence, the
present results suggest that inputs will be processed with minimal loss to decision performance --
despite the continued ignoring of weak components. This intuition can be generalized to apply to
a number of settings with non-stationary sensory streams.
Many cognitive functions evolve over time scales that are much longer than the perceptual
decisions we consider in this paper. Although we have focused on neural integration, it seems
likely that many other neural mechanisms are also prone to drift and instability. Hence, the need
for robustness may be more general. Yet, it is difficult to see how any mechanism can achieve
robustness without ignoring information. If so, our finding may provide some optimism. Although
we would not propose that ignorance is bliss, in the right measure it may be less costly than one
would expect.
References
Abbott LF, Dayan P (1999) The effect of correlated variability on the accuracy of a population
code. Neural Computation 11:91 -- 101.
Bair W, Zohary E, Newsome WT (2001) Correlated firing in macaque visual area MT: time scales
and relationship to behavior. Journal of Neuroscience 21:1676.
Billingsley P (1986) Probability and measure Wiley-Interscience.
Blackwell H (1953) Psychophysical thresholds; experimental studies of methods of measurement.
Engineering Research Institute, Univ. of Michigan.
Bogacz R, Brown E, Moehlis J, Holmes P, Cohen JD (2006) The physics of optimal decision
making: A formal analysis of models of performance in two-alternative forced-choice tasks. Psy-
chological Review 113:700 -- 765.
Britten KH, Shadlen MN, Newsome WT, Movshon JA (1992) The analysis of visual mo-
tion:
a comparison of neuronal and psychophysical performance. The Journal of Neuro-
science 12:4745 -- 4765.
Britten KH, Shadlen MN, Newsome WT, Movshon JA (1993) Responses of neurons in macaque
MT to stochastic motion signals. Visual Neuroscience 10:1157 -- 1169.
Brown E, Gao J, Holmes P, Bogacz R, Gilzenrat M, Cohen JD (2005) Simple neural networks
that optimize decisions.
International Journal of Bifurcation Chaos in Applied Sciences and
Engineering 15:803 -- 826.
Brown E, Holmes P (2001) Modelling a simple choice task: Stochastic dynamics of mutually
inhibitory neural groups. Stochastics and Dynamics 1:159 -- 191.
Burden R, Faires J (2011) Numerical analysis Brooks Cole.
Cannon SC, Robinson DA, Shamma S (1983) A proposed neural network for the integrator of the
oculomotor system. Biological Cybernetics 49:127 -- 136.
Churchland AK, Kiani R, Shadlen MN (2008) Decision-making with multiple alternatives. Nature
Neuroscience 11:693 -- 702.
27
Cohen MR, Newsome WT (2009) Estimates of the contribution of single neurons to perception
depend on timescale and noise correlation. Journal of Neuroscience 29:6635 -- 6648.
Donner TH, Siegel M, Fries P, Engel AK (2009) Buildup of Choice-Predictive Activity in Human
Motor Cortex during Perceptual Decision Making. Current Biology 19:1581 -- 1585.
Doya K, editor (2007) Bayesian brain: probabilistic approaches to neural coding The MIT Press.
Edwards W (1965) Optimal strategies for seeking information: Models for statistics, choice reac-
tion times, and human information processing. Journal of Mathematical Psychology 2:312 -- 329.
Fuchs AF (1967) Saccadic and smooth pursuit eye movements in the monkey. The Journal of
Physiology 191:609 -- 631.
Gardiner C (2002) Handbook of stochastic methods: for physics, chemistry and the natural sciences
Springer.
Gillespie DT (1996) Exact numerical simulation of the Ornstein-Uhlenbeck process and its integral.
Physical Review E 54:2084 -- 2091.
Gold JI, Shadlen MN (2002) Banburismus and the Brain Decoding the Relationship between
Sensory Stimuli, Decisions, and Reward. Neuron 36:299 -- 308.
Gold JI, Shadlen MN (2007) The neural basis of decision making. Annual Review of Neuro-
science 30:535 -- 574.
Goldman MS (2009) Memory without Feedback in a Neural Network. Neuron 61:621 -- 634.
Goldman MS, Levine JH, Major G, Tank DW, Seung HS (2003) Robust Persistent Neural
Activity in a Model Integrator with Multiple Hysteretic Dendrites per Neuron. Cerebral Cor-
tex 13:1185 -- 1195.
Green D, Swets J (1966) Signal detection theory and psychophysics Peninsula Pub.
Hesse CH (1991) The one-sided barrier problem for an integrated ornstein-uhlenbeck process.
Stochastic Models 7:447 -- 480.
Higham DJ (2001) An algorithmic introduction to numerical simulation of stochastic differential
equations. SIAM Review 43:525 -- 546.
Huk AC, Shadlen MN (2005) Neural Activity in Macaque Parietal Cortex Reflects Temporal
Integration of Visual Motion Signals during Perceptual Decision Making. Journal of Neuro-
science 25:10420 -- 10436.
Kemeny J, Snell J (1960) Finite markov chains D. Van Nostrand.
Kiani R, Hanks TD, Shadlen MN (2008) Bounded integration in parietal cortex underlies de-
cisions even when viewing duration is dictated by the environment. The Journal of Neuro-
science 28:3017 -- 3029.
Kim S, Hwang J, Lee D (2008) Prefrontal Coding of Temporally Discounted Values during In-
tertemporal Choice. Neuron 59:161 -- 172.
Koulakov AA, Raghavachari S, Kepecs A, Lisman JE (2002) Model for a robust neural integrator.
Nature Neuroscience 5:775 -- 782.
28
Laming D (1968) Information theory of choice-reaction times Academic Press New York.
Link SW, Heath RA (1975) A sequential theory of psychological discrimination. Psychome-
trika 40:77 -- 105.
Lisman J, Fellous J, Wang XJ (1998) A role for NMDA-receptor channels in working memory.
Nature Neuroscience .
Luce R (1963) A Threshold Theory for Simple Detection Experiments. Psychological Re-
view 70:61 -- 79.
Luce R (1986) Response times: their role in inferring elementary mental organization Oxford
University Press, Oxford psychology series, no. 8.
Machens CK, Romo R, Brody CD (2005) Flexible control of mutual inhibition: A neural model
of two-interval discrimination. Science 307:1121 -- 1124.
Mazurek ME, Roitman JD, Ditterich J, Shadlen MN (2003) A Role for Neural Integrators in
Perceptual Decision Making. Cerebral Cortex 13:1257 -- 1269.
Mazurek ME, Shadlen MN (2002) Limits to the temporal fidelity of cortical spike rate signals.
Nature Neuroscience 5:463 -- 471.
McKoon G, Ratcliff R (2008) The diffusion decision model: Theory and data for two-choice
decision tasks. Neural Computation 20:873 -- 922.
Miller P (2006) Analysis of spike statistics in neuronal systems with continuous attractors or
multiple, discrete attractor states. Neural Computation 18:1268 -- 1317.
Miller P, Wang XJ (2006) Power-law neuronal fluctuations in a recurrent network model of
parametric working memory. Journal of Neurophysiology 95:1099 -- 1114.
Nikitchenko M, Koulakov A (2008) Neural integrator: A sandpile model. Neural Computa-
tion 20:2379 -- 2417.
Okamoto H, Fukai T (2009) Recurrent network models for perfect temporal integration of fluctu-
ating correlated inputs. PLoS Computational Biology 5:e1000404.
Palmer J, Huk AC, Shadlen MN (2005) The effect of stimulus strength on the speed and accuracy
of a perceptual decision. Journal of Vision 5:376 -- 404.
Pouget A, Latham P (2002) Digitized neural networks: long-term stability from forgetful neurons.
Nature Neuroscience 5:709 -- 710.
Purcell BA, Heitz RP, Cohen JY, Schall JD, Logan GD, Palmeri TJ (2010) Neurally constrained
modeling of perceptual decision making. Psychological Review 117:1113 -- 1143.
Ratcliff R (1978) A Theory of Memory Retrieval. Psychological Review 85:59 -- 108.
Ratcliff R, Rounder JN (1998) Modeling response times for two-choice decisions. Psychological
Science 9:347 -- 356.
Roitman JD, Shadlen MN (2002) Response of neurons in the lateral intraparietal area during a
combined visual discrimination reaction time task. Journal of Neuroscience 22:9475 -- 9489.
29
Romo R, Kepecs A, Brody CD (2003) Basic mechanisms for graded persistent activity: discrete
attractors, continuous attractors, and dynamic representations. Current Opinion In Neurobiol-
ogy 13:204 -- 211.
Salzman CD, Murasugi CM, Britten KH, Newsome WT (1992) Microstimulation in visual area
MT: effects on direction discrimination performance. The Journal of Neuroscience 12:2331 -- 2355.
Schall JD (2001) Neural basis of deciding, choosing and acting. Nature Reviews Neuro-
science 2:33 -- 42.
Seung HS (1996) How the brain keeps the eyes still. Proceedings of the National Academy of
Sciences 93:13339 -- 13344.
Seung HS, Lee DD, Reis BY, Tank DW (2000) Stability of the memory of eye position in a
recurrent network of conductance-based model neurons. Neuron 26:259 -- 271.
Shadlen MN, Britten KH, Newsome WT, Movshon JA (1996) A computational analysis of the
relationship between neuronal and behavioral responses to visual motion. The Journal of Neuro-
science 16:1486 -- 1510.
Shadlen MN, Newsome WT (1996) Motion perception: seeing and deciding. Proceedings of the
National Academy of Sciences 93:628 -- 633.
Shadlen MN, Newsome WT (2001) Neural basis of a perceptual decision in the parietal cortex
(area LIP) of the rhesus monkey. Journal of Neurophysiology 86:1916 -- 1936.
Smith P (2010) From Poisson shot noise to the integrated Ornstein-Uhlenbeck process: Neurally
principled models of information accumulation in decision-making and response time. Journal of
Mathematical Psychology 54:266 -- 283.
Smith PL, Ratcliff R (2004) Psychology and neurobiology of simple decisions. TRENDS in
Neurosciences 27:161 -- 168.
Smith PL, Vickers D (1989) Modeling evidence accumulation with partial loss in expanded judg-
ment. Journal of Experimental Psychology: Human Perception and Performance 15:797.
Swets JA (1961) Is there a sensory threshold. Science 134:168 -- 177.
Usher M, McClelland JL (2001) The time course of perceptual choice: the leaky, competing
accumulator model. Psychological Review 108:550 -- 592.
Wald A (1944) On cumulative sums of random variables. Annals Of Mathematical Statis-
tics 15:342 -- 342.
Wald A (1945) Sequential Tests of Statistical Hypotheses. The Annals of Mathematical Statis-
tics 16:117 -- 186.
Wald A, Wolfowitz J (1948) Optimum character of the sequential probability ratio test. The
Annals of Mathematical Statistics 19:326 -- 339.
Wang XJ (2002) Probabilistic decision making by slow reverberation in cortical circuits. Neu-
ron 36:955 -- 968.
Wang XJ (2008) Decision Making in Recurrent Neuronal Circuits. Neuron 60:215 -- 234.
30
Wong K, Huk A, Shadlen M, Wang XJ (2007) Neural circuit dynamics underlying accumula-
tion of time-varying evidence during perceptual decision making. Frontiers in Computational
Neuroscience 1.
Wong KF, Wang XJ (2006) A recurrent network mechanism of time integration in perceptual
decisions. Journal of Neuroscience 26:1314 -- 1328.
Zacksenhouse M, Bogacz R, Holmes P (2010) Robust versus optimal strategies for two-alternative
forced choice tasks. Journal of Mathematical Psychology 54:230 -- 246.
Zohary E, Shadlen MN, Newsome WT (1994) Correlated neuronal discharge rate and its impli-
cations for psychophysical performance. Nature 370:140 -- 143.
31
Title:
Supplementary Materials for Neural integrators for decision making: A favorable tradeoff between
robustness and sensitivity
Abbreviated Title:
Supplementary Materials for Neural integrators for decision making
Authors:
Nicholas Cain, Department of Applied Mathematics, University of Washington
Andrea K. Barreiro, Department of Applied Mathematics, University of Washington
Michael Shadlen, Department of Physiology and Biophysics, University of Washington
Eric Shea-Brown, Department of Applied Mathematics, University of Washington
Corresponding Author:
Eric Shea-Brown
Department of Applied Mathematics
Box 325420
Seattle, WA 98195-2420
[email protected]
Number of pages: 10
Number of figures: 7
Number of tables: 0
Conflict of Interest: None
Acknowledgements: This research was supported by a Career Award at the Scientific Interface
from the Burroughs-Wellcome Fund (ESB), the Howard Hughes Medical Institute, the National
Eye Institute Grant EY11378, and the National Center for Research Resources Grant RR00166
(MS), by a seed grant from the Northwest Center for Neural Engineering (ESB and MS), and by
NSF Teragrid allocation TG-IBN090004.
arXiv:1111.6573v1 [q-bio.NC] 28 Nov 2011
1
1 Derivation of effective model
Here we we describe a family of neural circuit models, parameterized by a robustness value R, that
produce dynamics similar to that of the central robust integrator model of the main paper (given by
Equation 5 in the main text). In particular, as R → 0, the dynamics reduce to a perfect integrator.
Our construction follows closely that of Koulakov et al. (2002) and Goldman et al. (2003).
The circuits that we will study derive their robustness from multistability, which follows from
recurrent self-excitation among multiple subpopulations (or subunits). We begin with a system of
N differential equations, one for each subunit. Depending on the activity in the rest of the network,
each individual subunit can be become bistable, so that its eventual steady-state value (i.e., "On"
or "Off") depends on its past. This effect, known as hysteresis, underlies several robust integrator
models (Koulakov et al., 2002; Goldman et al., 2003).
The circuit integrates inputs via sequential activation of subunits, in an order determined by
graded levels of "background" inputs (or biases) to each subunit. Following Goldman et al. (2003),
we collapse the N differential equations that describe individual subunits into a equation that
approximates the dynamics of the entire integrator. This expression for the total firing rate E(t)
averaged over all subpopulations reduces to the robust integrator model of the main text (Equation
5 in the main text).
1.1 Firing rate model
The firing rate ri(t) of the ith bistable subunit (i ∈ {1, 2, ..., N}) is modeled by a firing rate equation:
τE
dri
dt
= −ri + f (ri)
(1)
N
X
i6=j
rj + a∆I − bi
f (ri) = r− + (r+ − r−)Hp ∗ ri + q(1 + β)
The parameters and variables in this equation are as follows:
• ri(t): Firing rate of subunit (or pool)
• a∆I(t) : Input signal, with synaptic weight
• p: Within-pool "synaptic" weight
• q: Between-pool "synaptic" weight
• β: Fractional mistuning of feedback connectivity
• N : Number of subunits in the integrator network
• r−: Minimum firing rate of a subunit
• r+: Maximum firing rate of a subunit
• τE : Time constant governing firing rate dynamics
We also define:
• E(t) = 1
NPN
i=1 ri(t): Overall integrator activity
2
(A)
(B)
Figure 1: Simultaneous plots of the identity line and the "feedback line," G(E) for two circuits with
different numbers of subunits. (A) Here N = 1, and so the feedback line G(E) = f (E) is exactly
determined as a function of E = r1 (see Equation (3)). The two intersections c0 and c1 are stable
fixed points. In this way, the subunit's firing rate is bistable, and the value attained will depend on
the history of the circuit activity. As ∆I is changed, this translates f (E), eventually eliminating
either co or c1 and forcing the subunit to the remaining stable fixed point (here, c0 corresponds to
E = 8 Hz. and c1 to 40 Hz). In dashed curves, the feedback line is plotted for two such values
of ∆I. (B) Now N > 1 and so the feedback line G(E) is no longer unambiguously specified as
a function of its argument. The function is instead the sum of N potentially bi-valued functions,
who's actual values will depend on the stimulus history. We represent this fact by plotting the
feedback line as a set of stacked boxes, representing the potential contribution of the ith subunit to
the total integrator dynamics (Goldman et al., 2003).
Figure 1(A) demonstrates the firing rate dynamics of a circuit composed of a single subunit
(N = 1). Here we refer to Equation (1) and plot two curves vs. E = r1. The first is the identity line,
corresponding to the first "decay" term in Equation (1). The second is the "feedback" line, defined
by the second term in Equation (1). Since N = 1, this simplifies to f (r1). The two intersections
marked c0 and c1 are stable fixed points (which we refer to as "On" and "Off" respectively). Thus,
the subunit shown is bistable. Importantly, however, the location of the step in f (E) varies with
changes in the input signal (as per Equation 1).
In particular, substantial values of ∆I(t) will
(perhaps transiently) eliminate one of the fixed points, forcing the subunit into either the "On" or
the "Off" state with ri = r+ or ri = r−, respectively. Moreover, the change is self-reinforcing via
the recurrent excitation pri. The range over which a given subunit displays bistability is affected by
the mistuning parameter β, which scales the total recurrent excitation from the rest of the circuit.
We now derive the dynamics of the overall firing rate E(t). After summing both sides of Equation
(1) over i we have:
where
G(E) = r− +
(r+ − r−)
N
N
τE
dE
dt
= −E + G(E)
H [(p − q(1 + β))ri + N q(1 + β)E + a∆I − bi]
.
(2)
(3)
X
i=1
At this point, we almost have a differential equation for a single variable, E(t). However, Equa-
tion (3) still depends on the N activities ri of the individual subunits, and at any particular time
their values are not uniquely determined by the value of E; we can only bound their values as
r− ≤ ri ≤ r+. We will return to discuss the dynamics of Equation 2 below.
3
(A)
(B)
Figure 2: Plot of possible equilibria for Equation 2. (A) The extent of each multivalued feedback
function defines the minimum input necessary to perturb the system away from equilibrium, defining
the "fixation" lines. As N → ∞, the stable fixed points become more tightly packed on the interval
(r−, r+). (B) When the integrator is mistuned, the fixation lines and the feedback line are no longer
parallel. The rate that the integrator accumulates input is approximated by the distance between
the center line of the feedback subunits ( G(E)), and the feedback line. However, integration only
occurs when the "fixation condition" is no longer satisfied, i.e. when the feedback line is no longer
bounded by the fixation lines.
1.2 Bias term
The bias term for the ith subunit, bi, is set by analyzing the range of values of E for which the
exact value of the feedback function f (ri) is unknown. In the case of an integrator composed of
only a single subunit, we choose the bias term to cause the positive input needed to force the unit
to be on, and the negative input needed to force the unit to be off, to take the same values. This
yields b1 = p(r++r−)
.
2
The general case of N subunits is more complicated. Now the feedback contribution of the ith
unit, f (ri), is no longer a simple function of the population activity E. Instead, it has additional
dependence on its own activity ri. We see this clearly in Equation 3, where the values of ri that
contribute to the definition of G(E) are unspecified. However, we do know that each ri is trapped
between r+ and r−. Therefore, we can plot G(E) as the sum of a sequence of bivalued functions of
E; see Figure 1(B) and (Goldman et al., 2003). The contribution from each pool is then represented
by the shaded region in the plot. Finally, the bias terms are chosen to center these shaded boxes
over the identity line. The correct biases in this general case are:
bi = p(cid:18) r− + r+
2
(cid:19) + q((i − 1)r+ + (N − i)r−) .
(4)
1.3 Fixation lines
We next define the "fixation" lines ( G+(E) and G−(E)), which are the consequence of this multi-
valued property of the integrator. These lines define a region (the fixation region) that runs across
the outermost corners of the "stacked boxes" in Figure 2.
G+(E) = E(1 + β) +
a∆I
N q −
βr+
N
+
p(r+ − r−)
2N q
(5)
4
G−(E) = E(1 + β) +
a∆I
N q −
βr−
N −
p(r+ − r−)
2N q
.
(6)
The term fixation region refers to the following property: if the input ∆I is such that the identity
line lies within the fixation region, then the integrator will possess a range of closely spaced fixed
points (where E = G(E)). Thus, E is not expected to change from its current value; in other
words, integration of ∆I will not occur (Goldman et al., 2003; Koulakov et al., 2002). Recall that
∆I acts to shift these boxes leftward or rightward relative to the identity line, just as in the analysis
of Figure 1. As a consequence, it is weak inputs that fail to be integrated.
From this analysis, we can see that integration by the system as a whole relies on two con-
cepts. The first is a condition on ∆I necessary to eliminate fixed points; we call this the "fixation
condition." The second is the question of how quickly to integrate once this condition is no longer
satisfied. We address this next.
1.4
Integration
Based on the analysis above, we derive a reduced model that approximately captures the dynamics
of the "full" model indicated by Equation 2. We call this the "effective" model. The rate of change
of E -- i.e., the rate of integration -- is given by the distance between the the current value of G(E)
and the identity line. We approximate this by the distance between the middle of the fixation lines,
which we define as G(E), and the identity line. This is pictured in Figure 2(B), and yields:
τE
dE
dt
= −E + G(E) ≈ −E + G(E)
G(E) = (1 + β)E +
a∆I
N q − β
(r+ + r−)
2N
(7)
(8)
We emphasize that integration by this equation only occurs when the "fixation" condition is no
longer satisfied, i.e. when the fixation lines no longer bound the the identity line.
1.5 Fixation condition
The last step in defining the 1-dimensional "effective" model is determining the fixation condition.
We must solve for the values of E that cause the feedback line to lie between the two fixation lines:
No Change in E ⇐⇒ G−(E) < E < G+(E)
⇐⇒ G(E) −
⇐⇒ (cid:12)(cid:12)(cid:12)(cid:12)βE +
(r+ − r−)(p − qβ)
2N q
(cid:12)
(cid:12)
(r+ + r−)
(cid:12)
(cid:12) <
a∆I
N q − β
2N
(r+ − r−)(p − βq)
2N q
< E < G(E) +
(r+ − r−)(p − qβ)
2N q
(9)
(10)
(11)
If this condition is violated with ∆I = 0, the integrator displays runaway integration (β > 0) or
leak (β < 0). If it is satisfied when ∆I = 0, we have a condition on the level of ∆I that must be
present for integration to occur. This yields a piecewise-defined differential equation, corresponding
to when integration can and cannot occur:
5
0
βE + a∆I
N q − β (r++r−)
2N
:(cid:12)(cid:12)(cid:12)βE + a∆I
:
N q − β (r++r−)
otherwise
2N (cid:12)(cid:12)(cid:12) < (r+−r−)(p−βq)
2N q
τE
dE
dt ≈
2N q
We now simplify this equation as follows. First, we assume that p is much larger than βq, so
2N q = R, where R is the robustness parameter in the main text. Next, we
becomes negligible (here the synaptic weights
N q . This
that (r+−r−)(p−βq)
note that as N increases, the additive term β (r++r−)
a and p can adjusted so that the remaining terms do not vanish). Finally, we define κ = a
yields the central Equation 5 of the main text:
≈ (r+−r−)p
2N
τE
dE
dt
=(cid:26) 0
βE + κI
: βE + κ∆I ≤ R
: otherwise
(12)
1.6 Quality of the model reduction
Figure 3 compares the reduced "effective" model, defined by Equation 12 and denoted by the
variable E, and the "full model" of multiple subunits, defined by Equation 1 and denoted by the
variable E. Results are given for three different values of the normalized robustness limit R 1 are
displayed, all in response to the same input signal ∆I(t) for ease of comparison. As R increases,
the ability of the effective model to track the full model decreases. The quality of the reduction
can be quantified by examining the relative error between the full and effective models:
(t) =
E(t) − E(t)
E(t)
(13)
Histograms of (t) evaluated at t = 500 sec. are given in Figure 3(D-F). The agreement of the
two models is within roughly 20% across a range of robustness values R. This is sufficient for
our purposes of demonstrating an approximate connection between the simplified integrator model
used in the main text and one of its many possible neural substrates.
2 Additional indicators that robust integration is compatible with
empirical data from decision tasks
As emphasized in the main text, we make several assumptions about how evidence is integrated
over time in our model of perceptual decision making. The first is that the tuning of connectivity in
the integrator circuit is inherently imprecise from trial to trial. This imprecision in feedback tuning
results in spurious activity, in the form of either exponential growth or decay. To counter this, we
consider the presence of a robustness mechanism meant to ameliorate this problem by allowing the
integrator to fixate at various levels of activity despite mistuning.
These assumptions differ from, e.g., the drift-diffusion model for two-alternative decision mak-
ing, which has been shown to explain psychophysical and physiological data. An important question
is therefore whether the model at hand will also be able to capture these data. In the main text we
find that the answer is yes, for behavioral data on decision accuracy and speed (Figures 14 and 16
in the main text). We next expand this analysis by considering several other aspects of the data:
1This is the robustness parameter divided by the steady-state standard deviation of the input signal; see main
text.
6
Figure 3: Quality of the reduction from the "full" model to the "effective" model. For three different
values of the robustness limit R, traces E(t) for the full model (Equation 1), and its reduction E(t)
(Equation 1.5) are compared (Panels A, B, and C). Here, the same signal is integrated by each
model. Panels D, E, and F show histograms (N=100) of the relative error between the two models
at t = 500 ms. (See Equation 13). The mean µ and standard deviation σ for each distribution
are also indicated. At these levels of R, the effective model gives an approximation of the full
subpopulation-based model with low to moderate error (see text).
7
Integrator
Traces
Relative Error
t (ms)
(A)
(B)
Figure 4: Reaction time histograms from Figures 15 and 16 of the main text, sorted by correct
and incorrect responses and plotted in Panels (A) and (B), respectively. Correct trials are pictured
above, while incorrect trials are below; settings as in the table included in Figure 14 of the main
text. (A) Mean reaction time for error trials is 740 ms., vs. 671 ms.
for correct trials; here a
non-robust integrator is precisely tuned (σβ = 0, and R = 0) (B) For a robust integrator with
imprecise tuning, error trials are now shorter (644 vs. 652 ms.).
error-trial vs. correct-trial reaction time histograms, controlled duration accuracy performance,
decision-triggered stimulus predictions, and traces of neural firing rates vs. time.
2.1 Asymmetry of reaction times between correct and incorrect trials
Roitman and Shadlen (2002) and many other authors report differences in mean reaction times
when responses are sorted by correct vs.
incorrect responses. In Figure 4, we perform the same
analysis on the data in Figure 14 of the main text. First, Panel (A) shows that our model of a
perfectly tuned, non-robust integrator produces the shows the same qualitative asymmetry, with
error trials on average having longer reaction times. With the addition of imprecise tuning and
robustness, however this asymmetry can be eliminated (Panel (B)).
Numerous changes to the model could recover the longer error RTs. One possibility is to
introduce a bias in mistuning parameters: choosing ¯β > 0 has the desired effect. Others include
inclusion of collapsing bounds, urgency signals, and other forms of trial-to-trial variability (for
example Ratcliff and Rounder (1998)).
2.2 Model consistency with controlled duration data
In the main text, we showed that an integrator model with imprecise tuning and robustness can
reproduce the chronometric and psychometric curves reported in Roitman and Shadlen (2002) for
the free-response task. We also considered the consequences of both robustness and mistuning in
a controlled duration task. Here we show that the same parameters (Table, Figure 14 of the main
text), this model can reproduce behavioral data reported in Kiani et al. (2008) for the controlled
duration task.
We assume bounded integration, where an integration threshold determines which alternative
is selected, and the subject waits until the end of the trial to report the selection. The comparison
is demonstrated in Figure 5, where model the accuracy functions from the model are overlaid on
8
500
1000
1500
RT (ms)
2000
P[Correct]
P[Incorrect]
500
1000
1500
RT (ms)
2000
P[Correct]
P[Incorrect]
(A)
(B)
Figure 5: Comparison of model with controlled duration accuracy, C=12.8. The perfectly tuned
model (A), and imprecise/robust model (B), both predict accuracy that is comparable with those
reported in (Kiani et al., 2008) (Dotted lines, overlaid). In both cases, the CD task is modeled
with bounded integration, with threshold set as in the table of Figure 14, main text.
the data provided in the Kiani et al. study, with no parameter changes (including threshold) from
the data reported in the fit to the Roitman et al. reaction time task data in the main text.
2.3 Decision-triggered stimuli
Kiani et al. (2008) also report the time-course of motion evidence leading to the eventual decision
in the controlled duration task. Specifically, the authors study long duration trials with neutral
stimuli (i.e., zero coherence, so neither alternative is "correct"). In our model, motion evidence is
encoded by the stimulus ∆I(t) (see main text). In Figure 6 we sort 400 trials (coherence C=0) based
on which alternative was selected. We average the stimuli, again assuming bounded integration in
a controlled duration task.
We observe that the results of our model, both perfectly tuned and imprecisely tuned with
nonzero robustness, are consistent with those reported by Kiani et al. Specifically, the early separa-
tion of motion energy profiles for rightward (red) and leftward (blue) choices are indicative bounded
accumulator, and do not qualitatively change when mistuning and robustness are included in the
model.
2.4 Trial-sorted traces of firing rates vs. time
Firing rates of LIP neurons reflect the time course of accumulated sensory evidence (Gold and
Shadlen, 2007). We verified that our model displays analogous ramping activity when averaged
over the similar number of trials.
In particular, we checked this both for the perfectly tuned,
non-robust version of our model (Figure 7(A-B)) and in the presence of mistuning and robustness
(Figure 7(C-D)).
Integrator activity averaged across multiple trials is averaged, and aligned to both the onset
9
T (sec)
Accuracy
T (sec)
Accuracy
(A)
(B)
Figure 6: Time course of decision-triggered stimuli. Stimulus realizations ∆I(t) are sorted by "left"
vs "right" alternative selected (red vs. blue; neither alternative is "correct" as coherence C=0), and
averaged across trials. Again assuming perfect tuning, (A), and imprecise tuning with robustness,
(B), we find that both results are consistent with the motion energy profiles reported in Kiani et
al. (2008). All integrator settings are the same as reported in the table in Figure 14 of the main
text.
(A)
(C)
(B)
(D)
Figure 7: Time course of model integrator activity E(t), during stimulus integration in reaction
time tasks. Integration of input signal at various coherence values, with responses aligned at the
beginning (A, C) and end (B, D), averaged across 400 trials. As expected, activity ramps up faster
as dot coherence increases in both the perfectly tuned (A, B) and imprecisely tuned/robust (C, D)
cases, in accordance with physiological observations. (For example, Roitman and Shadlen (2002).)
10
of stimulus and upon reaching threshold in both model setups. Activity ramps up (integrates)
faster at higher dot coherence values, consistent with physiological measurements. This provides
further evidence that our two additional model assumptions are consistent with known physiology
for circuits involved in perceptual decision making.
References
Gold JI, Shadlen MN (2007) The neural basis of decision making. Annual Review of Neuro-
science 30:535 -- 574.
Goldman MS, Levine JH, Major G, Tank DW, Seung HS (2003) Robust Persistent Neural
Activity in a Model Integrator with Multiple Hysteretic Dendrites per Neuron. Cerebral Cor-
tex 13:1185 -- 1195.
Kiani R, Hanks TD, Shadlen MN (2008) Bounded integration in parietal cortex underlies de-
cisions even when viewing duration is dictated by the environment. The Journal of Neuro-
science 28:3017 -- 3029.
Koulakov AA, Raghavachari S, Kepecs A, Lisman JE (2002) Model for a robust neural integrator.
Nature Neuroscience 5:775 -- 782.
Ratcliff R, Rounder JN (1998) Modeling response times for two-choice decisions. Psychological
Science .
Roitman JD, Shadlen MN (2002) Response of neurons in the lateral intraparietal area during a
combined visual discrimination reaction time task. Journal of Neuroscience 22:9475 -- 9489.
11
|
1606.00821 | 2 | 1606 | 2017-01-09T19:42:45 | Towards a statistical mechanics of consciousness: maximization of number of connections is associated with conscious awareness | [
"q-bio.NC"
] | It has been said that complexity lies between order and disorder. In the case of brain activity, and physiology in general, complexity issues are being considered with increased emphasis. We sought to identify features of brain organization that are optimal for sensory processing, and that may guide the emergence of cognition and consciousness, by analysing neurophysiological recordings in conscious and unconscious states. We find a surprisingly simple result: normal wakeful states are characterised by the greatest number of possible configurations of interactions between brain networks, representing highest entropy values. Therefore, the information content is larger in the network associated to conscious states, suggesting that consciousness could be the result of an optimization of information processing. These findings encapsulate three main current theories of cognition, as discussed in the text, and more specifically the conceptualization of consciousness in terms of brain complexity. We hope our study represents the preliminary attempt at finding organising principles of brain function that will help to guide in a more formal sense inquiry into how consciousness arises from the organization of matter. | q-bio.NC | q-bio | Towards a statistical mechanics of consciousness: maximization of
number of connections is associated with conscious awareness
R. Guevara Erra1, D. M. Mateos2, R. Wennberg3, J.L. Perez Velazquez2∗
1 Laboratoire Psychologie de la Perception, CNRS and Université Paris Descartes, Sorbonne Paris Cité, Paris, France.
2 Neuroscience and Mental Health Programme, Division of Neurology, Hospital for Sick Children.
Institute of Medical Science and Department of Paediatrics, University of Toronto, Toronto, Canada.
3 Krembil Neuroscience Centre, Toronto Western Hospital, University of Toronto, Toronto, Canada.
* [email protected]
January 11, 2017
Abstract
It has been said that complexity lies between order and disorder. In the case of brain activity, and
physiology in general, complexity issues are being considered with increased emphasis. We sought to
identify features of brain organization that are optimal for sensory processing, and that may guide the
emergence of cognition and consciousness, by analysing neurophysiological recordings in conscious
and unconscious states. We find a surprisingly simple result: normal wakeful states are characterised
by the greatest number of possible configurations of interactions between brain networks, representing
highest entropy values. Therefore, the information content is larger in the network associated to
conscious states, suggesting that consciousness could be the result of an optimization of information
processing. These findings encapsulate three main current theories of cognition, as discussed in
the text, and more specifically the conceptualization of consciousness in terms of brain complexity.
We hope our study represents the preliminary attempt at finding organising principles of brain
function that will help to guide in a more formal sense inquiry into how consciousness arises from
the organization of matter
1 Introduction
How consciousness arises from the organization of matter is a subject of debate that spans several
disciplines, from philosophy to physics. Multitude of studies focus on the investigation of patterns
of synchrony in brain activity based on magnitudes of a variety of synchrony indices, and thus the
search for organising principles of brain function is today more crucial than ever. We sought to identify
global features of brain organization that are optimal for sensory processing and that may guide the
emergence of conscious awareness. Our results provide a (very simple) answer to the question of what the
magnitudes of synchrony indices represent in terms of the structure of brain activity. Neurophysiological
recordings of brain activity demonstrate fluctuating patterns of cellular interactions, variability that
allows for a wide range of states, or configurations of connections of distributed networks exchanging
information, and support the flexibility needed to process sensory inputs. Recent years have seen a surge
in the study of fluctuations in brain coordinated activity, studies that have raised conceptual frameworks
1
7
1
0
2
n
a
J
9
]
.
C
N
o
i
b
-
q
[
2
v
1
2
8
0
0
.
6
0
6
1
:
v
i
X
r
a
such as that of metastable dynamics [1] and that have motivated interest in the practical application
of assessments of nervous system variability for clinical purposes [2, 3]. A prominent question is how
to describe the organizing principles of this collective activity, which allow features associated with
consciousness to emerge. What is the optimal brain organization that allows it to adequately process
sensory stimuli and enable the organism to adapt to its environment?
Previous studies have revealed values of different indicators of brain coordinated activity, such as
synchronization, associated with healthy and pathologic states by comparison of baseline values and
those in, for instance, unconscious states like coma and epileptic seizures [4, 5, 6]. These observations
prompt the question of what physiological organization underlies the specific values of the synchrony
indices found in normal alert states and other conditions; in other words, what is special about these
values found in conscious states? We believe that we have provided an answer to this question in
our work. We propose that there is a certain general organization of brain cell ensembles that will
be optimal for conscious awareness, that is, for the processing of sensory inputs. As an extension of
previous work [7] where it was proposed that a general organising principle of natural phenomena is
the tendency toward maximal – more probable – distribution of energy/matter, we venture that the
brain organization optimal for conscious awareness will be a manifestation of the tendency towards a
widespread distribution of energy (or, equivalently, maximal information exchange). Whereas we do
not deal with energy or information in our work, we instead focus on the number of (micro)states,
or combinations of connected signals derived from specific types of neurophysiologic recordings. We
use the term "information" in the intuitive sense that normally permeates neuroscience: cell ensembles
that are functionally connected to process/exchange information; furthermore, the equivalence between
information exchange and energy transactions has been the subject of several studies, more specifically
in [7] (see also [8, 9]). The question then becomes: how do we capture the nature of these organizations
of cell interactions?
We have followed the classic approach in physics when it comes to understanding collective be-
haviours of systems composed of a myriad of units: the assessment of the number of possible config-
urations, or microstates, that the system can adopt. In our study we focus on the collective level of
description and assume that coordinated patterns of brain activity evolve due to interactions of meso-
scopic areas ( [10, 11]). Thus we use several types of brain recordings in conscious and unconscious
states, evaluating the number of "connections" between these areas and the associated entropy and
complexity. We present evidence that conscious states result from higher entropy and complexity in
the number of configurations of pairwise connections. The number of pairwise channel combinations is
near the maximum of all possible configurations when the individual is processing sensory inputs in a
normal manner (e.g. with open eyes). Our interpretation is that a greater number of configurations of
interactions allows the brain to optimally process sensory information, fostering the necessary variabil-
ity in brain activity needed to integrate and segregate sensorimotor patterns associated with conscious
awareness.
2 Materials and Methods
Electrophysiological recordings
Recordings were analysed from 9 subjects, using magnetoencephalography (MEG), scalp electroen-
cephalography (EEG) or intracranial EEG (iEEG). Three epilepsy patients were studied with MEG;
2
1 epilepsy patient was studied with iEEG; 3 epilepsy patients were studied with simultaneous iEEG
and scalp EEG; and 2 nonepileptic subjects were studied with scalp EEG. For the study of seizures
versus alert states, the 3 subjects with MEG recordings and the one with iEEG were used. Details of
the patients' epilepsies and seizure types have been presented in previous studies (MEG patients in [5];
iEEG patients in [12]). For the study of sleep versus alert states, the 3 patients with combined iEEG
and scalp EEG have been described previously (patients 1, 3, 4 in [13]); the 2 subjects studied with
scalp EEG alone had been investigated because of a suspected history of epilepsy, but both were ulti-
mately diagnosed with syncope, with no evidence of epilepsy found during prolonged EEG monitoring.
In brief, the MEG seizure recordings were obtained in one patient with primary generalized absence
epilepsy, in one patient with symptomatic generalized epilepsy, and in one patient with frontal lobe
epilepsy. The iEEG seizure recordings were obtained from a patient with medically refractory temporal
lobe epilepsy as part of the patient's routine clinical pre-surgical investigation. MEG recordings were
obtained using a whole head CTF MEG system (Port Coquitlam, BC, Canada) with sensors covering
the entire cerebral cortex, whereas iEEG electrodes were positioned in various locations including, in
the temporal lobe epilepsy patient, the amygdala and hippocampal structures of both temporal lobes.
EEG recordings were obtained using an XLTEK EEG system (Oakville, ON, Canada). The details of
the acquisitions varied from patient to patient (e.g., acquisition rate varied from 200 to 625 Hz) and
were taken into consideration for the data analyses. The duration of the recordings varied as well: for
the seizure study, the MEG sample epochs were of 2 minutes duration each, with total recording times
of 30-40 minutes; the iEEG patient sample was of 55 minutes duration. The sleep study data segments
were each 2-4 minutes in duration, selected from continuous 24-hour recordings.
Data analysis
The only pre-processed data were those of the scalp EEG recordings. These were processed using a
Laplacian derivation [11], to avoid the potential effects of the common reference electrode on synchro-
nization [14] using the DSC algorithm [15]. Initially a phase synchrony index was calculated from all
possible pairwise signal combinations, for which we use the standard procedure of estimating phase dif-
ferences between two signals from the instantaneous phases extracted using the analytic signal concept
via the Hilbert transform. To compute the synchrony index, several central frequencies, as specified
in the text and figure legends, were chosen with a bandpass filter of 2 Hz on either side, hence, for
one value of the central frequency f, the bandpass is f ± 2 Hz. The central frequencies were chosen
according to the relevant behavioural states and some analytical limitations– thus for the sleep studies
we choose 4 Hz (not lower because the extraction of the instantaneous phase was not optimal for central
frequencies lower than 4). To see whether similar results were obtained with different frequencies, we
chose others (see figures) provided there was power at those values. The phase synchrony index (R)
was calculated using a 1-second running window and was obtained from the phase differences using
the mean phase coherence statistic which is a measure of phase locking and is defined as R = hei∆θi
where ∆θ is the phase difference between two signals. This analytical procedure has been described in
great detail elsewhere [4, 5, 14]. The calculation of the index R was done for all possible signal pairs.
The mean value of the R index thus obtained was then estimated for the desired time length. For
instance, for the sleep recordings, the time period was the whole episode, which was, as noted above,
between 2 and 4 minutes. For the seizure recordings, the periods to obtain the mean synchrony varied
depending on the behavioural condition, for instance, in figure 1C the whole ictal event (labelled 'Sz')
3
and the initial 10-second portion of it ('10 sec Sz') were taken for the reason explained in the text
and figure legend. The calculation of the number of connected signals and the entropy associated is
described in the Results section. We note here that to assess entropy we assume that the different
pairwise configurations are equiprobable, thus the entropy is reduced to the logarithm of the number
of states, S = lnC (see notation in the main text). However, the estimation of C (the combinations of
connections between diverse signals), is not feasible due to the large number of sensors; for example, for
35 sensors, the total possible number of pairwise connections is [1442] = 10296, then if we find in the
experiment that, say, 2000 pairs are connected, the computation of [102962000] has too large numbers
for numerical manipulations, as they cannot be represented as conventional floating point values in,
for instance, MATLAB. To overcome this difficulty, we used the well-known Stirling approximation for
large n : ln(n!) = n ln(n)n. The Stirling approximation is frequently used in statistical mechanics
to simplify entropy-related computations. Using this approximation, and after some basic algebra, the
equation for entropy reads, S = N ln(N/N − p) − p ln(p/N − p), where N is the total number of pos-
sible pairs of channels and p the number of connected pairs of signals in each experiment (see Results
for details and notation). Because this equation is derived from the Shannon entropy, it indicates the
information content of the system as well (37).
In addition to entropy, we used another measure of complexity, the Lempel-Ziv (L-Z) complexity,
based in the Kolmogorov deterministic complexity [16]. This complexity measure the amount of non-
redundant information in a string by estimating the minimal size of the "vocabulary" necessary to
describe the string [17]. Strings with high L-Z complexity require a large number of different patterns
("words") to be reproduced, while strings with low complexity can be largely compressed with a few
patterns employed to eliminate redundancy with no loss of information. For this purpose, values of
the matrix B (defined in Results) were placed in a one-dimensional vector and its L-Z complexity
determined.
3 Result
Guided by proposals that consciousness requires medium values of certain features of cell assemblies,
e.g. not too high or low synchrony or correlations [18, 19, 3], or halfway between order and disorder [20],
we chose to quantify the number of possible configurations the brain can adopt in different behavioural
conditions. Our basic approach consists in the estimation of the number of possible pairwise connections
between recorded brain signals. Signals included MEG, iEEG and scalp EEG recordings; details can
be found in Methods. We are limited to pairwise combinations of the signals because of the manner
in which phase synchrony is computed – as phase differences between two signals – and we use phase
synchronization as the means to determine "connectivity" between the two signals. Once the number
of "connected" signals is known, we estimate the entropy of those pairwise combinations. The results
obtained with recordings acquired during conscious states are compared with those acquired during
unconscious states, which included sleep (all stages) and epileptic seizures. To determine connectivity,
we use an accepted approach of computing a phase synchronization index (details in Methods).
It must be noted that, while many studies use the words synchrony and connectivity as synonymous,
in reality phase synchrony analysis reveals only a correlation between the phases of the oscillations
between two signals, and not a real connectivity which depends on several other factors (this matter
has been covered in detail in [21, 22]. Nevertheless, due to the unfeasibility of an accurate, realistic
4
estimation of connectivity which would necessitate individual cell recordings from entire cell ensembles
as well as structural connectivity details, we use an accepted version under the assumption that the
phase relations may represent, at least, some aspect of a functional connectivity. Hence, in order to
evaluate interactions ("connections"), we take each sensor/channel as one "unit", and define a pair
of signals as "connected" if the phase synchrony index is larger than a threshold. The threshold is
determined for each individual, and is the average synchrony index in the most normal alert state,
the 'awake eyes-open' condition, when the individual is fully alert and processing the sensorium in a
regular fashion. Because our data include the three recording methodologies aforementioned, we have
the opportunity to assess the reproducibility of the results in various types of recordings. While we work
at the signal level we will make the reasonable assumption that the MEG and scalp EEG sensors record
cortical activity underlying those sensors and thus throughout the text we will use the terms brain
signals or brain areas/networks as synonymous. The iEEG, obviously, records signals at the source
level. Note that we are not interested in the specific pattern of connectivity among brain sources/areas,
but rather in the global states. These points are further discussed in the Discussion section.
Phase synchronization for specific frequencies (details in Methods) is calculated for each pair of
channels and a "connectivity" matrix S is obtained, whose entries are the average values of the syn-
chrony index during a certain time period for each pairwise configuration. From this matrix, a Boolean
connectivity matrix B is calculated, with 0 entry if the corresponding synchrony index is lower than
a threshold, and 1 if higher. We define two channels as "connected" if the corresponding entry in
matrix B is 1. Then we use the combinations of connected channels as a 'complexity' measure. The
total number of possible pairs of channels given a specific channel montage is given by the binomial
coefficient N = N c!/2! (N c − 2)! where N c is the total number of channels in the recording montage,
normally 144− 146 in the case of MEG sensors, and between 19 and 35 with iEEG and scalp EEG. For
instance, in our MEG recordings we haveN c = 144, thus N = 10, 296possible pairs of connected sensors
are obtained. For each subject we calculate p, the number of connected pairs of signals in the different
behavioural states, using the aforementioned threshold of the synchrony index (which varies for each
subject), and estimate C, the number of possible combinations of those p pairs, using the binomial
coefficient again: C = N!/p!(N − p)! where N is the aforesaid value. In sum, all these calculations
represent the relatively simple combinatorial problem we are trying to solve: given a maximum total of
N pairs of connected signals, in how many ways can our experimental observation of p connected pairs
(that is, the number of 1's in matrix B) be arranged. We then compute the entropy and Lempel-Ziv
complexity associated with those p values.
Figure 1 depicts the entropy (S), which, assuming equiprobable states, is the logarithm of the
number of states, S = ln C (see Methods for the estimation of entropy using very large C values) in
three epileptic patients. We note that this equation using the natural logarithm allows for a calculation
of both the Gibbs and the Shannon entropy, if needed, which differ from S by a constant multiplicative
factor k (the Boltzmann constant) and 1/(ln2) respectively.
In reality the entropy estimation does
not provide any further information, as the main, crucial result is the number of configurations, C.
However, we have done it since it is a standard manner to quantify the "complexity" of the number of
microstates. The entropy data points are graphed on the curve that represents the entropy of all points
in the binomial distribution, where the maximum number of configurations (that is, maximum entropy)
occurs in the middle of the graph. Note in that figure that during conscious states, when patients are
not having generalised seizures with loss of awareness, the entropy is close to the maximum, whereas
entropy is lower (more distant from the top) for the seizure states. The values during the seizures fall
5
on the right-hand side of the graph because, due to the higher synchrony during ictal (seizure) events
such that the number of coupled channels (note that the x-axis is the number of connected signals)
is larger than during interictal (between seizure) activity, there are fewer pairwise configurations and
thus lower entropy. This phenomenon seems associated with the level of consciousness since when the
seizures are not generalised (Figure 1C and 1D), and the patients remain responsive and conscious, the
entropy values are similar to those of interictal (baseline) activity. Where in the curve the data points
are located depends of course on the synchrony index. Because seizures had higher synchrony than
interictal periods ("baseline"), the number of coupled signals is greater and the number of combinations
is lower.
We observe similar trends in the case of sleep. Figure 2 depicts some examples. Note how during
wakefulness the entropy is closer to the maximum of the curve, whereas the deeper the sleep stage,
the more distant from the maximum the values are. The deepest sleep stage, slow wave 3-4 ('sws3-4'),
has consistently the lowest entropy. Interestingly, the entropy during REM sleep is very close, in most
cases, to the normal, alert state. This result may not be as surprising as it sounds if we consider the
mental activity during REM episodes that are normally associated with dreams. It is worth noting
too that in recordings taken when the subjects had their eyes closed, the entropy is much lower than
during the eyes open condition, and sometimes it is as low as during slow-wave 3-4 sleep. The results
for two central frequencies (4 and 8 Hz) are shown in Figure 2A and 2C, to demonstrate that it is not
always the case of high synchrony having lower entropy; sometimes it is lower synchrony (e.g., results
at 8 Hz) that resulted in fewer channel combinations and thus lower entropy as compared to fully alert
states. In Figures 1 and 2 we have shown results with iEEG and MEG (Figure 1) or iEEG and scalp
EEG (Figure 2) recordings to demonstrate that the same qualitative results are obtained with these
three recordings techniques. Thus these results do not depend on recording methodology. Shown in
Figure 3 is an example of the time course of the entropy before and during an ictal event. Hence,
we have demonstrated that the specific values of the synchrony index R in fully alert states represent
the largest number of combinations (microstates, see Discussion) of pairwise signal configurations.
This method solves the potential problem of the interpretation of the different R values obtained with
various recordings techniques. For example, in our experiments, average values in baseline conditions
are 0.248 ± 0.2 for MEG, 0.428 ± 0.04for iEEG, and 0.46 ± 0.05 for scalp EEG, nevertheless, in our
study the number of combinations are "normalised" to the number of recording sensors and show a final
result (the entropy) that is independent of the structure and synchrony magnitudes of the recording
methodology. The main idea derived from these results is represented in Fig. 4.
To further explore whether the decrease in entropy has a parallel with a decrease in other forms
of complexity, the Lempel-Ziv complexity of the number of configurations was assessed. Tables 1 and
2 illustrate the Lempel-Ziv complexity estimated for the B matrices, where a complementary result
can be seen: unconscious states have lower values of complexity. A decrease of complexity in the
raw electrophysiological signals was obtained too (data not shown), hence this may be a phenomenon
observable at different levels of description.
4 Discussion
Our attempts at seeking features of brain organization that allow for adequate processing of sensory
stimuli have provided evidence that a greater number of possible configurations of interactions between
6
brain networks is associated with alert states, representing high entropy associated with the number of
those combinations, whereas lower entropy (and thus fewer combinations of connections) is characteristic
of either unconscious states or altered states of alertness (eyes closed). This observation reflects a
relatively simple general organising principle at this collective level of description, which results in the
emergence of properties associated with consciousness.
With the advent of the 'Big Data' era and the related torrent of empirical observations, the search
for organising principles that result in the emergence of biological phenomena seems more crucial than
ever. We tried to uncover relatively simple laws that capture the bounds in the global organization
of a biological system that enable it to become adaptable (i.e., responsive) to an environment, or, in
neuroscientific terms, the features of optimal brain organization (in terms of connections) that allow
brains to adequately process sensory stimuli. We focused on the global states and did not investigate
specific patterns of connectivity between brain areas as a variety of other studies have assessed these
inter-regional interactions in conscious and unconscious states [23, 24]. The fact that our results are
similar, independent of recording methodology, demonstrates the robustness of the analysis. On that
note, we remark that while we have used the term 'connectivity', in reality the analysis reveals only
correlation between phases of oscillation, as already discussed in Results.
The present study can be considered an extension of previous work where it was proposed that a
general organising principle of natural phenomena is the tendency toward a maximal, or more probable,
distribution of energy [7], which can be encapsulated by the notion of the maximization of information
transfer [25]. As well, the notions of information and energy exchange are conceptually related: "the
common currency paying for all biological information is energy flow" [26, 9].
In the final analysis,
information exchange implies energy exchange, hence we interpret information exchange as energy re-
distribution as proposed in [7], even though our study is focused not on energy considerations but on
the number of states. Other studies have assessed the importance of brain synchronization to optimally
transfer information [27]. We interpret our observation that the number of pairwise channel combina-
tions – that we take as interactions/connections between brain networks – occurs near the maximum
of possible configurations in periods with normal alertness, as that greater number of configurations
of interactions represents the most probable distribution of energy/information resulting in conscious
awareness. The configuration entropy we calculate measures the information content of the functional
network, and has been used in other works for the purpose of quantifying information [28]. One some-
what surprising result is the low entropy during the eyes-close condition. One could argue that having
the eyes closed does not change much the conscious state, and yet we observe reduced entropy as com-
pared with eyes-open condition. Hence our observations may indicate not only states of consciousness
but also optimality of sensory processing – considering the great importance of visual processing in
humans, interrupting visual inputs should result in considerable changes in the dynamics; for instance,
an apparent alteration of brain dynamics upon eye closure is the appearance of alpha waves most clearly
in parieto-occipital regions. Interesting too is the relatively high entropy associated with REM episodes,
perhaps a reflection of the partial awareness during dreams. Perhaps the main difference between dream
and awake states is psychological, but they share similar brain dynamics.
It has been proposed that aspects of awareness emerge when certain levels of complexity are reached
[29]. It is then possible that the organization (complexity) needed for consciousness to arise requires
the maximum number of configurations that allow for a greater variety of interactions between cell
assemblies because this structure leads to optimal segregation and integration of information, two fun-
damental aspects of brain information processing [20] . Our results help cast the study of consciousness
7
and cognition into more of a physics framework that may provide insight into simple principles guiding
the emergence of conscious awareness, and perhaps the proposed thermodynamics for a network of
connected neurons [18] can be extended to explain cognition. Some classical studies [8] have already
characterized biological order in terms of functions of the number of states of a system. It is tempt-
ing to link our observations with the typical chemical equilibrium that, despite being composed of a
myriad of microstates, when viewed at the macroscopic level produces some useful laws, like the law
of mass action describing chemical balance, and thus the perspective we develop here may help guide
research to uncover organising principles in the neurosciences. In physics, microstates that yield the
same macrostate form an ensemble. A system tends to approach the most probable state, maximising
entropy under present constraints, and the resulting macrostate will be represented by the maximum
number of microstates. Hence, the macrostate with higher entropy (see scheme in Fig. 4) we have
defined, composed of many microstates (the possible combinations of connections between diverse net-
works, our C variable defined in Results), can be thought of as an ensemble characterised by the largest
number of configurations. Here we define an ensemble of microstates as all possible configurations of
connectivity leading to the same macrostate (having the same number of connected pair of signals, p).
The entropy of this macrostate, given by the logarithm of the number of combinations (our C), is the
number of microstates that are compatible with the given macrostate (assuming all microstates have
same statistical weight). In neurophysiological terms, each microstate represents a different connectiv-
ity pattern and thus is associated with, in principle, different behaviours or cognitive processes. The
macrostate that we find associated with wakeful normal states (eyes open) is the most probable because
it has the largest entropy (largest number of combinations of connections). While many elementary cel-
lular microscopic processes are far from equilibrium (e.g., ionic gradients), at the macroscopic level the
system tends towards equilibrium, as most natural phenomena remain in near-equilibrium conditions
[30]. At the same time, the ensemble of microstates associated with normal sensory processing features
the most varied configurations and therefore offers the variability needed to optimally process sensory
inputs. In this sense, our results support current views on the metastability of brain states [1] in that
the states should not be too stable for efficient information processing, hence the larger the number of
possible interactions, the more variability is possible.
Equally, the results are consistent with the global workspace theory [31] in that the most widespread
distribution of information leads to conscious awareness. Furthermore, these observations relate as well
to the information integrated theory [32], in that consciousness increases in proportion to the system's
repertoire of states, thus the more combinations possible, the more states available, and here we can de-
fine states as configurations of interactions. Additionally, the results support computational/theoretical
studies showing that patterns of organised activity arise from the maximization of fluctuations in syn-
chrony [33], and that transitions between conscious states are achieved by just varying the probability
of connections in neural nets [34]. In general, our observations highlight the fundamental importance of
fluctuations in neuronal activity as the source of healthy brain dynamics. More specifically, our results
have a precise parallel with the work of Hudetz et al. (2014), where they graph a dispersion index
versus an activation level (their figure 7B) and propose that consciousness resides at the top of the
curve, and anaesthetic states and seizures to lower and higher activation levels respectively, as we show
in Figures 1 and 2. Their 'activation level' could correspond to the number of signals that take part in
the combinations (our x axes in the graphs), and their 'dispersion index' to our number of combinations
(the y axes). Other studies have proposed as well that consciousness requires medium values of certain
features of cell assemblies [32, 35].
8
In sum, along with others, we consider cognition/consciousness not a static property but a dynamic
process with constant flux of energy, or information exchange [25]. Even though we have talked above
about a macrostate, this should not be taken as a fixed state, rather it contains dynamic processes
represented by the microstates, the re-arrangements of connections among brain cell ensembles. The
emergent features of cognitive phenomena that can be termed "conscious" arise once an efficient web of
connections endowed with certain complexity appears. The fact that values of phase synchrony during
fully alert states gave us the largest entropy of the number of pairwise signal combinations explains
in part the neurophysiological organization underlying these specific values of the quantification of
brain synchrony. Studies at this level of description may help to understand how consciousness arises
from organization of matter. In our view, consciousness can be considered as an emergent property
of the organization of the (embodied) nervous system, especially a consequence of the most probable
distribution that maximizes information content of brain functional networks.
On a technical note, we remark that in our analysis we had to choose a threshold to define "connec-
tivity", and therefore a suitable baseline had to be chosen. This procedure is obviously biased in that a
signal corresponding to a baseline has to be ascribed as the one providing the threshold, and we chose
the signals corresponding to the (psychological) state that is most suited for the purposes of adaptabil-
ity: fully alert and receiving all sensory inputs (awake with open eyes). By this choice, and according
to our methods, that signal is already ascribed as having maximum entropy. We tried another less
prejudiced method, using surrogates of the original signals, and then computing the average synchrony
index among the surrogate population (10 phase-randomised surrogates per original channel/signal)
that is the threshold to define connectivity. It turns out that the value of the magnitude of synchrony of
the surrogates is close to the one for the aforementioned baseline chosen, so the results do not vary; nev-
ertheless this new method still assigns the largest entropy to the random signals (surrogates), so there
is still the assumption that the average synchrony of the stochastic signals is a good approximation to
define connections among brain networks. Our current purpose is to completely avoid using thresholds
of synchrony indices, for which we are presently working on a scheme that assigns connectivity in the
time domain.
5 Conclusions and conjectures on the structure of brain-behaviour-
environment
It is tempting to speculate, based on these results and the conclusions of a previous study [7] that there
could be a universal logic ruling the evolution of natural phenomena - biological and nonbiological-
and the nervous system in particular: patterns emerge from a central theme captured by maximising
information exchange. Because, in the final analysis, all exchange of information implies exchange
of energy, natural phenomena tend towards the most probable distribution of energy, and thus the
interactions among system constituents tends to be maximized.
Because the brain functions to maintain a predictive model of the environment (the reason the
brain evolved is to model the environment, after all), then perhaps the brain's global configuration
has to "copy" what is out there: and out there energy distributes in all possible microstates (second
principle of thermodynamics). Then to process such variability in nature, the nervous system should
have same structure, and the result is the 'inverted U' that has appeared in our analysis and has
been theoretically proposed in other publications [20], the top of the curve representing more possible
9
combinations to handle information/energy exchanges. On the other hand, in the extremes of this
curve we find fewer microstates, thus these are not optimal situations to process the many microstates
in the environment. The key then is not to reach the maximum number of units interacting (which
would be all-to-all connections and thus only one possible microstate), but rather the largest possible
number of configurations allowed by the constraints.
In a similar fashion, it has been argued that
the brain needs to show criticality because natural phenomena possess critical dynamics [36]. Then,
perchance, consciousness can be considered as an emergent property of the organization of the embodied
nervous system submerged in an environment, consequence of the most probable distribution of energy
(information exchange) in the brain. In this regard, consciousness (like biochemistry) may represent
thus an optimal channel for accessing sources of (free) energy.
10
References
[1] JA. S. Kelso. Dynamic patterns: The self-organization of brain and behavior. MIT press, 1997.
[2] D. D Garrett, G. R. Samanez-Larkin, S. WS. MacDonald, U. Lindenberger, A. R McIntosh, and
C. L Grady. Moment-to-moment brain signal variability: a next frontier in human brain mapping?
Neuroscience & Biobehavioral Reviews, 37(4):610–624, 2013.
[3] V. Nenadovic, J. L. Perez Velazquez, and J. S. Hutchison. Phase synchronization in electroen-
cephalographic recordings prognosticates outcome in paediatric coma. PloS one, 9(4):e94942, 2014.
[4] F. Mormann, K. Lehnertz, P. David, and C. E. Elger. Mean phase coherence as a measure for
phase synchronization and its application to the eeg of epilepsy patients. Physica D: Nonlinear
Phenomena, 144(3):358–369, 2000.
[5] L.. Garcia Dominguez, R. A. Wennberg, W. Gaetz, D. Cheyne, O. C. Snead, and J. L. Perez Ve-
lazquez. Enhanced synchrony in epileptiform activity? local versus distant phase synchronization
in generalized seizures. The Journal of neuroscience, 25(35):8077–8084, 2005.
[6] D. C. Shields, J. W. Leiphart, D. L. McArthur, P. M. Vespa, M. L. Quyen, J. Martinerie, and J. R.
Soss. Cortical synchrony changes detected by scalp electrode electroencephalograph as traumatic
brain injury patients emerge from coma. Surgical neurology, 67(4):354–359, 2007.
[7] J. L. Perez Velazquez. Finding simplicity in complexity: general principles of biological and non-
biological organization. Journal of biological physics, 35(3):209–221, 2009.
[8] H. J. Morowitz. Some order-disorder considerations in living systems. The bulletin of mathematical
biophysics, 17(2):81–86, 1955.
[9] E. Smith. Thermodynamics of natural selection i: Energy flow and the limits on organization.
Journal of theoretical biology, 252(2):185–197, 2008.
[10] J.J. Wright and D.T.J. Liley. Dynamics of the brain at global and microscopic scales: Neural
networks and the eeg. Behavioral and Brain Sciences, 19(02):285–295, 1996.
[11] P. L. Nunez. Toward a quantitative description of large-scale neocortical dynamic function and
eeg. Behavioral and Brain Sciences, 23(03):371–398, 2000.
[12] J. L. Perez Velazquez, Garcia Dominguez L., V. Nenadovic, and R. A. Wennberg. Experimen-
tal observation of increased fluctuations in an order parameter before epochs of extended brain
synchronization. Journal of biological physics, 37(1):141–152, 2011.
[13] R. Wennberg. Intracranial cortical localization of the human k-complex. Clinical Neurophysiology,
121(8):1176–1186, 2010.
[14] R. Guevara, J. L. Pérez Velazquez, V. Nenadovic, R. Wennberg, G Senjanović, and L. G.
Dominguez. Phase synchronization measurements using electroencephalographic recordings. Neu-
roinformatics, 3(4):301–313, 2005.
11
[15] J. Kayser and C. E. Tenke. Principal components analysis of laplacian waveforms as a generic
method for identifying erp generator patterns: I. evaluation with auditory oddball tasks. Clinical
neurophysiology, 117(2):348–368, 2006.
[16] A. N. Kolmogorov. Three approaches to the quantitative definition of information. Problems of
information transmission, 1(1):1–7, 1965.
[17] A. Lempel and J. Ziv. On the complexity of finite sequences. Information Theory, IEEE Transac-
tions on, 22(1):75–81, 1976.
[18] G. Tkacik, T. Mora, O. Marre, D. Amodei, II Berry, J. Michael, and W. Bialek. Thermodynamics
for a network of neurons: Signatures of criticality. arXiv preprint arXiv:1407.5946, 2014.
[19] V Nenadovic, J. S. Hutchison, L. G. Dominguez, H. Otsubo, M. P. Gray, R Sharma, J. Belkas, and
J. L. Perez Velazquez. Fluctuations in cortical synchronization in pediatric traumatic brain injury.
Journal of neurotrauma, 25(6):615–627, 2008.
[20] G. Tononi, G. M. Edelman, and O. Sporns. Complexity and coherency: integrating information in
the brain. Trends in cognitive sciences, 2(12):474–484, 1998.
[21] J. L. Perez Velazquez, R. G. Erra, R. Wennberg, and L. G. Dominguez. Correlations of cellular
activities in the nervous system: physiological and methodological considerations. In Coordinated
Activity in the Brain, pages 1–24. Springer, 2009.
[22] J. L. Perez Velazquez and M. Frantseva. The Brain-Behavior Continuum: The Subtle Transition
Between Sanity and Insanity. World Scientific, 2011.
[23] G Dumermuth and D Lehmann. Eeg power and coherence during non-rem and rem phases in
humans in all-night sleep analyses. European neurology, 20(6):429–434, 1981.
[24] K. Wang, M. L. Steyn-Ross, D.A. Steyn-Ross, M. T. Wilson, and J. W. Sleigh. Eeg slow-wave
coherence changes in propofol-induced general anesthesia: experiment and theory. Frontiers in
systems neuroscience, 8, 2014.
[25] Hermann Haken. Information and self-organization: A macroscopic approach to complex systems.
Springer Science & Business Media, 2006.
[26] H. J Morowitz. Energy flow in biology; biological organization as a problem in thermal physics.
1968.
[27] A Buehlmann and G Deco. Optimal information transfer in the cortex through synchronization.
PLoS Comput Biol, 6(9):e1000934, 2010.
[28] W. Bialek, F. Rieke, R de Ruyter van Steveninck, and D Warland. Spikes: exploring the neural
code. MIT. Roddey, JC, Girish, B., & Miller, JP (2000). Assessing the performance of neural
encoding models in the presence of noise. Journal of Computational Neuroscience, 8(95):112, 1997.
[29] M. Gell-Mann. Consciousness, reduction, and emergence. Annals of the New York Academy of
Sciences, 929(1):41–49, 2001.
12
[30] I. Prigogine. Introduction to thermodynamics of irreversible processes. New York: Interscience,
1967, 3rd ed., 1, 1967.
[31] B. J. Baars. A cognitive theory of consciousness. Cambridge University Press, 1993.
[32] G. Tononi. An information integration theory of consciousness. BMC neuroscience, 5(1):42, 2004.
[33] V. Vuksanović and P. Hövel. Dynamic changes in network synchrony reveal resting-state functional
networks. Chaos: An Interdisciplinary Journal of Nonlinear Science, 25(2):023116, 2015.
[34] D. W. Zhou, D. D. Mowrey, P. Tang, and Y. Xu. Percolation model of sensory transmission and
loss of consciousness under general anesthesia. Physical review letters, 115(10):108103, 2015.
[35] A. A Fingelkurts, A. A. Fingelkurts, S. Bagnato, C. Boccagni, and G. Galardi. Do we need a
theory-based assessment of consciousness in the field of disorders of consciousness? Frontiers in
human neuroscience, 8:402, 2014.
[36] D. R. Chialvo. Emergent complex neural dynamics. Nature physics, 6(10):744–750, 2010.
13
State
Subject #1 Baseline
Seizure
Subject #2 Baseline
Seizure
Subject #3 Baseline
Seizure
Subject #4 Baseline
Seizure
L-Z complexity
0.7
0.2
0.61
0.39
0.86
0.71
0.61
0.61
Table 1: Values of Lempel-Ziv (L-Z) complexity derived from the string of connections (details in Meth-
ods) in conscious (baseline) and unconscious (seizure) states, in four patients. Note lower complexity
during seizures in patients # 1 − 3; patient # 4 (Figure 1D) did not have fully generalised seizures.
Subject #1 Alert eyes open
State
SWS 2
SWS 3-4
REM
Subject #2 Alert eyes open
Alert eyes close
SWS 3-4
Subject #3 Alert eyes open
Subject #4 Alert eyes open
Alert eyes closed
SWS 2
SWS 3
REM
Subject #5 Alert eyes open
iEEG L-Z complexity
0.81
0.42
0.018
0.73
0.94
0.94
0.33
N/A
N/A
N/A
N/A
0.88
0.55
0.57
0.05
0.6
N/A
N/A
N/A
N/A
N/A
scalp L-Z complexity
1.01
1.0
0.0
0.96
1.06
0.85
0.22
1.07
1.07
1.07
0.97
1.0
1.0
9.8
0.0
1.06
0.97
1.08
1.04
0.8
1.12
SWS 2
SWS 3-4
REM
SWS 1
SWS 2
SWS 3-4
REM
Table 2: Values of Lempel-Ziv (L-Z) complexity derived from the string of connections in different sleep
stages. When both recordings were obtained from a subject, both iEEG and scalp EEG data were
analysed (subjects # 3 and # 5 did not have iEEG recordings). The L-Z complexity is consistently
lower, regardless of recording methodology, in the deepest sleep stage (SWS 3-4). These results parallel
those of entropy estimations shown in figure 2
14
Figure 1: Graphs representing the entropy of the number of pairwise configurations of signals in epileptic
patients during conscious (baseline) and unconscious (generalised seizure) states. A, derived from MEG
recordings in a patient with primary generalised epilepsy, shows entropy associated with a normal alert
period (baseline, 'Base') and a generalised absence seizure ('Sz'), estimated from synchrony values at two
central frequencies (defined in Methods) of 5 and 12 Hz. The curve in this and other graphs here and in
Figure 2 represents the possible entropy values of all possible numbers of pairwise combinations, yielding
an inverted U or, for very large numbers, a Gaussian. Note that here as well as in all generalised seizures
analysed, the entropy values associated with alert, baseline conditions were closer to the maximum (top
of the curve) than those associated with the seizures. B, entropy values of two seizures and their
corresponding baseline ('Base') activity (computed using a time period of 30-40 minutes before the
ictus) in a patient with secondary (symptomatic) generalised epilepsy (MEG recordings). C, derived
from iEEG recordings in a patient with temporal lobe epilepsy, shows the entropy during the alert
state ('baseline'), during the initial 10 seconds of the seizure when the patient was still responsive and
alert ('10 sec Sz'), and during the rest of the seizure when it became generalised and the patient was
unresponsive ('Sz'). Note that when the ictus has not yet generalised, the entropy is similar to that of
normal alertness. D, another example of a non-generalised seizure in a patient with frontal lobe epilepsy
(MEG recordings).
15
2000400060008000100001000300050007000BaselineSz2000400060008000100001000300050007000S (entropy)Bae zs 5HSz @ 5 HzBase@ 12 HzSz @ 12 HzABase #2Sz # 12000400060008000100001000300050007000Base #1Sz # 2B10 sec Sz5101520253035510152025BaselineSzCDS (entropy)Number of coupled channels Same graph types as in Figure 1, using sleep recordings.
Figure 2:
In each subject, data samples
were of 2-4 minutes duration during wakefulness with eyes open ('awake') or eyes closed ('Eyecl'), and
sleep stages slow-wave 2 ('Sws2'), slow-wave 3-4 ('Sws3-4') and rapid eye movement ('REM'). A, results
derived from iEEG recordings in a subject investigated with bilateral frontal and temporal electrodes
and simultaneous scalp EEG. Entropy estimated from synchrony values at two central frequencies of 4
and 8 Hz. As occurred in the patient recordings shown in Figure 1, the baseline, alert state (in this case
labelled 'awake') is closer to the top of the curve, having greater entropy. The deepest sleep stage, slow
wave 3-4 ('sws3-4'), has the lowest entropy. B, same subject but using the scalp EEG recordings for the
calculations, showing similar trend (evaluated at central frequency of 4 Hz). C, results derived from a
different subject investigated with right frontal electrodes and simultaneous scalp EEG, with synchrony
evaluated at two central frequencies of 4 and 8 Hz using the iEEG signals. Note that the eyes closed
('Eyecl') condition has lower entropy than that of the normal alert state with open eyes. Depending on
the frequency of analysis, the entropy in 'Eyecl' falls toward the left or right side of the curve; e.g., at
8 Hz, because the synchrony is higher (more coupled channels) due to alpha waves at 8-10 Hz and the
entropy is reduced due to fewer combinations of connections, as occurs similarly during seizures (Figure
1). D, same subject but using the scalp EEG signals, showing similar results to those obtained with
iEEG (evaluated at central frequency of 4 Hz).
16
200400600800100012000100200300400500600700800S (entropy)awakeSws2Sws3-4REMawakeSws2Sws3-4REM4 Hz8 Hz*AawakeSws2Sws3-45010015020025030050100150200250REMBREM10020030040050060050100150200250300350400awakeEyeclSws2Sws3-4awakeEyeclREMSws2Sws3-44 Hz8 Hz*CREM5010015020025030050100150200250awakeEyeclSws2Sws3-4DS (entropy)Number of coupled channels Figure 3: Time course of the entropy of the number of configurations of connected MEG signals before,
during and after a generalised absence seizure. MEG signal from one channel is shown at top, the ictus
('Sz') occurs towards the end of the 2 minute recording. Notice the drop in entropy during the seizure.
17
SzTime (sec)600040002000329664168048112S (entropy) Figure 4: Proposed general scheme of the relation between global brain connectivity and behavioural
states. Normal alertness resides at the top of the curve representing the number of configurations of
connections the system can adopt, or the associated entropy. The maximisation of the configurations
(microstates) provides the variability in brain activity needed for normal sensorimotor action. Abnor-
mal, or unconscious states, are located farther from the top, and are characterised by either large or
small number of "connected" networks therefore exhibiting lower number of microstates (hence lower
entropy) that are not optimal for sensorimotor processing.
18
number networks "connected"conscious/alertunconsciousEntropy (S) or number of configurations low Shigh SÞnumber microstates |
1804.01906 | 3 | 1804 | 2018-05-23T12:40:58 | An Accelerated LIF Neuronal Network Array for a Large Scale Mixed-Signal Neuromorphic Architecture | [
"q-bio.NC",
"cs.ET",
"physics.bio-ph",
"physics.comp-ph"
] | We present an array of leaky integrate-and-fire (LIF) neuron circuits designed for the second-generation BrainScaleS mixed-signal 65-nm CMOS neuromorphic hardware. The neuronal array is embedded in the analog network core of a scaled-down prototype HICANN-DLS chip. Designed as continuous-time circuits, the neurons are highly tunable and reconfigurable elements with accelerated dynamics. Each neuron integrates input current from a multitude of incoming synapses and evokes a digital spike event output. The circuit offers a wide tuning range for synaptic and membrane time constants, as well as for refractory periods to cover a number of computational models. We elucidate our design methodology, underlying circuit design, calibration and measurement results from individual sub-circuits across multiple dies. The circuit dynamics match with the behavior of the LIF mathematical model. We further demonstrate a winner-take-all network on the prototype chip as a typical element of cortical processing. | q-bio.NC | q-bio |
1
An Accelerated LIF Neuronal Network Array for
a Large Scale Mixed-Signal Neuromorphic
Architecture
Syed Ahmed Aamir∗, Student Member, IEEE, Yannik Stradmann∗, Paul Muller, Christian Pehle, Andreas Hartel,
Andreas Grubl, Johannes Schemmel, Member, IEEE, and Karlheinz Meier, Member, IEEE
Abstract-We present an array of leaky integrate-and-fire (LIF)
neuron circuits designed for the second-generation BrainScaleS
mixed-signal 65-nm CMOS neuromorphic hardware. The neu-
ronal array is embedded in the analog network core of a scaled-
down prototype HICANN-DLS chip. Designed as continuous-
time circuits, the neurons are highly tunable and reconfigurable
elements with accelerated dynamics. Each neuron integrates input
current from a multitude of incoming synapses and evokes a
digital spike event output. The circuit offers a wide tuning
range for synaptic and membrane time constants, as well as
for refractory periods to cover a number of computational
models. We elucidate our design methodology, underlying circuit
design, calibration and measurement results from individual sub-
circuits across multiple dies. The circuit dynamics match with the
behavior of the LIF mathematical model. We further demonstrate
a winner-take-all network on the prototype chip as a typical
element of cortical processing.
Keywords-Analog integrated circuits, Neuromorphic, Leaky
Integrate and Fire, 65nm CMOS, Spiking neuron, OTA, Opamp,
Tunable resistor, Winner-take-all network
I .
I N T R O D U C T I O N
T HE architecture of digital microprocessors is fundamen-
tally different from that of the central nervous system.
While the brain is a massively parallel structure of neurons in-
terconnected through synapses [1], microprocessors are mostly
based on a von Neumann architecture [2], [3] with logic gates
as the elementary primitives. The human brain consumes only
approximately 20 W [4], while its performance as a general-
purpose problem solver is still unmatched by any computer
algorithm.
Taking inspiration from this biological feat, neuromorphic
architectures not only adopt a non-von Neumann architecture
by collocating memory close to the computational element,
but also introduce massive parallelism, high energy efficiency,
reconfigurability, fault tolerance, and integrate computational
*Both authors contributed equally to this work.
Manuscript submitted August 29, 2017, revised January 20, 2018; April 4,
2018; May 14, 2018, accepted May 21, 2018. This work has received funding
from the European Union Seventh Framework Programme ([FP7/2007-2013])
under grant agreement nos. 604102 (HBP), 269921 (BrainScaleS), the Horizon
2020 Framework Programme ([H2020/2014-2020]) under grant agreement no.
720270 (HBP) as well as from the Manfred Stark Foundation.
All authors are with the Kirchhoff Institute for Physics, Heidelberg Univer-
sity, D-69120 Heidelberg, Germany.
Email: {aamir,yannik.stradmann}@kip.uni-heidelberg.de.
models of neural elements using CMOS technologies [5]–
[7]. They compute at biological timescales or at a specified
speed-up ("acceleration") factor. In particular, analog imple-
mentations integrate bio-physically inspired neuron models as
computational elements in order to capture the rich temporal
dynamics of the neuronal membrane.
Since their emergence in the late eighties [5], neuromorphic
circuits and systems have evolved with a variety of large-scale
architectures, model implementations and technologies [4], [8].
For example, [9]–[14] describe systems that adhere to the ini-
tial idea of exploiting weak-inversion characteristics of CMOS
circuits to emulate ion flow dynamics. More recently, with the
advancements in cognitive computing driven in part by the
speech and object recognition made possible by deep neural
networks, major industrial players [15], [16] explore neuro-
morphic architectures and integrate digital phenomenological
neuron models in ultra deep-submicron technologies - the
most recent being Intel's 14 nm FinFET Loihi chip [16].
Within EU's framework for brain-inspired computing [17],
[18], the SpiNNaker and BrainScaleS systems present large-
scale neuromorphic platforms. While SpiNNaker [19] uses
arrays of many-core interconnected ARM-based microproces-
sor nodes and implements a software-defined neuron model,
our approach as described within the BrainScaleS hardware
system [20], [21] is a wafer-scale implementation with an
analog physical neuron model and accelerated-time operation.
The system utilizes an entire post-processed CMOS wafer for
large scale integration by interconnecting multiple identical on-
wafer chips [21].
In the second-generation BrainScaleS architecture [22], [23],
we present, for the first time, a mixed-signal neuromorphic
computing core that integrates a custom Single Instruction
Multiple Data (SIMD) processor with the Analog Network
Core (ANC) (comprised of the neuron array and synapse
matrix) in 65 nm CMOS. We have previously described the
plasticity subsystem of the chip [23], [24]. Here we present
the first silicon implementation and characterization of the
neuronal array [25] based on the Leaky Integrate and Fire (LIF)
neuron model. We demonstrate its high tunability, reconfig-
urability, reliable parameter mapping through calibration and
the network operation on a scaled-down prototype chip. The
neuron array is tested in tight integration with the synapses and
dedicated analog bias storage [26], [27] providing voltage and
current biases for tuning the neuron's behavior. The presented
chip features a decreased neuronal and synaptic count (32
©2018 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works.
neurons interconnected with 32×32 synapses instead of 512
neurons and 256×512 synapses), but integrates all essential
elements of the scaled-up version (see Sec. II). We demonstrate
a winner-take-all network as an example application. The
detailed circuit description of synapses, the analog bias storage
and the plasticity mechanism of the chip is omitted in this
work.
In the subsequent sections, we describe the chip architecture
of the HICANN-DLS prototype, as well as the design pre-
considerations (Sec. II & Sec. III). The design, measurement,
calibration and performance results of the neuron as well
as of the individual sub-circuits are described in Sec. IV.
A demonstration of a winner-take-all network is detailed in
Sec. V. We discuss and compare our results with other large-
scale neuromorphic systems in Sec. VI before concluding the
paper in Sec. VII.
I I . T H E H I C A N N - D L S C H I P
The BrainScaleS wafer-scale system [21], [28] is built in the
first phase with 180 nm CMOS HICANN neuromorphic chips,
operated 104–105 times faster than biological timescale. The
second-generation hardware features the enhanced High Input
Count Analog Neural Network with Digital Learning System
(HICANN-DLS) chips as the fundamental building block [22].
HICANN-DLS is a 65 nm CMOS mixed-signal system on-chip
solution with integrated analog and digital cores and operated
one thousand times faster than biological real-time. The chip
is under development and an enhanced scaled-up version of
the current prototype will replace the HICANN chip in the
second-generation wafer-scale platform.
A simplified architecture of the initial chip prototype con-
nected to its off-chip measurement system is sketched in
Fig. 1a. The left half shows the ANC arranged in a columnar
architecture of the edge-connected neuron array, the synapse
matrix as well as the analog on-chip Capacitive Memory
(Capmem) cells for the storage of tunable neuron voltage and
current biases. The right half shows the digital part comprising
of the SIMD plasticity processor and digital control logic. In
this prototype version, the ANC contains 32 columns, each
containing 32 synapse rows relaying current pulse events to
a single neuron per column. Each neuron is configured by
14 current and 4 voltage biases, all of which are individually
tunable per neuron as well as a 15-bit digital configuration bus.
One global voltage bias is common to all neurons and provided
by an additional global Capmem column, see Fig. 1a.
The Capmem voltage and current cells have 10-bit resolution
and provide the neuron with tunable biases of 15 nA to 1 µA
(current cells) and 200 mV to 1.8 V (voltage cells). The
average drift of the capacitive storage is less than 1 LSB/20 ms
and the cells are periodically refreshed at a rate of 1–2 kHz
to minimize its effect [27]. The columnar arrangement of neu-
rons, synapses and dedicated tunable biases make the system
a highly configurable non-von Neumann architecture. Data
transfer between the digital backend and various components
of ANC is carried out
in the form of data packets via
the On-chip Multi-master Non-bursting Interface Bus-fabric
(OMNIBUS) [24] (blue connections in Fig. 1a). The packets
2
encode synaptic addresses, synapse matrix's row enables and
the digital configuration of neurons and the Capmem. During
network operation, incoming events reach the synapses via
OMNIBUS containing target synapse addresses and row-wise
enables. If a packet's synapse address matches the target
synapse address, the target enables a 6-bit current-mode Digital
to Analog Converter (DAC) for a short duration, essentially
relaying a current pulse event to the neuron on either of the
two dedicated lines per neuron column (2×32 in total) - one
for excitatory events, another for inhibitory events. The digital
input code of the 6-bit DAC inside each synapse corresponds to
the synaptic weight. This is further highlighted in the left half
of Fig. 1b. The digital output events evoked by each neuron
are routed off-chip to the FPGA via a Serializer/Deserializer
(SerDes), from where they are re-routed back into the synapse
array. To provide a debug interface, the entire neuron array is
connected to the chip pads by two dedicated pins, labeled Istim
and VreadOut (described in Sec. IV-D).
The system performs hybrid learning using parallel ana-
log processing in the synapses together with arbitrarily pro-
grammable learning rules in the custom SIMD plasticity pro-
cessor [23]. The synapses feature a dedicated implementation
of Spike-timing-dependent plasticity (STDP) [29], where they
measure the temporal correlation of pre– and post-synaptic
events. The post-synaptic events from all neurons are therefore
routed to the synapses. To facilitate STDP, the temporal corre-
lation of input/output events (pre-post or post-pre) is stored as
analog voltages inside each synapse. These analog voltages are
digitized by an Analog to Digital Converters (ADCs) before
being read by the plasticity processor [23]. The processor
can then alter the the synaptic weights (stored in SRAM)
in accordance with the implemented rule, thereby performing
online learning.
The processor is a 32-bit implementation based on power
instruction-set architecture with a custom SIMD unit. It pro-
cesses 128-bit wide vectors of either eight or sixteen elements
using vector slices. The current implementation features a sin-
gle slice that processes 16 columns in parallel and iterates twice
for 32 columns. All synapse rows are processed sequentially.
The processor can be used for implementing arbitrary learning
rules that make use of the available observables and modify
the structure and parameters of a running network. It can
additionally be used to perform local calibration algorithms
that can be executed in parallel on large multi-chip systems.
Within the described architecture, the total area of all ADC
channels, the neurons and the number of vector slices in the
processor scale linearly with the number of columns.
I I I . D E S I G N M E T H O D O L O G Y
When it comes to neuron design, every large scale system
adheres to certain parameters governed mainly by the system
specifications and target applications. The design phase of this
accelerated neuron circuit in general took the following pre-
considerations:
A. Neuron Model
Our choice of neuron model comes from the trade-off
between a design that encapsulates most of the biological
3
Fig. 1: (a) Architecture of the HICANN-DLS prototype chip and the measurement system. (b) The full circuit schematic of a
single integrated neuron. (c) The schematic of the resistor used inside the synaptic input. (d) The architecture of the excitatory
synaptic input (swapped terminals for inhibitory synaptic input). (e) The schematic of the spike pulse generator and reset circuit
(implements refractory period duration). (f) The two-stage opamp schematic used inside the read-out buffer. (g) The schematic
of the source-degenerated Operational Transconductance Amplifier (OTA) (used inside synaptic input and leak).
computational power versus the silicon area and design com-
plexity. This entails that the emulated neuron models should
be rich enough to replicate most computational studies, as
well as sufficiently simple to integrate enough neurons for
small functional networks on a single die. This is a good
compromise to envision large-scale systems, also endorsed
by other architectures, e.g., [30]. Further, Integrate and Fire
(I&F) models with refractoriness and adaptation are known
to reproduce the biological traces with good accuracy [31],
[32]. Our models of choice are therefore low dimensional
4
mode is featured that evokes a single output spike per input
synaptic event. The aforementioned digitally switchable terms
are realized by the transmission gate switches Sn in Fig. 1b.
An integrated voltage buffer reads out the membrane potential
as well as the activity on both shared synaptic input lines within
each neuron circuit.
D. Power and Area
Power reduction for large-scale integration is made possible
at two levels: First, the Capmem in this chip prototype can
provide bias currents as low as 15 nA. This saves significant
power consumption, since the circuits shift to moderate or weak
inversion. Further, the unused sub-circuits are power-gated.
The sub-circuit design has been optimized for area/power by
utilizing MOS gate capacitors, usage of thin and thick-oxide
transistors and dual voltage supply options, wherever possible.
As a result, the single neuron circuit consumes approximately
10 µW power (14.4 µW if both synaptic inputs are used).
E. Calibration
A major goal of the presented design was the possibility to
calibrate individual sub-circuits as well as the entire neuron.
The circuit has therefore been verified pre-silicon with focus
on mismatch compensation using the available Monte Carlo
models. Post-silicon, all 32 neurons have been calibrated for
mismatch across multiple dies, the respective results are shown
in Sec. IV. Utilizing these datasets, we can provide a mapping
between LIF model quantities and circuit-level parameters by
using fractional polynomial fits for the individual sub-circuits.
I V. C I R C U I T D E S I G N A N D M E A S U R E M E N T S
The neuron schematic designed for the current prototype
chip is shown in Fig. 1b connected to a single column of the
synapse matrix. Starting from the left, it shows a single synapse
column from which two output lines emanate - one carry-
ing excitatory pre-synaptic events and another inhibitory pre-
synaptic events. Each pre-synaptic event enables a 6-bit DAC
for a configurable duration of 10 ns–320 ns, which outputs a
current pulse on either of the two synaptic lines. At the neuron
side, these current pulses are received by the synaptic input
circuits (one excitatory, one inhibitory). These circuits model
the synaptic dynamics with an exponential kernel inside each
neuron circuit (Fig. 1d). The latter integrate current from both
synaptic inputs onto a capacitor Cmem that models the neuronal
membrane. The membrane can discharge itself through a
separate Leak circuit towards the resting potential Vleak. Once
the voltage reaches a threshold Vthresh, the circuit SpikeGen
triggers a digital pulse event (see fire in Fig. 1b,e) and brings
the membrane back to a reset potential Vreset via a Reset
module. This circuit also adds the respective refractory period,
essentially by clamping the membrane to Vreset for a duration
governed by the time it takes to toggle the inverter again. An
Analog IO block reads out debug voltages and injects (test)
stimulation current onto the membrane. The neuron schematic
also highlights the digitally controlled interconnecting switches
labeled S0−11. These are realized as transmission gates and are
threshold-based I&F models such as [33], [34]. Therefore,
in this initial prototype, we implement the LIF model as a
base computational element. The subthreshold response of the
LIF model can be defined by Kirchhoff's current law for the
conservation of charge:
Cmem
dVmem
dt
= −gleak · (Vmem − Vleak) + I
(1)
and if Vmem ≥ Vthresh,
Vmem → Vreset
(2)
where Vmem is the potential across the neuronal membrane,
Cmem is the membrane capacitance, Vreset and Vthresh are well
defined reset and threshold potentials, Vleak models the leak
potential and gleak is the leak conductance. I is the sum of
external current (Istim) and excitatory (Isyn,exc) and inhibitory
(Isyn,inh) synaptic currents. The synaptic inputs integrating
these currents are exponentially decaying current-based inputs,
whose time course can be defined as:
i )
(3)
τsyn (cid:19)Θ(t − tf
fi
wie−(cid:18) t−t
Isyn(t) =Xi Xf
where wi
is the weight of the synapse connecting a pre-
synaptic neuron to a post-synaptic neuron, tf
i denotes fth spike
at a synapse i, Θ(x) is the Heaviside step function and τsyn is
the synaptic time constant.
B. Specifications and Parameter Range
A selected set of computational modeling studies [35]–[46]
defines the tuning range and target specifications of the individ-
ual sub-circuits and eventually the neuron model parameters.
Table I summarizes these identified ranges for membrane and
synaptic time constants (τmem, τsyn) as well as the refractory
period (τrefr) in biological real-time. The presented hardware
Parameter Min. [ms] Max. [ms]
τmem
τsyn
τrefr
50
100
10
7
1
0
TABLE I: Target specifications of the neuron model parameters.
implementation uses current-based synaptic input. Because the
reference studies employ conductance-based synapse models,
an increased leak conductance by a factor of five is required
to be able to emulate the reduction of the effective mem-
brane time constant that is associated with high conductance
states [47], [48].
C. Debug and Testability
The neuron circuit provides the possibility to debug individ-
ual sub-circuits or disconnect individual terms for verification
and calibration. For example, a digitally configurable bypass
meant to test, debug and switch-off individual terms whenever
needed.
Each of the incoming input events to either excitatory or
inhibitory synaptic inputs can be made to directly trigger output
events (see fireout in Fig. 1b), bypassing the current integration
on the neuron membrane. Inverters with shifted trip-points
driven directly by the voltage drop on the input synaptic lines
make this possible - they are enabled by digital inputs S9,10
and tri-state inverters at the output. This "bypass mode" is
useful for testing certain modules of the system (e.g. event
routing) without relying on configured neurons. It is disabled
during nominal neuron operation.
Fig. 2: (a) Chip micrograph. (b) Prototype measurement sys-
tem.
The integrated array of 32 neuron circuits occupies 200 µm
× 376 µm of die area. Each neuron in turn occupies 200 µm
× 11.76 µm. The described prototype chip has an area of
1.9 × 1.9 mm2 and is fabricated in a low-K 1P9M 65-nm
low-power digital CMOS process. The bonded die and the
measurement setup are shown in Fig. 2. All measured analog
data presented within this work has been acquired using a
Keithley 2635B Sourcemeter for static current measurements
and a LeCroy Wavesurfer 44Xs digital oscilloscope for dy-
namic signals. The chip measurement framework comprises
of an on-board Xilinx Spartan-6 FPGA system that takes
command packets via a USB interface. The chip and all neuron
parameters are directly programmable via a C++/Python based
software environment. The neuron block specifications are
summarized in Table VII. Before demonstrating the neuron
operation, we describe the design and measurements of each
individual sub-circuit.
A. Synaptic Input
The synaptic input circuit provides the exponential synap-
tic dynamics to incoming events inside each neuron circuit
(Eq. III-A). The circuit integrates short current pulses (synaptic
events) arriving from 32 synapse circuits in a single column
onto an RC integrator before converting them into an equiv-
alent current with the help of a linear transconductor. The
architecture is shown in Fig. 1d. The current pulse events
from 32 synapse circuits are received on the shared line
labeled Isyn,exc in Fig. 1b and shown as an input terminal in
Fig. 1d. Each incoming input pulse event drops the voltage
5
on this line (nominally at 1.2 V), which is then recovered
with the synaptic time constant τsyn. This voltage drop is
proportional to the strength of the incoming pulse event. The
synaptic time constant is varied by the tunable resistor Rsyn,
while the integration capacitor Csyn is fixed at 1 pF. In the
current prototype, this capacitance is mostly contributed by
a metal capacitor, but it will eventually be realized from the
line parasitics alone - as the number of input synapses will
increase in the scaled-up final chip. The second OTA input Vsyn
is kept at approximately 1.2 V, the precise value can be altered
for calibration purposes. The realized amplifiers can cancel the
effect of input offset voltage at its output using a tunable bias
current IbiasOff.
1) Transconductor: The designed transconductor is a source-
degenerated OTA architecture, whose output current has a
linear dependence on input differential voltage within a limited
range, such that Iout = Gm(Vin+ − Vin-). Where Gm is the
OTA transconductance. The OTA's main bias current (labeled
Ibias) helps to vary this transconductance, whereas a second
bias (labeled IbiasSd) adjusts the proper biasing point of the
source degeneration pair. The OTA architecture is shown in
Fig. 1g. Transistors M1-M4, and M5-M8 form cascode current
mirror loads, whereas M9,10 form the input pair. Transistors
M11,12 are a source degeneration pair, acting therefore as
degenerating resistors to lower the gain. Their bias points are
tunable externally using a separate bias IbiasSd.
Note the presence of two separate stages at the output: First,
the output stage current multiplier formed by transistors M2h,4h
and M14h,16h - it provides extra output current in parallel to
the existing output stage formed by M2,4 and M14,16. A digital
switch labeled enhiCon enables this current multiplication. This
is useful when realizing a higher conductance - for example
in the leak term to ensure short membrane time constants.
The leak term (see Sec. IV-C) also uses this OTA and the
output current multiplier is implemented only there. Secondly,
another parallel circuit - a current mirror formed by M19,20
and M21,22 is used as an offset calibration circuit at the OTA
output. Ioffset is then the calibration current that is set equal to
the residual offset current, caused for example as a result of
input offset voltage between the OTA terminals. The current
mirror steals this current from the output stage of the OTA that
directly sends current to the membrane.
2) Synaptic Resistor: The synaptic resistor Rsyn is a tunable
resistor designed using bulk-drain connected devices [49]–[51]
and shown in Fig. 1c. The architecture of the resistor uses
a series of four pmos bulk-drain connected devices (labeled
M1 to M4) between the two terminals VinP and VinN. Each
pmos device connects its bulk terminal to its drain instead of
the nominal bulk configuration. They provide a linear increase
in drain current with increasing source-drain voltage. Series
stacked devices are used to reduce the source-drain potential
well below the threshold voltage. The total voltage drop across
the two resistor terminals can be as large as a couple of
hundred millivolts for the case of larger incoming events. For
example, a 10 µA event (if 4 ns long), can drop the synaptic
input line Isyn,exc/inh to 1 V, creating a drop of 200 mV across
the terminals. The resistor therefore provides almost linear
operation within this range. The pull-up nature of this synaptic
(k)
2
−500
Synaptic Offset Current [nA]
0
500
−2
6
(d)
Chips 20, 24
(e)
1.0
50
Synaptic Time Constant [µs]
1.5
10
30
(j)
Chip 20
25
20
15
10
05
14
12
10
02468
Counts
Counts
(c)
Chips 20, 22, 24
50
10
1
TimeConstant[µs]
(a)
(b)
Chip 20
0
75 150
0
75 150
∆V [mV]
0
100 200 300 400 500
Resistor Bias Current [nA]
(h)
Chip 22, N15
(i)
500
400
300
200
100
0
Counts
(f)
Chips 20, 22, 24
(g)
10 20 30
10 20 30
Synaptic Time Constant [µs]
10
20
30
10
20
30
Synaptic Time Constant [µs]
60
50
40
30
20
10
0
Current[nA]
180
160
140
120
100
80
60
40
20
0
Counts
Fig. 3: (a) Variation of the characteristic curves for the 64 synaptic input resistors of a single chip, with bias set to 80 nA
(pre-calibrated). (b) Post-calibration resistor curves with time constant of all synaptic inputs set to equal the mean of those in (a).
(c) Tuning the synaptic time constants of 192 synaptic input circuits by varying the resistor bias current (d,e) Distribution of the
minimum and maximum range of the achieved synaptic time constants. (f) Pre-calibration distribution of synaptic time constants
with a mean of 5 µs, 10 µs and 20 µs. (g) Post calibration: the target value of (f) is processed through individual polynomial fits.
(h) Trial-to-trial variation of synaptic time constants for programmed values of 5 µs, 10 µs and 25 µs. (i) Variation in synaptic
time constants measured from a train of 500 continuous input events. (j) Pre-calibrated synaptic output offset current with Vsyn
at 1.2 V for all OTAs. (k) Residual synaptic current after individually adjusting Vsyn and IbiasOff.
input resistor ensures that source potentials are always higher
than the drains. This prevents the two terminals from getting
swapped and possibly consume higher currents, since it is
then a nominal pmos configuration (bulk connected to source
terminal). A single bias current Ibias tunes the gate voltage
Device
M1,2,3,4
M1c,2c,3c,4c
M1b,2b,3b,4b
M1a,2a,3a,4a
M5,6
Width/Length [µm/µm]
0.15/1.0
0.15/5.0
0.15/0.5
0.15/0.5
0.75/0.5
TABLE II: The device dimensions of the bulk-drain connected
synaptic resistor.
of all subsequent bulk-drain connected devices via cascode
current mirrors, labeled M5,6 and MXa–MXb, where X denotes
the respective devices of each pmos stage. The devices marked
MXc help tune the transistor bias points. Table II summarizes
the transistor dimensions for the designed resistor. Notice that
in order to provide larger resistance the channel lengths of the
bulk-drain connected devices M1−4 are relatively long. We
have measured synaptic input circuits on up to three different
dies. The characteristic curves of the resistor within the desired
voltage range are shown in Fig. 4. The figure plots the
current vs. the potential difference by varying the resistor's bias
(Ibias in Fig. 1c). The equivalent swept Ibias and the resulting
resistances determined from a linear fit are shown. Note that at
very low bias currents the resistance is very large (bottom-blue
curve) and the resistance values are inclined for an exponential
increase as a function of bias current. Notice the presence of
finite voltage offsets, as the plotted traces cut the zero-nA line
at a potential difference of about 15 mV across the terminals.
This is due to supply drop and can be corrected for by tuning
the input Vsyn (second OTA terminal) during operation.
We further show our measurement results from all 64 re-
sistors (both synaptic inputs), configured for the same mid-
range resistance of about 4 MΩ in the desired voltage range in
Fig. 3a. The traces show triode curves for typical MOS devices,
since the bulk-drain connected devices are in fact biased in
linear region to adjust the resistive range. The variation in the
traces are due to device mismatch, largely contributed from
the biasing stages (MXa,Xb,Xc) that help set the bias points
of each individual device. This is minimized in Fig. 3b after
applying the calibration described in Sec. IV-A4.
7
variation of 500 incoming events for the aforementioned three
configurations.
B. Spike Generator and Reset
This circuit is responsible for evoking a digital output event
as an indication of spike occurrence, once the membrane
reaches a specified voltage threshold Vthresh. It also resets the
membrane voltage to a fixed reset potential Vreset and adds a
refractory period τrefr before the membrane starts integrating
again. The circuit consists of a comparator whose output stage
is reset after a finite delay via a feedback loop, as shown in
Fig 1e. Once the membrane Vmem reaches the threshold Vthresh,
the comparator outputs a logic 'high' (VOH). After a finite
time delay tdelay, the comparator output stage is pulled down to
logic 'low' (VOL). This essentially generates a digital voltage
pulse signal fire. The fire pulse marks the spike occurrence
and is routed as a single event through to the digital backend.
Furthermore, it also clamps the membrane Vmem to Vreset by
toggling the switch Srst (see Fig 1e). This happens as the
voltage over the capacitor Crefr (which constantly integrates
a tunable bias current Irefr) is reset by the incoming pulse.
This initiates the neuron's refractory period (τrefr) that lasts
until the voltage across the capacitor Crefr triggers the inverter
again to disconnect the switch Srst (Fig 1e). This marks the
end of refractory period and the membrane Vmem is then free
to integrate again.
The delay-cell designed for the spike pulse generator (la-
beled tdelay) is a tunable current-starved delay element [52]
typically set to 100 ns. The comparator is a simple two-stage
architecture similar to an uncompensated two-stage op-amp
that provides high gain to drive outputs to the desired output
levels (VOH or VOL). To realize long refractory periods without
a large Crefr, the bias current Irefr from the respective Capmem
cell is divided 10 times (fixed), reducing the effective bias
currents to 1.5 nA (min.) and 100 nA (max.).
Our measurements of two dies (32 circuits each) suggest
that the refractory period can be tuned successfully over a
range from 1.11 µs up to about 137 µs, as shown in Fig. 5a,b
(1σ quantile, see also Table III). The variation in longer
time constants is relatively large due to very small input bias
current (1.5 nA) being integrated on Crefr. The circuit has
been designed for much longer refractory periods than what
is known through biology to cater for artificial models such
as [53].
1) Calibration: Since the duration of the refractory period
is solely determined by the current Irefr (Fig. 1e), this variation
is catered for by measuring τrefr as a function of Irefr and
applying a second order fractional polynomial fit for each neu-
ron. In Fig. 5c,d, the distributions of three different refractory
periods are shown pre- and post-calibration. Notice the larger
spread towards smaller bias currents and therefore higher time
constants, which is significantly reduced after calibration (c.f.
Tab. IV). Fig. 5e shows the calibrated membrane potential
of all 32 neurons on a single die after spike occurrence. All
neurons are configured to a reset potential of 200 mV, a spike
threshold of 1.2 V, a resting potential of 1 V and - in order to
produce steep, comparable edges - to a short membrane time
Bias, Resistance
15nA, 14.66MΩ
130nA, 3.64MΩ
240nA, 2.36MΩ
355nA, 1.82MΩ
465nA, 1.52MΩ
580nA, 1.32MΩ
695nA, 1.19MΩ
805nA, 1.08MΩ
920nA, 1.00MΩ
200
150
100
50
0
Current[nA]
0.00
0.05
0.15
0.20
0.10
∆V [V]
Fig. 4: Tunability of the synaptic resistor. The current flowing
through Rsyn is measured as a function of the voltage drop
across the resistor terminals, as the resistor bias is swept.
3) Full Circuit: Finally, the measured results demonstrating
the range of possible synaptic time constants τsyn are shown
in Fig. 3c. The data has been acquired by measuring three
separate dies, each having 64 synaptic input circuits. The plot
shows a range of available time constants as well as their
mean value as the bias of the tunable resistor is swept from
0 to 0.5 µA. The plot shows a non-linear tuning curve which
steepens substantially in the low current regime. It corresponds
to the tuning behavior of synaptic resistor shown in Fig. 4.
The variation in different achievable time constants, hence is
larger when we tune longer synaptic time constants, compared
to shorter ones. This is also evident from the histogram plot of
Fig. 3d,e, where a distribution of both the shortest and longest
synaptic time constants (bias settings of 1 µA and 15 nA) are
plotted for 128 samples. Given these distributions, the range
of available synaptic time constants is between 1.24 µs and
20.5 µs for one-sigma single sided quantiles (see Table III).
4) Calibration: The input-referred offset of the synaptic
input OTA, if not canceled, can make the neuron membrane
integrate unwanted output offset current. To calibrate this, we
use the OTA reference voltage Vsyn as well as the offset
compensating bias IbiasOff, both of which are locally tunable
parameters. Using Vsyn we cancel the output offset to a first
approximation. The residual offset is trimmed by sweeping
IbiasOff and searching the root of a linear fit. Fig. 3j,k shows the
measurement of all 64 synaptic inputs' output offset currents
on a single die pre- and post-calibration. The post-calibration
residual current is below 5 nA for all samples, which corre-
sponds to approximately 20 mV deviation in the membrane
resting potential at a membrane time constant of 10 µs. The
curves shown in Fig. 3c are fitted with second order fractional
polynomials, which reduces the relative spread of individual
time constants. Fig. 3f,g shows the spread of three different
synaptic time constants pre- and post-calibration from three dif-
ferent dies. The standard deviation for these datasets is depicted
in Tab. IV. Finally, we demonstrate the trial-to-trial variation
for three different time constants within a single neuron. In
Fig. 3h, the mean time constant of 500 incoming synaptic
events is measured after reprogramming from a different time
constant. Most samples are clearly restricted to the same bin.
Fig. 3i on the other hand shows the synaptic time constant
8
40
(j)
0 10 20 30 40
10 20 30
Refractory Time [µs]
0
20
Time [µs]
1.0
0.5
0.0
Membrane[V]
(d)
(e)
(c)
Chips 22, 24
Chips 20, 22, 24
75
50
25
0
Counts
(h)
100
(i)
(g)
Chips 20, 24
70
60
50
40
30
20
10
0
Counts
20
10
Counts
0
0.3
0.4
25 50 75
Membrane Time Constant [µs]
20
0
Membrane Time Constant [µs]
40
0
10
20
(b)
Chips 20, 24
(a)
20
15
10
Counts
1.5
1.0
Refractory Time [µs]
150 200
B: 280, SD: 70
B: 280, SD: 90
B: 280, SD: 140
B: 375, SD: 465
B: 515, SD: 465
B: 655, SD: 465
B: Ibias [nA]
SD: IbiasSd [nA]
0.5
(f)
05
012
0.4
1.2
0.0
Membrane Voltage [V]
0.8
−1
−2
Current[µA]
Fig. 5: (a,b) The distribution of minimum and maximum achievable refractory time periods. (c,d) Pre- and post-calibration results
of refractory period circuits, set for a mean of 5 µs, 15 µs and 25 µs respectively. (e) Calibrated membrane traces showing a
post-spike response of different on-chip neurons with refractory times of 10 µs and 25 µs. (f) Output current of the leak term
as the membrane potential is swept with the resting potential fixed at 0.52 V. (g,h) Distribution of the minimum and maximum
achievable membrane time constants. (i,j) Pre- and post-calibration distribution for τmem of 1 µs, 10 µs and 20 µs.
constant of 250 ns. The figure shows two traces per neuron,
one programmed for a refractory time of 10 µs and the other
for 25 µs. Note how all neurons reach the resting potential
with little remaining spread in the time constant. The visible
variations of the resting potential are due to residual offset
currents from the synaptic inputs (c.f. Sec. IV-A4), as well
at least 31 mV of input offsets of the leak OTAs (estimated
through simulation using Monte Carlo device models).
C. Membrane Leak and Capacitance
The membrane conductance has been realized using a linear
transconductor whose output terminal is fed back to its negative
input as shown in Fig. 1b. In this configuration, the total
conductance across the input and output terminals equals the
transconductance Gm. The usage of linear transconductors to
realize conductance is already well known in neuromorphic
architectures [5], its architecture is described in Sec. IV-A1.
A two-bit configurable gate oxide capacitor provides a total
of 2.36 pF membrane capacitance. This is implemented in
multiples of 590 fF for layout symmetry, switchable between
2.36 pF, 1.77 pF or 590 fF via S7-8 (see Fig. 1b). To minimize
distortions introduced by the inherently non-linear behavior
of MOS capacitors, the respective transistors were chosen to
be thick-oxide devices. Within the membrane's full dynamic
range of up to 1.2 V, the bulk-gate drop can therefore be kept
above 1.3 V and the device in inversion region where the gate
capacitance is constant.
Parameter 1 min (1σ, 3σ)[µs] max (1σ, 3σ)[µs]
τrefr
τsyn
τmem
2
1.11, 1.24
1.24, 1.41
0.35, 0.41
137.5, 104.5
20.5, 13.6
16.6, 11.3
1 Single-sided 1σ (3σ) quantiles of 84.13% (99.86%)
2 Measured using Cmem = 2.36 pF; The min./max.
τmem (1σ, 3σ) for Cmem = 570 fF is estimated to
be 0.08, 0.10 µs and 4.11, 2.82 µs respectively
TABLE III: Measured range of time constants from data
acquired across two dies.
In Fig. 5g,h, we show the distribution of achieved membrane
time constants with the largest selectable membrane capaci-
tance. The distribution highlights the minimum and maximum
achievable time constants (τmem) from a total of 64 leak circuits
on two different dies. It is shown that the neuron can set time
constants, as small as 0.35 µs and up to 16.6 µs (one-sigma
quantiles) and can be further reduced by selecting different
capacitor configurations (c.f. Table III).
Fig. 1b shows how the leak term uses the transconductor in a
feedback configuration, as opposed to the synaptic input where
the OTA is in open loop. In feedback configuration, the OTA
transconductance Gm is the overall conductance of the leak
term, defining membrane time constant as τmem = Cmem/Gm.
target
5 µs
15 µs
25 µs
5 µs
10 µs
20 µs
1 µs
10 µs
20 µs
pre-calibration
0.50 µs (9.95%)
1.84 µs (12.3%)
4.88 µs (18.9%)
0.75 µs (15.2%)
2.49 µs (20.7%)
4.01 µs (19.9%)
0.19 µs (19.3%)
3.14 µs (40.9%)
8.34 µs (42.3%)
post-calibration
0.14 µs (2.83%)
0.12 µs (0.77%)
0.51 µs (2.06%)
0.14 µs (2.63%)
0.18 µs (1.75%)
1.12 µs (5.56%)
0.03 µs (3.03%)
0.46 µs (4.53%)
2.27 µs (11.5%)
τrefr
τsyn
τmem
TABLE IV: Pre- and post calibration standard deviations for
the histograms shown in Fig. 3 and Fig. 5.
Fig. 5f shows six measured traces plotting the output current
of the leak term as a function of the membrane voltage.
Three of them sweep the main bias current Ibias keeping
IbiasSd constant, while the remaining three sweep IbiasSd, while
Ibias is held constant. The traces highlight how lowering the
source degeneration bias linearizes the response and allows
for a wider range of membrane voltages. This is because the
degeneration resistors, implemented as transistors M11, M12,
are shifted from linear region to the saturation region of the
weak inversion regime, where their on-resistance Rs is as large
as few MΩ at low IbiasSd. This effectively linearizes the curve
as the OTA's transconductance Gm approximates to 1/Rs.
1) Calibration: The membrane time constant can be tuned
by three parameters: A two-bit variable membrane capacitor
and the two biases (Ibias, IbiasSd) of the leak OTA. Despite
the low gain curves and consequently wide range for lower
bias values of IbiasSd (Fig. 5f), tuning is done using the former
two parameters. The circuit achieves membrane time constants
between 0.35 µs and 16.6 µs within a linear range of about
230 mV. The reason for not relying on IbiasSd is twofolds -
first, low values of IbiasSd contribute a large input offset, causing
an output residual current, that is not calibrated for by tuning
Vleak. Secondly, at lower values of IbiasSd, overall transconduc-
tance Gm contributed by the OTA is a non-monotonic function
of Ibias. To avoid both effects, we calibrate the membrane
time constant by adjusting the OTA's bias current (Ibias) and
therefore trade linear range for robustness. We measure the
relation between τmem and Ibias at fixed IbiasSd of 1 µA for
a voltage decay from 600 mV towards a resting potential
of 400 mV. The resulting curves are fitted by second degree
fractional polynomials. Pre- and post-calibration variations of
τmem are depicted in Fig. 5i,j for targets of 1 µs, 10 µs and 20 µs,
the related standard deviations are shown in Tab. IV. It may
be noted that not all neurons can achieve the maximum time
time constant of 20 µs. These neurons are visible as outliers
to the left of the depicted distribution.
D. Analog Input/Output
We can inject constant current into the membrane, hold
it to a fixed potential or read out debug voltages via the
analog I/O sub-circuit (see Istim and VreadOut
in Fig. 1b).
9
Output range
Load
–1-dB bandwidth
Gain
Power
Input offset
Slew rate
Compensation cap.
0.1–2.1 V
≥ 16 pF k10 MΩa
1.15 MHzb
62.6 dBc
127 µWc
14 mV
2 V/µs
600 fF
TABLE V: Measured specifications of the read-out buffer.
aEstimated off-chip load
bMeasured after the LPF formed by output Tx-gate and parasitic load
cSimulation only
Device
M1a,1b,2a,2b
M3,4
M6, M7
M12, M5
M9,11
M8,10
Width/Length [µm/µm]
2.4/0.28
0.8/0.38
4.0/0.38 & 4.0/0.39
0.15/0.75 & 1.6/0.39
3.6/0.38
0.6/0.38
TABLE VI: The device sizes of the read-out amplifier.
The monitored output voltage is multiplexed from among the
membrane potential or incoming synaptic activity at any of the
two synaptic input lines. The Opamp designed for the buffer is
a two-stage amplifier with split-length indirect compensation
[54], [55] and is shown in Fig. 1f. The buffer has a –1-dB
bandwidth of 1.15 MHz when driving large off-chip loads
at a maximum power consumption of 127 µW. The close-
loop opamp bandwidth is filtered by the output low-pass filter
formed by transmission gate and the off-chip load. In the
current design, each neuron buffer directly drives the shared
read-out line that terminates at the output pad. In comparison to
Miller compensation, the use of indirect compensation ensures
smaller area, higher bandwidth, and reduced power consump-
tion [55]–[57]. During nominal neuron operation, the amplifier
is switched off, the stated power is only consumed during
analog read-out (therefore not reported in Sec. III-D). The
measured amplifier specifications are summarized in Table V
and the transistor dimensions used in this design are listed in
Table VI.
E. Full Circuit Characterization
Here we describe the measurements and performance of the
complete neuron circuit. Fig. 6 shows equidistant incoming
events at the excitatory synaptic input line (bottom trace) and
the resulting membrane trace (top). The reset potential is set to
0.2 V, the leak potential to 0.58 V and the spiking threshold to
1.2 V. The input synaptic lines (bottom trace) have a constant
1.2 V potential unless an input event arrives. At event arrival,
this level drops and is recovered with a time constant τsyn. Each
incoming event increases the membrane potential, eventually
10
0
20
40
80
60
Time [µs]
100
120
140
Fig. 6: Measured results - bottom: synaptic input potential
indicating incoming events as evaluated on Isyn,exc line, top:
resulting membrane response to the synaptic events.
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
1.2
1.1
1.0
Membrane[V]
Input[V]
Leaky I&F
32
11.76 × 200 µm2
2.5/1.2 V
65-nm CMOS
Neuron model
No. of neurons
Area of single neuron
Voltage supply
Process
Speed-up (acceleration) factor ×1000
Global parameters
Local (individual) parameters
Configurability
Membrane capacitor (max.)
Input synaptic event (max.)
1 voltage bias 1
18 (14 I-bias, 4 V-bias)1
15-bit digital bus
2.36 pF (2-bit config.)
10 µA, 4 ns pulses2
1 available from on-chip tunable capacitive memory
2 amplitude and length of each current pulse emitted by
cells
the synapse circuit
b) Software Neuron: Membrane Potential
c) Hardware Neuron: Synaptic Input
d) Hardware Neuron: Membrane Potential
200
100
0
Istim[nA]
900
800
700
600
Vmem[mV]
120
60
0
∆Vsyn[mV]
900
800
700
600
Vmem[mV]
0.0
0.2
0.4
0.6
Time [ms]
0.8
1.0
Fig. 7: A one-to-one comparison of the emulated neuron vs.
software simulation.
F
. Power Consumption
In order to estimate the chip prototype's power consumption,
the total current drawn by the chip during operation of the
network described in Sec. V has been measured. From this
measurement, we infer a mean full-chip power consumption
of (48.62± 0.04) mW. This figure includes prototype-specific
consumers - especially the 2.5 V full-swing IO channels
that communicate each spike off-chip to the controlling FPGA.
This IO power is expected to scale sublinearly with increasing
chip size.
evoking a spike, followed by a membrane reset to initiate the
refractory period duration.
We further demonstrate the neuron parameter set from a
typical modeling study [43] and compare the response proper-
ties of our emulated on-chip neuron, with that of a simulated
accelerated LIF model neuron (Eq. III-A). We use the Brian
spiking neural network simulator [58] to simulate the ideal LIF
neuron. The on-chip neuron circuit here is uncalibrated and
tuned manually such that the most pronounced non-ideal effects
are canceled for. This includes the output residual current of
synaptic input due to input offset voltage between the OTA
terminals. Further, time-constants are fine-tuned and synaptic
weights are adjusted to match the spike times of the software
simulation. The synaptic and membrane time constants (τsyn,
τmem) are set as 1.5 µs and 10 µs, with a refractory period
τrefr of 2 µs (×1000 speed-up from biological real-time). The
threshold is set to 870 mV and the leak and reset potentials
are 600 mV. A random spike train stimulus containing 32× 20
events (each synapse relaying 20 pre-synaptic events) is in-
jected into both neurons and their corresponding membrane
traces are plotted in Fig. 7. Fig. 7a shows the input current
as a result of random input stimulus of the software neuron,
while Fig. 7b shows its corresponding membrane voltage. In
the hardware domain, we monitor the on-chip voltage on the
excitatory synaptic input line, as its voltage drop (defined as
∆Vsyn := VSyn − Vsyn,exc, see Fig. 1d) is in proportion to the
charge of each incoming synaptic event. A larger voltage drop
corresponds here to higher input synaptic current (Fig. 7c).
Fig. 7d shows the measured membrane voltage of the on-chip
neuron. Comparing the dynamics of both membrane traces, one
finds the RMS of their common-mode corrected difference to
be 45 mV. We see that, when presented with the same input,
the on-chip neuron has similar spike-times and the membrane
voltage follows the time course of the software simulation.
TABLE VII: A summary of neuron array specifications.
a) Software Neuron: Synaptic Input
11
0
20 40 60 80 100 120
Stimulation [kHz]
0
20 40 60 80 100 120
Stimulation [kHz]
120
100
80
60
40
20
0
(c)
Response[kHz]
40
30
20
10
0
(b)
Response[kHz]
(a)
Fig. 8: (a) Schematic diagram of the implemented network topology. All neurons in the network form self- and direct-neighbor
excitatory projections, as well as to central inhibitory neurons. Both inhibitory neurons in turn inhibit the ring neurons and do
not make connections between themselves. (b) Network response for two competing input stimuli r1 and r2 before the network
routing has been established: One neuron receives Poisson noise with an average rate of r1 = 50 kHz, another neuron receives
Poisson noise with variable frequency between r1 = 0 kHz and r1 = 120 kHz. The mean firing rate of these neurons and their
next neighbors is depicted as an average over 10 trials with different random seeds for the input stimulus. (c) Network response
for two competing input stimuli r1 and r2 after network routing has been established: The Poisson input is as specified in Fig. 8b.
An estimation of the mean energy cost of a single spike has
been compiled by comparing the aforementioned result against
the power consumption of an idle chip (without stimulating
spikes) with the same network configuration. Due to the greatly
reduced number of expensive outgoing traffic when the system
is idle, the resulting number has to be understood as an upper
bound for the on-chip spike cost. We find Espike ≤ (790 ±
170) pJ for the presented network.
V. A W I N N E R - TA K E - A L L N E T W O R K
We demonstrate a soft winner-take-all circuit [59]–[61]
using the on-chip LIF neuronal array. From an array of 32
neurons, 20 are used to form a ring of excitatory neurons, each
having self-excitation, nearest-neighbor excitatory projection,
and excitatory projections to two common inhibitory neurons.
The inhibitory pool of two neurons in turn inhibit all neurons
in the ring and do not make any projection between themselves.
The resulting network topology is shown in Fig. 8a. All
neurons are chosen randomly and are calibrated as described
in the previous sections. To verify the network's behavior,
two opposite (equidistant) neurons in the ring network are
stimulated with Poisson stimuli of various frequencies between
0 kHz and 120 kHz. Fig. 8c shows the average firing rates
of both competing sub-populations. A sub-population here is
defined as the stimulated neuron and its two direct neighbors.
The stimulus to one is fixed at 50 kHz, while to the other is
being swept. The swept stimulus has an increment of 0.5 kHz
between separate runs, accounting to 240 different stimuli runs
within the specified range. Each experiment has been run for
100 ms. The line/highlighted region in Fig. 8c shows the
average and standard deviation over 10 independent runs; each
with a different random seed for the Poisson input. Due to the
speed-up of the presented neuron array (compared to biological
time scales), the total network emulation time of four minutes
(10 trials, each consisting of 240 different stimuli `a 100 ms)
corresponds to 2.8 days of biological network activity. The
mean firing rate for both populations during 10 control runs
with disabled spike connectivity (routing) is shown in Fig. 8b.
As in Fig. 8c, the mean firing frequency of one single sub-
population is plotted - as there are no connections, both
neighbors are quiet and population firing rate is scaled roughly
by a factor of three.
V I . D I S C U S S I O N
It has been demonstrated that
We have presented an array of LIF neurons designed for
the initial prototype of our HICANN-DLS chip. The pa-
per summarized comprehensive measurement and calibration
results to demonstrate the usability of the realized neuron
array.
the statistical spread
of the time constants can be minimized significantly, once
they are calibrated for device mismatch and corner variations.
The transconductor architecture features an offset cancellation
mechanism that reduces the residual synaptic output offset by
two orders of magnitude. The usability of the circuit has been
demonstrated by its low trial-to-trial variation. In comparison
to synaptic time constants, the membrane time constants show
wider spread, since the former was tuned via the synaptic
resistor which has a better control compared to leak OTA biases.
We conclude that both synaptic as well as membrane time con-
stants need improvement to cover longer durations in a future
revision. The circuit evokes a digital event as an output spike
that also initiates the refractory period - a very broad duration
has been demonstrated with a statistical spread that is widely
reduced after calibration. The presented prototype is ready to
i
i-1
i+1
excitatory
inhibitory
CMOS tech. [nm]
28
Architecture
Speed-up
Neuron Model
TrueNorth Neurogrid
180
analog
subth.
×1
quad.
I&F
Yes
digital
×1
augmented
LIF
No
This
Work
65
analog
×1000
LIF
Yes
Biophysical
Dynamicsa
Neuron Area [µm2]
Local Parameters
Parameter Area [µm2]
1800
No
–
2352c
Yes
2688d
2900b
Yes
–
a tunable τrefr, τmem, τsyn
b multiplexed 256 times per time step
c including excitatory and inhibitory synaptic inputs
d area of 18 analog on-chip biases per neuron for current
work; not comparable to TrueNorth's 124-bit SRAM-
based parameter config. or Neurogrid's shared biases
TABLE VIII: An overview of neuron model specifications in
large-scale neuromorphic architectures.
run local calibration routines using the embedded processor,
thereby enabling highly parallel calibration of multiple neurons
in large-scale systems. Such local calibration algorithms are
subject to future development.
To verify the characteristics of the emulated physical neuron,
a comparison with a software-simulated ideal LIF neuron is
made: the number of spikes as well as the membrane dynamics
are found to be in close correspondence. Finally, using the chip
routing resources, the FPGA and system software stack, a small
soft winner-take-all network is realized. A basic functionality
of the circuit
to two
competing sub-populations. It is pertinent to note that usage of
accelerated hardware reduced the network run-time (compared
to biological real-times) by three orders of magnitude, enabling
a platform feasible to study long-term developmental processes
in the nervous system.
is shown by feeding Poisson input
We further compare our chip's neuron implementation with
two state-of-the-art neuromorphic implementations, namely the
IBM's TrueNorth system [15] and Stanford University's Neu-
rogrid architecture [9]. The TrueNorth neuron is a phenomeno-
logical LIF model that implements various synthetic, arithmetic
and logical operations along with real-time spiking behaviors.
Being a phenomenological model, the temporal dynamics of
the synaptic and membrane time constants (τsyn, τmem) are not
tunable implementations. Despite a 28 nm CMOS implementa-
tion, the digital circuit occupies more area (2900 µm2), reduced
effectively by the possibility of time-multiplexing. Alternate
phenomenological implementations such as the one presented
by Loihi's 14 nm architecture offer slightly worse neuron
density [16]. Compared to this, Neurogrid - like BrainScaleS
12
TrueNorth Neurogrid Hi-DLS
Die Area [mm2]
Neurons/core [#]
Neurons/timea[#]
Synapses/core [#]
SynOps/sec [GSOPS]
430
4k
1M
64k
58
168
64k
64k
256kb
–
32
512
512k
131k
16k
a emulated neurons per timescale: HICANN-DLS acceler-
ates ×1000; TrueNorth multiplexes ×256
b each synapse population (4× per neuron) counted as
individual synapse
TABLE IX: An overview of single-chip characteristics of large-
scale neuromorphic systems with scaled-up HICANN-DLS
chip. Data plotted in Fig. 9.
- offers biologically plausible dynamics and tunable time
constants. However, all 64 k neurons in a single Neurocore
share their parameters, whereas the current work has dedicated
neuron parameters. Neurogrid implements subthreshold MOS
dynamics, which typically allow for compact neurons due to
small currents leading to less capacitor area for given time
constants. Both Neurogrid and TrueNorth operate in real-time,
as opposed to our accelerated approach. The neuron area occu-
pied here is slightly larger than Neurogrid neuron - in which
about 500 µm2 is occupied by membrane capacitor (MOScap).
The neuron model we implement in the current prototype
is a LIF model, compared to Neurogrid's two compartment
quadratic model. The presented modular neuron architecture
has been enhanced to an Adaptive-Exponential I&F (AdEx)
model with multiple compartments in a recently submitted chip
prototype [62]–[65]. Table VIII summarizes this comparison
against the TrueNorth and Neurogrid neurons.
The measured energy per output spike (Sec. IV-F) includes
the off-chip routing and has been reduced by providing local
(on-die) routing in the next prototype. We have not measured
energy per synaptic activation in our chip which has been
measured at 941 pJ and 26 pJ for implemented networks in
Neurogrid and TrueNorth systems. We expect that the upper
bound of (790 ± 170) pJ of energy/spike will reduce with
more neurons and input synapses in the scaled-up system.
The energy consumed in the stated neuromorphic architectures
are orders of magnitude lower than, for example, a quad-core
Xeon CPU [66] simulating a spiking neural network and still
more efficient compared to the network emulated on many-core
embedded systems, e.g. on the SpiNNaker system [67].
Fig. 9 presents a multi-dimensional overview of this proto-
type chip with the scaled-up HICANN-DLS as well as the
single-chip TrueNorth and Neurogrid implementations. The
chip area, number of physical neurons and synapses are shown
against peak synaptic operations per second (SOPS) and em-
ulated neurons in a given time-scale. The resulting pentagons
reflect the architectural preference of each implementation.
Data about peak SOPS was not available for Neurogrid. The
TrueNorth chip with the largest die area (430 mm2) time-
muliplexes the 4096 physical neuron instances to realize the
13
system architect and conceived the initial OTA and resistor
architectures, later implemented by S.A.A. The conceptual
advice was given by K.M.
The authors would like to extend their gratitude to Matthias
Hock for support during design and measurement.
R E F E R E N C E S
[1] E. R. Kandel, J. H. Schwartz, and T. M. Jessell, Principles of Neural
[2]
[3]
Science, 4th ed. New York: McGraw-Hill, 2000.
J. von Neumann, "First draft of a report on the edvac," IEEE Annals of
the History of Computing, vol. 15, no. 4, pp. 27–75, 1993.
J. von Neumann, The computer and the brain. New Haven, CT, USA:
Yale University Press, 1958.
[4] K. Meier, "Special report : Can we copy the brain? - the brain as
computer," IEEE Spectrum, vol. 54, no. 6, pp. 28–33, June 2017.
[5] C. A. Mead, Analog VLSI and Neural Systems. Reading, MA: Addison
Wesley, 1989.
[6] G. E. Moore, "Cramming more components onto integrated circuits,"
Proceedings of the IEEE, vol. 86, no. 1, pp. 82–85, 1998.
[7] G. Indiveri and S.-C. Liu, "Memory and information processing in
neuromorphic systems," Proceedings of the IEEE, vol. 103, no. 8, pp.
1379–1397, 2015.
[8] S. Furber, "Large-scale neuromorphic computing systems," Journal of
Neural Engineering, vol. 13, no. 5, p. 051001, 2016.
[9] B. V. Benjamin et al., "Neurogrid: A mixed-analog-digital multichip
system for large-scale neural simulations," Proceedings of the IEEE,
vol. 102, no. 5, pp. 699–716, 2014.
[10] R. J. Vogelstein et al., "Dynamically reconfigurable silicon array of
spiking neurons with conductance-based synapses," IEEE Transactions
on Neural Networks, vol. 18, no. 1, pp. 253–265, Jan 2007.
[11] Y. Wang and S.-C. Liu, "A two-dimensional configurable active silicon
dendritic neuron array," IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 58, no. 9, pp. 2159–2171, 2011.
[12] S. Brink, S. Nease, P. Hasler, S. Ramakrishnan, R. Wunderlich, A. Basu,
and B. Degnan, "A learning-enabled neuron array IC based upon
transistor channel models of biological phenomena," IEEE Transactions
on Biomedical Circuits and Systems, vol. 7, no. 1, pp. 71–81, 2013.
[13] N. Qiao et al., "A reconfigurable on-line learning spiking neuromorphic
processor comprising 256 neurons and 128k synapses," Frontiers in
Neuroscience, vol. 9, p. 141, 2015.
[14] E. Chicca, F. Stefanini, C. Bartolozzi, and G. Indiveri, "Neuromorphic
electronic circuits for building autonomous cognitive systems," Proceed-
ings of the IEEE, vol. 102, no. 9, pp. 1367–1388, Sept 2014.
[15] P. A. Merolla et al., "A million spiking-neuron integrated circuit with a
scalable communication network and interface," Science, vol. 345, no.
6197, pp. 668–673, 2014.
[16] M. Davies et al., "Loihi: A neuromorphic manycore processor with on-
chip learning," IEEE Micro, vol. 38, no. 1, pp. 82–99, January 2018.
[17] H. Markram, "The human brain project," Scientific American, vol. 306,
no. 6, pp. 50–55, 2012.
[18] FACETS, "Fast Analog Computing with Emergent Transient States –
project website," http://www.facets-project.org, 2010.
[19] S. B. Furber et al., "The SpiNNaker project," Proceedings of the IEEE,
vol. 102, no. 5, pp. 652–665, May 2014.
[20] K. Meier, "A mixed-signal universal neuromorphic computing system,"
[21]
[22]
in Proc. IEDM, Dec 2015, pp. 4.6.1–4.6.4.
J. Schemmel, D. Bruderle, A. Grubl, M. Hock, K. Meier, and S. Millner,
"A wafer-scale neuromorphic hardware system for large-scale neural
modeling," in Proc. ISCAS, 2010, pp. 1947–1950.
J. Schemmel, "BrainScales 2: A novel architecture for analog acceler-
ated neuromophic computing and hybrid plasticity," Internal document,
ASIC Lab., Kirchhoff-Institute for Physics, 2017.
maximum number of neurons and synapses, while Neurogrid
operates in real-time and features the largest physical neuron
count (65 k per Neurocore). The HICANN-DLS chip features
a 6-bit STDP enabled synapse circuit, which cannot be com-
pared with Neurogrid's analog-diffusor network and synapse-
population circuits (which we count here as individual synapse)
- and neither with TrueNorth's 1-bit crossbar synapses. For the
least die area, HICANN-DLS maximizes synaptic operations
per second due to the time-acceleration and possibility of high
input data rates. We emphasize that Fig. 9 and its correspond-
ing data given in Table IX may infer a very limited set of
architectural and operational features integrated by each large-
scale system. It highlights how the scaled-up chip of current
prototype will feature against other single-chip building blocks.
A direct comparison pertaining to varying architectural designs
and system capabilities is not possible.
Emulated neurons
per time
7
10
1014
Synaptic
operations
[SOPS]
This prototype
HICANN-DLS
TrueNorth
Neurogrid
Die Area
[m2]
1 0 − 6
0
10
1
0
6 Physical
neurons
Fig. 9: Overview of the single-chip implementations in con-
temporary large-scale neuromorphic architectures.
V I I . C O N C L U S I O N
We have presented the design, measurement, calibration and
application of our mixed-signal 65-nm CMOS LIF neuronal
array for the second-generation BrainScaleS platform. The
neurons cover a wide tunable parameter range to act as a
general-purpose element for many computational models. Hav-
ing demonstrated a functional scaled-down prototype of the 65-
nm chip in tight-integration with synaptic matrix and analog
capacitive memory, we march on towards the development of
an enhanced scaled-up HICANN-DLS chip for our large-scale
bio-physically inspired neuromorphic system.
A U T H O R C O N T R I B U T I O N A N D A C K N O W L E D G M E N T
S.A.A. wrote the manuscript, designed the neuron circuits
and implemented the on-chip array. Y.S. performed chip mea-
surements, post tape-out calibration and the network demon-
stration. P.M. devised the pre tape-out calibration and provided
biological benchmarks. C.P. implemented the external FPGA
spike routing. A.H. and A.G. implemented the digital backend
and performed top-level verification. J.S. was the overall
10
9
Synapses
[23] S. Friedmann, J. Schemmel et al., "Demonstrating hybrid learning in a
flexible neuromorphic hardware system," IEEE Trans. BioCAS, vol. 11,
no. 1, pp. 128–142, Feb 2017.
[24] S. Friedmann, "A new approach to learning in neuromorphic hardware,"
Ph.D. dissertation, Ruprecht-Karls-Universitat Heidelberg, 2013.
[25] S. A. Aamir, P. Muller, A. Hartel, J. Schemmel, and K. Meier, "A
highly tunable 65-nm CMOS LIF neuron for a large scale neuromorphic
system," in Proc. ESSCIRC, Sept 2016, pp. 71–74.
[26] M. Hock, "Modern semiconductor technologies for neuromorphic hard-
ware," Ph.D. dissertation, Ruprecht-Karls-Universitat Heidelberg, 2014.
[27] M. Hock, A. Hartel, J. Schemmel, and K. Meier, "An analog dynamic
memory array for neuromorphic hardware," in Proc. ECCTD, Sep. 2013,
pp. 1–4.
[28] C. Bartolozzi et al., Neuromorphic Systems. John Wiley and Sons, Inc.,
2016.
[29] H. Markram, J. Lubke, M. Frotscher, and B. Sakmann, "Regulation of
synaptic efficacy by coincidence of postsynaptic aps." Science, vol. 275,
pp. 213–215, 1997.
J. V. Arthur and K. A. Boahen, "Silicon-neuron design: A dynamical
systems approach," IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 58, no. 5, pp. 1034–1043, 2011.
[30]
[31] R. Jolivet et al., "Integrate-and-fire models with adaptation are good
enough: predicting spike times under random current injection," NIPS,
vol. 18, pp. 595–602, 2006.
[32] W. Gerstner, W. Kistler, R. Naud, and L. Paninski, Neuronal Dynamics.
Cambridge University Press, 2014.
[33] L. Lapicque, "Recherches quantitatives sur l'excitation electrique des
nerfs traitee comme une polarization," Journal de Physiologie et
Pathologie General, vol. 9, pp. 620–635, 1907.
[34] R. Brette and W. Gerstner, "Adaptive exponential integrate-and-fire
model as an effective description of neuronal activity," J. Neurophysiol.,
vol. 94, pp. 3637 – 3642, 2005.
[35] A. Destexhe, "Self-sustained asynchronous
and
Up/Down states in thalamic, cortical and thalamocortical networks of
nonlinear integrate-and-fire neurons." Journal of Computational Neuro-
science, vol. 3, pp. 493 – 506, 2009.
irregular
states
[36] A. Destexhe and D. Contreras, "Neuronal computations with stochastic
network states," Science, vol. 314, no. 5796, pp. 85–90, 2006.
[37] B. Nessler et al., "Bayesian computation emerges in generic cortical
microcircuits through spike-timing-dependent plasticity," PLoS Compu-
tational Biology, vol. 9, no. 4, p. e1003037, 2013.
[38] T. P. Vogels and L. F. Abbott, "Signal propagation and logic gating in
networks of integrate-and-fire neurons," J Neurosci, vol. 25, no. 46, pp.
10 786–95, Nov 2005.
[39] G. Deco and V. K. Jirsa, "Ongoing cortical activity at rest: critical-
ity, multistability, and ghost attractors," The Journal of Neuroscience,
vol. 32, no. 10, pp. 3366–3375, 2012.
[40] R. Naud, N. Marcille, C. Clopath, and W. Gerstner, "Firing patterns in
the adaptive exponential integrate-and-fire model," Biological Cybernet-
ics, vol. 99, no. 4, pp. 335–347, Nov 2008.
[41] M. A. Petrovici, B. Vogginger, P. Muller, O. Breitwieser et al., "Charac-
terization and compensation of network-level anomalies in mixed-signal
neuromorphic modeling platforms," PloS one, vol. 9, no. 10, 2014.
[42] M. A. Petrovici, I. Bytschok, J. Bill, J. Schemmel, and K. Meier, "The
high-conductance state enables neural sampling in networks of LIF
neurons," BMC Neuroscience, vol. 16, no. Suppl 1, p. O2, 2015.
J. Kremkow, L. Perrinet, G. Masson, and A. Aertsen, "Functional
consequences of correlated excitatory and inhibitory conductances in
cortical networks." J Comput Neurosci, vol. 28, pp. 579–594, 2010.
[43]
[44] M. Pospischil, Z. Piwkowska, M. Rudolph, T. Bal, and A. Destexhe,
"Calculating event-triggered average synaptic conductances from the
membrane potential," J. Neurophysiology, vol. 97, p. 2544, 2007.
[45] T. Masquelier and G. Deco, "Network bursting dynamics in excitatory
14
cortical neuron cultures results from the combination of different
adaptive mechanism," PloS one, vol. 8, no. 10, p. e75824, 2013.
[46] S. Millner, "Development of a multi-compartment neuron model emu-
lation," Ph.D. dissertation, Heidelberg University, 2012.
[47] D. Pare et al., "Impact of spontaneous synaptic activity on the resting
properties of cat neocortical pyramidal neurons in vivo," J Neurophysiol,
vol. 79, no. 3, pp. 1450–60, 1998.
[48] A. Destexhe et al., "The high-conductance state of neocortical neurons
in vivo," Nature Rev. Neurosci., vol. 4, pp. 739–751, 2003.
[49] F. Cannillo et al., "Bulk-drain connected load for subthreshold MOS
current-mode logic," Electron. Lett., vol. 43, no. 12, June 2007.
[50] F. Cannillo et al., "Nanopower subthreshold MCML in submicrometer
CMOS technology," IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 56, no. 8, pp. 1598–1611, Aug 2009.
[51] A. Tajalli, E. J. Brauer, Y. Leblebici, and E. Vittoz, "Subthreshold source-
coupled logic circuits for ultra-low-power applications," IEEE J. Solid-
State Circuits, vol. 43, no. 7, pp. 1699–1710, July 2008.
[52] D. K. Jeong, G. Borriello, D. A. Hodges, and R. H. Katz, "Design
of PLL-based clock generation circuits," IEEE J. Solid-State Circuits,
vol. 22, no. 2, pp. 255–261, Apr 1987.
[53] M. Petrovici, J. Bill, I. Bytschok, J. Schemmel, and K. Meier, "Stochastic
inference with spiking neurons in the high-conductance state," Physical
Review E, vol. 94, no. 4, October 2016.
[54] V. Saxena and R. J. Baker, "Indirect compensation techniques for three-
stage fully-differential op-amps," in Proc. MWCAS, Aug 2010, pp. 588–
591.
[55] S. A. Aamir, P. Harikumar, and J. J. Wikner, "Frequency compensation
of high-speed, low-voltage CMOS multistage amplifiers," in Interna-
tional Symposium on Circuits and Systems, May 2013, pp. 381–384.
split-length transistors," in Proc. MWSCAS.
[56] V. Saxena and R. J. Baker, "Compensation of CMOS op-amps using
IEEE, 2008, pp. 109–112.
[57] S. A. Aamir, P. Angelov, and J. J. Wikner, "1.2-v analog interface for a
300-msps HD video digitizer in core 65-nm CMOS," IEEE Transactions
on VLSI Systems, vol. 22, no. 4, pp. 888–898, April 2014.
[59]
[58] D. Goodman and R. Brette, "Brian: a simulator for spiking neural
networks in Python," Front. Neuroinform., vol. 2, no. 5, 2008.
J. Lazzaro, S. Ryckebusch, M. Mahowald, and C. Mead, "Winner-take-
all networks of o (n) complexity," in NIPS, vol. 1, 1988, pp. 703–711.
[60] R. Hahnloser et al., "Digital selection and analogue amplification coexist
in a cortex-inspired silicon circuit," Nature, vol. 405, no. 6789, pp. 947–
951, 2000.
[61] T. Pfeil et al., "Six networks on a universal neuromorphic computing
substrate," Frontiers in Neuroscience, vol. 7, p. 11, 2013.
[62] S. A. Aamir, "Mixed-signal circuit implementation of spiking neuron
models," Ph.D. dissertation, Karlsruhe Institute of Technology, 2018.
[63] S. A. Aamir, P. Muller, L. Kriener, G. Kiene, J. Schemmel, and K. Meier,
"From LIF to AdEx neuron models: accelerated analog 65 nm CMOS
implementation," in Proc. BioCAS, Oct. 2017, pp. 1–4.
[64] S. A. Aamir, P. Muller, G. Kiene, L. Kriener, Y. Stradmann, A. Grubl,
J. Schemmel, and K. Meier, "A mixed-signal structured AdEx neuron
for accelerated neuromorphic cores," IEEE Transactions on Biomedical
Circuits and Systems, 2018.
J. Schemmel, L. Kriener, P. Muller, and K. Meier, "An accelerated analog
neuromorphic hardware system emulating NMDA- and calcium-based
non-linear dendrites," in Proc. IJCNN, May 2017, pp. 2217–2226.
[65]
[66] P. S. Paolucci et al., "Power, energy and speed of embedded and
server multi-cores applied to distributed simulation of spiking neural
networks: ARM in NVIDIA Tegra vs Intel Xeon quad-cores," CoRR,
vol. abs/1505.03015, 2015.
[67] E. Stromatias et al., "Scalable energy-efficient, low-latency implemen-
tations of trained spiking deep belief networks on SpiNNaker," in Proc.
IJCNN, July 2015, pp. 1–8.
|
1012.0490 | 2 | 1012 | 2011-01-10T16:42:46 | Testing of information condensation in a model reverberating spiking neural network | [
"q-bio.NC",
"cs.NE"
] | Information about external world is delivered to the brain in the form of structured in time spike trains. During further processing in higher areas, information is subjected to a certain condensation process, which results in formation of abstract conceptual images of external world, apparently, represented as certain uniform spiking activity partially independent on the input spike trains details. Possible physical mechanism of condensation at the level of individual neuron was discussed recently. In a reverberating spiking neural network, due to this mechanism the dynamics should settle down to the same uniform/periodic activity in response to a set of various inputs. Since the same periodic activity may correspond to different input spike trains, we interpret this as possible candidate for information condensation mechanism in a network. Our purpose is to test this possibility in a network model consisting of five fully connected neurons, particularly, the influence of geometric size of the network, on its ability to condense information. Dynamics of 20 spiking neural networks of different geometric sizes are modelled by means of computer simulation. Each network was propelled into reverberating dynamics by applying various initial input spike trains. We run the dynamics until it becomes periodic. The Shannon's formula is used to calculate the amount of information in any input spike train and in any periodic state found. As a result, we obtain explicit estimate of the degree of information condensation in the networks, and conclude that it depends strongly on the net's geometric size. | q-bio.NC | q-bio |
Testing of information condensation in a model
reverberating spiking neural network1
A. K. Vidybida
Department of Synergetics, Bogolyubov Institute for Theoretical Physics
Metrologichna str., 14-B, 03680 Kyiv, Ukraine
E-mail: [email protected]
Abstract
Information about external world is delivered to the brain in the form of structured in time spike trains.
During further processing in higher areas, information is subjected to a certain condensation process, which
results in formation of abstract conceptual images of external world, apparently, represented as certain uniform
spiking activity partially independent on the input spike trains details. Possible physical mechanism of
condensation at the level of individual neuron was discussed recently.
In a reverberating spiking neural
network, due to this mechanism the dynamics should settle down to the same uniform/periodic activity in
response to a set of various inputs. Since the same periodic activity may correspond to different input spike
trains, we interpret this as possible candidate for information condensation mechanism in a network. Our
purpose is to test this possibility in a network model consisting of five fully connected neurons, particularly,
the influence of geometric size of the network, on its ability to condense information. Dynamics of 20 spiking
neural networks of different geometric sizes are modelled by means of computer simulation. Each network was
propelled into reverberating dynamics by applying various initial input spike trains. We run the dynamics
until it becomes periodic. The Shannon's formula is used to calculate the amount of information in any
input spike train and in any periodic state found. As a result, we obtain explicit estimate of the degree of
information condensation in the networks, and conclude that it depends strongly on the net's geometric size.
1 Introduction
The ability of a biological object to obtain enough
information about external world is essential for
the object's survival. The information is deliv-
ered to the central nervous system through vari-
ous sensory pathways in the form of spike trains.
The pathways' information throughput has been
studyed for a long time in theoretical [22] and ex-
perimental [25] research. The paradigm of those
research is consistent with the assumption that the
best situation is when the brain receives maximum
information from sensory systems. But compare
this with [10]. Another paradigm, which is as well
mature, [23], concentrates on self-organization of
spike trains when primary sensory activity spreads
to higher brain areas.
Self-organization is ac-
companied with information loss, [15], Indeed, a
kind of standartization of activity evoked by var-
ious primary sensory inputs is obsereved exper-
imentally in higher brain areas during olfactory
[6, 38] and auditory [4] perception. In visual sys-
1Accepted in the International Journal of Neural Systems, http://www.worldscinet.com/ijns/ijns.shtml
2
A. Vidybida
tem, simple examples are the transformation of
scene representation to viewpoint-invariant, [7] or
retinotopic-invariant [32] coordinates. Here, spike
trains in the optic nerve must depend on certain
information (retinal position, viewpoint), which
is removed in higher areas of conceptual repre-
sentation.
In the context of cognitive physiol-
ogy, the process of reduction of information aimed
at conceptual representation/recognition of exter-
nal objects is known as information condensation,
[21]. Usually pattern recognition phenomenon,
which is closely related to information condensa-
tion,
is considered in parallel with training, see
[12, 17, 24, 28, 30, 31, 39], learning, [1, 11, 13, 21],
or other plasticity, [16], in the corresponding net-
work. In a biological network, the learning mecha-
nism involves biosynthesis [27] and is therefore very
slow process, which requires seconds or minutes,
[3]. At the same time, recognition of objects in vi-
sual scene can be accomplished within 150 ms, or
faster, citeThorpe1,Thorpe. During this short pe-
riod of time, the network has constant structure,
but it is the spiking activity which evolves within
it from information-rich at the sensory periphery
into information-poor, representing concrete enti-
ties/concepts at higher brain areas.
We now put a question: What could be the
physical mechanism of information condensation in
a spiking neural system? One possible mechanism,
[37] which operates at the level of single neuron,
was discussed, see Fig. 1, 2 and n. 2.1.12.1.1, be-
low.
In a reverberating spiking network, due to
this mechanism the spiking dynamics should set-
tle down to the same definite periodic activity in
response to any stimulus from a definite set of var-
ious inputs, and to another periodic activity in re-
sponse to members of another set of inputs. If so,
then the definite periodic dynamical state can be
considered as an abstract representation of a fea-
ture, which all stimuli from the definite set have in
common. And this is just what is expected from
the condensation of information.
Our purpose in this work is to study how the
ability of a simple spiking neural net to condense
information in the above described sense depends
on its physical parameter -- the net's geomet-
ric size. For this purpose we simulate dynamics
of a net composed of five binding neurons placed
equidistantly on a circle. The circle's radius, R,
characterizes the net's geometric size. The net is
fully connected, and propagation velocity is taken
the same for all connections and all values of R.
Thus, the variations in R are expressed exclusively
in the variations in the interneuronal propagation
times.
Initially, the net is stimulated by a spike
train of input impulses triggering each of five neu-
rons at times {t0, t1, . . . , t4}. Afterwards, the spik-
ing dynamics is allowed to go freely until it settles
down to a periodic one. This allows to figure out
sets of input stimuli bringing about the same pe-
riodic dynamics. The number of different periodic
dynamics and the number of input stimuli in a set,
which corresponds to a definite periodic dynamics,
characterize the net's ability to condense informa-
tion. We found that this ability depends strongly
on the net's size.
2 Methods
2.1 The Binding Neuron Model
The understanding of mechanisms of higher brain
functions expects a continuous reduction from
higher activities to lower ones, eventually, to ac-
tivities in individual neurons, expressed in terms
of membrane potentials and ionic currents. While
this approach is correct scientifically and desirable
for applications, the complete range of the reduc-
tion is unavailable to a single researcher/engineer
due to human brain limited capacity. In this con-
nection, it would be helpful to abstract from the
rules by which a neuron changes its membrane po-
tentials to rules by which the input impulse sig-
nals are processed into its output impulses. The
coincidence detector, and temporal integrator are
the examples of such an abstraction, see discus-
sion by Konig et al., [20]. One more abstraction,
the binding neuron (BN) model, is proposed as sig-
nal processing unit, [34] which can operate as ei-
ther coincidence detector, or temporal integrator,
depending on quantitative characteristics of stim-
ulation applied. This conforms with behavior of
real neurons, see, e.g. work by Rudolph & Des-
texhe, [26]. The BN model describes functioning
of a neuron in terms of discret events, which are
input and output impulses, and degree of tempo-
ral coherence between the input events, see Fig.
1. Mathematically, this can be realized as follows.
Each input impulse is stored in the BN for a fixed
time, τ . The τ is similar to the tolerance inter-
val discussed by MacKay, [23]. All input lines are
excitatory and deliver identical impulses. The neu-
ron fires an output impulse if the number of stored
impulses, Σ, is equal or higher than the threshold
value, N0. In this model, inhibition is expressed in
decreased τ value. It is clear, that BN is triggered
when a bunch of input impulses is received in a
narrow temporal interval. In this case, the bunch
could be considered as compound event, and the
output impulse -- as an abstract representation of
this compound event. One could treat this mech-
anism as binding of individual input events into
a single output event, provided the input events
are coherent in time. Such interpretation is sug-
Testing of Information Condensation
3
elementary
event
elementary
event
...
...
elementary
event
✲
✲
✲
Binding of the
elementary
events based
on their
temporal
coherence
④
elementary event
for secondary neurons
✲
(represents the bound event)
✚✚❃
✲✚
❍
❍❍❥
inhibition controls binding
Figure 1: Signal processing in the binding neuron model.[35, 36]
gested by binding of features/events in largescale
neuronal circuits, [5, 8, 9, 36].
In our case,
It would be interesting to characterize the BN
input-output relations in the form of transfer func-
tion, which allows exact calculation of output in
terms of input.
input is the se-
quence of discrete arriving moments of standard
impulses: Tin = {l1, l2, l3, l4, . . . }. The output is
the sequence of discrete firing moments of BN:
Tout = {f1, f2, . . . }. It is clear that Tout ⊂ Tin. The
transfer function in our case could be the function
σ(l), l ∈ Tin, which equals 1 if l is the firing mo-
ment, l ∈ Tout, and 0 otherwise. For BN with
threshold N0 the required function can be con-
structed as follows. It is clear that the first N0 − 1
input impulses are unable to trigger neuron, there-
fore σ(l1) = 0, . . . , σ(lN0−1) = 0. The next input
is able to trigger if and only if all N0 inputs are
coherent in time:
σ(lN0 ) = 1 if and only if lN0 − l1 ≤ τ.
In order to determine σ(lN0+k), k ≥ 1, one must
take into acount all previous input moments, there-
fore we use notation σTin instead of σ. The values
of σTin (lN0 +k) can be determined recursively:
σTin (lN0+k) = 1 if and only if lN0+k − lk+1 ≤ τ and
σTin (li) = 0 for all i ∈ {k + 1, . . . , N0 + k − 1}.
The function σTin describes completely the BN
model for arbitrary threshold value N0 ≥ 2 .
2.1.1 Information Condensation in a
Single Neuron
It is worth noticing, that any firing (triggering)
moment of a spiking neuron is determined by the
moment of last input impulse, which just ensures
that the triggering condition is satisfied. In a neu-
ron, which needs more than one input impulse to
fire, variations of temporal position of impulses, re-
ceived just before the triggering one, do not influ-
ence the moment of emitting the output impulse,
provided those variations are in resonable limits
and arrival moment of the triggering impulse re-
mains the same.
τ
t25 t24 t23
t22
t21
t15 t14
t13
t12 t11
Figure 2: Example of two different inputs into
a single neuron, which produce identical out-
puts.
Thus, different input spike trains can produce ex-
actly the same output. This looks like if some de-
tailes of the input stimulus, which is composed of
several impulses, were reduced/condensed in the
output.
In the BN, the triggering condition is
that the number of impulses in the BN's inter-
nal memory equals to N0. Consider BN with
N0 = 4, which is stimulated with two different
spike trains with input impulses arrival times S1 =
{t11, t12, t13, t14, t15}, t11 < t12 < t13 < t14 < t15,
and S2 = {t21, t22, t23, t24, t25}, t21 < t22 < t23 <
t24 < t25. Let the arrival moments satisfy the fol-
lowing conditions:
t14 − t11 < τ,
t25 − t22 < τ,
t24 − t21 > τ,
t14 = t25 = to .
4
A. Vidybida
In this case, both S1 and S2, if fed to the BN, will
produce exactly the same output, namely, the sin-
gle impulse at moment to. This is illustrated in
Fig. 2.
2.2 The Network Model
As a reverberating spiking neural net we take the
net of five neurons placed equidistantly at a circle
of radius R, see Fig, 3. Each neuron has threshold
N0 = 4, and internal memory, τ = 10 ms. The
net is fully connected. The connection lines are
characterized with length and propagation veloc-
ity, v, which is the same in all lines. For R fixed,
there are two types of connection line, the short
one, with propagation delay d, and the long one
with propagation delay D. Each neuron has addi-
tionally the external stimulus input line, which is
used to start the net dynamics. Single impulse in
the stimulus input line delivers to its target neuron
just threshold excitation. This causes firing at the
moment of the stimulus impulse arrival. For nu-
merical simulations we use 20 networks of different
sizes, see Table 1. The propagation velocity in any
interconnection line is taken v = 0.1 m/s.
2.2.1 Numerical Simulation
As programming language we use Python under
Linux operating system. The dynamics was mod-
elled by advancing time with step dt = 200 µs.
The delay values d and D, when measured in the
dt units, were rounded to the nearest from below
integers, see Table 1. As a result, the simulating
program operates in whole numbers with no round-
ing errors involved.
Each single tick of the program, the network's
tick, advances time by dt, and consists of three par-
tial ticks, which are performed in the given order.
Namely, (i) input tick, which advances time in the
input lines, (ii) axonal tick, which advances time
in the internal connection lines, (iii) neuronal tick,
which advances time in the neurons.
This manner of updating states can be treated
as synchronous in a sense that each component of
the network has the same physical time when the
network's tick is complete. On the other hand,
viewed as interneuronal communication process,
the dynamics should be treated as asynchronous
due to nonzero propagation delays.
During the step (ii), a neuron can get impulse
into its internal memory. If a neuron appears in the
state "Fire" as a result of the network tick, then
the output impulse it produces can appear in the
connection lines only during the next network tick.
This introduces effective delay of one dt between
delivering the triggering impulse to a neuron and
emitting output impulse by that neuron.
2.3 Data Acquisition Algorithm
2.3.1 Set of Stimuli
The net was entrained to reverberating dynamics
by applying initial input spike train of five im-
pulses, one triggering impulse per neuron, at times
(in dt units) {t0 = 1, t1, t2, t3, t4}. The triggering
moment of neuron # 0 is taken 1 for all stimuli
in order to exclude rotational symmetry between
the stimuli applied. Other four triggering moments
run independently through the set {1, 2, . . . , tmax},
where tmax is choosen proportional to R for each
net size. In choosing tmax, we follow two different
paradigms.
In the first paradigm of short stim-
uli we restrict the overall duration of the stim-
ulus train with the value tmax = d. Thus, the
set of stimuli has d4 different stimuli.
If ti ≤ d,
i = 0, . . . , 4, then any neuron in the net never
obtains impulse from other neurons before it ob-
tains its external input stimulation. In the second
paradigm of extended stimuli, we restrict the over-
all duration of the stimulus train with the value
tmax = M , which is about three times longer than
d for each network (see Table 1). Here, all stim-
uli {t0 = 1, t1, t2, t3, t4}, which were presented to a
network, cover the set of M 4 different trains, which
equals from 625 different stimuli for net #1 to
100 000 000 different stimuli for net #20 (see Table
1). The stimuli were sampled in accordance with
standard algorithm of 4-digit counter. Namely, we
started from stimulus {1, 1, 1, 1, 1}, the next stim-
ulus is obtained by advancing t1 by 1, and so on.
The stimulus next to {1, M, 1, 1, 1} is {1, 1, 2, 1, 1},
the one next to {1, M, M, 1, 1} is {1, 1, 1, 2, 1} and
so on, until stimulus {1, M, M, M, M } is presented.
In the extended paradigm, the late external in-
put impulse can enter corresponding neuron after
it received impulses from neurons already triggered
by early external input impulses.
The second paradigme is in concordance with
visual information processing,
[2] where activity
from higher brain areas, which was invoked due to
visual stimulation at earlier time, is retroinjected
to areas V1 and V2 in the primary visual cortex,
where it interacts with activity invoked by visual
input at later time during perception.
2.3.2 Figuring out Periodic States
After the last input impulse from the train {t0 =
1, t1, t2, t3, t4} reachs its target neuron, the pro-
gram begins appending at each time step the in-
stantaneous state of the net to a Python list. The
instantaneous state consists of states of all 20 con-
nection lines and all 5 neurons (see Fig. 4). Before
appending, the program checks if the current in-
stantaneous state was already included in the list.
Testing of Information Condensation
5
t
2
d
n2
t1
n1
D
n3
t 3
n4
t
4
t0 = 1
n0
Figure 3: The network, used for simulations. Here {t0, t1, t2, t3, t4} -- is the input spike train, d, D -- are the
propagation delays in the connection lines. Any line can be either empty, or propagating one impulse.
net #
R, mm 0.029
1
d, dt
D, dt
M , dt
net #
R, mm 0.314
1
2
5
11
d, dt
D, dt
M , dt
18
29
55
2
3
4
5
6
7
8
9
10
0.057
0.086
0.114
0.143
0.171
0.200
0.229
0.257
0.286
3
5
10
12
5
8
15
13
6
10
20
14
8
13
25
15
10
16
30
16
11
19
35
17
13
21
40
18
15
24
45
19
16
27
50
20
0.343
0.371
0.400
0.429
0.457
0.486
0.514
0.543
0.571
20
32
60
21
35
65
23
38
70
25
40
75
26
43
80
28
46
85
30
48
90
31
51
95
33
54
100
Table 1: Dimensions of networks used for simulations; dt = 200 µs.
net # 1
31
2
61
104
123
net # 11
301
494
12
331
544
3
91
154
187
246
13
361
584
4
111
184
227
296
14
391
634
5
141
234
287
376
15
411
6
171
284
347
456
16
441
7
201
324
407
526
17
471
8
221
364
447
9
251
414
507
10
281
454
567
18
491
19
521
20
551
Table 2: Distinct periods in dt units of found periodic states in the short stimuli paradigm. Superscript denotes
the number of different periodic states with this period.
6
A. Vidybida
If it was, then the periodic dynamical state is found
with its complete cyclic trajectory covered by in-
stantaneous states between the inclusion and the
last record in the list, inclusively. Measures are
taken in order not to count the same cyclic trajec-
tory, which was entrained at its different points, as
different periodic states.
The data for each net were stored in two
MySQL tables. Table STATES included the se-
rial number of any periodic state found, one ele-
ment from the corresponding cyclic trajectory, and
period of the state (see Fig. 4). Single record
in the INPUTS table included the input stimulus
{t0 = 1, t1, t2, t3, t4}, the serial number of the pe-
riodic state it leads to (this number is 0 for fading
dynamics), and relaxation time, namely, the time,
which is spent between the last external input im-
pulse is delivered and the net's entrance moment
into the periodic regime.
3 Results
3.1 Characterization of Periodic
States Found
After initial stimulation, networks from #1 to #7
entrain to periodic activity after any stimulation,
and networks #8 to #20 either entrain to peri-
odic dynamics, or stops from any activity after
some time. This takes place for both short and
extended stimuli.The number of different periodic
states found for each network is shown in Fig. 5.
Here, the maximal number of periodic states ob-
tained with short stimuli is 18, and this number is
achieved for net numbers from 3 to 7. Exact values
of periods, and number of different periodic states
with this period is shown in Table 2 for short stim-
uli. We omit similar table for extended stimuli.
The maximal number of different periodic states
obtained with extended stimuli is 485, which is
achieved in the net number 9. The maximal num-
ber of different periodic states with the same pe-
riod is 294 for period 50·dt in net #9. It should be
mentioned that two periodic states, which can be
turned into eachother by suitable renumeration of
neurons, were considered as different.
It is evident that in the network of five neurons
with threshold 4, each neuron is triggered the same
number of times during period.
Indeed, expect
that neuron n4 fires k4 times, and any of the other
four fires less during period: k4 > ki, i = 0, 1, 2, 3.
In order to be triggered k4 times, n4 must obtain
not less than 4k4 input impulses during period.
But it can obtain only k0 +k1 +k2 +k3, which is less
than required. Similarly, situation when k4 = k3
and k3 > ki, i = 0, 1, 2 leads to contradiction, and
so on.
This number of triggering is either 1 or 2 for
trajectories found, see examples in Fig. 6. Some
nets have only one periodic state, which coresponds
to synchronous firing of all 5 neurons and symmet-
rical states of connection lines at any moment of
time. This is the case for nets number 1 and from
15 to 20 for short stimuli, and for nets number 19
and 20 for extended stimuli.
3.2 Condensation of Information
In order to estimate the degree of information con-
densation in the course of transformation of an ex-
ternal spike train into a certain periodic state of the
net, one needs to calculate information amount,
which is delivered by specifying a spike train, and
which is delivered by specifying the state, it leads
to.
23
0, 1
4
2, 3
0, 4
state #319
50
1
state #107
50
0, 3
1, 2
4
0, 3
1, 2
4
state #199
66
0
0
0
Figure 6: Examples of periodic states found
for net number 9 in the extended stimuli
paradigm. Spikes indicate the firing moments,
labels near each spike give numbers of neurons,
firing at this moment. The two upper trains
show states with period 10 ms, the lower one -
with period 13.2 ms.
This can be done by wellknown Shannon's formula
[29]
H = − X
pi log2 pi,
i
(1)
where pi is the probability to obtain case num-
ber i from a set of cases. At the input end we
have the set of d4, or M 4 different external input
spike trains. In our statement of the problem, it
is natural to consider all external input trains as
equally probable. If so, then information delivered
by specifying certain train is
Hs = 4 log2 d,
for the short stimuli paradigm, and
He = 4 log2 M,
(2)
(3)
for the extended stimuli paradigm.
While estimating information, delivered by
specifying certain periodic state, one should take
into account that probabilities of different periodic
Testing of Information Condensation
7
mysql> select * from STATES_5_9 where num=269;
+------+------------------------------------------------------------------------------
----------------------------------------------+--------+
num state
period
+------+------------------------------------------------------------------------------
----------------------------------------------+--------+
269 0 1 1 0 8 8 17 17 24 15 15 24 8 8 0 0 8 17 17 8
0 False False
0 False False
0 False False
0 False False
0 False True 49 1
82
+------+------------------------------------------------------------------------------
----------------------------------------------+--------+
1 row in set (0.02 sec)
mysql>
Figure 4: Example of single record in the MySQL table STATES. The first field (num) is numerical, and gives
the serial number of periodic state found. The second field (state) is a string, which describes instantaneous
state from the periodic state found (a point from the cyclic trajectory, which represents the whole trajectory, or
periodic state). The first 20 numbers in the string describe states of all connection lines: '0' means that the line
is empty, positive number specifies after how many ticks the propagating impulse will rich the targeted neuron.
The next five chunks confined between the "next line" symbols describe states of neurons. The first number in
each chunk is the "kick" -- the total number of impulses obtained by neuron after the axonal tick was complete.
During the neuronal tick, corresponding to that axonal tick, the "kick" is utilized and set to zero. The next
boolean in the chunk indicates if the neuron is in the "Fire" state. The next boolean indicates if the neuron has
any impulses in its internal memory. If it has, then next couples of numbers (up to three couples) describe those
impulses. In this example, neuron #4 has 1 impulse with time to leave 49·dt. The third field (period) specifies
period (in dt units) of this periodic state.
s
e
t
a
t
s
c
i
d
o
i
r
e
p
f
o
r
e
b
m
u
n
20
15
10
5
0
2
4
6
8
10
net #
12
14
16
18
20
s
e
t
a
t
s
c
i
d
o
i
r
e
p
f
o
r
e
b
m
u
n
500
450
400
350
300
250
200
150
100
50
0
2
4
6
8
10
net #
12
14
16
18
20
Figure 5: Number of different periodic states found for each net with short (left panel), and extended (right
panel) stimuli.
8
A. Vidybida
states are not the same. In order to calculate prob-
ability pn of a periodic state Cn, we calculate the
number of input spike trains leading to the Cn,
namely, Tn, and divide this number by the total
amount of different input stimuli (see histogram
for Tn in Fig. 7)
pn =
Tn
d4 ,
for the short stimuli paradigm, and
pn =
Tn
M 4 ,
(4)
(5)
for the extended stimuli paradigm. Then we use
Eq.
(1) with probabilities of individual periodic
states found in accordance with (4), (5) to cal-
culate information which should be ascribed to
any periodic state.
In this calculations, we treat
uniformly with others the external input stimuli,
which lead to fading dynamics. Correspondingly,
the state with no activity is treated uniformly with
periodic states. This is in the contrast with data
presented in Fig. 5, and Table 2, where the state
with no activity is excluded.
n
i
b
e
n
o
n
i
s
n
i
a
m
o
d
f
o
r
e
b
m
u
n
90
80
70
60
50
40
30
20
10
0
s
e
t
a
t
s
c
i
d
o
i
r
e
p
f
o
n
o
i
t
a
m
r
o
f
n
i
8
7
6
5
4
3
2
1
0
e
s
2
4
6
8
10
net #
12
14
16
18
20
Figure 8: Dependence of information amount,
which is ascribed to periodic state, on the net
size. Curve 'e' corresponds to extended stimuli
paradigm, 's' -- to the short stimuli one.
information of a periodic state, calculated in accor-
dance with Eq. (1) with probabilities found due to
Eqs.
(4), (5), varies between 6.93 and 7.33 bits
for extended stimuli, and between 3.17 and 3.46
for short stimuli. In the plateau, the degree of in-
formation condensation calculated as input infor-
mation divided by the periodic state information,
varies between 9.02 and 12.27 for extended stim-
uli, and between 11.2 and 19.31 for short stimuli.
Out of a plateau range, for greater net sizes the
amount of information in a periodic state drops
sharply due to simplification of the set of peri-
odic states. Namely, the total number of periodic
states decreases to one, with the second one with
no activity, and the probability is distributed very
unevenly between this two states. This leads to
the degree of information condensation as high as
41000 for extended stimuli, and 690 for short ones.
0
10000
20000
30000
40000
size of conceptual domain
4 Discussion
Figure 7: Histogram of conceptual domain
sizes for net #9 in extended stimuli paradigm.
The bin size is 518. 23 domains with sizes from
50764 to 193732 are not presented.
As it could be expected,
the information
amount in an external input spike train increases
as logarithm of the net size, in accordance with
Eqs. (2), (3), varying from 25 to 81 bits for short
paradigm, and from 37 to 106 bits for extended
paradigm.
Information, which could be ascribed
to a periodic state, depends on the net size in a
more complicated manner, see Fig. 8. A remark-
able feature is a kind of plateau between net #3
and #9 for both short, and extended stimulation
paradigme. In the plateau, the
Any reverberating spiking neural net can repre-
sent complicated dynamical behavior. If the net's
instantaneous states can be described with whole
numbers, then the net will inevitably either en-
train to periodic dynamics, or stop its activity at
all. In this study, it is appeared that a very simple
net of Fig. 3 can be engaged into a considerably
large set of different periodic activities. It is not
clear which part of all possible in this net periodic
states was discovered in our simulation. As it fol-
lows from comparison between short and extended
stimuli paradigm, the number of different periodic
states found increases with increasing range of in-
put stimuli. Certainly, this increase must saturate
somewhere. This is because any two different peri-
odic regimes are represented by their cyclic trajec-
tories, which has no common points (instantaneous
states). On the other hand, the total number of
instantaneous states the network can have is finite
Testing of Information Condensation
9
due to finiteness of the set of states of each element
the network is composed of.
The number of periodic states found in a net
depends on the net's geometric size. Variations
in the net's size display themselves exclusively in
variations of interneuronal propagation delays d
and D. On the other hand, the duration of neu-
ronal internal memory, τ , is the same for net of any
size. Thus, it is namely the relationships between
the times an impulse spends for travelling between
neurons, and time it is allowed to spend in a neu-
ron waiting for additional impulses, which controls
possible number of periodic states.
It is worth noticing that each net has one com-
pletely synchronized periodic state. The com-
pletely synchronozed state is stable and achieved
during finite time. This is in the contrast to the
case of pulse-coupled oscillators with delayed exci-
tatory coupling, (see, e.g. Ref. [40]).
In the four-dimensional set of stimuli we used,
the neighbouring stimuli differ from eachother by
one dt in one of four dimensions. This can be
treated as analogous representation of some reality.
The set of periodic states should be considered as
a set of discrete entities due to qualitative differ-
ence between any two states. This conforms with
a paradigm discussed in cognitive physiology, [21].
The process of transformation of initial analogous
inputs into a discret set of periodic states implies a
loss of information and can be treated as conden-
sation of information.
If we take a set of input stimuli, any of which
leads to the same periodic state, then that periodic
state can be considered as an abstract/conceptual
representation of a feature, which all stimuli from
the set have in common, and the corresponding set
could be named as "conceptual domain". What
kind of feature or concept does the conceptual do-
main represent? If our net was trained to recognize
a certain real feature, then it would be that feature.
In the context of this study, the common feature is
that all stimuli from the conceptual domain engage
namely this net into namely this periodic dynam-
ics.
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
42
0
349
2
8
125
42
262
273
42
0
349
300
8
125
42
262
273
41
256
42
8
172 182
371
42
125
0
133
262
273
270
201
41
40
92
35
133
270
178
178
269
268
267
266
265
264
221
220
221
264
265
266
267
268
269
270
35
270
270
203
212
202 92
259
201
40
259
260
259
92
256
42
125
42
8
172 182
371
41 133
262
273
356
354
352
350
92
41
256 348
133
270
178
35
35
270
178
269
268
267
266
265
264
221
220
221
264
265
266
267
268
269
270
256
178
270
228
45
132
256
177
178
170
270
228
270
178 181 186 190 193 276 275 274 240 239
Figure 9: 2-dimensional cross-section of the 4-
dimensional space of inputs for network #9 in
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
358
295
2
295
330
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
300
302
301
297
301
295
295
0
extended stimuli paradigm.
left, cross-section
is made with plane (t3 = 23, t4 = 23), right,
(t3 = 21, t4 = 23). Origin for (t1, t2) is in
the upper left corner. Both t1 and t2 run
through the set {1, 2, . . . , 45} of values. Num-
bers in polygons indicate serial numbers of cor-
responding periodic states.
It would be interesting to have a look on
the topology of the conceptual domains in the 4-
dimensional space of all stimuli. For this pur-
pose we figured out 2-dimensional cross-section
of the input stimuli (see Fig. 9).
In the cross-
section, a typical conceptual domain is composed
of several coherent clusters disconnected with ea-
chother. The histogram of sizes of conceptual do-
mains found is given in Fig. 7.
Decomposition of the whole set of input stimuli
into a number of conceptual domains represented
by corresponding periodic dynamics resembles ap-
proach in analyses of multivariate datasets, see re-
view in [14]. The difference is that here the 4-
dimensional dataset (a conceptual domain) is rep-
resented by unidimensional cyclic trajectory, which
corresponds to the domain, and the trajectory is
composed of points/vectors, which have other di-
mension than the dataset vectors (see Fig.
4).
Nevertheless, having the network, the whole cyclic
trajectory can be reproduced starting from its any
single point. Thus, here the datasets are reduced
down to individual points. This is in concordance
with the information condensation idea.
Why do we stick ourselves with namely the pe-
riodic dynamics? The answer is related to the
memory/learning problem, even if we do not con-
sider any plasticity in this study. It is known [3]
that modification of synaptic strength may hap-
pen due to repetitive delivery of impulses to those
synapses. Periodic dynamical states are just well
suited for such repetitive delivery. All other dy-
namical behaviors are of transient type, and have
less chances to cause plastic changes in biological
network. On the other hand, successful perception
expects ability to report about what was perceived,
which is impossible without memory.
Unfortunately, we cannot draw this biological
analogy too far. Real biological network includes
not only excitatory, but about 4% of inhibitory
neurons, [18]. Representing this characteristic in
a model network requires to have at least 25 neu-
rons in it. This dramatically increases the number
of possible external stimuli, which requires a quali-
tative change in our approach as regards the speed
of simulations and analysis of obtained data.
Finally, what happens if we use another neu-
ronal model in the network? Our opinion is that re-
sults will be qualitatively similar. Using the bind-
ing neuron here is natural, since it represents in
10
A. Vidybida
refined form what a spiking neuron does with sig-
nals it receives. Additionally, the BN model easily
allows to develop a program operating in whole
numbers. This excludes possible dynamical arte-
facts due to rounding errors.
In future work, it would be interesting to com-
pare results,
if another spiking neuron model is
used in the network, to study the topology of
conceptual domains and how the topology could
change if a plasticity is introduced in the network
model.
5 Conclusions
A network composed of spiking neurons is able to
condense information due to the fact that differ-
ent initial stimuli could lead the network to the
same periodic dynamics. This happens by means
of initial/basic condensation of information in spik-
ing neurons, as it is described in n. 2.1.12.1.1,
above. The network's geometric size, which deter-
mines the interneuronal transmission delays, has
considerable influence on the net's ability to con-
dense information, mainly due to influence on the
number of different periodic states the network can
have, see Fig. 5. The latter has influence on the
amount of information, which should be ascribed
to a single periodic state, see Fig. 8. As a re-
sult, the degree of information condensation varies
between 9 and 41 000, see n. 3.23.2 for details.
The networks considered here are too primi-
tive to have reliable biological implications. At the
same time, numerical parameters, see Table 1 and
n. 2.22.2, such as network sizes and spike propaga-
tion velocity, are taken corresponding to biological
data. The threshold value 4 does not contradict
to biological reality, as experimentally registered
thresholds are between 1 and 300. The BN inter-
nal memory duration, τ , is commensurable with
halfdecay time of the excitatory postsynaptic po-
tentials (EPSP). Thus, in the framework of this
extremely simple model, one could expect that the
ability of biological neural network to condense in-
formation should depend on its geometric size, or
on the relationships between interneuronal trans-
mission delays and the EPSP halfdecay time.
Acknowledgments This work was supported
by the Program of basic research of the National
Academy of Science of Ukraine.
Content of this work was partially published
in an abstract form in the abstract book of
the 2nd International Biophysics Congress and
Biotechnology at GAP & 21th National Biophysics
Congress, (5-9 Oct.
2009) Diyarbakır, Turkey,
http://www.ibc2009.org/
References
[1] Acharya, R., Chua, E.C.P., Chua, K.C.,
Min, L.C., and Tamura, T. (2010), Anal-
ysis and Automatic Identification of Sleep
Stages using Higher Order Spectra, Interna-
tional Journal of Neural Systems, 20:6.
[2] J. Bullier, "Integrated model of visual pro-
cessing," Brain Res. Rev., 36, 96 -- 107 (2001).
[3] D. V. Buonomano and M. M. Merzenich,
"Cortical plasticity:
to
maps," Annu. Rev. Neurosci., 21, 149 -- 186
(1998).
from synapses
[4] P. Cariani, "Temporal codes, timing nets,
and music perception," J. New Music Res.,
30, 107 -- 135 (2001).
[5] A. R. Damasio, "The brain binds entities and
events by multiregional activation from con-
vergence zones," Neural Comput., 1, 123 --
132 (1989).
[6] P. Duchamp-Viret and A. Duchamp, "Odor
processing in the frog olfactory system,"
Prog. Neurobiol., 53, 561 -- 602 (1997).
[7] R. Durbin and G. Mitchison, "A dimension
reduction framework for understanding cor-
tical maps," Nature, 343, 644 -- 647 (1990).
[8] R. Eckhorn, R. Bauer, W. Jordan, M.
Brosch, W. Kruse, M. Munk and H. J. Reit-
boeck, "Coherent oscillations: a mechanism
for feature linking in the visual cortex?,"
Biol. Cybern., 60, 121 -- 130 (1988).
[9] A. K. Engel, P. Konig, A. K. Kreiter, C.
M. Gray and W. Singer, "Temporal coding
by coherent oscillations as a potential so-
lution to the binding problem: physiolog-
ical evidence," In Schuster, H.G., Singer,
W. (ed), Nonlinear Dynamics and Neuronal
Networks, 3 -- 25. VCH Weinheim (1991).
[10] J. Feldman, "Ecological expected utility and
the mythical neural code," Cognitive Neuro-
dynamics, 4, 25 -- 35 (2010).
[11] Ghosh-Dastidar, S. and Adeli, H. (2007), Im-
proved Spiking Neural Networks for EEG
Classification and Epilepsy and Seizure De-
tection, Integrated Computer-Aided Engi-
neering, Vol. 14, No. 3, pp. 187-212.
[12] S. Ghosh-Dastidar and H. Adeli, "Spiking
neural networks," Intern. J. Neural Sys. 19,
295 -- 308 (2009).
[13] S. Ghosh-Dastidar and H. Adeli, A New
Supervised Learning Algorithm for Multiple
Spiking Neural Networks with Application in
Epilepsy and Seizure Detection, Neural Net-
works 22, 1419 -- 1431 (2009).
Testing of Information Condensation
11
[14] A. Gorban and A. Zinovyev, "Principal man-
ifolds and graphs in practice: from molecu-
lar biology to dynamical systems," Intern. J.
Neural Sys., 20, 219 -- 232 (2010).
[15] H. Haken, "Brain Dynamics. Synchroniza-
tion and Activity Patterns in Pulse-Coupled
Neural Nets with Delays and Noise,"
Springer, Berlin 2007.
[16] J. Iglesias and A. E. P. Villa, "Emergence
of preferred firing sequences in large spik-
ing neural networks during simulated neu-
ronal development," Intern. J. Neural Sys.,
18, 267 -- 277 (2008).
[17] Johnston, S.P., Prasad, G.,Maguire, L. and
McGinnity, T.M. (2010), An FPGA Hard-
ware/software co-design methodology - to-
wards evolvable spiking networks for robotics
application, Intern. J. Neural Sys., 20:6.
[18] E. R. Kandel, J. H. Schwartz and T. M. Jes-
sell, "Principles of Neural Science," (Fourth
ed.), New York: McGraw-Hill (2000).
[19] H. Kirchner and S. J. Thorpe, "Ultra-rapid
object detection with saccadic eye move-
ments: visual processing speed revisited,"
Vision Res., 46, 1762 -- 1776 (2006).
[20] P. Konig, A. K. Engel and W. Singer, "Inte-
grator or coincidence detector? The role of
the cortical neuron revisited," Trends Neu-
rosci., 19, 130 -- 137 (1996).
[21] P. Konig and N. Kruger, "Symbols as self-
emergent entities in an optimization process
of feature extraction and predictions," Biol.
Cybern., 94(4), 325 -- 334 (2006).
[22] D. M. MacKay and W. S. McCulloch, "The
limiting information capacity of a neuronal
link," Bull. Math. Biophys., 14, 127 -- 135,
(1952).
[23] D. M. MacKay, "Self-organization in the
time domain," In M. C. Yovitts, G. T. Jacobi
and G. D. Goldstein (ed), Self-Organizing
Systems, Spartan Books, Washington, 37 -- 48
(1962).
[24] Nichols, E., McDaid, L.J., and Siddique,
N.H. (2010), Case Study on Self-organizing
Spiking Neural Networks for Robot Naviga-
tion, Intern. J. Neural Sys., 20:6.
[25] Ch. L. Passaglia and J. B. Troy, "Informa-
tion transmission rates of cat retinal gan-
glion cells," J. Neurophysiol., 91, 1217 -- 1229
(2004).
[26] M. Rudolph and A. Destexhe, "Tuning neo-
cortical pyramidal neurons between integra-
tors and coincidence detectors, " J. Comput.
Neurosci., 14, 239 -- 251 (2003).
[27] M. T. Scharf, N. H. Woo, K. M. Lattal, Z.
Young, P. V. Nguyen and T. Abel, "Protein
synthesis is required for the enhancement of
long-term potentiation and long-term mem-
ory by spaced training," J. Neurophysiol., 87
2770 -- 2777 (2002).
[28] Schliebs, S., Kasabv, N., and Defoin-Platel,
M. (2010), On the Probabilistic Optimiza-
tion of Spiking Neural Networks, Intern. J.
Neural Sys., 20:6.
[29] C. E. Shannon, "A mathematical theory of
communication," Bell System Technical J.,
27, 379 -- 423 and 623 -- 656, July and October,
(1948).
[30] Soltic and S. Kasabov, N. (2010), Knowledge
extraction from evolving spiking neural net-
works with rank order population coding ,
Intern. J. Neural Sys., 20:6.
[31] Strain, T.J., McDaid, L.J., Maguire, L.P.,
and T.M. McGinnity, T.M.
(2010), An
STDP Training Algorithm for a Spiking Neu-
ral Network with Dynamic Threshold Neu-
rons, Intern. J. Neural Sys., 20:6.
[32] S. Tang, R. Wolf, S. Xu and M. Heisenberg,
"Visual pattern recognition in Drosophila is
invariant for retinal position," Science, 305,
1020 -- 1022 (2004).
[33] S. Thorpe, D. Fize and C. Marlot, "Speed
of processing in the human visual system,"
Nature, 381, 520 -- 522 (1996).
[34] A. K. Vidybida, "Neuron as time coherence
discriminator," Biol. Cybern., 74, 539 -- 544
(1996).
[35] A. K. Vidybida,
"Information process-
ing in a pyramidal-type neuron," Proc.
BioNet'96
Biologieorientierte Informatik
und pulspropagierende Netze, 3-d Workshop.
Ed.: Heinz G., pp. 96 -- 99. (1996).
[36] A. K. Vidybida, "Inhibition as binding con-
troller at the single neuron level," BioSys-
tems, 48, 263 -- 267 (1998).
[37] A. K. Vidybida, "Output stream of binding
neuron with instantaneous feedback," Eur.
Phys. J. B, 65, 577 -- 584 (2008); Eur. Phys.
J. B, 69, 313 (2009).
[38] M. Wehr and G. Laurent, "Odor encoding
by temporal sequences of firing in oscillat-
ing neural assemblies" Nature, 384, 162 -- 166
(1996).
[39] B. Widrow, "Generalization and information
storage in networks of adaline 'neurons' " in:
Yovitz,M.C., Jacobi,G.T., Goldstein,G.(ed),
Self-Organizing Systems, 435 -- 461 (1962).
12
A. Vidybida
[40] W. Wu and T. Chen,
"Impossibility
of asymptotic synchronization for pulse-
coupled oscillators with delayed excitatory
coupling," Intern. J. Neural Sys., 19, 425 --
435 (2009).
s
n
i
a
m
o
d
f
o
r
e
b
m
u
n
5
4
3
2
1
0
1
10
100
1000
10000
100000
1e+06
size of conceptual domain
|
0905.2935 | 3 | 0905 | 2010-01-18T02:39:52 | Experimental demonstration of associative memory with memristive neural networks | [
"q-bio.NC",
"cond-mat.mes-hall",
"q-bio.MN"
] | When someone mentions the name of a known person we immediately recall her face and possibly many other traits. This is because we possess the so-called associative memory, that is the ability to correlate different memories to the same fact or event. Associative memory is such a fundamental and encompassing human ability (and not just human) that the network of neurons in our brain must perform it quite easily. The question is then whether electronic neural networks (electronic schemes that act somewhat similarly to human brains) can be built to perform this type of function. Although the field of neural networks has developed for many years, a key element, namely the synapses between adjacent neurons, has been lacking a satisfactory electronic representation. The reason for this is that a passive circuit element able to reproduce the synapse behaviour needs to remember its past dynamical history, store a continuous set of states, and be "plastic" according to the pre-synaptic and post-synaptic neuronal activity. Here we show that all this can be accomplished by a memory-resistor (memristor for short). In particular, by using simple and inexpensive off-the-shelf components we have built a memristor emulator which realizes all required synaptic properties. Most importantly, we have demonstrated experimentally the formation of associative memory in a simple neural network consisting of three electronic neurons connected by two memristor-emulator synapses. This experimental demonstration opens up new possibilities in the understanding of neural processes using memory devices, an important step forward to reproduce complex learning, adaptive and spontaneous behaviour with electronic neural networks. | q-bio.NC | q-bio | Experimental demonstration of associative memory
with memristive neural networks
Yuriy V. Pershin and Massimiliano Di Ventra
1
0
1
0
2
n
a
J
8
1
]
.
C
N
o
i
b
-
q
[
3
v
5
3
9
2
.
5
0
9
0
:
v
i
X
r
a
Abstract—Synapses are essential elements for computation and
information storage in both real and artificial neural systems. An
artificial synapse needs to remember its past dynamical history,
store a continuous set of states, and be ”plastic” according to the
pre-synaptic and post-synaptic neuronal activity. Here we show
that all this can be accomplished by a memory-resistor (memristor
for short). In particular, by using simple and inexpensive off-
the-shelf components we have built a memristor emulator which
realizes all required synaptic properties. Most importantly, we
have demonstrated experimentally the formation of associative
memory in a simple neural network consisting of three electronic
neurons connected by two memristor-emulator synapses. This
experimental demonstration opens up new possibilities in the
understanding of neural processes using memory devices, an
important step forward to reproduce complex learning, adaptive
and spontaneous behavior with electronic neural networks.
Index Terms—Memory, Resistance, Neural network hardware,
Neural networks.
I. INTRODUCTION
W HEN someone mentions the name of a known person
we immediately recall her face and possibly many
other traits. This is because we possess the so-called asso-
ciative memory - the ability to correlate different memories to
the same fact or event [1]. This fundamental property is not
just limited to humans but it is shared by many species in the
animal kingdom. Arguably the most famous example of this
are experiments conducted on dogs by Pavlov [2] whereby
salivation of the dog’s mouth is first set by the sight of food.
Then, if the sight of food is accompanied by a sound (e.g., the
tone of a bell) over a certain period of time, the dog learns to
associate the sound to the food, and salivation can be triggered
by the sound alone, without the intervention of vision.
Since associative memory can be induced in animals and
we, humans, use it extensively in our daily lives, the network
of neurons in our brains must execute it very easily. It is
then natural to think that such behavior can be reproduced
in artificial neural networks as well - a first important step
in obtaining artificial
intelligence. The idea is indeed not
novel and models of neural networks have been suggested
over the years that could theoretically perform such function
[3], [4], [5], [6]. However,
their experimental realization,
especially in the electronic domain, has remained somewhat
difficult. The reason is that an electronic circuit that simulates
a neural network capable of associative memory needs two
important components: neurons and synapses, namely connec-
tions between neurons. Ideally, both components should be of
nanoscale dimensions and consume/dissipate little energy so
that a scale-up of such circuit to the number density of a typical
human brain (consisting of about 1010 synapses/cm2) could be
feasible. While one could envision an electronic version of the
first component relatively easily, an electronic synapse is not
so straightforward to make. The reason is that the latter needs
to be flexible (”plastic”) according to the type of signal it
receives, its strength has to depend on the dynamical history
of the system, and it needs to store a continuous set of values
(analog element).
In the past, several approaches with different levels of ab-
straction were used in order to implement electronic analogues
of synapses [7]. For instance, one of the first ideas involved
the use of three-terminal electrochemical elements controlled
by electroplating [8]. While some of these approaches do not
involve synaptic plasticity at all [9], [10], [11], [12], the latter
is generally implemented using a digital (or a combination of
analog and digital) hardware [13], [14], [15], [16], [17], [18],
[19]. The common feature of synaptic plasticity realizations
is the involvement of many different circuit elements (such as
transistors) and, therefore, occupation of a significant amount
of space on a VLSI chip. Thus, the amount of electronic
synapses in present VLSI implementations is much lower than
the amount of synapses relevant to actual biological systems.
Novel, radically different approaches to resolve this issue
would be thus desirable.
A recently demonstrated resistor with memory (memristor
Input 1
“ i h
d”
“sight of food”
f f
N
1
1
Input 2
p
“sound”
N
2
S
S
1
1
S
S
2
N
N
3
Output
“salivation”
Yu. V. Pershin is with the Department of Physics and Astronomy and USC
Nanocenter, University of South Carolina, Columbia, SC, 29208
e-mail: [email protected].
M. Di Ventra is with the Department of Physics, University of California,
San Diego, La Jolla, California 92093-0319
e-mail: [email protected].
Manuscript received November XX, 2009; revised January YY, 2010.
Fig. 1. Artificial neural network for associative memory. Real neurons and
their networks are very complex systems whose behavior is not yet fully
understood. However, some simple brain functions can be elucidated studying
significantly simplified structures. Here, we consider three neurons (N1, N2
and N3) coupled by two memristive synapses (S1 and S2). The output signal
is determined by input signals and strengths of synaptic connections which
can be modified when learning takes place.
for short [20], [21]) based on TiO2 thin films [22], [23] offers a
promising realization of a synapse whose size can be as small
as 30×30×2 nm3. Using TiO2 memristors, a fabrication of
neuromorphic chips with a synapse density close to that of the
human brain may become possible. Memristors belong to the
larger class of memory-circuit elements (which includes also
memcapacitors and meminductors) [24], [25], namely circuit
elements whose response depends on the whole dynamical
history of the system. Memristors can be realized in many
ways, ranging from oxide thin films [22], [23], [26] to spin
memristive systems [27]. However, all these realizations are
limited to the specific material or physical property respon-
sible for memory, and as such they do not easily allow for
tuning of the parameters necessary to implement the different
functionalities of electronic neural networks.
In the present paper, we describe a flexible platform al-
lowing for simulation of different types of memristors, and
experimentally show that a memristor could indeed function as
a synapse. We have developed electronic versions of neurons
and synapses whose behavior can be easily tuned to the func-
tions found in biological neural cells. Of equal importance,
the electronic neurons and synapses were fabricated using
inexpensive off-the-shelf electronic components resulting in
few dollars cost for each element, and therefore can be realized
in any electronic laboratory. Clearly, we do not expect that
with such elements one can scale up the resulting electronic
neural networks to the actual brain density. However, due to
their simplicity reasonably complex neural networks can be
constructed from the two elemental blocks developed here and
we thus expect several functionalities could be realized and
studied.
For the purpose of this paper we have built the neural
network shown in Fig. 1. We have then shown that such
circuit is capable of associative memory. In this network,
two input neurons are connected with an output neuron by
means of synapses. As an example of the functionality that
this network can provide, we can think about the animal
memory we have described above [2] in which the first input
neuron (presumably located in the visual cortex) activates
under a specific visual event, such as ”sight of food”, and
the second input neuron (presumably located in the auditory
cortex) activates under an external auditory event, such as a
particular ”sound”. Depending on previous training, each of
these events can trigger ”salivation” (firing of the third output
neuron). If, at a certain moment of time, only the ”sight of
food” leads to ”salivation,” and subsequently the circuit is
subjected to both input events, then, after a sufficient number
of simultaneous input events the circuit starts associating the
”sound” with the ”sight of food”, and eventually begins to
”salivate” upon the activation of the ”sound” only. This process
of learning is a realization of the famous Hebbian rule stating,
in a simplified form, that ”neurons that fire together, wire
together”.
II. RESULTS AND DISCUSSION
A. Electronic neuron
Biological neurons deal with two types of electrical signals:
receptor (or synaptic) potentials and action potentials. A stimu-
2
a
V
IN+
ADC
I
I
V
IN-
NC
O
O
-2.5V
+2.5V
Microcontroller
b
)
V
(
e
g
a
t
l
o
V
5
4
3
2
1
0
-1
-2
-3
V
out
V
in
0.0
0.5
1.0
1.5
2.0
2.5
3.0
Time (s)
Electronic neuron. a, Main components of the electronic neuron
Fig. 2.
proposed here are an analog-to-digital converter (ADC) and a microcontroller.
NC means not connected. If the voltage value Vin on the input terminal
(I) exceeds a threshold voltage VT (in our experiments VT = 1.5V), the
microcontroller connects the input pin to -2.5V and output pin (O) to 2.5V
for 10ms, thus sending forward and backward pulses. After that, it waits for
a certain amount of time δt and everything repeats again. If Vin < VT ,
then the microcontroller just continuously samples Vin. The waiting time
was selected as δt = τ − γ · (Vin − VT ) + λ (η − 0.5), where τ = 60ms,
γ = 50ms/V, λ = 10ms and η is a random number between 0 and 1. In
our experimental realization, we used microcontroller dsPIC30F2011 from
Microchip with internal 12bits ADC and a possibility of pin multiplexing, so
that the only additional elements were two 10k resistors. Actual measurements
were done using 5V voltage span, with further assignment of the middle
level as zero. The value of resistors shown in a is 10kΩ. b, Response of the
electronic neuron on input voltages of different magnitude, the red dashed
line is the threshold voltage and is only a guide to the eye. When Vin < VT ,
no firing occurs. When Vin > VT , the electronic neuron sends pulses, with
the average pulse separation decreasing with increasing Vin. Vout is shifted
for clarity.
lus to a receptor causes receptor potentials whose amplitude is
determined by the stimulus strength. When a receptor potential
exceeding a threshold value reaches a neuron, the latter starts
emitting action potential pulses, whose amplitude is constant
but their frequency depends on the stimulus strength. The
action potentials are mainly generated along axons in the
forward direction. However, there is also a back-propagating
part of the signal [28], [29] which is now believed to be
responsible for synaptic modifications (or learning) [29], [30].
We have implemented the above behaviour in our electronic
scheme as shown in Fig. 2a using an analog-to-digital con-
verter and a microcontroller. The input voltage is constantly
monitored and once it exceeds a threshold voltage, both
forward and backward pulses are generated whose amplitude
is constant (we set
it here for convenience at 2.5V), but
pulse separation varies according to the amplitude of the input
signal. In Fig. 2b we show the response of the electronic neu-
ron when three resistors of different values are subsequently
connected between the input of the neuron and 2.5V. When
the resulting voltage (determined by the external resistor and
internal resistor connected to the ground in Fig. 2a) is below
the threshold voltage (t < 1s in Fig. 2b), no ”firing” occurs (no
change in the output voltage). When the input voltage exceeds
the threshold (t > 1s in Fig. 2b), the electronic neuron sends
forward- and back-propagating pulses. The pulse separation
decreases with increase of the input voltage amplitude as it is
evident in Fig. 2b.
B. Electronic synapse
As electronic synapse we have built a memristor emulator,
namely an electronic scheme which simulates the behaviour
of any memory-resistor. In fact, our memristor emulator can
reproduce the behaviour of any voltage- or current-controlled
memristive system. The latter is described by the following
relations
y (t) = g (x, y, t) u (t) ,
x = f (x, u, t) ,
(1)
(2)
where y(t) and u(t) are input and output variables, such
as voltage and current, g(x, u, t) is a generalized response
(memresistance R, or memductance G), x is a n-dimensional
vector describing the internal state of the device, and f (x, u, t)
is a continuous n-dimensional vector function [21], [25].
As Fig. 3a illustrates schematically, our memristor emulator
consists of the following units: a digital potentiometer, an
analog-to-digital converter and a microcontroller. The A (or
B) terminal and the Wiper of the digital potentiometer serve
as the external connections of the memristor emulator. The
resistance of the digital potentiometer is determined by a code
written into it by the microcontroller. The code is calculated by
the microcontroller according to Eqs. (1) and (2). The analog-
to-digital converter provides the value of voltage applied to
the memristor emulator needed for the digital potentiometer
code calculation. The applied voltage can be later converted
to the current since the microcontroller knows the value of the
digital potentiometer resistance.
In our experiments, we implemented a threshold model
of voltage-controlled memristor previously suggested in our
earlier paper [31] and also discussed in Ref. [25]. In this
model, the following equations (which are a particular case
of Eqs. (1,2)) were used:
G = x−1,
x = (βVM + 0.5 (α − β) [VM + VT − VM − VT ])
(3)
×θ (x − R1) θ (R2 − x) ,
(4)
where θ(·) is the step function, α and β characterize the
rate of memristance change at VM ≤ VT and VM > VT ,
respectively, VM is a voltage drop on memristor, VT is a
V
VM
b
AA
W
B
a
0.2Hz
1Hz
5Hz
0.2
0.1
0.0
)
A
m
(
I
3
VIN+
ADC
VIN-
Microcontroller
Microcontroller
R
0
+
V(t)
-
R
0.5
1.0
-0.1
-0.2
-1.0
-0.5
0.0
V
M (V)
Fig. 3. Electronic synapse. a, Schematic of the main units of the memristor
emulator. The memristor emulator consists of a digital potentiometer, ADC
and microcontroller. The digital potentiometer unit represents an element
whose resistance is defined by a digital code written in it. Two terminals
of this unit (A and W) are the external connection terminals of the memristor
emulator. The differential ADC converts the voltage between A and W
terminals of the digital potentiometer into a digital value. The microcontroller
reads the digital code from ADC and generates (and writes) a code for
the digital potentiometer according to predefined functions g(x, u, t) and
f (x, u, t) and Eqs. (1-2). These operations are performed continuously. In
our circuit, we used a 256 positions 10kΩ digital potentiometer AD5206
from Analog Device and microcontroller dsPIC30F2011 from Microchip
with internal 12bits ADC. b, Measurements of memristor emulator response
when V (t) = V0 cos(2πωt) with V0 ≃ 2V amplitude is applied to the
circuit shown in the inset with R0 = 10kΩ. The following parameters
determining the memristor emulator response (see Eqs. (3,4)) were used:
α = β = 146kΩ/(V·s), VT = 4V, R1 = 675Ω, R2 = 10kΩ. We noticed
that the initial value of RM (in the present case equal to 10kΩ) does not
affect the long-time limit of the I-V curves. The signals were recorded using
a custom data acquisition system.
threshold voltage and R1 and R2 are limiting values of
memristance. In Eq. (4), the θ-functions symbolically show
that the memristance can change only between R1 and R2. In
the actual software implementation, the value of x is monitored
at each time step and in the situations when x < R1 or
x > R2, it is set equal to R1 or R2, respectively. In this way,
we avoid situations when x may overshoot the limiting values
by some amount and thus not change any longer because of
the step function in Eq. (4). In simple words, the memristance
changes between R1 and R2 with different rates α and β
below and above the threshold voltage. This activation-type
model was inspired by recent experimental results on thin-
film memristors [23] and as we discuss below it reproduces
synapse plasticity.
18 a
18 b
16
14
12
)
10
V
(
e
g
a
t
l
o
V
8
6
4
2
0
Input 1
"sight of food"
Input 2
"sound"
V
3, in
Output
"salivation"
Probing
Learning Probing
Input 1
Input 2
V
3, in
Output
Learning
16
14
12
)
10
V
(
e
g
a
t
l
o
V
8
6
4
2
0
-2
-2
0
3
6
9
12
Time (s)
15
9.0
9.5
10.0
10.5
11.0
Time (s)
4
11.5
12.0
Fig. 4. Development of associative memory. a, Using electronic neurons and electronic synapses (memristor emulators), we have built an electronic circuit
corresponding to the neural network shown in Fig. 1. We have used the following parameters defining the operation of memristor emulators: VT = 4V,
α = 0, β = 15kΩ/(V·s). At the initial moment of time, the resistance of S1 was selected equal to R1 = 675Ω (lowest resistance state of memristor) and
the resistance of S2 was selected equal to R2 = 10kΩ (highest resistance state of memristor). Correspondingly, in the first probing phase, when Input 1 and
Input 2 signals are applied without overlapping, the output signal develops only when Input 1 signal is applied. In the learning phase (see also b for more
detailed picture), Input 1 and Input 2 signals are applied simultaneously. According to Hebbian rule, simultaneous firing of input neurons leads to development
of a connection, which, in our case, is a transition of the second synapse S2 from high- to low-resistance state. This transition is clearly seen as a growth
of certain pulses in the signal V3,in (voltage at the input of third neuron) in the time interval from 10.25s to 11s in b. In the subsequent probing phase, we
observe that firing of any input neuron results in the firing of output neuron, and thus an associative memory realization has been achieved. The curves in a
and b were displaced for clarity.
To test that our emulator does indeed behave as a memristor,
we have used the circuit shown in the inset of Fig. 3b, in
which an ac voltage is applied to the memristor emulator,
R, connected in series with a resistor R0 which was used
to determine the current. The obtained current-voltage (I-
V) curves, presented in Fig. 3b, demonstrate typical features
of memristive systems. For instance, all curves are pinched
hysteresis loops passing through (0,0) demonstrating no energy
storage property of memristive systems [21], [24], [25]. More-
over, the frequency dependence of the curve is also typical for
memristive systems: the hysteresis shrinks at low frequencies,
when the system has enough time to adjust its state to varying
voltage, and at higher frequencies, when the characteristic
timescale of system variables change is longer than the period
of voltage oscillations.
C. Associative memory
Using the electronic neurons and electronic synapses de-
scribed above, we have built an electronic scheme correspond-
ing to the neural network depicted in Fig. 1. In this scheme, we
directly connect the memristor terminals to the third neuron
input. In such configuration, our network behaves essentially
as a linear perceptron [32], [33], [34], although different
connection schemes are possible. For example, putting a
capacitor between the ground and input of the third neuron,
we would obtain an integrate-and-fire model [35]. In such a
circuit (and also in perceptron networks with many synapses)
it is important to ensure that the current from a synapse
does not spread to its neighbors. This can be achieved by
placing diodes between the right terminals of synapses in
Fig. 1 and the third neuron’s input. For our neural network
containing only two synapses the effect of current spreading
between synapses is not important. Moreover, we would like
to highlight that our neural network is fully asynchronous, in
distinction to a scheme suggested by Snider [30] based on a
global clock. This makes our approach free of synchronization
issues when scaling up and closer to a bio-inspired circuit.
An asynchronous memristor-based network was also discussed
recently [36].
Fig. 4 demonstrates the associative memory development in
the present network. Our experiment consists in application
of stimulus signals to the first (”sight of food”) and second
(”sound”) neurons, and monitoring of the output signal on the
third (”salivation”) neuron. We start from a state when the
first synaptic connection is strong (low resistance state of the
first memristor) and second synaptic connection is weak (high
resistance state of the second memristor).
In the first ”probing phase” (t < 9s, Fig. 4) we apply
separate non-overlapping stimulus signals to the ”sight of
food” and ”sound” neurons. This results in the ”salivation”
neuron firing when a stimulus signal is applied to the ”sight of
food” neuron, but not firing when a stimulus signal is applied
to the ”sound” neuron. Electronically, it occurs because pulses
generated by the ”sight of food” neuron exceed the threshold
voltage of the ”salivation” neuron (due to a low resistance
of the first memristor synapse) while the voltage on the
”salivation” neuron input due to the ”sound” neuron pulses
is below the threshold voltage of the ”salivation” neuron.
In this phase there is no memristor state change since the
first memristor is already in its limiting state (with minimal
resistance allowed) and its resistance cannot decrease further,
and voltage drop on the second memristor is below its voltage
threshold.
In the ”learning phase” (9s< t <12s, Fig. 4), stimulus
voltages are applied simultaneously to both input neurons, thus
generating trains of pulses. The pulses from different neurons
are uncorrelated, but sometimes they do overlap, owing to a
random component in the pulse separation (see Fig. 2 caption
for details). During this phase, in some moments of time,
back-propagating pulses from the ”salivation” neuron (due
to excitation from the ”sight of food” neuron) overlap with
forward propagating pulses from the ”sound” neuron causing
a high voltage across the second memristor synapse. As this
voltage exceeds the memristor threshold, the second synapse
state changes and it switches into a low resistance state. It
is important to note that this change is possible when both
stimuli are applied together (in other words they correlate).
As a result, an association between input stimuli develops and
the circuit ”learns” to associate the ”sight of food” signal to
the ”sound” signal.
Our measurements during the second probing phase (t >
12s, Fig. 4) clearly demonstrate the developed association. It
is obvious that, in this phase, any type of stimulus - whether
from the ”sight of food” or from the ”sound” neurons - results
in the ”salivation” neuron firing.
Mention should be made about the memristor evolution
function given by Eq. 4. In the present neural network the
resistance of the second synapse can only decrease, since only
negative or zero voltages can be applied to the memristor. A
process of synapse depression (corresponding to an increased
synapse resistance in our electronic network) can be easily
taken into account in different ways. The simplest way is to
add a small positive constant to the first line of Eq. 4 (in
the parentheses) for the conditioned memristor (S2 in Fig. 1).
Then, even at zero value of VM , the resistance of the memristor
will slowly increase. Another way is to apply a small positive
bias to the conditioned memristor (S2) in the circuit without
any changes in Eq. 4 (this would also require taking a non-zero
α).
III. CONCLUSION
We have shown that the electronic (memristive) synapses
and neurons we have built can represent important function-
alities of their biological counterparts, and when combined
5
together in networks - specifically the one represented in
Fig. 1 of this work - they give rise to an important function
of the brain, namely associative memory. It is worth again
mentioning that, although other memristors (e.g., those built
from oxide thin films [37]) could replace the emulator we
have built,
the electronic neurons and synapses proposed
here are electronic schemes that can be built from off-the-
shelf inexpensive components. Together with their extreme
flexibility in representing essentially any programmable set of
operations, they are ideal to study much more complex neural
networks [38] that could adapt to incoming signals and ”take
decisions” based on correlations between different memories.
This opens up a whole new set of possibilities in reproducing
different types of learning and possibly even more complex
neural processes.
ACKNOWLEDGMENT
M.D. acknowledges partial support from the National Sci-
ence Foundation (DMR-0802830).
REFERENCES
[1] J. R. Anderson, Language, memory, and thought. Hillsdale, NJ: Erlbaum,
1976.
[2] I. P. Pavlov, Conditioned Reflexes: An Investigation of the Physiological
Activity of the Cerebral Cortex (translated by G. V. Anrep). London:
Oxford University Press, 1927.
[3] J. J. Hopfield, ”Neural networks and physical systems with emergent
collective computational properties,” Proc. Nat. Acad. Sci. (USA), vol.
79, pp. 2554-2558, 1982.
[4] K. Gurney, An Introduction to Neural Networks. UCL Press (Taylor &
Francis group), 1997.
[5] L. N. Cooper, ”Memories and memory: a physicist’s approach to the
brain,” Int. J. Modern Physics A, vol. 15, pp. 4069-4082, 2000.
[6] T. Munakata, Fundamentals of the New Artificial Intelligence: Neural,
Evolutionary, Fuzzy and More, 2nd ed. Springer, 2008.
[7] N. Lewis and S. Renaud, ”Spiking neural networks ”in silico”: from
single neurons to large scale networks”, Systems, Signal and Devices
Conference, SSD 2007, Hammamet, Tunisia, March 19-22 2007.
[8] B. Widrow, “Rate of adaptation of control systems”, ARS Journal, pp.
1378-1385 (1962).
[9] M. Mahowald, and R. Douglas, A silicon neuron, Nature, vol.354, pp.
515-518, 1991.
[10] R. Jung, E. J. Brauer, and J.!J. Abbas Real-time interaction between
a neuromorphic electronic circuit and the spinal cord, IEEE Trans. on
Neural Systems and Rehab. Eng., vol. 9, pp.319326, 2001.
[11] G. LeMasson, S. Renaud, D. Debay, and T. Bal, Feedback inhibition
controls spike transfer in hybrid thalamic circuits, Nature, vol. 4178, pp.
854-858, 2002.
[12] M. Sorensen, S. DeWeerth, G. Cymbalyuk, R. L. Calabrese, Using a
hybrid neural system to reveal regulation of neuronal network activity by
an intrinsic current, J. Neurosci., vol.24, pp. 5427-5438, 2004.
[13] R.J. Vogelstein, U. Malik, G. Cauwenberghs, Silicon spikebased synaptic
transceiver, Proceedings of ISCAS04, vol.5,
array and address-event
pp.385-388, 2004.
[14] B. Glackin, T.M. McGinnity, L.P. Maguire, QX Wu, A. Belatreche,
A novel approach for the implementation of large scale spiking neural
networks on FPGA hardware, Proc. IWANN 2005 Computational Intellin-
gence and Bioinspired Systems, pp.552-563, Barcelona, Spain, June 2005.
[15] J. Schemmel, K. Meier, E. Mueller, A new VLSI model of neural mi-
crocircuits including spike time dependent plasticity, Proc. ESANN2004,
pp. 405-410, 2004.
[16] G. Indiveri, E. Chicca, and R. Douglas. ”A VLSI array of low-
power spiking neurons and bistable synapses with spike-timing dependent
plasticity”, IEEE Transactions on Neural Networks, vol. 17, pp. 211-221,
2006.
[17] J. V. Arthur and K. Boahen, ”Learning in Silicon: Timing is Every-
thing”, Advances in Neural Information Processing Systems, vol. 18, Eds.
B. Sholkopf and Y. Weiss, MIT Press, 2006.
6
[18] A. Bofill and A. F. Murray. ”Circuits for VLSI implementation of
temporally asymmetric Hebbian learning”, Advances in Neural Informa-
tion processing systems, vol. 14, Eds. T. G. Dietterich, S. Becker, and
Z. Ghahramani, editors, MIT Press, Cambridge, MA, 2001.
[19] B. Linares-Barranco, E. Snchez-Sinencio, A. Rodrguez-Vzquez, and
J. L. Huertas, ”A CMOS Analog Adaptive BAM with On-Chip Learning
and Weight Refreshing”, IEEE Trans. Neural Networks, vol. 4, pp. 445-
455, 1993.
[20] L. O. Chua, ”Memristor - The Missing Circuit Element,” IEEE Trans.
Circuit Theory, vol. 18, pp. 507-519, 1971.
[21] L. O. Chua and S. M. Kang, ”Memrisive devices and systems,” Proc.
IEEE, vol. 64, pp. 209-223, 1976.
[22] D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, ”The
missing memristor found,” Nature (London), vol. 453, pp. 80-83, 2008.
[23] J. J. Yang, M. D. Pickett, X. Li, D. A. A. Ohlberg, D. R. Stewart, and
R. S. Williams, ”Memristive switching mechanism for metal/oxide/metal
nanodevices,” Nature Nanotechnology, vol. 3, pp. 429-433, 2008.
[24] M. Di Ventra, Yu. V. Pershin, and L. Chua, ”Putting memory into circuit
elements: memristors, memcapacitors and meminductors,” Proc. IEEE,
vol. 97, pp. 1371-1372, 2009.
[25] M. Di Ventra, Yu. V. Pershin, and L. Chua, ”Circuit elements with
memory: memristors, memcapacitors and meminductors,” Proc. IEEE,
vol. 97, pp. 1717-1724, 2009.
[26] T. Driscoll, H.-T. Kim, B.-G. Chae, M. Di Ventra, and D. N. Basov,
”Phase-transition driven memristive system,” Appl. Phys. Lett. vol. 95,
pp. 043503/1-3, 2009.
[27] Yu. V. Pershin and M. Di Ventra, ”Spin memristive systems: Spin
memory effects in semiconductor spintronics,” Phys. Rev. B, Condens.
Matter, vol. 78, pp. 113309/1-4, 2008.
[28] G. Stuart, M. Spruston, B. Sakmann, and M. Husser, ”Action potential
initiation and backpropagation in neurons of the mammalian CNS,”
Trends Neurosci., vol. 20, pp. 125-131, 1997.
[29] J. Waters, A. Schaefer, and B. Sakmann, ”Backpropagating action
potentials in neurons: measurement, mechanisms and potential functions,”
Prog. Biophys. Mol. Biol., vol. 87, pp. 145-170, 2005.
[30] G. S. Snider, ”Cortical Computing with Memristive Nanodevices,”
SciDAC Review, vol. 10, pp. 58-65, 2008.
[31] Yu. V. Pershin, S. La Fontaine and M. Di Ventra, ”Memristive model
of amoeba’s learning,” Phys. Rev. E, vol. 80, pp. 021926/1-6, 2009.
[32] R. Rojas, ”Neural Networks - A Systematic Introduction” (Springer-
Verlag, Berlin), 1996.
[33] F. Rosenblatt, ”The Perceptron: A Probabilistic Model for Information
Storage and Organization in the Brain”, Psych. Rev., vol. 65, pp. 386-408,
1958.
[34] M. L. Minsky and S. A. Papert, ”Perceptrons” (Cambridge, MA: MIT
Press), 1969.
[35] W. Gerstner and W. Kistler, ”Spiking neuron models” (Cambridge
university press), 2002.
[36] B. Linares-Barranco1 and T. Serrano-Gotarredona,
”Memristance
can explain Spike-Time-Dependent-Plasticity in Neural Synapses”,
http://hdl.handle.net/10101/npre.2009.3010.1, 2009.
[37] J. Borghetti, Z. Li, J. Straznicky, X. Li, D. A. A. Ohlberg, W. Wu,
D. R. Stewart, and R. S. Williams, ”A hybrid nanomemristor/transistor
logic circuit capable of self-programming,” PNAS, vol. 106, pp. 1699-
1703, 2009.
[38] K. Likharev, A. Mayr, I. Muckra, and O. Turel, ”CrossNets: High-
Performance Neuromorphic Architectures for CMOL Circuits,” Ann. NY
Acad. Sci., vol. 1006, pp. 146-163, 2003.
|
1808.10315 | 2 | 1808 | 2018-09-04T08:12:34 | Deep Learning for Quality Control of Subcortical Brain 3D Shape Models | [
"q-bio.NC"
] | We present several deep learning models for assessing the morphometric fidelity of deep grey matter region models extracted from brain MRI. We test three different convolutional neural net architectures (VGGNet, ResNet and Inception) over 2D maps of geometric features. Further, we present a novel geometry feature augmentation technique based on a parametric spherical mapping. Finally, we present an approach for model decision visualization, allowing human raters to see the areas of subcortical shapes most likely to be deemed of failing quality by the machine. Our training data is comprised of 5200 subjects from the ENIGMA Schizophrenia MRI cohorts, and our test dataset contains 1500 subjects from the ENIGMA Major Depressive Disorder cohorts. Our final models reduce human rater time by 46-70%. ResNet outperforms VGGNet and Inception for all of our predictive tasks. | q-bio.NC | q-bio |
Deep Learning for Quality Control of
Subcortical Brain 3D Shape Models
The ENIGMA Consortium
The full author list appears at the end of the paper
Abstract. We present several deep learning models for assessing the
morphometric fidelity of deep grey matter region models extracted from
brain MRI. We test three different convolutional neural net architectures
(VGGNet, ResNet and Inception) over 2D maps of geometric features.
Further, we present a novel geometry feature augmentation technique
based on parametric spherical mapping. Finally, we present an approach
for model decision visualization, allowing human raters to see the areas of
subcortical shapes most likely to be deemed of failing quality by the ma-
chine. Our training data is comprised of 5200 subjects from the ENIGMA
Schizophrenia MRI cohorts, and our test dataset contains 1500 subjects
from the ENIGMA Major Depressive Disorder cohorts. Our final models
reduce human rater time by 46-70%. ResNet outperforms VGGNet and
Inception for all of our predictive tasks.
Keywords: deep learning, subcortical shape analysis, quality checking
1
Introduction
Quality control (QC) has become one of the main practical bottlenecks in big-
data neuroimaging. Reducing human rater time via predictive modeling and
automated quality control is bound to play an increasingly important role in
maintaining and hastening the pace of scientific discovery in this field. Recently,
the UK Biobank publicly released over 10,000 brain MRIs (and planning to
release 90,000 more); as other biobanking initiatives scale up and follow suit,
automated QC becomes crucial.
In this paper, we investigate the viability of deep convolutional neural nets
for automatically labeling deep brain regional geometry models of failing qual-
ity after their extraction from brain MR images. We compare the performance
of VGGNet, ResNet and Inception architectures, investigate the robustness of
probability thresholds, and visualize decisions made by the trained neural nets.
Our data consists of neuroimaging cohorts from the ENIGMA Schizophrenia and
Major Depressive Disorder working groups participating in the ENIGMA-Shape
project [1]. Using ENIGMAs shape analysis protocol and rater-labeled shapes,
we train a discriminative model to separate FAIL(F) and PASS(P) cases. Fea-
tures are derived from standard vertex-wise measures.
For all seven deep brain structures considered, we are able to reduce human
rater time by 46 to 70 percent in out-of-sample validation, while maintaining
FAIL recall rates similar to human inter-rater reliability. Our models generalize
across datasets and disease samples. Our models' decision visualization, partic-
ularly ResNet, appears to capture structural abnormalities of the poor quality
data that correspond to human raters' intuition.
With this paper, we also release to the community the feature generation
code based on FreeSurfer outputs, as well as pre-trained models and code for
model decision visualization.
2 Methods
Our goal in using deep learning (DL) for automated QC differs somewhat from
most predictive modeling problems. Typical two-class discriminative solutions
seek to balance misclassification rates of each class. In the case of QC, we focus
primarily on correctly identifying FAIL cases, by far the smaller of the two
classes (Table 1). In this first effort to automate shape QC, we do not attempt
to eliminate human involvement, but simply to reduce it by focusing human
rater time on a smaller subsample of the data containing nearly all the failing
cases.
2.1 MRI processing and shape features
Our deep brain structure shape measures are computed using a previously de-
scribed pipeline [2] [3], available via the ENIGMA Shape package. Briefly, struc-
tural MR images are parcellated into cortical and subcortical regions using
FreeSurfer. Among the 19 cohorts participating in this study, FreeSurfer ver-
sions 5.1 and 5.3 were used. The binary region of interest (ROI) images are then
surfaced with triangle meshes and spherically registered to a common region-
specific template [4]. This leads to a one-to-one surface correspondence across
the dataset at roughly 2,500 vertices per ROI. Our ROIs include the left and
right thalamus, caudate, putamen, pallidum, hippocampus, amygdala, and nu-
cleus accumbens. Each vertex p of mesh model M is endowed with two shape
descriptors:
Medial Thickness, D(p) = (cid:107)cp− p(cid:107), where cp is the point on the medial curve
c closest to p.
mapping, J : Tφ(p)Mt → TpM.
LogJac(p), Log of the Jacobian determinant J arising from the template
Since the ENIGMA surface atlas is in symmetric correspondence, i.e., the left
and right shapes are vertex-wise symmetrically registered, we can combine the
two hemispheres for each region for the purposes of predictive modeling. Though
we assume no hemispheric bias in QC failure, we effectively double our sample.
The vertex-wise features above are augmented with their volume-normalized
counterparts: {D, J}normed(p) =
} . Given discrete area elements of the
3At(p)J(p)D(p).
template at vertex p, At(p), we estimate volume as V = (cid:80)
, 2
3
{D,J}(p)
V
{ 1
3
We normalize our features subject-wise by this volume estimate to control for
subcortical structure size.
p∈vrts(M)
2.2 Human quality rating
Human-rated quality control of shape models is performed following the ENIGMA-
Shape QC protocol (enigma.usc.edu/ongoing/enigma-shape-analysis). Briefly,
raters are provided with several snapshots of each region model as well as its
placement in several anatomical MR slices. A guide with examples of FAIL
(QC=1) and PASS (QC=3) cases is provided to raters, with an additional cat-
egory of MODERATE PASS (QC=2) suggested for inexperienced raters. With
sufficient experience, the rater typically switches to the binary FAIL/PASS rat-
ing. In this work, all QC=2 cases are treated as PASS cases, consistent with
ENIGMA shape studies.
2.3 Feature mapping to 2D images
Because our data resides on irregular mesh vertices, we first interpolate the fea-
tures from an irregular spherical mesh onto an equiangular grid. The interpolated
feature maps are then treated as regular 2D images by Mercator projection. Our
map is based on the medial curve-based global orientation function (see [5]),
which defines the latitude (θ) coordinate, as well as a rotational standardization
of the thickness profile D(p) to normalize the longitudinal (φ) coordinate. The
resulting map normalizes the 2D appearance of D(p), setting the poles to lie
at the ends of the medial curve. In practice, the re-sampling is realized as ma-
trix multiplication based on trilinear mesh interpolation, resulting in a 128×128
image for each measure.
2.4 Data augmentation
Although our raw sample of roughly 13,500 examples is exceptionally large by the
standards of neuroimaging, this dataset may not be large enough to train gen-
eralizable CNNs. Standard image augmentation techniques, e.g. cropping and
rotations, are inapplicable to our data. To augment our sample of spherically
mapped shape features, we sample from a distribution of spherical deforma-
tions, i.e. changes in the spherical coordinates of the thickness and Jacobian
features. To do this, we first sample from a uniform distribution of vector spher-
ical harmonic coefficients Blm, Clm, and apply a heat kernel operator [4] to the
generated field on T S2. Change in spherical coordinates is then defined based
on the tangential projection of the vector field, as in [4]. The width σ of the
heat kernel defines the level of smoothness of the resulting deformation, and
the maximum point norm M defines the magnitude. In practice, each random
sampling is a composition of a large magnitude, smooth deformation (σ = 10−1,
M = 3 × 10−1) and a smaller noisier deformation (σ = 10−2, M = 3 × 10−2).
Once the deformation is generated, it is applied to the spherical coordinates of
the irregular mesh, and a new sampling matrix is generated, as above.
2.5 Deep learning models
We train VGGNet [6], ResNet [7] and Inception [8] architectures on our data. We
chose these architectures as they perform well in traditional image classification
problems and are well-studied.
2.6 Model decision visualization
Deep learning models tend to learn superficial statistical patterns rather than
high-level global or abstract concepts [9]. As we plan to provide a tool that both
(1) classifies morphometric shapes, and (2) allows a user to visualize what the
machine perceives as a 'FAIL', model decision visualization is an important part
of our work. Here, we use Prediction Difference Analysis [10], and Grad-CAM
[11] to visualize 'bad' and 'good' areas for each particular shape in question.
2.7 Predictive model assessment
We use two sets of measures to evaluate the performance of our models. To
assess the validity of the models' estimated 'FAIL' probabilities, we calculate
the area under the ROC curve (ROC AUC). We also use two supplementary
measures: FAIL-recall and FAIL-share. In describing them below, we use the
following definitions. TF stands for TRUE FAIL, FF stands for FALSE FAIL,
TP stands for TRUE PASS, and FP stands for FALSE PASS. Our first measure,
F-recall = T F
T F +F P , shows the proportion of FAILS that are correctly labeled
by the predictive model with given probability threshold. The second measure,
F-share =
Number of observations , shows the proportion of the test sample labeled
as FAIL by the model. Ideal models produce minimal F-share, and an F-recall
of 1 for a given set of parameters.
T F +F F
3 Experiments
For each of the seven ROIs, we performed three experiments defined by three
DL models (VGGNet-, ResNet- and Inception-like architecture).
3.1 Datasets
Our experimental data from the ENIGMA working groups is described in Table
1. Our predictive models were trained using 15 cohorts totaling 5218 subjects'
subcortical shape models from the ENIGMA-Schizophrenia working group. For
a complete overview of ENIGMA-SCZ projects and cohort details, see [12].
To test our final models, we used data from 4 cohorts in the Major Depres-
sive disorder working group (ENIGMA-MDD), totaling 1509 subjects, for final
out-of-fold testing. A detailed description of the ENIGMA-MDD sites and its
research objectives may be found here [13].
Train mean±std 3.4±4.7
max
size
16.4
10431
Test mean±std 4.7±4.5
FAIL % accumbens caudate hippocampus thalamus putamen pallidum amygdala
0.8±1.0 0.6±0.6 2.3±3.6 0.9±0.9
3.4
2.6
10436
10436
1.4±1.5 0.4±0.8 1.9±2.0 0.8±0.9
3.5
3018
0.9±0.7 2.0±1.1
2.1
4.2
10433
10436
1.4±1.5 4.9±4.8
3.5
3018
1.5
10436
13.8
10435
10.5
3017
max
size
11.4
3018
1.6
3017
3.8
3018
2.1
3018
Table 1: Overview of FAIL percentage mean, standard deviation and maximum
for each site. Minimum is equal to 0 for all regions and sites except for hippocam-
pus on train (FAIL percentage 5%). Sample sizes for each ROI vary slightly due
to FreeSurfer segmentation failure.
3.2 Model validation
All experiments were performed separately for each ROI. The training dataset
was split into two parts referred to as 'TRAIN GRID' (90% of train data) and
'TRAIN EVAL.' (10% of the data). The two parts contained data from each
ENIGMA-SCZ cohort, stratified by the cohort-specific portion of FAIL cases.
Each model was trained on 'TRAIN GRID' using the original sampling ma-
trix and 30 augmentation matrices resulting in 31x augmented train dataset.
We also generated 31 instances of each mesh validation set using each sampling
matrix and validated models' ROC AUC on this big validation set during the
training.
As models produce probability estimates of FAIL (PF AIL), we studied the
robustness of the probability thresholds for each model. To do so, we selected
PF AIL values corresponding to regularly spaced percentiles values of F-share,
from 0.1 to 0.9 in 0.1 increments. For each such value, we examined F-recall
the evaluation set.
Final thresholds were selected based on the lowest F-share on the TRAIN
EVAL set, requiring that F-recall ≥ 0.8, a minimal estimate of inter-rater re-
liability. It is important to stress that while we used sample distribution infor-
mation in selecting a threshold, the final out-of-sample prediction is made on an
individual basis for each mesh.
4 Results
Trained models were deliberately set to use a loose threshold for FAIL detection,
predicting 0.2-0.5 of observations as FAILs in the 'TRAIN EVAL' sample. These
predicted FAIL observations contained 0.85-0.9 of all true FAILs, promising to
reduce the human rater QC time by 50-80%. These results largely generalized to
the test samples: Table 2 shows our best model and the threshold performance
for each ROI. When applied to the test dataset, the models indicated modest
over-fitting, with the amount of human effort reduced by 46-70%, while cap-
turing 76-94% of poor quality meshes. The inverse relationship between FAIL
percentage and F-share (Figure 1) may indicate model failure to learn gener-
alizable features on smaller number of FAIL examples. ROC AUC and F-recall
performance generalize across the test sites. Since 68% of our test dataset is
comprised of the Munster cohort, it is important that overall test performance
is not skewed by it.
ROI
Model
Eval
AUC
Test
AUC
Eval
Test
Eval
Test
F-share
F-share
F-recall
F-recall
ResNet 0.8
ResNet 0.9
0.3
Accumbens ResNet 0.86 0.8
0.75 0.5
Amygdala
0.84 0.2
Caudate
Hippocampus ResNet 0.85 0.93 0.3
ResNet 0.86 0.91 0.3
Pallidum
ResNet 0.88 0.7
Putamen
0.3
0.87 0.4
Thalamus
ResNet 0.8
0.35
0.54
0.3
0.36
0.32
0.52
0.47
0.83
0.92
0.82
0.81
0.81
0.93
0.82
0.78
0.8
0.78
0.92
0.91
0.76
0.94
Table 2: Test performance of the best models for each region. ResNet performs
the best in all cases. Overall models' performance generalizes to out-of-sample
test data.
Our experiments with decision visualization (see Fig. 2) indicate that in most
FAIL cases, the attention heat map generated by Grad-CAM corresponds to
human raters' intuition while Prediction Difference Analysis tend to concentrate
on local 'bumps' on shapes.
5 Discussion and Conclusion
We have presented potential deep learning solutions for semi-automated quality
control of deep brain structure shape data. We believe this is the first DL ap-
proach for detecting end-of-the-pipeline feature failures in deep brain structure
geometry. We showed that DL can robustly reduce human visual QC time by
46-70% for large-scale analyses, for all seven regions in question, across diverse
MRI datasets and populations. Qualitative analysis of models decisions shows
promise as a potential training and heuristic validation tool for human raters.
There are several limitations of our work. Our planar projection of vertex-
wise features introduces space-varying distortions and boundary effects that can
affect training, performance and visualization. Recently proposed spherical con-
volutional neural nets [14] may be useful to fix this issue. Second, our models'
Fig. 1: Scatter plots of F-recall vs. ROC AUC on test datasets and F-share vs.
proportion of predicted FAIL cases on test datasets (F-share). Left: F-recall
vs ROC AUC. Right: Fail F-share vs FAIL percentage. F-share was calculated
based on thresholds from Table 2. Mark size shows the dataset size. Mark shape
represents dataset (site): (cid:13) - CODE-Berlin (N=176); (cid:3) - Munster(N=1033); (cid:52)
- Stanford (N=105); (cid:53) - Houston(N=195)
.
Fig. 2: QC report for human raters (left) and decision visualization example
based on Grad-CAM for the ResNet model (right). Red colors correspond to
points maximizing the model's FAIL decision in the last layer. Decision visu-
alization corresponds to the observable deviations from underlying anatomical
boundaries indicative of a "FAIL" rating according to an experienced rater.
.
decision visualization only partly matches with human raters' intuition. In some
cases, our models do not consider primary "failure" areas, as assessed by a human
rater. Finally, our models are trained on purely geometrical features and do not
include information on shape boundaries inside the brain. In rare cases, human
0.60.70.80.91.0F-recall60708090100ROC AUCAccumbensAmygdalaCaudateHippocampusPallidumPutamenThalamus0.00.20.40.60.8F-share024681012FAIL %AccumbensAmygdalaCaudateHippocampusPallidumPutamenThalamus raters pass shapes with atypical geometry because their boundaries look reason-
able, and conversely mark normal-appearing geometry as failing due to poor a
fit with the MR image. Incorporating intensity as well as geometry features will
be the focus of our future work.
References
1. Gutman, B., Ching, C., Andreassen, O., Schmaal, L., Veltman, D., Van Erp, T.,
Turner, J., Thompson, P.M., et al.: Harmonized large-scale anatomical shape anal-
ysis: Mapping subcortical differences across the ENIGMA Bipolar, Schizophrenia,
and Major Depression working groups. Biological Psychiatry 81(10) (2017) S308
2. Gutman, B.A., Jahanshad, N., Ching, C.R., Wang, Y., Kochunov, P.V., Nichols,
T.E., Thompson, P.M.: Medial demons registration localizes the degree of ge-
netic influence over subcortical shape variability: An n= 1480 meta-analysis. In:
Biomedical Imaging (ISBI), 2015 IEEE 12th International Symposium on, IEEE
1402 -- 1406
3. Roshchupkin*, G.V., Gutman*, B.A., et al.: Heritability of the shape of subcortical
brain structures in the general population. Nature Communications 7 (2016) 13738
4. Gutman, B.A., Madsen, S.K., Toga, A.W., Thompson, P.M.: 24. In: A Family of
Fast Spherical Registration Algorithms for Cortical Shapes. Volume 8159 of Lecture
Notes in Computer Science. Springer International Publishing (2013) 246 -- 257
5. Gutman, B.A., Yalin, W., Rajagopalan, P., Toga, A.W., Thompson, P.M.: Shape
matching with medial curves and 1-d group-wise registration. In: Biomedical Imag-
ing (ISBI), 2012 9th IEEE International Symposium on. 716 -- 719
6. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556 (2014)
7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: Proceedings of the IEEE conference on computer vision and pattern recognition.
(2016) 770 -- 778
8. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.A.: Inception-v4, inception-resnet
and the impact of residual connections on learning. In: AAAI. Volume 4. (2017)
12
9. Jo, J., Bengio, Y.: Measuring the tendency of cnns to learn surface statistical
regularities. arXiv preprint arXiv:1711.11561 (2017)
10. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network
decisions: Prediction difference analysis. arXiv preprint arXiv:1702.04595 (2017)
11. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-
cam: Visual explanations from deep networks via gradient-based localization. See
https://arxiv. org/abs/1610.02391 v3 7(8) (2016)
12. van Erp, T.G.M., Hibar, D.P., et al.: Subcortical brain volume abnormalities in
2028 individuals with schizophrenia and 2540 healthy controls via the ENIGMA
consortium. Molecular psychiatry (2015)
13. Schmaal, L., Hibar, D.P., et al.: Cortical abnormalities in adults and adolescents
with major depression based on brain scans from 20 cohorts worldwide in the
enigma major depressive disorder working group. Mol Psychiatry 22(6) (2017)
900 -- 909
14. Cohen, T.S., Geiger, M., Koehler, J., Welling, M.: Spherical cnns. arXiv preprint
arXiv:1801.10130 (2018)
Full author list
Dmitry Petrov1,2, Boris A. Gutman2,3, Egor Kuznetsov45, Theo G.M. van
Erp4, Jessica A. Turner6, Lianne Schmaal24,25, Dick Veltman25, Lei Wang5,
Kathryn Alpert5, Dmitry Isaev1, Artemis Zavaliangos-Petropulu1, Christopher
R.K. Ching1, Vince Calhoun40, David Glahn7, Theodore D. Satterthwaite8,
Ole Andreas Andreassen9, Stefan Borgwardt10, Fleur Howells11, Nynke
Groenewold11, Aristotle Voineskos11, Joaquim Radua13,34,35,36, Steven G.
Potkin4, Benedicto Crespo-Facorro14,38, Diana Tordesillas-Gutirrez14,38, Li
Shen15, Irina Lebedeva16, Gianfranco Spalletta17, Gary Donohoe18, Peter
Kochunov19, Pedro G.P. Rosa20,33, Anthony James21, Udo Dannlowski26,
Bernhard T. Baune31, Andr Aleman32, Ian H. Gotlib27, Henrik Walter28,
Martin Walter29,41,42, Jair C. Soares30, Stefan Ehrlich43, Ruben C. Gur8, N.
Trung Doan9, Ingrid Agartz9, Lars T. Westlye9,37, Fabienne Harrisberger10,
Anita Riecher-Rossler10, Anne Uhlmann11, Dan J. Stein11, Erin W. Dickie12,
Edith Pomarol-Clotet13,34, Paola Fuentes-Claramonte13,34, Erick Jorge
Canales-Rodrguez13,34,39,46, Raymond Salvador13,34, Alexander J. Huang4,
Roberto Roiz-Santiaez14,38, Shan Cong15, Alexander Tomyshev16, Fabrizio
Piras17, Daniela Vecchio17, Nerisa Banaj17, Valentina Ciullo17, Elliot Hong19,
Geraldo Busatto20,33, Marcus V. Zanetti20,33, Mauricio H. Serpa20,33, Simon
Cervenka22, Sinead Kelly23, Dominik Grotegerd26, Matthew D. Sacchet27, Ilya
M. Veer28, Meng Li29, Mon-Ju Wu30, Benson Irungu30, Esther Walton43,44,
and Paul M. Thompson1, for the ENIGMA consortium
1 Imaging Genetics Center, Stevens Institute for Neuroimaging and Informatics,
University of Southern California, Marina Del Rey, CA, USA
2 The Institute for Information Transmission Problems, Moscow, Russia
3 Department of Biomedical Engineering, Illinois Institute of Technology, Chicago,
IL, USA
4 Department of Psychiatry and Human Behavior, University of California Irvine,
Irvine, CA, USA
5 Department of Psychiatry, Northwestern University, Chicago, IL, USA
6 Psychology Department & Neuroscience Institute, Georgia State University,
7 Yale University School of Medicine, New Haven, CT, USA
8 Department of Psychiatry, University of Pennsylvania School of Medicine,
Atlanta GA, USA
Philadelphia, PA, USA
9 CoE NORMENT, KG Jebsen Centre for Psychosis Research, Division of Mental
Health and Addiction, Oslo University Hospital & Institute of Clinical Medicine,
University of Oslo, Oslo, Norway
10 Department of Psychiatry, University of Basel, Basel, Switzerland
11 MRC Unit on Risk & Resilience to Mental Disorders, Department of Psychiatry
and Mental Health, University of Cape Town, Cape Town, South Africa
12 Centre for Addiction and Mental Health, Toronto, Canada
13 FIDMAG Germanes Hospitalaries Research Foundation, Barcelona, Spain
14 University Hospital Marqus de Valdecilla, IDIVAL, Department of Psychiatry,
School of Medicine, University of Cantabria, Santander, Spain
15 Department of Radiology and Imaging Sciences, Indiana University School of
Medicine, Indianapolis, IN, USA
16 Mental Health Research Center, Moscow, Russia
17 Laboratory of Neuropsychiatry, Santa Lucia Foundation IRCCS, Rome, Italy
18 School of Psychology, NUI Galway, Galway, Ireland
19 Maryland Psychiatric Research Center, University of Maryland School of
20 Department of Psychiatry, Faculty of Medicine, University of Sao Paulo, Sao
Medicine, Baltimore
21 University of Oxford, Oxford, United Kingdom
Paulo, Brazil
22 Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska
23 Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA
24 Orygen, The National Centre of Excellence in Youth Mental Health, Melbourne,
Institutet, Stockholm, Sweden
25 Department of Psychiatry, VU University Medical Center, Amsterdam, The
Australia
Netherlands
26 Department of Psychiatry and Psychotherapy, University of Munster, Germany
27 Department of Psychology, Stanford University, Stanford, CA, USA
28 Charit´e Universitatsmedizin Berlin, corporate member of Freie Universitat Berlin,
Humboldt-Universitat zu Berlin, and Berlin Institute of Health, Department of
Psychiatry and Psychotherapy CCM, Berlin, Germany
29 Clinical Affective Neuroimaging Laboratory, Leibniz Institute for Neurobiology,
Magdeburg, Germany
30 University of Texas Health Science Center at Houston, Houston, TX, USA
31 Discipline of Psychiatry, Adelaide Medical School, The University of Adelaide
32 Interdisciplinary Center Psychopathology and Emotion regulation (ICPE),
Neuroimaging Center (BCN-NIC), University Medical Center Groningen, University
33 Center for Interdisciplinary Research on Applied Neurosciences (NAPNA),
of Groningen, Groningen, The Netherlands
34 CIBERSAM, Centro Investigacin Biomdica en Red de Salud Mental, Barcelona,
University of So Paulo, So Paulo, Brazil
Spain
35 Department of Clinical Neuroscience, Centre for Psychiatric Research, Karolinska
Institutet, Stockholm, Sweden
36 Department of Psychosis Studies, Institute of Psychiatry, Psychology and
Neuroscience, King's College London, United Kingdom
37 Department of Psychology, University of Oslo, Oslo, Norway
38 CIBERSAM, Centro Investigacin Biomdica en Red Salud Mental, Santander,
Spain
39 Radiology department, University Hospital Center (CHUV), Lausanne,
40 The Mind Research Network, Albuquerque, NM, USA
41 Leibniz Institute for Neurobiology, Magdeburg, Germany
Switzerland
42 Department of Psychiatry and Psychotherapy, University of Tbingen, Tbingen,
Germany
43 Division of Psychological and Social Medicine and Developmental Neurosciences,
Faculty of Medicine, TU Dresden, Germany
44 Psychology Department, Georgia State University, Atlanta, GA, USA
45 Skolkovo Institute of Science and Technology, Moscow, Russia
46 Signal Processing Laboratory 5 (LTS5), ´Ecole Polytechnique F´ed´erale de
Lausanne (EPFL), Lausanne, Switzerland
|
1011.2797 | 3 | 1011 | 2012-02-01T16:14:01 | When are microcircuits well-modeled by maximum entropy methods? | [
"q-bio.NC",
"cond-mat.dis-nn",
"cs.IT",
"cs.IT",
"physics.data-an"
] | Describing the collective activity of neural populations is a daunting task: the number of possible patterns grows exponentially with the number of cells, resulting in practically unlimited complexity. Recent empirical studies, however, suggest a vast simplification in how multi-neuron spiking occurs: the activity patterns of some circuits are nearly completely captured by pairwise interactions among neurons. Why are such pairwise models so successful in some instances, but insufficient in others? Here, we study the emergence of higher-order interactions in simple circuits with different architectures and inputs. We quantify the impact of higher-order interactions by comparing the responses of mechanistic circuit models vs. "null" descriptions in which all higher-than-pairwise correlations have been accounted for by lower order statistics, known as pairwise maximum entropy models.
We find that bimodal input signals produce larger deviations from pairwise predictions than unimodal inputs for circuits with local and global connectivity. Moreover, recurrent coupling can accentuate these deviations, if coupling strengths are neither too weak nor too strong. A circuit model based on intracellular recordings from ON parasol retinal ganglion cells shows that a broad range of light signals induce unimodal inputs to spike generators, and that coupling strengths produce weak effects on higher-order interactions. This provides a novel explanation for the success of pairwise models in this system. Overall, our findings identify circuit-level mechanisms that produce and fail to produce higher-order spiking statistics in neural ensembles. | q-bio.NC | q-bio |
When Do Microcircuits Produce Beyond-Pairwise
Correlations?
Andrea K. Barreiro1,4,∗, Julijana Gjorgjieva3,5, Fred Rieke2, and Eric Shea-Brown1
1 Department of Applied Mathematics, University of Washington
2Department of Physiology and Biophysics, University of Washington
3 Department of Applied Mathematics and Theoretical Physics, University of Cambridge
4 present affiliation: Department of Mathematics, Southern Methodist University
5 present affiliation: Center for Brain Science, Harvard University
∗ corresponding author
Abstract
Describing the collective activity of neural populations is a daunting task: the
number of possible patterns grows exponentially with the number of cells, resulting
in practically unlimited complexity. Recent empirical studies, however, suggest a vast
simplification in how multi-neuron spiking occurs: the activity patterns of some cir-
cuits are nearly completely captured by pairwise interactions among neurons. Why are
such pairwise models so successful in some instances, but insufficient in others? Here,
we study the emergence of higher-order interactions in simple circuits with different
architectures and inputs. We quantify the impact of higher-order interactions by com-
paring the responses of mechanistic circuit models vs. "null" descriptions in which
all higher-than-pairwise correlations have been accounted for by lower order statistics,
known as pairwise maximum entropy models.
We find that bimodal input signals produce larger deviations from pairwise predic-
tions than unimodal inputs for circuits with local and global connectivity. Moreover,
recurrent coupling can accentuate these deviations, if coupling strengths are neither
too weak nor too strong. A circuit model based on intracellular recordings from ON
parasol retinal ganglion cells shows that a broad range of light signals induce unimodal
inputs to spike generators, and that coupling strengths produce weak effects on higher-
order interactions. This provides a novel explanation for the success of pairwise models
in this system. Overall, our findings identify circuit-level mechanisms that produce and
fail to produce higher-order spiking statistics in neural ensembles.
Author Summary
Neural populations can, in principle, produce an enormous number of distinct multi-cell
patterns -- a number so large that the frequency of these patterns could never be mea-
1
sured experimentally. Remarkably, the activity of many circuits is well captured by simpler
probability models that rely only on the activity of single neurons and neuron pairs. These
pairwise models remove higher-order interactions among groups of more than two cells in a
principled way. Pairwise models succeed even in cases where circuit architecture and input
signals seem likely to create a more complex set of outputs. We develop a general approach
to understanding which network architectures and input signals will lead such models to
succeed, and which will lead them to fail.
As a specific application, we consider the remarkable empirical success of pairwise mod-
els in capturing the activity of a class of retinal ganglion cells -- the output cells of the
retina. Our theory provides a direct explanation for these findings based on the filtering
and spike generation properties of ON parasol retinal circuitry, in which collective activity
arises through common feedforward inputs to spiking cells together with relatively weak gap
junction coupling. Specifically, filtering of light input upstream of ON parasol cells shapes
inputs to these cells in such a way that when they are processed by parasol cells, the output
spiking patterns are closely fit by a pairwise model.
Introduction
Information in neural circuits is often encoded in the activity of large, highly interconnected
neural populations. The combinatoric explosion of possible responses of such circuits poses
major conceptual, experimental, and computational challenges. How much of this potential
complexity is realized? What do statistical regularities in population responses tell us about
circuit architecture? Can simple circuit models with limited interactions among cells capture
the relevant information content? These questions are central to our understanding of neural
coding and decoding.
Two developments have advanced studies of synchronous activity in recent years. First,
new experimental techniques provide access to responses from the large groups of neurons
necessary to adequately sample synchronous activity patterns [1]. Second, maximum en-
tropy approaches from statistical physics have provided a powerful approach to distinguish
genuinely higher-order synchrony (interactions) from that explainable by pairwise statistical
interactions among neurons [2 -- 4]. These approaches have produced diverse findings. In some
instances, activity of neural populations is extremely well described by pairwise interactions
alone, so that pairwise maximum entropy models provide a nearly complete description [5,6].
In other cases, while pairwise models bring major improvements over independent descrip-
tions, it is not clear that they fully capture the data [3,7 -- 12]. Empirical studies indicate that
pairwise models can fail to explain the responses of spatially localized triplets of cells [11,13],
as well as the activity of populations of ∼100 cells responding to natural stimuli [13]. Over-
all, the range of empirical results highlights the need to understand the network and input
features that control the statistical complexity of synchronous activity patterns.
Several themes have emerged from efforts to link the correlation structure of spiking activ-
ity to circuit mechanisms using generalized [14 -- 17] and biologically-based models [3, 18, 19].
Two findings are particularly relevant for the present study. First, thresholding nonlinearities
2
in circuits with Gaussian input signals can generate correlations that cannot be explained
by pairwise statistics [14]; the deviations from pairwise predictions are modest at moder-
ate population sizes [16], but may become severe as population size N grows large [14, 20].
Cluster sizes (i.e., the number of cells firing simultaneously), in particular, may be poorly
fit by pairwise models. For large population sizes, a widespread distribution of cluster sizes
requires interactions of all orders [14], although for population sizes explored in experimental
data third or fourth order interactions may be sufficient [10]. Poor fitting of cluster sizes by
the pairwise model is also noted in networks of recurrent integrate-and-fire units with adapt-
ing thresholds and refractory potassium currents [19]. Small groups of cells that perform
logical operations can be shown to generate higher-order interactions by introducing noisy
processes with synergistic effects [4], but it is unclear what neural mechanisms might produce
similar distributions. These diverse findings point to the important role that circuit features
and mechanisms -- input statistics, input/output relationships, and circuit connectivity --
may play in regulating higher-order interactions. Nevertheless, we still lack a systematic
understanding that links these features and their combinations to the success and failure of
pairwise statistical models.
Second, perturbation approaches can explain why maximum entropy models with purely
pairwise interactions capture circuit behavior when the population firing rate is low (i.e. the
total number of firing events from all cells in the same small time window is small) [17,21,22].
The fraction of multi-information [4] captured by the pairwise model, a common metric of
success, is necessarily low in this regime [17]. In this regime, as well, higher-order interactions
cannot be introduced as an artifact of under-sampling the network [22]. The studies which
find that pairwise models are successful in capturing multivariate spiking data, however,
include cases which either extend beyond [5, 6], or are marginally within [7], the low spikes
per time bin regime. Nor do perturbation results necessarily account for the success of models
where pairwise connections are restricted to nearest-neighbor connections [5]. Therefore, an
explanation for the empirical successes of pairwise models remains incomplete.
Here, we aim to bridge some of the gaps between our understanding of mechanistic and
statistical models. Our strategy is to systematically characterize the ability of pairwise max-
imum entropy (PME) models to capture the responses of several simple circuits (see Figure
1). In each case, we search exhaustively over the entire parameter space for networks with
a simple thresholding model of spike generation. Studies of feedforward circuits reveal that
the success of PME models does not always bear a simple relation to network architecture
-- we find examples in which networks with local connections can deviate substantially from
PME predictions, while networks with global connections can be well approximated by PME
models. A more consistent determinant of the success of PME models is the unimodal vs.
bimodal profile of inputs to the circuit. In addition to departures driven by these inputs,
we show that excitatory recurrent coupling can increase departures from PME models by a
further 3-5 fold.
We apply these general results to responses of specific networks based on measured prop-
erties of primate ON parasol ganglion cells. We find that these responses are closely ap-
proximated by PME models for a wide range of input types. PME models are successful in
3
this case because the temporal filtering properties of parasol cells induce unimodal synaptic
inputs for a broad range of light inputs and because the measured coupling strengths are
insufficient to produce large effects on higher-order interactions. This provides insight into
why the measured activity patterns in these cells are well captured by PME models [5, 6].
Results
Our aim is to determine how network architecture and input statistics influence the success
of PME models. We start by considering networks of three cells, which both allow us to
make substantial analytical progress and to visualize the results geometrically. We then
extend these results in two ways: (1) to models based on the measured properties of primate
ganglion cells, one system in which PME models have been very successful [5, 6]; and (2) to
larger networks.
A geometric approach to identifying higher-order interactions among
triplets of cells
One strategy to identify higher-order interactions is to compare multi-neuron spike data
against a description in which any higher-order interactions have been removed in a principled
way -- that is, a description in which all higher-order correlations are completely described by
lower-order statistics. Such a description may be given by a maximum entropy model [2, 23,
24], which determines how much of the potential complexity of response patterns produced by
large neural populations can be captured by a given set of constraints. The idea is to identify
the most unstructured, or maximum entropy, distribution consistent with the constraints.
Comparing the predicted and measured probabilities of different responses tests whether
the constraints used are sufficient to explain the network activity, or whether additional
constraints need to be considered. Such additional constraints would produce additional
structure in the predicted response distribution, and hence lower the entropy.
A common approach is to limit the constraints to a given statistical order -- for example,
to consider only the first and second moments of the distributions, which are determined by
the mean and pairwise interactions. In the context of spiking neurons, we denote µi ≡ E[xi]
as the firing rate of neuron i and ρij ≡ E[xixj] as the joint probability that neurons i and j
will fire. The distribution with the largest entropy for a given µi and ρij is often referred to
as the pairwise maximum entropy (PME) model. The problem is made simpler if we consider
only permutation-symmetric spiking patterns, in which the firing rate and correlation do not
depend on the identity of the cells; i.e. µi = µ, ρij = ρ for i (cid:54)= j. Thus the PME problem is
to identify the distribution that maximizes the response entropy given the constraints µ and
ρ. In this section, we review some results that help provide a geometric, and hence visual,
approach to this problem (see, for example [19]).
We consider a permutation-symmetric network of three cells with binary responses. We
assume that the response is stationary and uncorrelated in time. From symmetry, the pos-
4
sible network responses are
p0 = P (0, 0, 0)
p1 = P (1, 0, 0) = P (0, 1, 0) = P (0, 0, 1)
p2 = P (1, 1, 0) = P (1, 0, 1) = P (0, 1, 1)
p3 = P (1, 1, 1),
where pi denotes the probability that a particular set of i cells spike and the remaining 3− i
do not. Possible values of (p0, p1, p2, p3) are constrained by the fact that P is a probability
distribution, meaning that the sum of pi over all eight states is one. We will rearrange these
response probabilities to define a more convenient coordinate system below.
Possible solutions to the PME problem take the form of exponential functions character-
ized by two parameters, λ1 and λ2, which serve as Lagrange multipliers for the constraints,
P (x1, x2, x3) =
1
Z
exp[λ1(x1 + x2 + x3) + λ2(x1x2 + x2x3 + x1x3)].
(1)
The factor Z normalizes P to be a probability distribution. We can combine the individual
probabilities of events
p0 =
p1 =
p2 =
p3 =
1
Z
1
Z
1
Z
1
Z
exp(λ1)
exp(2λ1 + λ2)
exp(3λ1 + 3λ2)
(cid:18) p2
(cid:19)3
p1
to yield the equation
p3
p0
=
.
(2)
This is equivalent to the condition that the strain measure defined in [25] be zero (in particu-
lar, the strain is negative whenever p3/p0 − (p2/p1)3 < 0, a condition identified in [25] as cor-
responding to sparsity in the neural code). Equation 2 defines, implicitly, a two-dimensional
surface in the three-dimensional space of possible probability distributions which we call the
maximum entropy surface. The family of distributions consistent with a given µ and ρ forms
a line (the iso-moment line) in this space [19]. The PME fit is the intersection of this line
with the surface defined by Equation 2.
This geometrical description of the PME problem takes a particularly simple form in an
alternative coordinate space:
fp = p3 + p0
f1p =
f1m =
p3
p3 + p0
p2
p2 + p1
.
5
(3)
This set of coordinates separates network responses based on whether they are "pure" (all
cells either spike, or do not) or "mixed" (only a subset of cells spike). fp is the fraction of
observed responses that are pure; f1p is the fraction of pure responses with more cells spiking
than not (p3 vs. p0). f1m is the fraction of mixed responses with more cells firing than not
(p2 vs. p1). Possible probability distributions are contained within a cube in this coordinate
space: 0 ≤ fp, f1p, f1m ≤ 1. The iso-moment line for a given µ and ρ is still a line in this
coordinate space (see Methods), and the PME approximation is given by the intersection of
this line with the maximum entropy surface.
The convenience of this coordinate system is apparent when the maximum entropy con-
straint (Equation 2) is rewritten:
f1p =
f 3
1m
1 − 3f1m + 3f 2
1m
.
(4)
This surface is independent of fp -- i.e. the maximum entropy surface forms a curve when
projected into the (f1p, f1m)-plane. In addition, each iso-moment line lies in a constant fp
plane; the distance of an observed distribution P from the surface is thus easily visualized.
This geometric view of the relation between the activity of a given network and its PME fit is
particularly useful in visualizing results from exhaustive searches across network parameters,
as described below.
We use the Kullback-Leibler divergence, DKL(P, P ), to quantify the accuracy of the
PME approximation P to a distribution P . This measure has a natural interpretation as
the contribution of higher-order interactions to the response entropy S(P ) [2, 4], and may in
this context be written as the difference of entropies S( P ) − S(P ). In addition, DKL(P, P )
is approximately − log2 L, where L is the average likelihood -- that is, the average (i.e. per
observation) relative likelihood that a sequence of data drawn from the distribution P was
instead drawn from the model P [5, 26]. For example, if DKL = 1, the average likelihood
that a single sample from P -- i.e. a single network response -- in fact came from P is
2−1 (we use the base 2 logarithm in our definition of the Kullback-Leibler divergence, so all
numerical values are in units of bits).
An alternative measure of the quality of the pairwise model comes from normalizing
DKL(P, P ) by the corresponding distance of the distribution P from an independent max-
imum entropy fit DKL(P, P1), where P1 is the highest entropy distribution consistent with
the mean firing rates of the cells (equivalently, the independent model P1 is given by the
product of single-cell marginal firing probabilities) [2]. Many studies [5 -- 7, 17] use
∆ =
Dind − Dpair
Dind
= 1 − DKL(P, P )
DKL(P, P1)
,
(5)
where following [17] we define Dind ≡ DKL(P, P1) and Dpair ≡ DKL(P, P ). A value of ∆ = 1
(100%) indicates that the pairwise model perfectly captures the additional information left
out of the independent model, while a value of ∆ = 0 indicates that the pairwise model gives
6
no improvement over the independent model. Because we are interested in cases where the
pairwise model fails to capture circuit outputs, at first it seems we would wish to identify
circuits which produce a low ∆. However, in the circuits we explore, we find that the lowest
values of ∆ are achieved for nearly independent -- and therefore "uninteresting" -- spike
patterns. Therefore, we have chosen to use DKL(P, P ) as our primary measure of the quality
of the pairwise model, while reporting values of ∆ in parallel.
To get an intuitive picture of DKL(P, P ) throughout the cube of possible distributions
P , we view this quantity along constant-fp slices in Figure 2. DKL(P, P ) increases with
distance from the constraint curve (Equation 4); along the iso-moment line for a given
(µ, ρ), DKL(P, P ) is convex with a minimum of zero at P = P (detailed calculations are
given in Materials and Methods). Therefore, for any choice of µ and ρ, DKL(P, P ) increases
monotonically as a function of the distance along an iso-moment line. The distance, which is
easily visualized, thus gives an indication of how close P comes to being a pairwise maximum
entropy distribution. The observed distribution with the maximal deviation from its pairwise
maximum entropy approximation will occur at one of the two points where the iso-moment
line reaches the boundary of the cube. The global maximum of DKL(P, P ) will, therefore,
also occur on the boundary.
To assess the numerical significance of DKL(P, P ), we can compare it with the maximal
achievable value for any symmetric distribution on three spiking cells. For three cells, this
value is 1 (or 1/3 bits per neuron), achieved by the XOR operation [4] -- i.e. the average
probability that a single output of the XOR circuit came instead from the PME approxima-
tion is 2−1. This distribution, along with its position in the (fp, f1p, f1m) coordinate space, is
illustrated in Figure 2 (fp = 0.25 slice) and the right column of Figure 3. We will find that
distributions produced by permutation-symmetric networks fall far short of this value.
In summary, we have shown that identifying high-order interactions in the joint firing
patterns of three cells is equivalent to showing that spiking probabilities lie a substantial
distance from a constraint surface that is easy to visualize. Given this geometric description
of the problem, we next consider how the distance from the constraint surface depends on
circuit connectivity, nonlinear properties of the individual circuit elements, and the statistics
of the input signals (Figure 1).
When do triplet inputs produce higher-order interactions in spike
outputs?
We first considered a simple feedforward circuit in which three spiking cells sum and threshold
their inputs. Each cell j received an independent input Ij and a "triplet" -- or global --
input Ic that is shared among all three cells. Comparison of the total input Sj = Ic + Ij with
a threshold Θ determined whether or not the cell spiked in that time bin. The nonlinear
threshold can produce substantial differences between input and output correlations [27 -- 31].
An additional parameter, c, identified the fraction of the total input variance σ2 originating
from the global input; that is, c ≡ Var[Ic]/Var[Ic + Ij].
What types of inputs might be natural to consider? The constraint surface (Equations
7
2 and 4) can give us some intuition into what types of inputs will be more or less likely to
produce spiking responses that will deviate from the pairwise model. For example, suppose
that Ic can take on values that cluster around two separated values, µA < µB, but that it
is very unlikely that Ic ever takes on values in the interval between; that is, the distribution
of Ic is bimodal. If µB is large enough to push the cells over threshold but µA is not, then
we see that any contribution to the right-hand side of Equation 2, p2/p1, depends only
on the distribution of the independent inputs Ij; if either one or two cells spike, then the
common input must have been drawn from the cluster of values around µA, because otherwise
all three cells would have spiked. Keeping the values of µA and µB the same, therefore,
but changing the relative likelihood of drawing the common input from one cluster or the
other, would change the ratio p3/p0 without changing the ratio p2/p1. Hence the constraint
specifying those network responses exactly describable by PME models can be violated when
the common input is bimodal.
In contrast, we may instead consider a unimodal input -- one whose distribution has a
single peak or range of most likely values -- of which a Gaussian input is a natural example.
Here, the distribution of the common input Ic is completely described by its mean and
variance; both parameters can impact the ratio p3/p0 (by altering the likelihood that the
common input alone can trigger spikes) and the ratio p2/p1. Each value of Ic is consistent
with both events p1 and p2, with the relative likelihood of each event depending on the
specific value of Ic; it is no longer clear how to separate the two.
Motivated by these observations, we choose our inputs to come from either one of three
types of unimodal distributions -- Gaussian, skewed, or uniform -- or a bimodal distribution.
The one-sided skewed shape, in particular, is chosen to mimic qualitative features of inputs to
retinal ganglion cells (e.g. [32] and below). We indeed find both qualitative and quantitative
differences in the capacity of each class of inputs to generate deviations from the pairwise
model.
Unimodal inputs fail to produce higher-order interactions in three cell feedfor-
ward circuits
We first considered unimodal inputs, which were chosen from a distribution with a single
peak (or range of most likely values). Gaussian inputs provide a natural example. If Ic and
each Ij are Gaussian, then the joint distribution of S = (S1, S2, S3) is multivariate normal,
and therefore characterized entirely by its means and covariances. Because the PME fit to a
continuous distribution is precisely the multivariate normal that is consistent with the first
and second moments, every such input distribution on S exactly coincides with its PME
fit. However, even with Gaussian inputs, outputs (which are now in the binary state space
{0, 1}3) will deviate from the PME fit [14, 16]. As shown below, non-Gaussian unimodal
inputs can produce outputs with larger deviations. Nonetheless, these deviations in all cases
are small, and PME models are quite accurate descriptions of circuits with a broad range of
unimodal inputs.
In more detail, we considered a circuit of three cells with inputs Ic and Ij that could be
Gaussian, uniform, or skewed. For each type of input distribution, we probed the output
8
distribution across a range of values for c, σ, and Θ that explored "all" possible activity
patterns. In particular, we covered a full range of firing rates (or equivalently spikes per
time bin), not limited to the low firing rate regime treated in [17]. Figure 4A-C shows
observed distributions for different marginal input statistics (left column). The central col-
umn compares all observed distributions with the PME constraint curve, projected into the
(f1p, f1m)-plane.
The right column of Figure 4A-C shows DKL(P, P ) as a function of c and σ for the value
of Θ that maximized DKL(P, P ) (or one of them, if multiple such values exist). Different
scales are used to emphasize the structure of the data. For the unimodal cases shown, DKL
peaked in regions with comparatively low input variance (σ < 1) and large relative strength
of common input (c > 0.5). However, DKL(P, P ) never reached a very high numerical value
for unimodal inputs; the maximal values achieved for Gaussian, skewed, and uniform distri-
butions are 0.00376 (∆ = 0.989), 0.0152 (∆ = 0.999), and 0.0186 (∆ = 0.9428) respectively
(compare with Figure 3). Thus hundreds or thousands of samples would be required to
reliably distinguish the outputs of these networks from the PME approximation.
Clear patterns emerged when we viewed DKL(P, P ) as a function of output spiking statis-
tics rather than input statistics. Figure 5A-B show the same data contained in the center
column of Figure 4, but now plotted with respect to the output firing rate, which is the same
for all cells. The data were segregated according to the correlation coefficient ρ between
the responses of cell pairs, with lighter shades indicating increasing correlation. For a fixed
correlation, there was generally a one-to-one relationship between firing rate and DKL(P, P ).
For unimodal distributions (Figure 5A-B), DKL(P, P ) showed a double-peaked relationship
with firing rate, with larger values attained at low and high firing rates, and a minimum in
between. Additionally, DKL(P, P ) had a non-monotonic relationship with spike correlation:
it increased from zero for low values of correlation, obtained a maximum for an intermediate
value, and then decreased. These limiting behaviors agree with intuition: a spike pattern
that is completely uncorrelated can be described by an independent distribution (a special
case of PME model), and one that is perfectly correlated can be completely described via
(perfect) pairwise interactions alone.
Bimodal triplet inputs can generate higher-order interactions in three cell feed-
forward circuits
Having shown that a wide range of unimodal common inputs produced spike patterns that are
well-approximated by PME fits, we next examined bimodal inputs. Figure 4D shows results
from a simple ensemble of bimodal inputs -- Bernoulli-distributed common and independent
inputs -- that produced moderate deviations from the pairwise approximation. The common
input was "1" with probability p and "0" with probability 1 − p. The independent inputs
were each chosen to be "1" with probability q and "0" with probability 1− q. The threshold
of the cells was between 1 and 2, so that spiking required both common and independent
inputs to be active. The space of possible spiking distributions was explored by varying p
and q.
This circuit produced response distributions that deviated modestly from PME fits, and
9
these distributions preferentially lie on one side of the constraint curve (Figure 4D, center).
The largest values of DKL(P, P ) occurred where moderate correlated input is coupled with
strong background input (q > 0.5; Figure 4D, right), and reached values that are five times
higher than those found for a unimodal distribution (the maximal value achieved is 0.091).
The location of the distribution generating this maximum value is demonstrated in the center
column of Figure 3.
Both of these observations can be explained by direct calculation of the spiking probabili-
ties. Substituting the probabilities of different events -- p0 = 1−p+p(1−q)3, p1 = pq(1−q)2,
p2 = pq2(1−q) and p3 = pq3 -- into the constraint surface specifying PME models (Equation
2) and dividing by q3, we can write
p
1 − p + p(1 − q)3 =
1
(1 − q)3
(6)
which gives us an intuition for how to violate the constraint; for a fixed q, we manipulate
the left-hand side by changing p.
Another way to view this is by observing that the right hand side of Equation 2 can be
written without reference to the probability of common input; because P [1 spike Ic = 0] = 0
and P [2 spikes Ic = 0] = 0, one may write
p2
p1
=
=
P [2 spikes Ic = 1]P [Ic = 1]
P [1 spike Ic = 1]P [Ic = 1]
P [2 spikes Ic = 1]
P [1 spike Ic = 1]
(7)
which has no dependence on the statistics of the common input. So the left hand side of the
constraint equation (Eqn. 2) can be manipulated by shifting p, without making any changes
to the right hand side.
In Figure 5C, we again present values of DKL(P, P ) as a function of the firing rate
and pairwise correlation elicited by the full range of possible bimodal inputs. We see that
DKL(P, P ) is maximized at a single, intermediate firing rate, and for correlation values near
0.6.
We find distinctly different patterns when we view ∆ (i.e. DKL normalized by the distance
of P from an independent maximum entropy fit, Equation 5), for these same simulations, as
a function of output spiking statistics (Figure 5D-F). For unimodal distributions (Figure 5D-
E), ∆ is very close to 1, with the few exceptions at extreme firing rates. For bimodal inputs
(Figure 5F), ∆ may be appreciably far from 1 -- as small as 0.5 -- with the smallest numbers
(suggesting a poor fit of the pairwise model) occurring for low correlation ρ. This highlights
one interesting example where these two metrics for judging the quality of the pairwise
model, DKL(P, P ) and ∆, yield contrasting results.
An analytical explanation for unimodal vs. bimodal effects
We next develop analytical support for the distinct impact of unimodal vs. bimodal inputs
on fits by the PME model. Specifically, we calculate DKL(P, P ) for small deviations from the
10
PME constraint surface of Equation 2. We summarize the results of this calculation here;
details are in Materials and Methods. We considered narrow distributions of common input
Ic, with a small parameter c indicating their variance. By approximating the distribution
of network outputs by a Taylor series in c, we found that the leading order behavior of
DKL(P, P ) depended on c3 for unimodal distributions -- i.e the low order terms in c dropped
out (for symmetric distributions, such as Gaussian, the growth was even smaller: c4). For
bimodal distributions, on the other hand, the leading order term of DKL(P, P ) grew like
c2. The key point is that, as the strength of common input signals increased, circuits with
bimodal inputs diverged from the PME fit much more rapidly than circuits with unimodal
inputs.
When do pairwise inputs produce higher-order interactions in spike
outputs?
In the previous sections, we considered permutation-symmetric distributions generated by
a single, global, common input. Another class of permutation-symmetric distributions can
be generated when common inputs are shared pairwise -- i.e. by two cells but not three at
once. We now show that significant departures from the pairwise maximum entropy model
(PME) can be generated with pairwise bimodal inputs. This provides a specific example in
which network architecture and output statistics are not in simple correspondence.
Our circuit setup included three cells, each of which received and summed two inputs
and spiked if this sum exceeded a threshold. We denote the inputs I12, I23, I13 so that cell 2
received I12 and I23, and so forth, as illustrated in Figure 1. Each input was chosen from a
binary distribution with parameters m and r so that P [Iij = m] = r and P [Iij = 0] = 1 − r.
Without loss of generality, we chose m = 1 and chose the threshold such that 1 < Θ < 2.
Therefore, both pairwise inputs to a cell must be active in order for a cell to fire. It is not
possible in this circuit for precisely two cells to fire; for two cells to fire (say cell 1 and cell
2), both inputs to each cell must be active. However, this implies that both inputs to cell
3 (I13 and I23) are active as well. If two cells fire, then the third must fire as well; that is,
p2 = 0.
The remaining probabilities are easily computed by itemizing and computing the proba-
bilities for each event and are as follows: p3 = r3, p1 = r2(1−r), and p0 = 3r(1−r)2+(1−r)3.
This distribution has a unique PME fit consistent with both the first and second moments.
However, the PME fit can be far from the actual distribution; we found that DKL(P, P ) de-
pended on the input rate r and could exceed 0.5. Thus observation of the network response
(single draw from P ) would, on average, have a likelihood ratio of 0.7 of coming from P
versus P , and observation of 15 network responses would have a likelihood ratio of less than
0.01.
We will revisit the contrast between global and pairwise common inputs below; we also
refer to the pairwise case as local inputs.
11
Recurrent coupling produces modest effects on higher-order inter-
actions
Neural circuits are rarely purely feedforward, and recurrent connectivity can play an impor-
tant role in shaping the resulting activity. Therefore, we next modify our thresholding model
to incorporate the effects of recurrent coupling among the spiking cells. We take all-to-all
coupling among our N = 3 cells, and assume that this coupling is rapid compared with
the timescale over which inputs arrive at the cells, so that coupling has its full impact on
the spike events that are recorded in a single realization of the model. We limit our study
to excitatory interactions, anticipating an application below to the effects of gap junctions
known to mediate recurrent coupling between retinal ganglion cells.
In more detail, our coupling model may be described as follows: first, inputs arrive at
each cell as for the cases without coupling. If the inputs elicit any spikes, there is a second
stage in which the input to each neuron receiving a connection from a spiking cell is increased
by an amount g. This represents a rapid depolarizing current, assumed for simplicity to add
linearly to the input currents. If the second stage results in additional spikes, the process is
repeated: recipient cells receive an additional current g, and their summed inputs are again
thresholded. The sequence terminates when no new spikes occurred on a given stage; e.g.,
for N = 3, there are a maximum of three stages. The spike pattern that is recorded on a
given trial is the total number of spikes fired across all of the stages.
For the simplest case of globally structured inputs and all-to-all coupling, the probability
of each spike pattern (or count) can be written down analytically, by evaluating the proba-
bilities along the tree of possible events that unfold through the stages just described. For
other cases, these trees are more complex, and we instead evaluate probabilities via Monte-
Carlo sampling (unimodal inputs), or through enumeration of a finite number of possible
input states (bimodal inputs).
We begin by describing our results for globally structured, unimodal inputs. We range
over parameters c, σ, and Θ as for the uncoupled cases above, and additionally vary g over a
wide range of values, from g = 0 to g = 2.4 (comparable to the maximum threshold value).
Our basic finding is that coupling has only modest effects on DKL values, when compared
with the theoretical range of values that could be obtained. Specifically, the largest value
found over all parameters tested was DKL = 0.0099 for the case of Gaussian inputs, 0.108
for uniform, and 0.0599 for skewed. All of these values are roughly 3 − 5× values obtained
for the uncoupled cases, and, for Gaussian and skewed inputs, remain much smaller than
values found for bimodal inputs. The largest value achieved over any unimodal input and for
any coupling strength (0.108) is comparable to the largest value found with bimodal inputs
without coupling.
Figure 6A shows the data for Gaussian inputs as a scatter plot of DKL values vs. coupling
strengths g, showing that intermediate values of coupling -- for which the size of coupling
terms is about the same as the standard deviation of the inputs -- have the greatest ability to
drive departures from the PME model. Moreover, when coupling strength was varied while
keeping other parameters fixed, DKL was also generally maximized at intermediate values of
the coupling. In Figure 6B, we show DKL(P, P ) vs. coupling strength g, for representative
12
values of the circuit parameters c, σ and Θ. For each case of global Gaussian, skewed,
uniform, and bimodal inputs (thin lines), we see a peak in DKL(P, P ) at medium values of
g.
We next introduce coupling to circuits with pairwise, locally structured unimodal inputs
(as in the preceding section). Here, we range over parameters σ and Θ (c is fixed at 0.5), and
additionally vary g over the same wide range of values above. For local inputs, the maximum
value of DKL found could be increased modestly by including coupling (from 0.0013 to 0.0086
with Gaussian inputs, from 0.0150 to 0.0157 with skewed inputs, and from 0.0233 to 0.0547
with uniform inputs), but remains substantially below the values found for global inputs with
coupling. As in circuits with global inputs, DKL was generally maximized at intermediate
values of the coupling; this is illustrated for uniform inputs in the third panel of 6B (note
that global and local inputs are identical for Gaussian inputs). An exception was observed
for skewed inputs, illustrated in the top panel of 6B, where DKL shows a maximum at g = 0.
Finally, we introduce coupling to circuits with bimodal inputs. We note that because
the inputs can only take on a finite number of configurations, with all other parameters
fixed, the output distribution of the network can only change at a few discrete values of g.
Moreover, large values of g induce "all-or-none" firing behavior, which is perfectly matched
by pairwise maximum entropy. For global inputs, we find modest increases in DKL which
are maximized at an intermediate value of the coupling; for pairwise inputs, the output
distribution attains its maximum level of higher order interactions at g = 0 and decreases
only when g is sufficiently strong enough to create an "all-or-none" firing pattern (bottom
panel, Fig. 6B).
In summary, both input statistics and connectivity shape the development of higher-
order interactions in our basic circuit model. Other parameters being equal, bimodal inputs
can generate a larger DKL(P, P ) than unimodal inputs. For a particular choice of input
marginals, global inputs can generate greater deviations than purely pairwise inputs (with
the exception of one case, that locally structured bimodal inputs). However, the goodness of
the PME fit alone does not distinguish between global and pairwise anatomical projections.
Adding fast recurrent excitation increases the accessible magnitude of higher-order interac-
tions; for a fixed Θ, the range of accessible DKL increases -- and then decreases -- with
the magnitude of recurrent excitation. With these principles in hand, we now study a more
biologically realistic network model.
An experimentally constrained model for correlated firing in retinal
ganglion cells
PME approaches have been effective in capturing the activity of small retinal ganglion cell
(RGC) populations [5 -- 7]. This success does not have an obvious anatomical correlate -- i.e.
there are multiple opportunities in the retinal circuitry for interactions among three or more
ganglion cells. Why do these apparently fail to generate higher-order interactions in output
spiking data? To answer this question, we explored the properties of circuits composed of
cells with input statistics, recurrent connectivity and spike-generating mechanisms based
13
directly on experiment. We based our model on ON parasol RGCs, one of the RGC types
for which PME approaches have been applied extensively [5, 6]. We first describe the RGC
model, then apply this to feedforward circuits, and finally consider recurrent connections.
RGC model
We modeled a single ON parasol RGC in two stages (for details see Materials and Methods).
First, we characterized the light-dependent excitatory and inhibitory synaptic inputs to cell
k (gexc
k (t)) in response to randomly fluctuating light inputs s(t) via a linear-nonlinear
model, e.g.:
k (t), ginh
k (t) = N exc[Lexc ∗ sk(t) + ηexc
gexc
k ],
(8)
where N exc is a static nonlinearity, Lexc is a linear filter, and ηexc
is an effective input noise
that captures variability in the response to repetitions of the same time-varying stimulus.
These parameters were determined from fits to experimental data collected under condi-
tions similar to those in which PME models have been tested empirically. The modeled
excitatory and inhibitory conductances captured many of the statistical features of the real
conductances, particularly the correlation time and skewness.
k
Second, we used Equation 8 and an equivalent expression for ginh
k (t) as inputs to an
integrate-and-fire model incorporating a nonlinear voltage and history-dependent term to
account for refractory interactions between spikes [33]. The voltage evolution equation was
of the form
dV
dt
= F (V, t − tlast) +
Iinput(t)
,
C
(9)
where F (V, t − tlast) was allowed to depend on the time of the last spike tlast (more details
in Methods). We fit the parameters of this model to a dynamic clamp experiment [34, 35] in
which currents corresponding to gexc(t) and ginh(t) were injected into a cell and the resulting
voltage response measured. Excitatory and inhibitory synaptic currents injected during
one time step were determined by scaling the conductances by driving forces based on the
measured voltage in the previous time step. Recurrent connections were implemented by
adding an input current proportional to the voltage difference between the two coupled cells.
The prescription above provided a flexible model that we used to study the responses of
small RGC networks to a wide range of light inputs and circuit connectivities. Specifically,
we simulated RGC responses to light stimuli that were (1) constant, (2) time-varying and
spatially uniform, and (3) varying in both space and time. Correlations between cell inputs
arose from shared stimuli, from shared noise originating in the retinal circuitry [32], or
from recurrent connections [32, 36]. Shared stimuli were described by correlations among
sk. Shared noise arose via correlations in ηk as described in Materials and Methods. The
recurrent connections were chosen to be consistent with observed gap-junctional coupling
between ON parasol cells. We also investigated how stimulus filtering by Lexc and Linh
influenced network statistics.
14
For each network model, we determined whether the accuracy of a PME fit to the outputs
was predictable based on the structure of the input distributions, using the results developed
above for the idealized circuit model. We focused on excitatory conductances because they
exhibit stronger correlations than inhibitory conductances in ON parasol RGCs [32]. To
compare our results with empirical studies, constant light and spatially and temporally
fluctuating checkerboard stimuli were used as in [5, 6].
Feedforward RGC circuits
k
k
and ηinh
We start by considering networks without recurrent connectivity and without temporal mod-
ulations in the light input. Thus we set sk(t) = 0 for k = 1, 2, 3, so that the cells received
only Gaussian correlated noise ηexc
and constant excitatory and inhibitory conduc-
tances. Time-dependent conductances were generated and used as inputs to a simulation of
three model RGCs. Simulation length was sufficient to ensure significance of all reported
deviations from PME fits (see Materials and Methods). Under these conditions the excita-
tory conductances were unimodal and broadly Gaussian. As expected from earlier results
on threshold models, the spiking distributions were well-modeled by a PME fit, as shown
in Figure 7A; DKL(P, P ) is 2.90 × 10−5 bits. This agrees with the very good fits found
experimentally in [5] under constant light stimulation.
For full-field simulations, each cell received the same stimulus, sk(t) = s(t), where s(t)
refreshes every few milliseconds with an independently chosen value from one of several
marginal distributions. The shared stimulus produced strong pairwise correlation between
conductances of neighboring cells. However, results from our threshold model (Fig. 4) suggest
that this is not the overall determining factor in whether or not spiking outputs will be
well-modeled by a PME fit; rather, the shape of the marginal distribution of inputs (here,
conductances) should be more important than the strength of pairwise correlations.
We first examined the effects of different marginal statistics of light stimuli, standard
deviation of full-field flicker, and refresh rate on the marginal distributions of excitatory
conductances. For a short refresh rate (8 ms) and small flicker variance (1/6 or 1/4 of
baseline light intensity), temporal averaging via the filter Lexc and the approximately linear
form of N exc over these light intensities produced a unimodal, modestly skewed distribution
of excitatory conductances, regardless of whether the flicker is drawn from a Gaussian or
binary distribution (see Figure 7B-C, center panels). For a slower refresh rate (100 ms)
and large flicker variance (1/3 or 1/2 of baseline light intensity), excitatory conductances
had multi-modal and skewed features, again regardless of whether the flicker is drawn from
a Gaussian or binary distribution (Figure 7D). Other parameters being equal, binary light
input produced more skewed conductances. While some conductance distributions had mul-
tiple local maxima, these were never well-separated, with the envelope of the distribution
still resembling a skewed distribution.
As expected from our studies with the simple thresholding model of spike generation,
the largely unimodal shape of input distributions was reflected in the ability of PME fits
to accurately capture spiking distributions. DKL(P, P ) values computed from the observed
distributions were small, never exceeding 0.0067; these are numbers comparable to what is
15
achievable by skewed inputs in our simple thresholding circuit model. To test the sensitivity
of this conclusion to the finite sampling in our simulations, we performed an analysis in which
the data was divided into 20 subsets and the maximum entropy analysis was performed indi-
vidually on each subset. The resulting KL-distance remained small, never exceeding 0.0089.
In summary, even high-contrast, bimodal, highly spatially correlated stimulus variations do
not produce a large departure from the PME fit.
When we examined all of the spiking distributions produced in this sequence of simula-
tions, we found a common pattern in the way in which the PME fit deviated from observed
distributions. Single spiking events were over-predicted by PME fits, whereas double spiking
events were under-predicted. We note that this is the same situation observed in our simple
threshold model with bimodal global inputs (see Figure 3 and Materials and Methods), and
corresponds to the case of negative strain identified by Ohiorhenuan et al. [25]. This find-
ing is extremely robust; upon perturbing the distributions by estimated standard errors, as
described in Materials and Methods, only 22 out of 840 perturbed distributions showed a
positive strain while the remainder had negative strain.
Overall, high-pass filtering -- a consequence of the differentiating linear filter in Equation
18 and illustrated in Figure 7D -- was responsible for significantly reducing the bimodality
of the input stimuli.
In Figure 7E, we produce a filter without the biphasic, high-pass
shape of Equation 18 (i.e., without the negative dip at longer time lags) by simply rectifying
the experimentally determined filter. This filter produced conductance distributions that
more completely reflect the bimodal shape of binary light inputs (Figure 7E, center panel).
The resulting simulation produces six-times greater DKL(P, P ) (Figure 7E, right panel).
This raises the intriguing suggestion that greater DKL(P, P ) may occur for other cell types
primarily characterized via monophasic filters (such as midget cells), or for different light
stimuli (e.g., at different mean levels) for which the retinal circuit acts to primarily integrate
rather than differentiate over time.
In Figure 8A, we examine this effect over all full-field stimulus conditions by plotting
deviations from the pairwise model, in terms of DKL(P, P ), in simulations with a rectified
filter against the same quantity with a non-rectified filter. An increase in DKL(P, P ) was
observed across stimulus conditions, with a markedly higher effect for longer refresh rates.
Large effects were accompanied by a striking increase in the bi- or multi-modality of ex-
citatory conductances (see Figure 8B, illustrating binary stimulus at 100 ms refresh rate
and standard deviation of 1/4), while small effects were accompanied by small or no dis-
cernible change in the marginal distribution of excitatory synaptic conductances (see Figure
8C, illustrating binary stimulus at 8 ms refresh rate and standard deviation of 1/2). This
consistent change could not be attributed to changes in lower order statistics; there was no
consistent relationship between the change in pairwise model performance and either firing
rate or pairwise correlations (data not shown).
Moving beyond full field light stimuli, we asked whether pairwise maximum entropy
models capture RGC responses to stimuli with varying spatial scales. We fixed stimulus
dynamics to match the two cases that yielded the highest DKL(P, P ) under the full field
protocol: for both Gaussian and binary stimuli, this was an 8 ms refresh rate and σ = 1/2.
16
The stimulus was generated as a random checkerboard with squares of variable size; each
square in the checkerboard, or stixel, was drawn independently from the appropriate marginal
distribution and updated at the corresponding refresh rate. The conductance input to each
RGC was then given by convolving the light stimulus with its receptive field, where the
stimulus is positioned with a fixed rotation and translation relative to the receptive fields.
This position was drawn randomly at the beginning of each simulation and held constant
throughout (see Figure 9D,G for examples, and Materials and Methods for further details).
The RGC spike patterns remained very well described by PME models for the full range
of spatial scales. Figure 9A shows this by plotting DKL(P, P ) vs. stixel size. Values of
DKL(P, P ) increased with spatial scale, sharply rising beyond 128 µm, where a stixel is
approximately the same size as a receptive field center. The points at 512 µm are from
the corresponding full field simulations, illustrating that introducing spatial scale via stixels
produces even closer fits by PME models.
Values reported in Figure 9A are averages of DKL(P, P ) produced by 5 random stimulus
positions. At stixel sizes of 128 µm and 256 µm, the resulting spiking distributions differed
significantly from position to position; in Figure 9B, we show the probabilities of the distinct
singlet (e.g. P (1, 0, 0)) and doublet (e.g. P (1, 1, 0)) spiking events produced at 256 µm.
Each stimulus position creates a "cloud" of dots (identified by color); large dots show the
average over 20 sub-simulations. Each sub-simulation is identified by a small dot of the
same color; because the simulations are very well-resolved, most of these are contained
within the large dots (and hence not visible here). Heterogeneity across stimulus positioning
is indicated by the distinct positioning of differently colored dots. At smaller spatial scales,
the process of averaging stimuli over the receptive fields results in spiking distributions that
are largely unchanged as stimulus position changes, as shown in Figure 9C, where singlet
and doublet spiking probabilities are plotted for 60 µm stixels. Thus, filtered light inputs
are largely homogeneous from cell to cell, as each receptive field samples a similar number of
independent, statistically identical inputs: Figure 9D shows the projection of input stixels
onto cell receptive fields from an example with 60 µm stixels. The resulting excitatory
conductances (Figure 9E) and spiking patterns (Figure 9F) are very close to cell-symmetric.
By contrast, spiking patterns showed significant heterogeneity from cell to cell,when the
stixel size was large; this arises because each cell in the population may be located differently
with respect to stixel boundaries, and therefore receive a distinct pattern of input activity.
Figure 9G shows the projection of input stixels onto cell receptive fields from one example,
with 256 µm stixels. The difference in statistical properties of inputs is reflected in the
marginal distributions of excitatory conductances that each cell receives, shown in Figure
9H. It is also apparent in the observed output spiking patterns, where one particular doublet
spiking pattern (here, 110) is significantly more likely to occur than the others (see Figure
9I). However, PME models gave excellent fits to data regardless of heterogeneity in RGC
responses, as illustrated for this particular example in Figure 9I; over all 20 sub-simulations
(as above), and over all individual stixel positions, we found a maximal DKL(P, P ) value of
0.00811.
17
Recurrent connectivity in the RGC circuit
We next considered the role of recurrence in shaping higher order interactions by incorpo-
rating gap junction coupling into our simulations. We did this separately for each full field
stimulus condition: Gaussian and binary marginal distributions, with variances 1/16, 1/12,
1/8, 1/6, 1/4, 1/3, or 1/2 of baseline light intensity, and refreshed at either 8, 40, or 100 ms
intervals. In each case, we added gap junction coupling with strengths from 1 to 16 times
an experimentally measured value [32], and compared the resulting DKL with that obtained
without recurrent coupling. Results are shown in Figure 10.
At the measured coupling strength (ggap = 1.1 nS) itself, the fit of the pairwise model
barely changed (Figure 10B). At twice the measured coupling strength (ggap = 2.2 nS),
recurrent coupling had increased higher order interactions, as measured by larger values of
DKL for all tested stimulus conditions. Higher order interactions could be further increased,
particularly for long refresh rates (100 ms), by increasing the coupling strength to 4 or 8
times its baseline level (ggap = 8.8 nS; see Figure 10C, D). Consistent with our intuition that
very strong coupling leads to "all-or-none" spiking patterns, DKL(P, P ) decreased as ggap
increased further, often to a level below what was seen in the absence of coupling (Figure
10F).
Again, we considered the possibility that firing rates could explain the increase; however,
firing rates actually decreased with the addition of coupling. Our findings are consistent with
our results for the threshold model (Figure 6), in that the impact of coupling is maximized
at intermediate values of the coupling strength. Moreover, the impact of recurrent coupling
on the maximal values of DKL evoked by visual stimuli is small overall, and almost negligible
for experimentally measured coupling strengths.
Scaling of higher-order interactions with population size
The results above identify the input statistics -- particular unimodality vs bimodality -- as
a key determinant of the success of PME models. How robust is this conclusion to network
size? The permutation-symmetric architectures we have considered can be scaled up to more
than three cells in several natural ways; for example, we can consider N cells with a global
common input. The pairwise (local) input structure can also be scaled up to consider N cells
on a ring, with each pair of adjacent cells receiving a common, pairwise input (see Figure 1).
We first considered a sequence of models in which a set of N threshold spiking units
received global input Ic (with mean 0 and variance σ2c) and an independent input Ij (with
mean 0 and variance σ2(1− c)). As for the three cell networks above, the output of each cell
was determined by summing and thresholding these inputs. The probability distribution of
network outputs was computed as described in the Methods and then fit with a pairwise
maximum entropy distribution. As also for the three cell networks, we explored a range of σ,
c, and Θ and recorded the maximum value of DKL(P, P ) between the observed distribution
P and its PME fit P . Figure 11 shows this DKL/N (i.e. entropy per cell [16]) for Gaussian,
uniform, skewed, and bimodal input distributions.
We found that the maximum DKL(P, P ) increased roughly linearly with N for bimodal
18
inputs, and superlinearly for unimodal inputs. The relative ordering found at N = 3 --
that the maximal achievable DKL(P, P ) is lowest for Gaussian inputs, followed by skewed,
uniform, and bimodal inputs consecutively -- remained the same. The sidebar of Figure 11
shows that the probability distributions produced by these inputs qualitatively agree with
this trend: departures from PME were more visually pronounced for global bimodal inputs
(top histogram) than for global unimodal inputs (third histogram from top). At N = 16,
the value DKL/N ≈ 0.1 for bimodal global inputs corresponds to a likelihood ratio of 0.33
that a single draw from P (single network output) in fact came from the PME fit P versus
P ; a likelihood < 0.01 is reached for 4 draws.
We next considered pairwise inputs for N > 3 cells by adopting a ring structure with
nearest-neighbor common inputs (illustrated in Figure 1). For unimodal inputs, we computed
DKL(P, P ) while varying σ and Θ; for bimodal inputs, we varied the probability of each
Bernoulli input. Figure 11 shows the maximal DKL(P, P ) per neuron. Overall, circuits
with bimodal pairwise inputs showed appreciable values that are about half of that found
for bimodal global inputs. The relatively large deviation at N = 3 receded, replaced by
deviations that were similar to those seen for global, unimodal inputs. For pairwise unimodal
inputs, values of DKL(P, P ) remained very small.
Finally, we extended our recurrent model to N > 3 cells. As for the three cell networks,
we explored a range of σ, c, and Θ for unimodal inputs, or s, p, and Θ for bimodal inputs.
In addition, the coupling strength, g, was varied for each type of input. Coupling was either
all-to-all (global) or applied pairwise in a ring structure (local). Figure 12 shows the maximal
DKL(P, P ) per neuron, for each type of input and network architecture, up to population
size N = 8. Recurrent coupling increases the available range of higher order interactions
in most settings from the level achieved with purely feed forward connections. However,
the maximal value of DKL(P, P ) per cell is steady or decreasing with N , suggesting that
recurrence has less impact as population size grows.
To summarize, the greater impact of bimodal vs. unimodal input statistics on maximal
values of DKL(P, P ) that can be obtained in a given circuit persists from N = 3 up to
N = 16 (tested up to N = 8 in the recurrent case). Moreover, agreeing with intuition, global
inputs of a given type can generate greater deviations than purely pairwise inputs (with
the exception of one case, N = 3). Overall (again excepting this one case), for the circuit
parameters producing maximal deviations, it becomes easier to statistically distinguish be-
tween spiking distributions and their PME fits as N increases in feedforward networks. In
contrast, the maximal attainable DKL(P, P ) per cell appears to decrease with N in many
recurrent networks.
Discussion
We use simple mechanistic models to identify which combinations of network architectures
and input signals produce spike patterns with higher-order interactions and which do not.
Deviations in circuit outputs from pairwise maximum entropy (PME) predictions were much
smaller than the maximum theoretically attainable values for a general spiking pattern.
19
Moreover, output statistics were not simply related to network architecture. Nonetheless,
several simple principles emerge that determine the strength of higher-order interactions.
First, bimodal input distributions produced stronger higher-order interactions than unimodal
distributions. Second, networks with shared inputs among all cells produced greater higher-
order interactions than those with pairwise inputs (except for the case of three cell networks
receiving bimodal input). Third, recurrent excitatory or gap junction coupling could produce
a further, moderate increase higher-order correlations; the effect was greatest for coupling of
intermediate strength. Our overall results held for networks with nonlinear integrate-and-fire
units based on measured properties of retinal ganglion cells. Together with the facts that
ON parasol cell filtering suppresses bimodality in light input, and that coupling among ON
parasol cells is relatively weak, our findings provide an explanation for why their population
activity is well captured by PME models.
Comparison with empirical studies
How do our maximum entropy fits compare with empirical studies? In terms of DKL(P, P )
-- equivalently, the logarithm of the average relative likelihood that a sequence of data drawn
from P was instead drawn from the model P -- numbers obtained from our RGC models are
very similar to those obtained by experiments on retinal ganglion cells [5, 6]. We find that
DKL(P, P ) = 2.90× 10−5 bits under constant light conditions, compared to an experimental
value of 0.0008 [5] (inferred from a reported likelihood ratio of 0.99944). Under full-field,
time varying light conditions, as well as spatiotemporally varying stixel simulations, we find
average log-likelihood ratios of up to one order of magnitude larger -- bounded above by 0.007.
We can view this as a model of the checkerboard experiments of [5], for which close fits by
PME distribution were also observed (likelihood numbers were not reported). Similarly, the
values of ∆ that are produced by our RGC model are close to those found by [5, 7] under
comparable stimulus conditions. We obtain ∆ = 99.5% (for cell group size N = 3) under
constant illumination, which is near the range reported by [5] (97.8 − 99.2%, N = 3 − 7).
For full-field stimuli we find a range of numbers from 95.7 − 99.3% (N = 3).
The simple threshold models that we have developed, meanwhile, give us a roadmap for
how circuits could be driven in such a way as to lower ∆. Figure 5D-F show ∆ plotted
as a function of firing rate for the data presented in Figure 5A-C: circuits of N = 3 cells
receiving global common inputs. We observe that ∆ ≈ 1 for Gaussian and skewed inputs
over a broad range of firing rates and pairwise correlation coefficients, but that values of ∆
can be depressed by 10-15% in the presence of a bimodal common input. Indeed, Shlens
et al. [5] showed that adding global bimodal inputs to a purely pairwise model can lead to
a comparable departure in ∆. Our results are consistent with this finding, and explicitly
demonstrate that the bimodality of the inputs -- as well as their global projection -- are
characteristics that lead to this departure.
While meaningful in an experimental study with non-neglible pairwise correlations, we
caution that using ∆ as a metric can be problematic when an idealized circuit is explored
over its full range of parameters, because it may flag "uninteresting" cases in which cells are
nearly independent, and a pairwise model adds little additional value. Specifically, if Dind
20
is small (i.e., the true distribution is well-approximated by the independent distribution
P1), then ∆ may be appreciably far from 1 although Dpair is small. Thus, a poor pairwise
maximum entropy fit, as measured by ∆ (that is, ∆ < 1) is not necessarily indicative of a
poor performance in DKL(P, P ). For example, in the bimodal common input case (Figure
5F), the very lowest values of ∆ are achieved for low correlation ρ; in essence, when the
independent model already does a good job of representing the output distribution. As
suspected this performance is not reflected in Figure 5C, where low correlation gives low
DKL(P, P ). In summary, ∆ can be as low as 0.5 for distributions that are barely perceptibly
different when measured by the Kullback-Leibler divergence.
Figure 13A-B extend our observations to a circuit of N = 12 cells forced by Gaussian
and skewed inputs respectively. We find that small ∆ occurs for σ (cid:28) 1, coincident with
low population firing rates. Here P is nearly independent (because it is dominated by 0
spiking events, the distribution is well-modeled by independent non-spiking neurons) and
the improvement of the pairwise model over the independent model is negligible. The low
population firing rate regime (N ν < 1, where ν is the single cell firing rate per time bin and
N is the population size) is also the regime where 1− ∆ is linear in N − 2 [17]. If a nontrivial
deviation of ∆ (from 1) is observed for N ν < 1 (such as, in Figure 5, N = 3), then 1 − ∆
must continue to grow as N grows; equivalently, ∆ must decrease. The growth of 1− ∆ with
N for particular points in this region is illustrated in Figure 13C.
Using correlation structure to infer anatomical structure
We address two questions about the relationship between the architecture of feedforward
circuits and the statistical structure of the spike patterns that they produce, based on our
comparisons between global-input and pairwise nearest-neighbor network architectures in
the Results.
First, if a circuit produces spike patterns that deviate substantially from pairwise maximum-
entropy (PME) predictions, can we conclude that it has beyond-pairwise anatomical pro-
jections -- that is, common inputs received by more than two cells? For a small group of
N = 3 cells, the answer is no: we find that, among all cases we study, the largest deviation
from PME predictions occurs for purely pairwise (binary) inputs, so that departures from
PME models do not imply departures from pairwise nearest-neighbor network architectures.
For larger N , the answer is a qualified yes: for marginal statistics of a given type, we show
in Figure 11 that the greatest deviations from PME models do correspond to global com-
mon (as opposed to purely pairwise) inputs. However, without knowing input marginals,
values of DKL are still not predictive of anatomy: for example, if N = 16, then roughly the
same values of DKL(P, P ) follow from global inputs with uniform marginals as for pairwise
nearest-neighbor inputs with binary marginals.
Second, if a circuit produces spike patterns that are well-described by PME models, does
this imply that it has a pairwise architecture? Again the answer is no, as the success of
PME models depends on N and marginal statistics. For N > 3 and known input marginals,
better fits by PME imply pairwise nearest-neighbor connectivity; otherwise, such inferences
cannot be made.
21
Scope and open questions
Our first set of findings are for a set of circuit models with a simple thresholding nonlinearity
at each cell. These models were chosen to be simple enough to allow analytical insights and
a complete parametric study.
While our retinal ganglion cell model demonstrates that these findings, based on a simple
threshold model, do carry over to describe the spiking statistics of a more realistic spiking
model (here, a time-dependent, nonlinear integrate-and-fire system), there are many aspects
of circuits left unexplored by our study. Prominent among these is heterogeneity. Only a
few of our simulations produce heterogeneous inputs to model RGCs, and all of our stud-
ies apply to cells with identical response properties. This is in contrast to studies such
as [7] which examine correlation structures among multiple cell types. For larger networks,
feedforward connections with variable spatial profiles also occur, between the extremes of
"nearest neighbor" and global input connections examined here. It is also possible that more
complex input statistics could lead to greater higher-order interactions [37]. Finally, Figure
11 indicates that some trends in DKL(P, P ) vs. N appear to become nonlinear for N (cid:39) 10;
for larger networks, our qualitative findings could change.
Our study also leaves largely open the role of different retinal filters in generating higher-
order interactions. We have found that the specific filtering properties of ON parasol cells
suppress bimodality in light inputs, suggesting that other classes of retinal ganglion cells
may produce more robust higher-order interactions (compare Figures 7D, E). This predicts
a specific mechanism for the development of higher-order interactions in preparations that
include multiple classes of ganglion cells [7]. Finally, we considered circuits with a single
step of inputs and simple excitatory or gap junction coupling; a plethora of other network
features could also lead to higher-order interactions, including multilayer feedforward struc-
tures, together with lateral and feedback coupling. We speculate that, in particular, such
mechanisms could contribute to the higher-order interactions found in cortex [8, 10, 11, 38].
A final outstanding area of research is to link tractable network mechanisms for higher-
order interactions with their impact (or lack of impact) on information encoded in neural
populations [10, 38]. A simple starting point is to consider rate-based population codes in
which each stimulus produces a different "tuned" average spike count (see, e.g., Ch. 3 of [39]).
One can then ask whether spike responses can be more easily decoded to estimate stimuli
for the full population response (i.e., P ) to each stimulus or for its pairwise approximation
( P ). In our preliminary tests where higher-order correlations were created by inputs with
bimodal distributions, we found examples where decoding of P vs. P differed substantially.
However, a more complete study would be required before general conclusions about trends
and magnitudes of the effect could be made; such a study would include complementary
approach in which the full spike responses P are themselves decoded via a "mismatched"
decoder based on the pairwise model P [38]. Overall, we hope that the present paper, as
one of the first that connects circuit mechanisms to higher-order statistics of spike patterns,
will contribute to future research that takes these next steps.
22
Materials and Methods
DKL and distance from PME surface
We can see that the curve along which µ and ρ are constant -- the iso-moment line -- remains
a line in the (fp, f1p, f1m) coordinate space by inverting the equations µ = p1 + 2p2 + p3,
ρ = p2 + p3, and 1 = p0 + 3p1 + 3p2 + p3. The first coordinate, fp ≡ p0 + p3 = 3(ρ − µ) − 1,
is constant for a fixed (µ, ρ). Similarly, f1m and f1p can be written as linear functions of the
remaining free parameter.
To see that DKL(P, P ) is convex along an iso-moment line, we consider DKL(P, P ) as P
varies so as to remain on an iso-moment line. Letting f1m and f1p be the coordinates of the
PME fit, and defining dm = f1m − f1m and dp = f1p − f1p, we find
where dx is an increment of distance along the iso-moment line. Inverting Equations 3 and
substituting the results into the definition of DKL, we can write
(cid:33)
(cid:19)
1 − dp
1 − f1p
(cid:18)
(cid:19)
1 − dm
1 − f1m
dp =
dm =
(cid:113)
(cid:113)
(cid:32)
(cid:18)
dp
f1p
(cid:18)
dx
dx
(fp − 1)/3
p + (1 − fp)2/9
f 2
fp
p + (1 − fp)2/9
f 2
(cid:32)
(cid:33)
(cid:19)
1 − f1p
1 − dm
1 − f1m
1 +
log
(cid:18)
(cid:18)
log
(cid:19)
(cid:32)
dm
f1m
1 +
(cid:33)
log
dm
f1m
(cid:33)
dp
f1p
(cid:18)
(cid:19)
dx
β
(cid:19)
dx
β
DKL(P, P ) = fp(1 − f1p)
1 − dp
+ (1 − fp)(1 − f1m)
+ (1 − fp) f1m
(cid:32)
+ fp
f1p
1 +
log
1 +
.
(10)
This is a convex function of dx; we can see this by observing that each of the four terms is
a function of the form
F (dx) = α
1 +
log
1 +
(in the first term, for example, we have α = fp(1 − f1p) and β = −(1 − f1p)). This can be
readily shown to be convex by taking the second derivative with respect to dx and verifying
that it is positive:
(cid:16)
α
1 + dx
β
(cid:17),
F (cid:48)(cid:48)(dx) =
β2
23
where we can verify that α > 0 and dx < β. The sum of convex functions is likewise convex.
Because DKL(P, Q) is non-negative for any distributions P and Q, DKL(P, P ) achieves its
unique minimum along an iso-moment line at P = P , and it must monotonically increase as
a function of dx.
To get an intuitive picture of DKL(P, P ) as it varies in the (fp, f1p, f1m)-coordinate space,
we view this quantity along constant-fp slices (Figure 2). DKL(P, P ) increases with distance
from the constraint curve. Generally, the range of this distance peaks at fp = 0.25.
To further quantify the relationship between distance and DKL, we approximate the
logarithms in Equation 10 for small arguments and find that DKL increases quadratically
with distance for small arguments:
DKL(P, P ) ≈ (dx)2C(fp, f1m) + O(dx3)
(11)
where
C(fp, f1m) =
+
fp(1 − fp)2/9
p + (1 − fp)2/9
f 2
p (1 − fp)
f 2
p + (1 − fp)2/9
f 2
1m)2
(1 − 3 f1m + 3 f 2
1m(1 − f1m)3
f 3
1
f1m(1 − f1m)
Numerical sampling of 3 cell network
For general circuit set-ups, it may be necessary to probe the output distribution by sampling.
In the case of global input, however, it is more computationally efficient and accurate to
compute the output spiking probability distribution using quadrature. To be concrete, a set
of N = 3 threshold spiking units is forced by a common input Ic (drawn from a probability
distribution PC(y)) and an independent input Ij (drawn from a probability distribution
PI(y)). The output of each cell xj is determined by summing and thresholding these inputs:
xj = H(Ij + Ic − Θ)
Conditioned on Ic, the probability of each spike is given by:
Prob[xj = 1 Ic = a] = Prob[Ij + a − Θ > 0]
= Prob[Ij > Θ − a]
=
PI(y) dy
Similarly, we have the conditioned probability that xj = 0:
Prob[xj = 0 Ic = a] = Prob[Ij + a − Θ < 0]
= Prob[Ij < Θ − a]
(cid:90) ∞
Θ−a
(cid:90) Θ−a
−∞
PI(y) dy
=
24
Because these are conditionally independent, the probability of any spiking event (x1, x2, x3) =
(A1, A2, A3) is given by the integral of the product of the conditioned probabilities against
the density of the common input.
Prob[x1 = A1, x2 = A2, x3 = A3] =
dy PC(y)
Prob[xj = Aj Ic = y]
(12)
(cid:90) ∞
−∞
3(cid:89)
j=1
The integrand in the previous equation is numerically evaluated via an adaptive quadrature
routine, such at Matlab's quad. This is easily generalized to an arbitrary number of cells N .
Unimodal inputs Ij, Ic were chosen from different marginals with mean 0 and variance
σ2. For Gaussian input, P (x) ∝ e−x2/2σ2; for uniform inputs, P (x) ∝ 1 for x <
3σ2 and 0
otherwise. For skewed input, P (x) ∝ (x + µ)e−(x+µ)2/2a, for x > −µ, where the parameter a
sets the variance 2a(1 − π
2 ensures that the mean of P (x) is zero.
When sampling was necessary in thresholding models, for N ≤ 14 the number of samples
was chosen so that the mean bin occupancy of each unique circuit output was > 100 (and
usually > 1000). For N = 16 -- the most poorly resolved case -- the mean bin occupancy
was lower ( 40).
4 ) and shifting by µ =(cid:112) aπ
√
Bimodal triplet inputs always generate distributions with negative
strain
Another information theoretic quantity that relates to the ability of a distribution to be
characterized by a pairwise model is the strain:
(cid:18) P (1, 1, 1)P (1, 0, 0)P (0, 1, 0)P (0, 0, 1)
(cid:19)
γ =
1
8
log
P (0, 0, 0)P (1, 1, 0)P (1, 0, 1)P (0, 1, 1)
.
Indeed, this quantity must be zero for any distribution that satisfies the PME constraint
(Equation 2). Negative values of strain occur to the left side of the PME constraint curve
in the (f1p, f1m)-plane, whereas positive values occur to the right.
For a circuit forced by common binary inputs, the simplicity of our setup allows us to
show why observed distributions occur to the left side of the PME constraint curve. We
approach this by showing that given the f1m coordinate of an observed distribution, the
f1p coordinate is less than the PME fit would predict. A point on the constraint surface
corresponding to a particular value of f1m may be written
f1p =
f 3
1m
1 − 3f1m + 3f 2
1m
whereas
=
q3
q3 + (1 − q)3
f1p =
pq3
pq3 + (1 − p) + p(1 − q)3
25
and
1
f1p
=
=
This makes it clear that 1
f1p
≥ 1
f1p
1 − p + p(q3 + (1 − q)3)
1 − p
pq3 +
pq3
1
f1p
with equality if and only if p = 1. Therefore
f1p < f1p
(13)
unless p = 1, in which case they coincide.
According to Equation 13, p0 is under-predicted by the PME model, whereas p3 is over-
predicted ; this is precisely the condition of "sparse coding" [11, 25].
An analytical explanation for unimodal vs. bimodal effects
We consider an analytical argument to support the our numerical results that bimodal inputs
generate larger deviations from PME model fits than unimodal inputs. As a metric, we
consider DKL(P, P ) -- where P and P are again the true and model distributions respectively
-- when we perturb an independent spiking distribution by adding a common, global input
of variance c. To simplify notation, the small parameter in the calculation will be denoted
=
√
We begin by observing that when P is a maximum entropy distribution; that is, its
logarithm is given as a sum over functionals whose averages over P must also be satisfied by
P , or
c.
log(p(x)) =
p(x)fµ(x) =
the KL-distance may be written as a difference of entropies:
(cid:88)
x
p(x)fµ(x);
(cid:33)
µ
x
p(x)
p((cid:126)x) log
(cid:88)
(cid:18) p(x)
λµfµ(x) − log Z,
(cid:88)
(cid:19)
DKL(P, P ) ≡ (cid:88)
(cid:88)
p(x) log(p(x)) −(cid:88)
(cid:32)
= −S(P ) −(cid:88)
= −S(P ) + log Z −(cid:88)
= −S(P ) + log Z −(cid:88)
= −S(P ) −(cid:88)
p(x)
λµ
λµ
µ
µ
x
x
=
x
p(x) log(p(x))
p(x) log(p(x))
(cid:88)
x
− log Z +
λµfµ(x)
(cid:88)
(cid:88)
x
x
µ
p(x)fµ(x)
p(x)fµ(x)
= −S(P ) + S( P ).
x
26
Here, the entropy of a probability distribution P is given
S(P ) = −p0 log(p0) − 3p1 log(p1) − 3p2 log(p2) − p3 log(p3)
(14)
if we use the fact that the distributions are permutation-symmetric (i.e. p1 ≡ P (1, 0, 0) =
P (0, 1, 0) = P (0, 0, 1)). We take the logarithms in the definitions of entropy S and KL-
divergence DKL to be base 2, so that any numerical values of these quantities are in units
of bits.
We now compute S(P ) and S( P ) by deriving a series expansion for each set of event prob-
abilities. We can compute the true distribution P using the expressions derived in Equation
12; to recap, let the common input have probability density p(c), and the independent input
to each cell have density ps(x). Let θ be the threshold for generating a spike (i.e., a "1"
response). For each cell, a spike is generated if x + c > θ, i.e., with probability
(cid:90) ∞
θ−c
d(c) =
ps(x)dx .
Given c, this is conditionally independent for each cell. We can therefore write our proba-
bilities by integrating over c as follows:
−∞
(cid:90) ∞
(cid:90) ∞
(cid:90) ∞
(cid:90) ∞
−∞
−∞
−∞
p0 =
p1 =
p2 =
p3 =
p(c)(1 − d(c))3 dc
p(c)d(c)(1 − d(c))2 dc
p(c)d(c)2(1 − d(c)) dc
p(c)d(c)3 dc
(15)
(cid:16)c
(cid:17)
p(c) =
1
f
We develop a perturbation argument in the limit of very weak common input. That is, p(c)
is close to a delta function centered at c = 0. Take p(c) to be a scaled function
(16)
(cid:82) ∞
We place no constraints on f (c), other than that it must be normalized (E[1] = 1) and
that its moments must be finite (so E[c], E[c2], and so forth must exist, where E[g(c)] ≡
−∞ g(c)f (c) dc).
For the moment, assume that the function f (c) has a single maximum at c = 0. To
evaluate the integrals above, we Taylor-expand d(c) around c = 0. Anticipating a sixth-
order term to survive, we keep all terms up to this order. This gives, for small y,
d(y) ≈ d(0) +
akyk + O(y7)
6(cid:88)
k=1
27
where a1 = ps(θ) (the other coefficients a2-a6 can be given similarly in terms of the indepen-
dent input distribution at θ). Substituting this into the expressions for p0, etc., above, with
p(c) given as in Equation 16, gives us each event as a series in ; for example,
0 +(cid:0)3a1d2
0 E[c](cid:1) +(cid:0)(3a2
p3 = d3
0) E[c2](cid:1) 2 + ...
1d0 + 3a2d2
The entropy S(P ) is now given by using these series expansions in Equation 14.
We note that our derivation does not rely on the fact that the distribution of common
input is peaked at c = 0 in particular. For example, we could have a common input centered
around µ. The common input distribution function would be of the form
(cid:18) c − µ
(cid:19)
p(c) =
1
f
Changing regulates the variance, but doesn't change the mean or the peak (assuming,
without loss of generality, that the peak of f occurs at zero). The peak of p(c) now occurs
at µ, and the appropriate Taylor expansion of d(y) is
6(cid:88)
d(y) ≈ d(µ) +
bk(y − µ)k + O(y7),
k=1
where the coefficients bk now depend on the local behavior of d around µ. The expectations
that appear in the expansion of p3, and so forth, are now centered moments taken around
µ; the calculations are otherwise identical. In other words, perturbation expansion requires
the variance of the common input to be small, but not the mean.
For bimodal inputs, we consider a common input with a probability distribution of the
following form:
p(c) = (1 − 2)
1
f
(cid:16)x
(cid:17)
+ 2 1
f
(cid:18) x − 1
(cid:19)
so that most of the probability distribution is peaked at zero, but there is a second peak of
higher order (here taken at c = 1, without loss of generality). Again, we approximate the
integrals given in Equations 15, and therefore the entropy S(P ), by Taylor expanding d;
6(cid:88)
6(cid:88)
k=1
d(c) ≈ d(0) +
≈ d(1) +
akck + O(c7);
(c ≈ 0)
bk(c − 1)k + O((c − 1)7);
(c ≈ 1)
k=1
around the two peaks 0 and 1 respectively. For each integral we have the same contributions
from the unimodal case, multiplied by (1 − 2), as well as the corresponding contributions
from the second peak multiplied by 2 (these weightings are chosen so that the common
28
input has variance of order 2, as in the unimodal case). This makes clear at what order
every term enters.
We now construct an expansion for the PME model P :
P (x1, x2, x3) =
1
Z
exp (λ1(x1 + x2 + x3) + λ2(x1x2 + x2x3 + x1x3))
We approach this problem by describing λ1 and λ2 as a series in . We match coefficients by
forcing the first and second moments of P to match those of P -- as they must. Specifically,
take
6(cid:88)
λ1 = λ +
kuk + O(7)
6(cid:88)
k=1
kvk + O(7)
λ2 =
where λ1 = λ, λ2 = 0 are the corresponding parameters from the independent case. The
events p0, p1, p2 and p3 can be written as a series in . We then require that the mean and
centered second moments of P match those of P ; that is
k=1
p1 + 2p2 + p3 = p1 + 2p2 + p3
p2 + p3 − (p1 + 2p2 + p3)2 = p2 + p3 − (p1 + 2p2 + p3)2.
At each order k, this yields a system of two linear equations in uk and vk; we solve, inductively,
up to the desired order; we now have P , and therefore S( P ), as a series in .
Finally, we combine the two series to find that in the unimodal case,
DKL(P, P ) = S( P ) − S(P )
(cid:20)a6
(cid:21)
= 6
1(2 E[c]3 − 3 E[c] E[c2] + E[c3])2
2(1 − d0)3d3
0
+ O(7)
(17)
If the first two odd moments of the distribution are zero (something we can expect for
"symmetric" distributions, such as a Gaussian), then this sixth-order term is zero as well.
For the bimodal case
DKL(P, P ) = S( P ) − S(P )
(cid:20) (d1 − d0)6
2(1 − d0)3d3
0
(cid:21)
+ O(5)
= 4
This last term depends on the distance d1 − d0, in other words, how much more likely the
independent input is to push the cell over threshold when common input is "ON". We
can also view this as depending on the ratio d1−d0
, which gives the fraction of previously
1−d0
non-spiking cells that now spike as a result of the common input.
The main point here, of course, is that DKL(P, P ) is of order 4 rather than 6. So, as
the strength of a common binary vs. unimodal input increases, spiking distributions depart
from the PME more rapidly.
29
Experimentally-based model of a RGC circuit
We model the response of a individual RGC using data collected from a representative
primate ON parasol cell, following methods in [32, 35]. Similar response properties were
observed in recordings from 16 other cells. To measure the relationship between light stimuli
and synaptic conductances, the retina was exposed to a full-field, white noise stimulus. The
cell was voltage clamped at the excitatory (or inhibitory) reversal potential VE = 0 mV
(VI = −60 mV), and the inhibitory (or excitatory) currents were measured in response to
the stimulus. These currents were then turned into equivalent conductances by dividing by
the driving force of ±60 mV; in other words
I exc = gexc(V − VE);
I inh = ginh(V − VI);
V − VE = −60 mV
V − VI = 60 mV
The time-dependent conductances gexc and ginh were now injected into the same cell using
a dynamic clamp (i.e., input current was varied rapidly to maintain the correct relationship
between the conductance and the membrane voltage) and the voltage was measured at a
resolution of 0.1 ms.
To model the relationship between the light stimulus and synaptic conductances, the
current measurements I exc and I inh were fit to a linear-nonlinear model:
I exc(t) = N exc[Lexc ∗ s(t) + ηexc],
I inh(t) = N inh[Linh ∗ s(t) + ηinh]
where s is the stimulus, Lexc (Linh) is a linear filter, N exc (N inh) is a nonlinear function, and
ηexc (ηinh) is a noise term. The linear filter was fit by the function
Lexc(t) = Pexc(t/τexc)nexc exp(−t/τexc) sin(2πt/Texc)
(18)
and the nonlinear filter by the polynomial
N exc = Aexcx2 + Bexcx + Cexc;
k , ηinh
Linh and N inh were fit using the same parametrization. The noise terms ηexc
k were fit to
reproduce the statistical characteristics of the residuals from this fitting. We simulated the
noise terms ηexc and ηinh using Ornstein-Uhlenbeck processes with the appropriate parame-
ters; these were entirely characterized by the mean, standard deviation, and time constant
of autocorrelation τη,exc (τη,inh), as well as pairwise correlation coefficients for noise terms
entering neighboring cells. The noise correlation coefficients were estimated from the dual
recordings of [32].
Linear filter parameters computed were Pexc = −8 × 104 pA/s, nexc = 3.6, τexc = 12
ms, Texc = 105 ms, and Pinh = −1.8 × 105 pA/s, ninh = 3.0, τinh = 16 ms, Texc = 120
ms. Nonlinearity parameters were Aexc = −5 × 10−5, Bexc = 0.42, Cexc = −57, and Ainh =
1 × 104, Binh = 0.37, Cinh = 250. Noise parameters were measured to be mean(ηexc
k ) = 30,
30
k ) = 500, τη,exc = 22 ms, and mean(ηinh
std(ηexc
addition, excitatory (inhibitory) noise to different cells ηexc
coefficient of 0.3 (0.15).
the filter function; i.e. Lexc,R(t) = Lexc(t).
k ) = −1200, std(ηinh
k , ηexc
j
k ) = 780, τη,inh = 33 ms. In
(ηinh
) had a correlation
k , ηinh
j
To obtain the rectified filter shown in Figure 7E, we simply took the absolute value of
Model fitting
We create a model of the cell as a nonlinear integrate-and-fire model using the method of
Badel et al. [33], in which the membrane voltage is assumed to respond as
dV
dt
= F (V, t − tlast) +
Iinput(t)
C
(19)
where C is the cell capacitance, tlast is the time of the last spike before time t, and Iinput(t)
is a time-dependent input current. We use the current-clamp data, which yields cell voltage
in response to the input current Iinput(t) = gexc(t)(V − VE) + ginh(V − VI), to fit a function
F (V, t). When voltage data is segregated according to the (binned) time since the last spike,
the I − V curve is well fit by a function of the form
(cid:0)EL − V + ∆T e(V −VT )/∆T(cid:1)
F (V ) =
1
τm
(20)
The membrane time constant τm, resting potential (EL), spike width ∆T and knee of the
exponential curve VT are parameterized as a function of t−tlast. Our model neuron comprises
Equations (19, 20) for V < Vthreshold; a spike was detected when V reached Vthreshold = −30
mV, with a voltage reset to Vreset = −55 mV after 2 ms. In addition, the cell was unable to
spike for an absolute refractory period of τabs = 3 ms.
The capacitance was inferred from the voltage trace data by finding, at a voltage value
where the voltage/membrane current relationship is approximately Ohmic, the value of C
that minimizes error in the relation Equation 19 [33]. The estimated value was C = 28 pF.
Recurrent model
Gap junction coupling was introduced as an additional current on the right-hand side of
Equation 19:
Igap,j
C
= −ggap
C
(Vj − Vk)
(cid:88)
k(cid:54)=j
The coupling strength ggap was held constant during a simulation. When coupling was
present (i.e. when ggap (cid:54)= 0), ggap was varied from a baseline level (1.1 nS) [32] to 16 times
this value (17.6 nS) between simulations.
For simulations that include electrotonic coupling, the spike trajectory was modeled with
greater care. A typical (averaged) spike waveform was extracted from voltage traces of a
31
primate ON parasol cell. The spike waveform was used to replace 1 ms of the membrane
voltage trajectory during and after a spike; at the end of the 1 ms, the voltage was released
at approximately −58 mV. The cell was unable to spike for an absolute refractory period of
τabs = 3 ms. A relative refractory period was induced by introducing a declining threshold
for the period of 3-6 ms following a spike, after which Vth limits to -30 mV.
Cell receptive field
We defined each cell's stimulus as the linear convolution of an image with its receptive field.
The receptive fields include an "on" center and an "off" surround, as in [40]:
sj((cid:126)x) = exp
((cid:126)x − (cid:126)xj)Q((cid:126)x − (cid:126)xj)
− k exp
r((cid:126)x − (cid:126)xj)Qr((cid:126)x − (cid:126)xj)
(cid:18)
−1
2
(cid:19)
(cid:18)
−1
2
(cid:19)
where the parameters k and 1/r give the relative strength and size of the surround. Q
specifies the shape of the center and was chosen to have a 1 standard deviation (SD) radius
of 50 µm and to be perfectly spherical. The receptive field locations (cid:126)x1, (cid:126)x2, and (cid:126)x3 were
chosen so that the 1 SD outlines of the receptive field centers will tile the plane (i.e. they
just touch). Other parameters used were k = 0.3, r = 0.675.
Numerical methods and convergence testing
For each simulation, Equations 19, 20 were integrated using the Euler method for > 105 ms
with a time step of 0.1 ms. For each stimulus condition, 20 simulations (or sub-simulations)
were run, for a total integration time of > 20 × 105 ms. The synaptic noise terms, ηexc
and
ηinh
k , as well as the light input, were generated independently for each sub-simulation. To
discretize spiking outputs, 5 ms bins were used.
k
These 20 sub-simulations were used to estimate standard errors in both the probabil-
ity distribution over spiking events and DKL(P, P ). For example, in the constant light
case, we generated the following distribution on spiking events: P (0, 0, 0) = 0.816 ± 0.004,
P (0, 0, 1) = 0.0457 ± 0.0015; P (0, 1, 0) = 0.0448 ± 0.0015, P (1, 0, 0) = 0.0459 ± 0.0017,
P (0, 1, 1) = 0.00554± 0.00054, P (1, 0, 1) = 0.00545± 0.00051, P (1, 1, 0) = 0.00545± 0.00056,
and P (1, 1, 1) = 0.00116 ± 0.00020. Numbers reported in the Results are, unless specified
otherwise, produced by collating the data from the 20 simulations.
To test our finding that the observed distributions were well-modeled by the PME fit, we
also performed the PME analysis on each of the 20 simulations for each stimulus condition.
While in general DKL(P, P ) can be quite sensitive to perturbations in P , the numbers re-
mained small under this analysis. To confirm that our results for DKL(P, P ) are sufficiently
resolved to remove bias from sampling, we performed an analysis in which we collect the 20
simulations in subgroups of 1, 2, 4, 5, 10, and 20, and plot the mean DKL with estimated
standard errors. As expected (e.g. [41]), bias decreases as the length of subgroup increases
and asymptotes at -- or before -- the full simulation length. Results are shown in Figure
S1 for the RGC simulations under full-field stimulation, as well as two representative cases
with "stixel" stimuli.
32
For the thresholding model, sampling was used for pairwise input cases in Figure 11,
and all data presented in Figure 12. In all but three simulations, the sampling frequency
was chosen so that the mean bin occupancy (that is, the average frequency of each unique
spiking pattern) was > 100 (and usually > 1000). The exceptions are for the cases where
the network size N = 16 and unimodal pairwise inputs are used, for which the mean bin
occupancy was ≈ 40. We show the bias for N = 16, pairwise Gaussian inputs in Figure S1A,
which confirms that the bias is close to its asymptotic value of zero.
To provide a cross-validation test, we divided our data into halves (which we denote P1
and P2, each including data from 10 subsimulations) and performed the PME analysis on
one half (say P1) to yield a model P1. We then computed DKL(P2, P1) and DKL(P2, P1)
(as in [42]), which we refer to the cross-validated and empirical likelihood respectively. The
former tests whether the PME fit is robust to over-fitting; the latter tests how well-resolved
our "true" distribution is in the first place. In Figure S2, we plot these numbers vs. the
numbers reported in the paper. Most cross-validated likelihoods fall on or near the identity
line; most empirical likelihoods are close to zero (and importantly, significantly smaller than
either DKL(P, P ) or DKL(P2, P1), indicating that DKL(P, P ) is accurately resolved). We
conclude that the deviations that we observe when these conditions are met can not be
accounted for by the differences in testing and training data.
Finally, to test the robustness of our finding that the strain γ < 0 for the full-field
RGC simulations, we perturbed our spiking event distributions randomly, with perturbations
weighted by the estimated standard errors. This was repeated 20 times for each stimulus
condition. Out of the resulting 840 perturbations (20 each for 42 stimulus conditions), γ > 0
in only 22 trials.
Financial Disclosure
This research was supported by NSF grant DMS-0817649 and by a Career Award at the
Scientific Interface from the Burroughs-Wellcome Fund (ESB), by the Howard Hughes Med-
ical Institute and by NIH grant EY-11850 (FMR), and a Cambridge Overseas Research
Studentship (JG). The funders had no role in study design, data collection and analysis,
decision to publish, or preparation of the manuscript.
33
References
[1] Baudry M, Taketani M, editors (2006) Advances in Network Electrophysiology Using
Multi-Electrode Arrays. New York: Springer Press.
[2] Amari S (2001) Information geometry on hierarchy of probability distributions. IEEE
Transactions on Information Theory .
[3] Martignon L, Deco G, Laskey K, Diamond M, Freiwald W, et al. (2000) Neural coding:
higher-order temporal patterns in the neurostatistics of cell assemblies. Neural Comp
12: 2621 -- 53.
[4] Schneidman E, Still S, Berry (II) MJ, Bialek W (2003) Network information and con-
nected correlations. Physical Review Letters 91: 238701.
[5] Shlens J, Field GD, Gauthier JL, Grivich MI, Petrusca D, et al. (2006) The structure of
multi-neuron firing patterns in primate retina. Journal of Neuroscience 26: 8254 -- 8266.
[6] Shlens J, Field GD, Gauthier JL, Greschner M, Sher A, et al. (2009) The structure
of large-scale synchronized firing in primate retina. Journal of Neuroscience 29: 5022 --
5031.
[7] Schneidman E, Berry (II) MJ, Segev R, Bialek W (2006) Weak pairwise correlations
imply strongly correlated network states in a neural population. Nature 440: 1007 -- 1012.
[8] Tang A, Jackson D, Hobbs J, Smith JL, Patel H, et al. (2008) A maximum entropy model
applied to spatial and temporal correlations from cortical networks in vitro. Journal of
Neuroscience 28: 505 -- 518.
[9] Yu S, Huang D, Singer W, Nikolic D (2008) A small world of neuronal synchrony.
Cerebral Cortex 18: 2891 -- 2901.
[10] Montani F, Ince RAA, Senatore R, Arabzadeh E, Diamond ME, et al. (2009) The
impact of high-order interactions on the rate of synchronous discharge and information
transmission in somatosensory cortex. Philosophical Transactions of the Royal Society
A 367: 3297 -- 3310.
[11] Ohiorhenuan IE, Mechler F, Purpura KP, Schmid AM, Hu Q, et al. (2010) Sparse coding
and high-order correlations in fine-scale cortical networks. Nature 466: 617 -- 21.
[12] Santos GS, Gireesh ED, Plenz D, Nakahara H (2010) Hierarchical interaction structure
of neural activities in cortical slice cultures. The Journal of Neuroscience 30: 8720 -- 8733.
[13] Ganmor E, Segev R, Schneidman E (2011) Sparse low-order interaction network un-
derlies a highly correlated and learnable population code. Proceedings of the National
Academy of Sciences USA 108: 9679 -- 9684.
34
[14] Amari S, Nakahara H, Wu S, Sakai Y (2003) Synchronous firing and higher-order inter-
actions in neuron pool. Neural Computation 15: 127 -- 142.
[15] Krumin M, Shoham S (2009) Generation of spike trains with controlled auto- and cross-
correlation functions. Neural Comp 21: 1642 -- 64.
[16] Macke JH, Berens P, Ecker AS, Tolias AS, Bethge M (2009) Generating spike trains
with specified correlation coefficients. Neural Computation 21: 397 -- 423.
[17] Roudi Y, Nirenberg S, Latham PE (2009) Pairwise maximum entropy models for study-
ing large biological systems: When they can work and when they cant. PLoS Compu-
tational Biology 5: e1000380.
[18] Roudi Y, Tyrcha J, Hertz J (2009) Ising model for neural data: Model quality and
approximate methods for extracting functional connectivity. Physical Review E 79:
051915.
[19] Bohte SM, Spekreijse H, Roelfsema PR (2000) The effects of pair-wise and higher-order
correlations on the firing rate of a postsynaptic neuron. Neural Computation 12: 153 --
179.
[20] Macke JH, Opper M, Bethge M (2011) Common input explains higher-order correlations
and entropy in a simple model of neural population activity. Physical Review Letters
106: 208102.
[21] Cocco S, Leibler S, Monasson R (2009) Neuronal couplings between retinal ganglion
cells inferred by efficient inverse statistical physics methods. Proceedings of the National
Academy of Sciences USA 106: 14058 -- 14062.
[22] Tkacik G, Schneidman E, Berry II MJ, Bialek W (2009) Spin glass models for a network
of real neurons. ArXiv:0912.5409.
[23] Jaynes ET (1957) Information theory and statistical mechanics. Physiol Rev 106: 620-
630.
[24] Jaynes ET (1957) Information theory and statistical mechanics II. Physiol Rev 108:
171-190.
[25] Ohiorhenuan IE, Victor JD (2010) Information-geometric measure of 3-neuron firing
patterns characterizes scale-dependence in cortical networks. Journal of Computational
Neuroscience : 10.1007/s10827-010-0257-0.
[26] Cover TM, Thomas JA (1991) Elements of information theory. New York: Wiley.
[27] de la Rocha J, Doiron B, Shea-Brown E, Josic K, Reyes A (2007) Correlation between
neural spike trains increases with firing rate. Nature 448: 802 -- 806.
35
[28] Vilela RD, Lindner B (2009) Comparative study of different integrate-and-fire neurons:
Spontaneous activity, dynamical response, and stimulus-induced correlation. Physical
Review E 80: 031909.
[29] Shea-Brown E, Josi´c K, Doiron B, de la Rocha J (2008) Correlation and synchrony
transfer in integrate-and-fire neurons: Basic properties and consequences for coding.
Phys Rev Lett 100: 108102.
[30] Marella S, Ermentrout GB (2008) Class-II neurons display a higher degree of stochastic
synchronization than class-I neurons. Physical Review E 77: 041908.
[31] Barreiro AK, Shea-Brown ET, Thilo EL (2010) Timescales of spike-train correlation for
neural oscillators with common drive. Physical Review E 81: 011916.
[32] Trong PK, Rieke F (2008) Origin of correlated activity between parasol retinal ganglion
cells. Nature Neuroscience 11: 1343 -- 1351.
[33] Badel L, Lefort S, Brette R, Petersen CCH, Gerstner W, et al. (2007) Dynamic I-V
curves are reliable predictors of naturalistic pyramidal-neuron voltage traces. Journal
of Neurophysiology 99: 656 -- 666.
[34] Sharpe LT, Whittle P, Nordby K (1993) Spatial integration and sensitivity changes in
the human rod visual system. J Physiol 461: 235-246.
[35] Murphy GJ, Rieke F (2006) Network variability limits stimulus-evoked spike timing
precision in retinal ganglion cells. Neuron 52: 511-524.
[36] Dacey D, Brace S (1992) A coupled network for parasol but not midget ganglion cells
in the primate retina. Vis Neurosci 9: 279-290.
[37] Bethge M, Berens P (2008) Near-maximum entropy models for binary neural represen-
tations of natural images. In: Advances in Neural Information Processing Systems 20.
pp. 97 -- 104.
[38] Oizumi M, Ishii T, Ishibashi K, Okada M (2010) Mismatched decoding in the brain.
Journal of Neuroscience 30: 4815 -- 4826.
[39] Dayan P, Abbot L (2001) Theoretical Neuroscience: Computational and Mathematical
Modeling of Neural Systems. MIT Press, 1st edition.
[40] Chichilnisky EJ, Kalmar RS (2002) Functional asymmetries in ON and OFF ganglion
cells of primate retina. Journal of Neuroscience 22: 2737 -- 2747.
[41] Paninski L (2003) Estimation of entropy and mutual information. Neural Computation
15: 1191 -- 1253.
[42] Yu S, Yang H, Nakahara H, Santos G, Nikoli´c D, et al. (2011) Higher-order interactions
characterized in cortical activity. Journal of Neuroscience 31: 17514 -- 17526.
36
Figure 1: Schematic showing different axes on which we explore microcircuits in this study.
(Top left) Network architecture: global vs. pairwise inputs and scaling up system size
N . (Top right) Presence vs. absence of reciprocal coupling. (Bottom left) Input statistics:
unimodal vs. bimodal marginal statistics. (Bottom right) varying strength of common input.
37
vsvsvsIcI1I2I3I13I12I23I1I2IN-1INI1I12I23IN-1,NIN,1I2IN1 0 0 1 0 ...0 1 0 1 0 ...0 1 0 0 1 ...IcIcI1I2I3vsIcI1I2I3 Figure 2: Geometrical organization of DKL(P, P ) within the space of three-cell spiking
distributions. Outer plots: Slices of DKL(P, P ) along surfaces fp = constant. Cen-
ter: schematic of the (fp, f1p, f1m) coordinate space. Counterclockwise from lower right:
fp = 0.1, 0.25, 0.58, 0.76, 0.9. fp = 0.25 contains the maximal attainable DKL(P, P ) over
all admissible P . fp = 0.616 contains the maximal attainable DKL from pairwise bimodal
inputs (see Results).
38
fpf1pf1mf1mf1pfp00.5100.20.40.60.81f1mf1pfp00.5100.20.40.60.81= 0.76f1mf1pfp00.5100.20.40.60.81= 0.58f1mf1pfp00.5100.20.40.60.81= 0.25f1mf1pfp00.5100.20.40.60.81= 0.1= 0.9 Figure 3: Examples of spiking distributions with small, intermediate, and large deviations
from PME models (as quantified by the KL- divergence, DKL(P, P ), between the observed
distribution P and its PME fit P ). (Top row) Bar plot contrasting three distributions with
their pairwise maximum entropy (PME) fits. The probability shown is the total probability
of all events with a certain number of spikes. From left: global skewed input, global binary
input, XOR operator. DKL(P, P ) is, from left, 0.0013 (skewed), 0.091 (bimodal), and 1
(XOR). (Middle row) The same distributions (crosses) projected into the (f1m, f1p)-plane
and their corresponding PME fits (circles). The cyan line is the iso-moment line of all
distributions with the same first and second moments, and the black curve is the PME
constraint surface. (Bottom row) DKL(P, P ) along the iso-moment line (cyan solid) and the
quadratic approximation derived in the Methods. (black dashed).
39
00.20.40.60.8100.20.40.60.8100.20.40.60.8100.20.40.60.81012300.20.40.60.81# spikes# spikes# spikesP(s)f1mf1mf1mf1pf1mf1mf1mbitsglobal skewed inputglobal bimodal inputXOR operationObservedPMEObservedPMEObservedPMED = 0.0013D = 0.091D = 1KLKLKL Figure 4: Deviation from PME fit for circuits receiving independent and (global) common
input. Results shown for N = 3; (A) Gaussian, (B) uniform, (C) skewed, and (D) bimodal.
For each choice of marginal input statistics, possible input parameters are varied over a
broad range as described in the Results; over c ∈ [0, 1], σ ∈ [0, 4], and Θ ∈ [−1, 3] (unimodal
inputs), over p ∈ [0, 1] and q ∈ [0, 1] (bimodal inputs). (Left column) Schematic of input dis-
tributions. (Center column) Projection of all distributions onto the (f1p, f1m)-plane. (Right
column) For the value of Θ for which the maximal value is achieved, a slice of DKL(P, P )
(unimodal inputs); contour plot of DKL(P, P ) where 1 < Θ < 2 (bimodal inputs).
40
qABCD Figure 5: Relationship between measures of higher-order interactions and other output firing
statistics. (A-C) DKL(P, P ) versus firing rate E[xj], for data obtained from N = 3 threshold-
ing cells. For each choice of marginal input statistics, possible input parameters were varied
over a broad range as described in the Results; over c ∈ [0, 1], σ ∈ [0, 4], and Θ ∈ [−1, 3] for
unimodal inputs, and over p ∈ [0, 1] and q ∈ [0, 1] for bimodal inputs. In each panel, data
is organized by ρ, from dark (ρ ∈ (0, 0.1)) to light (ρ ∈ (0.9, 1)); (A) Gaussian inputs, (B)
skewed inputs, (C) bimodal inputs. (D-F) The fraction of multi-information captured by the
PME model, ∆, versus firing rate E[x1], for these same distributions. Data is organized by
correlation coefficient ρ, as in panels (A-C); D) Gaussian, E) skewed, F) bimodal.
41
Figure 6: Effect of recurrent coupling on PME fits of circuit outputs for the thresholding
model. Left panel: scatter plot of DKL achieved with Gaussian inputs vs. coupling strength
g, for N = 3 cells. Right panel: another view of DKL(P, P ) vs. coupling strength g, for
N = 3 cells. In the top three panels, for unimodal inputs, c, σ, and Θ were fixed at a single
representative value, while g was varied. The top panel shows the results from circuits with
Gaussian inputs (global and pairwise inputs are equivalent) and the next two for skewed and
uniform inputs with either global (thin lines) or local (heavy lines) inputs. For all unimodal
inputs, (c, σ, Θ) were (0.5, 1, 1). Analogous results from circuits with either global (thin line)
or local (heavy line) bimodal inputs are shown in the bottom panel. Here, p, q, and Θ were
fixed at a single representative value, while g was varied. The parameters (p, q, Θ) were
(0.81, 0.5, 1.5) and (0.5, 0, 1.5) for global and pairwise inputs respectively.
42
00.511.522.530123456789x 10-3g (coupling strength)D (P, P)KL~A00.511.5200.0050.0100.010.020.0500.511.50.20.40000.511.5200.511.52g (coupling strength)D (P, P)~KLB 43
-1500-500050010002468101214-1000x 104-2000-1000010000.511.522.533.544.5x 104-2000-10000100002468101214x 104-3000-2000-100001000024681012141618x 104ABCD00.20.40.60.8ObservedPME-1-0.500.51050100150200−1001000Lexc050100150200−1000100Lexc0400800120016002000-1-0.500.51E−3000−2000−10000100000.511.52x 10500.20.40.60.8000100110111ObservedPME00010011011100010011011100.20.40.60.8ObservedPME00010011011100010011011100.20.40.60.8ObservedPME00.20.40.60.8ObservedPME−10400800120016002000-2-1.5-1-0.500.511.520.511.52502004006008001000-1-0.500.512.5x 10time (ms)% dev. from meantime (ms)-2-1.5-1-0.500.511.52012x 1002004006008001000-1-0.500.5143% dev. from meantime (ms)-2-1.5-1-0.500.511.525101502004006008001000-1-0.500.51x 104% dev. from meantime (ms)time (ms)% dev.% dev.% dev.countcountcount% dev./pA% dev./pA1000-20001000-2000pAcountpAcountpAcountpAcountpAcount Figure 7: Composite of results for RGC simulations with constant light and full field flicker.
In rows (A-C), we have (left) a histogram and time series of stimulus, (center) a histogram
of excitatory conductances (multiplied by driving voltage −60 mV, and therefore in units
of pA) and (right) the resulting distribution on spiking patterns. (A) Gaussian noise only.
(B) Gaussian input, standard deviation 1/6, refresh rate 8 ms. (C) Binary input, standard
deviation 1/3, refresh rate 8 ms. (D-E) Binary input, standard deviation 1/3, refresh rate
100 ms. In the top left panels, the excitatory filter Lexc(t) is shown instead of a stimulus
histogram; in the bottom left panels, the normalized excitatory conductance (in pA -- red
dashed line) is superimposed on the stimulus (blue solid). (D) Filter fit from data, parameters
given in Methods. Both the filter and conductance trace illustrate that the LN model that
processes light input acts as a (time-shifted) high pass filter. (E) As in (D), but with rectified
filter, producing a bimodal distribution of conductances.
(A) DKL(P, P ) for original vs.
simulations incorporating a rectified filter.
Figure 8: Comparison of RGC simulations computed with the original ON parasol filter,
vs.
rectified
filter. Data is organized by stimulus refresh rate (8, 40, and 100 ms) and marginal statistics
(Gaussian vs. binary). (B-C) Histograms of excitatory conductances (multiplied by driving
voltage −60 mV, and therefore in units of pA) for representative simulations, under original
(left) and rectified (right) filters. (B) Binary stimulus with standard deviation of 1/4 and
refresh rate 100 ms. The histogram shows a significant shift towards a bimodal structure,
with a corresponding increase in DKL(P, P ). (C) Binary stimulus with standard deviation of
1/2 and refresh rate 8 ms. The histogram shows little qualitative change under the change
in filter, and little change in DKL(P, P ).
44
00.0020.0040.0060.0080.010.01200.0020.0040.0060.0080.010.0128 ms40 ms100 msGaussianBinaryD (P, P), original filterKL~D (P, P), rectified filterKL~ABCcountpAcountpApApA 45
02004000246x 10−3A00.0500.08stixel size = 256 mµP(1,0,0), P(1,1,0)P(0,0,1), P(0,1,1)0.060.040.020.040.080P(0,1,0),P(1,0,1)B0.0500.040.08stixel size = 60 mµP(1,0,0), P(1,1,0)P(0,1,0),P(1,0,1)0.060.02000.040.08P(0,0,1), P(0,1,1)C321D−2002040600510x 105−2002040600510x 105−2002040600510x 105132E00.20.40.6p0p1p2p3ObservedPMEF321G−2002040600510x 105−2002040600510x 105−2002040600510x 105123H00.20.40.6p0p1p2p3ObservedPMEI Figure 9: Results for RGC simulations with light stimuli of varying spatial scale ("stixels").
With the exception of (A), we show data from the binary light distributions; results from
the Gaussian case are similar. (A) Average DKL(P, P ) as a function of stixel size. Values
were averaged over 5 stimulus positions, each with a different (random) stimulus rotation
and translation; 512 µm corresponds to full field stimuli. (B,C) Probability of singlet and
doublet spiking events, under stimulation by movies of 256 µm (B) and 60 µm (C) stixels.
Event probabilities are plotted in 3-space, with the x, y, and z axes identifying the singlet
(doublet) events 001 (011), 010 (101), and 100 (110) respectively. The black dashed line
indicates perfect cell-to-cell homogeneity (e.g. P (1, 0, 0) = P (0, 1, 0) = P (0, 0, 1)). Both
individual runs (dots) and averages over 20 runs (large circles) are shown, with averages
outlined in black (singlet) and gray (doublet). Different colors indicate different stimulus
positions. (D-F) Results for one stimulus position, with stixel size 60 µm. (D) Contour lines
of the three receptive fields (at 0.5, 1, 1.5 and 2 SD; and at the zero contour line) superim-
posed on the stimulus checkerboard (for illustration, pictured in an alternating black/white
pattern). (E) Marginal distributions of the excitatory conductances, for each cell. (F) Spike
pattern distribution; the three different probabilities labeled p1 correspond to, e.g, P (1, 0, 0),
P (0, 1, 0), and P (0, 0, 1)), demonstrating heterogenous responses among the RGCs. (G-I)
As in (D-F), but for stixel size 256 µm.
46
Figure 10: The impact of recurrent coupling on RGC networks with full-field visual stimuli.
The strength of gap junction connections was varied from a baseline level (relative magnitude
g = 1, or absolute magnitude ggap = 1.1 nS) to an order of magnitude larger (g = 16,
or ggap = 17.6 nS). (A) DKL(P, P ) obtained from two independent simulations without
coupling, for each of 42 different stimulus ensembles (deviation of this data from the line
y = x thus gives a control for the statistical variability of the results in later panels). (B-F)
DKL(P, P ) obtained with coupling, plotted versus the value obtained for the same stimulus
ensemble without coupling, for each of 42 different stimulus ensembles.
47
000.0050.010.0050.01ggap= 0 nSggap= 1.1 nSggap= 2.2 nSggap= 4.4 nSggap= 8.8 nSggapDKL, no coupling= 17.6 nS00.0050.010.0050.01000.0050.010.0050.01000.0050.010.0050.01000.0050.010.0050.01000.0050.010.0050.01DKL, with couplingDKL, with couplingDKL, no couplingDKL, no coupling8 msGaussianBinaryABCFED Figure 11: Maximal deviation from PME fit for thresholding circuit model, forced by either
global or local input against background noise, for increasing network size N . For each N ,
possible input parameters are ranged as described in the Results: over c ∈ [0, 1], σ ∈ [0, 4],
and Θ ∈ [−1, 3] (unimodal inputs), over s ∈ [0, 1] and p ∈ [0, 1] (bimodal inputs). The
sidebar shows sample distributions with maximal DKL(P, P ) for N = 16; from top, global
bimodal inputs, pairwise bimodal inputs, global Gaussian inputs, and pairwise Gaussian
inputs.
48
4681012141600.020.040.060.080.10.120.140.16pairwise, bimodalpairwise, uniformpairwise, skewedpairwise, gaussianglobal, bimodalglobal, uniformglobal, skewedglobal, gaussianbits per neuronN04812160.20.4004121600.20.48ObservedPME048121600.10.2048121600..13 Figure 12: Maximal deviation from PME fit for circuit forced by either global or local
input against background noise and either global or local recurrent coupling. Plot shows
the maximal DKL achieved over all parameters as N increases. For each N , possible input
parameters are ranged described as in the Results; over c ∈ [0, 1], σ ∈ [0, 4], and Θ ∈ [−1, 3]
(unimodal inputs), over s ∈ [0, 1], p ∈ [0, 1], and Θ ∈ [0.3, 1.9] (bimodal inputs). In addition,
recurrent coupling strength g is varied over [0, 2.4]. The sidebar shows sample distributions
with maximal DKL(P, P ) for N = 8; from top, global bimodal inputs with global coupling;
pairwise bimodal inputs with global coupling; global Gaussian inputs with pairwise coupling.
49
Nbits per neuronglobal input, global couplingglobal input, local couplinglocal input, global couplinglocal input, local couplingbimodaluniformskewedgaussian00.10.30.50.7246802468000.10.30.52468000.20.40.60.8ObservedPME Figure 13: The fraction of multi-information captured by the PME model, ∆, is low for
some network parameters as population size increases. Here, we show data obtained from
feedforward networks of N = 12 thresholding cells. For each choice of marginal input
statistics, possible input parameters were varied over correlation coefficient c ∈ [0, 1] and
input variance σ ∈ [0, 4]. The threshold was held at Θ = 1. (A-B) ∆ for (A) Gaussian and
(B) skewed input, for N = 12. (C) 1 − ∆ vs. N for Gaussian, skewed and uniform inputs at
a fixed value of c and σ (c = 0.56; for Gaussian and skewed, σ = 0.2; for uniform, σ = 0.6).
Because of the low firing rate, 1 − ∆ must grow linearly with N (see text).
50
Figure S1: Mean and estimated standard errors of DKL(P, P ), as a function of subgroup size.
The 20 simulations for each circuit condition (see Materials and Methods) were collected into
subgroups of M = 1, 2, 4, 5, 10, and 20. M = 20 corresponds to the full simulation length --
2× 106 ms in (B), 2× 105 ms in (C,D) -- reported in the text. As expected, bias decreases as
the length of subgroup increases and asymptotes at -- or before -- the full simulation length.
(A) N = 16, Gaussian pairwise inputs, for the sum-and-threshold model. (B) Full-field RGC
simulations. (C) Spatially variable RGC simulations, binary stimulus, stixel size = 4 µm.
Different colors signify different positions of stimulus relative to receptive field. (D) As in
(C), but stixel size = 256 µm.
51
0510152000.40.81.21.62x 10−4M (# simulations per subgroup)M (# simulations per subgroup)M (# simulations per subgroup)M (# simulations per subgroup)D (P, P)KL~D (P, P)KL~D (P, P)KL~D (P, P)KL~ABCD Figure S2: Cross-validated and empirical values of DKL(P, P ), vs. values reported in the
paper. (A) Full field, (B) Rectified, (C) Stixel simulations with binary inputs, (D) Same as
(C), zoomed into origin.
52
Cross−validated8 ms40 ms100 msGaussianBinaryACross−validated8 ms40 ms100 msGaussianBinaryB0246x 10−301234567x 10−3Cross−validatedEmpirical model4-32 x 10 m60 x 10 m128 x 10 m256 x 10 m-6-6-6-6CD4-32 x 10 m60 x 10 m128 x 10 m256 x 10 m-6-6-6-6 |
1611.05918 | 1 | 1611 | 2016-11-17T22:08:20 | Morphology of Fly Larval Class IV Dendrites Accords with a Random Branching and Contact Based Branch Deletion Model | [
"q-bio.NC"
] | Dendrites are branched neuronal processes that receive input signals from other neurons or the outside world [1]. To maintain connectivity as the organism grows, dendrites must also continue to grow. For example, the dendrites in the peripheral nervous system continue to grow and branch to maintain proper coverage of their receptor fields [2, 3, 4, 5]. One such neuron is the Drosophila melanogaster class IV dendritic arborization neuron [6]. The dendritic arbors of these neurons tile the larval surface [7], where they detect localized noxious stimuli, such as jabs from parasitic wasps [8]. In the present study, we used a novel measure, the hitting probability, to show that the class IV neuron forms a tight mesh that covers the larval surface. Furthermore, we found that the mesh size remains largely unchanged during the larval stages, despite a dramatic increase in overall size of the neuron and the larva. We also found that the class IV dendrites are dense (assayed with the fractal dimension) and uniform (assayed with the lacunarity) throughout the larval stages. To understand how the class IV neuron maintains its morphology during larval development, we constructed a mathematical model based on random branching and self-avoidance. We found that if the branching rate is uniform in space and time and that if all contacting branches are deleted, we can reproduce the branch length distribution, mesh size and density of the class IV dendrites throughout the larval stages. Thus, a simple set of statistical rules can generate and maintain a complex branching morphology during growth. | q-bio.NC | q-bio |
Morphology of Fly Larval Class IV Dendrites Accords with a
Random Branching and Contact Based Branch Deletion Model.
Sujoy Ganguly1, Olivier Trottier1, Xin Liang2, Hugo Bowne-Anderson1, and Jonathon
Howard1
1Department of Molecular Biophysics and Biochemistry, Yale University, New Haven
2Tsinghua-Peking Joint Center for Life Sciences, School of Life Sciences, Tsinghua
University, Beijing, China
April 16, 2018
Abstract
Dendrites are branched neuronal processes that receive input signals from other neurons or the
outside world [1]. To maintain connectivity as the organism grows, dendrites must also continue to
grow. For example, the dendrites in the peripheral nervous system continue to grow and branch
to maintain proper coverage of their receptor fields [2, 3, 4, 5]. One such neuron is the Drosophila
melanogaster class IV dendritic arborization neuron [6]. The dendritic arbors of these neurons tile
the larval surface [7], where they detect localized noxious stimuli, such as jabs from parasitic wasps
[8]. In the present study, we used a novel measure, the hitting probability, to show that the class IV
neuron forms a tight mesh that covers the larval surface. Furthermore, we found that the mesh size
remains largely unchanged during the larval stages, despite a dramatic increase in overall size of the
neuron and the larva. We also found that the class IV dendrites are dense (assayed with the fractal
dimension) and uniform (assayed with the lacunarity) throughout the larval stages. To understand
how the class IV neuron maintains its morphology during larval development, we constructed a
mathematical model based on random branching and self-avoidance. We found that if the branching
rate is uniform in space and time and that if all contacting branches are deleted, we can reproduce
the branch length distribution, mesh size and density of the class IV dendrites throughout the
larval stages. Thus, a simple set of statistical rules can generate and maintain a complex branching
morphology during growth.
In our brains, billions of neurons interact with each other to build a nervous system of unparalleled
complexity and computational power. Neurons have dendrites, which are branched structures that receive
synaptic or sensory inputs, and an axon, which send outputs to other neurons. The shape or morphology
of individual neurons sets the number and types of interactions that a neuron can have and provides the
structural basis of neuronal computation [9, 10, 11, 12, 13, 14].
Since many organisms continue to enlarge after the establishment of the body plan, it is critical
for axons and dendrites to maintain their morphology as they grow. For example, interneurons of the
grasshoper [2], motor neurons in moths [3] and mice [4] grow drastically in size yet maintain connections
to their target cells. Futhermore, dendrites in the perpherial nervous system, like those of gold fish
retinal ganglion cells [5], and dendritic arborization (da) sensory neurons of the fly larva [6], which are
the subject of this work, grow to continually maintain coverage of their receptor fields. In this paper,
we investigate the growth rules that are required to maintain the correct branching morphology as a
dendrite grows.
The da sensory neurons of the fly larva are a model system for studying dendritic arborization
[15, 16, 17]. These dendrites innervate the extracellular matrix, which lies between the outer cuticle
and the inner epidermal cell layer [7]. They tile the surface on the fly larva in a highly stereotyped
manner and have four distinct morphological classes [18] (Fig. 1 A). Since it is easy to identify and
image individual da neurons, these neurons have proven to be a powerful model system for studying
dendrite morphology [15, 16]. In this paper, we address the question of how the morphology of the class
IV da neurons (Fig. 1 B) is maintained during the larval stages.
The class IV da neuron has highly branched dendrites [18], which detect potentially harmful stimuli,
such as the ovipositor barb of parasitic wasps [8, 19]. The dendrites of the class IV neuron begin
1
morphogenesis during late embryogenesis ∼ 16 hrs After Egg Lay (AEL). By the time the larva hatches
(∼ 22 hrs AEL at 25 ◦C), the class IV dendrites nearly cover its surface. The dendrites then continue to
expand and branch as the larva grows (22 − 126 hrs AEL), so that the neuron maintains its coverage of
the larval surface [6].
In this work, we are seeking the growth rules that allow class IV dendrites to maintain their dense
coverage of the larval surface. To this end, we have used a novel measure, the hitting probability, that
quantifies the mesh size and two well-known measures of branching morphology: the fractal dimension [20]
and lacunarity [21, 22] (see Definition of Morphometrics). We show that these measures remain largely
invariant over larval stages, despite a several fold increase in larval length. Furthermore, we demonstrate
that a model with simple rules for branching and self-avoidance can capture essential features of the
establishment and maintenance of the dendrite's morphology.
1 Experimental Results
To characterize the morphology of fully-developed class IV dendrites, we imaged larvae expressing Cd4-
tdGFP under the ppk promoter (ppk-cd4-tdGFP ) during the third instar stage (Fig. 1) using a laser-
scanning confocal microscope (See Material and Methods for details). Using NeuronStudio [23] and Fiji
we traced the branches of the dendrites to produce skeletons. These skeletons were then analyzed to
obtain the the mesh size, density and uniformity of class IV dendrites using parameters defined in the
next section.
Definition of Morphometrics
Here we include simple definitions of the relevant morphometrics to aid comprehension.
Hitting Probabiltiy H(B): The probability that a box of size B hits the dendrite.
Mesh Size BH: The length at which 50% of all boxes hit the neuron.
Fractal Dimension df: A measure of the space-fillingness of a shape. For a completely filled box df = 2,
for a straight line df = 1, for branched shapes 1 ≤ df ≤ 2.
Lacunarity Λ(B): A measure of density fluctuations as a function of length scale B.
Lacunarity Length BΛ: The length at which Λ(B = BΛ) = 0.25, i.e. the length at which the neuron
is uniform. The larger BΛ, the more variable the density of the neuron.
Radius of Gyration Rg: A length scale that measures how spread out a shape is from its center. The
larger Rg the more spread out the neuron.
Persistence Length β: The characteristic length at which a branch bends.
For mathematical definitions see Appendix.
1.1 Class IV dendrites have a small mesh size
To characterize the mesh size of the dendrites, we developed a novel measure called the hitting probability
H(B). H(B) measures the probability H that a randomly placed box of size B hits the dendrite (see
Appendix for details). The hitting probability generalizes an earlier metric called the coverage index [6]
by allowing for any box location and any box size. A typical hitting probability curve of a neuron (Fig.
1 D) H(B) increases monotonically with B, eventually reaching H = 1 as B approaches the size of the
neuron.
We define the characteristic mesh size BH as the box size at which half of all boxes hit the dendrite. In
other words, BH is the maximum size of a stimulus that would go untouched, or undetected, on average,
by the neuron. BH is similar to the mesh size in a cross-linked polymer network[24].
We found that BH = 8.4 ± 0.5 µm (mean ± SD, n = 14 neurons) for the mature dendrites of the class
IV neuron. Thus, the mesh size is approximately equal to the diameter of the ovipositor barb of wasps
that lay eggs in Drosophila larva (∼ 10 µm, [8]). This indicates that the class IV dendrite has a high
chance of detecting a wasp attack. Furthermore, the mesh size is small compared to the overall size of
the neuron (∼ 500 µm) and is similar to the mean branch length (see below).
2
Figure 1: Morphometrics of class IV and class III dendrites. (A) 3rd instar larval expressing
GFP-tagged membrane protein (ppk-cd4-tdGFP ) in class IV neurons. (B) The skeleton of a class IV
neuron from a third instar larva with an example of a box used to calculate the hitting probability
(blue) and a circle used to calculate the correlation dimension dc (magenta). (C) The skeleton of a
class III neuron from a third instar larva with an example of a set of boxes used to compute the box
dimension db and lacunarity function Λ. See Definition of Morphometrics and Appendix for
details (D) Example hitting probability H versus box size B. (E) The correlation function κ versus
the radius R. Fits used to determine the correlation dimension dc plotted in solid lines. See Appendix
for details of fits. The curves for dc = 1 and dc = 2 are plotted in blue for reference. (F) The number
of boxes needed to cover a neuron versus the box size B. Fits used to determine −db are plotted in
solid lines. See Appendix for details of fits. The curves for db = 1 and db = 2 are plotted in blue for
reference. (G) Lacunarity function Λ versus box size B for a class IV neuron (black) and a class III
neuron (red).
1.2 Class IV dendrites are dense and uniform
To understand the morphological basis underlying BH we quantified the density and uniformity of the
class IV dendrites during the third instar stage. The fractal dimension df is a commonly used measure
of how a branched structure fills space. A solid square, for example, has a df = 2, while a straight line
has a df = 1 (See Appendix for mathematical definitions of df). We found that the fractal dimension of
class IV dendrites was df = 1.80± 0.04 (mean± SD, n = 14), indicating that the dendrites are dense and
space filling.
To establish a small mesh, a dendrite needs to not only be space filling (i.e. df ∼ 2), but uniformly so.
To measure the uniformity of the dendritic arbor, we used the lacunarity function Λ(B). This measures
the density variation as a of function length scale B. To compare the lacunarity of different cells we
calculated the length BΛ (Λ(B = BΛ) = 0.25 see Definition of Morphometrics). The smaller BΛ, the
more uniform the dendritic density. We found that BΛ = 32.6 ± 16.8 µm (mean ± SD, n = 14) for third
instar larva. In other words, on spatial scales larger than 33 µm the density of the neuron is uniform,
whereas below 33 µm it is variable. While the BΛ is larger than the mesh size BH, it is much smaller than
the dendrite size, indicating that the coverage is uniform and arbors can be considered homogeneous.
In summary (Tab. 1), mature class IV neurons have dense (large df) and uniform (small BΛ) dendritic
arbors, which is consistent with a small mesh size (small BH).
3
BBBCDEFG010020000.250.50.751HittingProbability(H)BoxSize(B)[µm]100 102 104 NumberofBoxesBoxSize(B)[µm]010020000.20.4Lacunarity(Λ)BoxSize(B)[µm]500 µm30 µm30 µmA10.110100100010.1101001000100 103 10CorrelationFunction(κ)Radius(R)[µm]61Class IVClass IIIClass IVClass IIIClass IVClass IIIClass IVClass III 1.3 Class IV dendrites have a tighter mesh, are denser and more uniform
than class III dendrites
To assess the ability of these measures to quantify dendritic morphology, we also imaged class III cells
(Fig. 1 C). Class III cells are gentle touch sensors and use a different set of mechanotransduction channels
[25]. The class III dendrites (Fig. 1 C) are substantially less branched and have smaller branches, on
average, than the class IV dendrites (Tab 1). We find that H(B) is much smaller for class III neurons
than for class IV neurons (Fig. 1 D) for most box sizes. Consequently, we find that the mesh size BH is
much larger in class III neurons (BH ∼ 24.7 µm) than in class IV neurons (BH ∼ 8.4 µm) (Tab. 1). In
other words, class III dendrites have larger gaps in coverage than class IV dendrites.
Table 1: Properties of class IV and class III neurons. All numbers are mean ± SD.
Mean branch length, (cid:104)L(cid:105)
Mesh size, BH
Fractal dimension (correlation method), dc
Fractal dimension (box method), db
Lacunarity length, BΛ
class IV (n = 14)
12.54 ± 0.04 µm
8.37 ± 1.80 µm
1.80 ± 0.04
1.79 ± 0.04
32.6 ± 16.8 µm
class III (n = 8)
7.49 ± 0.8 µm
22.7 ± 4.8 µm
1.42 ± 0.03
1.43 ± 0.04
130.5 ± 31.2 µm
We find that class III dendrites have a fractal dimension of df ∼ 1.42 (Tab 1), which is substantially
smaller than the class IV neuron. In other words, class III dendrites are sparser, at all scales, than class
IV dendrites (Fig. 1 E and F). We find that the lacunarity Λ decays much slower (with B) for class
III dendrites compared to class IV dendrites (Fig. 1 G). Consequently BΛ is much larger (Tab. 1) for
class III dendrites than for class IV neurons (Tab. 1) showing that they are less uniform than class IV
neurons. These results (Tab. 1), show that class III neurons are sparser (small df) and less uniform
(large BΛ) than class IV dendrites, which is consistent with having larger mesh size (large BH). These
differences likely reflect the different mechanoreceptor functions of class III and class IV neurons (see
Discussion).
1.4 Class IV dendrites maintain a tight mesh, high density and uniformity
throughout larval stages
To determine how the morphology of class IV dendrites changes with time, we imaged and skeletonized the
class IV dendrites, as before, from early first instar (30 hrs AEL) to wandering third instar (126 hrs AEL)
(Fig. 2 A). We used the larval body segment length as a proxy for time since each data point comes
from a unique larva, and S increases linearly with time [26] (SI).
The mesh size BH increased modestly from 4.4 ± 0.3 µm (1st instar) to 8.5 ± 1.4 µm (third instar)
(Fig. 2 C). This two-fold increase is less than the six-fold growth in larval body segment size and
indicates that the mesh size remains small during development. Interestingly, the ratio between BH and
the mean branch length (cid:104)L(cid:105) is nearly constant during development showing that the shape of the arbor
has remarkable conservation during development
The morphologies of the class IV neurons remained dense and uniform during the larval stages of
development. We found that the fractal dimension (Fig. 2 D) in the just hatched larvae (∼ 30hrs AEL)
was df ∼ 1.7 and increased to 1.75 within 24 hours, eventually rising to about 1.8 during the next four
days. The lacunarity slightly increased with time (Fig. 2 E). By rescaling B by the radius of gyration
Rg (i.e. overall neuronal size see Eq. 1), we found that Λ follows the same curve at all developmental
times, i.e., collapses when scaled by Rg (Fig. 2 D inset). The nearly constant fractal dimension and
the collapse of the lacunarity curves indicate that the morphological pattern of the class IV neurons is
mostly established during embryogenesis.
2 Dynamic Model of Class IV Development
The maintenance of a small mesh size throughout larval stages raises the question: how can the class IV
dendrites achieve this, despite the six-fold increase in the larval segment size? To answer this question,
we developed a mathematical model of class IV dendrite morphogenesis during the larval stages. Our
model consists of three rules that determine 1) branch creation, 2) the direction and speed that this new
branch grows, and 3) how branches avoid crossing the other branches (self-avoidance).
4
Figure 2: Morphometrics of class IV neurons during larval growth. (A)Examples of class IV
neurons during larval stages. All scale bars 30 µm, all time stamps hours After Egg Lay. (B) The
hitting probability H is plotted versus the box size B for the 5 neurons shown in Fig. 2 A. In the inset,
we plot H versus B/(cid:104)L(cid:105). We define the mesh size BH such that H(B = BH) = 0.5. BH ∼ 0.72(cid:104)L(cid:105)
throughout the larval stages. (C) BH is plotted versus larva body segment length S (red 28 cells). For
−1 and α = 102 µm (blue). In the inset, we have plotted the mean
simulated neurons, ω = 0.2 min
branch length of the class IV dendrites versus the larva body segment length S. (D) The fractal
dimension df is plotted against S. We have binned the data by body segment length S with bin widths
of 77 µm. Simulation parameters are same as before. (E) The lacunarity Λ is plotted against box size
B for the five neurons in Fig. 2. In the inset, we plot Λ versus B/Rg. We define the lacunarity length
BΛ as the box size at which Λ(B = BΛ) = 0.25.
2.1 Branch Creation
We found that the number of branch points increases with time during larval stages (Fig. 3A), and this in-
crease was well described by a linear function with a net branch creation rate ωnet ∼ 0.1 branch points min
(Tab. 2). This observation led us to model branch creation as a time invariant process with a branching
frequency ω. Since ωnet can include the removal of branches (see below), it is a lower bound on the
branching creation frequency ω.
−1
We also measured the branch length distribution, and found that it was well described be an expo-
nential distribution (Fig. 3 B). Furthermore, we found that that the branch length was independent of
the branch depth, defined as the number of branch points between a branch and the soma along the
shortest path (Fig. 3 C). Motivated by these observations, we modeled branch creation as a random
process that was uniform along the neuron and constant in time.
2.2 Tip Elongation
Since the growth of class IV dendrites occurs at the branch tips [27, 6], we modeled neuron growth as
branch tip elongation. We measured the overall size of class IV dendrites using the radius of gyration Rg
(Eq. 1). Rg measures spread of neuron mass from its center. We found that Rg increases linearly during
development (Fig. 3 D). Therefore, we assume that the tip elongation rate v is constant in time and
v ∝ Vg, where Vg is the growth speed of the class IV dendrite. The assumption of a constant growth speed
assumes that the simulation time scale is much larger than the fluctuation times scale. We estimated Vg
−1 (Tab 2).
from the change in dendrite size during development (Fig. 3 D), and found Vg ∼ 0.04 µm min
To determine the direction of branch growth, we measured how much the path of a branch changes
as a function of path length by using the persistence length β; the distance over which the orientation
5
020040000.20.4Lacunarity(Λ)BoxSize(B)[µm]012B/Rg00.20.4ΛBΛ/Rg30 hr56 hr80 hr96 hr126 hr02004006001.61.71.8FractalDimension(df)SegmentLength(S)[µm]EC05010000.250.50.751HittingProbability(H)BoxSize(B)[µm]0123B/(cid:31)L(cid:30)00.250.50.751HBH/(cid:31)L(cid:30)30 hr56 hr80 hr90 hr126 hrBA0200400600051015SegmentLength(S)[µm]BH[µm]0200400600051015(cid:31)L(cid:30) [µm] S [µm]30 hrs56 hrs80 hrs96 hrs126 hrsDdbdcSimulation30 µm Figure 3: Branching Rules (A) Number of branches versus larval body segment length S for class IV
dendrites at different larval stages. The slope of the fit is 1.64 branches µm−1 for the line passing
through the S axis at 50 µm, the segment length at the onset of dendritogenesis. (B) The probability
density P (L) that a branch of a dendrite (at a particular time) has length L. The probability density is
rescaled by multiplying by (cid:104)L(cid:105) and the branch length is rescaled by dividing by (cid:104)L(cid:105) (of the particular
dendrite). The superposition of the data indicates that the branch lengths are well described be an
exponential distribution (red line) at all larval stages. (C) The branch length autocorrelation Cl (Eq.
3) vs depth difference ∆d averaged over all branches. The near-zero values for ∆d ≥ 1 imply that
branch length is independent of depth. (D) The radius of gyration Rg (Eq. 1) of 28 class IV neurons
versus larval body segment length S. The slope of the fit is 0.39, where the S−intercept is set such
that Rg = 0 at the onset of class IV dendrite morphogenesis. (E) The tangent vector autocorrelation
function Ct (Eq. 2), averaged over all branches, versus the path length lag ∆s (Eq. 2). The red line is
an exponential fit to the data Ct(∆s) = e∆s/β, where β = 44.8 ± 1.5 µm is the persistence length.
of the direction of growth of a branch changes (see Appendix for mathematical details). We find that
β = 45 ± 2 µm (Tab. 2, Fig. 3 E), which is much greater than the mean branch length ((cid:104)L(cid:105) ∼ 12.5 Tab.
1), indicating that branches tend to be straight.
2.3 Self-Avoidance
Previous work has shown that the branches of the class IV neuron do not cross each other. Furthermore,
it is known that this self-avoidance is contact mediated and it has been proposed that branches retract
after contact [27, 28, 29, 30]. Therefore, we modeled self-avoidance by having growing branches retract
at a constant speed v (same as elongation rate), if they contact an existing branch. For simplicity, the
retraction length was assumed to be exponentially distributed with a mean retraction length α.
Table 2: Parameters in the model. All errors are SD (n=28 neurons).
Parameter
Branching frequency
Tip elongation rate, v
Persistence length, β
Retraction length, α
0.12 ± 0.03 min
0.04 ± 0.02 µm · min
−1 (ωnet)
−1
Measured Value
44.8 ± 1.5 µm
not measured
Simulation Value
−1(ω)
0.01 − 2 min
−1
0.08 µm · min
70 µm
0 − 1000 µm
2.4 Model Implementation
Using these rules, we arrived at a four parameter model; ω the branch creation rate, v the growth speed,
β persistence length and α to retraction scale. We implemented our model on a hexagonal lattice, with
lattice spacing = 0.4 µm. We chose to use a lattice model to facilitate contact detection as the lattice
spacing was set to be equal to the thickness of the branches. Since we implemented the rules on a
hexagonal lattice, the possible branching angles are limited. Therefore we ignored the role of branching
angles on morphology. The lattice spacing and the growth speed act as scale factors, i.e. setting the
overall size of the dendrite but not changing the shape. Therefore, we chose v such that the size of
a simulated dendrite matched that of a real one, which left us with three free parameters. However,
we found that as long as β is large, i.e., as long as the branches are straight, β has little effect on the
morphology (SI). Therefore, we focused on how branching frequency ω and the contact-based retraction
scale α control the morphology of dendrites.
6
03006000150300Segment Length (S) [µm]R g [µm]024600.51Depth (∆d)05101500.51Path Length (∆s) [µm]Ct(∆s)Cl(∆d)0240.010.11L/(cid:31)L(cid:30)P(L)(cid:31)L(cid:30)030060005001000Number of Branch PointsSegment Length (S) [µm]ABCDEDataFitDataFitDataFitDataFit 2.5 Model Results
Figure 4: Simulations of dendritic growth (A) Nine representative simulations of dendrties for a
range of branching frequencies ω and retraction scales α. For all simulations, the persistence length was
−1 (dashed box) we
β = 70 µm and simulation time was T = 100 hrs. For α (cid:29) 1 µm and ω ∼ 0.2min
have subjective agreement with the morphology of real neurons. (B) The branch length distribution
for simulated neurons in the dashed black box in (A) is approximately exponentially distributed as
observed. (C) The mesh size BH in a contour plot versus the branching frequency ω and retraction
scale α. Note that the retraction scale α is plotted in a log scale. In pink we have highlighted the
observed value of BH = 7 µm (see Tab. 1). The white box indicates the values used for the time series
−1 and α = 102 µm). (D) The fractal dimension df in a contour
plotted in Fig. 2 C and D (ω = 0.2 min
plot versus the branching frequency ω and the retraction scale α. In pink we have highlighted the
physiological value of df = 1.8 (see Tab. 1) (E) The lacunarity scale BΛ in a contour plot versus the
branching frequency ω and the retraction scale α. We highlighted in pink the observed value
BΛ = 33 µm (see Tab. 1). We find that this value is not found in the range of values used in our
simulations.
We generated simulated neurons for a range of branching frequencies ω and retraction scales α (Fig.
4 A) and analyzed their shape.
Importantly, we found that the distribution of branch lengths was
exponential (Fig. 4 B) and the branch lengths were uncorrelated with branch depth, for a limited range
of parameters (dashed box in Fig. 4 A), which agrees with the experimental observations (Fig. 3 B and
C).
To test whether our model provides a good description of the morphology, we measured the mesh size,
fractal dimension and lacunarity of the simulated dendrites. The mesh size BH decreased with increasing
ω and decreasing α (Fig. 4 C) and appeared to saturate as ω increases. There was a small region of ω− α
space where there was quantitative agreement between the simulated and observed values of BH (Fig. 4
C, pink). We found, that df increased monotonically with ω and saturated for large values of ω. As for
BH, we found a narrow band of ω − α space where we have a quantitative agreement between the model
and the class IV neuron for df (Fig. 4 D, pink). Crucially, there was a small region of ω − α space where
the model recapitulated both BH and df (Fig. 4 C and D, white box). Taking values from this small
−1 and α = 102 µm), we recapitulated the third instar and even found agreement
region (ω = 0.2 min
throughout most of the larval development (Fig. 2 B and 2 C). Thus, the model recapitulates both the
mesh size and fractal dimension throughout development. It did not recapitulate the lacunarity, which
−1 and α = 102 µm was larger than the observed values. In other
for the parameter values ω = 0.2 min
words, the model arbors are not as uniform as the observed ones (see Discussion). In summary, this
model recapitulates most, but not all aspects of the dendritic morphology of class IV neurons.
7
Branching Frequency (ω) [min-1]10240.010.11P(L)(cid:31)L(cid:30)L/(cid:31)L(cid:30)Retraction Scale (α) [µm]00.51.52110010110210360120180240BΛ [µm] Branching Frequency (ω) [min-1]Retraction Scale (α) [µm]00.51.5211001011021031.71.8Fractal Dimension (dc)Branching Frequency (ω) [min-1]Retraction Scale (α) [µm]00.51.52110010110210336912BH [µm]Branching Frequency (ω) [min-1]Retraction Scale (α) [µm]0.020.22101001000CBDEA 3 Discussion
Our key experimental finding is that the morphology of class IV neurons, as characterized by the branch
length, mesh size, fractal dimension, and lacunarity, remarkably remains constant during development.
Indeed, from the early first instar larva (30 hours after egg lay) to the late third instar larva (126 hours
after egg lay), as both the segment size and the number of branches increases approximately six-fold, the
mean length, the mesh size, and the lacunarity only increase around two-fold and the fractal dimension
is almost unchanged. When we normalize the mesh size by the mean branch length, it is virtually
unchanged throughout development. Thus, as these cells grow, key aspects of their morphology are
invariant, even as the overall size the number of branches increases by nearly an order of magnitude.
As a first step toward probing the mechanism underlying this geometric invariance, we developed a
simple computational model for branched morphogenesis. The model assumes that the branching rate is
constant over development (consistent with the observed linear increase in the number of branches), that
the rate was independent of position and that growing tips retract random, exponentially distributed
distance after contacting other branches. The model reproduced many of the key features of the growth
of class IV cells including branch lengths, mesh sizes, and fractal dimensions. However, it was unable to
capture the lacunarity (the model predicted a higher relative density in the center than was observed,
see SI). Importantly, the data constrained the values of the branching frequency and mean retraction
distance. Thus, the model provides a framework for understanding the changes in the morphology of
these cells during development.
One of our most striking experimental and theoretical findings was that the branching frequency
in class IV dendrites was independent of total dendrite length. Naively, we might have expected that
the mean number of branches added per unit time would increase with total dendrite length, as the
longer the dendrites, the more positions on which branches could form. However, this would have led
to an exponential increase in the number of branches, rather than the observed linear increase. Our
modeling shows that even if the retraction length is much larger than the mean branch length, an
exponential increase in branch length is still observed, and the distribution of branch lengths deviates
from the observed exponential distribution (see SI). The constant branching rate suggests that branching
is limited by the production of a nucleating factor that is produced at a constant rate. Furthermore,
our finding that branching is uniform in space implies that the putative nucleation factor would be
distributed widely and uniformly throughout the cell.
We also found that for our model to recapitulate class IV-like morphologies, contact-based retraction
needs to lead to complete branch deletion, i.e., the mean retraction distance (α) is much larger than the
mean branch length ((cid:104)L(cid:105)). By deleting the branches whose tips collide with other branches, gaps of a
size similar to the mean branch length are created and maintained. Also, dense regions where the gap
size is less than or equal to the mean branch length will not increase in density. Such overfilling is seen in
the simulations for small retraction lengths (Fig. 4 A). Thus, our model constrains both the branching
frequency and the retraction distance.
Finally, we note that the mesh size of class IV dendrites, 4 − 8 µm, is well suited for detecting highly
localized nociceptive stimuli such as punctures by the 10 µm diameter ovipositor barb of parasitic wasps
([8]). This acuity is maintained throughout development. Thus, the small mesh size is consistent with
the class IV neuron being a harsh touch sensor. Indeed, the theory of contact dynamics predicts that the
indentation h of the surface of an elastic body poked by a probe with a cross-sectional radius R (pushing
normal to the surface) is h ∝ F/R, where F is the applied force [31]. Therefore, the smaller R, the larger
h (for a fixed force); the more local the stimuli, the more sensitive to localized forces. Thus the small
mesh size suggests that the class IV neuron is well adapted to sensing harsh touch throughout larval
stages. In contrast, the class III neuron, a soft touch sensor, would need to capture diffuse stimuli, i.e.,
the mesh size can be large. Thus, the morphologies of both class IV and class III neurons are well-suited
for their mechanosensitive functions.
4 Materials and Methods
Drosophila Stocks
All flies were maintained on standard medium at 23oC. The strain ppk-cd4-tdGFP was a kind gift from
Dr. Han Chun (Cornell University).
8
Imaging and Skeletonization
The larvae were mounted in 50% glycerol in PBS between a glass slide and a cover slip. The sam-
ple was imaged using a confocal laser scanning microscope (Zeiss, LSM780) with 63x objective. The
600x600 µm images were stitched together offline using Fiji and the stitched images were processed using
the NeuronStudio [23] to obtain the skeleton a one pixel wide tracings of the dendritic arbors.
A Mathematical Definitions of Morphometrics
Radius of gyration of the Neuron
The radius of gyration is defined as
Rg =
(cid:118)(cid:117)(cid:117)(cid:116) 1
M
M(cid:88)
j=1
(rj − rm)2,
(1)
where M is the total number of occupied pixels, rj is the position of the jth occupied pixel and rm is
the mean position of all occupied pixels. Rg measures the standard deviation of the dendrite pixels, i.e.,
the spread of the imaged neuron from its center.
Path Correlation and Persistence Length
The deviation of a branch path from a straight line can be quantified using the tangent vector autocor-
relation function
Ct(∆s) ∼ (cid:104)t(s) · t(s + ∆s)(cid:105)s,
(2)
where t(s) is the tangent vector as a function of the path length s. Ct measures the angular change of
the t as a function of path length, i.e., how bent the branch is. If Ct = 1, the path is straight and if
Ct = 0, there is a 90o turn in the path.
Branch Length Correlation Function
The branch length autocorrelation function is
Cl(∆d) = (cid:104)L(d)L(d + ∆d)(cid:105)d − (cid:104)L(d)(cid:105)2
d
(cid:104)L2(d)(cid:105)d − (cid:104)L(d)(cid:105)2
d
,
(3)
where L(d) is the branch length at depth d and ∆d is the depth difference. Depth is defined as the
number of branch points between the branch and the soma, along the shortest path from the branch to
the soma. (cid:104). . .(cid:105)d represents the average over d.
Hitting Probability
Consider a box with side length R centered anywhere in the receptor field of the neuron (Fig. 1 C). We
then ask: 'what is the probability PH (b, R) of having b pixels in a box of size R?'. Using this probability,
we can determine the probability that a box of size R contains at least n pixels
(cid:90) M
Hn(R) =
PH (b, R)db,
n
(4)
where M is the total number of neuron pixels. We define the mesh size BH such that H1(R = BH) = 0.5,
i.e., the mesh size is the box width such that there is a 50% chance that the box contains at least one
pixel from the skeleton.
9
Fractal Dimension
In this paper, the fractal dimension is measured using two different methods: the correlation dimension
(Fig. 1 E) and box counting (Fig. 1 F) method. In the box counting method, we determine the number
of boxes N of side length R that are needed to cover the neuron (Fig. 1 B). The number of boxes needed
to cover a line of length l is N = l
R ; therefore N ∝ R−1. The number of boxes needed to cover a square
; therefore N ∝ R−2. In general, N ∝ R−db, where db is the box counting
of side length l is N =(cid:0) l
measure of the fractal dimension.
(cid:1)2
R
In the correlation method [32], we determine how many pixels are contained within a circle of radius
R. Let each point xi on the neuron be the center of a circle of radius R (Fig. 1 B). Then N (xi, R) is
the number of skeleton pixels in the circle (green pixels in Fig. 1 C). Averaging over all possible centers
(i.e. skeleton pixels) xi, gives
κ(R) = (cid:104)N (xi, R)(cid:105)xi.
(5)
In general, κ(R) ∝ Rdc where dc is the correlation measure of the fractal dimension.
The relation f (R) ∝ Rd is called a scaling law and is only valid in a finite range of R (e.g. for small
R we approach the scale of one pixel, and for large R we approach the total dimension of the neuron).
For the neurons the minimum scale is half the mean branch length and the maximum scale is the radius
of gyration.
Lacunarity
Consider the set of boxes of linear dimension R used in the box counting method (see figure 1 B). Instead
of asking how many boxes are needed to cover the shape, we ask 'what is the probability PB(b, r) of
having b pixels in a box of size R?'. PB(b, r) differs from PH (b, R) since it only considers boxes that have
at least one pixel (b ≥ 1).
The moments of PB are defined as
(cid:90) M
µn(R) =
1
bnPB(b, R)db,
where M is the total number of skeleton pixels. This then allows us to look at the coefficient of variation
of PB(b, R)
CV (R) =
σ2
µ2
1
=
µ2 − µ2
1
µ2
1
,
(6)
where σ2 is the variance of PB(b, R). CV (R) is also called the lacunarity function. The more uniform a
shape is, the smaller CV and the more variable the shape, the larger the CV . For example, a uniform
shape would have a CV ∼ 0. How large CV needs to be for a shape/neuron to be consider variable is
somewhat arbitrary. The more important point is that the larger CV , the larger the variation in the
neuron. It also allows us to see how these variations change with length scale. Thus, the lacunarity
function measures the variations and assigns them a typical length scale.
Acknowledgments
This work was partially supported by the NIH Pioneer Award (Award Number), National Natural Science
Foundation of China (NSFC Grant 31671389, to X.L.) and Max-Planck Partner Program (to X.L.). SG
was supported by an EMBO Long-Term Fellowship, and OT is supported by the Fonds de Reshershe du
Qubec - Nature et technologies. Thanks to Dr. Han Chun (Cornell University) for the fly strains.
References
[1] Greg Stuart, Nelson Spruston, and Michael Hausser, editors. Dendrites. Oxford University Press,
2nd edition, 2007.
[2] D Bentley and A Toroian-Raymond. Embryonic and postembryonic morphogenesis of a grasshopper
interneuron. J Comp Neurol, 201(4):507–18, Oct 1981.
10
[3] J W Truman and S E Reiss. Hormonal regulation of the shape of identified motoneurons in the
moth manduca sexta. J Neurosci, 8(3):765–75, Mar 1988.
[4] Yan Li, Diana Brewer, Robert E Burke, and Giorgio A Ascoli. Developmental changes in spinal
motoneuron dendrites in neonatal mice. J Comp Neurol, 483(3):304–17, Mar 2005.
[5] P F Hitchcock. Constant dendritic coverage by ganglion cells with growth of the goldfish's retina.
Vision Res, 27(1):17–22, 1987.
[6] Jay Z Parrish, Peizhang Xu, Charles C Kim, Lily Yeh Jan, and Yuh Nung Jan. The microrna
bantam functions in epithelial cells to regulate scaling growth of dendrite arbors in drosophila
sensory neurons. Neuron, 63(6):788–802, Sep 2009.
[7] Rolf Bodmer and Yuh Nung Jan. Morphological differentiation of the embryonic peripheral neurons
in drosophila. Roux's archives of developmental biology, 196(2):69–77, 1987.
[8] Richard Y Hwang, Lixian Zhong, Yifan Xu, Trevor Johnson, Feng Zhang, Karl Deisseroth, and
W Daniel Tracey. Nociceptive neurons protect drosophila larvae from parasitoid wasps. Curr Biol,
17(24):2105–16, Dec 2007.
[9] H Wassle and B B Boycott. Functional architecture of the mammalian retina. Physiol Rev,
71(2):447–80, Apr 1991.
[10] M Hausser, N Spruston, and G J Stuart. Diversity and dynamics of dendritic signaling. Science,
290(5492):739–44, Oct 2000.
[11] D Attwell and S B Laughlin. An energy budget for signaling in the grey matter of the brain. J
Cereb Blood Flow Metab, 21(10):1133–45, Oct 2001.
[12] Gordon M G Shepherd, Armen Stepanyants, Ingrid Bureau, Dmitri Chklovskii, and Karel Svoboda.
Geometric and functional organization of cortical circuits. Nat Neurosci, 8(6):782–90, Jun 2005.
[13] Nelson Spruston. Pyramidal neurons: dendritic structure and synaptic integration. Nat Rev Neu-
rosci, 9(3):206–21, Mar 2008.
[14] Quan Wen and Dmitri B Chklovskii. A cost-benefit analysis of neuronal morphology. J Neurophysiol,
99(5):2320–8, May 2008.
[15] Yuh-Nung Jan and Lily Yeh Jan. Branching out: mechanisms of dendritic arborization. Nat Rev
Neurosci, 11(5):316–28, May 2010.
[16] S Lawrence Zipursky and Wesley B Grueber. The molecular basis of self-avoidance. Annu Rev
Neurosci, 36:547–68, Jul 2013.
[17] Xintong Dong, Kang Shen, and Hannes E Bulow. Intrinsic and extrinsic mechanisms of dendritic
morphogenesis. Annu Rev Physiol, 77:271–300, Feb 2015.
[18] Wesley B Grueber, Lily Y Jan, and Yuh Nung Jan. Tiling of the drosophila epidermis by multiden-
dritic sensory neurons. Development, 129(12):2867–78, Jun 2002.
[19] Jessica L Robertson, Asako Tsubouchi, and W Daniel Tracey. Larval defense against attack from
parasitoid wasps requires nociceptive neurons. PLoS One, 8(10):e78704, 2013.
[20] Benoit Mandelbrot. How long is the coast of britain? statistical self-similarity and fractional
dimension. Science, 156(3775):636–638, 1967.
[21] Benoit B. Mandelbrot. The fractal geometry of nature. W.H. Freeman, 1 edition, August 1982.
[22] C. Allain and M. Cloitre. Characterizing the lacunarity of random and deterministic fractal sets.
Phys. Rev. A, 44:3552–3558, Sep 1991.
[23] S L Wearne, A Rodriguez, D B Ehlenberger, A B Rocher, S C Henderson, and P R Hof. New
techniques for imaging, digitization and analysis of three-dimensional neural morphology on multiple
scales. Neuroscience, 136(3):661–80, 2005.
11
[24] Masao Doi and Sam F Edwards. The theory of polymer dynamics, volume 73. oxford university
press, 1988.
[25] Zhiqiang Yan, Wei Zhang, Ye He, David Gorczyca, Yang Xiang, Li E Cheng, Shan Meltzer, Lily Yeh
Jan, and Yuh Nung Jan. Drosophila nompc is a mechanotransduction channel subunit for gentle-
touch sensation. Nature, 493(7431):221–5, Jan 2013.
[26] Michael Ashburner, Kent G Hawley Golic, et al. Drosophila: a laboratory handbook. Technical
report, 2005.
[27] Benjamin J Matthews, Michelle E Kim, John J Flanagan, Daisuke Hattori, James C Clemens,
S Lawrence Zipursky, and Wesley B Grueber. Dendrite self-avoidance is controlled by dscam. Cell,
129(3):593–604, May 2007.
[28] Peter Soba, Sijun Zhu, Kazuo Emoto, Susan Younger, Shun-Jen Yang, Hung-Hsiang Yu, Tzumin
Lee, Lily Yeh Jan, and Yuh-Nung Jan. Drosophila sensory neurons require dscam for dendritic
self-avoidance and proper dendritic field organization. Neuron, 54(3):403–16, May 2007.
[29] Katie M Hutchinson, Fernando Vonhoff, and Carsten Duch. Dscam1 is required for normal den-
drite growth and branching but not for dendritic spacing in drosophila motoneurons. J Neurosci,
34(5):1924–31, Jan 2014.
[30] Benjamin J Matthews and Wesley B Grueber. Dscam1-mediated self-avoidance counters netrin-
dependent targeting of dendrites in drosophila. Curr Biol, 21(17):1480–7, Sep 2011.
[31] Ian N Sneddon. The relation between load and penetration in the axisymmetric boussinesq problem
for a punch of arbitrary profile. International Journal of Engineering Science, 3(1):47–57, 1965.
[32] Peter Grassberger and Itamar Procaccia. Measuring the strangeness of strange attractors. Physica
D: Nonlinear Phenomena, 9(1–2):189 – 208, 1983.
12
|
1012.1611 | 2 | 1012 | 2011-03-05T17:28:20 | Broken chaotic clocks of brain neurons and depression | [
"q-bio.NC",
"cond-mat.dis-nn",
"nlin.CD"
] | Irregular spiking time-series obtained in vitro and in vivo from singular brain neurons of different types of rats are analyzed by mapping to telegraph signals. Since the neural information is coded in the length of the interspike intervals and their positions on the time axis, this mapping is the most direct way to map a spike train into a signal which allows a proper application of the Fourier transform methods. This analysis shows that healthy neurons firing has periodic and chaotic deterministic clocks while for the rats representing genetic animal model of human depression these neuron clocks might be broken, that results in decoherence between the depressive neurons firing. Since depression is usually accompanied by a narrowing of consciousness this specific decoherence can be considered as a cause of the phenomenon of the consciousness narrowing as well. This suggestion is also supported by observation of the large-scale chaotic coherence of the posterior piriform and entorhinal cortices' electrical activity at transition from anesthesia to the waking state with full consciousness. | q-bio.NC | q-bio |
Broken chaotic clocks of brain neurons and depression
ICAR, P.O.B. 31155, Jerusalem 91000, Israel
A. Bershadskii
Irregular spiking time-series obtained in vitro and in vivo from singular brain neurons of different
types of rats are analyzed by mapping to telegraph signals. Since the neural information is coded
in the length of the interspike intervals and their positions on the time axis, this mapping is the
most direct way to map a spike train into a signal which allows a proper application of the Fourier
transform methods. This analysis shows that healthy neurons firing has periodic and chaotic de-
terministic clocks while for the rats representing genetic animal model of human depression these
neuron clocks might be broken, that results in decoherence between the depressive neurons firing.
Since depression is usually accompanied by a narrowing of consciousness this specific decoherence
can be considered as a cause of the phenomenon of the consciousness narrowing as well. This sug-
gestion is also supported by observation of the large-scale chaotic coherence of the posterior piriform
and entorhinal cortices' electrical activity at transition from anesthesia to the waking state with full
consciousness.
PACS numbers: 87.19.L, 87.19.ll, 87.19.lm
I.
INTRODUCTION
All types of information, which is received by sen-
sory system, are encoded by nerve cells into sequences
of pulses of similar shape (spikes) before they are trans-
mitted to the brain. Brain neurons use such sequences
as main instrument for intercells connection. The infor-
mation is reflected in the time intervals between succes-
sive firings (interspike intervals of the action potential
train, see Fig. 1). There need be no loss of information
in principle when converting from dynamical amplitude
information to spike trains [1] and the irregular spike se-
quences are the foundation of neural information process-
ing. Although understanding of the origin of interspike
intervals irregularity has important implications for elu-
cidating the temporal components of the neuronal code
and for treatment of such mental disorders as depression
and schizophrenia, the problem is still very far from its
solution.
The mighty Fourier transform method, for instance, is
practically non-applicable to the spike time trains. The
spikes are almost identical to each other and the neural
information is coded in the length of the interspike in-
tervals and the interspike intervals positions on the time
axis, therefore it is the most direct way to map the spike
train into a telegraph time signal, which has values -1
from one side of a spike and values +1 from another side
of the spike with a chosen time-scale resolution. An ex-
ample of such mapping is given in figure 1. While the
information coding is here the same as for the corre-
sponding spike train, the Fourier transform method is
quite applicable to analysis of the telegraph time-series.
On the other hand, recent dynamical models of neu-
ron activity revealed new and complex role of regimes of
a (deterministic) chaotic irregularity in the neuron spike
trains (see, for instance, [2]-[5]). Therefore, we have to
use all available mathematical tools in order to study the
experimental data on the deterministic chaos properties
(and, especially, in order to separate between determin-
h
p
a
r
g
e
e
t
l
s
e
k
p
s
i
1
0
-1
1
0
0
a
b
5
10
15
20
25
t [s]
FIG. 1: Mapping of a spike train (figure 1b) into a telegraph
signal (figure 1a).
istic chaos and stochasticity in the experimental signals).
In present paper we have analyzed three types of exper-
imentally obtained spike trains: a) obtained in vitro from
a spontaneous activity in CA3 hippocampal slice culture
of a healthy Wistar/ST rat (the raw data and the de-
tail description of the experiment can be found online at
http://hippocampus.jp/data and in Refs. [6],[7]), b) ob-
tained in an electrophysiological in vivo experiment from
neurons belonging to red nucleus of a healthy (Sprague-
Dawley) rat's brain (see Ref.
[8] for more details of the
experiment and a preliminary discussion of the data), and
c) obtained in an electrophysiological in vivo experiment
from neurons belonging to red nucleus of a genetically
depressed (Flinders Sensitive Rat Line) rat's brain (see
Ref. [8] for more details and a preliminary discussion of
the data).
In the in vitro experiment a) a functional imaging
technique with multicell
loading of the calcium fluo-
rophore was used in order to obtain the spike trains of
)
f
(
E
25
20
15
10
5
0
0
0.5
1.5
2
1
f [s-1]
100
10
1
0.1
)
f
(
E
2
-1.7
0
2
4
6
8
10
12
14
16
τ [s-1]
0.01
0.01
0.1
1
10
f [s-1]
)
τ
(
C
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
FIG. 2: Autocorrelation function for the telegraph signal cor-
responding to the cell-21 (800 spikes). Insert in the Fig. 2
shows corresponding spectrum.
FIG. 3: Spectrum of the telegraph signal corresponding to the
cell-21 in log-log scales. The dashed straight line indicates a
power law: E(f ) ∼ f −1.7.
spontaneously active singular neurons in the absence
of external
In the in vivo experiments
b) and c) the rats were anesthetized and the extracel-
lular recordings were processed from the singular cells [8].
input [6],[7].
Motivation to study the hippocampus and red nucleus
areas of brain in relation to the psychomotor aspects of
depression is based on the recently discovered evidences
of their deep involvement in this mental disorder. The
hippocampus is a significant part of a brain system re-
sponsible for behavioral inhibition and attention, spa-
tial memory, and navigation. It is also well known that
spatial memory and navigation of the rats is closely re-
lated to the rhythms of their moving activity. On the
other hand, the hippocampus of a human who has suf-
fered long-term clinical depression can be as much as 20%
smaller than the hippocampus of someone who has never
been depressed [9]. Inputs to the Red Nucleus arise from
motor areas of the brain and in particular the deep cere-
bellar nuclei (via superior cerebellar peduncle; crossed
projection) and the motor cortex (corticorubral; ipsilat-
eral projection). On the other hand it is known that
humans with deep depression have intrinsic locomotors
problems. Therefore, investigation of Red Nucleus for
genetically defined rat model of depression (these rats
partially resemble depressed humans because they ex-
hibit reduced appetite and psychomotor function) can
be useful for understanding the mental disorder origin.
II. NEURON CLOCK
In the in vitro experiment with spontaneous activ-
ity of the hippocampal pyramidal cells different levels
of activity were observed for different neurons [6],[7].
We take for our analysis the two most active neurons
(http://hippocampus.jp/data - Data-006, cell-21, with
800 spikes in the time-series; and cell-25, with 692 spikes).
100
10
1
0.1
)
f
(
E
-1.7
0.01
0.01
0.1
1
10
f [s-1]
FIG. 4: As in Fig. 3 but for cell-25.
The spike trains were mapped to telegraph signals as it is
described above (see also [10]). Figure 2 shows autocor-
relation function for the telegraph signal corresponding
to the cell-21 (800 spikes).
Insert in the Fig. 2 shows
corresponding spectrum. Both the correlation function
and the spectrum provide clear indication of a strong
periodic component in the signal (the oscillations in the
correlation function and the peak in the spectrum). The
periodic component can be seen at frequency f0 ≃ 0.3Hz.
Figure 3 shows the spectrum in log-log scales. One can
see that at high frequencies the spectrum exhibits a scal-
ing behavior (power law: E(f ) ∼ f −1.7, as indicated by
the dashed straight line). The real power law can be more
pronounced but under the experimental conditions indi-
vidual spikes emitted at firing rates higher than 5Hz were
experimentally inseparable [6],[7]. Figure 4 shows spec-
trum of the telegraph signal corresponding to the spike
train obtained for the cell-25 (692 spikes). The spectrum
is rather similar to the spectrum shown in Fig. 3 (for
cell-21). The more broad peak in Fig. 4 can be related
3
20
15
10
5
0
-5
X
-10
-15
0
threshold
2000
4000
t
6000
8000
30
25
20
Z
15
10
5
0
0
50
100
150
t
200
250
300
FIG. 5: X-component fluctuations of a chaotic solution of the
Rossler system Eq. (1) (a = 0.15, b = 0.20, c = 10.0).
FIG. 6: Z-component fluctuations of a chaotic solution of the
Rossler system.
to the poorer statistics for the cell-25 in comparison with
cell-21. The spectra were calculated using the maximum
entropy method (because it provides an optimal spectral
resolution even for small data sets [11]).
III. THRESHOLD EFFECT
Now let us speculate about physics which could result
in the spectra observed in Figs. 3 and 4. And let us recall
some basic electrochemical properties of neuron. Nerve
cells are surrounded by a membrane that allows some
ions to pass through while it blocks the passage of other
ions. When a neuron is not sending a signal it is said to
be "at rest". At rest there are relatively more sodium
ions onside the neuron and more potassium ions inside
that neuron. The resting value of the membrane electro-
chemical potential P (the voltage difference across the
neural membrane) of a neuron is about -70mV. If some
event (a stimulus) causes the resting potential to move
toward 0mV and the depolarization reaches about -55mV
(a "normal" threshold) a neuron will fire an action poten-
tial. The action potential is an explosive release of charge
between neuron and its surroundings that is created by
a depolarizing current. If the neuron does not reach this
critical threshold level, then no action potential will fire.
Also, when the threshold level is reached, an action po-
tential of a fixed size will always fire (for any given neu-
ron the size of the action potential is always the same).
Depending on different types of voltage-dependent ion
channels, different types of action potentials are gener-
ated in different cells types and the qualitative estimates
of the potentials and time periods can be varied. Recent
reconstructions of a driver of the membrane potential us-
ing the neuron spike trains indicate the Rossler oscillator
as the most probable (and simple) candidate (see, for in-
stance, Refs.
[10],[12]-[17]). Figure 5 shows as example
the x-component fluctuations of a chaotic solution of the
Rossler system [18]
100
10
1
0.1
)
f
(
E
-1.7
0.01
0.001
0.01
0.1
1
f
FIG. 7: Spectrum of the telegraph signal corresponding to
the spike train generated by the x-component fluctuations
overcoming the threshold x = 7. The dashed straight line
indicates a power law: E(f ) ∼ f −1.7.
dx
dt
= −y − z;
dy
dt
= x + ay;
dz
dt
= b + xz − cz
(1)
where a, b and c are parameters). At certain values of
the parameters a,b and c the z-component of the Rossler
system is a spiky time series [19],[20]: Fig. 6.
It can
be shown that the Rossler system and the well known
Hindmarsh-Rose model [21] of neurons are subsystems of
the same differential model with a spiky component [20].
Previously the 'spiky' component of such models was in-
terpreted and studied as a simulation of a neuronal out-
put. For the spontaneous neuron firing (without external
stimulus), however, we suggest to reverse the approach
and consider the spiky variable as the main component of
the electrical input (which naturally should have a 'spiky'
character, see above) to the neuron under consideration.
For each neuron the height of the spikes, which the neu-
ron generates, is about the same. However, the heights
E(f) ~ exp-f/fe
4
E(f) ~ exp-f/fe
Te=1/fe=0.9 s
3
2
1
0
)
f
(
E
n
l
0
0.1
0.2
0.3
0.4
0.5
f
0
1
3
4
2
f [s-1]
)
f
(
E
n
l
10
5
0
-5
-10
-15
-20
FIG. 8: Spectrum of the x-component fluctuations shown in
Fig. 5. We used the semi-log axes in order to indicate expo-
nential decay of the spectrum.
FIG. 9: Spectrum of the telegraph signal corresponding to a
healthy red nucleus cell. The data is shown in the semi-log
scales in order to indicate the exponential decay Eq. 2 (the
straight line).
of the spikes generated by different neurons are differ-
ent. Also the signals coming from different neurons to
the neuron under consideration have to go through the
electrochemical passes with different properties. There-
fore, the spiky z-time-series (Fig. 6) can naturally rep-
resent a multineuron signal, which can be considered as
a spontaneous input for the neuron under consideration.
If we use the usual interpretation of the x-component
as a driver of the membrane potential P (x) and the y-
component as that taking into account the transport of
ions across the membrane through the ion channels [21],
then the position of the input (the component z) in the
first equation of the system Eq. (1) has a good phys-
ical background (cf. Ref.
[21]). Then, the quadratic
nonlinearity in the third equation of the system Eq. (1)
can be interpreted as a simple (in the Taylor expansion
terms) feedback of the neuron to the main component
of the neuronal input. This model with the strong non-
linear feedback can be relevant to the most active neu-
rons of a spontaneously active brain (see below results of
an in vitro experiment with a spontaneous brain activ-
ity). The details of the function P (x) is not significant
for the threshold firing process, what really matters is
that the membrane potential function P (x) reaches its
firing value when (and only when) its argument x crosses
certain threshold from below. In this simple model the
driving variable x may overcome its threshold value (Fig.
5) due to the deterministic (chaotic) spontaneous stimu-
lus. Let us consider an output spike signal resulting from
overcoming a threshold value x = 7, for instance. Fig. 7
shows spectrum of the telegraph signal corresponding to
the spike signal.
One can compare Fig. 7 with Figs. 3 and 4 to see very
good reproduction of the main spectral properties.
In order to understand what is going on here we show
in figure 8 spectrum of the x-component itself.
The semi-log scales are used in these figures in order to
indicate exponential decay in the spectra (in the semi-log
scales this decay corresponds to a straight line):
E(f ) ∼ e−f /fe
(2)
While the peak in the spectrum corresponds to the funda-
mental frequency, f0, of the Rossler chaotic attractor, the
rate of the exponentional decay (the slope of the straight
line in Fig. 8 provides us with and additional charac-
teristic frequency fe. Thus Rossler chaotic attractor has
two clocks: periodic with frequency f0 and chaotic with
frequency fe. If one compares Fig. 8 and Fig. 7 one can
see that the periodic clock survived the threshold cross-
ing (with a period doubling, see Appendix). The chaotic
clock, however, did not survive the threshold crossing at
spontaneous activity: the exponential decay in Fig. 8 has
been transformed into a scaling (power law) decay in Fig.
7, which has no characteristic frequency (scale invari-
ance). The scaling exponent value '-1.7' is not sensitive
to a reasonable variation of the threshold value (∼ 20%)
and even to Gaussian fluctuations of the threshold value.
Therefore, it is not just a coincidence that the scaling law
in the Rossler case agrees with results of the in vitro ex-
periment (cf. also [22],[23],[24] and Fig. 16a).
It should
be noted that for a wide class of deterministic systems a
broad-band spectrum with exponential decay is a generic
feature of their chaotic solutions Refs. [11],[25]-[27]. It is
shown in Ref. [25] that the characteristic frequency
fe = X
λ+
i
i
(3)
where λ+
system.
i are positive Lyapunov exponents of the chaotic
)
f
(
E
n
l
4
3
2
1
0
-1
-2
0
E(f) ~ exp-f/fe
Te=1/fe=0.9s
1
2
f [s-1]
3
4
FIG. 10: As in Fig. 9 but for another healthy cell.
IV. CHAOS VS. STOCHASTICITY IN NEURON
FIRING
Both stochastic and deterministic processes can result
in the broad-band part of the spectrum, but the decay in
the spectral power is different for the two cases. An expo-
nential decay with respect to frequency refers to chaotic
time series while a power-law decay indicates that the
spectrum is stochastic.
Figure 9 shows a power spectrum obtained by the fast
Fourier transform method applied to a telegraph signal
mapped from a spike train measured in the red nucleus
of a healthy rat (we can use the fast Fourier transform
here due to sufficiently large number of spikes in the spike
train: 2170). The spike train corresponds to a singular
neuron firing. Figure 10 shows analogous spectrum ob-
tained from another healthy red nucleus's neuron (2139
spikes). The semi-log scales are used in these figures in
order to indicate exponential decay in the spectra (unlike
the situation discribed above for a spontaneous activity):
in the semi-log scales this decay corresponds to a straight
line - Eq. 2. The characteristic frequency fe ≃ 1.1Hz in
the both cases. Figure 11 shows a power spectrum ob-
tained by the fast Fourier transform method applied to a
telegraph signal mapped from a spike train (2022 spikes)
measured in the red nucleus of a genetically depressive
(the "Flinders" line) rat. The spike train corresponds
to a singular neuron firing. Figure 12 shows analogous
spectrum obtained from another genetically depressive
red nucleus's neuron (2048 spikes). The log-log scales
are used in these figures in order to indicate a power-
law decay in the spectra (in the log-log scales this decay
corresponds to a straight line):
E(f ) ∼ f −α
(4)
In this (scaling) situation there is no characteristic time
scale. The scaling exponent α ≃ 1.5 ± 0.1 and ≃ 1.4 ± 0.1
for these two cases.
5
-1.5
-2
-1
0
1
ln f [s-1]
5
4
3
2
1
0
-1
-2
)
f
(
E
n
l
FIG. 11: Spectrum of the telegraph signal corresponding to
a genetically depressed red nucleus cell. The data are shown
in the log-log scales in order to indicate the power law decay
Eq. 4 (the straight line).
)
f
(
E
n
l
5
4
3
2
1
0
-1
-2
-1.4
-2
-1
0
1
ln f [s-1]
FIG. 12: As in Fig. 11 but for another genetically depressed
red nucleus cell.
V. CHAOTIC NEURAL COHERENCE AND
DEPRESSION
In order to work together the brain neurons have to
make adjustment of their rhythms. The main problem
for this adjustment is the very noisy environment of the
brain neurons.
If their work was based on pure peri-
odic inner clocks this adjustment would be impossible
due to the noise. The nature, however, has another op-
tion. This option is a chaotic clock. In chaotic attrac-
tors certain characteristic frequencies can be embedded
by broad-band spectra, that makes them much more sta-
ble to the noise perturbations [28].
In the light of presented results one can conclude that
for the considered cases the healthy neurons firing has de-
terministic clocks (periodic and chaotic), while the genet-
ically depressive red nucleus's neurons exhibited a pure
1
0.8
0.6
0.4
0.2
y
c
n
e
r
e
h
o
c
healthy
depressed
0
0
0.005
0.01
0.02
0.025
0.03
0.015
f [s-1]
6
E(f) ~ exp-f/fe
Te=1/fe=0.1s
0
20
40
60
80
100
f [s-1]
-140
-150
-160
-170
-180
-190
-200
]
B
d
[
)
f
(
E
FIG. 13: Comparison of coherency in firing for the healthy
(solid curve) and for the genetically depressed (doted curve)
neuron pairs in a low-frequency domain (the in vivo experi-
ments).
FIG. 14: Spectrum of local field potentials for the posterior
piriform (the data were taken from Ref. [34]). The data are
shown in the semi-log scales in order to indicate the exponen-
tial decay Eq. (2) (the straight line).
stochastic firing and it seems that their background de-
terministic clocks were broken. The existence of the back-
ground clocks can be utilized by the healthy neurons for
synchronization of their activity [2],[6],[7],[29]-[31].
In order to compare coherent properties of the healthy
and the depressive neuron pairs we will use cross-spectral
analysis. The cross spectrum E1,2(f ) of two processes
x1(t) and x2(t) is defined by the Fourier transformation of
the cross-correlation function normalized by the product
of square root of the univariate power spectra E1(f ) and
E2(f ):
E1,2(f ) = Pτ hx1(t)x2(t − τ )i exp(−i2πf τ )
2πpE1(f )E2(f )
(5)
the bracket h...i denotes the expectation value. The cross
spectrum can be decomposed into the phase spectrum
φ1,2(f ) and the coherency C1,2(f ):
E1,2(f ) = C1,2(f )e−iφ1,2(f )
(6)
Because of the normalization of the cross spectrum the
coherency is ranging from C1,2(f ) = 0, i.e. no linear
relationship between x1(t) and x2(t) at f , to C1,2(f ) = 1,
i.e. perfect linear relationship.
Figure 13 shows comparison of coherency in firing for
the healthy (solid curve) and for the genetically depressed
(doted curve) neuron pairs in a low-frequency domain
(the in vivo experiments). Despite of the deep anesthesia
the healthy neurons exhibit bands of rather high (> 0.5)
coherency in the low-frequency domain, while the depres-
sive neurons activity is rather decoherent in this domain.
VI. LONG-RANGE CHAOTIC COHERENCE
The chaotic coherence can involve a large number of
the healthy neurons and may be the entire brain. The
multi-second oscillations, for instance, are known to be
synchronized nearly brain-wide [32],[33]. In the case of
depression, however, the chaotic neuron clocks can be
broken in a significant part of the brain neurons. That
can result in certain decoherence in different parts of
the brain. Since depression is usually accompanied by
a narrowing of consciousness (and a distorted sense of
time) this specific decoherence could be considered as a
cause of the phenomenon of the consciousness narrowing
as well. The coherence is important for attention,
sensorimotor processing, etc.. In humans, in particular,
being low in attentional flexibility magnified the effects
of private self-focused attention so typical for depressive
persons.
In order to support the possibility of the extended
chaotic coherence we will use analysis of simultaneously
recorded local field potentials from three sites along
the olfactory-entorhinal axis (the anterior piriform,
posterior piriform, and entorhinal cortices: aPIR, pPIR
and Ent C) reported in a recent paper [34]. The
measurements reported in the Ref.
[34] were performed
in lightly anesthetized healthy rats (the Long-Evans rats
with electrode bundles implanted in their anterior and
posterior cortices, and with vertical, silicon probes in
their entorhinal cortices), which were emerged from the
anesthesia to the waking state with full consciousness.
Since the measured local field potentials time series
are not spiky ones one does not need in the special
data mappings in this case. The authors of the Ref.
[34] discovered a new form of coherent neural activity
across the three widely separated brain sites, which they
named Synchronous Frequency Bursts (SFBs). The
high-energy bursts of spontaneous momentary synchrony
were observed across widely separated olfactory and
entorhinal sites (which have also different architecture:
the 6 layers of the entorhinal cortex vs. the three layers
of the piriform cortices). Moreover, a significant rate of
0.7
0.6
0.5
0.4
0.3
0.2
0.1
y
c
n
e
r
e
h
o
c
0
0
5
10
20
25
30
15
f [s-1]
FIG. 15: Coherency calculated for the SFBs in the posterior
piriform and entorhinal cortices (the data were taken from
Ref. [34]).
the SFBs simultaneous occurrences was also observed
across the different functional processing systems: motor
and olfactory ones.
The stereotypical duration of the SFBs was about
250 ms and the power spectra taken across the events
were exponentially decaying. Figure 14 shows a typical
spectrum for the posterior piriform area. The straight
line is drawn in this figure in order to indicate the
exponential decay Eq.
(2) in the semi-log scales (cf.
Figs. 9 and 10 for the singular neuron firing). The
exponential decay indicates a chaotic nature of these
bursts (see above). The decay rate Te = 1/fe ≃ 0.1s is
significantly smaller than that observed for the singular
neurons (Figs. 9 and 10). Taking into account Eq. (3)
one can conclude that the chaotic mixing in the phase
space (determined by the Lyapunov's exponents) is
much more active for the multineuron activity than for
the singular neuron firing (that seems quite natural).
This more active mixing shifts the spectrum into more
high frequency range. Moreover, one can expect that
expansion (globalization) of the chaotic coherence on the
larger brain areas will shift the coherent chaotic activity
even into the higher frequency ranges (cf. below).
The authors of the Ref.
[34] computed coherence
across SFBs in a pair of brain regions. Figure 15 shows
the coherency calculated for the SFBs in the posterior
piriform and entorhinal cortices (which are separated in
brain space by about 8mm). One can compare this figure
with the Fig. 13 (where the coherency was calculated
for a pair of neighboring neurons). In this case the fre-
quency bands of high coherency can be observed as well.
The coherent frequency-range is shifted considerably in
the high frequency direction for the multineuron case
(see above for a reason of this shift). Actually, "the
main purpose of SFBs might be to coordinate multiple
frequency bands across different processing subsystems"
[34].
Such coordination provides a sufficient level of
7
coherence for the work of these separated subsystems
with speed and efficiency impossible in the case of
transmission of a specific behavioral content. This can
be considered as the main advantage of the chaotic
coherence. The hardware for these effective 'manage-
ment' can be provided (at least partially) by recently
discovered in the cortex and hippocampus interneuronal
networks with long-range axonal connections [35],[36]
and for the high frequency γ-range (30-90Hz) oscillations
"via neurons (and glia) inter-connected by electrical
synapses called gap junctions which physically fuse and
electrically couple neighboring cells."[37].
The authors of the Ref.
[34] observed also that the
SFBs occurrence is a function of level of consciousness.
They found "that the SFBs occurred far more often
under light anesthesia than deeper anesthetic states,
and were especially prevalent as the animals regained
consciousness". They did not observe the SFBs after
the rats regained full alertness, but as they comment
this can be a technical problem of inferring the specific
signal from the highly complex local field potential of
the awake state. Therefore, one cannot rule out the
possibility that the phenomenon is still in a full swing
also in the fully consciousness state (at least at certain
conditions).
Finally, it should be noted that the transitional states
of consciousness (emerging and decaying) have a very
interesting relationship to associative human creativity
(H. Poincare called these states as semi-somnolent condi-
tions, see Ref.
[38], Chapter: Mathematical discovery).
The very creative and unexpected associative ideas
that come in these states can have the above described
long-range chaotic coherence as their direct physical
background. Moreover, the same mechanism can also be
in work at full consciousness (see previous paragraph).
In this case, however, its results are considered as ones
coming from the 'clear sky' and we tend to interpret
them (may be wrongly) as a result of a prolonged period
of unconscious work.
In the full consciousness state
these results are more often turn out to be adequate
ones, unlike of those obtained in the transitional states
[38]).
"A new result has value, if any, when, by establish-
ing connections between elements that are known but
until then dispersed and apparently unrelated to one an-
other, order is immediately created where chaos seemed
to reign" [38].
VII. ACKNOWLEDGMENTS
I thank Dremencov E. and Ikegaya Y. for sharing the
data and discussions and also Allegrini P. and Grigolini
P. for comments and suggestions. I thank Greenberg A.
for help in computing.
VIII. APPENDIX
In order to understand how the fundamental chaotic
clock survives the threshold firing let us consider a very
simple and rough model, which allows analytic calcula-
tion of its autocorrelation function.
In this model the
spike firing takes place at discrete moments: tn = nT +ζ,
where ζ is an uniformly distributed over the interval [0, T ]
random variable, n = 1, 2, 3... and T is a fixed period.
Then let us consider a telegraph signal constructed for
this spike train as it has been described above. If p is
a probability of the spike firing at a current moment
(0 ≤ p < 1), then the autocorrelation function of such
telegraph signal is:
C(τ ) = (n − τ /T )(2p − 1)n−1 + (τ /T − (n − 1))(2p − 1)n
(A1)
in the interval (n − 1)T ≤ τ < nT . Figure 16b shows
the autocorrelation function Eq. (A1) calculated for p =
0.25, as an example. For comparison figure 16a shows
also a superposition of the autocorrelation functions for
the telegraph signals corresponding to the cell-21 (the
solid line) and to the spike train generated by the the
Rossler attractor fluctuations overcoming the threshold
x = 7 (circles).
In order to make the autocorrelation
functions comparable a rescaling has been made for the
Rossler attractor generated autocorrelation function.
8
neuron
Rossler
a
0
2
4
6
8
10
12
)
τ
(
C
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
)
τ
(
C
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
τ [s-1]
model
b
T
2T
3T
4T
5T
6T
τ
FIG. 16: a) Autocorrelation functions for the telegraph sig-
nals corresponding to the cell-21 (solid curve) and to the spike
train generated by the Rossler attractor fluctuations overcom-
ing the threshold x = 7 (circles). b) Autocorrelation function
for the simple model telegraph signal: Eq. (A1) with p = 0.25.
[1] T. Sauer, Chaos, 5, 127 (1995).
[2] E. M. Izhikevich, Dynamical Systems in Neuroscience:
The Geometry of Excitability and Bursting (MIT, Cam-
bridge, MA, 2006).
[3] G. S. Medvedev, Phys. Rev. Lett., 97, 048102 (2006).
[4] A. Shilnikov and G. Cymbalyuk, Phys. Rev. Lett., 94,
(2002)
[15] T. Gedeon , M. Holzer, and M. Pernarowski, Physica D,
178, 267 (2007)
[16] N. Crook, W.J. Goh, and M. Hawarat, BioSystems, 87,
267 (2007).
[17] T. Pereira , M.S. Baptista, and J. Kurths, Phys. Lett. A,
048101 (2005).
362 159 (2007).
[5] H. Korn and P. Faure, C. R. Biologie, 326, 787 (2003).
[6] T. Sasaki, N. Matsuki and Y. Ikegaya, J. Neuroscience,
[18] O.E. Rossler, Phys. Lett. A, 57 397 (1976).
[19] C. Lainscsek, C.Letellier and I. Gorodnitsky, Phys. Lett.
27, 517 (2007).
A, 314 409 (2003).
[7] N. Takahashi, T. Sasaki, W. Matsumoto, N. Matsuki,
and Y. Ikegaya, PNAS, 107, 10244 (2010).
[8] A. Bershadskii, E. Dremencov, D. Fukayama and G. Ya-
did, Phys. Lett. A, 289, 337 (2001).
[9] J.D. Bremner, M. Narayan, E.R Anderson et al., Am. J.
Psychiatry, 157, 115 (2000).
I. Gorodnitsky
dynamics
and C. Letellier,
[20] C. Lainscsek,
from amplitude mea-
Reconstructing
8th Joint Symposium
sures of
spiky time-series,
on Neural Computation,
(2001)
at
http://www.its.caltech.edu/∼jsnc/2001/Proceedings/).
[21] J.L. Hindmarsh and R.M. Rose, Proc. R. Soc. Lond. B,
(available
[10] A. Bershadskii and Y. Ikegaya, arXiv:1010.4722 (avail-
221, 87 (1984).
able at http://arxiv.org/abs/1010.4722) (2010).
[11] N. Ohtomo, K. Tokiwano, Y. Tanaka et. al., J. Phys. Soc.
Jpn. 64 1104 (1995).
[12] R. Castro and T. Sauer, Phys. Rev. E, 59, 2911 (1999).
[13] N. Masuda, and K. Aihara, Phys. Rev. Lett., 88 248101
(2002).
[14] K. Aihara and I. Tokuda, Phys. Rev. E, 66, 026212
[22] P. Allegrini, D. Menicucci, R. Bedini, A. Gemignani, and
P. Paradisi, Complex intermittency blurred by noise: the-
ory and application to neural dynamics Phys. Rev. E, 82,
015103(R) (2010).
[23] M. Lukovi´c and P. Grigolini, Power spectra for both in-
terrupted and perennial aging processes, J. Chem. Phys.
129, (2008) 184102.
9
[24] P. Grigolini, G. Aquino, M. Bologna, M. Lukovi´c, and
B. J. West, A theory of 1/f noise in human cognition,
Physica A, 388, (2009) 4192-4204.
[25] D.E. Sigeti, Phys. Rev. E, 52, 2443; Physica D, 82, 136
(1995).
[26] J. D. Farmer, Physica D, 4, 366 (1982).
[27] U. Frisch and R. Morf, Phys. Rev., 23, 2673 (1981).
[28] M. I. Rabinovich, and H. D. I. Abarbanel, Neuroscience,
061904 (2005).
[32] D. Contreras, and M. Steriade, J. Neurosci., 15, 604
(1995).
[33] M. Steriade, C Neuroscience 101 243 (2000).
[34] R. Hermer-Vazquez, L. Hermer-Vazquez, and S. Srini-
vasan, Brain Res. Bull., 79, 6(2009).
[35] T. Klausberger, P. Somogyi, Science, 321 53 (2008).
[36] S. Jinno, T. Klausberger, L.F. Marton, et al., J. Neurosci.
87, 5 (1998).
27 8790 (2007).
[29] D.I. Abarbanel, R. Huerta, M. I. Rabinovich, et al.,, Neu-
ral Comput., 8, 1567 (1996).
[37] S. Hameroff, J. Biol. Phys., 36, 71-93 (2010).
[38] H. Poincare, Science and Method (Courier Dover Publi-
[30] F. Mormann, K. Lehnertz, P. David, et al., Physica D,
cations, NY, 2003).
144, 358 (2000).
[31] E. Rossoni, Y. Chen, M. Ding, et al., Phys. Rev. E, 71,
|
1803.08109 | 4 | 1803 | 2018-08-09T15:56:54 | A mechanistic model of connector hubs, modularity, and cognition | [
"q-bio.NC",
"q-bio.QM"
] | The human brain network is modular--comprised of communities of tightly interconnected nodes. This network contains local hubs, which have many connections within their own communities, and connector hubs, which have connections diversely distributed across communities. A mechanistic understanding of these hubs and how they support cognition has not been demonstrated. Here, we leveraged individual differences in hub connectivity and cognition. We show that a model of hub connectivity accurately predicts the cognitive performance of 476 individuals in four distinct tasks. Moreover, there is a general optimal network structure for cognitive performance--individuals with diversely connected hubs and consequent modular brain networks exhibit increased cognitive performance, regardless of the task. Critically, we find evidence consistent with a mechanistic model in which connector hubs tune the connectivity of their neighbors to be more modular while allowing for task appropriate information integration across communities, which increases global modularity and cognitive performance. | q-bio.NC | q-bio | A mechanistic model of connector hubs, modularity, and
cognition
Maxwell A. Bertolero1,2*, B.T. Thomas Yeo3,4, Danielle S. Bassett2,5-7, Mark D'Esposito1
1Helen Wills Neuroscience Institute and the Department of Psychology, University of California, Berkeley, CA 94720-3190, USA
2Department of Bioengineering, School of Engineering and Applied Sciences, University of Pennsylvania, Philadelphia, PA 19104
USA
3Electrical and Computer Engineering, ASTAR-NUS Clinical Imaging Research Centre, Singapore Institute for Neurotechnology &
Memory Networks Programme, National University of Singapore, Singapore 119077, Singapore
4NUS Graduate School for Integrative Sciences and Engineering, National University of Singapore, Singapore
5Department of Electrical & Systems Engineering, School of Engineering and Applied Sciences, University of Pennsylvania,
Philadelphia, PA 19104, USA
6Department of Physics & Astronomy, College of Arts and Sciences, University of Pennsylvania, Philadelphia, PA 19104 USA
7Department of Neurology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104 USA.
*corresponding author ([email protected])
Abstract
The human brain network is modular -- comprised of communities of tightly
interconnected nodes1. This network contains local hubs, which have many connections
within their own communities, and connector hubs, which have connections diversely
distributed across communities2,3. A mechanistic understanding of these hubs and how
they support cognition has not been demonstrated. Here, we leveraged individual
differences in hub connectivity and cognition. We show that a model of hub connectivity
accurately predicts the cognitive performance of 476 individuals in four distinct tasks.
Moreover, there is a general optimal network structure for cognitive performance --
individuals with diversely connected hubs and consequent modular brain networks
exhibit increased cognitive performance, regardless of the task. Critically, we find
evidence consistent with a mechanistic model in which connector hubs tune the
connectivity of their neighbors to be more modular while allowing for task appropriate
information integration across communities, which increases global modularity and
cognitive performance.
Main
The human brain is a complex network that can be parsimoniously summarized by a set
of nodes representing brain regions and a set of edges representing the connections
between brain regions. In network models of fMRI data, each edge represents the
strength of functional connectivity -- the temporal correlation of fMRI activity levels --
between the two nodes (Figure 1). This network model can be used to study global and
local brain connectivity patterns. Brain networks contain communities -- groups of nodes
that are more strongly connected to members of their own group than to members of
other groups(Figure 1)1,4,5. This feature of networks is termed modularity and can be
quantified by the modularity quality index Q (see Methods for equation).
Modularity is ubiquitously observed in complex systems in Nature -- a modular structure
is observed consistently across the brains of very different species, from c elegans to
humans6. Given its ubiquity, modular network organizations are potentially naturally
selected because they reduce metabolic costs. Functional and structural connectivity is
metabolically expensive7-11. A modular architecture with anatomically segregated and
functionally specialized communities reduces the average length and number of
connections -- the network's wiring cost. Moreover, the brain's genotype-phenotype map
is modular, forming groups of phenotypes, including brain communities (e.g., the visual
community)12,13, that are co-affected by groups of genes14. Modularity, at the genetic and
phenotypic level, allows systems to quickly evolve under new selection pressures15,16.
As we noted in earlier work2, modularity potentially increases fitness in information
processing systems 17-19, and network simulations show that modularity allows for robust
network dynamics, in that the connections between nodes can be reconfigured without
sacrificing information processing functions, a process necessary for the evolution of a
network20. Artificial intelligence research has shown that modular networks also solve
tasks faster and more accurately and evolve faster than non-modular networks 21 with
lower wiring costs than non-modular networks22. Critically, modularity is also
behaviorally relevant -- modularity predicts intra-individual variation in working memory
capacity23 and how well an individual will respond to cognitive training24,25.
Within each of these communities, local hubs exist that have strong connectivity to their
own community. The within community strength can be used to measure a node's
locality (see Methods for equation; Figure 1 for a schematic). A high within community
strength reflects that a node has strong connectivity within its own community and is
thus a local hub. Local hubs are ideally wired for segregated processing. Because the
connections of local hubs are predominantly concentrated within their own community
and their functions are likely specialized and segregated, damage to local hubs tends to
cause relatively specific cognitive deficits26,27 and does not significantly alter the modular
organization of the network27. Supporting their more segregated and discrete role in
information processing, their activity levels do not increase as more communities are
involved in a task1.
Yet, a completely modular organization renders the brain extremely limited in function --
without connectivity between communities, information from, for example, visual cortex
could never reach motor cortex and therefore visual information could not be used to
inform movements. Thus, how is information integrated across these mostly segregated
communities? The interdependence between modular communities and integration is a
modern rendition of one of the first observations in neuroscience -- Cajal's conservation
principle, which states that the brain has been naturally selected and is thus organized
by an economic trade-off between minimizing the wiring cost of the network, which
leads to modularity, and more costly connectivity patterns that increase fitness, like the
integrative functions afforded by connections between communities 28-30.
Connector hubs have diverse connectivity across different communities. The
participation coefficient can be used to measure a node' diversity (see Methods for
equation, Figure 1 for a schematic). A high participation coefficient reflects that a node
has connections equally distributed across the brain's communities and is thus labeled
a connector hub. Connector hubs are ideally wired for integrative processing1,28,31-34. In
human brain networks, connector hubs have a particular cytoarchitecture35, are
implicated in a diverse range of cognitive tasks36,37, and are physically located in
anatomical areas at the boundaries between many communities33. Moreover, damage to
connector hubs causes widespread cognitive deficits26 and a decrease in the modular
structure of the network27. During cognitive tasks, connector hubs appear to coordinate
connectivity changes between other pairs of nodes -- activity in connector hubs predicts
changes in the connectivity of other nodes, particularly the connectivity between nodes
in different communities38-40. Connector hubs are also strongly interconnected to each
other, forming a diverse club -- tightly interconnected connector hubs5. Connector hubs
also have connections to almost every community in the network. Thus, they have
access to information from every community. Finally, connector hubs exhibit increased
activity if more communities are engaged in a task, which suggests that connector hubs
are involved in processes that are more demanding as more communities are engaged1.
Connector hubs might be Nature's cheapest solution to integration in a modular
network. Generative models suggest that the diverse club -- tightly interconnected
connector hubs -- potentially evolved to balance modularity and efficient integration5.
However, given the amount of wiring required to link to many different and distant
communities, connector hubs' connectivity pattern dramatically increases wiring costs11.
Despite this cost, connector hubs potentially provide a necessary function -- connector
hubs could be the conductor of the brain's neural symphony.
A parsimonious mechanistic model of these findings is that connector hubs tune
connectivity between communities. Neuronal tuning refers to cells selectively
representing a particular stimulus, association, or information. We introduce the
mechanistic concept of network tuning, in which connections between nodes are
organized to achieve a particular network function or topology, like the integration of
information across communities or decreased connectivity between two communities.
We propose that diverse connectivity across the network's communities allows
connector hubs to tune connectivity between communities to be modular but also allows
for task appropriate information integration across communities. This facilitates a global
modular network structure in which local hubs and nodes within each community are
dedicated to mostly autonomous local processing. The modular network structure
afforded by diversely connected connector hubs -- connector hubs that are wired well for
network tuning -- is potentially optimal for a wide variety of cognitive processes. Thus,
despite their cost, strong and diverse connector hubs might be critically necessary for
integrative processing in complex modular neural networks.
Local and connector hubs have been exhaustively studied by network science and their
functions have been inferred from their topological locations in the network5. Moreover,
individuals' brain network connectivity has been shown to be predictive of task
performance41-45,46 and is able to "fingerprint" individuals47. However, no study has
leveraged these individual differences to test a mechanistic model of hub function and
direct evidence supporting a mechanistic model of these hubs and how they support
human cognition remains absent. Moreover, it is currently unknown if there is a hub and
network structure that is optimal for a diverse set of tasks or if different hub and network
structures are optimal for different tasks. Here, we analyze how individual differences in
the locality and diversity of hubs during the performance a task relates to network
connectivity, modularity, and performance on that task as well as subject measures
collected outside of the scanner, including psychometrics (e.g., fluid intelligence,
working memory) and other behavioral measures (e.g., sleep quality and emotional
states). We test a mechanistic model in which connector hubs tune their neighbors'
connectivity to be more modular, which increases the global modular structure of the
network and task performance, regardless of the particular task.
We leveraged the size and richness of fMRI data from 476 (S500 release) subjects that
participated in the Human Connectome Project48. A network was built for each subject
using fMRI data collected during seven different cognitive states (Resting-State,
Working Memory, Social Cognition, Language & Math, Gambling, Relational, Motor, see
Methods). Thus, each subject has seven different networks for analysis. Each edge
represents the strength of functional connectivity between each pair of 264 nodes49. In
every network, Q and a division of nodes into communities was calculated (see
Methods). Next, in every network, for each node, locality and diversity was measured.
Within community strength measures a node's locality and the participation coefficient
measures a node's diversity, respectively (see Methods for equations). Figure 1
displays this processing workflow.
Figure 1 Functional connectivity and network science processing workflow. a, The mean signal across time is extracted from
264 cortical, sub-cortical, and cerebellar regions, three of which are shown here. b, The time series of the three nodes is shown. To
measure functional connectivity, the Pearson r correlation coefficient between the time series of node i and the time series of node j
for all i and j is calculated. c, The strongest (e.g., the top 5% percent r values) functional connections serve as weighted edges in
the graph (a range of graph densities was explored, see Methods for details). d, The Infomap community detection algorithm is
applied, generating a community assignment for each node, displayed here in different colors in a schematic (top) and the mean
graph across subjects (bottom). e, Given that particular community assignment and network, nodes' participation coefficients are
calculated. Red nodes are high participation coefficient nodes, shown here in a schematic (top) and the mean graph (bottom). f,
Within community strengths are also calculated. Purple nodes are high within community strength nodes, shown here in a schematic
(top) and the mean graph (bottom). The graphs along the bottom are laid out using the force-atlas algorithm, where nodes are
repelling magnets and edges are springs, which causes nodes in the same community to cluster together, nodes that are diversely
connected across communities (connector hubs) to be in the center of the graph, and nodes that are strongly connected to a single
community (local hubs) in the middle of that community. d, lower, A single community (light blue) and its connections to the rest of
the graph is extracted and enlarged, with nodes colored by community. Note that the nodes within each community are more
strongly connected to each other than to nodes in other communities. e, lower, A node (and its connections) with a high participation
coefficient is extracted and enlarged, with nodes colored by community. Note that the connector hub is connected to many different
communities. f, A node (and its connections) with a high within community strength is extracted and enlarged, with nodes colored by
community. Note that the local hub is strongly connected to its own community.
In the proposed mechanistic model of hub connectivity, connector hubs, via their
diverse connectivity, tune the network to preserve or increase global modularity and
local hubs' locality, which, in turn, increases task performance. If this model is true,
hubs' connectivity in the network and network modularity should be predictive of task
performance. Thus, the first test of this model involved using hub diversity and locality,
network connectivity, and modularity to predict task performance. A predictive multilayer
perceptron model (three layers plus the input layer and the output layer (enough for
non-linear relationships); eight neurons per layer (one per feature, with two layers
containing 12 neurons, allowing for higher dimensional expansion)) was used to predict
subjects' task performance (Supplementary Figure 1, Figure 2). Known as deep neural
networks, these predictive models are constructed by tuning the weights between
neurons across adjacent layers to achieve the most accurate relationship between the
features (input) and the value the model is trying to predict (output). The predictive
model's features (n=8) captured how well subjects' nodes' diversity and locality, network
connectivity (i.e., edge weights in the network), and modularity (Q) are optimized for the
performance of a task. For example, for the feature that captures how optimized the
diversity of subjects' nodes' are for task performance, for each node, the Pearson r
across subjects between that node's participation coefficients (which measures
diversity) and task performance values was calculated (Supplementary Figure 1a). We
call this r value the node's diversity facilitated performance coefficient. The feature,
then, for a given subject, is the Pearson r across nodes between each node's diversity
facilitated performance coefficient and each node's participation coefficient in that
subject, representing how optimized the diversity of that subject's nodes' are for
performance in the task (Supplementary Figure 1). Critically, for each subject's feature
calculation, the diversity facilitated performance coefficients are calculated without that
subject's data. The same procedure is executed for locality (using the within community
strengths) and edge weights; instead of participation coefficients, within community
strengths or edge weights are used. Finally, the Q values of the networks are included
in the model.
The predictive model was fit for each of the four cognitive tasks that subjects performed
in the Human Connectome Project for which performance was measured (Working
Memory, Relational, Language and Math, Social tasks; see Methods for task
performance measures). For this and other subject performance analyses, we could not
analyze the Gambling, Motor, or Resting-State tasks, as there was no performance
measured for these tasks. Each predictive model was fit to the subjects' networks
constructed during the performance of each task as well as the resting state (four
features from each). The inclusion of the resting-state and the cognitive task state
allowed the model to capture the subjects' so-called intrinsic network states as well as
the subjects' task driven network states. Using a leave-one-out cross-validation
procedure, the features were constructed, and the model was fit with data from all
subjects except one. The predictive model was then used to predict the left-out subject's
task performance (Supplementary Figure 1c). To test the accuracy of the model, the
Pearson r between the observed and predictive performance of each subject was
calculated (Figure 2, Supplementary Figure 2).
This predictive model significantly (p<0.001, Bonferroni corrected (n tests = 4))
predicted performance in all four tasks (Figure 2). Also, using a predictive model with
only nodes' diversity and locality and modularity features (i.e., ignoring individual
connections in the network) did not dramatically decrease the models' prediction
accuracies (Supplementary Figure 2a,b). Given that head motion is a concern when
analyzing fMRI data, scrubbing techniques were applied to remove motion artifacts from
the fMRI data and the mean frame-wise displacement was regressed out from task
performance. Neither of these additional analyses dramatically decreased the predictive
models' prediction accuracies (Supplementary Figure 2c-f). Finally, in each task,
modularity (Q) alone was only weakly correlated with task performance (Working
Memory, Pearson's r (dof=471):0.303, p<0.001, CI:0.219,0.383; Relational, Pearson's r
(dof=455):0.106, p:0.095, CI:0.014,0.196; Language & Math, Pearson's r
(dof=469):0.085, p:0.259, CI:-0.005,0.174; Social, Pearson's r (dof=471):0.084, p:0.275,
CI:-0.006,0.173, all confidence intervals=95%). These results suggest that the diversity
and locality of nodes, in combination with the modular connectivity structure of the
network, are highly predictive of task performance.
The Human Connectome project contains psychometrics and other behavioral
measures collected outside of the MRI scanner; for clarity and to differentiate these
measures from the task performance measures and the tasks' corresponding networks,
we call these "subject measures" 50. If a particular hub and network structure is
generally optimal for many different types of cognition and many different behaviors (a
component of the mechanistic model of hub function), then the tasks' optimal hub and
network structures should be similarly optimal across subject measures -- sub-optimal
for negative measures like poor sleep, sadness, and anger and optimal for positive
measures like life satisfaction and processing speed.
Figure 2 Hub diversity and locality, modularity, and network connectivity predict cognitive performance. a, for each task,
the correlation between task performance and the performance predicted by a predictive model of hub diversity and locality,
modularity, and network connectivity. Each data point represents the (y-axis) true performance (see Methods, each task's
performance value scale is unique) of the subject and the (x-axis) predicted performance of the subject by the neural network.
Shaded areas represent 95 percent confidence intervals. In every task, the predictive model significantly predicted task performance
(p < 1e-3, Bonferroni corrected (n tests=4), N=Working Memory: 473, Relational: 457, Language & Math: 471, Social: 473). b, we
correlated the tasks' feature correspondence values (see Supplementary Figure 3 for each task's feature correspondence with each
subject measure) -- measuring if the two tasks' optimal hub and network structures are also optimal for the same subject measures.
High correlations mean that the two tasks' hub and network structures are similarly optimal for the same subject measures (all
results significant at p < 1e-3, Bonferroni corrected (n tests = 4), dof=45, N=47, the number of subject measures, while the feature
correspondence N =Working Memory: 473, Relational: 457, Language & Math: 471, Social: 473).
For each task, the predictive model constructs features that capture how optimal each
subject's hub and network structure is for performance on that task. Using the subjects'
networks from a given task, the predictive model of hub and network structure can
construct features that capture how optimal each subject's hub and network structure is
for a given subject measure collected outside of the scanner instead of task
performance. This was executed using the networks from each of the four tasks for all
subject measures. Thus, the predictive model constructs features that capture how
optimal each subject's hub and network structure (measured during the performance of
a task (e.g. Working Memory)) is for a given subject measure (e.g., Delayed
Discounting). The correspondence between the features in the two models -- how
similarly optimal subjects' hub and network structure are for the task and a given subject
measure -- can then by calculated by, across subjects, computing the correlation
between the features in the two predictive models. Specifically, for each feature, the
correlation, across subjects, between the feature in the predictive model fit to task
performance (e.g., Working Memory) and that feature in the predictive model that was fit
to a subject measure (e.g., Delayed Discounting) is computed. The mean correlation
across the edge, locality, and diversity features (n=6, three features from resting-state
and three features from the task) is then calculated, which we call feature
correspondence. The Q feature was ignored, as the Q feature remains constant
regardless of what the model is fit to. Thus, this value determines if each task's optimal
hub and network structure is optimal for other subject measures and if all the tasks'
optimal hub and network structures are similarly optimal for other subject measures.
For each task, the hub and network structures that were optimal for that task were
typically also optimal for positive subject measures and sub-optimal for negative subject
measures (Supplementary Figure 3). Next, the similarity by which two tasks' optimal
hub and network structures generalized to other subject measures can be measured by
correlating the feature correspondence values (for example, the Working Memory and
Social columns in Supplementary Figure 3). High Pearson r correlations were found
between all tasks (r (dof=45) values between 0.82 and 0.96, p<0.001 Bonferroni
corrected (n tests = 4), Figure 2b). Finally, the predictive model was able to significantly
predict most subject measures (Supplementary Figure 4). These results demonstrate
that, if an individual has a particular brain state during a given task, as defined by the
connectivity of the network's hubs, that is optimal for that given task, it also likely
optimal for other subject measures. Critically, different tasks' optimal hub and network
structures are similarly optimal for other subject measures. Moreover, these findings
demonstrate that the predictive model captures hub connectivity patterns in the network
that are relevant for behavior and cognition in general, instead of overfitting hub
connectivity patterns that are only related to a particular cognitive process or behavior.
Having established relationships between hub locality and diversity, modularity, and
task performance, we sought to test the mechanistic claim that diverse connector hubs
increase modularity by analyzing how individual differences in a node's diversity within
the network are predictive of individual differences in brain network modularity (Q; see
Methods for mathematical definition). Typically, the result of damage to a region can be
used to infer the function of that region -- if a region is damaged and modularity
decreases, the region is putatively involved in preserving modularity. Here, we analyze
the other direction -- when a hub is diversely connected across the brain (i.e., strong), if
modularity increases, the region's diverse connectivity is putatively involved in
preserving modularity (Supplementary Figure 5).
Thus, we first tested if, across subjects, a node's participation coefficients are positively
correlated with modularity (Q). For each node, the Pearson r between that node's
participation coefficients and the network's modularity values (Q) across subjects was
calculated. Intuitively, higher r values indicate that the node's diversity (i.e., the
participation coefficient) is associated with higher network modularity. This is an
important feature that can be used to distinguish the roles of different brain regions. For
ease of presentation, we refer to each node's r value as the diversity facilitated
modularity coefficient, as it measures how the diversity of the node's connections
facilitates (we use this term to remain causally agnostic) the modularity of the network.
For every node, the Pearson r between the within community strengths and Q values
across subjects was also calculated. Intuitively, higher r values indicate that the node's
locality (i.e., the within community strength) is associated with higher network
modularity. We refer to each node's r value (between within community strengths and Q
values across subjects) as the locality facilitated modularity coefficient, as it measures
how the locality of the node's connections facilitates the modularity of the network.
We performed these computations separately for all seven distinct cognitive states. In
all states, the diversity facilitated modularity coefficients of connector hubs (top 20
percent highest participation coefficient nodes) were shown to be significantly higher
than other nodes in a Bonferroni-corrected independent two-tailed t-test (Figure 3a,
Working Memory t(dof:262):7.182, p<0.001, Cohen's d:1.104, CI:0.062,0.117, Gambling
t(dof:262):4.101, p:0.0004, Cohen's d:0.63, CI:0.025,0.052, Language & Math
t(dof:262):7.292, p<0.001, Cohen's d:1.12, CI:0.062,0.102, Motor t(dof:262):7.354,
p<0.001, Cohen's d:1.13, CI:0.088,0.13, Relational t(dof:262):4.457, p:0.0001, Cohen's
d:0.685, CI:0.038,0.075, Resting State t(dof:262):3.947, p:0.0007, Cohen's d:0.606,
CI:0.029,0.096, Social t(dof:262):3.716, p:0.0017, Cohen's d:0.571, CI:0.022,0.051. P
values Bonferroni corrected (n tests=7), all confidence intervals=95%). Moreover, in all
cognitive states, the locality facilitated modularity coefficients of local hubs (top 20
percent highest within community strength nodes) were shown to be significantly higher
than other nodes in a Bonferroni-corrected independent two-tailed t-test (Figure 3b,
Working Memory t(dof:262):5.415, p<0.001, Cohen's d:0.832, CI:0.045,0.093, Gambling
t(dof:262):4.959, p<0.001, Cohen's d:0.762, CI:0.034,0.074, Language & Math
t(dof:262):6.428, p<0.001, Cohen's d:0.988, CI:0.045,0.085, Motor t(dof:262):9.822,
p<0.001, Cohen's d:1.509, CI:0.101,0.146, Relational t(dof:262):6.131, p<0.001,
Cohen's d:0.942, CI:0.036,0.07, Resting State t(dof:262):0.966, p:1.0, Cohen's d:0.148,
CI:-0.014,0.038, Social t(dof:262):4.54, p:0.0001, Cohen's d:0.698, CI:0.026,0.06. P
values Bonferroni corrected (n tests = 7), all confidence intervals=95%). While the
diversity facilitated modularity coefficients of connector hubs were not always positive,
they were typically close to or above zero. This means that a diverse connector hub can
be associated with increased integrative connectivity between communities without
decreasing the modularity of the network. To more fully understand the relationship
between nodes' diversity and the networks' modularity, the Pearson r between each
node's mean participation coefficient across subjects (which defines a connector hub)
and the node's diversity facilitated modularity coefficient was calculated (Supplementary
Figure 6). Moreover, the Pearson r between each node's mean within community
strength across subjects (which defines a local hub) and the node's locality facilitated
modularity coefficient was calculated (Supplementary Figure 6). In every task, there was
a significant positive correlation between a node's mean participation coefficient and
that node's diversity facilitated modularity coefficient (Supplementary Figure 6a). In
every task, there was also a significant positive correlation between a node's mean
within community strength and its locality facilitated modularity coefficient
(Supplementary Figure 6b). These analyses demonstrate that connector hubs' strong
diverse connectivity to many communities and local hubs' strong local connectivity is
associated with higher brain network modularity, regardless of the subjects' cognitive
state. Thus, these results are consistent with the mechanistic model of connector hub
function, where connector hubs preserve the modular structure of the network via
diverse connectivity.
We confirmed the reliability and reproducibility of these results and demonstrated that
they are not driven by analytically necessary relationships. First, the mean participation
coefficient and within community strength was calculated in one half of the subjects and
the diversity and locality facilitated modularity coefficients were calculated in the other
half of the subjects, testing 10,000 splits (Supplementary Figure 7). Next, four null
models were tested to ensure the current results were not driven by analytically
necessary relationships (Supplementary Figure 8). Other analyses ensured the current
results are not driven by the number of communities (Supplementary Figure 9,
Supplementary Figure 10). Finally, to justify the use of the Pearson r to calculate the
coefficients, the relationship between nodes' diversity and Q was confirmed as typically
linear (Supplementary Figure 11).
Having found evidence supporting a mechanistic model in which connector hubs tune
their neighbors' connectivity to be more modular, thereby increasing the global modular
structure of the network, we next asked if diverse hubs concurrently facilitate higher
modularity and higher task performance. To address this question, the Pearson r
between each node's participation coefficient (or within community strength) and task
performance was calculated. A positive r at a node indicates that a subject with a higher
participation coefficient (or within community strength) at that node performs better on
the task. We refer this r value as the node's diversity facilitated performance coefficient
for participation coefficients, and locality facilitated performance coefficient for within
community strengths. Note that these are the same r values used in the construction of
the predictive performance model features (Supplementary Figure 1). However, for the
predictive model, we calculated these r values with the subject whose behavior was to
be predicted held out. Here, we calculate these r values across all subjects.
In all tasks, the diversity facilitated performance coefficients of connector hubs (top 20
percent strongest) were shown to be significantly higher than other nodes in a
Bonferroni-corrected independent two-tailed t-test, with only the Language & Math task
at p=0.0677 after Bonferroni correction (uncorrected p=0.0169) (Figure 3c, Working
Memory t(dof:262):5.378, p<0.001, Cohen's d:0.826, CI:0.03,0.071, Language & Math
t(dof:262):2.404, p:0.0677, Cohen's d:0.369, CI:0.005,0.04; Relational t(dof:262):2.959,
p:0.0135, Cohen's d:0.455, CI:0.01,0.037; Social t(dof:262):4.744, p<0.001, Cohen's
d:0.729, CI:0.025,0.053. All p values Bonferroni corrected (n tests=4), all confidence
intervals=95%). The locality facilitated performance coefficients of local hubs (top 20
percent strongest) were shown to be significantly higher than other nodes in a
Bonferroni-corrected independent two-tailed t-test (Figure 3d, Working Memory
t(dof:262):2.712, p:0.0285, Cohen's d:0.417, CI:0.008,0.054, Language & Math
t(dof:262):2.864, p:0.0181, Cohen's d:0.44, CI:0.006,0.043; Relational t(dof:262):0.327,
p:1.0, Cohen's d:0.05, CI:-0.016,0.021; Social t(dof:262):1.862, p:0.2547, Cohen's
d:0.286, CI:0.0,0.031. All p values Bonferroni corrected (n tests=4), all confidence
intervals=95%). Moreover, the correlation between each node's diversity facilitated
performance coefficient and each node's mean participation coefficient was positive and
significant (Supplementary Figure 6), suggesting that, for connector hubs, a higher
participation coefficient is associated with higher task performance. The Pearson r
correlation between each node's locality facilitated performance coefficient and each
node's mean within community strength was also positive and significant
(Supplementary Figure 6; all tasks except Relational, Bonferroni (n tests=4) p=0.081,
uncorrected p=0.02), suggesting that, for local hubs, a higher within community strength
is also associated with higher task performance. Finally, there was a significant positive
correlation between a node's diversity facilitated modularity coefficient and a node's
diversity facilitated performance coefficient (Figure 3e) as well as a significant positive
correlation between a node's locality facilitated modularity coefficient and a node's
locality facilitated performance coefficient (Figure 3f). Thus, diverse connector hubs
facilitate higher task performance in proportion to how much they facilitate higher
modularity, suggesting a strong link between the increased modularity afforded by
diverse hubs and increased task performance.
Figure 3 Connector hubs and local hubs concurrently facilitate increased modularity and task performance. For each task,
diversity and locality facilitated modularity coefficients, a measure of how the diversity and locality (respectively) of a node facilitates
modularity, were calculated. In every task, the diversity and locality facilitated modularity coefficients of connector (a) and local hubs
(b), compared to other nodes, is significantly (except Resting-State for locality) higher, demonstrating that strong connector and
local hubs facilitate the modular structure of brain networks. For each task, diversity and locality facilitated performance coefficients
were calculated. In every task the diversity and locality facilitated performance coefficients of connector (c) and local hubs (d),
compared to other nodes, is significantly (except Language for diversity (p=0.0677 after Bonferroni correction (uncorrected
p=0.0169)), Relational and Social for locality) higher, demonstrating that strong connector and local hubs facilitate increased task
performance. For a-d, the mean and quartiles are marked in each violin. Each task's distribution of coefficients was tested for
normality using D'Agostino and Pearson's omnibus test k2. No evidence was found (k2>0.0 for all tasks) that these distributions were
not normal. N=264, the number of nodes in the graph. e, The correlation between a node's diversity facilitated modularity coefficient
and a node's diversity facilitated performance coefficient. f, The correlation between the node's locality facilitated modularity
coefficient and the node's locality facilitated performance coefficient. In panels e,f, N=264, the number of nodes in the graph.
Shaded areas represent 95 percent confidence intervals. All p values are Bonferroni corrected (n tests = 4).
Next, we tested the mechanistic network tuning claim of the model: "do connector hubs
increase Q by tuning the connectivity of their neighbors' edges to be more modular?"
This relationship should only hold for connector hubs, not local hubs, as previous
studies suggest that connector hubs tune connectivity between communities and
maintain a modular structure1,36,38-40. We therefore examined connector hubs for which
their diversity facilitated modularity coefficients were positive. This analysis had two
aspects. First, do connector hubs increase modularity by tuning within community edge
strengths? Second, are connector hubs tuning the within community edge strengths of
their neighbors in order to increase global modularity?
In order to test the first aspect of the neural tuning mechanism -- if within community
edges are tuned by connector hubs in order to increase global modularity -- we used a
canonical division of nodes into communities (Figure 4d displays this division, see
Methods for link to division36,49). We assessed, for each edge in the network, how the
edge's weights related to modularity values (Q) across subjects (Figure 4a). Next, we
calculated how well each connector hub's participation coefficients correlate with each
edge's weights across subjects (Figure 4b). Higher Q values and higher connector hub
participation coefficients are associated with decreased connectivity between the visual,
sensory/motor hand, sensory/motor mouth, auditory, ventral attention, dorsal attention,
and cingulo-opercular communities. These communities were also more strongly
connected to fronto-parietal, default mode, salience, and sub-cortical communities in
networks with higher modularity values and higher connector hub participation
coefficients.
Given these observations, we sought to find the edges that mediate between connector
hubs' increased participation coefficients and modularity (Q), as these are the edges
that connector hubs likely tune in order to increase Q. Specifically, a mediation analysis
was performed for each connector hub, with an edge weight mediating the relationship
between the connector hub's participation coefficients and the Q indices of the networks
across subjects. An edge's mediation value of a connector hub's participation
coefficients and Q is the regression coefficient of the edge's weights by the connector
hub's participation coefficients across subjects multiplied by the regression coefficient of
Q indices by the edge's weights, controlling for the connector hub's participation
coefficients, across subjects. Each edge's mean mediation value across connector hubs
is shown in Figure 4c. We found that edges between the visual, sensory/motor hand,
sensory/motor mouth, auditory, ventral attention, dorsal attention, and cingulo-opercular
communities, as well as edges between those communities and the fronto-parietal,
default mode, and sub-cortical communities, mediate the relationship between
connector hubs' participation coefficients and Q indices. These results are consistent
with a mechanistic model in which diverse connector hubs tune connectivity to increase
segregation between sensory, motor, and attention systems, which increases the global
modularity of the network.
Next, we tested the second aspect of the network tuning mechanism -- if the relationship
between a connector hub's participation coefficients and Q indices is mediated primarily
by that connector hub's neighbors' edge pattern increasing Q. Neighbors were defined
based on edges present between the two nodes in a graph at a density of 0.15 (as it
was our densest cost explored). The mediation values calculated above each represent
an edge mediating between a node i's participation coefficients and Q values. Thus, for
each connector hub i, there is the set of arrays of absolute mediation values of node i's
neighbors' edges (n=263 for each neighbor j's array) and the set of arrays of the
absolute mediation values of node i's non-neighbors' edges (n=263 for each non-
neighbor j's array). Edges of node i in every array were ignored, as we were only
interested in how the participation coefficients of connector hub i modulate Q via the
mediation of j's connectivity to the rest of the network, not j's connectivity to connector
hub i (thus, n=264-1). If a connector hub is primarily modulating Q via the tuning of its
neighbors' edges, then the absolute mediation values in the neighbors' arrays should be
greater than the absolute values in the non-neighbors' arrays. The distribution of t-
values between the two sets of arrays for all connector hubs (neighbors versus non-
neighbors) is shown in Supplementary Figure 12; across tasks, the mediation values
were consistently and significantly higher for connector hubs' neighbors' edges than
non-neighbors edges. Moreover, these same t-values can be calculated for local hubs,
using the within community strength instead of the participation coefficient; thus, an
edge mediates between a local hub's within community strengths and Q. Across tasks,
connector hubs' neighbors' mediation t values were shown to be higher than local hubs'
neighbors' mediation t values with a two-tailed independent student's t-test (t(dof:1358):
3.892, p:0.0001, Cohen's d:0.219, CI:1.887,6.62), demonstrating that this result is
specific to connector hubs. All distributions were confirmed as normal (k2>100.0,
p<0.00001 for all tasks). We also performed an alternative analysis that confirmed these
relationships (see Methods, Supplementary Figure 13). These results suggest that each
connector hub, not local hub, tunes their neighbors' connectivity to be more modular. A
connector hub's high diversity facilitated modularity coefficient does not largely reflect
diffuse global connectivity changes. Instead, connector hubs are likely connected in a
way that allows them to directly tune the connectivity of their neighbors to be more
modular, thereby increasing global network modularity. Thus, the locality facilitated
modularity coefficients are likely a downstream effect of connector hub modulation.
Supporting this interpretation, we found that, when connector hubs have high
participation coefficients, local hubs have high within community strengths
(Supplementary Figure 14).
Figure 4 Connectivity between primary sensory, motor, dorsal attention, ventral attention, and cingulo-opercular
communities mediate the relationship between connector hubs and modularity. a, Each entry is the Pearson correlation
coefficient, r, across subjects (N=476), between modularity (Q) and that edge's weights. b, For each connector hub, the Pearson r
between the hub's participation coefficients and each edge's weights across subjects (N=476) was calculated. The matrix in b is the
mean of those matrices across connector hubs. c, To investigate the relationship between connector hubs' participation coefficients,
edge weights, and Q, a mediation analysis was performed for each connector hub, with an edge's weights mediating the relationship
between the connector hub's participation coefficients and Q indices (N=476). Each edge's mean mediation value between
connector hubs' participation coefficients and Q is shown. d, The anatomical locations of each node and community on the cortical
surface36,49.
In the series of analyses we report here, we explicitly and comprehensively tested a
mechanistic model by leveraging individual differences in connectivity and cognition in
humans. Specifically, a model of the diversity and locality of hubs, the modularity of the
network, and the network's connectivity was highly predictive of task performance and a
range of subject measures. Critically, the diversity and locality of nodes optimal for each
task were also similarly optimal for positive subject measures. Thus, it appears that
there is a hub and network structure that is generally optimal for cognitive processing.
We found evidence that diverse connector hubs preserve or increase the modularity of
brain networks. Moreover, diverse connector hubs tune the connectivity of their
neighbors to be more modular. Finally, we found that the diversity of connector hubs
simultaneously facilitated higher modularity and task performance. Thus, connector
hubs appear to contribute to the maintenance of an optimal modular architecture during
integrative cognition without greatly increasing the wiring cost or decreasing
modularity5,29,51. In sum, these data are consistent with a mechanistic model of hub
function, where connector hubs integrate information and subsequently tune their
neighbors' connectivity to be more modular, which increases the global modularity of
the network, allowing local hubs and nodes to perform segregated processing.
Across individuals, we found that diverse connector hubs increase modularity and task
performance, regardless of the task. In all seven tasks, the subjects with the most
diversely connected connector hubs also had the highest modularity and, in the four
tasks for which performance was measured, the highest task performance. Thus, we
propose that connector hubs are likely critical for integrating information and tuning their
neighbors' connectivity to be more modular, regardless of the task. Although connector
hubs are more active during tasks that require many communities1, as their functions
are likely more computationally demanding during these tasks, it is likely that every task
requires the functions of connector hubs, as supported by our finding that their diverse
connectivity predicts performance in all of the tasks analyzed here.
Our findings compliment many previous task-based fMRI studies that have identified
regions that are more active during a particular cognitive process. We have
demonstrated that, while different regions are more or less active in different tasks,
including connector hubs, the diverse connectivity and integration and tuning functions
of connector hubs are consistently required across different cognitive processes. Our
findings are also consistent with neuropsychological studies of patients with focal brain
lesions. It has been found that damage to connector hubs decreases modularity and
causes widespread cognitive deficits, while damage to local nodes does not decrease
modularity and causes more isolated deficits, such as hemiplegia, or aphasia26,27. While
connector hubs are not likely critical for only one specific cognitive process, their
functions and diverse connectivity are required to maintain a cognitively optimal modular
structure across cognitive processes. Thus, as we observed here, individual differences
in the diversity of connector hubs' connectivity is predictive of cognitive performance
across a range of very different tasks. Although diversely connected connector hubs are
critical for successful performance in many different tasks, any given task nevertheless
recruits very different cognitive and neural processes; each task likely engages
connectivity patterns that are specifically optimal to that task. Future analyses should
seek to understand both the general optimal connectivity patterns of connector hubs
found here and the connectivity patterns that are optimal to a single task, including if
and how these connectivity patterns interact.
Methods
Data and Preprocessing We used fMRI data from the Human Connectome Project48
S500 release. For the task-based fMRI data, Analysis of Functional NeuroImages
(AFNI) was used to preprocess the images52. The AFNI command 3dTproject was used,
passing the mean signal from the cerebral spinal fluid mask, the white matter mask, the
whole brain signal, and the motion parameters to the "-ort" options, which removes
these signals via linear regression. Within AFNI, the "-automask" option was used to
generate the masks. The "-passband 0.009 0.08" option, which removes frequencies
outside of 0.009 and 0.08, was used. Finally, the "-blur 6" option, which smooths the
images (inside the mask only) with a 6mm FWHM filter after the time series filtering.
Given the short length of the Emotion task (176 frames; Resting-State:1200, Social:
274, Relational:232, Motor:284, Language:316, Working Memory:405, Gambling:253) it
was not included in our analyses. For the fMRI data collected at rest, we used the
images that were previously preprocessed by the Human Connectome Project with ICA-
FIX. We also used the AFNI command 3dBandpass to further preprocess these images.
We used it to remove the mean whole brain signal and frequencies outside 0.009 and
0.08 (explicitly, "-ort whole_brain_signal.1D -band 0.009 0.08 -automask"). We did not
regress out stimulus or task effects from the time series of each node, because how
nodes' low frequency oscillations respond to stimulus or task effects is meaningful.
Moreover, other investigators have noted that task effect regression has minimal
effects53.
As subject head motion during fMRI can impact functional connectivity estimates and
has been shown to bias brain-task performance relationships54, performance prediction
analyses were executed with scrubbing (removing frames with high motion) executed
on frames with frame-wise displacement greater than 0.2 millimeters, including the
frame before and after the movement. Frame-wise displacement measures movement
of the head from one volume to the next, and was computed as the sum of the absolute
values of the differentiated rigid body realignment estimates (translation and rotation in
x, y, and z directions) at every time point with rotation values evaluated with a radius of
50 mm54. Frames were removed after all preprocessing was executed. Subjects with
more than 75 percent of frames removed were not analyzed. Moreover, we executed all
analyses after regressing out mean frame-wise displacement from the task performance
values (Supplementary Figure 2).
Graph Theory Analyses The Power atlas49 was used to define the 264 nodes in our
graph because it was the only atlas that met all of the following requirements: (1) Given
that the homogeneity of nodes in this atlas is high and they do not share physical
boundaries, it will not overestimate the local connectivity of regions, (2) it is the only
atlas that is defined based both on functional connectivity and studies of task activations
making it optimal for our current analyses, (3) it accurately divides nodes into
communities observed with other approaches (e.g., at the voxel level), and this division
has been used in many studies33,36,49,55. A canonical division of nodes into communities
aides in the interpretation and generalizability of our results. It can be found at:
http://www.nil.wustl.edu/labs/petersen/Resources_files/Consensus264.xls. Moreover,
we used this division to calculate within and between community edge weight changes
across subjects. (4) It has anatomical coverage of cortical, subcortical, and cerebellar
regions.
All graph theory analyses were executed with our own custom python code
(www.github.com/mb3152/brain_graphs) that uses the iGraph library. All analysis code
is also publicly available (github.com/mb3152/hcp_performance/). For each task (both
LR and RL encoding directions were used) and for each subject, the mean signal from
264 regions in the Power atlas was computed. The Pearson r between all pairs of
signals was computed to form a 264 by 264 matrix, which was then Fisher z
transformed. We chose Pearson r values to represent functional connectivity (i.e.,
edges) between nodes, for its simplicity in interpretation and ubiquity in human network
neuroscience56. However, more complex statistical measures could be employed,
including measures that attempt to estimate the directionality of each edge. The LR and
RL matrices were then averaged. The mean matrix was then thresholded, retaining
edge weights, at a range of costs (0.05 to 0.15 at 0.01 intervals), a common range and
interval in graph theory analyses1,27,33,49. The maximum spanning tree was calculated to
ensure all nodes had at least one edge. No negative correlations were included in our
analyses. The matrix was then normalized to sum to a common value across subjects,
and was used to represent the edges in the graph. Thus, all graphs had the same
number of edges and sum of edge weights.
For each cost, the InfoMap algorithm57 was run. While this method has been shown to
be highly accurate on benchmark networks with known community structures, it is still a
heuristic, as community detection is NP-hard58. While InfoMap does not explicitly
maximize Q, it has been shown to estimate community structure accurately in several
test cases59, rendering the Q value, the participation coefficients, and within community
strengths computed based on the community structure accurate and valid. Moreover, in
biological networks, InfoMap achieves Q values that are similar to algorithms that
maximize Q 60; in the current resting-state data, InfoMap Q values and Fast-Greedy Q61
values were correlated at Pearson r=0.87 (dof=474, p<0.001, 95% CI: 0.84, 0.89);
InfoMap Q values were found to be higher than Fast-Greedy Q values with a student's
independent t-test (t(dof:952):16.027, p:<0.001, Cohen's d:0.775, 95% CI:0.024,0.03).
InfoMap Q values and Louvain Q 62 values were correlated at Pearson r=0.98 (dof=474,
p<0.001, 95% CI: (0.97, 0.98)); Louvain Q values were higher than InfoMap Q values.
When comparing InfoMap Q values to the distribution that includes both Louvain and
Fast-Greedy, two algorithms that explicitly maximize Q, InfoMap Q values were shown
to be significantly higher with a student's independent t-test (t(dof:1429):5.304,
p:<0.001, Cohen's d:0.222, 95% CI:0.005,0.011). Regardless, we found that it detects a
community structure with Q values highly similar to other methods (Supplementary
𝛿8𝑐",𝑐#<.
And modularity (Q) can be written as:
Figure 15). Moreover, previous work has demonstrated the stability of community
detection and the participation coefficient across community detection methods5.
The participation coefficients, within community strengths, and Q were calculated at
each cost. Q is written analytically as follows. Consider a weighted and undirected
graph with n nodes and m edges represented by an adjacency matrix 𝐀 with elements
𝐀"#=edge weight between i and j.
Thus, the strength of a node is given by 𝑘"=3𝐀"#
#
𝑄= 12𝑚38𝐀"#−𝛾𝑝"#<
"=#
Here, 𝑝"# is the probability that nodes i and j are connected in a random null network
𝑃"#=𝑘"𝑘#2𝑚,
γ is the resolution parameter, and 𝑐" is the community to which node 𝑖 belongs to and
𝛿(𝛼,𝛽)=1 if 𝛼=𝛽 and 𝛿(𝛼,𝛽)=0 if 𝛼≠𝛽.
𝑃𝐶"= 1− 3J𝐾"L𝐾"MN
OP
LQR
Given a particular community assignment, the participation coefficient of each node can
be calculated. The participation coefficient (PC) of node i is defined as:
where 𝐾" is the sum of i 's edge weights, 𝐾"Lis the sum of i 's edge weights to community
s, and NM is the total number of communities. Thus, the participation coefficient is a
measure of how evenly distributed a node's edges are across communities. A node's
participation coefficient is maximal if it has an equal sum of edge weights to each
community in the network. A node's participation coefficient is 0 if all of its edges are to
a single community.
Finally, we calculate the within community strength value for each node as follows:
𝑧"=𝑘"−𝑘TLU
𝜎WXU
Where 𝑘" is the number of links of node 𝑖 to other nodes in its community 𝑠", 𝑘TLU is the
average of 𝑘 over all the nodes in 𝑠", and 𝜎WXU is the standard deviation of 𝑘 in 𝑠". Thus,
The within community strength measures how well-connected node 𝑖 is to other nodes
in the community relative to other nodes in the community.
Each subject's participation coefficient, within community strength, and Q were the
mean of those values across the range of costs. All analyses were executed and all
prediction models were fit separately for each task.
Tasks The following descriptions for each task have been adapted for brevity from the
Human Connectome Project Manual63.
Working Memory. The category specific representation task and the working memory
task are combined into a single task paradigm. Participants were presented with blocks
of trials that consisted of pictures of places, tools, faces and body parts (non-mutilated
parts of bodies with no "nudity"). Within each run, the 4 different stimulus types were
presented in separate blocks. Also, within each run, 1⁄2 of the blocks use a 2-back
working memory task and 1⁄2 use a 0-back working memory task (as a working memory
comparison). A 2.5 second cue indicates the task type (and target for 0-back) at the
start of the block. Each of the two runs contains 8 task blocks (10 trials of 2.5 seconds
each, for 25 seconds) and 4 fixation blocks (15 seconds). On each trial, the stimulus is
presented for 2 seconds, followed by a 500 ms inter-task interval (ITI).
Gambling. Participants play a card guessing game where they are asked to guess the
number on a mystery card (represented by a "?") in order to win or lose money.
Participants are told that potential card numbers range from 1-9 and to indicate if they
think the mystery card number is more or less than 5 by pressing one of two buttons on
the response box. Feedback is the number on the card (generated by the program as a
function of whether the trial was a reward, loss or neutral trial) and either: 1) a green up
arrow with "$1" for reward trials, 2) a red down arrow next to -$0.50 for loss trials; or 3)
the number 5 and a gray double headed arrow for neutral trials. The "?" is presented for
up to 1500 ms (if the participant responds before 1500 ms, a fixation cross is displayed
for the remaining time), following by feedback for 1000 ms. There is a 1000 ms ITI with
a "+" presented on the screen. The task is presented in blocks of 8 trials that are either
mostly reward (6 reward trials pseudo randomly interleaved with either 1 neutral and 1
loss trial, 2 neutral trials, or 2 loss trials) or mostly loss (6 loss trials pseudo- randomly
interleaved with either 1 neutral and 1 reward trial, 2 neutral trials, or 2 reward trials). In
each of the two runs, there are 2 mostly reward and 2 mostly loss blocks, interleaved
with 4 fixation blocks (15 seconds each).
Motor. Participants are presented with visual cues that ask them to either tap their left or
right fingers, or squeeze their left or right toes, or move their tongue to map motor
areas. Each block of a movement type lasted 12 seconds (10 movements), and is
preceded by a 3 second cue. In each of the two runs, there are 13 blocks, with 2 of
tongue movements, 4 of hand movements (2 right and 2 left), and 4 of foot movements
(2 right and 2 left). In addition, there are 3 15-second fixation blocks per run.
Language & Math. The task consists of two runs that each interleave 4 blocks of a story
task and 4 blocks of a math task. The lengths of the blocks vary (average of
approximately 30 seconds), but the task was designed so that the math task blocks
match the length of the story task blocks, with some additional math trials at the end of
the task to complete the 3.8 minute run as needed. The story blocks present
participants with brief auditory stories (5-9 sentences) adapted from Aesop's fables,
followed by a 2-alternative forced- choice question that asks participants about the topic
of the story. For example: "after a story about an eagle that saves a man who had done
him a favor, participants were asked, "Was that about revenge or reciprocity?" The math
task also presents trials auditorily and requires subjects to complete addition and
subtraction problems. The trials present subjects with a series of arithmetic operations
(e.g., "fourteen plus twelve"), followed by "equals" and then two choices (e.g., "twenty-
nine or twenty- six"). Participants push a button to select either the first or the second
answer. The tasks are adaptive to try to maintain a similar level of difficulty across
participants.
Social (Theory of Mind). Participants were presented with short video clips (20 seconds)
of objects (squares, circles, triangles) that either interacted in some way, or moved
randomly on the screen. After each video clip, participants judge whether the objects
had a mental interaction (an interaction that appears as if the shapes are taking into
account each other's feelings and thoughts), Not Sure, or No interaction (i.e., there is no
obvious interaction between the shapes and the movement appears random). Each of
the two task runs has 5 video blocks (2 Mental and 3 Random in one run, 3 Mental and
2 Random in the other run) and 5 fixation blocks (15 seconds each).
Relational. The stimuli are 6 different shapes filled with 1 of 6 different textures. In the
relational processing condition, participants are presented with 2 pairs of objects, with
one pair at the top of the screen and the other pair at the bottom of the screen. They are
told that they should first decide what dimension differs across the top pair of objects
(differed in shape or differed in texture) and then they should decide whether the bottom
pair of objects also differ along that same dimension (e.g., if the top pair differs in
shape, does the bottom pair also differ in shape). In the control matching condition,
participants are shown two objects at the top of the screen and one object at the bottom
of the screen, and a word in the middle of the screen (either "shape" or "texture"). They
are told to decide whether the bottom object matches either of the top two objects on
that dimension (e.g., if the word is "shape", is the bottom object the same shape as
either of the top two objects. For both conditions, the subject responds yes or no using
one button or another. For the relational condition, the stimuli are presented for 3500
ms, with a 500 ms ITI, and there are four trials per block. In the matching condition,
stimuli are presented for 2800 ms, with a 400 ms ITI, and there are 5 trials per block.
Each type of block (relational or matching) lasts a total of 18 seconds. In each of the two
runs of this task, there are 3 relational blocks, 3 matching blocks and 3 16-second
fixation blocks.
Performance measures. All performance measures were chosen a priori. In the
working memory task, we used the mean accuracy across all n-back conditions (face,
body, place, tool). In the relational task, we used mean accuracy across both the
matching and the relational conditions. For the language task, we took the maximum
difficulty level that the subject achieved across both the math and language conditions.
We did not use accuracy, because the task varies in difficulty based on how well the
subject is doing, making accuracy an inaccurate measure of performance for these
tasks. For the social task, given that almost all subjects correctly identified the social
interactions as social interactions, we used the percentage of correctly identified
random interactions.
Deep neural network model. A deep neural network is a supervised learning algorithm
that can learn a non-linear function for regression or classification. Unlike logistic
regression, there are one or more non-linear layers, called hidden layers, between the
input and the output layer. Thus, the model is trained to relate a set of input features to
outputs by learning weights between neurons across adjacent layers (Supplementary
Figure 1). Our implementation uses the sklearn python library. Explicitly, a prediction for
subject z is calculated as:
model = sklearn.neural_network.MLPRegressor(hidden_layer_sizes=(8,12,8,12))
model.fit(x[subjects!=z], y[subjects!=z])
prediction = model.predict(x([z])
where x is the set of features across subjects and y is the task performance across
subjects.
Analytic quality of diversity and locality facilitated modularity coefficients To
further understand diversity and locality facilitated modularity coefficients, we performed
an iterative split-half analysis. Specifically, we estimated the mean within community
strength or participation coefficient of each node in one half of subjects, and each
node's locality and diversity facilitated coefficient in the other half, testing 10,000
random splits of subjects. All relationships were reliably observed in every cognitive
state (Supplementary Figure 7). Next, we sought to determine if this relationship was a
necessary feature of the underlying mathematics, or whether it was a phenomenon
specific to the neurophysiology of brain networks. To address this question, we tested
four null model networks and observed that none of them exhibited a significant
relationship between mean participation coefficient and diversity facilitated modularity
coefficient (Supplementary Figure 8). As a third check, we assessed whether the
number of communities identified in the network was inadvertently biasing our results.
We observed that the number of communities in each network was negatively
correlated with the modularity value Q (Supplementary Figure 9). After regressing out
the number of communities in each network from the modularity value, we observed
that our findings remained qualitatively unchanged (Supplementary Figure 10). Finally,
we tested whether the relationships between variables of interest were linear (and
therefore appropriate to examine with Pearson r correlation coefficients), or nonlinear.
To address this question, we analyzed individual 1st, 2nd, and 3rd order curve fits of the
relationship between participation coefficients and modularity values. We observed that
many relationships were well-captured by a first order fit, with the connector hub's
maximal participation coefficients corresponding to maximal Q indices, with only a few
showing a more nonlinear relationship (Supplementary Figure 11).
Alternative analysis of connector hubs' tuning the connectivity of their neighbors
We executed an alternative analysis to test if connector hubs tune the connectivity of
their neighbors to be more modular. For each node i we calculated a matrix, where the
j-kth entry is the Pearson r correlation coefficient that captures how well the participation
coefficients of node i correlates with the edge weights between nodes j and k in the
network across subjects. These Pearson r values allowed us to test whether a node's
participation coefficients correlate positively with its neighbors' increased connectivity to
its own community and decreased connectivity to other communities. We subtracted the
sum of r values in the matrix corresponding to node i's participation coefficients and
node j's between community edge weights from the sum of r values in the matrix
corresponding to node i's participation coefficients and node j's within community edge
weights. Thus, this value measures how well the participation coefficients of node i are
correlated with the increased modular (within community) connectivity of node j. We
used the partition of nodes into communities that was created along with the nodes
themselves (Figure 6d)49. Edges between node i and node j were ignored in this
calculation, as the participation coefficients of node i is likely highly correlated with the
edge weights between node i and node j, and we were only interested in how the
participation coefficient of node i modulates node j's connectivity to the rest of the
network, not node j's connectivity to node i. Edges that were not positive on average
across subjects were not included in this analysis, as the interpretation of negative
edges in fMRI-based networks is not obvious (results were similar only including the top
25 percent of edges (Supplementary Figure 13). Correlations between the edge
strength between nodes i and j and the amount of modulation of j's modularity by i were
calculated such that the set of nodes i were either connector hubs or non-connector
hubs. A positive correlation means that a node is biased to modulate the connectivity of
its neighbors versus its non-neighbors to be more modular. In all cognitive states, these
correlations were only positive and significant (Pearson's r>0.17, p<0.001, Bonferroni
corrected (n tests = 7)) for connector hubs (Supplementary Figure 13) suggesting that
connector hubs tune the connectivity of their neighbors to be more modular.
To test if this relationship existed for local hubs, we calculated a similar matrix, where,
for each node i, the j-k th entry is the Pearson r value that captures how well the within
community strengths of node i correlate with the edge weight between nodes j and k in
the network across subjects. Correlations between the edge strength between nodes i
and j and the amount of modulation of j's modularity by the within community strength of
i were calculated such that nodes i were either local hubs or non-local hubs. None of
these correlations were robust (-0.1>r<0.1). These analyses add to our conclusion,
demonstrated in the Results, that connector hubs facilitate higher modularity by tuning
the connectivity of their neighbors to be more modular.
Statistical Methods
The number of subjects was determined by the number of subjects released by the
Human Connectome Project at the start of the analyses. As this dataset represented the
largest dataset of its kind at that time and the number of subjects is greater than many
similar analyses47, no power analysis was computed. Total N=Working Memory: 475,
Gambling: 473, Relational:458, Motor:475, Language & Math:472, Social:474, Resting
State: 476. However, as we only analyzed subjects with both Resting-State and the task
scans, N=Working Memory:473, Relational:457, Language & Math:471, Social:473. This
results in a unique N=476 across tasks, in that 476 different subjects had a resting state
scan and at least one task scan. As scrubbing (which removes frames with large head
motion) can cause too many frames to be removed from the time series, subjects with
less than 75 percent of remaining frames were not included in the analyses that
implemented scrubbing; thus, for analyses using scrubbed data, N=Working
Memory:351, Relational:335, Language & Math:348, Social:358.
All confidence intervals (CI) are reported with alpha=0.05. For Pearson r correlation
coefficients CIs, the interval of r values is given by Fisher transforming r to z, computing
the interval, and then Fisher reverse transforming the z intervals back to r intervals. For
t-tests, the confidence interval represents the largest and smallest differences in means
across the two distributions. For all t-tests, distributions were confirmed as normal
(p<0.001) or exhibiting no significant evidence as not normal (k2>0.0) using D'Agostino
and Pearson's omnibus test k2. All p values are two sided tests.
All p values that are part of a family of tests are Bonferroni corrected for multiple
comparisons. For example, we test if two tasks' hub and network structures are similarly
optimal for the same subject measures, testing across a large number of subject
measures. In this case, we applied a Bonferroni correction to the p-values to determine
whether the effect remained true for particular subject measures. Here, the number of
tests is equal to the number of subject measures, 47. Individual subject networks were
built independently for each task and task performance is different for each task. Thus,
these tests are not strictly in the same family. However, to be conservative, we still
Bonferroni corrected these p-values. In these cases, the family size is either 4 or 7,
depending on the number of tasks analyzed. Unless otherwise stated, all p values are
Bonferroni corrected.
Many statistical tests are calculated here without reported p values. For example,
Pearson r values are used to calculate functional connectivity. Here, only the r values
are of interest -- more precisely, individual differences in the r values across subjects,
and how these differences relate to individual differences in cognition. This treatment of
multiple comparisons in the context of functional connectivity and individual differences
in cognition is common and recommended47,64. We extend this notion to other analyses
here as well. For example, we use the Pearson correlation coefficient r to compare how
well different nodes' participation coefficients across subjects explain variance in
network modularity or task performance (the diversity facilitated modularity and
performance coefficients). In these cases, we relate these r-values to other measures,
and are only concerned with how these r-values explain another distributions' variance
(here, we find a positive correlation between these r-values and a node's mean
participation coefficient across subjects). We are not concerned with the statistical
significance any particular r-value as estimated by the p-value. We care about the
distribution of r-values, not the distribution of p-values, and we do not make any claims
about any single r-value. Thus, the p-values are neither reported nor corrected for
multiple comparisons. This is precisely how functional connectivity is treated
statistically.
Acknowledgements This work was supported by NIH Grant NS79698 and the National
Science Foundation Graduate Research Fellowship Program under Grant no. DGE
1106400 to MAB and MD. MAB would also like to acknowledge NIH T32 Ruth L.
Kirschstein Institutional National Research Service Award (5T32MH106442-02). BTTY
was also supported Singapore MOE Tier 2 (MOE2014-T2-2-016), NUS Strategic
Research (DPRT/944/09/14), NUS SOM Aspiration Fund (R185000271720), Singapore
NMRC (CBRG/0088/2015), NUS YIA and the Singapore National Research Foundation
(NRF) Fellowship (Class of 2017). DSB would also like to acknowledge support from the
John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the
Army Research Laboratory and the Army Research Office through contract numbers
W911NF-10-2-0022 and W911NF-14-1-0679, the National Institute of Health (2-R01-
DC-009209-11,1R01HD086888-01, R01-MH107235, R01-MH107703, and R21-M MH-
106799), the Office of Naval Research, and the National Science Foundation (BCS-
1441502, PHY-1554488, and BCS-1631550). The funders had no role in study design,
data collection and analysis, decision to publish, or preparation of the manuscript.
Data availability
Data were provided by the Human Connectome Project, WU-Minn Consortium
(Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded
by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience
Research and by the McDonnell Center for Systems Neuroscience at Washington
University. The content is solely the responsibility of the authors and does not
necessarily represent the official views of any of the funding agencies. All analyses
were executed in accordance with the authors' institutions' relevant ethical regulations
as well as the WU-Minn HCP Consortium Open Access Data Use Terms. Informed
consent was obtained from all participants.
Competing Interests
The authors declare no competing interests.
Code availability
All graph theory analyses were executed with our own custom python
code (www.github.com/mb3152/brain_graphs) that uses the iGraph library. All analysis
code is also publicly available (github.com/mb3152/hcp_performance/).
Author Contributions M.A.B. conceived the analyses. M.A.B., B.T.T.Y., D.S.B., and
M.D. collaboratively designed the analyses. M.A.B. executed the analyses. M.A.B.,
B.T.T.Y., D.S.B., and M.D. collaboratively wrote the paper.
Supplementary Figure 1. Workflow of the deep neural network features and model construction. a, For each node, we
measured the diversity facilitated performance coefficient. For example, the Pearson r is calculated between node 9's participation
coefficients across subjects (black outline) and task performance across subjects (black outline), resulting in a diversity facilitated
performance coefficient of 0.23 (black outline) for that node. b, The optimality of a subject's participation coefficients for task
performance is measured by, for example, calculating the Pearson r between subject 9's participation coefficients (black outline) and
the diversity facilitated performance coefficients (black outline), resulting in a diversity feature for the model of 0.43 (black outline).
The same procedures in a and b are also executed using within community strengths and edge weights instead of participation
coefficients. Modularity (Q) values are also used as features. These four features are derived separately from resting-state and the
task network for which task performance is being predicted, resulting in eight features. All calculations in panels a and b are
calculated without data from the subject for which the prediction is being made. These eight features are then used in a deep neural
network model (c) to learn the relationship between the features and task performance (again, without data from the subject for
which the prediction is being made) by adjusting the weights between nodes in adjacent layers. The features are then calculated for
the left-out subject (e.g., Pearson r between the left-out subject's participation coefficients and the previously calculated diversity
facilitated performance coefficients that did not include data from the left-out subject) and are used in the deep neural network
model to generate a prediction for the left-out subject.
Supplementary Figure 2 Hub diversity and locality, modularity, and network connectivity predicts task performance. a,
predictive model of task performance (see Methods for performance measures), as shown in Figure 2a. b, Predictive model of task
performance without using network connectivity (i.e., only hub diversity and locality and modularity). N=Working Memory: 473,
Relational: 457,Language & Math: 471, Social: 473). Each dot represents a subject's prediction. Shaded areas represent 95 percent
confidence intervals. c,d, As in a,b, except the frames with high motion were removed from the time series before the network of
that individual was constructed (i.e., scrubbing), N=Working Memory: 351, Relational: 335, Language & Math: 348, Social: 358, as
subjects with less than 75 percent of frames after scrubbing were not analyzed. e,f, As in a,b, except mean frame wise
displacement was regressed out from the task performance values, N=Working Memory: 473, Relational: 457,Language & Math:
471, Social: 473). All p values are Bonferroni corrected (n tests=4).
Supplementary Figure 3. Feature correspondence between tasks and subject measures. To test if certain hub and network
structures that are optimal to each task (Language & Math, Relational, Social, or Working Memory) are also optimal for other subject
measures, we measured the feature correspondence between subjects' features in the task model with the subjects' features in the
same model that was fit to a given subject measure (e.g., Delayed Discounting) instead of performance in the task. The feature
correspondence is shown for each task; high feature correspondence means a similar hub and network structure is optimal for the
task and subject measure. Colors represent z-scored (within column) feature correspondence values. Results significant at p<1-e3
uncorrected, p<0.05 Bonferroni corrected (N tests= 47) are marked with an asterisk. N subjects = Working Memory: 473, Relational:
457,Language & Math: 471, Social: 473).
Supplementary Figure 4 Prediction of subject measures by the model of hub diversity and locality, modularity, and network
connectivity. As in Figure 1, we predicted each subject measure using a model of hub diversity and locality, modularity, and
network connectivity. This was executed for networks constructed based on the data from each task (columns). For each task and
subject measure, the Pearson r between the real subject measure values and the predicted subject measure values are shown.
Colors represent the z-scored (within each task) Pearson r. Predictions that are significant (p<1-e3 uncorrected, p<0.05 after
Bonferroni correction, N tests = 47) are marked with an asterisk. N= Working Memory: 473, Relational: 457, Language & Math: 471,
Social: 473).
Supplementary Figure 5 Analyzing individual differences of hubs' diversity and locality and modularity. When connector
hubs are connected to many communities (as indicated by high participation coefficients), they can integrate information and tune
connectivity optimally, allowing other regions to perform more modular local processing. Thus, across subjects, we predicted that
the increased participation coefficients (i.e., diversity) of connector hubs would be correlated with preservations or increases in the
modularity (Q) of the network. This prediction is illustrated above (a). Each graph represents the network structure of an individual
subject. A single connector hub, shown in red, is identified based on its high mean participation coefficient across individuals;
subject-level participation coefficient values are shown inside the red nodes. Local neighborhoods or communities are shown in
green, blue, and pink, and subject-level values of modularity(Q) are shown above each graph. Across subjects, we predict that
higher participation coefficients of connector hubs will occur in more modular networks in which the neighbors of connector hubs
(pink, dashed-outline) display more local connectivity. b, To test this prediction, for each node we measured the diversity facilitated
modularity coefficient; for example, the Pearson r is calculated between node 9's participation coefficients across subjects (black
outline) and modularity (Q) across subjects (black outline), resulting in a diversity facilitated modularity coefficient of 0.27 (black
outline).
Supplementary Figure 6 Connector and local hubs facilitate network modularity and task performance. a, correlations
between a node's mean participation coefficient (across subjects) and the node's diversity facilitated modularity coefficient (the
Pearson r correlation coefficient between that node's participation coefficients and modularity values (Q) across subjects).
Correlations were calculated for each cognitive state. Each dot represents a node in the brain's functional network. Shaded areas
represent 95 percent confidence intervals. In all cognitive states, there was a significant positive correlation between a node's mean
participation coefficient and the node's diversity facilitated modularity coefficient. This demonstrates that connector hubs facilitate
increased modularity. b, correlations between a node's mean within community strength (across subjects) and the node's locality
facilitated modularity coefficient (Pearson r correlation coefficient between that node's within community strengths and modularity
values (Q) across subjects). In all cognitive states, there was a significant positive correlation between a node's within community
strength and the node's locality facilitated modularity coefficient. This demonstrates that local hubs facilitate increased modularity. c,
correlations between a node's mean participation coefficient (across subjects) and the node's diversity facilitated performance
coefficient (the Pearson r correlation coefficient between that node's participation coefficients and task performance across
subjects). Correlations were calculated for each cognitive state. Each dot represents a node in the brain's functional network.
Shaded areas represent 95 percent confidence intervals. In all cognitive states, there was a significant positive correlation between
a node's mean participation coefficient and the node's diversity-facilitated performance coefficient. This demonstrates that connector
hubs facilitate increased task performance. d, correlations between a node's mean within community strength (across subjects) and
the node's locality facilitated performance coefficient (Pearson r correlation coefficient between that node's within community
strengths and task performance across subjects). In all cognitive states (except Relational, Bonferroni (n tests=4) p=0.081,
uncorrected p=0.02), there was a significant positive correlation between a node's within community strength and the node's locality
facilitated performance coefficient. Bonferroni corrected p values are shown (n tests=7(a,b), and 4(c,d). N=264, the number of nodes
in the graph. This demonstrates that local hubs facilitate increased task performance.
Supplementary Figure 7 Cross validation of diversity and locality facilitated modularity coefficients. Connector hubs and
local hubs are defined by a higher participation coefficient and within community strength, respectively, across subjects on average.
To avoid potential dependencies, we estimated the mean within community strength or participation coefficient of each node in one
half of subjects, and the correlation between a nodes' within community strength or participation coefficients and Q indices -- the
node's locality and diversity facilitated modularity coefficients -- in the other half, testing 10,000 split of subjects. On average,
correlations were still significant in every state. a, The distribution of Pearson r values between a node's mean participation
coefficient (defined in half the subjects) and the diversity facilitated modularity coefficient (defined in the other half of the subjects).
b, The distribution of Pearson r values between a node's mean within community strength (defined in half the subjects) and the
locality facilitated modularity coefficient (defined in the other half of the subjects). N=264, the number of nodes in the graph, for each
r value calculation, (dof=262). The mean and quartiles are marked in each violin.
Supplementary Figure 8 Null Models. We tested four null models of diversity facilitated modularity coefficients. (1) Random Edges,
Real Community utilizes the true partition of nodes into communities that was uncovered by the application of community detection
to each subject's intact resting-state graph (thresholded at a weighted graph density of 0.10, chosen based on this cost being the
median cost from our original analyses), but we randomly permuted the edges uniformly after the partition was identified. (2)
Random Community, Real Edges utilizes the partition of nodes into communities that was uncovered by the application of
community detection to each subject's intact resting-state graph (thresholded at a weighted graph density of 0.10), but we then
randomly permuted the assignment of nodes to communities (this retains the same number of communities and sizes of
communities as the true partition). The edges remain in their true locations. (3) Random Edges, Clustered permutes the edges in
each subject's resting-state graph uniformly at random. We then applied community detection to this permuted graph to identify a
partition of nodes into communities, and then calculated the participation coefficient of each node based on that partition. We used a
cost of 0.05, as denser random graphs result in just one community, which results in trivial participation coefficients of 0. For each
instantiation (n=100) of each of these three models, we generated a graph for each subject using the subject's original graph and
that particular null model. Finally, (4) we generated Wattz-Strogatz small world graphs. Each graph had 264 nodes with each node
initially connected to its 7 neighbors in the lattice. We set the rewiring probability to 0.25. This results in Q values of roughly 0.40 and
a binary density of roughly 0.05. 100 graphs were generated for each instantiation(n=100) of the model. For each null model, we
then calculated the diversity facilitated modularity coefficient. Violin plots are shown for the distribution of the correlation between
each node's mean participation coefficient and each node's diversity facilitated modularity coefficient across 100 instantiations of
each model. For comparison, the distribution from the original analysis across tasks (Real Edges, Real Community) is also shown.
N=264, the number of nodes in the graph, dof=262.
Supplementary Figure 9 Participation coefficients, the number of communities in the network, and modularity (Q). a,
Negative correlation between the number of communities in subjects' graphs (averaged across graph densities in the range 0.05-
0.15) and the Q value of the graphs (also averaged across graph densities in the range 0.05-0.15). b, We performed a mediation
analysis between each node's participation coefficients, the number of communities in the graphs, and the Q values of the graphs,
with the number of communities being the mediator. Mediation values are plotted for each node on the x-axis. The y-axis is the
Pearson r between each node's participation coefficients and Q indices across subjects' graphs -- the diversity facilitated modularity
coefficient. Shaded areas represent a 95 percent confidence interval. N=264, the number of nodes in the graph, dof=262.
Supplementary Figure 10 Diversity and locality facilitated modularity coefficients controlling for the number of
communities. a, Correlations between a node's mean participation coefficient (across subjects) and the node's diversity facilitated
modularity coefficient (the Pearson r correlation coefficient between that node's participation coefficients and modularity values (Q)
across subjects after regressing out the number of communities in the network from Q). Correlations were calculated for each
cognitive state. Each dot represents a node in the brain's functional network. Shaded areas represent 95 percent confidence
intervals. In all cognitive states, there was a significant positive correlation between a node's mean participation coefficient and the
node's diversity-facilitated modularity coefficient. b, Correlations between a node's mean within community strength (across
subjects) and the node's locality facilitated modularity coefficient (Pearson r correlation coefficient between that node's within
community strength and modularity values (Q) across subjects after regressing out the number of communities in the network from
Q. In all cognitive states (except Resting state (uncorrected p=0.0132, Bonferroni corrected p=0.093), there was a significant
positive correlation between a node's within community strength and the node's locality facilitated modularity coefficient. All p values
are Bonferroni corrected (n tests=7). N=264, the number of nodes in the graph. N subjects=Working Memory:475, Gambling:473,
Relational: 458, Motor:475, Language:472, Social: 474, Rest: 476.
Supplementary Figure 11 Relationships between individual connector hubs' participation coefficients and Q. To test if the
relationship between a connector hub's participation coefficients and Q -- the connector hubs' diversity facilitated modularity
coefficient -- was linear, we fit three regression models to individual connector hubs' participation coefficients and Q values across
subjects. In each plot, each line is the relationship between a single connector hub's participation coefficients and Q values across
subjects. Only nodes with positive Pearson r values are shown. a, Locally weighted scatter-plot smoother fit. b, 2nd order fit. c, 3rd
order fit. The relation between many connector hubs' participation coefficients and Q indices across subjects is well captured by a
first order fit, with the hub's maximal participation coefficients (0.7; the mathematically upper limit is 0.99T) corresponding to maximal
Q. However, for example, some connector hubs' participation coefficients correspond to the maximal Q at a participation coefficient
of 0.4-0.5, and then Q decreases at higher participation coefficients of that connector hub. Shaded areas represent 95 percent
confidence intervals. N=476.
Supplementary Figure 12 Connector hubs' mediation of neighboring nodes' edges. To investigate if the relationship between a
connector hub's participation coefficient and Q is mediated primarily by that connector hub's neighbors' edge pattern increasing Q
(per our prediction), we executed, for each connector hub node (i), a t-test between the absolute mediation values of node (i)'s
neighbors' edges versus the absolute mediation values of node (i)'s non-neighbors' edges (neighbors were defined based on edges
present between the two nodes in a graph at a density of 0.15, which was used because it is the densest density we utilized in our
analyses; neighbors' and non-neighbors' edges connecting to the node (i) were ignored). Mediation values were based on the edge
mediating between node (i)'s participation coefficients and modularity values (Q); the distributions of mediation values was tested for
normality using D'Agostino and Pearson's omnibus test k2. All distributions were confirmed as normal (k2>100.0, p<0.001 for all
tasks). The distribution of t-values for connector hubs is shown for each task. The mean quartiles are marked. In general, connector
hubs showed higher mediation values with edges of its neighbors compared to the edges of its non-neighbors (i.e., t>0). Moreover,
these t-values were significantly higher than the same t-values calculated with local hubs (t(dof:1358): 3.892, p:0.0001, Cohen's
d:0.219, CI:1.887,6.62).
Supplementary Figure 13 Connector hubs' relationships with individual edges' weights. a,b, For each task, for each pair of
nodes i and j, we calculated the correlation between (x) how strongly node i increased the within community edge strength of node j
(the sum of Pearson r values between the participation coefficients of node i and the within community edges of node j minus the
sum of Pearson r values between the participation coefficients of node i and the between community edges of node j) with (y) the
average connectivity edge weight between node i and node j. a Includes all positive edges, while b includes the strongest 25
percent of edges. Bonferroni corrected p values are shown in all plots (n tests = 7). These results show that connector hubs are
predominately tuning the connectivity of their neighbors. Shaded areas represent 95 percent confidence intervals. N=dof+2 for each
panel.
Supplementary Figure 14 Connector hubs and local hubs are interrelated. For each task, across subjects, the correlation
between the subject's connector hubs' mean participation coefficient and the subject's local hubs' mean within community strengths.
In every task, if a subject's connector hubs were diverse, local hubs were local. Here, connector and local hubs were defined based
on nodes for which their diversity and locality (respectively) facilitated modularity coefficients were positive. For this calculation,
participation coefficients were z-scored within each subject, as within community strengths are z-scored within each subject.
Bonferroni corrected p values are reported in all plots(n tests = 7). Shaded areas represent 95 percent confidence intervals.
N=Working Memory:475, Gambling:473, Relational: 458, Motor:475, Language:472, Social: 474, Rest: 476.
Supplementary Figure 15 Modularity quality indices from three community detection algorithms. We used Q indices
repeatedly in our analyses. However, the community detection algorithm that we utilized, InfoMap, does not explicitly maximize Q.
To see if this could potentially impact our analyses, we compared Q values from InfoMap to two popular algorithms, Fast-Greedy
and Louvain, that explicitly maximize Q. The mean Q value, as in our analyses, was taken across costs of 0.05 to 0.15. a,
Distribution, across subjects, of Q values of each algorithm. b, Correlation between Q values across algorithms. Shaded areas
represent 95 percent confidence intervals. In all panels, N=476, dof=474.
2.
3. Guimerà, R., Guimera, R., Amaral, L. A. N. & Amaral, L. Functional cartography of
4. Meunier, D., Lambiotte, R. & Bullmore, E. T. Modular and hierarchically modular
5.
6.
Bertolero, M. A., Yeo, B. T. T., & D'Esposito, M. The modular and integrative
functional architecture of the human brain. Proc. Natl. Acad. Sci. U.S.A. 112,
E6798 -- 807 (2015).
Bertolero, M. A., Yeo, B. T. T., & D'Esposito, M. The diverse club. Nature
Communications 8, 1277 (2017).
complex metabolic networks. Nature 433, 895 -- 900 (2005).
organization of brain networks. Front. Neurosci. 4, 200 (2010).
Bertolero, M. A., Yeo, B. T. T., & D'Esposito, M. The diverse club. Nature
Communications 8, 1 -- 10 (2017).
Bassett, D. S. et al. Efficient Physical Embedding of Topologically Complex
Information Processing Networks in Brains and Computer Circuits. PLoS Comput
Biol 6, e1000748 (2010).
Laughlin, S. B., de Ruyter van Steveninck, R. R. & Anderson, J. C. The metabolic
cost of neural information. Nature Neuroscience 1, 36 -- 41 (1998).
Lord, L.-D., Expert, P., Huckins, J. F. & Turkheimer, F. E. Cerebral Energy
Metabolism and the Brain's Functional Network Architecture: An Integrative
Review. Journal of Cerebral Blood Flow & Metabolism 33, 1347 -- 1354 (2013).
Raichle, M. E. & Gusnard, D. A. Appraising the brain's energy budget.
Proceedings of the National Academy of Sciences 99, 10237 -- 10239 (2002).
Neuroscience 32, 356 -- 371 (2012).
development. Proceedings of the National Academy of Sciences 111, 13010 --
13015 (2014).
12. Krienen, F. M., Yeo, B. T. T., Ge, T., Buckner, R. L. & Sherwood, C. C.
Transcriptional profiles of supragranular-enriched genes associate with
corticocortical network architecture in the human brain. Proc. Natl. Acad. Sci.
U.S.A. 113, E469 -- 78 (2016).
13. Hawrylycz, M. et al. Canonical genetic signatures of the adult human brain.
Nature Neuroscience 18, 1832 -- 1844 (2015).
14. Wagner, G. P. & Zhang, J. The pleiotropic structure of the genotype-phenotype
map: the evolvability of complex organisms. Nature Reviews Genetics 12, 204 --
213 (2011).
Biol. Sci. 280, 1 -- 9 (2013).
modularity and network motifs. 102, 13773 -- 13778 (2005).
doi:10.1007/978-1-4899-0718-9_31
15. Clune, J., Mouret, J.-B. & Lipson, H. The evolutionary origins of modularity. Proc.
16. Kashtan, N., Kashtan, N., Alon, U. & Alon, U. Spontaneous evolution of
17. Simon, H. A. in Facets of Systems Science 457 -- 476 (Springer US, 1991).
18. Fodor, J. & Fodor, J. The Modularity of Mind. (MIT Press, 1983).
19. Coltheart, M. Modularity and cognition. Trends in Cognitive Sciences 3, 115 -- 120
9.
10. Harris, J. J. & Attwell, D. The Energetics of CNS White Matter. Journal of
11. Kuzawa, C. W. et al. Metabolic costs and evolutionary implications of human brain
1.
7.
8.
(1999).
20. Robinson, P. A., Henderson, J. A., Matar, E., Riley, P. & Gray, R. T. Dynamical
23. Stevens, A. A., Tappon, S. C., Garg, A. & Fair, D. A. Functional brain network
28.
29. Rubinov, M., Ypma, R. J. F., Watson, C. & Bullmore, E. T. Wiring cost and
Reconnection and Stability Constraints on Cortical Network Architecture. Physical
Review Letters 103, 108104 (2009).
21. Clune, J., Mouret, J.-B. & Lipson, H. The evolutionary origins of modularity. Proc.
Biol. Sci. 280, 20122863 -- 20122863 (2013).
22. Tosh, C. R., Tosh, C. R., McNally, L. & McNally, L. The relative efficiency of
modular and non-modular networks of different size. Proc. Biol. Sci. 282,
20142568 -- 20142568 (2015).
modularity captures inter- and intra-individual variation in working memory
capacity. PLOS ONE 7, e30468 (2012).
24. Arnemann, K. L. et al. Functional brain network modularity predicts response to
cognitive training after brain injury. Neurology 84, 1568 -- 1574 (2015).
25. Modular Brain Network Organization Predicts Response to Cognitive Training in
Older Adults. PLoS ONE (2016).
26. Warren, D. E. et al. Network measures predict neuropsychological outcome after
brain injury. Proc. Natl. Acad. Sci. U.S.A. 111, 14247 -- 14252 (2014).
27. Gratton, C., Nomura, E. M., Pérez, F. & D'Esposito, M. Focal Brain Lesions to
Critical Locations Cause Widespread Disruption of the Modular Organization of
the Brain. Journal of Cognitive Neuroscience 24, 1275 -- 1285 (2012).
van den Heuvel, M. P. & Sporns, O. Network hubs in the human brain. Trends in
Cognitive Sciences 17, 683 -- 696 (2013).
topological participation of the mouse brain connectome. Proc. Natl. Acad. Sci.
U.S.A. 112, 10032 -- 10037 (2015).
30. Ferreira, F. R. M., Nogueira, M. I. & Defelipe, J. The influence of James and
Darwin on Cajal and his research into the neuron theory and evolution of the
nervous system. Front Neuroanat 8, 1 (2014).
transportation network: Anomalous centrality, community structure, and cities'
global roles. Proceedings of the National Academy of Sciences 102, 7794 -- 7799
(2005).
32. Guimerà, R., Sales-Pardo, M. & Amaral, L. A. N. Classes of complex networks
defined by role-to-role connectivity profiles. Nature Physics 3, 63 -- 69 (2007).
33. Power, J. D., Schlaggar, B. L., Lessov-Schlaggar, C. N. & Petersen, S. E.
Evidence for Hubs in Human Functional Brain Networks. 79, 798 -- 813 (2013).
van den Heuvel, M. P. & Sporns, O. An Anatomical Substrate for Integration
34.
among Functional Networks in Human Cortex. 33, 14489 -- 14500 (2013).
35. Scholtens, L. H., Schmidt, R., de Reus, M. A. & van den Heuvel, M. P. Linking
Macroscale Graph Analytical Organization to Microscale Neuroarchitectonics in
the Macaque Connectome. 34, 12192 -- 12205 (2014).
36. Cole, M. W. et al. Multi-task connectivity reveals flexible hubs for adaptive task
control. Nature Neuroscience 16, 1348 -- 1355 (2013).
31. Guimera, R., Mossa, S., Turtschi, A. & Amaral, L. A. N. The worldwide air
43. Sala-Llonch, R. et al. Brain connectivity during resting state and subsequent
44. Sadaghiani, S., Poline, J.-B., Kleinschmidt, A. & D'Esposito, M. Ongoing
37. Yeo, B. T. T. et al. Functional Specialization and Flexibility in Human Association
Cortex. Cereb Cortex 25, 3654 -- 3672 (2015).
38. Bassett, D. S., Yang, M., Wymbs, N. F. & Grafton, S. T. Learning-induced
autonomy of sensorimotor systems. Nature Neuroscience 18, 744 -- 751 (2015).
39. Spadone, S. et al. Dynamic reorganization of human resting-state networks during
visuospatial attention. Proc. Natl. Acad. Sci. U.S.A. 112, 8112 -- 8117 (2015).
40. Gratton, C., Laumann, T. O., Gordon, E. M., Adeyemo, B. & Petersen, S. E.
Evidence for Two Independent Factors that Modify Brain Networks to Meet Task
Goals. Cell Reports 17, 1276 -- 1288 (2016).
41. Yamashita, M., Kawato, M. & Imamizu, H. Predicting learning plateau of working
memory from whole-brain intrinsic network connectivity patterns. Scientific
Reports 5, 7622 (2015).
42. Baldassarre, A. et al. Individual variability in functional connectivity predicts
performance of a perceptual task. Proc. Natl. Acad. Sci. U.S.A. 109, 3516 -- 3521
(2012).
working memory task predicts behavioural performance. Cortex 48, 1187 -- 1196
(2012).
dynamics in large-scale functional connectivity predict perception. Proc. Natl.
Acad. Sci. U.S.A. 112, 8463 -- 8468 (2015).
Network States during Cognitive Task Performance. Neuron 92, 544 -- 554 (2016).
abstinence effects on craving and cognitive function. JAMA Psychiatry 71, 523 --
530 (2014).
47. Finn, E. S. et al. Functional connectome fingerprinting: identifying individuals
using patterns of brain connectivity. Nature Neuroscience 18, 1664 -- 1671 (2015).
48. Van Essen, D. C. et al. The WU-Minn Human Connectome Project: an overview.
NeuroImage 80, 62 -- 79 (2013).
49. Power, J. D. et al. Functional Network Organization of the Human Brain. Neuron
72, 665 -- 678 (2011).
50. Smith, S. M. et al. A positive-negative mode of population covariation links brain
connectivity, demographics and behavior. Nature Neuroscience 18, 1565 -- 1567
(2015).
51. Bullmore, E. & Sporns, O. The economy of brain network organization. Nat Rev
Neurosci 74, 47 -- 14 (2012).
52. Cox, R. W. AFNI: software for analysis and visualization of functional magnetic
resonance neuroimages. Comput. Biomed. Res. 29, 162 -- 173 (1996).
53. Cole, M. W., Bassett, D. S., Power, J. D., Braver, T. S. & Petersen, S. E. Intrinsic
and Task-Evoked Network Architectures of the Human Brain. 83, 238 -- 251 (2014).
54. Siegel, J. S., Mitra, A., Laumann, T. O. & Seitzman, B. A. Data quality influences
observed links between functional connectivity and behavior. Cerebral … (2016).
45. Shine, J. M. et al. The Dynamics of Functional Brain Networks: Integrated
46. Lerman, C. et al. Large-scale brain network coupling predicts acute nicotine
55. Gu, S., Satterthwaite, T. D. & Medaglia, J. D. Emergence of system roles in
56. Zalesky, A., Fornito, A. & Bullmore, E. On the use of correlation as a measure of
57. Rosvall, M., Rosvall, M., Bergstrom, C. T. & Bergstrom, C. T. Maps of random
58. Brandes, U., Delling, D. & Gaertler, M. On modularity clustering. IEEE
59. Lancichinetti, A. & Fortunato, S. Community detection algorithms: A comparative
60. Berenstein, A. J., Piñero, J., Furlong, L. I. & Chernomoretz, A. Mining the Modular
61. Clauset, A., Newman, M. & Moore, C. Finding community structure in very large
62. Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. Fast unfolding of
normative neurodevelopment. in (2015).
network connectivity. NeuroImage 60, 2096 -- 2106 (2012).
walks on complex networks reveal community structure. 105, 1118 -- 1123 (2008).
transactions on … (2008).
analysis. Phys. Rev. E 80, 056117 -- 11 (2009).
Structure of Protein Interaction Networks. PLOS ONE 10, e0122477 (2015).
networks. Phys. Rev. E (2004).
communities in large networks. Journal of Statistical Mechanics: Theory and
Experiment 2008, P10008 (2008).
https://www.humanconnectome.org/storage/app/media/documentation/s1200/HC
P_S1200_Release_Reference_Manual.pdf
64. Shen, X. et al. Using connectome-based predictive modeling to predict individual
behavior from brain connectivity. Nat Protoc 12, 506 -- 518 (2017).
63. WU-Minn HCP 1200 Subjects Data Release Reference Manual. 1 -- 213 (2018).
|
1712.02142 | 1 | 1712 | 2017-12-06T11:46:02 | Maps of Visual Importance | [
"q-bio.NC"
] | The importance of an element in a visual stimulus is commonly associated with the fixations during a free-viewing task. We argue that fixations are not always correlated with attention or awareness of visual objects. We suggest to filter the fixations recorded during exploration of the image based on the fixations recorded during recalling the image against a neutral background. This idea exploits that eye movements are a spatial index into the memory of a visual stimulus. We perform an experiment in which we record the eye movements of 30 observers during the presentation and recollection of 100 images. The locations of fixations during recall are only qualitatively related to the fixations during exploration. We develop a deformation mapping technique to align the fixations from recall with the fixation during exploration. This allows filtering the fixations based on proximity and a threshold on proximity provides a convenient slider to control the amount of filtering. Analyzing the spatial histograms resulting from the filtering procedure as well as the set of removed fixations shows that certain types of scene elements, which could be considered irrelevant, are removed. In this sense, they provide a measure of importance of visual elements for human observers. | q-bio.NC | q-bio |
Maps of Visual Importance
Xi Wanga,∗, Marc Alexaa
aTechnische Universitat Berlin, Germany
Abstract
The importance of an element in a visual stimulus is commonly associated with the
fixations during a free-viewing task. We argue that fixations are not always correlated
with attention or awareness of visual objects. We suggest to filter the fixations recorded
during exploration of the image based on the fixations recorded during recalling the
image against a neutral background. This idea exploits that eye movements are a spatial
index into the memory of a visual stimulus. We perform an experiment in which we
record the eye movements of 30 observers during the presentation and recollection of
100 images. The locations of fixations during recall are only qualitatively related to
the fixations during exploration. We develop a deformation mapping technique to align
the fixations from recall with the fixation during exploration. This allows filtering the
fixations based on proximity and a threshold on proximity provides a convenient slider
to control the amount of filtering. Analyzing the spatial histograms resulting from the
filtering procedure as well as the set of removed fixations shows that certain types of
scene elements, which could be considered irrelevant, are removed. In this sense, they
provide a measure of importance of visual elements for human observers.
Keywords: Visual importance, Mental imagery, Eye tracking
1. Introduction
The relative importance of visual elements presented on a display is useful in a
variety of applications. In the fields of visualization and human-computer interaction,
it is often a design goal to make specific elements clearly visible. Consider for example
∗Corresponding author at Computer Graphics, Technische Universitat Berlin, Germany.
Email address: [email protected] (Xi Wang)
1
the visualization of medical imaging data: it is important that anomalies are recognized
by the physician inspecting the data. Likewise, in computer graphics or image and
video coding, it is of interest what visual objects are important or unimportant in order
to allocate resources such as computation time (for example, for rendering 3D objects
in an augmented reality scenario) or bandwidth (as in video coding).
Note that in these applications we are not just interested in whether visual elements
have been processed by an observer, but we need a map of the relative importance of
regions of the stimulus. The common approach for creating such maps is based on eye
tracking and accumulating fixations from several observers (Bruce and Tsotsos, 2005;
Judd et al., 2009, 2012; Coutrot and Guyader, 2014; Xu et al., 2014; Borji and Itti,
2015) The underlying reasoning is that most visual information is processed during
fixating (Matin, 1974), and saccade target selection is likely optimized to attend to the
most important visual elements first (Itti and Koch, 2000; Parkhurst et al., 2002). A va-
riety of factors have been shown to affect the selection (Krieger et al., 2000; Einhauser
et al., 2008; Kienzle et al., 2009; Tatler et al., 2011; Schomaker et al., 2017).
We are nonetheless unconvinced that fixations alone are sufficient for creating maps
of importance of visual elements, for several reasons: fixated locations are not neces-
sarily attended; and the frequency and duration of fixations is not always correlated
with the importance of an object as assigned by an observer. Here, we suggest a new
experimental setup that creates a map based on whether visual elements have been
encoded and deemed important.
1.1. Fixations vs. visual attention
It is well known that observers often fail to detect seemingly prominent visual ob-
jects, typically because they are given a task for which the object is irrelevant – an effect
called inattentional blindness (Simons, 2000). One may speculate that the specific task
leads to a lack of fixations on the unexpected object, yet this is not the case. In partic-
ular, Memmert (2006) repeated the famous Gorilla experiment while tracking the gaze
of the observers. They found that observers fixate the Gorilla a significant amount of
time and that recognizing the Gorilla was independent of the fixations. Koivisto et al.
(2004) used a more controlled setup with a screen presentation of the stimulus. Their
2
particular findings are that visual attention and fixations were uncorrelated under the
specific experimental conditions (they note, however, that this might not be common
under natural viewing conditions).
1.2. Fixation statistics vs. importance
A variety of factors influence the statistics of fixations, some of which may be un-
correlated with the importance of a scene element that is later attributed by an observer.
To begin with, a confounding factor is the experimental setup for gathering fixa-
tions: it is common to only consider fixations in the first few seconds after the onset of
the stimuli. This is based on the intuitive idea that the most important elements would
attract visual attention first. Indeed, it has been shown that fixation locations relate
less to image statistics over time (Itti and Koch, 2000; Parkhurst et al., 2002). This
approach, on the other hand, limits the number of locations extracted, as each fixation
has a certain minimal duration (typically not less than 150ms (Pannasch et al., 2008)).
Also the distance between fixations (the saccade amplitude) is limited to about 7◦ vi-
sual angle. These characteristics of fixations, saccades, and the time in which fixations
are collected have an important consequence: only few elements can be visited and,
since the distribution of fixations is constrained but not the distribution of scene ele-
ments, it is not clear that all fixations indeed correspond to visually important elements.
The common strategy for coping with this potential problem is to average over the data
of repeated trials and several observers (e.g. Van Der Linde et al., 2009; Ramanathan
et al., 2010; Koehler et al., 2014). It remains unclear why averaging would result in a
measure of importance.
More importantly, it is known that fixations may extend or accumulate for a variety
of reasons, such as generally being biased towards the center of the field of view (Tatler,
2007), being longer for scene elements that are unlikely or improbable given the scene
context (Loftus and Mackworth, 1978; De Graef et al., 1990), or following the gaze of
humans depicted in the scene (Ricciardelli et al., 2002; Hutton and Nolte, 2011). We
also speculate that visual elements that are complicated (e.g., cluttered or obstructed)
require longer or more fixations. In all of these cases, the importance of the visual
element depends on a variety of factors, and assigning high importance based on just
3
Figure 1: An example of mirror scene.
the number or duration of fixations is unwarranted. As an example for this statement
consider a mirror in the scene (see Figure 1): the mirror likely draws fixations because
it reflects a variety of interesting objects. We would suspect, once the region is being
decoded as a mirror it likely decreases in importance, either because the objects are
visible in the scene and can be attended to directly, or because understanding the scene
through a mirror is too complicated to be attempted during a brief exposure to the
stimulus.
1.3. Importance of fixation based on recall
In this work, we distinguish among fixated regions that do and do not make it into
the visual memory; and among the ones being stored we create a map that reflects their
relative importance. The contribution of the paper is an experimental setup and an
algorithm for post-processing that leads to such maps.
The idea is to base importance on the motion of the eyes during the recall of the
image. It is know that the eyes move during recalling an image. In the absence of other
visual features (i.e. while looking at nothing), the motion of the eyes is related to the
motion of the eyes while the stimulus was present. In particular, fixations during recall
are related to fixations during exposure with the stimulus (Johansson and Johansson,
2014; Laeng et al., 2014). We want to use fixations during recall as indicator of the
presence in visual memory, and by averaging over multiple observer as a measure for
4
their importance.
The fixations during recalling an image while looking at nothing lack a reference
frame and cannot be used directly to identify important elements in a visual stimulus.
Our idea is to record the exploration of an image using a calibrated eye tracker as well
as the immediate recall of the image. The fixations during exploration of the present
stimulus can be mapped to locations in the image. And the fixations during recall can
be used to validate the fixations during exploration.
Despite the large body of work on eye motions during recall while looking at noth-
ing, we were unable to find data from experiments. In order to analyze our idea, we
have performed an experiment based on standard image data, in which we record eye
movements during exploration and recall of those images. The complete data set for
pairs of exploration and recall sequences will be made publicly available.
1.4. Eye movements during recall
The looking at nothing phenomenon, i.e. the effect of eye movements during the
recall of a scene while looking at a neutral visual stimulus, has been known for a long
time (Moore, 1903), yet was mostly considered an odd coincidence. Laeng and Teodor-
escu (2002) are usually credited with having established that eye movements during re-
call play a functional role in memory retrieval. Their experiment is based on inhibiting
eye motion (here, by asking observers to fixate on a point) during either exploration
or recall. Inhibiting eye motion during exploration lead to reduced eye motion during
recall, and inhibiting eye motion during recall let to decreased recall performance. A
variety of other setups have confirmed and refined these findings (Johansson and Jo-
hansson, 2014; Laeng et al., 2014; Scholz et al., 2015; Bochynska and Laeng, 2015;
Pathman and Ghetti, 2015; Scholz et al., 2016).
In particular, it has been shown that the fixations during recall of the stimulus are
connected to the location of the objects. Ferreira et al. (2008) noted that observers'
gaze would move to the location of previously observed visual elements when cued by
keywords, even if the visual element was not present anymore. de Vito et al. (2014)
confirmed that inhibiting eye motion during recall decreased memory performance,
however, they refined the results and showed that only memory of spatial location
5
was affected, while recollection of visual content was independent of eye movements.
Martarelli et al. (2017) showed that gaze still indicates spatial location if observers
were asked to imagine a scene with objects being replaced by other similar objects.
All these experiments provide evidence that the eye position during recall of a stimulus
reveals the location of an object.
1.5. Accounting for displacements in recall data
We want to use the fixations during recall to validate the fixations during explo-
ration. The problem in this task is that the fixations in the recall phase are only qual-
itatively related to the stimulus – without the stimulus being present there is no frame
of reference. We also find that the motion of the eye during recall contains significant
local displacement, i.e. the spatial encoding in eye position contains error. We develop
an approach for the approximate matching between the fixations in the recall sequence
and those in the corresponding exploration sequence. It is based on computing a de-
formation mapping for the locations of fixations in the recall phase. After applying the
deformation we retain fixations from the exploration sequence that are close enough to
a fixation in the mapped recall. A threshold on the distance between deformed fixa-
tions in the recall and fixations in the exploration sequence allows us to steer the filter.
This leads to varying maps of importance, going from considering all fixations in the
original exploration sequences to subsets. We explore that resulting spatial importance
maps and observe a variety of meaningful effects.
2. Method
We collect eye movement data during the exploration and recall of images. We first
describe how we acquire this data and motivate our design choices for the experiment.
To our knowledge, this is the first experiment in this direction. We then provide statis-
tics of the acquired data and show that it is consistent with other data collected for the
free-viewing task.
6
2.1. Experiment
The basic setup for our experiment is to present observers with an image as stim-
ulus and let them recall the image immediately after the presentation of the stimulus.
Their head remains in the same position during stimulus presentation and recall. In the
following we discuss the details.
2.1.1. Participants
We recruited 30 participants for our experiment (mean age = 26, SD = 4, 9 female).
Importantly, all participants are
All reported normal or corrected-to-normal vision.
naive w.r.t. the aim of using the fixations during the recall phase to gauge the impor-
tance of fixations during exploration. All studies have been carried out in accordance
with the Code of Ethics of the World Medical Association (Declaration of Helsinki).
Participants were informed about the procedure before giving their written consent and
could stop the experiment at any time. Their time was compensated and all data is used
anonymously.
2.1.2. Apparatus
The experiment was conducted in a dark and quiet room. A 24-inch display (0.52m×
0.32m) with a resolution of 1920×1200 pixels was in front of the observer at a distance
of 0.7m. We used an EyeLink1000 desktop mount system (SR Research, Canada) to
record the eye movements at a sampling rate of 1000Hz. A chin and forehead rest was
used for stabilization. All experiments were conducted in binocular viewing condition,
but only the movements of the dominant eye were recorded.
2.1.3. Stimuli
We use 100 natural images randomly selected from the MIT data set (Judd et al.,
2012). This set includes both indoor and outdoor scenes of various complexity (see Fig-
ure 4 for example images). These images have been used in eye tracking experiments
with a free viewing task and the data is publicly available (Judd et al., 2012). This
allows us to compare the eye tracking data we gathered during the exploration phase to
existing data. All images were presented at the center of display in their original size
with the largest dimension being 1024 pixels.
7
2.1.4. Procedure
The experiment consists of 100 basic trials, and a single image was presented in
each trial. The details of the presentation in one trial are: Prior to the presentation
of the image, the screen is black and shows a white dot in the center (1◦ visual angle).
Observers are asked to fixate at the white dot so that the eye motion starts in a consistent
way for all images and observers. Then the image is presented for 5 seconds. Observers
are asked to explore the image in order to later be able to recognize it (a free viewing
task). After presentation of the image white noise is shown for 0.5 seconds to suppress
the after image. Then the screen is set to neutral gray for 5 seconds. During this phase,
observers are asked to immediately recall the image from memory. After that the screen
turns black for 1.5 seconds before the procedure is repeated for the next image. Each
observer saw all 100 stimuli and the order of them is randomized for each participant.
Instructions as to encourage observers to explore the mental image during the recall
phase were given at the beginning of the experiment. The 100 trials were divided into
five blocks. Each block of 20 trials started with a standard 9-point calibration proce-
dure. We repeated the calibration until the average accuracy reported in the following
validation was below 0.5◦ and no validation point had an error larger than 1.0◦. After
a successful calibration 20 trials were performed. This procedure required roughly 5
minutes. All participants performed the experiment, however, for two of them we were
unable to achieve the desired calibration accuracy (probably due to glasses they had to
use). This leaves us with data sets from 28 observers.
An important design choice was the duration of the exploration and recall peri-
ods. In particular the recall phase cannot be too short, because retrieving the image
from memory requires time. We decided to follow experiments in cognitive psychol-
ogy (Laeng et al., 2014) and allow 5 seconds for recall. For reasons of temporal equiv-
alence, we also present the stimulus for 5 seconds. The basic task of recalling an image
turns out to be significantly more demanding than just exploring the image. To avoid
exhaustion we limited the number of presentations to 100 while still having a suffi-
cient number of repeats for our analysis. Participants were allowed to take a break of
arbitrary length after each block.
8
The whole experiment lasted about one hour. At the end of the experiment partic-
ipants were shown 10 images, half of which were part of the 100 stimuli used in the
experiment. Images were presented one after the other in a randomized order, and par-
ticipants had to decide if the images were among the 100 presented to them. All parts
of the experiment were explained in detail at the beginning, including the memory test
at the end.
2.2. Data processing and analysis
We congregated and analyzed the data from 28 observers for the exploration and
recall phase of each image.
2.2.1. Processing raw data and fixation detection
We base our analysis on fixations. To extract them from the raw eye tracking data,
we use a dispersion based algorithm (Holmqvist et al., 2011), which involves setting
two thresholds: minimum duration τ and maximum dispersion ϕ. Setting these pa-
rameters is a standard procedure for eye-tracking experiments with the stimulus being
present, we feel experimentation is warranted for the recall phase. We tested a range of
parameter combinations for values discussed in the literature and Figure 2 shows the
classification results of one sequence under different parameter settings. We found that
also during recall the settings were not critical and within the common range detected
fixations did not vary much. Thus we fixed the parameters for the following analysis.
The minimum duration of fixation is set to be 100 ms and maximum dispersion is set
to be 0.5◦ visual angle.
We detect fixations in all exploration and recall sequences. During exploration the
median and mean number of fixations is 16 (SD=2.8), and during recall the median
and mean number of fixations is 11 (SD=3.6). The fewer fixations in recall have a
correspondingly longer duration (avg = 452.2 ms, SD = 308.0 ms) than fixations in
exploration (avg = 278.0 ms, SD=73.4 ms). Fixation durations are plotted as a function
of the starting time of the fixation during the trial in Figure 3. Exploration phase is
separated from the recall phase. Each fixation in the dataset from all 28 observers is
depicted as a cross and the black curve shows the averaged duration in each phase.
9
Figure 2: Fixation classification results for different parameter settings. The minimum duration threshold τ
ranges from 50 ms to 200 ms and the maximum dispersion threshold ϕ ranges from 0.25◦ to 1.0◦. Raw
samples are depicted in green and fixations in red whose radius correspond to the durations. Fixation number
(#fix) and mean duration (md) are reported in each plot. As shown in the first row, small dispersion ϕ results
in more fixations with shorter durations and a large ϕ tends to fuse fixations together. Second row depicts
the influence of τ, the minimum duration threshold. Small τ leads to over segmentation and large τ results
in much less fixations.
After the first two fixations, there is no significant variation of fixation duration
during the trials for both exploration and recall (the durations getting shorter towards
the end of the trial is an artifact of stopping data collection after 5 seconds). On average
fixations in recall are roughly two times longer than the ones in exploration.
The fixations from all participants for a single image can be summarized in a spatial
histogram, leading to a so-called heat map for the image. We do this for the data from
our experiment and the publicly available data for the 100 images. Similar to Judd et al.
(2012) we remove the first center fixation from each sequence and apply a Gaussian
filter with a kernel size equivalent to 1 degree of visual angle.
2.2.2. Data from exploration
As a sanity check, we compare the heat maps resulting from the publicly available
data to the heat maps from the exploration phases in our experiment. To measure
the similarity of corresponding maps we use Pearson's correlation coefficient (CC),
10
Figure 3: Comparison of fixation durations in exploration (a) and recall (b). Fixation durations are plotted as
a function of starting time . All fixations from 28 observers for 100 trials are depicted as crosses. The black
curves indicate the mean durations and the center 50 percent intervals are depicted in blue.
following Bylinskii et al. (2016). Averaging over all 100 pairs of saliency maps, we
find a mean CC of 0.766 with standard deviation 0.115. We also perform a qualitative
comparison. Figure 4 shows maps for different ranks among the correlations. Most
heat maps are very similar. For those that are dissimilar, we see a stronger center
bias in the existing data. We speculate that this is due to the shorter presentation time
of 3 seconds vs. the 5 seconds in our experiment. We conclude that our results are
qualitatively similar to previous experiments with the same data.
2.2.3. Data from recall
Compared to exploration sequences, recall sequences have fewer but longer fixa-
tions (see. Figure 3) The last row in Figure 4 shows the heat maps generated from
the recall phases. As can be seen, while they roughly resemble the maps from the ex-
ploration phase, they are more biased towards the center and typically fail to exactly
correspond with features in the image. In particular, they cannot be used directly as
heat maps for the images.
In some cases, the correspondence between fixations in exploration and recall is
clear. Then we found that the temporal order is generally not preserved (see Figure 5
for some examples). This is consistent with previous results in cognitive psychol-
ogy (Johansson et al., 2012).
11
Figure 4: Comparison of heat maps. Images are ranked based on correlation scores between the heat maps
generated using the existing data from Judd et al. (2012) and the data we collected during presentation of the
images. Seven examples are provided. Each column shows the original image, the MIT heat map, and the
heat maps created from the exploration and recall phases of our experiment.
Figure 5: Pairs of exploration and recall fixation sequences from four observers. First row shows fixation
sequences collected during exploration and second row shows the corresponding recall sequences. The
temporal order of each sequence is indicated by the numbers and consecutive fixations are connected by
lines.
We also notice that observers tend to 'stall' during recall. This means they stop
moving their eyes, leading to fixations that are unlikely corresponding to image con-
tent. This sometimes happens for long durations at the center. For 5 out of the 28
participants, the number of fixations in recall is less than half compared to exploration.
One participant reported after the experiment that he changed his strategy through the
12
ImageMIT Heat mapViewingHeat mapcc = 0.448cc = 0.662cc = 0.748cc = 0.926cc = 0.858cc = 0.838cc = 0.789rank 1rank 17rank 34rank 50rank 67rank 83rank 100RecallHeat map Figure 6: Subset selection criteria. We base our selection on the similarity between the averaged exploration
and recall heat maps. Data sets are ranked according to correlation scores. Two different types of inconsis-
tencies are visible in the least similar pairs: the distributions of recall fixations are either peaked at one point
or spread more randomly.
experiment and only recalled the single most interesting element of the image.
2.3. Selecting a subset of observers
The analysis above suggests that not all observers consistently moved their eyes to
revisit important features in the image. Without ground truth it is impossible to decide
if the data from a particular observer is useful. In order to avoid basing the analysis on
low quality data, we focus our analysis on only half of the observers.
For selection, we take the fixations for all 100 images from one observer and com-
pare the spatial distribution of fixations during exploration and recall. Figure 6 shows
heat maps for several observers. The similarity of the spatial distributions varies; some
show significantly different patterns (see the three examples on the right hand side in
Figure 6). The inconsistencies are of two types: 1) the spatial distribution in recall has
smaller variance compared to exploration (indicating observers fixated mostly at one
location), 2) the spatial distribution has much higher variance. Both types of differ-
ences hamper further using the fixations from recall for matching against the fixations
in exploration.
To account for the idiosyncrasies of individual observers we compare the spatial
distribution from exploration to the spatial distribution from recall for each observer
(rather than judging only the spatial distribution during recall). As an unbiased method
to measure their similarity we compute Pearson's correlation coefficient. This results
in a ranking of the observers - the images in Figure 6 are sorted based on this ranking.
We select the upper half of this ranking for the following processing and analysis. The
13
Figure 7: Example of the location mismatch between fixations in exploration and recall. Fixations from
exploration are shown in green (a); recall fixations are subset of the fixations during exploration and are
shown in red. Their displacement (b) contains global rigid transformation (c) and deformation (d). The
deformation mapping is based on the yellow consensus locations (e).
publicly available data will include all observers.
3. Filtering the fixations based on recall witnesses
Our idea for exploiting the information in the recall is to match the fixations during
exploration with the fixations during recall. A fixation in recall matched to a fixation
in exploration serves as a witness, testifying that the visual element corresponding to
the fixation is important.
3.1. Deformation mapping
The problem of establishing the matching is that spatial locations of the fixations
in recall are distorted relative to the locations of features in the image. This distor-
tion could be decomposed into a global rigid transformation (due to a lack of reference
frame) and local inaccuracies. Figure 7 provides an illustration: The six points in
a) depict a sequence of fixations during exploration, of which only five are being re-
called. The observed set of fixations during recall projected into the same reference
frame is shown in b). We may think of the locations as being composed of the rigid
transformation in c) and the deformation in d). The deformation contains no global
rigid transformation in the sense that moving or rotating the fixations would not make
the sum of the squared distances to the matched fixations in exploration any smaller.
Note that fixations in recall and exploration that are corresponding cannot be found by
simply considering their distances.
14
When data is corrupted by global transformation and deformation, it is common
to approximate the effects by minimizing the squared distances between matched data
points. Yet, this is a chicken and egg problem: The estimated mapping is supposed
to improve the matching, while the matching is needed to estimate the mapping. The
common solution to this dilemma is to start with a first guess about the matching, then
compute the deformation, and based on the deformation make a better guess about the
matching, and so on until the method converges. A common approach of this type
is Iterated Closest Point (ICP) (Besl and McKay, 1992), which is used to compute a
global rigid transformation to match two partially overlapping point sets.
The approach we suggest for matching the recall fixations to the fixations in ex-
ploration is inspired by ICP, yet extends it in two important ways: First, rather than
using a closest point matching, it matches fixations in recall to consensus locations in
the exploration data. Second, rather than computing a global rigid transformation it is
based on a deformation mapping that has controllable overall deviation from a global
rigid transformation.
The first modification is motivated by the situation in Figure 7 b), namely that
matching the recall fixations to the closest fixations in exploration may lead to wrong
assignments. These wrong assignments will steer the estimated mapping away from
the desired solution. Given the data, we believe several situations have to be accom-
modated:
1. One fixation in recall maps to exactly one fixation from exploration.
2. One fixation in recall maps to several close fixations from exploration, i.e. the
observer recalls just one scene element that drew several fixations.
3. A fixation in recall may have no corresponding fixation in exploration (i.e. be-
cause the fixation is unrelated to the process of recalling, or the spatial error is
too large to be rectified).
We suggest to accommodate all cases by computing a consensus location for each
recall fixation, based on computing a weighted average of the fixation locations in
exploration. The weights decay exponentially with distance. This has the effect that
if only one fixation in the exploration sequence is close while the others are far away
15
(case 1) the closest location will receive a large weight, while the others relatively small
weights, so that the consensus location will be the matching fixation. In case several
fixations in exploration are close (case 2), all of them receive equal weights, and the
consensus location is the center of these fixations. If no fixation is found for matching
(case 3), all fixations receive little weight, and the consensus location is roughly where
the recall fixation already is. Figure 7 e) shows the consensus locations for the situation
in b). The computation of consensus locations is explained in Appendix A. It can be
controlled by a parameter wp, measured in visual angle, which could be interpreted as
distance of fixations that contribute significantly to the weighted averaging procedure.
We set the parameter to wp = 2◦ and discuss this choice in Section 6.
The consensus locations for recall fixations are used to compute a deformation
mapping D : R2 (cid:55)→ R2 from the current positions of the recall fixations to the desired
ones. Allowing deformation overcomes the problem that the best we could achieve
with computing a rigid transformation is to be left with the deformation, i.e. the situa-
tion depicted in Figure 7 c). However, some global rigidity needs to be preserved, i.e.
the positions should not deform arbitrarily. This is important, as we would otherwise
always match all recall fixations to some fixations in exploration. For example, if we
allowed arbitrary scale, it would always be possible to scale the set of recall fixations
to a single point and then match it to one of the fixations in exploration. To avoid
such degenerated solutions it is important to restrict the mapping to preserve the global
structure. The computation of the mapping based on consensus locations and the cur-
rent positions of the recall fixations is explained in Appendix A. The global rigidity of
the mapping can be controlled by a parameter wd, measured in visual angle. Roughly
speaking, wd describes the distance of points that may be transformed by two rigid
transformations that differ significantly. If wd is large, fixations are transformed by
very similar rigid transformations, restricting the deformation; if it is small, fixations
are transformed by independent rigid transformations, allowing deformation. We set
the parameter to wd = 10◦ and discuss this choice in Section 6.
The necessary steps to compute the deformation mapping are given in pseudocode
in Appendix A. Once the mapping is computed, we simply use distance as the sole
criterion for testimony: a fixation in exploration has a witness, if there is a mapped
16
Figure 8: Quantitative results of filtering relative to witness radius ∈ [1◦, 10◦]. Results from sequences in
the selected subset of observers are shown in orange and the results from the the excluded subset are shown
in blue. (a) Relative number of fixations retained after filtering. Each observer's data is scattered in the
graph and the light color marks the center 50% region of the distribution. The effect of the witness radius is
uniform across different images and observers. (b) Percentage of all filtered maps whose most salient spots
remain within 4◦ after filtering.
fixation in recall that is closer than the witness radius .
4. Quantitative results of filtering
Filtering is supposed to reduce the number of fixations in a meaningful way. It
is affected by the choice of witness radius . Before we analyze specific effects on a
qualitative level we present global statistics to show that the effect of filtering is not
random.
One statistic is the number of fixations that remain after filtering. We expect that
even for relatively small witness radii a significant number of fixations remain. The
left graph in Figure 8 depicts the number of retained fixations relative to the witness
radius. Here, we differentiate between the subset of observers we selected for further
processing vs. the subset of excluded observers. Note that the effect of the threshold
is quite similar across different images and observers. For values larger than = 10◦
we effectively retain all fixations. For very small values the amount of noise in the
recall data leads to results that seem to contain no useful information. In the subset
of observers we consider more reliable fixations remain after filtering. This means
17
that more fixations in the exploration sequences found a closest match in the recall
sequences. This is what we expect, as the distributions of the fixations in exploration
and recall are similar.
We also analyze if the sets of fixations change in the sense of their relative impor-
tance within an image. For this we identify in each image the region with the most
fixations during exploration based on the smoothed spatial histogram. We do the same
after filtering and consider that the point has shifted if the difference in spatial location
exceeds 4◦. The right graph in Figure 8 shows the result relative to the witness radius.
We observe that, as expected, for large witness radii there is no significant change, yet
for smaller radii the most salient object changes in about 40% of the images. We take
this as indication that filtering is not random: on one hand, we expect filtering to affect
the relative importance of elements in the scene; on the other hand, in a significant
number of images the most fixated object is indeed important.
Figure 9 illustrates the effect of varying the threshold on the resulting maps for
two of the images using the selected subset. We show heat maps for the values =
1◦, 2◦, 4◦ and observe different effects: some regions that attracted a lot of fixations
appear to have been deemed relatively unimportant, for example the sign to the left
in the top image. Other regions are boosted in importance after filtering such as the
objects in the foreground in the top image or the tower to the right of the road in the
bottom image. In the following section we attempt a systematic discussion of these
effects based on the selected data.
5. Main results
We apply the deformation process to each pair of fixation sequences for all images
from the subset of selected observers. Then we apply the recall based filtering based
on witness radius = 1◦.
In other words, a fixation is retained only if there is a
corresponding fixation in the deformed recall sequence that is within 1◦ visual angle.
This is a rather strict setting, removing a lot of fixations. We chose it for achieving a
significant effect, at the expense of possibly introducing more noise.
Our analysis is based on comparing the heat maps from the unfiltered and filtered
18
Figure 9: Filtered maps using different thresholds. We show examples of = 1◦, 2◦, 4◦. Original heat
map shows the result based on all fixations during exploration of the image. The color coding of the scaled
maps is based on the highest value appearing in all heat maps, which is naturally the one generated without
filtering. The normalized map is color coded in its own range.
images averaged over the selected 14 observers. For each image, we detect the region
showing the largest (absolute) difference in the heat maps (these difference maps are
also shown in all illustrations). The inspection of these 100 difference maps showed
that the removal of fixations from the exploration sequences can be explained in most
cases based on established effects. Before we exemplify these effects, we present an
explanation for the removal of a fixation that appears to apply to virtually all cases.
For a region to be removed it necessarily received many fixations during exploration
but few during recall. Indeed, most of these regions show complex visual patterns,
which explains why they have received a lot of fixations during exploration. Then one
of two cases seems to apply: 1) despite the effort, the region could not be decoded into
a semantic object and is, consequently, not stored; 2) an object has been recognized,
but is deemed unimportant (relative to the rest of the scene, or at least relative to the
19
Image Stimulus ε = 1°ε = 2° ε = 4°Original Heat MapScaled by Original Heat MapNormalized Normalized Image Stimulus ε = 1°ε = 2° ε = 4° ε = 1°ε = 2° ε = 4°Original Heat Map ε = 1°ε = 2° ε = 4°Scaled by Original Heat Map small amount of time allowed for recalling). In the following we discuss mostly effects
of the second type. Clearly, these effects may vary among observers – in particular, the
importance of a recognized object is highly subjective. For effects in the whole dataset
of 100 images see the supplementary material.
5.1. Low-level features
Low-level features have been shown to contribute to saccade target selection (Bad-
cock et al., 1996; Krieger et al., 2000; Kienzle et al., 2009). Consequently, they con-
tribute significantly to fixation-based attentional models. In our model that considers
also attention, awareness, and relative importance of the visual object, their strength is
no more relevant for the relative importance of a visual element.
Figure 10 shows three examples of such cases. A white box in the right corner over
a black background in the first image draws a lot of fixations due to its high contrast. It
is less dominant in the filtered heat map as shown in both the third and last columns. We
would speculate that observers either were unable to decode the object or decide that a
white box-like object is not important for the scene. Similar effects can be observed in
the second example on the red car (which is hard to be recognized as such) and in the
third example on the ball (which is relatively unimportant compared to its colorfulness
and contrast). Instead we see fixations on the train and the shape of the lamp become
relatively more important.
5.2. High-level features
Among high-level features, we observe noticeable effects on text, signs, and people.
5.2.1. Text and signs
Text and signs commonly draw attention as they provide a lot of information in
small spatial region. Some signs are purposefully designed to be conspicuous. As
shown in Figure 11, they have relative high attendance during exploration. Their im-
portance depends significantly on the result of decoding their meaning and then on the
experimental conditions, for example the task, but also the experience of the observer.
Based on our selection of examples, the elements in Figure 11 significantly decreased
20
Figure 10: Strong low-level features after filtering. The second column shows the heat map based on all
fixations during exploration of the image. Removed fixations are used to generate the removal map (or
difference map) and remaining fixations are used to generate the filtered heat map, which is essentially the
difference between original heat map and the removal map. The color coding of the removal map is scaled by
the highest value in original heat map, while the filtered heat map is normalized in its own range. Fixations
triggered by low-level features are largely removed in the filtered heat maps, if the underlying visual object
was difficult to understand or turned out to be of little relevance.
in importance. We believe the text in the first row and the sign in the last row were
undecipherable, while in the other two examples the text (whether legible or not) and
the exit sign were deemed unimportant for the scene.
5.2.2. People
As shown in Figure 12, observers look at humans in images regardless of their size.
The relative importance of visual objects recognized as people apparently depends on
the scene.
In the two examples above the dashed line, people are not a significant
element of the scene – they are expected and not the dominating elements. In the other
two examples, humans are contributing more to the composition of the scene, and they
remain important after filtering.
21
Image StimuliOriginal Saliency MapNormalized Filtered Saliency MapRemoval Map Figure 11: The effect of filtering on text and signs. The meaning of the color coded images is as before.
Illegible or unimportant text (upper rows) as well as signs without particular importance for the scene (lower
rows) lose importance.
5.3. Relative importance of similar items
Similar objects in a scene may be assigned equal or different importance, indepen-
dent of their visual representation. We provide several examples in Figure 13. Similar
objects may well be recalled similarly, as is demonstrated in the first example. In the
second row, the larger boat dominates the fixations. Based on recall, the smaller boat
in the middle row gains in relative importance. Among a set of faces, all of which draw
fixations, the one that stands out indeed becomes more important.
5.4. Other effects
When looking at an image we often follow the apparent gaze of people in the scene
because it might provide additional information about people's intention or the local
22
Image StimuliOriginal Saliency MapNormalized Filtered Saliency MapRemoval Map Figure 12: Filtering may or may not affect people. The meaning of the color coded images is as before.
People in the scene are observed in exploration but their importance in the filtered heat maps depends on
the scene composition. In the upper rows, the importance of people decreased, while in the lower rows they
remain important.
environment (Corkum and Moore, 1998; Hutton and Nolte, 2011; Gallup et al., 2012;
Bayliss et al., 2013). This effect is notoriously hard to model computationally (Xu
et al., 2014; Recasens et al., 2015). In our model, the object found in these locations
determines the importance of the scene element, and not the mere fact that some person
in the scene appears to be looking at the region (see Figure 14).
Other noteworthy visual elements are horizons and mirrors. Both attract a lot of
fixations, however, appear of little importance when the image is being recalled (see
Figure 15).
23
Image StimuliOriginal Saliency MapNormalized Filtered Saliency MapRemoval Map Figure 13: Filtering results for elements with similar meaning. The meaning of the color coded images is
as before. In the top row, similar items preserve their similar importance. The smaller boat in the middle
row gains in relative importance, while still being dominated by the larger boat. In the last row, the face that
stands out among other faces increases in importance.
Figure 14: The point of gaze of people in the stimulus is effectively filtered. The meaning of the color coded
images is as before. In our recall based attention model, the apparent gaze seems to have less impact and the
resulting heat map rather depends on the scene elements being gazed at.
24
Image StimuliOriginal Saliency MapNormalized Filtered Saliency MapRemoval MapImage StimuliOriginal Saliency MapNormalized Filtered Saliency MapRemoval Map Figure 15: Other effects in the filtered saliency maps. The meaning of the color coded images is as before.
Observers scan the horizon, yet only recall elements if they were dominant. Likewise, mirrors are complex
visual stimuli, but once they are decoded they become unimportant.
6. Discussion
6.1. Data collection
While the fixations we extracted from the collected eye tracking data in the explo-
ration phase are similar to the reference (Judd et al., 2012), there are also some images
with considerably different resulting heat maps. Our experimental conditions differ
only in the length of exposure, namely 3s vs. 5s in our setup. As the data is used for
judging the success of computational models for the prediction of saccade targets, we
wonder about the variation in the data across observers, and minor variation in experi-
mental conditions, relative to the performance difference of different models.
We noticed that recall is a strenuous task and not all observers are willing or able
to continuously move their eyes. It remains unclear if a lack of motion means that
observers paused the recall, however, the fact that the recall performance correlated
with motion of the eyes in other experiments (Johansson and Johansson, 2014; Laeng
et al., 2014; de Vito et al., 2014) suggests that this is the case. It is unclear how to
motivate observers to engage more actively in recall.
Observers only had limited time for recall, thus limiting the number of elements
that could potentially being recalled. This situation is similar to the influence of time
for recording the fixations during exploration: shorter times leads to fewer fixations and
more variation across observers, while longer times increases the amount of noise (Itti
25
Image StimuliOriginal Saliency MapNormalized Filtered Saliency MapRemoval Map and Koch, 2000; Parkhurst et al., 2002). It is not clear how our choice of 5 seconds for
recall affected the results.
If it was acceptable to present the stimulus for a significantly longer time, a sub-
stantially different protocol for collecting information could be used: observers use
standard interaction devices such as computer mice or touch screens to mark inter-
esting points or objects (Jiang et al., 2015; Kim et al., 2015, 2017). This higher-level
interaction leads to selection based on semantics of features. It has been recently shown
to be inconsistent with fixations (Tavakoli et al., 2017).
6.2. Subset selection
We selected only half of the observers based on the correlation of the spatial his-
tograms between exploration and recall. The choice of 14 is arbitrary and it appears
that varying it has almost no effect on the outcome. However, we do observe that the
data of few observers is different and would have obscured the results we show.
For ranking observers we also considered using the pair correlation function (Illian
et al., 2008; Oztireli and Gross, 2012) as it has been recently demonstrated to char-
acterize the spatial distribution of fixations well (Trukenbrod et al., 2017). Yet, we
felt there is no obvious way to compare the distributions. Depending on the choice of
parameters and method for measuring the similarity the results varied, while largely re-
sembling the order computed with the much simpler approach of taking the correlation.
Consequently, we opted for correlation for the subset selection.
6.3. Deformation mapping
As the true matching is unknown, it is impossible to optimize parameters control-
ling the matching and deformation automatically. We had asked 3 participants in the
experiment to manually mark the correspondences among the fixations, in this way
generating a reference, yet their data are unreliable (consistent with the literature (Vo
et al., 2016)).
On the other hand, the results of the deformation process are stable across a wide
range of parameters (see Figure 16). We found that the radius for matching recall
fixations to exploration fixations wp should be chosen in the range of 2◦ − 4◦ visual
26
Figure 16: Matching results using different parameters. Original recall fixations (in green) are matched to
fixations during exploration (in yellow) and matched fixations are shown in fuchsia. Matching is stable across
various parameters. The rightmost column shows cases of undesired deformation for extreme parameter
settings. More fixations are considered for matching with a larger wp while the deformation is more rigid
with a larger wd.
Figure 17: The deformation model performs well on different examples. Parameters wp is set to 2◦ and wd
to 10◦. Each example shows the fixations during exploration in yellow and recall fixations in green. Each
pair overlaid with the stimulus is shown at the top-left corner. Matched fixations are shown in fuchsia.
angle. This means we expect that matching fixations are usually not separated by more
than twice this amount. To limit the deformation of the mapping we have tried values of
wd ∈ [4◦, 16◦]. Based on experimentation we have settled for wp = 2◦ and wd = 10◦.
Figure 17 shows the results of three examples using this parameter setting.
In order to understand if deformation was really necessary to achieve the results
we have presented, we perform the filtering based on the original, undeformed recall
sequences. Similar to Figure 8, Figure 18 plots the percentage of remaining fixations
as a function of the witness radius varying from 1◦ to 10◦ visual angle, as well as
the relative changes of importance of fixations. Comparing the results we notice that
without applying deformation, much less fixations remain after filtering. For = 1◦
27
Figure 18: Results of filtering relative to witness radius ∈ [1◦, 10◦] based on original recall fixations are
shown in blue. No deformation mapping is applied. Results of filtering based on deformed recall fixations are
depicted in orange for comparison. (a) Relative number of fixations retained after filtering. Each observer's
data is scattered in the graph and the light color marks the center 50% region of the distribution. Much less
fixations are retained when is small. (b) Percentage of all stimuli maps whose most salient spots remain
within 4◦ after filtering. More images contain changes of importance without applying deformation. Small
in particular highlights the effects of deformation mapping.
nearly 20% of fixations during exploration remain after applying deformation, while
only 5% have matched fixations in the original recall. Similar effects can be observed
on the changes of the most important visual element: filtering based on the deformed
recall sequences lead to significantly more of the images being unaffected. We spec-
ulate that without deformation several important elements remain unmatched, causing
significant but unwarranted changes to the map.
6.4. Visual attention and visual working memory
Visual attention and visual working memory are entangled: attention controls en-
coding in working memory by selecting what is the relevant information, and visual
working memory biases the focus of attention (Downing, 2000; Schmidt et al., 2002;
Olivers et al., 2006; Hollingworth et al., 2013). Several studies suggest that visual
attention and visual working memory share the same representations (Farah, 1985;
Harrison and Tong, 2009; Hollingworth and Hwang, 2013; Albers et al., 2013), and it
has been debated whether they should be regarded as one cognitive function since they
also share some of the same mechanisms (see Olivers (2008) for a review).
28
Our results suggest that the prioritization in visual attention differs from the priori-
tization in working memory: the spatial importance maps are affected by filtering, and
filtering is based on a functional connection to the spatial memory of visual elements.
While more experiments are needed to understand this difference, we hope the exper-
imental setup based on immediate recall of previously presented stimuli provides an
additional useful tool.
7. Conclusions & Outlook
The 'looking-at-nothing' phenomenon is an interesting way to test the awareness
of a feature in an image previously fixated during exploration. This gives rise to a new
way of measuring the importance of visual elements as assigned by a human observer.
We have generated a data set based on eye tracking experiment that allows studying
this idea and other researchers can use it as a basis for additional analysis.
Overall, this importance model appears to conform to the idea of decomposing
the image into relevant pieces - low-level features are only important for areas to be
visited during exploration. We argue that classical saliency maps overemphasize the
role of fixations. And concluding that objects which drew a lot attention based on
fixations are indeed recognized by the observer may be wrong.
The main challenge in this approach is the large amount of distortion, noise, and
artifacts in the recall data. There seems to be two ways to address this problem: better
processing of the data hopefully resulting in better matching; and better data collection.
It might be possible to exploit the raw data to better match the recall to explo-
ration sequence, as apparently not just the fixations correspond but also partial scan
paths (Johansson et al., 2006). The matching might be improved using machine learn-
ing techniques if we had ground truth data. It remains unclear how to generate such
data. As mentioned, we have tried to get this information from the observers, but this
attempt has been unsuccessful.
We believe the experimental setup leaves little room for improvement in order to
get cleaner data. The main problem is to motivate participants to equalize recall and
eye movements without making the experiment significantly more complex. We made
29
one observation that would be worth exploiting: It appears that the recall performance
improves in repeated trials. Similar effects have been observed in other recall experi-
ments Mantyla and Holm (2006); Kaspar and Koenig (2011). A good protocol for such
an experiment would still have to be designed.
The setup clearly exploits short term memory. The results for long-timer recall,
specifically if locations of fixations still encode the locations of features in images, are
inconsistent (Martarelli and Mast, 2013; Laeng et al., 2014; Wantz et al., 2016). We
expect the amount of noise in the data to increase and, therefore, finding the matching
between recall and exploration to be even more difficult. Nonetheless, a setup similar
to ours, yet with considerably delayed recall might be worthwhile. Such an experiment
would be related to image memorability (Isola et al., 2014; Khosla et al., 2015; Bylin-
skii et al., 2015), which explores the memorability of visual objects, and the results
would likely become even more dependent on personal factors.
The most successful techniques for predicting fixations to date are based on neural
networks (Huang et al., 2015; Kruthiventi et al., 2015). We considered using similar
approaches to predict the filtered heat maps. Better prediction would indicate that our
results are more consistent. However, given the amount of noise in the data we use for
filtering and the high-level effects we found in our analysis, we are unsure about the
meaning of this analysis. We still think it would be generally fruitful to use filtered heat
maps for learning important features that are closer to what humans really encode.
Acknowledgments
We would like to thank Marianne Maertens for valuable advice. Furthermore, we
thank David Lindlbauer for help performing the experiment. This work has been par-
tially supported by the ERC through grant ERC-2010-StG 259550 (XSHAPE).
Appendix A. Computing the deformation mapping
Let pi ∈ R2 be the positions of the fixations in the exploration sequence and rj ∈
R2 the positions of fixations in the recall sequence (in a common coordinate system).
30
We wish to compute a deformation D : R2 (cid:55)→ R2 that is applied to the recall locations
rj with the aim to align the data with the fixations positions during exploration.
We need two ingredients for computing the deformation: partial matching and de-
formation mapping. For the deformation, we suggest Moving Least Squares (MLS) (Levin,
1998). Here, we use this framework applied to rigid transformations, i.e. local rigid
transformations are fitted using weighted least squares. This approach has become
popular in geometric modeling where it is usually derived as minimizing the deviation
of the mapping from being locally isometric (Schaefer et al., 2006; Sorkine and Alexa,
2007; Chao et al., 2010).
We model the deformation D as a rigid transformation that varies smoothly over
space:
D(x) = Rxx + tx.
(A.1)
The subscript x indicates that rotation and translation vary (smoothly) with the location
x in the plane. They are computed by solving a weighted special orthogonal Procrustes
problem (Gower and Dijksterhuis, 2004), where the weights depend on the distance of
the points to x. Assume the desired position for rj is the position qj, then Rx, tx are
computed by solving
(cid:88)
j
arg min
TRx=I,tx
Rx
θ((cid:107)x − rj(cid:107))(cid:107)Rxrj + tx − qj(cid:107)2
2.
(A.2)
Here, the weight function θ should be smoothly decaying with increasing distance. We
use the common choice
θd(x) = e
− x2
w2
d ,
(A.3)
which gives us control over the amount of deformation in the mapping with the parame-
ter wd. The minimization can be solved directly using the singular value decomposition
(SVD), see Sorkine-Hornung and Rabinovich (2016) for an accessible derivation.
Note that for computing the mapping we simply assumed the desired positions
qj were given. We compute them as the distance weighted centroid of exploration
fixations:
(cid:80)
(cid:80)
i θp((cid:107)D(rj) − pi(cid:107))pi
i θp((cid:107)D(rj) − pi(cid:107))
qj =
,
(A.4)
31
Figure A.19: Iterations of the deformation. The initial state shows two fixation sets of viewing (orange) and
recall (green). Radius of each circle correlates to the fixation duration. The set of fuchsia circles are the
deformed recall fixations in each iteration. Note that the matching converges quickly after a few iterations.
where θp(d) quickly decreases such that points further away are receiving relatively
insignificant contribution.
Note that we are considering the distances of the fixations pi to the deformed lo-
cations of the fixations in the recall sequence. This means that setting the target loca-
tions depends on the deformation mapping and computing the deformation mapping
depends on the target locations. Consequently, we alternate the two steps as shown
in Algorithm 1. We start this process with D being the identity. Then we compute
the desired positions qj as explained above. The procedure converges after very few
iterations (see Figure A.19). Note that the deformation D needs to be evaluated only
in the location rj. This means for the next step we only need to compute Rx and tx
for x = rj.
References
References
Albers, A.M., Kok, P., Toni, I., Dijkerman, H.C., de Lange, F.P., 2013. Shared repre-
sentations for working memory and mental imagery in early visual cortex. Current
Biology 23, 1427 – 1431. doi:10.1016/j.cub.2013.05.065.
Badcock, D.R., Hess, R.F., Dobbins, K., 1996. Localization of element clusters: Mul-
tiple cues. Vision Research 36, 1467 – 1472. doi:10.1016/0042-6989(95)
00205-7.
Bayliss, A.P., Murphy, E., Naughtin, C.K., Kritikos, A., Schilbach, L., Becker, S.I.,
2013. "Gaze leading": Initiating simulated joint attention influences eye movements
32
Algorithm 1: Deformation mapping
Data: Fixations locations in exploration p1, . . . , pn and recall r1, . . . , rm,
convergence criteria λ
1, . . . , r(cid:48)
Result: Locations of mapped recall fixations r(cid:48)
1 ∀(i, j) ∈ [1, m]2 : wij ← exp(−(cid:107)ri − rj(cid:107)2/w2
d)
2 for j ∈ [1, m] do
3
m
(cid:17)
(cid:16)(cid:80)m
(cid:17)
j ← rj
r(cid:48)
¯ri ←(cid:16)(cid:80)m
6 repeat
4
5
7
8
9
10
11
12
13
j=1 wijrj
/
j=1 wij
Ri = (wi1(r1 − ¯ri), . . . , wim(rm − ¯ri))
i−pj(cid:107)2/w2
p)pi
i−pj(cid:107)2/w2
p)
qi ←
σ ← 0
for i ∈ [1, m] do
(cid:80)n
for i ∈ [1, m] do
(cid:80)n
j=1 exp(−(cid:107)r(cid:48)
j=1 exp(−(cid:107)r(cid:48)
¯q ←(cid:16)(cid:80)m
(cid:16)(cid:80)m
(cid:17)
UΣVT ← SVD(cid:0)RiQT(cid:1)
Q = (q1 − ¯q, . . . , qm − ¯q)
j=1 wijqj
/
j=1 wij
(cid:17)
15
14
s ← VUT(ri − ¯ri) + ¯q
σ ← σ + (cid:107)r(cid:48)
i ← s
r(cid:48)
17 until σ < λ
i − s(cid:107)
16
33
and choice behavior. Journal of Experimental Psychology: General 142, 76–92.
doi:10.1037/a0029286.
Besl, P.J., McKay, N.D., 1992. Method for registration of 3-d shapes, in: Robotics-DL
tentative, International Society for Optics and Photonics. pp. 586–606.
Bochynska, A., Laeng, B., 2015. Tracking down the path of memory: eye scanpaths
facilitate retrieval of visuospatial information. Cognitive Processing 16, 159–163.
doi:10.1007/s10339-015-0690-0.
Borji, A., Itti, L., 2015. Cat2000: A large scale fixation dataset for boosting saliency
research. arXiv preprint arXiv:1505.03581 .
Bruce, N.D.B., Tsotsos, J.K., 2005. Saliency based on information maximization, in:
Proceedings of the 18th International Conference on Neural Information Processing
Systems, MIT Press, Cambridge, MA, USA. pp. 155–162.
Bylinskii, Z., Isola, P., Bainbridge, C., Torralba, A., Oliva, A., 2015.
Intrinsic and
extrinsic effects on image memorability. Vision Research 116, Part B, 165–178.
doi:10.1016/j.visres.2015.03.005. computational Models of Visual At-
tention.
Bylinskii, Z., Judd, T., Oliva, A., Torralba, A., Durand, F., 2016. What do different
evaluation metrics tell us about saliency models? arXiv preprint arXiv:1604.03605 .
Chao, I., Pinkall, U., Sanan, P., Schroder, P., 2010. A simple geometric model for elas-
tic deformations. ACM Trans. Graph. 29, 38:1–38:6. doi:10.1145/1778765.
1778775.
Corkum, V., Moore, C., 1998. The origins of joint visual attention in infants. Develop-
mental psychology 34, 28–38. doi:10.1037/0012-1649.34.1.28.
Coutrot, A., Guyader, N., 2014. How saliency, faces, and sound influence gaze in
dynamic social scenes. Journal of Vision 14, 5. doi:10.1167/14.8.5.
34
De Graef, P., Christiaens, D., d'Ydewalle, G., 1990. Perceptual effects of scene con-
text on object identification. Psychological Research 52, 317–329. doi:10.1007/
BF00868064.
Downing, P.E., 2000. Interactions between visual working memory and selective at-
tention. Psychological Science 11, 467–473. doi:10.1111/1467-9280.00290.
pMID: 11202491.
Einhauser, W., Spain, M., Perona, P., 2008. Objects predict fixations better than early
saliency. Journal of Vision 8, 18. doi:10.1167/8.14.18.
Farah, M.J., 1985. Psychophysical evidence for a shared representational medium for
mental images and percepts. Journal of Experimental Psychology: General 114, 91.
doi:10.1037/0096-3445.114.1.91.
Ferreira, F., Apel, J., Henderson, J.M., 2008. Taking a new look at looking at nothing.
Trends in Cognitive Sciences 12, 405 – 410. doi:10.1016/j.tics.2008.07.
007.
Gallup, A.C., Hale, J.J., Sumpter, D.J.T., Garnier, S., Kacelnik, A., Krebs, J.R., Couzin,
I.D., 2012. Visual attention and the acquisition of information in human crowds.
Proceedings of the National Academy of Sciences 109, 7245–7250. doi:10.1073/
pnas.1116141109.
Gower, J.C., Dijksterhuis, G.B., 2004. Procrustes problems. volume 30. Oxford Uni-
versity Press on Demand.
Harrison, S.A., Tong, F., 2009. Decoding reveals the contents of visual working mem-
ory in early visual areas. Nature 458, 632. doi:10.1038/nature07832.
Hollingworth, A., Hwang, S., 2013. The relationship between visual working memory
and attention: retention of precise colour information in the absence of effects on
perceptual selection. Philosophical Transactions of the Royal Society of London B:
Biological Sciences 368. doi:10.1098/rstb.2013.0061.
35
Hollingworth, A., Matsukura, M., Luck, S.J., 2013. Visual working memory modulates
rapid eye movements to simple onset targets. Psychological Science 24, 790–796.
doi:10.1177/0956797612459767. pMID: 23508739.
Holmqvist, K., Nystrom, M., Andersson, R., Dewhurst, R., Halszka, J., van de Weijer,
J., 2011. Eye Tracking: A Comprehensive Guide to Methods and Measures. Oxford
University Press.
Huang, X., Shen, C., Boix, X., Zhao, Q., 2015. Salicon: Reducing the semantic gap in
saliency prediction by adapting deep neural networks, in: Proceedings of the IEEE
International Conference on Computer Vision, pp. 262–270.
Hutton, S.B., Nolte, S., 2011. The effect of gaze cues on attention to print advertise-
ments. Applied Cognitive Psychology 25, 887–892. doi:10.1002/acp.1763.
Illian, J., Penttinen, A., Stoyan, H., Stoyan, D., 2008. Statistical analysis and modelling
of spatial point patterns. volume 70. John Wiley & Sons.
Isola, P., Xiao, J., Parikh, D., Torralba, A., Oliva, A., 2014. What makes a photograph
IEEE transactions on pattern analysis and machine intelligence 36,
memorable?
1469–1482.
Itti, L., Koch, C., 2000. A saliency-based search mechanism for overt and covert shifts
of visual attention. Vision research 40, 1489–1506.
Jiang, M., Huang, S., Duan, J., Zhao, Q., 2015. Salicon: Saliency in context, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pp. 1072–1080.
Johansson, R., Holsanova, J., Dewhurst, R., Holmqvist, K., 2012. Eye movements dur-
ing scene recollection have a functional role, but they are not reinstatements of those
produced during encoding. Journal of Experimental Psychology: Human Perception
and Performance 38, 1289–1314.
Johansson, R., Holsanova, J., Holmqvist, K., 2006.
scriptions elicit similar eye movements during mental
Pictures and spoken de-
imagery, both in light
36
and in complete darkness. Cognitive Science 30, 1053–1079. doi:10.1207/
s15516709cog0000_86.
Johansson, R., Johansson, M., 2014. Look here, eye movements play a functional role
in memory retrieval. Psychological Science 25, 236–242.
Judd, T., Durand, F., Torralba, A., 2012. A benchmark of computational models of
saliency to predict human fixations. Technical Report MIT-CSAIL-TR-2012-001.
MIT.
Judd, T., Ehinger, K., Durand, F., Torralba, A., 2009. Learning to predict where humans
look, in: Computer Vision, 2009 IEEE 12th international conference on, IEEE. pp.
2106–2113.
Kaspar, K., Koenig, P., 2011. Viewing behavior and the impact of low-level image
properties across repeated presentations of complex scenes. Journal of Vision 11,
26–26.
Khosla, A., Raju, A.S., Torralba, A., Oliva, A., 2015. Understanding and predicting
image memorability at a large scale, in: Proceedings of the IEEE International Con-
ference on Computer Vision, pp. 2390–2398.
Kienzle, W., Franz, M.O., Scholkopf, B., Wichmann, F.A., 2009. Center-surround
patterns emerge as optimal predictors for human saccade targets. Journal of vision
9, 7–7.
Kim, N.W., Bylinskii, Z., Borkin, M.A., Gajos, K.Z., Oliva, A., Durand, F., Pfister, H.,
2017. Bubbleview: an alternative to eye-tracking for crowdsourcing image impor-
tance. arXiv preprint arXiv:1702.05150 .
Kim, N.W., Bylinskii, Z., Borkin, M.A., Oliva, A., Gajos, K.Z., Pfister, H., 2015. A
crowdsourced alternative to eye-tracking for visualization understanding, in: Pro-
ceedings of the 33rd Annual ACM Conference Extended Abstracts on Human Fac-
tors in Computing Systems, ACM. pp. 1349–1354.
37
Koehler, K., Guo, F., Zhang, S., Eckstein, M.P., 2014. What do saliency models pre-
dict? Journal of vision 14, 14–14. doi:10.1167/14.3.14.
Koivisto, M., Hyona, J., Revonsuo, A., 2004. The effects of eye movements, spatial
attention, and stimulus features on inattentional blindness. Vision Research 44, 3211
– 3221. doi:10.1016/j.visres.2004.07.026.
Krieger, G., Rentschler, I., Hauske, G., Schill, K., Zetzsche, C., 2000. Object and scene
analysis by saccadic eye-movements: an investigation with higher-order statistics.
Spatial Vision 13, 201–214. doi:10.1163/156856800741216.
Kruthiventi, S.S., Ayush, K., Babu, R.V., 2015. Deepfix: A fully convolutional neural
network for predicting human eye fixations. arXiv preprint arXiv:1510.02927 .
Laeng, B., Bloem, I.M., DAscenzo, S., Tommasi, L., 2014. Scrutinizing visual images:
The role of gaze in mental imagery and memory. Cognition 131, 263–283.
Laeng, B., Teodorescu, D.S., 2002. Eye scanpaths during visual imagery reenact those
of perception of the same visual scene. Cognitive Science 26, 207–231. doi:10.
1207/s15516709cog2602_3.
Levin, D., 1998. The approximation power of moving least-squares. Math. Comput.
67, 1517–1531. doi:10.1090/S0025-5718-98-00974-0.
Loftus, G., Mackworth, N., 1978. Cognitive determinants of fixation location dur-
ing picture viewing. Journal of Experimental Psychology: Human Perception and
Performance 4, 565–572. doi:10.1037/0096-1523.4.4.565.
Mantyla, T., Holm, L., 2006. Gaze control and recollective experience in face recogni-
tion. Visual Cognition 14, 365–386.
Martarelli, C.S., Chiquet, S., Laeng, B., Mast, F.W., 2017. Using space to repre-
sent categories: insights from gaze position. Psychological Research 81, 721–729.
doi:10.1007/s00426-016-0781-2.
Martarelli, C.S., Mast, F.W., 2013. Eye movements during long-term pictorial recall.
Psychological Research 77, 303–309. doi:10.1007/s00426-012-0439-7.
38
Matin, E., 1974. Saccadic suppression: a review and an analysis. Psychological bulletin
81, 899.
Memmert, D., 2006. The effects of eye movements, age, and expertise on inatten-
tional blindness. Consciousness and Cognition 15, 620 – 627. doi:10.1016/j.
concog.2006.01.001.
Moore, C.S., 1903. Control of the memory image. The Psychological Review: Mono-
graph Supplements .
Olivers, C.N., 2008. Interactions between visual working memory and visual attention.
Frontiers in Bioscience 13, 1182–1191. doi:10.2741/2754.
Olivers, C.N., Meijer, F., Theeuwes, J., 2006. Feature-based memory-driven at-
tentional capture: visual working memory content affects visual attention. Jour-
nal of Experimental Psychology: Human Perception and Performance 32, 1243.
doi:10.1037/0096-1523.32.5.1243.
Oztireli, A.C., Gross, M., 2012. Analysis and synthesis of point distributions based on
pair correlation. ACM Trans. Graph. 31, 170:1–170:10. doi:10.1145/2366145.
2366189.
Pannasch, S., Helmert, J.R., Roth, K., Herbold, A.K., Walter, H., 2008. Visual fixation
durations and saccade amplitudes: Shifting relationship in a variety of conditions.
Journal of Eye Movement Research 2.
Parkhurst, D., Law, K., Niebur, E., 2002. Modeling the role of salience in the allo-
cation of overt visual attention. Vision Research 42, 107 – 123. doi:10.1016/
S0042-6989(01)00250-4.
Pathman, T., Ghetti, S., 2015. Eye movements provide an index of veridical mem-
ory for temporal order. PLOS ONE 10, 1–17. doi:10.1371/journal.pone.
0125648.
Ramanathan, S., Katti, H., Sebe, N., Kankanhalli, M., Chua, T.S., 2010. An eye
fixation database for saliency detection in images, in: Daniilidis, K., Maragos,
39
P., Paragios, N. (Eds.), Computer Vision – ECCV 2010: 11th European Con-
ference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010,
Proceedings, Part IV, Springer Berlin Heidelberg, Berlin, Heidelberg. pp. 30–43.
doi:10.1007/978-3-642-15561-1_3.
Recasens, A., Khosla, A., Vondrick, C., Torralba, A., 2015. Where are they looking?,
in: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (Eds.), Ad-
vances in Neural Information Processing Systems 28. Curran Associates, Inc., pp.
199–207.
Ricciardelli, P., Bricolo, E., Aglioti, S.M., Chelazzi, L., 2002. My eyes want to look
where your eyes are looking: Exploring the tendency to imitate another individual's
gaze. NeuroReport 13, 2259–2264.
Schaefer, S., McPhail, T., Warren, J., 2006.
Image deformation using moving least
squares, in: ACM transactions on graphics (TOG), ACM. pp. 533–540.
Schmidt, B.K., Vogel, E.K., Woodman, G.F., Luck, S.J., 2002. Voluntary and automatic
attentional control of visual working memory. Perception & Psychophysics 64, 754–
763. doi:10.3758/BF03194742.
Scholz, A., von Helversen, B., Rieskamp, J., 2015. Eye movements reveal memory
processes during similarity- and rule-based decision making. Cognition 136, 228 –
246. doi:10.1016/j.cognition.2014.11.019.
Scholz, A., Mehlhorn, K., Krems, J.F., 2016. Listen up, eye movements play a role
in verbal memory retrieval. Psychological Research 80, 149–158. doi:10.1007/
s00426-014-0639-4.
Schomaker, J., Walper, D., Wittmann, B.C., Einhauser, W., 2017. Attention in natural
scenes: Affective-motivational factors guide gaze independently of visual salience.
Vision Research 133, 161 – 175. doi:10.1016/j.visres.2017.02.003.
Simons, D.J., 2000. Attentional capture and inattentional blindness. Trends in Cogni-
tive Sciences 4, 147 – 155. doi:10.1016/S1364-6613(00)01455-8.
40
Sorkine, O., Alexa, M., 2007. As-rigid-as-possible surface modeling, in: Proceed-
ings of the Fifth Eurographics Symposium on Geometry Processing, Eurographics
Association, Aire-la-Ville, Switzerland, Switzerland. pp. 109–116.
Sorkine-Hornung, O., Rabinovich, M., 2016. Least-squares rigid motion using svd.
Technical note.
Tatler, B.W., 2007. The central fixation bias in scene viewing: Selecting an opti-
mal viewing position independently of motor biases and image feature distributions.
Journal of Vision 7, 4. doi:10.1167/7.14.4.
Tatler, B.W., Hayhoe, M.M., Land, M.F., Ballard, D.H., 2011. Eye guidance in natural
vision: Reinterpreting salience. Journal of vision 11, 5–5. doi:10.1167/11.5.5.
Tavakoli, H.R., Ahmed, F., Borji, A., Laaksonen, J., 2017. Saliency revisited: Analysis
of mouse movements versus fixations. arXiv preprint arXiv:1705.10546 .
Trukenbrod, H.A., Barthelm´e, S., Wichmann, F.A., Engbert, R., 2017. Rigorous spatial
statistics for gaze patterns in scene viewing: Effects of repeated viewing. arXiv
preprint arXiv:1704.01761 .
Van Der Linde, I., Rajashekar, U., Bovik, A., Cormack, L., 2009. Doves: a
database of visual eye movements. Spatial Vision 22, 161–177. doi:10.1163/
156856809787465636.
de Vito, S., Buonocore, A., Bonnefon, J.F., Della Sala, S., 2014. Eye movements
disrupt spatial but not visual mental imagery. Cognitive processing 15, 543–549.
Vo, M.L.H., Aizenman, A.M., Wolfe, J.M., 2016. You think you know where you
looked? you better look again. Journal of experimental psychology: human percep-
tion and performance 42, 1477.
Wantz, A.L., Martarelli, C.S., Mast, F.W., 2016. When looking back to noth-
ing goes back to nothing. Cognitive Processing 17, 105–114. doi:10.1007/
s10339-015-0741-6.
41
Xu, J., Jiang, M., Wang, S., Kankanhalli, M.S., Zhao, Q., 2014. Predicting human gaze
beyond pixels. Journal of vision 14, 28–28.
42
|
1512.08309 | 2 | 1512 | 2016-08-16T19:21:32 | Identifying Seizure Onset Zone from the Causal Connectivity Inferred Using Directed Information | [
"q-bio.NC",
"cs.IT",
"cs.IT",
"stat.ME"
] | In this paper, we developed a model-based and a data-driven estimator for directed information (DI) to infer the causal connectivity graph between electrocorticographic (ECoG) signals recorded from brain and to identify the seizure onset zone (SOZ) in epileptic patients. Directed information, an information theoretic quantity, is a general metric to infer causal connectivity between time-series and is not restricted to a particular class of models unlike the popular metrics based on Granger causality or transfer entropy. The proposed estimators are shown to be almost surely convergent. Causal connectivity between ECoG electrodes in five epileptic patients is inferred using the proposed DI estimators, after validating their performance on simulated data. We then proposed a model-based and a data-driven SOZ identification algorithm to identify SOZ from the causal connectivity inferred using model-based and data-driven DI estimators respectively. The data-driven SOZ identification outperforms the model-based SOZ identification algorithm when benchmarked against visual analysis by neurologist, the current clinical gold standard. The causal connectivity analysis presented here is the first step towards developing novel non-surgical treatments for epilepsy. | q-bio.NC | q-bio | Identifying Seizure Onset Zone from the Causal
Connectivity Inferred Using Directed Information
Rakesh Malladi, Student Member, IEEE, Giridhar Kalamangalam, Nitin Tandon,
and Behnaam Aazhang, Fellow, IEEE
1
6
1
0
2
g
u
A
6
1
]
.
C
N
o
i
b
-
q
[
2
v
9
0
3
8
0
.
2
1
5
1
:
v
i
X
r
a
Abstract -- In this paper, we developed a model-based and a
data-driven estimator for directed information (DI) to infer the
causal connectivity graph between electrocorticographic (ECoG)
signals recorded from brain and to identify the seizure onset zone
(SOZ) in epileptic patients. Directed information, an information
theoretic quantity, is a general metric to infer causal connectivity
between time-series and is not restricted to a particular class of
models unlike the popular metrics based on Granger causality
or transfer entropy. The proposed estimators are shown to be
almost surely convergent. Causal connectivity between ECoG
electrodes in five epileptic patients is inferred using the proposed
DI estimators, after validating their performance on simulated
data. We then proposed a model-based and a data-driven SOZ
identification algorithm to identify SOZ from the causal connec-
tivity inferred using model-based and data-driven DI estimators
respectively. The data-driven SOZ identification outperforms the
model-based SOZ identification algorithm when benchmarked
against visual analysis by neurologist, the current clinical gold
standard. The causal connectivity analysis presented here is the
first step towards developing novel non-surgical treatments for
epilepsy.
Index Terms -- Epilepsy, directed information, seizure onset
zone, ECoG, causal connectivity.
I. INTRODUTION
Epilepsy is a common neurological disease affecting nearly
1% of the world's population. Epilepsy is characterized by
unprovoked seizures, which are periods of hypersynchronous
activity in the brain. The current treatment options include
medication, resective surgery and more recently, electrical
stimulation approaches like vagus nerve and responsive neuro
stimulation. However, medication is not able to stop seizures
in about one-third of the patients. The efficacy of the other
current neuro-modulation approaches is variable and almost
never results in a cure [3], [4]. The current approaches lack
specificity and suffer from negative side effects ( [5] and ref-
erences therein). Selective modulation of the epileptic circuits
in the brain via electrical stimulation [6], optogenetics and
designer receptive technologies [5] represent possible options
Copyright (c) 2016 IEEE. Personal use of this material
is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to [email protected].
This work is funded in part by grant 1406447 from National Science
Foundation and Texas Instruments. A portion of this work was presented
at Society of Neuroscience (SfN) 2014 [1] and International Conference on
Acoustics, Speech and Signal Processing (ICASSP) 2015 [2].
Rakesh Malladi and Behnaam Aazhang are with the Department of Elec-
trical and Computer Engineering, Rice University, Houston, TX, 77005 USA.
Giridhar Kalamangalam and Nitin Tandon are with Department of Neu-
rology and Department of Neurosurgery,
respectively, at University of
Texas Health Center, Houston, TX, 77005. E-mail: {rm17, aaz}@rice.edu,
{Giridhar.P.Kalamangalam, Nitin.Tandon}@uth.tmc.edu.
for better treatments for this disabling disease. A crucial initial
step in this endeavor is understanding how seizures originate
and spread within the brain. Effective or causal connectivity
[7] quantifies how the activity spreads between different brain
regions and can be used to characterize epileptic networks.
In addition, causal connectivity can also be used to identify
seizure onset zone (SOZ) (brain regions initiating seizures
[8]) and has been shown to predict the efficacy of resective
surgery [9], [10]. The main objective of this paper is to
develop estimators of directed information (DI), derive causal
connectivity between brain regions and identify the SOZ using
electrocorticographic (ECoG) data in patients with epilepsy.
Estimating causal connectivity from electrophysiological
recordings of brain has been the focus of many papers. A
good summary is provided in [9]. The causality referred to in
this paper is in the Wiener-Granger causal sense [11]. Metrics
based on Granger causality (GC) [12], [13] and information
theory like transfer entropy [14] are commonly used to es-
timate causal connectivity between continuous-valued ECoG
recordings. However, these techniques are well-suited only
for a specific model and subset of the recorded signals. For
instance, GC-based measures are applicable only for data from
multivariate autoregressive (MVAR) processes. ECoG data
are often modeled using linear MVAR model [9], [13], even
though associations between seizure loci detected by ECoG
recordings are likely nonlinear [15], [16].
We propose to develop a causal metric that would be
applicable to diverse models and different types of electro-
physiological recordings of brain. Directed information, used
to infer causal connections between spike trains in [17] -- [19],
can indeed be further developed into a general technique to
estimate causal connectivity. The definition of DI is based on
the underlying probability distribution and no assumptions are
imposed on the underlying distributions. Directed information
was developed for discrete-valued time-series in [20] -- [22]
and nonparametrically estimated in [23]. DI quantifies the
amount of causal information about one time-series that is
explained by the other time-series [17]. Modified time-lagged
directed information is proposed in [24], [25] to reduce the
computational complexity of estimating directed information.
DI is also used in many other applications [26] -- [28]. The
definition of DI is broadened to the class of continuous-valued
processes like ECoG signals in this paper [1], [24], [25], [29].
If the data is assumed to be from a MVAR process with
Gaussian white noise, DI is equivalent to Granger causality
[29] and if the data satisfies the general Markov condition, DI
is very closely related to transfer entropy [24], [25]. The main
2
advantage of DI over other existing techniques is that DI is
applicable to a large class of electrophysiological recordings
from brain, including spike trains, EEG and ECoG, and is not
restricted to a particular class of models.
We developed an almost surely convergent model-based
and data-driven estimators of DI in this paper, inspired by
prior work [17], [30]. The performance of the proposed DI
estimators was validated on linear and nonlinear simulated
models and compared with the Granger causality metric [12],
[31]. The statistical significance of the causal connection
inferred using DI and GC estimates was demonstrated using an
adaptation [32] of stationary bootstrap [33]. We then applied
both model-based and data-driven DI estimators to infer causal
connectivity graphs between ECoG channels from twelve
seizures in five patients with epilepsy.
The DI metric with model-based and data-driven estimators
proposed in this paper allows us the flexibility to simultane-
ously use both these estimators and identify which one leads
to more reliable causal connectivity graphs from real ECoG
data. This would also allows us to examine the appropriateness
of imposing linear MVAR assumption on ECoG data. We used
the model-based DI estimator with MVAR model assumption
to detect the linear causal interactions and the data-driven DI
estimator to detect both linear and nonlinear causal interactions
between ECoG channels. We observed that nonlinear causal
interactions between channels are stronger around the onset of
a seizure, as widely believed [16].
We then proposed a model-based and a data-driven SOZ
identification algorithm to identify SOZ from the causal con-
nectivity graphs inferred using model-based and data-driven
DI estimators respectively. The SOZ identified by model-based
and data-driven algorithms are respectively the isolated nodes
and strong sources in the corresponding causal connectivity
graphs. Despite the numerous SOZ identification algorithms
available [9], [10], [34] -- [36], the current clinical gold standard
is still the visual analysis of ECoG data by the neurologist. We
therefore compared the performance of both model-based and
data-driven SOZ identification algorithms with visual analysis
by the neurologist. We find that
the data-driven approach
outperforms the model-based approach and also leads to more
interpretable results. The methodology proposed here should
be extended to analyze the whole ECoG record to better
understand the evolution of seizure mechanisms over time.
The main algorithmic contributions of this paper are
• Developing an almost surely convergent model-based and
data-driven DI estimator, described in sections III and IV.
• Developing a MVAR model-based and a data-driven SOZ
identification algorithm, described in section VI.
II. DIRECTED INFORMATION
Consider the N samples recorded at a sampling frequency
Fs from each ECoG electrode implanted in an epileptic
patient. Without
let us focus on two
channels X and Y. The N samples recorded from the two
channels are denoted by XN = (x1, x2,··· , xN )T and YN =
(y1, y2,··· , yN )T. Also let W denote the matrix of samples
recorded from a group of channels excluding X and Y. For
loss of generality,
notational simplicity, the elements corresponding to the non-
positive subscripts are treated as empty sets and the subscripts
are not shown when equal to 1.
The directed information, I(cid:0)XN → YN(cid:1), from N samples
of continuous-valued random process X to those of Y is
defined as
I(cid:0)XN → YN(cid:1) = h(cid:0)YN(cid:1) − h(cid:0)YN(cid:107)XN(cid:1) ,
where h(cid:0)YN(cid:1) is the differential entropy of the N-dimensional
continuous random vector YN [30] and h(cid:0)YN(cid:107)XN(cid:1) is the
(1)
causally conditioned differential entropy of YN causally con-
ditioned on XN . The causally conditioned differential entropy
is defined as
h(cid:0)YN(cid:107)XN(cid:1) =
N(cid:80)
h(cid:0)ynYn−1, Xn(cid:1) .
(2)
n=1
of bits of uncertainty in one process that is causally explained
The definitions of DI and causally conditioned differential en-
tropy in (1) and (2) are obtained by broadening the definitions
of the same quantities from discrete-time, discrete-valued ran-
dom processes [21], [22] to discrete-time, continuous-valued
processes [2]. One of the main differences between discrete-
valued and continuous-valued random processes is that the
entropy of a discrete-valued process is always non-negative,
whereas the differential entropy of a continuous-valued pro-
cess can be negative [30]. However, DI is always non-negative
since conditioning cannot increase differential entropy [30],
i.e., I(cid:0)XN → YN(cid:1) ≥ 0. DI can be interpreted as the number
away by the other process. If I(cid:0)XN → YN(cid:1) = 0, then there
metric in general, i.e., I(cid:0)XN → YN(cid:1) (cid:54)= I(cid:0)YN → XN(cid:1).
mutual information [30], I(cid:0)yn; XnYn−1(cid:1), as
I(cid:0)XN → YN(cid:1) =
(cid:8)h(cid:0)ynYn−1(cid:1) − h(cid:0)ynYn−1, Xn(cid:1)(cid:9)
I(cid:0)yn; XnYn−1(cid:1) .
N I(cid:0)XN→ YN(cid:1)
N h(cid:0)YN(cid:1)−limN→∞ 1
N(cid:80)
N(cid:80)
I(cid:0)X→ Y(cid:1) =limN→∞ 1
Note that DI can also be expressed in terms of conditional
is no causal influence from X to Y. The DI is not a symmetric
Now, the DI between the time-series X and Y is defined as
N h(cid:0)YN(cid:107)XN(cid:1)
(3)
n=1
n=1
=
= h(cid:0)Y(cid:1)−h(cid:0)Y(cid:107)X(cid:1),
=limN→∞ 1
(4)
provided the limits exist. h (Y) and h (Y(cid:107)X) are respectively
the differential entropy of Y and the causally conditioned
differential entropy of Y given X. The DI from Y to X is
also similarly defined.
Furthermore, the DI defined earlier is easily extended to
define directed information from X to Y, causally conditioned
on W. Note that W comprises the samples recorded from a
group of ECoG channels excluding X and Y. The causally
conditioned DI, I (X → Y(cid:107)W), is defined as
I (X → Y(cid:107) W) = limN→∞ 1
N I(cid:0)XN → YN(cid:107)WN(cid:1)
(cid:8)h(cid:0)YN(cid:107)WN(cid:1)−h(cid:0)YN(cid:107)XN,WN(cid:1)(cid:9) ,
= limN→∞ 1
N
= h (Y(cid:107)W) − h (Y(cid:107)W, X),
(5)
where h(cid:0)YN(cid:107)XN, WN(cid:1) =
N(cid:80)
h(cid:0)ynYn−1, Xn, Wn(cid:1) is the
n=1
differential entropy of YN causally conditioned on XN and
WN , h (Y(cid:107)W) is the causal conditioned differential entropy
of Y given W and h (Y(cid:107)W, X) is the causally conditioned
differential entropy of Y given the causal past of W and
X. We use the directed information defined here to learn
the causal connectivity graph between all ECoG channels and
identify the SOZ of epileptic patients.
III. UNIVERSAL ESTIMATOR FOR DIRECTED
INFORMATION
A universal estimator for directed information between
channels X and Y, I (X → Y), and the causally conditioned
DI, I (X → Y(cid:107)W) is developed in this section. The proposed
estimator is universal and is shown to be almost surely
convergent assuming that the causal conditional likelihood
(CCL) is known. If CCL is not known and is estimated, then
the convergence of the proposed DI estimator is dependent on
the CCL estimator. The ideas used in developing the proposed
DI estimator are inspired by prior work [17], [30]. Without loss
of generality, we will first focus on estimating the pairwise
DI, I (X → Y). We will then outline the procedure to extend
this pairwise DI estimator to estimate the causally conditioned
DI, I (X → Y(cid:107)W). The inputs to the proposed pairwise DI
estimator are the observed N samples of time-series X and
Y. The main idea is to develop an almost surely convergent
estimator for the entropies in (4) and the difference between
the two entropy estimates is an almost surely convergent
estimate for I (X → Y). Let us first focus on the causally
conditioned differential entropy estimator, h (Y(cid:107)X).
Assumption 1 - The random processes X and Y are assumed
to be stationary, ergodic and Markovian in the observed time-
window. These are reasonable assumptions to model ECoG
data. First, an implicit assumption in the problem of estimating
the causal connectivity from a ECoG data segment is that
the causal connectivity does not vary in this segment, which
is mathematically captured by stationarity. The entire ECoG
data record is usually not stationary and stationary segments
are identified using either sliding windows [37] or change-
point detection algorithms [38]. We used the sliding window
approach in this paper. A crucial parameter in this process
is the length of the window in which data is assumed to be
stationary. It is also important to realize that we need a mini-
mum amount of data points to reliably estimate any unknown
parameters involved. It is recommended that the number of
data points should be much larger (as a thumb rule, at least
an order of magnitude larger) than the number of parameters
to be estimated [37]. Directed information is then estimated
in each stationary segment using the algorithm proposed in
this section. Directed information for the entire time-series is
the sum of the DI estimates from each stationary segment and
is interpreted as the total amount of uncertainty in one time-
series in the entire recording window that is explained by the
other time-series. Second, ergodicity is required to ensure that
the estimates from long-enough recording windows converge
to the true value. Finally, the Markovian assumption captures
the dependence of the current activity on the past activity
3
at different electrodes. Let the current sample of the time-
series Y depend on the past Jyy and past Kyx samples of
the time-series Y and X respectively. Note that (Jyy, Kyx)
are unknown and should be estimated from data. The explicit
model of the dependence is captured by the causal likelihood
of yn conditioned on the past activity at electrodes X and
Y. This CCL is denoted by P(cid:0)ynYn−1
(cid:1) and
, Xn
can be estimated using either a model-based or a data-driven
approach. Let us assume for now that CCL is known.
n−Kyx+1
n−Jyy
, Xl
(cid:16)
and Xl
ylYl−1
l−Jyy
(cid:17) ∈ R.
Assumption 2 - Let us also assume that differential entropy
of the first sample, y1, of time-series Y exists and that for
some time-index l ∈ [1, N ], the conditional differential entropy
of yl, conditioned on Yl−1
l−Kyx+1 also exists, i.e.,
l−Jyy
h (y1) , h
l−Kyx+1
Lemma 3.1. Let Assumptions 1 and 2 hold. Then h (Y) ,
h (Y(cid:107)X) , I (X → Y) exists and are in R.
Proof. Stationarity and the property that conditioning cannot
increase the differential entropy are the main ideas in the proof,
which is in the Appendix A.
Lemma 3.2. Let Assumptions 1 and 2 hold. Then for some
time-index l
N h(cid:0)YN(cid:107)XN(cid:1) =E(cid:104)−log P
(cid:16)
ylYl−1
l−Jyy
l−(Kyx−1)
(cid:17)(cid:105)
.
(6)
, Xl
1
Proof. The proof uses the definition of causally conditioned
differential entropy (2), the Markovian and the stationarity
assumptions. The proof is in the Appendix A.
Theorem 3.1. Let Assumptions 1 and 2 hold. Then the almost
surely convergent causally conditioned differential entropy
estimator is
h (Y(cid:107)X) = 1
(cid:110)−log P
(cid:16)
(cid:17)(cid:111)
N(cid:80)
, Xn
. (7)
ynYn−1
n−Jyy
n−(Kyx−1)
N
n=1
Proof. The proof is based on two observations: the first is
that the right-hand side of (6) does not depend on N and
therefore it is easy to compute its limit as N → ∞. The
second observation is that the strong law of large numbers
(SLLN) for Markov chains [39] can be applied to estimate the
expectation on the right-hand side of (6). The detailed proof
is in the Appendix A.
An almost surely convergent estimator for h (Y) can be
easily derived using Theorem 3.1, simply by modeling the
dependence of the current samples of Y on its own J(cid:48)
yy past
samples. This is equivalent to setting Kyx = 0. The difference
between the two estimators, h (Y) and h (Y(cid:107)X), is the almost
surely convergent estimator for DI from X to Y, I (X → Y).
This is stated in Theorem 3.2.
Theorem 3.2. Let Assumptions 1 and 2 hold. The universal
estimator for DI from time-series X to Y is
I (X → Y) = h (Y) − h (Y(cid:107)X) a.s.−−→ I (X → Y) .
(8)
Proof. We have from Theorem 3.1 I (X → Y) = h (Y) −
h (Y(cid:107)X) a.s.−−→ h (Y) − h (Y(cid:107)X) = I (X → Y) .
The DI estimator in Theorem 3.2 can be easily extended
to estimate the causally conditioned directed information,
4
(cid:16)
I (X → Y(cid:107)W). First, h (Y(cid:107)W) is estimated using Theo-
rem 3.1. We now need to estimate h (Y(cid:107)W, X). Let Jyy,
Kyw and Kyx respectively denote the number of past sam-
ples of Y, W and X that
influence the current sample
of Y. Let us also assume the causal conditional likelihood
is known. A model-
P
based and a data-driven approach to estimate this CCL is
described in the subsequent section. Then Theorem. 3.1 can
be easily extended to show that
ynYn−1
n−Jyy
n−Kyw+1, Xn
n−Kyx+1
, Wn
(cid:17)
h (Y(cid:107)W, X)
N(cid:80)
(cid:110)−log P
(cid:16)
= 1
N
n=1
ynYn−1
n−Jyy
,Wn
n−Kyw+1,Xn
n−Kyx+1
(cid:17)(cid:111)
(9)
is important
there is no causal
information X contains about Y. It
is an almost surely convergent estimate of h (Y(cid:107)W, X). From
(5), I (X → Y(cid:107)W) is the difference between the estimates,
h (Y(cid:107)W) and h (Y(cid:107)W, X). It
to note that
the
as the number of channels included in W increases,
computational complexity of the estimator also increases.
The DI estimate, I (X → Y), can be interpreted as the
amount of causal
is,
however, important to note that I (X → Y) is estimated from
N samples and is an estimate of the true value of DI from
X to Y. The statistical significance of the causal connection
from X to Y inferred from I (X → Y) is calculated using
an adaptation [32] of stationary bootstrap [33]. B stationary
bootstrap samples of X, denoted by X(b), are generated using
the algorithm described in [32] for b = 1, 2,··· , B. The DI
from bth stationary bootstrap sample X(b) to Y, denoted by
I (X(b) → Y), is estimated using the proposed DI estimator.
Note that
influence from any of these
bootstrap samples to Y by construction. Therefore the B
samples, I (X(b) → Y), for b = 1, 2,··· , B are from the null
hypothesis of no causal influence. The statistical significance
is determined by the P-value [40]. P-value is the probability
that DI estimate greater than or equal to I (X → Y) can
be observed under the null hypothesis of no causality from
X to Y and is computed from the empirical distribution of
I (X(b) → Y) for b = 1, 2,··· , B. If the P-value is less than
a predetermined significance level δ, the null hypothesis of no
causal connection from X to Y is rejected. On the other hand,
if the P-value is greater than δ, the null hypothesis cannot
be rejected and the causal connection from X to Y is not
statistically significant. Note that the empirical distribution of
I (X(b) → Y) is concentrated around 0, since the DI between
time-series that are not causally connected is zero. Therefore,
when the actual DI estimate is large enough, the P-value will
be less than δ and the statistical significance assessment is
not required. However, statistical significance assessment is
useful when the DI estimate is close to zero. The significance
assessment described here is applied to the simulated examples
in section V to identify the significant causal connections,
particularly useful when the DI estimates are close to zero.
The above discussion assumes CCL is known. The likelihood,
however, must be estimated from data in practice. A model-
based and a data-driven approach to estimate CCL is described
in the following section.
IV. ESTIMATING CAUSAL CONDITIONAL LIKELIHOOD
Estimating DI from X to Y using the proposed DI estimator
in section III requires estimating two CCLs, P(cid:0)ynYn−1(cid:1)
and P(cid:0)ynYn−1, Xn(cid:1), while estimating DI from Y to X
P(cid:0)ynYn−1, Wn(cid:1) and P(cid:0)ynYn−1, Wn, Xn(cid:1). Let us focus
on estimating P(cid:0)ynYn−1, Xn(cid:1) for n = 1, 2,··· , N, which
to extend this approach to estimate P(cid:0)ynYn−1, Wn, Xn(cid:1).
is required to estimate h (Y(cid:107)X). We will then describe how
causally conditioned on W requires estimating two CCLs,
The CCLs are estimated using either model-based or data-
driven techniques. The choice between model-based and data-
driven approaches is determined by the application from
which data is recorded. For instance, the time-series signals
obtained from electrophysiological recordings of brain or from
stock markets are commonly modeled using MVAR models
with Gaussian white noise. In this case, the CCL is easily
estimated from the MVAR model of the data. Usually the
parameters of the model are unknown and several classical
techniques to estimate the unknown parameters are described
in [41]. On the other hand, using model-based approaches to
estimate CCLs from data recorded from nonlinear systems or
systems without a prescribed linear model is non-trivial. This
is because estimating the CCLs using model-based approach
requires essentially inverting the nonlinear generative model,
which is not trivial. Data-driven approaches do not have this
limitation and are therefore preferred for nonlinear time-series
data. A good review of the various data-driven algorithms that
estimate probability distribution from data is provided in [42],
[43]. The model-based and the data-driven CCL algorithm
used in this paper are described in the remainder of this
section.
A. Model-based CCL Estimation
We will focus on estimating the CCL specifically for
multivariate autoregressive process with Gaussian white noise
in this paper. Let the time-series X and Y be sampled from
such processes. Then, the samples of Y can be expressed as
Jyy(cid:80)
Kyx(cid:80)
yn =
αjyn−j +
βkxn−k+1 +zn,n = 1, 2,··· ,N,
(10)
j=1
k=1
where zn is the additive white Gaussian noise with zero mean
z. Here αj for j = 1, 2,··· , Jyy and βk for k =
and variance σ2
1, 2,··· , Kyx are the parameters of the model and Jyy and
Kyx are the model orders representing how many past samples
of Y and of X respectively influence the current sample of
Y. It is easy to observe from (10) that
P(cid:0)ynYn−1,Xn(cid:1)∼N(cid:16)Jyy(cid:80)
(cid:17)
vector θ (Jyy, Kyx) =(cid:0)α1,··· , αJyy , β1,··· , βKyx, σ2
The two model orders, Jyy and Kyx, and the parameter
(cid:1)T are
βkxn−k+1, σ2
z
Kyx(cid:80)
αjyn−j +
not known apriori and need to be estimated from the N
observed samples of X and Y. The parameters and the
model orders are estimated using a maximum likelihood (ML)
estimator with minimum description length [44] penalty. ML
estimator is known to be asymptotically consistent. Mini-
mum description length is a model order selection procedure
(11)
k=1
j=1
.
z
to
solutions of the following problem:
(cid:1) = arg min
with good consistency properties [44] and proportional
(Jyy + Kyx). The optimal model orders (cid:0) Jyy, Kyx
(cid:1) are the
N logP(cid:0)YN(cid:107)XN;θ(cid:0)Jyy, Kyx
(cid:1)(cid:1)
(cid:0) Jyy, Kyx
(cid:8)− 1
log N(cid:9),
negative log-likelihood for a given(cid:0)Jyy, Kyx
(cid:1) and is obtained
N logP(cid:0)YN(cid:107)XN;θ (Jyy,Kyx)(cid:1) . (13)
(12)
where θ (Jyy, Kyx) is the value of θ which minimizes the
by solving
θ (Jyy, Kyx)=arg min
+ Jyy+Kyx
(Jyy,Kyx)
− 1
2N
θ
The ML estimation of θ for a given (Jyy, Kyx) in (13)
is equivalent
to the ML estimation of the parameters of
a standard linear regression model [41], since the CCL is
Gaussian distributed (11). The estimated parameters almost
surely converge to the true parameter values [45] resulting
in almost surely convergence of the proposed DI estimator.
The desired CCL is obtained by substituting the solutions
of (13), (12) in (11). The resultant CCL is then substituted
in (7) to estimate h (Y(cid:107)X), which is further simplified to
h (Y(cid:107)X) = 1
z is the estimate of the
noise variance from (12), (13).
2 log(cid:0)2πeσ2
(cid:1), where σ2
Let us focus on estimating P(cid:0)ynYn−1, Wn, Xn(cid:1), which is
The MVAR model-based CCL estimation algorithm de-
scribed above can be easily extended to estimate the CCLs re-
quired to estimate the causal conditional DI, I (X → Y(cid:107)W).
required to estimate h (Y(cid:107)W, X). Assuming MVAR model,
let Jyy, Kyw, Kyx respectively denote the number of past sam-
ples of Y,W, X that influence yn. Then for n = 1, 2,··· , N,
the current sample of Y can be expressed as
z
Jyy(cid:80)
Kyw(cid:80)
Kyx(cid:80)
yn=
αjyn−j+
γkwn−k+1+
βlxn−l+1+zn.
(14)
j=1
k=1
l=1
The only difference with (10) are the extra terms of the
time-series W. As a result, the CCL will still be Gaussian
distributed with same variance as the distribution in (11) and
whose mean contains the extra terms corresponding to the
samples of W. The unknown parameters under this model are
αj,γk,βl for j=1,··· , Jyy, k = 1,··· , Kyw,l = 1,··· , Kyx
and the model orders Jyy, Kyw, Kyx. Maximum likelihood
with minimum description length penalty can be used to
estimate these parameters similarly. The resulting parameter
estimates can then be used to calculate the CCL, which is
substituted in (9) to estimate h (Y(cid:107)W, X).
Let Jyy and Kyx denote the number of past samples of Y
and X that influence the current sample of Y. Then the CCL
and
, Xn
n−Kyx+1
(cid:17)
(cid:16)
B. Data-driven CCL Estimation
P(cid:0)ynYn−1, Xn(cid:1) is same as P
(cid:17)
(cid:16)
can be written as
ynYn−1
n−Jyy
n−Kyx+1
(cid:16)
,Xn
P
ynYn−1
n−Jyy
(cid:16)
(cid:16)
Yn
Yn−1
n−Jyy
n−Jyy
P
,Xn
,Xn
P
=
(cid:17)
(cid:17) .
n−Kyx+1
n−Kyx+1
(cid:17)
The joint distribution P
of Jyy + 1
and Kyx consecutive samples of Y and X respectively is
n−Kyx+1
n−Jyy
, Xn
Yn
5
Yn
, Xn
(cid:16)
n− Jyy
n− Kyx+1
(cid:1) are
learned using kernel density estimator [42] with Gaussian
kernels. This estimator is implemented in the 'ks' package
in R [46]. The true values of (Jyy, Kyx) are not known and
should be estimated. The joint density is learned for different
values of Jyy and Kyx and the optimal values(cid:0) Jyy, Kyx
(cid:17)
those that maximize the likelihood. The desired CCL is then
estimated by substituting P
in (15).
The denominator in (15) marginalizes the joint distribution
in numerator of (15) over yn. This marginalization is imple-
mented by approximating the integral with a Riemann sum
of the distribution over a partition of the range of yn. Note
that the convergence of the estimated CCL to the true CCL
depends on the underlying true data distribution [47]. h (Y(cid:107)X)
is obtained by substituting the estimated CCL in (7).
can be extended to estimate P(cid:0)ynYn−1, Wn, Xn(cid:1) as well.
P(cid:0)ynYn−1,Wn,Xn(cid:1)=
Let Jyy, Kyw, Kyx respectively denote the number of past
samples of Y, W, X that influence yn. Then
The data-driven CCL estimation algorithm described above
(cid:17)
(cid:17) .
(16)
(cid:16)
(cid:16)
,Wn
P
n−Kyw +1,Xn
n−Kyw +1,Xn
n−Kyx+1
n−Kyx+1
Yn
Yn−1
n−Jyy
n−Jyy
,Wn
P
The joint distribution in the numerator can be similarly es-
timated using kernel density estimator [42] with Gaussian
kernels using 'ks' package [46]. Note that the optimal values
of the model-orders Jyy, Kyw, Kyx are those that maximize
the likelihood. The denominator in (16) is then obtained
by marginalizing the distribution in the numerator similarly.
The resultant numerator and denominator probabilities are
is further substituted in (9) to estimate h (Y(cid:107)X, W).
substituted in (16) to estimate P(cid:0)ynYn−1, Wn, Xn(cid:1), which
above can be easily modified to estimate P(cid:0)ynYn−1(cid:1), which
is required to estimate h (Y). P(cid:0)ynYn−1(cid:1) is obtained from
The model-based and data-driven CCL algorithms described
either model-based or data-driven CCL by modeling the
dependence of the current sample of Y just on its own
past samples. I (X → Y) and I (X → Y(cid:107)W) can now be
estimated using the estimator proposed in section III.
The DI estimator obtained by using the proposed estimator
in Theorem. 3.2 with model-based CCL and data-driven CCL
estimation algorithms will henceforth be referred to as model-
based and data-driven DI estimator respectively. If data is
assumed to be drawn from MVAR model with Gaussian white
noise, then model-based DI will be referred to as MVAR
model-based DI estimator. Note that model-based approach
is not restricted to just MVAR models, it is feasible for all
those models from which we can estimate the appropriate
causal conditional likelihoods parametrically. We focused on
MVAR with Gaussian white noise in this paper because ECoG
is commonly modeled using this model in connectivity studies
[9], [13]. The performance of both the proposed DI estimators
on simulated time-series data is demonstrated in the following
section.
(15)
V. PERFORMANCE ON SIMULATED DATA
In this section, the performance of the proposed DI esti-
mators is demonstrated using simulated data generated from
linear (section V-A)
five models - two node bidirectional
6
Fig. 1. The true causal connectivity graphs of the simulated data models used
to validate the performance of the proposed model-based and data-driven DI
estimators.
and nonlinear (section V-B) causal network whose true con-
nectivity is depicted in Fig. 1a, a two node unidirectional
noisy chaotic polynomial (section V-C) causal network whose
true connectivity is shown in Fig. 1b, four node linear (sec-
tion V-D) and nonlinear (section V-E) causal network whose
true connectivity is depicted in Fig. 1c. A directed arrow in
Fig. 1 represents a causal connection. The causal connection
between two nodes, say from node A to B in Fig. 1c, implies
I (A → B) > 0 or equivalently, that the past samples of A
have some information about the current sample of B. We
also compared the performance of the proposed DI estimators
with the standard Granger causality (GC) [12]. GC estimate is
obtained from MVGC toolbox [31]. Let us now describe the
performance of the proposed DI estimators on the five models
considered in detail.
A. Two Node Bidirectional Linear Causal Network
Consider two time-series X and Y causally connected as
shown in Fig. 1a. The time-series Y is generated from
yn = β1xn + β2xn−1 + zn, for n = 1, 2,··· , N,
(17)
where xn and zn are sampled from an i.i.d Gaussian distri-
z respectively. The
bution with zero mean and variance σ2
samples of X and Z are independent. The true value of the
DI between X and Y in both directions is used to benchmark
the performance of the proposed model-based and data-driven
DI estimators.
x, σ2
2 log(cid:0)1 + σ2
Let us first look at the true value of DI for the model
by (17) in two special cases. When β1 = 1, β2 = 0, (17)
reduces to yn = xn + zn, and it is obvious that both X and
Y have equal causal information about each other. It is easy
to see that I (X → Y) = I (Y → X) = I (X; Y) = C,
where I (X; Y) is the mutual information between X and
Y and C = 1
when β1 = 0, β2 = 1 and in this case (17) reduces to
yn = xn−1 + zn. In this case, X has causal information
about Y, while Y has no causal information about X. More
precisely, I (X → Y) = I (X; Y) = C and I (Y → X) = 0.
For the remaining case of non-zero β1, β2,
the analytical
expressions for DI are
(cid:1). The other special case occurs
x
σ2
z
I(cid:0)X → Y(cid:1)=1
(cid:16)β1β2σ2
I(cid:0)Y → X(cid:1) = 1
2log
σ2
z
x
(cid:17)
(cid:16)
2 log
−1
+1
2cosh
1 σ2
1 + β2
σ2
z
x
(cid:18)(β2
(cid:17)
.
2)σ2
1 +β2
2β1β2σ2
x+σ2
z
x
(cid:19)
,
(18)
The derivation of (18) uses the tridiagonal matrix determinant
from [48] and is given in Appendix B. Note from (18) that
DI from Y to X does not depend on β2. It is because the
(a) β2 = β1
(b) β2 = 1 − β1
Fig. 2. DI estimates and their standard deviation for the two node network (in
Fig. 1a) generated from a linear model (17) using analytical expression (18),
proposed model-based and data-driven DI estimators for different values of
causal strength quantified by (β1, β2). The DI estimates are plotted against
β1 with β2 = β1 in Fig. 2a and with β2 = 1 − β1 in Fig. 2b.
uncertainty in the current sample of X does not depend on
β2, when causally conditioned on the past of X and Y.
x = 1, σ2
The DI from X to Y and vice versa is estimated from
N = 105 samples of X and Y generated with σ2
z = 1
using the proposed model-based and data-driven DI estimators.
The model-based DI estimator assumes that the time-series
are modeled by a MVAR model with Gaussian white noise,
whereas the data-driven CCL estimator does not impose any
model assumptions on the data. Assuming X, Y are from a
MVAR process and when xn is included in the past samples
of X, Granger causality estimate from X to Y is equal to
twice the MVAR model-based DI estimate from X to Y and
vice versa [29]. We therefore do not show the GC estimates for
linear MVAR models with Gaussian white noise. GC estimates
are plotted only for nonlinear simulated models in this paper.
Fig. 2 plots directed information values obtained from the
analytical expression in (18), I (X → Y) and I (Y → X)
from the proposed model-based and data-driven DI estimators
for different values of β1 ∈ (0, 1). The corresponding curves
are respectively referred to as theoretical, model-based and
data-driven. For the model-based and data-driven curves in
Fig. 2, multiple datasets of X, Y are generated using different
seeds for the random number generator. The mean and the
standard deviation of the resultant estimates are plotted in
Fig. 2. The average standard deviation across all (β1, β2) in
Fig. 2 is about 0.003 and 0.01 for the model-based and data-
driven DI estimators respectively. β2 = β1 in Fig. 2a and
β2 = 1 − β1 in Fig. 2b.
When β1 = β2, a larger β1 implies a stronger causal
connection between X and Y and this should result in a
larger DI. This expected trend is observed in Fig. 2a. This
implies that DI tracks the strength of the causal connection.
Also in the corner case of β1 = β2 = 0, DI is zero in both
directions as expected. In Fig. 2b, DI estimates in the corner
cases of β1 = 0, β2 = 1 and β1 = 1, β2 = 0 match with
the analytical expression as expected. Also as β1 increases
from 0 to 1, the causal information Y has about X increases,
and DI tracks this. This is demonstrated by observing that
I (Y → X) increases with β1 in Fig. 2b. Finally, it is clear
from Fig. 2 that the model-based estimate matches the correct
value of DI estimate from (18) and the data-driven estimator
follows the true value of DI. This validates the accuracy of the
proposed DI estimators. For this MVAR model with Gaussian
XY(a) Two Node Bidirectional NetworkXY(b) Two Node Unidirectional Network(c) Four Node Causal NetworkADCB00.51Connection Strength00.250.5TheoreticalModel-basedData-driven00.51Connection Strength00.20.4TheoreticalModel-basedData-driven white noise, the model-based DI estimator clearly performs
better than the data-driven DI estimator and also has a lower
run-time. We therefore use the MVAR model-based estimator
to estimate DI between data modeled by MVAR processes
with Gaussian white noise, instead of using the data-driven
estimator.
The adaptation of stationary bootstrap algorithm described
earlier is used to assess the significance of the inferred causal
connections for different values of (β1, β2). We observed that
the null hypothesis of no causality from Y to X cannot be
rejected for β1 ∈ {0, 0.1} (P-value > δ = 0.05) and can be
rejected at all other points (P-value < δ) in Fig. 2. This is not
surprising since I (Y → X) is small for β1 ∈ {0, 0.1} and
hence did not result in a significant causal connection from Y
to X. Similarly, we observed that statistically significant causal
connection from X to Y does not exist for β1 = 0, β2 = 0
(P-value > δ) and exits at all other points (P-value < δ) in
Fig. 2. This once again confirms our intuition that only large
positive values of DI imply a statistically significant causal
connection.
B. Two Node Bidirectional Nonlinear Causal Network
Now, consider time-series X and Y causally connected as
shown in Fig. 1a and are generated according to
yn = β1x2
n + β2x2
n−1 + zn, for n = 1, 2,··· , N,
(19)
1
1
where xn and zn are sampled from an i.i.d Gaussian distribu-
z respectively. Also, the
tion with zero mean and variance σ2
samples of X and Z are independent. It is very non-trivial to
estimate I (X → Y) and I (Y → X) using model-based DI
x, σ2
estimator. This is because estimating p(cid:0)xnXn−1
p(cid:0)ynYn−1
(cid:1) and
(cid:1) requires essentially inverting the non-linear, non-
, Yn
1
Gaussian generative model in (19) and this is very hard even
for this simple nonlinear model. These two probability densi-
ties are required to estimate h (X(cid:107)Y) and h (Y) respectively.
Therefore we only use the proposed data-driven DI estimator
to estimate the DI from X to Y and vice versa. However,
we can always assume that the data from the model in (19)
comes from a MVAR model with Gaussian noise, which is
incorrect and estimate DI using the proposed MVAR model-
based DI estimator. The resulting DI estimate will be half of
the Granger causality estimate between these two time-series,
GC (X → Y) and GC (Y → X). Note that GC also assumes
the data is generated from a MVAR process even though it
is incorrect. We will now compare the performance of data-
driven DI and GC estimates on this model.
x = 1, σ2
Directed information and Granger causality between X and
Y in both directions is estimated from N = 105 samples
generated with σ2
z = 1 for different values of (β1, β2)
and plotted in Fig. 3. The DI and GC estimates are plotted for
β2 = β1 and β2 = 1 − β1 in Fig. 3a and Fig. 3b respectively.
For each (β1, β2), multiple datasets of X, Y are generated
with different random number generator seeds. The mean and
the standard deviation of the resultant data-driven DI and GC
estimates are plotted in Fig. 3. The average standard deviation
across all (β1, β2) of the data-driven DI and GC estimates is
0.01 and 1.8× 10−5 respectively. In addition, the search space
7
(a) β2 = β1
(b) β2 = 1 − β1
Fig. 3. Data-driven DI and GC estimates, along with standard deviation of
the estimates, for the two node network (depicted in Fig. 1a) generated from
the nonlinear model (19) for different values of causal strength quantified by
(β1, β2). The estimates are plotted against β1 with β2 = β1 in Fig. 3a and
with β2 = 1 − β1 in Fig. 3b.
of the model order used by the Granger causality estimator
is up to 20, i.e, Jyy, Kyx ∈ [1, 20]. In Fig. 3a, I (X → Y)
increases with β1 as expected. DI estimates also behave as
expected in the corner cases of (β1, β2) = (0, 1) and (1, 0)
in Fig. 3b. I (Y → X) increases with β1 as expected. This
once again demonstrates that DI tracks the strength of causal
connections. On the other hand, Granger causality estimates
in both directions are almost zero (of the order of 10−5),
indicating that Granger causality cannot detect
the causal
connections in nonlinear models.
The statistical significance of the inferred causal connec-
tions by DI and GC estimates for different values of (β1, β2)
in Fig. 3 is assessed using the stationary bootstrap algorithm
described in section III. Using DI, the null hypothesis of no
causality from Y to X cannot be rejected for (β1, β2) ∈
{(0, 0) , (0, 1) (0.1, 0.1) , (0.1, 0.9)} and from X to Y cannot
be rejected for (β1, β2) = (0, 0) (P-value > δ = 0.05) in
Fig. 3. At all other points in Fig. 3, the null hypothesis of no
causality can be rejected (P-value < δ) using DI estimates.
This once again confirms our intuition that large values of DI
imply a statistically significant causal connection. For GC, the
null hypothesis of no causality cannot be rejected at all points
in Fig. 3 implying that GC could not find statistically signif-
icant causal connections in nonlinear models. This example
proves that DI is a more general causal connectivity metric
that is not restricted to some particular models.
C. Two Node Unidirectional Noisy Chaotic Polynomial Map
We now consider two unidirectionally coupled time-series
X and Y whose underlying causal connectivity is shown in
Fig. 1b. The time-series X and Y are generated from a noisy
chaotic polynomial map [49] according to
xn = 1.4 − x2
n−1 + 0.3xn−2,
(20)
yn = 1.4− (βxn−1 + (1 − β) yn−1) yn−1 + 0.3yn−2,
where β controls the amount of causal information flowing
from X to Y. The initial
two samples, x1, x2, y1, y2 are
randomly chosen. The two time-series become completely
synchronized for β > 0.7. Gaussian i.i.d measurement noise
of variance 0.01 is added to both time-series X and Y. For
β ∈ [0, 0, 7), strength of the causal connection from X to Y
should increase with β and there is no causal connection from
Y to X. For β ∈ (0.7, 1], since both time-series are completely
-100.51Connection Strength00.250.5^I(X!Y)^I(Y!X)^GC(X!Y)^GC(Y!X)-100.51Connection Strength00.20.4^I(X!Y)^I(Y!X)^GC(X!Y)^GC(Y!X) 8
Fig. 4. Data-driven DI estimates and
GC estimates, along with standard de-
viation of the estimates, for two node
unidirectional network in Fig. 1b gen-
erated from noisy chaotic polynomial
map (20) for different values of the
coupling parameter β.
(a) Model-based Estimator
(b) Data-driven Estimator
Fig. 5. The causal network along with connection strengths
between the four MVAR processes simulated from (21)
estimated by the MVAR model-based DI and the data-
driven DI estimators. The true causal connectivity graph
between these four time-series is depicted in Fig. 1c. It is
clear that both DI estimators correctly infer the underlying
causal network.
(a) Model-based Estimator
(b) Data-driven Estimator
Fig. 6. The causal network along with connection strengths
between the four time-series simulated from (22) estimated
by the MVAR model-based DI and the data-driven DI es-
timators. The true causal connectivity graph between these
four time-series is depicted in Fig. 1c. It is clear that unlike
MVAR model-based estimator, the data-driven estimator
correctly infers the underlying causal connectivity graph.
synchronized and because of the measurement noise, there is a
non-zero equally strong causal connection in both directions.
In the absence of measurement noise for β ∈ (0.7, 1), xn = yn
leading to causal conditional entropy estimate of negative
infinity and a DI estimate of infinity. The intuition behind this
is that once the past of X is known, there is no uncertainty left
in Y. On the other hand, GC estimates in the synchronized
range will be close to zero because the past of X used by GC
(unlike DI, GC does not include xn in the past of X) does not
contain any predictive information about yn resulting in a GC
estimate of zero from X to Y. Note that it is very non-trivial
to apply model-based DI on this model because of the same
reasons outlined in the previous simulated nonlinear model.
We therefore only compare the performance of data-driven DI
and GC estimates on this model.
DI and GC in both directions is estimated from N = 105
samples of X and Y (after discarding the initial transient
points) for different values of β ∈ [0, 1] and plotted in Fig. 4.
For each β, the time-series are generated from (20) using
different seeds of the random number generator. The mean and
the standard deviation of the resulting data-driven DI and GC
estimates are plotted in Fig. 4. The average standard deviation
across all β for the data-driven DI and GC estimates is 0.03
and 0.001 respectively. The standard deviation was largest at
β = 0.7, implying that it is very hard to estimate at the bound-
ary before and after complete synchronization. In addition, the
search space of the model order used by the Granger causality
estimator is up to 20, i.e, Jyy, Kyx ∈ [1, 20]. The DI estimate
is obtained by subtracting two non-negative numbers and it
can sometimes be a small negative number because of the
inaccuracies in estimation algorithms or insufficient data or
violation of stationarity assumptions [23] and in those cases,
we reset the DI estimate to be zero. For instance, the largest
negative DI estimate we obtained for this model is −0.06 from
Y to X at β = 0.6 and we reset this estimate to 0. It is clear
from Fig. 4 that DI estimates behave as expected. DI from X
to Y increase as β goes from 0 to 1. On the other hand, the DI
estimates from Y to X are very small numbers for β < 0.7 and
then there is a sudden jump in this estimate after β > 0.7. This
jump is because the time-series get synchronized for β > 0.7.
On the other hand, GC estimates in both directions are small
positive numbers (when compared to DI estimates) for the
whole range and become equal in value in the synchronized
range of β > 0.7.
The statistical significance of the causal connections in-
ferred by DI and GC estimates is assessed using the adaption
of stationary bootstrap. The null hypothesis of no causality
using DI estimates from Y to X cannot be rejected for β < 0.7
and cannot be rejected for the connection from X to Y for
β < 0.1. This implies DI correctly identifies the presence
of causal connection from X to Y for all β ≥ 0.1 and the
absence of causal connection from Y to X for β < 0.7. It can
also differentiate causally independent time-series (β = 0) and
completely identical time series (β ∈ (0.7, 1]). On the other
hand, the null hypothesis of no causality cannot be rejected
only for β = 0 using GC estimates. This implies GC identifies
the presence of a causal connection in both directions for all
non-zero β, which is incorrect. This example also shows DI
correctly infers causal connectivity from nonlinear models.
D. Four Node Linear Causal Network
Now, consider the four node causal network depicted in
Fig. 1c. The four time series A, B, C and D are generated
according to
bn = an−1 + an−2 + zb
n,
dn = an−2 + zd
n,
n, zc
n and zd
cn = bn−1 + zc
n,
for n = 1, 2,··· , N,
(21)
where an, zb
n are sampled from an i.i.d Gaus-
sian distribution with zero mean and unit variance. In this
network, A influences C indirectly through B. This is an
example of an 'indirect' causal connection, in contrast with
the connection from A to B, which is a 'direct' causal
connection. DI estimate between pairs of time-series cannot
differentiate between 'direct' and 'indirect' causal connections
[17]. For instance, the DI estimate from A to C is positive,
even though A does not directly influence C, but causally
influences C via B. A thorough discussion on the direct
and indirect
influences for point processes is in [17] and
is directly applicable here. Following the approach taken in
[29], [50],
influence from A to C is
non-zero, if and only if I (A → C(cid:107)B, D, ) > 0. However,
estimating the causally conditioned DI when the number of
channels recorded from is large (of the order of hundred's)
is difficult because of the curse of dimensionality [43]. To
overcome this, the pairwise DI is first estimated between all
the 'direct' causal
00.50.71Connection Strength0 0.30.6ADCBADCBADCBADCB pairs of channels. The indirect influences are then resolved
by first estimating only the required causal DI between two
processes, conditioned on one more process. Then if required,
the causal DI between two processes, conditioned on two more
processes, is estimated and so on. The termination condition is
determined by the desired degree of 'directness' in the inferred
causal network. In this simulated example, we are interested in
recovering the true 'direct' causal network depicted in Fig. 1c.
To infer the true causal network, DI is estimated between
these four time-series using both MVAR model-based and
data-driven DI estimators. Model-based DI estimator assumes
the data is generated from a linear causal MVAR model,
whereas data-driven DI estimator does not impose any para-
metric model assumptions on the data. The data is generated
from (21) using 20 different seeds to generate the Gaussian
noise and the resultant estimates are averaged. We will first
describe the performance using model-based DI estimator.
Model-based DI estimator is used to estimate the pair-
wise DI between all pairs of these four nodes, resulting
a 4 × 4 matrix with zeros on the diagonal. We found
that I (A → B) = 0.485 ± 0.009, I (A → C) = 0.314 ±
0.009 and I (B → C) = 0.658 ± 0.009. To determine if there
is an indirect causal connection from A to C or from B to
C, we estimated I (A → C(cid:107)B) and I (B → C(cid:107)A) using the
model-based causally conditioned DI estimator described in
section III, IV-A. We found that I (A → C(cid:107)B) = 0 and
I (B → C(cid:107)A) = 0.344 ± 0.009. Therefore, A to C is an
'indirect' connection via B. Causally conditional DIs are
estimated till the network is completely resolved and free of
any indirect influences. The estimated causal network along
with the strength and the standard deviation of the estimated
causal connections is depicted in Fig. 5a. It is clear from
Fig. 5a and Fig. 1c that model-based DI estimator infers the
true causal network correctly.
We now use the data-driven DI estimator to infer the true
causal network. The pairwise DI is estimated between all
pairs of these four nodes using the data-driven estimator,
resulting in a 4 × 4 matrix with zeros on the diagonal.
Using this DI estimator, we find that I (A → B) = 0.468 ±
0.009, I (A → C) = 0.296± 0.004and I (B → C) = 0.648±
0.008. To identify the presence of any indirect connections,
we estimated I (A → C(cid:107)B) and I (B → C(cid:107)A) using the
model-based causally conditioned DI estimator described in
section III, IV-B. We found that I (A → C(cid:107)B) = 0 and
I (B → C(cid:107)A) = 0.273 ± 0.009. Therefore, A to C is
an 'indirect' connection via B. This procedure is continued
to identify and remove all indirect causal connections. The
resultant estimated direct causal network is depicted in Fig. 5b.
It is clear that data-driven DI also recovers the true network
correctly. Moreover, it is clear from Fig. 5 that for this model,
both model-based and data-driven DI estimators correctly infer
the underlying causal network, which is not surprising since
the underlying model is a linear MVAR model.
E. Four Node Nonlinear Causal Network
We now use a nonlinear model to generate the four time-
series A, B, C and D whose underlying causal connectivity
9
graph is depicted in Fig. 1c. N samples from the four time-
series are generated according to
bn = a2
n−2 + zb
n−1 + a2
n,
dn = an−2 + zd
n,
cn = bn−1 + zc
n,
for n = 1, 2,··· , N,
(22)
n, zc
n and zd
where an, zb
n are sampled from an i.i.d Gaussian
distribution with zero mean and unit variance. The only
difference with the model in section V-D is that the causal
connection from A to B is now nonlinear.
First, we infer the true causal connectivity for this model
using the MVAR model-based DI estimator. This DI estimator
assumes that the data is drawn from a linear MVAR model,
which is not true for this model. It is clear from (22) that the
time-series B is not generated from a linear MVAR model.
Pairwise DI is estimated using this model between all pairs
of these four time-series resulting in a 4 × 4 matrix with
zeros on the diagonal. The only significant causal connections
estimated by the model-based DI estimator are from B to C
and from A to D. This process is repeated for data generated
using 20 different seeds and the resultant DI estimates are
averaged. We find that I (B → C) = 0.847 ± 0.013 and
I (A → D) = 0.348 ± 0.008. It is also clear that there are
no indirect connections to resolve in this case. The underlying
causal connectivity graph estimated by the model-based DI
estimator is depicted in Fig. 6a. It is clear from this figure that
model-based DI estimator could not recover this true network
correctly. This is not surprising since the MVAR model-
based estimator can only identify linear causal connections and
cannot identify the nonlinear causal connections. As result, the
connection from A to B is not identified by the model-based
DI estimator.
We now use data-driven DI estimator to infer the causal
connectivity from the simulated data. The pairwise DI is
estimated between all pairs of these four nodes using the
data-driven estimator, resulting in a 4 × 4 matrix with zeros
on the diagonal. In contrast to the model-based DI estima-
tor, we find that DI from A to B estimated using data-
driven DI is nonzero. Specifically, we find that I (A → B) =
0.433 ± 0.011. In addition, we also find that I (A → C) =
0.320 ± 0.010 and I (B → C) = 0.753 ± 0.009. To eliminate
indirect causal connections, we estimated I (A → C(cid:107)B) = 0
and I (B → C(cid:107)A) = 0.262 ± 0.037. Therefore, A to C is
an 'indirect' connection via B. This procedure is continued
to identify and remove all indirect causal connections. The
resultant estimated direct causal network is depicted in Fig. 6b.
It
is clear that data-driven DI estimator recovers the true
network correctly, while the model-based DI estimator could
not infer the true causal network correctly.
The five diverse simulated models considered in this section
demonstrate that the DI correctly infers the presence and tracks
the strength of a causal connection - large values of DI imply
a strong causal connection and vice versa. Using stationary
bootstrap, we also showed that only large positive DI estimates
correspond to statistically significant causal connections. We
also observed that model-based DI estimator cannot identify
nonlinear causal connections, whereas data-driven DI estima-
tor can correctly identify both linear and nonlinear causal
10
CLINICAL DETAILS OF THE PATIENTS ANALYZED.
TABLE I
P1
Patient ID Age/Sex Syndrome
20/M Nonlesional
temporal
Lesional
60/M
temporal
29/M Nonlesional
temporal
37/M Nonlesional
Lesional
temporal
20/F
P4
P2
P3
P5
extratemporal SPS+CPS
Seizure
Type
CPS
CPS
CPS
Electrode
Type
Surgery Outcome
of Surgery
D
Right TL Class I
Selective
Left HC Class II
G+D Right TL Class II
D
G
Right OC Class III
CPS
G+D
Left TL
Class I
CPS - complex partial seizures, SPS - simple partial seizures. D - depth
electrodes, G - subdural grid electrodes. TL - temporal lobectomy, HC -
hippocampectomy, OC - occipital corticosectomy. The outcomes are in Engel
epilepsy surgery outcome scale. "Class I - free of disabling seizures, class
II - Almost seizure-free, class III - worthwhile improvement, class IV - no
worthwhile improvement" [52].
connections. We now use both the MVAR model-based and
data-driven DI estimators to infer the causal connectivity graph
from ECoG data in epileptic patients. We only consider the
large DI estimates (large compared to the rest of the causal
connectivity graph) since they only imply a significant causal
connection. We propose a model-based and a data-driven SOZ
identification algorithm in the following section.
VI. SEIZURE ONSET ZONE IDENTIFICATION ALGORITHMS
Seizure onset zone (SOZ) is defined as the regions of the
brain that initiate seizures [8]. The current clinical standard is
for neurologists to identify SOZ from visual analysis of the
ECoG data. The SOZ identified in this way is removed during
resective surgery. However, visual analysis is time consuming,
subjective and potentially unreliable [36], [51]. We propose
two computationally derived SOZ identification algorithms -
model-based and data-driven SOZ identification algorithms.
We identified the SOZ in five patients with epilepsy using
these two algorithms and compared their performance with
visual analysis by the neurologist.
A. Clinical ECoG Data
The five patients analyzed here were all managed and
treated by our physician coauthors. The clinical details of
these patients are summarized in Table I. Three seizure records
each from patients P1, P2 and P5, two from patient P3 and
one from patient P4 were analyzed. Each seizure record was
approximately 10 minutes long and contained one seizure.
Each seizure on average lasted for a minute and was roughly
in the middle of the seizure record. The seizure start time
was identified by the neurologist. Each electrode records the
voltage waveform at a sampling frequency of 1 KHz. The
number of electrodes in these five patients varied from 120
to 150. Electrodes with artifacts likely due to either loose
contacts, patient movement or excessive line noise were not
included in the analysis.
Fig. 7. A 30s snapshot of ECoG signals from the 30 high energy channels
of P1. The seizure start time, represented by a vertical solid black line, is
identified by neurologist. Causal connectivity is estimated from this entire
30s window for this seizure record.
B. Proposed SOZ Identification Algorithms
The first stage of the proposed SOZ identification algorithm
is an energy detector which selects only M channels out of
all ECoG channels for further analysis. The main objective
of this stage is to reduce the computational complexity of
the proposed algorithms. The energy is l2-norm of the ECoG
signal computed from a window around the start of seizures
containing preictal and ictal recordings. Any channel involved
in seizure onset is expected to have interictal spikes before the
seizure starts and/or have high amplitude low-frequency ictal
activity once the seizure is fully developed, both of which will
increase the energy in the selected time-window. The time-
window was selected to be long enough to capture both spiking
and large ECoG amplitudes during seizures. The second stage
consisted of estimating the causal connectivity between every
pair of M channels selected in the first stage to form a M ×
M causal connectivity matrix. The causal connectivity was
estimated from a shorter time-window around the seizure start
time, since we are interested in estimating the seizure onset
electrodes. The following subsections describe the remaining
stages of the two proposed SOZ identification algorithms.
1) Model-based SOZ Identification Algorithm: In this ap-
proach, ECoG data is assumed to be derived from a MVAR
process with Gaussian white noise. This is a very common
assumption imposed to estimate causal connectivity between
ECoG data [9], [13]. The MVAR model-based DI estimator is
used to infer the causal connectivity between the selected M
high energy channels. The causal connectivity estimated using
this approach only represents the linear causal interactions
between the ECoG channels. However it is widely believed
that seizures are highly non-linear phenomenon during which
SOZ drives the rest of the network into a hypersynchronous
state [4], [8], [16]. As a result, we expect the seizure onset
electrodes in the causal connectivity graph to be isolated,
since model-based approach can only capture linear causal
interactions. The proposed model-based algorithm therefore
identifies the nodes in the causal connectivity graph with zero
degree (threshold was set to select only the strongest 10%
connections) as the estimated SOZ. If a patient had multiple
seizures, the electrodes identified across all seizures in that
patient form the estimated SOZ for that patient.
2412512612713020101Time (s)Channel Index 11
(a) From Patient P1
(b) From Patient P3
Fig. 9. Normalized net outward flow from the ECoG electrodes with positive
net information outflow using data-driven SOZ identification algorithm.
activity (150 past samples at Fs = 1KHz) at this channel and
other channels. This corresponds to restricting the model order
Jyy, Kyx search space to [1, 150] for the MVAR model-based
DI estimator. In addition, we need to capture the connectivity
just before and just after a seizure starts to estimate the SOZ.
Therefore, we used ECoG data from a 30s window (3 × 104
data points) that begins 20s before the start of the seizure to
be stationary. The same window was used for the data-driven
estimator as well. In addition, the past activity was down-
sampled by a factor of 50 for the data-driven estimator to
restrict the Jyy, Kyx search space to [1, 4] and also reduce its
computational complexity (i.e. the past activity of channel X
can include {xn, xn−50, xn−100, xn−150}). The exact values
of these parameters is not crucial as the algorithms seem to
be fairly robust to changes in these parameters.
Consider the second seizure record of patient P1. The
energy detector selected 30 high energy channels. Fig. 7
shows the recordings from these channels in the 30s window
in which causal connectivity graph is inferred. The inferred
graph by model-based and data-driven approaches is shown in
Fig. 8. The weighted adjacency matrix of the inferred causal
connectivity graphs, whose (i, j)th element is the DI estimate
from channel i to j for i, j ∈ [1, 30], is plotted in Fig. 8
using a image plot. It is clear from this figure that the mean
strength of the DI estimates using model-based approach is
smaller than using data-driven approach (colorbar ranges are
different in the two sub-figures). We observed this across
all
the twelve seizures analyzed. This indicates that data-
driven DI captured more causal information on average than
model-based DI, implying that non-linear causal interactions
are stronger around the beginning of a seizure. The nodes
with zero degree in the causal connectivity graphs from each
seizure in a patient are identified as the SOZ by the model-
based algorithm. The zero degree criterion used by model-
based algorithm is counterintuitive, since we except the SOZ to
drive the network to seizure state and not be weakly connected.
On the other hand, the data-driven algorithm selects electrodes
with large net outflows, which is very intuitive. The data-
driven algorithm computed the normalized net outward flow
for each node using (24). Fig. 9a plots the Φ for all electrodes
with positive net outward flows in patient P1. The electrodes
with Φ > 5% are the estimated SOZ for this patient P1 using
data-driven algorithm.
Table II summarizes the results from our analysis. The first
column in Table II identifies the patient ID and the number
(a) From model-based algorithm
(b) From data-driven algorithm
Fig. 8. Causal connectivity between 30 high energy channels estimated from
ECoG data between 241s and 271s from the second seizure of P1. The channel
indices with bluish rows and bluish columns (correspond to low DI estimates)
in Fig. 8a correspond to isolated nodes and are the estimated SOZ using
model-based algorithm. The corresponding channels in Fig. 8b have large net-
outflows of information and are the estimated SOZ from data-driven algorithm.
2) Data-driven SOZ Identification Algorithm: In this al-
gorithm, no parametric model assumptions were imposed on
ECoG data. The causal connectivity between the M high
energy channels selected in the first stage was inferred using
the data-driven DI estimator. This estimator inferred both
linear and nonlinear causal
interactions between channels.
Intuitively, activity at the SOZ electrodes drives the activity at
the other electrodes into a hypersynchronous state via linear
and nonlinear causal interactions [16]. We therefore expect the
SOZ electrodes to act as sources (with strong outgoing and
weak incoming causal connections) in the causal connectivity
graph inferred around the seizure start time using data-driven
DI. As a result, the SOZ nodes in the causal connectivity graph
are expected to have large net-outward flow of information.
The data-driven SOZ identification algorithm quantifies this
intuition to estimate SOZ. The net-outward flow (Φ) of causal
information from an electrode i is calculated using
{I(i → j) − I(j → i)} .
M(cid:80)
Φ(i) =
(23)
j=1,j(cid:54)=i
If a patient had multiple seizures, the net outward flow of an
electrode is the average net outward flow of that electrode
across all seizures recorded in that patient. Then the normal-
ized net outward flow ( Φ) of the electrode i is given by
Φ(i) = 100 ×
Φ(i)(cid:80)
Φ(j) .
(24)
j:Φ(j)>0
The electrodes with Φ > 5% are considered to have significant
net outward flow of information in the causal connectivity
graph and are identified as the seizure onset electrodes for
that patient by the data-driven SOZ identification algorithm.
C. Performance of Proposed SOZ Identification Algorithms
The energy detector selected the top M = 30 channels
with the largest energy computed from a 100s window com-
prising of 50s of activity immediately before and after the
seizure starts. The causal connectivity graph between these
high energy channels is then estimated using model-based
and data-driven DI estimators from a 30s window that begins
20s before the seizure start time and ends 10s into the start
of the seizure. We assumed that the current activity at an
ECoG channel does not depend on more than 150ms of past
Channel Index1 1530Channel Index1 153000.10.20.30.40.5Channel Index1 1530Channel Index1 15300 0.10.94.6 12
of seizures analyzed for that patient. The second, third and
fourth columns in Table II list the SOZ identified across all
the five patients using model-based, data-driven algorithms
and visual analysis respectively. We observed that all
the
channels identified as SOZ by visual analysis, except AST
2 in one seizure of P5, are included in the 30 high energy
channels selected from each seizure by the energy detector
in the first stage. The top 30 channels selected from two
seizures in patient P5 contained AST 2, but the 30 channels
picked from the third seizure did not contain AST 2. The
normalized net outflow Φ from AST 2 electrode for patient P5
using data-driven algorithm was 1% and hence this electrode
was not identified as SOZ (note that Φ has to exceed 5% to
be selected as SOZ). Expect for this one region, it is clear
from this table that the data-driven algorithm identifies all
the regions identified by the neurologist, whereas the model-
based algorithm misses some regions (for instance, RAMY
electrodes in P1, TP and AST electrodes in P5). Also, the
model-based algorithm incorrectly identified lateral temporal
(LT) electrodes as SOZ in patient P3, whereas data-driven
algorithm correctly identified posterior depth (PD) electrodes
in hippocampus as SOZ. Except in P3 and P4, both algorithms
do not have any false positives. The false positives in P4 could
be because only one seizure was analyzed in this patient.
Another advantage of the data-driven SOZ identification
algorithm over model-based algorithm and analysis by the
neurologist is that Φ could be used as a quantitative metric
to rank the electrodes in the decreasing order of clinical
relevance. Fig. 9 plots the Φ of all electrodes with positive net
outward flows in patients P1 and P3. Using our metric Φ, it is
clear from Fig. 9b that electrodes PD4, PD5 contribute much
more in generating and spreading seizures than LF28 and
LP4 electrodes even though Φ exceeds the chosen threshold
at all these four electrodes. Depending on the significance
level (5% is used here), the set of selected SOZ electrodes
varies. We observed in all five patients that the electrodes
with the highest Φ values were always the same as the ones
identified by the neurologist. Visual analysis by the neurol-
ogist can only give qualitative information about the SOZ
and cannot give quantitative information like the proposed
data-driven SOZ identification algorithm. In addition, data-
driven algorithm can also differentiate between electrodes
in close proximity - for example, Φ is negative for RPH2
electrode in P1 even though Φ is positive for both RPH3
and RPH4 (refer to Fig. 9a). The increased spatial-specificity
provided by our data-driven algorithm could be relevant for
next generation epilepsy treatments [5]. The main advantage
of the model-based algorithm over data-driven one is its lower
computational complexity. However, this is less critical with
today's powerful computers. To summarize, data-driven SOZ
identification algorithm outperforms model-based algorithm
and provides more interpretable results.
VII. DISCUSSION AND CONCLUSIONS
An almost surely convergent MVAR model-based and data-
driven estimators for DI are introduced in this paper. Linear
causal interactions between two time-series can be quantified
SEIZURE ONSET ZONE IDENTIFIED FROM THE PROPOSED ALGORITHMS
AND THE VISUAL ANALYSIS BY NEUROLOGIST.
TABLE II
Patient - #
of Seizures
P1 - 3
P2 - 3
P3 - 2
P4 - 1
P5 - 3
Model-based
Algorithm
Data-driven
Algorithm
RAH 1-3, RPH 2-4 RAH 1-2, RPH 4,
LAH 2-4, LPH 2
LAH 2-4, LPH 1-2
RAMY 2-3
LT 1-3, 10
LO 3, 14, 15, 25,
LO 12, 13, PST 3,
PST 1, MOG 27
MST 1, 2, HD 1
PD 4-5, LF 28, LP 4
LO 3, 14, 15, 12,
PST 1, MOG 23,
SOG 21, 36
MST 1, TP 1,
HD 1
Visual Analysis
RAH 1-3, RPH 2-4,
RAMY 2-3
LAH 2-4, LPH 1-2
PD 3-5
LO 3, 14, 15,
LO 25, PST 3
MST 1, 2, TP 1,
HD 1-3, AST 2
The label of an ECoG electrode comprises of an abbreviation of the brain
region it is implanted in and a number. For depth electrodes, smallest number
is assigned to deepest electrode from scalp. For instance, RAH1 - deepest
electrode contact in depth electrode in right anterior hippocampus and LO3 -
third electrode contact in subdural grid electrode over lateral occipital lobe.
RPH - right posterior hippocampus, RAMY - right amygdala, LF - lateral
frontal, LP - lateral parietal, LT - lateral temporal, PD - posterior hippocampal
depth, MOG - medial occipital grid, SOG - sub-occipital grid, PST - posterior
sub-temporal, MST - mid-subtemporal lobe AST - anterior sub-temporal lobe,
TP - temporo-polar, HD - hippocampal depth.
using MVAR model-based DI estimator, whereas both linear
and nonlinear causal interactions are quantified by data-driven
DI estimator. The resultant DI estimates can be used to infer
whether the data has (1) linear causal interactions or (2) both
linear and nonlinear causal interactions. If the MVAR model-
based DI estimate is comparable in value to data-driven DI
estimate, then the interaction is predominantly linear. This is
not feasible with existing metrics because they can be split into
two non-overlapping groups - the first group only detects linear
causal interactions (e.g., Granger causality, partial directed
coherence), whereas the second group detects both linear and
nonlinear causal interactions (e.g., transfer entropy). The DI
estimators proposed in this paper can be automatically adapted
to other types of electrophysiological data like EEG to learn
the causal connectivity.
Data-driven DI estimator seems to be more appropriate than
model-based DI estimator if the underlying data distribution
is not known, which is the case with most real data. The
main challenge with data-driven DI estimator is estimating
the causal conditional likelihood nonparametrically and its
computational complexity. We used kernel density estimators
in this paper to estimate causal conditional likelihood. Ker-
nel density estimators are asymptotically optimal [43]. Their
bias decreases with increasing number of data samples and
complexity increases with the dimensionality of the data,
just
like other nonparametric estimators. Even though we
selected optimal bandwidth using smoothed cross-validation
to minimize the asymptotic mean integrated squared error,
several other criteria could also be used [43], [46]. In addition,
data-driven entropy estimators based on adaptive partitioning,
nearest neighbors and m-spacing algorithms [24], [53] can also
be used to estimate DI nonparametrically. Another approach
to estimate DI nonparametrically is to extend the universal
DI estimator proposed for discrete-valued signals in [23] to
continuous-valued ECoG signals. Future work should also
include developing approximate data-driven DI estimators to
13
Fig. 10. Causal connectivity between
30 high energy channels depicted in
Fig. 7 estimated using PDC.
(a) From a segment before the seizure
(c) From a segment after the seizure
Fig. 11. Causal connectivity between the 30 high energy channels from the second seizure of patient P1 estimated
using data-driven DI estimator from ECoG data in three segments - one before seizure (181s -211s), one during seizure
(261s - 291s) and one after seizure (361s - 391s). This seizure starts at 261s and ends at 350s.
(b) From a segment during the seizure
further reduce computational complexity.
Directed information was used in this paper to estimate
causal connectivity between ECoG channels. The causal con-
nection identified between two channels could be due to the
effect of activity at other spatial locations in the brain. If
an ECoG electrode was implanted at these other locations,
causally conditioned DI can be used to remove their influ-
ence. This was demonstrated using the four node examples
in section V. On the other hand, if ECoG activity is not
recorded from these locations, then removing the effects of
these hidden nodes on the inferred causal connectivity is a
very hard problem in general. Future work should look into the
sensitivity of DI to volume conduction effects when compared
with synchronization metrics like phase lag index [54].
DI estimators proposed in this paper do not quantify the
amount of causal information between time-series at each
frequency, unlike partial directed coherence (PDC) or directed
transfer function (DTF). However, the advantage of DI is
that data-driven DI estimator can detect nonlinear causal
interactions, which PDC or DTF cannot detect. Metrics based
on PDC, DTF assume the data is drawn from a MVAR
model and can only detect linear causal interactions (similar
to MVAR model-based DI estimator proposed in this paper).
To demonstrate this, we estimated the causal connectivity
graph between the 30 channels depicted in Fig. 7 by PDC
using eMVAR toolbox [55]. The resultant 30 × 30 causal
connectivity matrix is plotted in Fig. 10,
in which (i, j)
element corresponds to the maximum value of PDC from
channel i to channel j. Note that causal connectivity estimates
from the proposed DI estimators for the same data is plotted
in Fig. 8. Comparing Fig. 10 with Fig. 8b, it is clear that net
outflow from the SOZ electrodes is not large in PDC when
compared to data-driven DI. This implies unlike data-driven DI
estimator, PDC cannot capture nonlinear causal interactions.
We also proposed model-based and data-driven algorithms
to identify the SOZ. The first stage of both these algorithms
is an energy detector. The chosen electrodes from the first
stage turned out to have large overlap (more than half) across
multiple seizures within a patient. All electrodes with low
rhythmic gamma activity in SOZ were selected by the energy
detector in all the patients analyzed. Note that other criteria
could also be used instead of energy detector. In particular, we
experimented with selecting channels displaying strong high-
frequency activity around the seizure start time (since channels
involved in seizure onset display strong high-frequency activ-
ity around the beginning of a seizure that typically develops
into high amplitude low-frequency activity). The time-window
used to estimate the high-frequency activity should be of
much smaller length than the one used with energy detector,
because the seizures typically display low amplitude rhythmic
high-frequency oscillations only for a very short duration.
The resulting performance with energy detector or the high-
frequency activity detector was similar. We therefore presented
the results only with the energy detector in this paper.
The causal connectivity graphs between the selected high
energy channels estimated using MVAR model-based DI and
data-driven DI from same time-window are not the same, since
both estimators capture different causal interactions in the
data - model-based captures linear interactions, whereas data-
driven captures both linear and nonlinear causal interactions.
Therefore the criterion used to estimate SOZ from the causal
connectivity graph was different for the two algorithms. In
the model-based approach, the SOZ nodes are isolated since
they drive the other brain regions into a seizure through
nonlinear interactions (which are not captured by model-based
DI estimator). Similar results were reported in other studies
using linear metrics [56], [57]. It is reported in [57] that SOZ
electrodes form an isolated focus using symmetric coherence
metric that captures linear interactions. On the other hand in
causal connectivity graphs estimated by data-driven DI, the
outgoing and incoming edges from SOZ electrodes have large
and small DI estimates respectively (refer to Fig. 9). This
is in accordance with our intuition that the SOZ drives the
seizure activity [4], [8], [16]. Also metrics closely related to
net outward flow were used in [35] to infer SOZ using transfer
entropy (which detects nonlinear interactions) by analyzing
hours of ECoG recordings (here we are only using recordings
from a 30s window).
A key advantage of ECoG recordings over other neu-
roimaging techniques is its good temporal resolution. The DI
estimators proposed in this paper can be applied to ECoG
recordings from different windows to learn the spatiotemporal
changes in causal connectivity networks during the course of
a seizure. The causal connectivity before, during and after the
second seizure of patient P1 estimated using data-driven DI
estimator from three 30s long windows is shown in Fig. 11.
It is clear from Fig. 11 that SOZ electrodes (corresponding
to rows with more red color or large DI values in Fig. 11b)
have large net outflows during seizure when compared with
before and after seizure (same rows have more blue color or
Channel Index1 1530Channel Index1 15300 0.10.20.30.4Channel Index1 1530Channel Index1 153000.71.42.12.8Channel Index1 1530Channel Index1 15300.10.41.86.3Channel Index1 1530Channel Index1 153000.61.21.8 14
smaller DI value in Fig. 11a , 11c). We are extending this
analysis to infer seizure mechanisms by examining the changes
in causal connectivity estimated from preictal, ictal, postictal
periods when compared with interictal periods. This is the
subject of our current and future work [58]. The results from
this analysis potentially could improve our understanding of
seizure mechanisms and lead to the development of novel non-
surgical treatments for epilepsy.
VIII. ACKNOWLEDGMENTS
The authors wish to thank Sugnaya Karunakaran for care-
fully reading the manuscript.
PROOF OF CAUSAL CONDITIONAL ENTROPY ESTIMATOR
APPENDIX A
A. Proof of Lemma 3.1
1
(cid:1)
(25)
n−K+1
n−J , Xn
(cid:1) ≥ ······
First, we will prove the existence of h (Y(cid:107)X). Since
conditioning reduces differential entropy, we have
increasing sequence that is upper bounded by h (y1). Also
let l = max (J + 1, K). Then for n ≥ l,
1, X2
1
n−K+1
n−J , Xn
(cid:1) ≥ h(cid:0)y2Y1
h (y1) ≥ h(cid:0)y1X1
≥ ··· ≥ h(cid:0)ynYn−1
Therefore the sequence h(cid:0)ynYn−1
h(cid:0)ynYn−1
(cid:1) = h(cid:0)ynYn−1
= h(cid:0)ylYl−1
the sequence h(cid:0)ynYn−1
(cid:1). Let an = h(cid:0)ynYn−1
h(cid:0)ylYl−1
N h(cid:0)YN(cid:107)XN(cid:1) = 1
N(cid:80)
(cid:1) is a non-
(cid:1)
(cid:1) ,
(cid:1) is lower bounded by
(cid:1) and bN =
where (26) is from the Markovian assumption and (27) is
from the stationarity assumption. Note that (27) also implies
N→∞ aN exists, from
Cesaro mean theorem [30] we have h (Y(cid:107)X) = lim
N→∞ bN also
exists. The above proof can be easily modified to prove h (Y)
exists. Therefore I (X → Y) = h (Y)− h (Y(cid:107)X) also exists.
n−J , Xn
l−J , Xl
an. Since the lim
n−K+1
l−K+1
n−J , Xn
(26)
(27)
l−J , Xl
n−K+1
l−K+1
, Xn
1
, Xn
1
n=1
N
1
1
1
1
B. Proof of Lemma 3.2
N h(cid:0)YN(cid:107)XN(cid:1) = 1
N(cid:80)
h(cid:0)ynYn−1
N(cid:80)
E(cid:2)− log P(cid:0)ylYl−1
n=1
= 1
N
= E(cid:2)− log P(cid:0)ylYl−1
n=1
l−J , Xl
N
n−J , Xn
n−K+1
(cid:1)(cid:3) ,
l−J , Xl
l−K+1
l−K+1
(cid:1)
(cid:1)(cid:3)
(28)
(29)
where (28) is from chain rule and Markovian assumption, and
(29) is due to stationarity.
(cid:16)
(cid:17)
C. Proof of Theorem 3.1
n−K+1
be a fixed function over the states of the Markov chain
= −log P(cid:0)ynYn−1
(cid:1)
(cid:1). From the strong law of large numbers for
(cid:0)Yn
Let gJ,K
n−J,Xn
n−J,Xn
n−K+1)
Yn
n−J,Xn
n−K+1
Markov chains [39] which states that for a fixed function g (.)
over the states of the Markov chain, the sample mean will
N(cid:80)
1
N
n=1
gJ,K
We also have
h(Y(cid:107)X)= lim
N→∞
almost surely converge to the expected value as N → ∞, we
have,
n−K+1
n−J,Xn
(cid:0)Yn
(cid:1) a.s.−−→E(cid:2)gJ,K
(cid:0)Yl
Nh(cid:0)YN(cid:107)XN(cid:1)= lim
E(cid:2)gJ,K
(cid:0)Yl
= E(cid:2)gJ,K
(cid:0)Yl
(cid:1)(cid:3) ,
N(cid:80)
N→∞
l−K+1
1
l−J,Xl
l−K+1
(cid:1)(cid:3) . (30)
(cid:1)(cid:3) (31)
(cid:17) a.s.−−→ h (Y(cid:107)X) .
l−K+1
l−J,Xl
l−J , Xl
(32)
where (31) is from Lemma. 3.2. We have from (30), (32),
h (Y(cid:107)X) = 1
(cid:16)
Yn
n−J , Xn
n−(K−1)
gJ,K
N
n=1
APPENDIX B
DERIVATION OF DI FOR LINEAR TWO NODE NETWORK
Consider the MVAR model in section V-A, described by
(17). Here we will derive the DI in both directions between
time-series X and Y for non-zero β1, β2. Appendix. B-C
considers the case when (β1, β2) ∈ {(1, 0) , (1, 0)}.
A. DI from X to Y
For the system described by (17), the causal conditional
entropy h (Y(cid:107)X) is given by
N h(cid:0)YN(cid:107)XN(cid:1) = 1
2 log(cid:0)2πeσ2
1
z
(cid:1) ,
h (Y(cid:107)X) = lim
N→∞
(33)
where γ = (cid:0)β2
because conditioned on (xn, xn−1), the only uncertainty in yn
z which is
is due to the i.i.d Gaussian noise Z of variance σ2
independent of X.
Now, from (17), we have (y1, y2,··· , yN )T ∼ N (0, ΣN ),
x. MN is a tridiagonal
where ΣN = δMN with δ = β1β2σ2
matrix whose main diagonal elements are D and non-zero
diagonal below and above the main diagonal are all 1. D = γ
δ ,
z. Upon further simplification
using the tridiagonal matrix determinant from [48], we have
ΣN = δN sinh((N +1)λ)
The unconditioned entropy of Y is now given by
2 log(2πeδ)+1
=1
2 λ,
−1(cid:16)D
, where λ = cosh
(cid:1) σ2
x + σ2
1 + β2
2
1
2Nlog
(cid:17)
(34)
(35)
sinh λ
h (Y)= lim
N→∞
2
.
obtained by expanding the hyperbolic sinh function in the
determinant ΣN in terms of exponentials and some basic
algebraic manipulations. Now, from (33) and (35), we have
I (X → Y) = 1
(cid:18) (β2
(cid:19)
x+σ2
+ 1
−1
x
z
.
2 cosh
2)σ2
1 +β2
2β1β2σ2
x
2 log
σ2
z
(cid:16)
(2πe)NΣN(cid:17)
(cid:16)β1β2σ2
(cid:17)
B. DI from Y to X
The causal conditional entropy, h (X(cid:107)Y) is given by
h (X(cid:107)Y) = lim
N→∞
h (xnxn−1, yn)
N(cid:80)
1
N
n=1
(36)
(37)
N(cid:80)
N(cid:80)
n=1
n=1
{h (xn, yn, xn−1) − h (xn−1, yn)}
(cid:8) 1
2 log (2πeΦ1) − 1
2 log (2πeΦ2)(cid:9)
= lim
N→∞
1
N
1
N
(cid:16)
= lim
N→∞
= 1
2 log
(cid:17)
2πe σ2
β2
1 σ2
xσ2
x+σ2
z
z
,
where Φ1 and Φ2 are the appropriate covariance matrices. The
reason for (36) is that conditioned on xn−1 and yn, xn is
independent of the other past samples of X and Y. Since
X is drawn from i.i.d. Gaussian distribution with mean zero
and variance σ2
x, the unconditional entropy of X is given by
h (X) = 1
(cid:1) . Therefore, the DI from Y to X is
2 log(cid:0)2πeσ2
x
I (Y → X)=h (X) − h (X(cid:107)Y)=1
2 log
1 + β2
1 σ2
σ2
z
x
.
(38)
(cid:16)
(cid:17)
C. Special cases
2 log
x + σ2
z
x + σ2
z
h(xn, yn)=1
differential entropy of xn and yn is
Consider the system in (17) with β1 = 1, β2 = 0. For this
system, yn are i.i.d. Gaussian distributed with mean zero and
1
(cid:1). Therefore the differential entropy of YN
variance(cid:0)σ2
is given by h(cid:0)YN(cid:1) = N
(cid:1)(cid:1). Also the joint
2 log(cid:0)2πe(cid:0)σ2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)
(cid:12)(cid:12)(cid:12)(cid:12)σ2
(cid:18)
2 log(cid:0)2πeσ2
(cid:1) . (39)
=⇒ h(cid:0)YN(cid:107)XN(cid:1) =
(cid:0)h (xn, yn) −
N(cid:80)
N(cid:80)
(cid:1) . Therefore the directed information
2 log(cid:0)2πeσ2
h (xn)(cid:1) = N
(cid:17)
(cid:16)
σ2
x
x + σ2
z
h (ynxn) =
(cid:0)h(cid:0)YN(cid:1)−h(cid:0)YN(cid:107)XN(cid:1)(cid:1)=1
z
from X to Y is given by
2 log
The DI from Y to X can be similarly derived.
I(X→Y)= lim
N→∞
x
σ2
x σ2
1+ σ2
x
σ2
z
(40)
xσ2
z
2πe
=1
n=1
n=1
.
Now, consider the system in (17) with β1 = 0, β2 = 1. For
this system, the DI from X to Y is computed by following
the approach described above. Let us derive I (Y → X). The
causal conditional entropy of XN given YN is given by
h(cid:0)xnXn−1,YN(cid:1)=
N(cid:80)
h(xn)=h(cid:0)XN(cid:1) , (41)
h(cid:0)XN(cid:107)YN(cid:1)=
N(cid:80)
n=1
n=1
since xn does not depend on the past samples of Y. Therefore,
the DI from Y to X is zero, i.e., I (Y → X) = 0.
REFERENCES
[1] R. Malladi, G. Kalamangalam, N. Tandon, and B. Aazhang, "Identifying
the epileptogenic zone using directed information," in Abstract presented
at Soc. for Neuroscience (SfN), 2014.
[2] R. Malladi, G. P. Kalamangalam, N. Tandon, and B. Aazhang, "Inferring
causal connectivity in epileptogenic zone using directed information," in
IEEE Int. Conf. on Acoust., Speech and Signal Process. (ICASSP), 2015,
pp. 822 -- 826.
[3] G. K. Bergey, M. J. Morrell, E. M. Mizrahi, A. Goldman, D. King-
Stephens, D. Nair, S. Srinivasan, B. Jobst, R. E. Gross, D. C. Shields
et al., "Long-term treatment with responsive brain stimulation in adults
with refractory partial seizures," Neurology, vol. 84, pp. 810 -- 817, 2015.
[4] F. Rosenow and H. Luders, "Presurgical evaluation of epilepsy," Brain,
vol. 124, no. 9, pp. 1683 -- 1700, 2001.
[5] E. Krook-Magnuson and I. Soltesz, "Beyond the hammer and the scalpel:
selective circuit control for the epilepsies," Nature Neuroscience, vol. 18,
no. 3, pp. 331 -- 338, 2015.
[6] S. Sunderam, B. Gluckman, D. Reato, and M. Bikson, "Toward rational
design of electrical stimulation strategies for epilepsy control," Epilepsy
& Behavior, vol. 17, no. 1, pp. 6 -- 22, 2010.
[7] K. J. Friston, "Functional and effective connectivity in neuroimaging: a
synthesis," Human Brain Mapping, vol. 2, no. 1-2, 1994.
[8] H. O. Luders, I. Najm, D. Nair, P. Widdess-Walsh, and W. Bingman,
"The epileptogenic zone: general principles," Epileptic Disorders, vol. 8,
2006.
[9] P. van Mierlo, M. Papadopoulou, E. Carrette, P. Boon, S. Vandenberghe,
K. Vonck, and D. Marinazzo, "Functional brain connectivity from EEG
in epilepsy: Seizure prediction and epileptogenic focus localization,"
Progress in neurobiology, vol. 121, pp. 19 -- 35, 2014.
15
[10] F. Pittau and S. Vulliemoz, "Functional brain networks in epilepsy:
recent advances in noninvasive mapping," Current opinion in neurology,
vol. 28, no. 4, pp. 338 -- 343, 2015.
[11] S. L. Bressler and A. K. Seth, "Wiener -- Granger causality: a well
established methodology," Neuroimage, vol. 58, pp. 323 -- 329, 2011.
[12] C. W. Granger, "Investigating causal relations by econometric models
and cross-spectral methods," Econometrica: J. of the Econometric Soc.,
1969.
[13] K. J. Blinowska, "Review of the methods of determination of directed
connectivity from multichannel data," Medical & Biological Eng. &
Computing, vol. 49, no. 5, 2011.
[14] T. Schreiber, "Measuring information transfer," Physical Review Lett.,
vol. 85, no. 2, 2000.
[15] E. Pereda, R. Q. Quiroga, and J. Bhattacharya, "Nonlinear multivari-
ate analysis of neurophysiological signals," Progress in Neurobiology,
vol. 77, no. 1, pp. 1 -- 37, 2005.
[16] K. Lehnertz, "Epilepsy and nonlinear dynamics," J. of biological physics,
vol. 34, no. 3-4, pp. 253 -- 266, 2008.
[17] C. J. Quinn, T. P. Coleman, N. Kiyavash, and N. G. Hatsopoulos, "Esti-
mating the directed information to infer causal relationships in ensemble
neural spike train recordings," J. of Computational Neuroscience, vol. 30,
no. 1, 2011.
[18] K. So, A. C. Koralek, K. Ganguly, M. C. Gastpar, and J. M. Carmena,
"Assessing functional connectivity of neural ensembles using directed
information," J. of Neural Eng., vol. 9, no. 2, 2012.
[19] N. Soltani and A. Goldsmith, "Directed information between connected
leaky integrate-and-fire neurons," in IEEE Int. Symp. on Inform. Theory
(ISIT), 2014, pp. 1291 -- 1295.
[20] H. Marko, "The bidirectional communication theory -- a generalization of
information theory," IEEE Trans. Commun., vol. 21, no. 12, 1973.
[21] J. Massey, "Causality, feedback and directed information," in Int. Symp.
on Inform. Theory Applications (ISITA), 1990.
[22] G. Kramer, "Directed information for channels with feedback," Ph.D.
dissertation, ETH Zurich, 1998.
[23] J. Jiao, H. H. Permuter, L. Zhao, Y.-H. Kim, and T. Weissman,
"Universal estimation of directed information," IEEE Trans. Inf. Theory,
vol. 59, no. 10, pp. 6220 -- 6242, 2013.
[24] Y. Liu, "Directed information for complex network analysis from
multivariate time series," Ph.D. dissertation, Michigan State University,
East Lansing, MI, USA, 2012.
[25] Y. Liu and S. Aviyente, "The relationship between transfer entropy
and directed information," in IEEE Statistical Signal Process. Workshop
(SSP), Aug 2012, pp. 73 -- 76.
[26] A. Rao, A. O. Hero, D. J. States, and J. D. Engel, "Inference of biolog-
ically relevant gene influence networks using the directed information
criterion," in IEEE Int. Conf. on Acoust., Speech and Signal Process.
(ICASSP), vol. 2, 2006.
[27] S. Tatikonda and S. Mitter, "The capacity of channels with feedback,"
IEEE Trans. Inf. Theory, vol. 55, no. 1, 2009.
[28] H. H. Permuter, Y.-H. Kim, and T. Weissman, "Interpretations of di-
rected information in portfolio theory, data compression, and hypothesis
testing," IEEE Trans. Inf. Theory, vol. 57, no. 6, 2011.
[29] P.-O. Amblard and O. J. Michel, "On directed information theory and
Granger causality graphs," J. of Computational Neuroscience, vol. 30,
no. 1, 2011.
[30] T. Cover and J. Thomas, Elements of information theory. Wiley, 2006.
[31] L. Barnett and A. K. Seth, "The MVGC multivariate Granger causality
toolbox: a new approach to Granger-causal inference," J. of Neuro-
science Methods, vol. 223, pp. 50 -- 68, 2014.
[32] C. Diks and J. DeGoede, "A general nonparametric bootstrap test for
Granger causality," in Global Analysis of Dynamical Systems, 2001.
[33] D. N. Politis and J. P. Romano, "The stationary bootstrap," J. of the
Amer. Statistical association, vol. 89, no. 428, pp. 1303 -- 1313, 1994.
[34] C. Wilke, G. Worrell, and B. He, "Graph analysis of epileptogenic
networks in human partial epilepsy," Epilepsia, vol. 52, no. 1, 2011.
[35] S. Sabesan, L. B. Good, K. S. Tsakalis, A. Spanias, D. M. Treiman, and
L. D. Iasemidis, "Information flow and application to epileptogenic focus
localization from intracranial EEG," IEEE Trans. Neural Syst. Rehabil.
Eng., vol. 17, no. 3, pp. 244 -- 253, 2009.
[36] F. Panzica, G. Varotto, F. Rotondi, R. Spreafico, and S. Franceschetti,
"Identification of the epileptogenic zone from stereo-EEG signals: a
connectivity-graph theory approach," Frontiers in Neurology, 2013.
[37] K. J. Blinowska and M. Kami´nski, "Multivariate signal analysis by para-
metric models," Handbook of Time Series Analysis: Recent Theoretical
Developments and Applications, p. 373, 2006.
16
[38] R. Malladi, G. P. Kalamangalam, and B. Aazhang, "Online bayesian
change point detection algorithms for segmentation of epileptic activity,"
in Asilomar Conf. on Signals, Systems and Computers, 2013.
[39] S. P. Meyn and R. L. Tweedie, Markov chains and stochastic stability.
Cambridge University Press, 2009.
[40] N.-Z. Shi and J. Tao, Statistical hypothesis testing: theory and methods.
[41] S. M. Kay, Fundamentals of statistical signal processing : estimation
World Scientific, 2008.
theory. Prentice-Hall, 2010.
[42] A. J. Izenman, "Review papers: recent developments in nonparametric
density estimation," J. of the Amer. Statistical Assoc., vol. 86, no. 413,
pp. 205 -- 224, 1991.
[43] D. W. Scott, Multivariate density estimation: theory, practice, and
visualization.
John Wiley & Sons, 2015.
[44] P. D. Grunwald, The minimum description length principle. MIT press,
2007.
[45] A. R. Barron and T. M. Cover, "Minimum complexity density estima-
tion," IEEE Trans. Inf. Theory, vol. 37, no. 4, 1991.
[46] T. Duong et al., "ks: Kernel density estimation and kernel discriminant
analysis for multivariate data in R," J. of Statistical Software, vol. 21,
no. 7, pp. 1 -- 16, 2007.
[47] D. Wied and R. Weissbach, "Consistency of the kernel density estimator:
a survey," Statistical Papers, vol. 53, no. 1, pp. 1 -- 21, 2012.
[48] G. Hu and R. O'Connell, "Analytical inversion of symmetric tridiagonal
matrices," J. of Physics A: Math. and General, vol. 29, no. 7, 1996.
[49] K. Ishiguro, N. Otsu, M. Lungarella, and Y. Kuniyoshi, "Comparison
of nonlinear Granger causality extensions for low-dimensional systems,"
Physical Review E, vol. 77, no. 3, p. 036217, 2008.
[50] C. Quinn, N. Kiyavash, and T. P. Coleman, "Directed information
graphs," arXiv preprint arXiv:1204.2003, 2012.
[51] P. Mierlo, E. Carrette, H. Hallez, R. Raedt, A. Meurs, S. Vandenberghe,
D. Roost, P. Boon, S. Staelens, and K. Vonck, "Ictal-onset localization
through connectivity analysis of intracranial EEG signals in patients with
refractory epilepsy," Epilepsia, vol. 54, no. 8, pp. 1409 -- 1418, 2013.
[52] C. Tonini, E. Beghi, A. T. Berg, G. Bogliun, L. Giordano, R. W. Newton,
A. Tetto, E. Vitelli, D. Vitezic, and S. Wiebe, "Predictors of epilepsy
surgery outcome: a meta-analysis," Epilepsy research, vol. 62, no. 1, pp.
75 -- 87, 2004.
[53] Q. Wang, S. R. Kulkarni, and S. Verd´u, "Universal estimation of
information measures for analog sources," Foundations and Trends in
Communications and Information Theory, vol. 5, pp. 265 -- 353, 2009.
[54] C. J. Stam, G. Nolte, and A. Daffertshofer, "Phase lag index: assessment
of functional connectivity from multi channel EEG and MEG with
diminished bias from common sources," Human brain mapping, vol. 28,
no. 11, pp. 1178 -- 1193, 2007.
[55] L. Faes and G. Nollo, Multivariate frequency domain analysis of causal
INTECH Open Access
time series.
interactions in physiological
Publisher, 2011.
[56] C. P. Warren, S. Hu, M. Stead, B. H. Brinkmann, M. R. Bower, and G. A.
Worrell, "Synchrony in normal and focal epileptic brain: the seizure
onset zone is functionally disconnected," J. of neurophysiology, vol. 104,
no. 6, pp. 3530 -- 3539, 2010.
[57] S. P. Burns, S. Santaniello, R. B. Yaffe, C. C. Jouny, N. E. Crone, G. K.
Bergey, W. S. Anderson, and S. V. Sarma, "Network dynamics of the
brain and influence of the epileptic seizure onset zone," Proceedings of
the National Academy of Sciences, vol. 111, no. 49, 2014.
[58] R. Malladi, G. Kalamangalam, N. Tandon, and B. Aazhang, "Identifying
seizure mechanisms from ECoG data using directed information," in
Cosyne Abstracts, Salt Lake City, USA, 2016.
Rakesh Malladi is a graduate student
in Electrical and Computer Engineering
Department at Rice University, Houston
since fall 2011. He graduated with a dual
degree (B.Tech & M.Tech) in Electrical
Engineering from Indian Institute of Tech-
nology Madras in 2011. He was a research
intern at Cyberonics, Houston, USA dur-
ing the summer of 2015, at Texas Instru-
ments, Dallas, USA during the summer
of 2013 and at IBM Research, Bangalore,
India during the summer of 2010. His research interests spans
topics in signal processing, machine learning, computational
neuroscience and information theory.
Giridhar Kalamangalam was born
in Salem, India in 1965. He received
the MBBS degree from the Jawahar-
lal Institute in Pondicherry, India in
1989, and the MSc and DPhil degrees
in applied mathematics from Oxford
University, UK,
in 1991 and 1995.
Following postgraduate general medi-
cal training (MRCP (UK)) he qualified
as a neurologist (Glasgow, UK) with subspecialty expertise
in epilepsy (Cleveland Clinic, USA). Since 2006, he has
been with the University of Texas Health Science Center
in Houston, TX, where he is currently Associate Professor
of Neurology. His research interests are in the physiological
dynamics and the neuroimaging of epilepsy and cognitive
function.
Nitin Tandon, MD, FAANS is a pro-
fessor of neurosurgery at University of
Texas Medical School at Houston. He
received his medical degree at the Armed
Forces Medical College in Pune, India,
followed by a residency in neurosurgery
at The University of Texas Health Science
Center in San Antonio and a fellowship in
epilepsy surgery at Cleveland Clinic. He
has been in practice for thirteen years at
Memorial Hermann TMC and on faculty
at the UT Health Medical School. He has
performed more than 3000 brain operations, with over 1200
for brain tumors and 600 for epilepsy. His lab performs multi-
modality assessments of cognitive functions combining and
correlating intracranial recordings, functional MRI, tractogra-
phy and direct cortical stimulation. The thrust of his research is
the development of optimal tools to characterize interactional
brain processes across regions using intracranial recordings
(www.tandonlab.org)
Behnaam Aazhang received his B.S.
(with highest honors), M.S., and Ph.D.
degrees in Electrical and Computer En-
gineering from University of Illinois at
Urbana-Champaign in 1981, 1983, and
1986, respectively. In August 1985, he
joined the faculty of Rice University,
Houston, Texas, where he is now the J.S.
Abercrombie Professor in the Department
of Electrical and Computer Engineering
Professor and Director of Center on Neuro-Engineering, a
multi-university research center in Houston, Texas. He is a
Fellow of IEEE and AAAS, and also a recipient of 2004 IEEE
Communication Society's Stephen O. Rice best paper award
for a paper with A. Sendonaris and E. Erkip. In addition,
the authors received IEEE Communication Society's 2013
Advances in Communication Award for the same paper. He has
been listed in the Thomson-ISI Highly Cited Researchers and
has been keynote and plenary speaker of several conferences.
|
1202.4482 | 2 | 1202 | 2013-02-09T21:34:51 | Metabolic cost as an organizing principle for cooperative learning | [
"q-bio.NC",
"cs.LG",
"nlin.AO"
] | This paper investigates how neurons can use metabolic cost to facilitate learning at a population level. Although decision-making by individual neurons has been extensively studied, questions regarding how neurons should behave to cooperate effectively remain largely unaddressed. Under assumptions that capture a few basic features of cortical neurons, we show that constraining reward maximization by metabolic cost aligns the information content of actions with their expected reward. Thus, metabolic cost provides a mechanism whereby neurons encode expected reward into their outputs. Further, aside from reducing energy expenditures, imposing a tight metabolic constraint also increases the accuracy of empirical estimates of rewards, increasing the robustness of distributed learning. Finally, we present two implementations of metabolically constrained learning that confirm our theoretical finding. These results suggest that metabolic cost may be an organizing principle underlying the neural code, and may also provide a useful guide to the design and analysis of other cooperating populations. | q-bio.NC | q-bio |
Metabolic cost as an organizing principle
for cooperative learning
David Balduzzi1, Pedro A. Ortega2, Michel Besserve2
1Swiss Federal Institute for Technology, ETH Zurich
2Max Planck Institute for Intelligent Systems, Tubingen, Germany
[email protected], {ortega, besserve}@tuebingen.mpg.de
Abstract
This paper investigates how neurons can use metabolic cost to facilitate
learning at a population level. Although decision-making by individual
neurons has been extensively studied, questions regarding how neurons
should behave to cooperate effectively remain largely unaddressed. Un-
der assumptions that capture a few basic features of cortical neurons, we
show that constraining reward maximization by metabolic cost aligns the
information content of actions with their expected reward. Thus, metabolic
cost provides a mechanism whereby neurons encode expected reward into
their outputs. Further, aside from reducing energy expenditures, impos-
ing a tight metabolic constraint also increases the accuracy of empirical
estimates of rewards, increasing the robustness of distributed learning. Fi-
nally, we present two implementations of metabolically constrained learning
that confirm our theoretical finding. These results suggest that metabolic
cost may be an organizing principle underlying the neural code, and may
also provide a useful guide to the design and analysis of other cooperating
populations.
1 Introduction
Rational decision making is typically formalized as optimizing a reward function [19,22,26].
This paper investigates neuronal learning from an optimization perspective. We assume that
both the brain as a whole and individual neurons are rational decision makers optimizing
reward functions of some kind.
Since the brain learns from finite samples, it is exposed to a tradeoff between over- and
under-fitting: increasing a model's capacity can improve its fit on training data, but poten-
tially worsens performance on future samples [29]. Remarkably, however, the human brain
effortlessly handles a wide-range of complex pattern recognition tasks suggesting it both
has a large capacity and, paradoxically, also generalizes extremely well. On the basis of
these conflicting observations, it has been argued that useful biases in the form of "generic
mechanisms for representation" must be hardwired into cortex [10]. Our goal in this paper
is to propose a bias that is both useful and biologically plausible.
Let us outline the problem. Neurons learn inductively. They can generalize from finite
samples and encode estimates of future outcomes (for example, rewards) into their spike-
trains [12]. Learning-theoretical results imply that generalizing successfully from small sam-
ples requires strong biases [29] or, in other words, specialization. Thus, at any given time
some neurons' specialties are more relevant than others. Since most of the data neurons
receive are other neurons' outputs, it is essential that they indicate which of their outputs
1
encode high quality estimates. Downstream neurons should then be biased to specialize on
these outputs, thereby reducing the effective search space that neurons explore.
The question we ask is: How can a single neuron near-maximize its expected reward and
simultaneously help downstream neurons do the same?
As a partial answer, we propose the following organizing principle. Neurons should consis-
tently label outputs as useful and not useful. Specifically, the optimizations performed by
neurons should be designed so that spikes are useful and silences are not. For example, it is
well known that, on average, the more spikes a neuron receives, the more it learns by modi-
fying its synapses. Although this fact is often taken for granted, it requires explanation. We
suggest that spikes drive learning because they are useful. By useful, we mean an output
that (i) predicts high reward with (ii) tight confidence intervals.
Distinguishing useful from irrelevant outputs is helpful to downstream neurons since it
reduces the size of the search space they are confronted with.
If spikes reliably predict
future reward, then neurons should be biased towards learning from spikes. In fact, this fits
well with experimental evidence [8, 9, 20]. Moreover, consistently biasing learning toward
outputs labeled as useful (i.e. spikes) also provides a principled way to reduce capacity, and
thus improve generalization guarantees, without sacrificing empirical performance.
The scope of this paper is limited to showing how neurons may systematically distinguish
useful from irrelevant outputs. Fleshing out the implications for learning at the population
level is deferred to future work. Furthermore, we only consider excitatory connections in
this paper; the role of inhibition is also deferred to future work.
Overview. We explore the consequences of a few basic assumptions regarding spikes,
metabolic cost and neuronal rewards, see §2. We consider a minimal model of biological
neural network from which we extract guiding principles that may apply, at least approx-
imately, to more complex, biologically realistic models. The most important assumptions
are that neurons aim to maximize reward after spiking, and that neurons have to operate
within a fixed metabolic budget that constrains how often they can spike in a given time
interval.
It turns out that these seemingly innocuous assumptions are the key to distinguishing useful
from irrelevant outputs. Our first result is that constrained reward maximization causes
neurons to maximize the information they encode into their spikes, see §3. Moreover, if
spikes are sufficiently rare, it turns out that spikes dominate the information communicated
by neurons, which has interesting implications for credit assignment in cortex, see §3 and [5].
If neurons attempt to maximize empirical reward after spiking, it follows that neurons encode
reward estimates into spikes. When a neuron produces a spike, it thus signals to downstream
neurons that it expects a positive neuromodulatory signal such as dopamine. This raises
questions concerning the quality of the reward estimates encoded in spikes, see §4. We
show that the more information neurons encode into spikes, the tighter the guarantees tying
empirical reward to expected reward.
We conclude by describing some implications of our results for cooperative optimization.
Theorems are proved in the appendices.
Related work. Neuronal plasticity and its implications for the neural code have been
intensively studied for many years. The work closest in spirit to this paper is Seung's
"hedonistic" synapses, which seek to increase average reward [24]. A second related line
of research applies the information bottleneck method – an alternate constraint to the one
considered here – to neuronal learning [7, 27].
An information-theoretic perspective on synaptic homeostasis that complements the results
in this paper is [5]. The consequences of constrained optimization for spike-timing dependent
plasticity (STDP) are presented in [2]. A practical implementation of regularized STDP,
inspired by the ideas presented here, can be found in [16].
Acknowledgements. We thank Yevgeny Seldin and Giulio Tononi for useful discussions.
2
2 A minimal model
The mammalian cortex contains between 107 and 1011 neurons (depending on the species)
that are guided by neuromodulators signaling pleasure, pain and other globally salient
events. Neurons communicate with each other via spiketrains – sequences of silences and
spikes they receive from and transmit to 103 to 104 other neurons through connections
called synapses. Neurons learn by increasing and decreasing the efficacy of their synapses
according to the timing of pre-synaptic (input) and post-synaptic (output) spikes, as well
as the presence or absence of neuromodulators [8, 9, 15, 20, 25].
Model neurons. The cortex can be modeled as a population of K neurons. Neuron nk
follows policy πk, where πk(as) is a Markov matrix specifying the probability the neuron
picks action, or output, a ∈ A = {0, 1} upon encountering situation s ∈ S. A situation, or
input, is a vector s = (a1, . . . , aK) whose entries are the actions of all the neurons in the brain
at the previous time step. Since each neuron is only exposed to a small fraction of the brain,
it ignores most entries in the vector. Thus, the policy of neuron k is πk(as) = πk(aΣk)
for some subset Σk ⊂ {a1, . . . , aK}. Let P (S) denote the prior over situations. We will
assume the prior is i.i.d. when providing guarantees on estimates, see Remark 4 for a brief
discussion.
Neurons are exposed to a global neuromodulatory signal ν ∈ N that signals the performance
of the population as a whole. Neuromodulators are drawn with probability P (νs).
We make the following assumptions.
Assumption 1 (reward maximization).
Neurons maximize a reward function that depends on neuromodulatory signals, input spikes
and output spikes:
Rk(si, ai, νi),
(1)
(cid:98)πk = arg max
π∈M
N(cid:88)
i=1
where ai is the output chosen by π in response to input si.
The set M is the set of possible neuronal mechanisms, a subset of the set Markov matrices
on S × A. Two examples of M that are relevant to our discussion are discrete threshold
neurons
M =(cid:8)H((cid:104)w, s(cid:105) − ϑ)(cid:12)(cid:12) w ∈ RK(cid:9)
where H(•) is the Heaviside function, see [2], and the full set of Markov matrices, see
discussion of Q-learning below.
Note that, since different neurons are exposed to different subsets of the total brain activity,
they have different reward structures, even for the same neuromodulators: in general Rk (cid:54)=
Rj for j (cid:54)= k since Σk (cid:54)= Σj.
Since we focus on the behavior of a single neuron, we will often drop the subscript k from
the notation below.
Assumption 2 (spikes gate rewards).
The reward function is gated by synaptic outputs:
R(s, a, ν) = R(s, ν) · Ia=1,
where Ia=1 =
is the indicator function.
(2)
(cid:26)1 a = 1
0
else
Neurophysiological evidence suggests that neurons only potentiate or depotentiate their
synapses shortly before or after producing spikes [8,9,15,25]. This encourages specialization:
neurons search for a small set of inputs that reliably predict future reward signals – other
inputs are ignored.
For simplicity we assume neurons have two outputs, spikes and silence. We use notations
a0 or a = 0 for silence and a1 or a = 1 for spikes.
3
Assumption 3 (metabolic budget).
Neurons have a fixed metabolic budget that determines the maximum frequency of spiking
over some (unspecified) time period:
where π(a1) =(cid:80)
s p(s) · π(a1s) is the spiking frequency under policy π.
π(a1) ≤ ρ
(3)
Although a soft constraint is more biologically plausible, imposing a hard constraint simpli-
fies the exposition without significantly altering the conclusions. The main effect of softening
the constraints is to allow the information carried by spikes and the capacity of neurons to
vary, thereby softening Theorems 2 and 3 below.
Since neurons only modify their synapses when they spike, it follows that neurons that spike
very infrequently learn very little if at all. Thus, we expect that there are mechanisms in
place ensuring that neurons not only stay within their metabolic budget, but also that they
come close to using all of it.
Spikes and silences are not abstract, interchangeable symbols. Spiking and responding to
spikes carries a much higher metabolic cost than not spiking [13]. This cost is significant
since the nervous system consumes a disproportionate share of an organism's total energy
budget [1]. It has been hypothesized that a function of sleep is to homeostatically regulate
synaptic strengths, so as to control metabolic expenditures associated with action potentials,
see [11, 14, 28, 30].
Unlike "agents in the wild", individual neurons have negligible impact on what happens
next: a single neuron has little effect on the neurons it targets – since each receives inputs
from thousands of other neurons. The inability of individual neurons to manipulate their
environment simplifies the optimization problems they face by stripping out the recursive
Bellman aspect [6]:
Assumption 4 (disempowerment).
An individual neuron has no immediate influence on its environment
(cid:12)(cid:12)(cid:12) s(t), ak(t)
(cid:17)
(cid:16)
(cid:12)(cid:12)(cid:12) s(t)
(cid:17)
p
s(t + 1, . . .)
= p
s(t + 1, . . .)
,
(4)
(cid:16)
where t refers to time. I.e. conditioning on the neuron's output makes no difference to the
distribution on subsequent actions by the rest of the brain.
The future situations a neuron encounters are essentially unaffected by its output. This
assumption fails at the population level – populations of neurons necessarily affect the
organism's actions. Nevertheless Eq. (4) is a reasonable assumption at the individual neuron
level since, for example, destroying a single neuron makes essentially no difference to brain
function.
3 Encoding useful information in spikes
This section considers the implications of our assumptions for the information content of
spikes. Mutual information provides a formal method for quantifying the information con-
tent of a channel. We present a related measure, effective information, that measures the
information content of a single output in terms of how much that output reduces prior
uncertainty regarding the set of inputs.
Importantly, we show that under the above as-
sumptions, spikes not only reduce uncertainty, but also signal that the neuron received an
input that was historically followed by high rewards. The information encoded in spikes is
thus indicative of expected future rewards.
Information.
It is useful to consider neurons as communication channels mapping situ-
ations to actions. The average information communicated by a neuron is then the mutual
information Iπ(S,A). However, since we are interested in the information communicated
by specific actions (in particular, spikes), we introduce effective information1 [3, 4] which
quantifies the information that a single output encodes about the input.
1We extend the definition of effective information in [3] to allow arbitrary priors instead of
restricting to the uniform distribution.
4
Definition 1 (effective information).
Given a neuron with policy π(as) and prior P (S) on situations, the effective information
generated by action a ∈ A is
ei(π, a) := D(cid:2)π(Sa)(cid:13)(cid:13) P (S)(cid:3) where π(sa) :=
π(as)
π(a)
· P (s)
is computed via Bayes' rule and D[•(cid:107)•] is the Kullback-Leibler divergence D[p(cid:107)q]
(5)
:=
(cid:80) pi log pi
qi
.
An interesting special case is when the prior on situations is the uniform distribution and
the policy is deterministic. It follows that
ei(π, a) = − log
π−1(a)
S
,
(6)
where • denotes cardinality and (since the policy is deterministic), π is a function π : S →
A. In Eq (6), effective information quantifies the selectivity of an output: the fraction of
inputs causing the policy to output a. The smaller the fraction, or alternatively the more
sensitive output a is to perturbations in the input, the higher effective information [5].
Remark 1. Suppose we have model PM(dh) that specifies the probability of observing data
given a hypothesis. Further suppose we have prior distribution P (h) on hypotheses. If we
observe data d, how much have we learned about the hypotheses? The Bayesian information
gain is
(7)
If we consider a neuron's policy as a model, with inputs as hypotheses and outputs as
evidence, then effective information quantifies the Bayesian information gained about the
inputs given an output.
Remark 2. The expectation of ei is mutual information: E
(cid:2)ei(π, a)(cid:3) = Iπ(S,A).
π(a)
D(cid:2)PM(Hd)(cid:13)(cid:13) P (H)(cid:3).
(cid:98)R(s, a) :=
1
N
(cid:88)
Information aligns with rewards (theory). We show that neurons implementing con-
strained reward maximization from Assumption 1 also maximize the effective information
of their spikes ei(π, a1), that we will call information per spike.
Definition 2 (empirical reward).
Given a finite sample of situations, actions and neuromodulators (si, ai, νi)N
pirical reward observed after performing action a in situation s be
i=1, let the em-
R(si, ai, νi).
{isi=s,ai=a}
We also introduce the empirical reward after spiking, 1
N
pirical reward per spike
(cid:1) and the em-
(cid:80){iai=1} R(cid:0)si, a1, νi
Assume that situations yield different empirical rewards, i.e. (cid:98)R(s, a1) (cid:54)= (cid:98)R(s(cid:48), a1) for all
s, s(cid:48) ∈ S. Recall that by Assumptions 1 and 3 an optimal policy (cid:98)π satisfies
Theorem 1 (maximizing reward/spike maximizes information/spike).
(cid:80){iai=1} R(cid:0)si, a1, νi
N(cid:88)
{iai=1}
(cid:1).
1
(cid:98)π =
arg max
{π∈Mπ(a1)≤ρ}
R(si, ai, νi),
i=1
If we also apply constraint π(a1) ≥ ρ, i.e. ρ is both upper and lower bound, then the optimal
policy (cid:98)π maximizes information per spike. More precisely, the optimal policy (cid:98)π satisfies
Remark 3. The optimal policy will satisfy(cid:98)π(a1) = ρ if there are enough inputs s satisfying
ei((cid:98)π, a1) ≥ ei(π, a1) for all π such that π(a1) = ρ.
Eν[R(s, ν, a1)] > 0. In other words, if there are enough situations where spiking, on average,
is followed by positive reward. Imposing the lower bound in the theorem means we compare
the optimal policy with alternate policies that spike with the same frequency.
Thus, the optimal policy necessarily maximizes both the empirical reward after spiking and
the effective information encoded in spikes. We illustrate this result with simulations below.
5
Information aligns with rewards (experiments). A learning algorithm implementing
constrained reward maximization is a modification of Q-learning [31]:
Example 1 (Metabolically constrained Q-learning).
If a neuron chooses action a in situation s and subsequently receives neuromodulator ν,
then it updates the Q-matrix by
Q(s, a) ← Q(s, a) + α ·(cid:104)(cid:98)R(s, a, ν) − Q(s, a)
(cid:105)
,
where α controls the rate. After updating Q, the neuron constructs new policy
π(as) = Mn(cid:16)
eQ(s,a)(cid:17)
.
Operation M(•)
s∈S π(as)P (s) = P (a) for all a, and then by Z(s) chosen such that (cid:80)
(cid:80)
first by Z(a) chosen such that
a∈A π(as) = 1
for all s. Setting n = 3 yields a policy that approximately implements the metabolic con-
straint.
the policy twice:
renormalizes
Figure 1 shows how effective information and empirical reward after spiking covary as neu-
rons Q-learn. We initialized 5000 neurons randomly and applied metabolic constraints
ρ ∈ {0.1, 0.3, 0.5}. Rewards are drawn randomly. As the neurons adapt, their policies be-
come both more deterministic and more likely to spike in situations yielding higher rewards,
so as neurons adapt they both encode more information into their spikes and predict higher
rewards after spiking. The tighter the metabolic constraint (i.e. the lower ρ), the higher the
empirical reward after spiking. Thus, the information encoded in spikes provides a reliable
guide to the empirical reward after spiking.
Spikes dominate information content. Theorem 1 shows information per spike is max-
imized by constrained reward maximization. Theorem 2 below consider how much of the
total information communicated by a neuron is carried by spikes.
Theorem 2 (spikes carry essentially all information).
Suppose a neuron has two actions (silence a0 and spike a1) and produces spikes infrequently:
π(a1) (cid:28) 1. Then the total information communicated by the neuron is approximately the
information it communicates using spikes alone:
Iπ(S;A) = π(a1) · ei(π, a1) + O(cid:0)π(a1)2(cid:1) .
(8)
If the metabolic constraint is tight, meaning ρ and so π(a1) are small, then for the optimal
policy encodes a lot of information into spikes. In this setting, Theorem 2 implies that the
information communicated by a neuron is (up to first order) carried by spikes alone.
Eq (6) and Theorem 2 together have interesting implications for credit assignment, see [5]
for details. In particular, if spikes carry most of the information in cortex, then spiking neu-
rons and synapses should reinforced in response to positive global signals such as dopamine,
and conversely for negative global signals. Neurons and synapses that are silent contribute
little to the information generated by cortex, and so should be neither potentiated nor depo-
tentiated. This fits neurophysiological evidence suggesting that spikes play a distinguished
role in synaptic potentiation and depotentiation [8, 9, 15, 20].
pirical rewards(cid:8)(cid:98)R(s, a1)(cid:12)(cid:12)s ∈ S(cid:9). The optimal policy is constructed as follows. First, rank
Encoding reward estimates in spikes. Finally, we briefly illustrate the effect of the
metabolic constraint ρ on the empirical reward per spike. Suppose we have sampled em-
states by their empirical reward. Let Sρ denote the states in the top ρth percentile. Define
(cid:98)π(a1s) =
s ∈ Sρ
else.
(cid:26)1
0
The optimal policy can be visualized as moving a window of fixed size and variable shape
over the input space, such that the underlying configuration maximizes reward.
6
Figure 2 shows an example. Situations are ranked according to their empirical reward. The
metabolic constraint is set at ρ = 15%, so the optimal policy spikes for the 15% of situations
with highest reward. The policy picking these situations results in a reward per spike of
0.89. Situations which do not cause spikes receive no reward ; if spikes and silences would
be exchanged, this reward would be .21 (we call it reward after silence).
We make two observations. First, tightening the metabolic constraint, so that the policy
spikes for (cid:28) 15% of situations, increases the reward per spike. Second, the variance in
reward per spike is much lower than after not spiking, and typically decreases with ρ.
As a general rule of thumb, tightening the metabolic constraint by decreasing ρ both in-
creases the empirical reward per spike and reduces the variance in empirical reward per
spike.
4 Guarantees on reward estimates
In a dynamically changing environment like the cortex, it is important that neurons reliably
represent high reward. This section shows that the reliability of spikes depends on how
much information is encoded in them.
Theoretical guarantees. We say that a neuron's spikes reliably represent reward when
there is a low variability in the empirical mean reward when the neuron spikes. For a given
policy π. The expected and empirical rewards per spike are
Rπ := E(cid:104)
(cid:98)Rπ :=
R(S, a1, N )
(cid:88)
s,n
(cid:105)
(cid:88)
:=
1
T1
{(st,νt)π(st)=1}
π(sa1) · P (νs) · R(s, a1, ν) and
R(cid:0)st, a1, νt
(cid:1) respectively,
where S is the total number of possible situations. Let πu(a) =(cid:80)
where T1 counts spikes produced by the neuron during [1, T ]. We measure sample size in
terms of spikes (rather than spikes and silences) because spikes are metabolically expensive,
so only spikes count towards the cost of collecting a finite sample.
Let us introduce notation for computations with respect to the uniform prior U (s) = 1S ,
s π(as)· U (s), πu(sa) =
π(as) U (s)
πu(a) . Note that eiu recovers the original notion
of effective information in [3], which Definition 1 generalizes.
π(a) , and eiu(π, a) =(cid:80)
s πu(sa) log πu(sa)
The following theorem is proved using a version of Occam's razor [23].
Theorem 3 (error bound for empirical reward).
Suppose the neuron chooses a deterministic policy π under the constraint that it spikes for
a fixed fraction of situations: πu(a1) = const. Further, suppose that situations are sampled
i.i.d. Without loss of generality,2 assume that rewards lie in [0, b]. Then with probability at
least 1 − δ,
(cid:115)
(cid:12)(cid:12)(cid:12)Rπ − (cid:98)Rπ
(cid:12)(cid:12)(cid:12) ≤ b ·
Guarantees improve as eiu increases since x+1
S · eiu(π, a1) + 1
2T1 · eeiu(π,a1) +
ex decreases as x increases.
log 2
δ
2T1
(9)
Note that ei and eiu covary since increasing the number of situations where a neuron spikes
decreases both ei and eiu; similarly, decreasing the number of situations where a neuron
spikes increases both ei and eiu.
Thus, tightening the metabolic constraint in Assumption 3, by choosing low ρ, yields policies
that have better guarantees on their reward estimates.
2Since only relative rewards affect the choice of optimal policy, it follows that negative rewards
can be stripped out of the optimization problem by introducing an additive constant.
7
Remark 4. The assumption that inputs are i.i.d.
is not realistic for cortical neurons. We
make two remarks. First, if rewards are only non-zero in the presence of neuromodulatory
signals, then the assumption states that situations directly preceding neuromodulator release
are i.i.d, which is more reasonable. Second, similar results have recently been obtained in
non-i.i.d. scenarios using more sophisticated PAC-Bayes methods [21] – and these may be
applicable to our setting.
Guarantees in practice. Figure 3 plots information encoded in spikes (x-axis) against
the difference between the normalized empirical and expected reward (y-axis). Rewards
were drawn randomly and the expected and empirical error for 16,000 deterministic policies
with S = 50, P (S) uniform and T1 = 20 were computed. Policies were sampled randomly
with k, the number of situations causing the policy to spike, varying uniformly across [1, 25].
The figure shows that both normalized error and the standard error of the error decrease
as ei increases. Figure 3 confirms that the bound in Theorem 3 is a reasonable guide to
performance in practice.
Thus, the more information encoded in spikes, the better a neuron's empirical estimate of its
expected reward per spike. The metabolic constraint in Assumption 3 controls the quality
of a neuron's empirical estimates of its expected reward.
5 Discussion
The space of possible policies that the cortex as a whole could implement is vast: it consists
in choosing synaptic weights of millions or billions of neurons each receiving inputs from
thousands or tens of thousands of other neurons. However, the space of policies that makes
sense biologically is probably much smaller due to biological as well as learning-theoretic
considerations. This paper has shown that metabolically constrained reward maximization
provides a biologically plausible way for neurons to distinguish useful outputs from those
that are not. It follows from our assumptions that spikes:
• are responsible for most of the information communicated by neurons;
• signal when, based on empirical estimates, neurons predict high reward; and
• come equipped with performance guarantees that increase as the metabolic con-
straint is tightened, and so more information is encoded in spikes.
This suggests that neurons should privilege spikes during learning; and indeed there is a
large body of experimental evidence that this is exactly what occurs [8, 9, 15, 25]: synaptic
plasticity is triggered by pre- and post-synaptic spikes, with the decision to potentiate or
depotentiate depending on their precise timing.
An important unanswered question is setting the metabolic constraint ρ. Clearly, if ρ is
too low, then neurons will barely fire at all, which is not desirable. Conversely, if ρ is too
high then information and empirical reward encoded in spikes, as well as the quality of the
empirical estimates, all degrade, which is also to be avoided. In this paper, we have simply
highlighted the metabolic constraint ρ as an important lever that neurons may actively
manipulate. Finding an optimal value or range of values for ρ is outside the scope of this
paper.
Biasing neuronal mechanisms. Q-learning is not a practical learning rule; it simply
constructs a giant lookup table. If we think of the set of inputs causing a neuron to fire as
a window of varying shape, then the metabolic constraint fixes the size of the window and
Q-learning places no other constraints on its shape.
However, if upstream neurons systematically encode information and reward in spikes, then
it makes sense to bias shape of downstream neuronal "firing windows" to take the asymmetry
between spikes and silences into account. This is exactly what we occurs in cortex. The
vast majority of synapses are excitatory: the more input spikes neurons receive, the more
likely they are to spike themselves.
8
Cortical neurons are biased toward firing for more spikes, which makes sense if spikes are
reliable predictors of future reward. Thus, neuronal "firing windows" are shaped such that if
a firing pattern causes a neuron to spike, then so does any firing pattern containing strictly
more spikes. Moreover, sufficiently many pre-synaptic spikes will (essentially) always cause
a neuron to fire.
Thus, neurons aggregate evidence for high reward (spikes from upstream neurons), and
modify their synaptic strengths to maximize their empirical reward. Fine-tuning is necessary
since no two neurons have exactly the same connectivity, and therefore no two neurons use
the same data to predict global neuromodulatory signals.
Inhibitory (GABA) synapses do not fit this picture. Extending our framework for plasticity
to include both types of neurons is challenging. However, we conjecture this incompatibility
can be overcome, with inhibition playing a complementary role; possibly centered on im-
posing sparse activity in the brain [17, 18] and selecting competing neural assemblies and
structures.
Global versus overlapping local optimizations. The results in this paper depend
on specific assumptions and model choices. A particularly important assumption is that
all neurons attempt to maximize the same global neuromodulatory reward signal. In this
scenario, since neurons differ in their connectivity, they have access to different subsets
of brain activity representing different environemental features and events, and therefore
specialize on different sources of reward.
The reality is more complicated with multiple overlapping neuromodulatory signals includ-
ing dopamine, noradrenaline, acetylcholine and others. Moreover, neurons involved in, say,
early visual processing may not require neuromodulatory guidance; rather, they may search
for stable invariants over short time frames (hundreds of milliseconds). There is likely a
diverse array of reward functions implemented across cortex.
It is thus unclear whether the brain can be accurately described as optimizing a single well-
defined reward function. Nevertheless, until decisively shown to be false, we believe the
optimization perspective to be a fruitful working hypothesis. Even if it does not apply to
the brain as a whole, it may nevertheless provide insight into how populations of neurons
in specific brain areas converge on useful behaviors.
A spiking currency. Finally, it is interesting to speculate on an analogy between spikes
and paper currency. Money plays many overlapping roles in an economy, including: (i)
focusing attention; (ii) stimulating activity; and (iii) providing a quantitative lingua franca
for tracking revenues and expenditures. Note that, as for the brain, it is unclear whether
an economy as a whole can be reduced to optimizing a single well-defined function.
Spikes may play similar roles in cortex to those of paper currency in an economy. Spikes
focus attention: STDP and other proposed learning rules are particularly sensitive to spikes
and spike timing. Spikes stimulate activity: input spikes cause output spikes. Finally, spikes
leave trails of (Calcium) traces that are used to reinforce and discourage neuronal behaviors
in response to neuromodulatory signals.
Neither money nor spikes are intrinsically valuable. Currency can be devalued by inflation.
Similarly, the information content and guarantees associated with spikes can be eroded
by overpotentiating synapses which reduces their selectivity (potentially leading to epileptic
seizures in extreme cases). Regulating the information content of spikes is therefore essential.
Assumption 3 provides a simple constraint that can be approximately imposed by regulating
synaptic weights. Indeed, there is evidence that one of the functions of sleep is precisely
this [11, 14, 28, 30].
Spikes with high information content are valuable because they come with strong guarantees
on their estimates. They are therefore worth paying attention to, worth responding to, worth
keeping track of, and worth learning from.
9
References
[1] Attwell, D. and Laughlin, S. B., An energy budget for signaling in the grey matter of
the brain, J Cereb Blood Flow Metab 21 (2001) 1133–45.
[2] Balduzzi, D. and Besserve, M., Towards a learning-theoretic analysis of spike-timing de-
pendent plasticity, Advances in Neural Information Processing Systems (NIPS) (2012).
[3] Balduzzi, D. and Tononi, G., Integrated Information in Discrete Dynamical Systems:
Motivation and Theoretical Framework., PLoS Comput Biol 4 (2008) e1000091.
[4] Balduzzi, D. and Tononi, G., Qualia: the geometry of integrated information, PLoS
Comput Biol 5 (2009) e1000462.
[5] Balduzzi, D. and Tononi, G., What can neurons do for their brain? Communicate
selectivity with spikes, Theory in Biosciences (2012).
[6] Bellman, R., Dynamic Programming (Princeton University Press, 1957).
[7] Buesing, L. and Maass, W., Simplified rules and theoretical analysis for information
bottleneck optimization and PCA with spiking neurons, in Adv in Neural Information
Processing Systems (NIPS) (2007).
[8] Dan, Y. and Poo, M.-M., Spike timing-dependent plasticity of neural circuits., Neuron
44 (2004) 23–30.
[9] Dan, Y. and Poo, M.-M., Spike timing-dependent plasticity: from synapse to percep-
tion., Physiol Rev 86 (2006) 1033–1048.
[10] Geman, S., Bienenstock, E., and Doursat, R., Neural Networks and the Bias/Variance
Dilemma, Neural Comp 4 (1992) 1–58.
[11] Gilestro, G. F., Tononi, G., and Cirelli, C., Widespread changes in synaptic markers
as a function of sleep and wakefulness in Drosophila, Science 324 (2009) 109–12.
[12] Gottfried, J. A., O'Doherty, J., and Dolan, R. J., Encoding predictive reward value in
human amygdala and orbitofrontal cortex, Science 301 (2003) 1104–7.
[13] Hasenstaub, A., Otte, S., Callaway, E., and Sejnowski, T. J., Metabolic cost as a
unifying principle governing neuronal biophysics, Proc Natl Acad Sci U S A 107 (2010)
12329–34.
[14] Maret, S., Faraguna, U., Nelson, A. B., Cirelli, C., and Tononi, G., Sleep and waking
modulate spine turnover in the adolescent mouse cortex., Nat Neurosci 14 (2011) 1418–
1420.
[15] Markram, H., Lubke, J., Frotscher, M., and Sakmann, B., Regulation of synaptic effi-
cacy by coincidence of postsynaptic aps and epsps, Science 275 (1997) 213–5.
[16] Nere, A., Olcese, U., Balduzzi, D., and Tononi, G., A neuromorphic architecture for
object recognition and motion anticipation using burst-STDP, PLoS One 7 (2012)
e36958.
[17] Olshausen, B. A. and Field, D. J., Sparse coding with an overcomplete basis set: a
strategy employed by v1?, Vision Res 37 (1997) 3311–25.
[18] Olshausen, B. A. and Field, D. J., Sparse coding of sensory inputs, Curr Opin Neurobiol
14 (2004) 481–7.
[19] Ortega, P. A. and Braun, D. A., Information, utility and bounded rationality, The
fourth conference on artificial general intelligence (2011) 269–274.
[20] Pawlak, V., Wickens, J. R., Kirkwood, A., and Kerr, J. N. D., Timing is not everything:
neuromodulation opens the STDP gate, Front. Syn. Neurosci 2 (2010).
[21] Rubin, J., Shamir, O., and Tishby, N., Trading Value and Information in MDPs, in
Decision Making with Imperfect Decision Makers, eds. Guy, T. V., K´arn´y, M., and
Wolpert, D. (Springer, 2011).
[22] Russell, S. and Norvig, P., Artificial Intelligence: A Modern Approach, 3rd edn. (Pren-
tice Hall, 2009).
[23] Seldin, Y. and Tishby, N., Multi-Classification by Categorical Features via Clustering,
in Proceedings of the 25th International Conference on Machine Learning (2008).
10
[24] Seung, H. S., Learning in Spiking Neural Networks by Reinforcement of Stochastic
Synaptic Transmission, Neuron 40 (2003).
[25] Song, S., Miller, K. D., and Abbott, L. F., Competitive Hebbian learning through
spike-timing-dependent synaptic plasticity, Nature Neuroscience 3 (2000).
[26] Sutton, R. S. and Barto, A. G., Reinforcement Learning: An Introduction (MIT Press,
1998).
[27] Tishby, N., Pereira, F., and Bialek, W., The information bottleneck method, in Proc.
of the 37-th Annual Allerton Conference on Communication, Control and Computing,
eds. Hajek, B. and Sreenivas, R. (1999).
[28] Tononi, G. and Cirelli, C., Sleep and synaptic homeostasis: a hypothesis, Brain Res.
Bull. 62 (2003) 143–150.
[29] Vapnik, V., Estimation of Dependencies Based on Empirical Data (Springer, 1982).
[30] Vyazovskiy, V. V., Olcese, U., Lazimy, Y., Faraguna, U., Esser, S. K., Williams, J. C.,
Cirelli, C., and Tononi, G., Cortical firing and sleep homeostasis, Neuron 63 (2009)
865–78.
[31] Watkins, C. and Dayan, P., Q-learning, Machine Learning 8 (1992) 279–292.
A.1 Proof of Theorem 1
Proof. First we show that deterministic policies maximize the information encoded in spikes,
then we show that deterministic policies maximize reward per spike.
Deterministic policies maximize effective information subject to π(a1) ≤ ρ. Observe that
(cid:88)
(cid:88)
s∈S
s∈S
π(sa) log
π(sa) log
− log π(a)
(cid:124)
(cid:123)(cid:122)
π(sa)
P (s)
π(as)
π(a)
(cid:125)
log frequency of output a
ei(π, a) =
=
=
(cid:88)
(cid:124)
s∈S
+
π(sa) log π(as)
.
(cid:125)
(cid:123)(cid:122)
stochasticity
The log-frequency, or surprise, term is nonnegative and the stochasticity term is non-
positive.
It is easy to see that the stochasticity term is maximized at 0 if and only if
output a is chosen deterministically – i.e. π(as) is either 0 or 1 for all s ∈ S.
Deterministic policies maximize reward per spike. Suppose there are N situations ordered
ministic policy spiking only for the ρ · N policies with highest empirical reward maximizes
empirical reward after spiking.
according to their empirical reward, so that (cid:98)R(s1) < ··· < (cid:98)R(sN ). It is clear that a deter-
A.2 Proof of Theorem 2
Proof. Observe that
P (s) − π(sa1) · π(a1)
P (sa0) =
Thus to first order in π(a1), π(sa0) is of the form p + δp where (cid:82) δp = 0. We can then
= P (s) + π(a1)(cid:0)P (s) − π(sa1)(cid:1) + O(cid:0)π(a1)2(cid:1).
1 − π(a1)
compute
D(cid:2)p + δp
(cid:13)(cid:13) p(cid:3) =
(cid:90)
(cid:90)
(p + δp) log2
= α
(p + δp)
δp
p
p + δp
(cid:18)
p
1 − δp
2p
(cid:19)
+ O(cid:0)(δp)3(cid:1) = α
(cid:90) (δp)2
2p
+ O(cid:0)(δp)3(cid:1) ,
11
ln 2 .
where α = 1
Substituting p = P (s) and δp = π(a1)(cid:0)P (s) − π(sa1)(cid:1) gives
D(cid:2)π(Sa0)(cid:13)(cid:13) P (S)(cid:3) = D(cid:2)p + δp(cid:13)(cid:13) p(cid:3) + O(cid:0)π(a1)2(cid:1)
(cid:90) (cid:0)P (s) − π(sa1)(cid:1)
Thus I(S; A) = π(a1)D(cid:2)π(Sa1)(cid:13)(cid:13) P (S)(cid:3) + O(cid:0)π(a1)2(cid:1).
2P (s)
= α · π(a1)2
+ O(cid:0)π(a1)2(cid:1).
A.3 Proof of Theorem 3
Occam's razor can be paraphrased to say that the simplest hypothesis should be preferred.
Suppose we have a setof hypotheses H with prior distribution P (h) on H. Let − log P (h)
denote the complexity of hypothesis h. Let L : X × H → [0, b] be a loss function. Then
Theorem 4 (Occam's razor).
For any data generating distribution on X and any prior distribution P (h) over H, with a
probability greater than 1− δ over drawing an i.i.d. sample from X of size T , for all h ∈ H:
(cid:12)(cid:12)(cid:12)L(h) −(cid:98)L(h)
(cid:12)(cid:12)(cid:12) ≤ b ·
(cid:115)− log P (h) + log 2
δ
.
2T
Proof. See [23].
Let H =(cid:8)π : S → A(cid:9) denote the set of deterministic policies, where A = {a0, a1} = {0, 1}.
Define loss function
L :
(cid:16)S × N(cid:17) × H −→ R : (s, ν) × π (cid:55)→ R(s, π(s), ν).
Further, set probability distribution P (s, ν) = P (s) · P (νs) on S × N . Theorem 4 holds
for any sampling distribution. In particular we may use the policy π to restrict samples to
situations that cause the neuron to spike to obtain P (s, νπ(s) = a1) = π(sa1) · P (νs). It
follows that
L(π) = Rπ = E(cid:104)
(cid:98)L(π) = (cid:98)Rπ =
1
T1
R(s, a1, ν)
(cid:12)(cid:12)(cid:12) π(sa1) · P (νs)
(cid:105)
(cid:88)
R(cid:0)st, a1, νt
{(st,νt)π(st)=a1}
is the expected reward and
(cid:1) is the empirical reward,
where T1 is the number of spike produced by the neuron during [1, T ].
Let H ⊃ Hk = (cid:8)π : S → A s.t. π−1(a1) = k(cid:9) denote policies that spike for exactly k
Theorem 3.
situations. Given the setup above, with probability at least 1 − δ,
(cid:115)
(cid:12)(cid:12)(cid:12)Rπ − (cid:98)Rπ
(cid:12)(cid:12)(cid:12) ≤ b ·
. By Stirling's approximation, log(cid:0)N
(cid:19)
(cid:18)N
k
Proof. Let N = S denote the number of possible situations. We put the uniform prior on
Hk, so P (π) = 1
(N
k )
− log P (π) = log
(cid:1), and it follows that
(cid:19)
(cid:1) ≤ k log(cid:0) N·e
≤ k log
(cid:18)
+ 1
log
·
k
.
N · e
k
= N · k
N
N
k
k
S · eiu(π, a1) + 1
2T1 · eeiu(π,a1) +
log 2
δ
2T1
The theorem follows since πu(a1) = k
N and eiu(π, a1) = log N
k .
12
(a) Effective information from spikes
(b) Empirical reward after spiking
Figure 1: Metabolically constrained Q-learning. As neurons learn, effective infor-
mation and empirical reward increase in qualitatively the same way. Tighter metabolic
constraints yield both higher effective information and greater rewards.
13
P(a1)=0.5P(a1)=0.3P(a1)=0.1P(a1)=0.5P(a1)=0.3P(a1)=0.1Rπ Figure 2: Concentration of reward. The x-axis lists situations, ranked by the empirical
reward the neuron would receive if it spiked. Situations are grouped into two categories:
the top 15%, which cause spikes, and the rest, which do not. The average and variance of
the empirical reward in each category is displayed.
Figure 3: Empirical versus expected reward. The normalized difference between ex-
pected and empirical reward, plotted against effective information.
14
Empirical reward after spikingEmpirical reward after silence15 percentileμ = 0.21σ(cid:31) = .054μ = 0.89σ(cid:31) = .003EMPIRICAL REWARDSILENTSPIKE |
1611.06197 | 1 | 1611 | 2016-11-18T18:53:45 | An Empirical Study of Continuous Connectivity Degree Sequence Equivalents | [
"q-bio.NC",
"cs.NE"
] | In the present work we demonstrate the use of a parcellation free connectivity model based on Poisson point processes. This model produces for each subject a continuous bivariate intensity function that represents for every possible pair of points the relative rate at which we observe tracts terminating at those points. We fit this model to explore degree sequence equivalents for spatial continuum graphs, and to investigate the local differences between estimated intensity functions for two different tractography methods. This is a companion paper to Moyer et al. (2016), where the model was originally defined. | q-bio.NC | q-bio |
An Empirical Study of Continuous Connectivity
Degree Sequence Equivalents
Daniel Moyer, Boris A. Gutman, Joshua Faskowitz, Neda Jahanshad, and
Paul M. Thompson
Imaging Genetics Center, University of Southern California
[email protected]
Abstract. In the present work we demonstrate the use of a parcellation
free connectivity model based on Poisson point processes. This model
produces for each subject a continuous bivariate intensity function that
represents for every possible pair of points the relative rate at which we
observe tracts terminating at those points. We fit this model to explore
degree sequence equivalents for spatial continuum graphs, and to inves-
tigate the local differences between estimated intensity functions for two
different tractography methods. This is a companion paper to [11], where
the model was originally defined.
Keywords: Human Connectome, Diffusion MRI, Non-Parametric Esti-
mation
1
Introduction
In the past decade, graph theoretic analyses have rapidly propagated through
neuroimaging literature. Following advances in diffusion and functional MRI,
the rise of connectomics has popularized the use of network representations of
brain architecture and activity. Such analyses usually equate physical regions
of the cortical surface with nodes in a graph, and use structural or function
measurements as proxies for edge weights.
Increased use of the network representation of brain connectivity has been
accompanied by the use of network statistics in disease-oriented neuroscience;
nodal measures such as modularity, centrality, and degree sequences are all com-
mon descriptors for connectomes [2], and the effect of various diseases on these
descriptors has been the focus of recent study [4]. The popularity of these mea-
sures stems in part from their theoretical underpinnings. In particular, the distri-
bution of nodal degrees has deep implications for the organization and topology
of a random network.
For brain networks the choice of a particular delineation of physical regions
(parcellation) and thus the choice of nodes is non-trivial. Multiple studies have
shown that the choice of parcellation influences the summary measures including
nodal degree for both structural and functional networks [13,17,18]. It remains
unclear which of the many parcellations is optimal, or whether or not a single
overall "best" parcellation exists [12], assuming a criterion for quality can be
agreed upon. Furthermore, the study of local or multi-scale grained phenomena
often require specific resolutions and thus specific parcellations. This presents a
dilemma to the community: do we choose the best parcellation for their specific
study, or do we choose a parcellation that generalizes well to other literature?
It is therefore valuable to explore alternative representations that avoid these
issues. In particular, it is useful to construct representations of cortical connec-
tivity that are independent of the choice of parcellation, yet still theoretically
and computationally tractable both for estimation as well as statistical analy-
sis, as well as retaining the ability to construct parcellation based connectivity.
Furthermore, the exploration of continuous equivalents to currently popular net-
work statistics and their similarities and/or deviations from current empirical
observations is of interest.
In the current work we explore the empirical properties of such a model,
described in a companion paper [11]. We explore a continuous representation
of cortical connectivity that describes the observation of white matter tract1
endpoints using a random process defined on the gray matter/white matter
interface (the inner cortical surface). This form of connectivity generalizes tra-
ditional connectomes to a parcellation-independent representation from which,
given any particular parcellation, the discrete connectome may be recovered.
We reproduce degree distribution results similar to those described by discrete
representation analyses using our continuous model, and note the instances of
discrepancy between the two. We further investigate differences between two
tractography methods in the degree distribution (marginal intensity function).
2 Continuous Connectivity Model
In order to ensure that the reader may understand the results in Section 3,
we first review the theory and motivation behind this particular continuous
connectivity model; more details can be found in [11]. In particular we introduce
the framework itself and its general terminology, focusing on the key piece of
our model, the Poisson point process and its intensity function. We then define
the analogous statistic to degree for the continuous framework, which will serve
as the empirical focus of the next section.
2.1 Model Description and Theoretical Discussion
A point process is a random process in which collections of discrete points are
generated randomly on a measurable space. The Poisson process is the most
basic of these, assuming that these points are generated independently (i.e. the
appearance of one point does not affect the probability of observing another) and
with some relative rate proportional to an intensity function λ : Domain → R+.
For any subset of the domain the distribution of the number of points in that
1 It is important to distinguish between white matter fibers (fascicles) and observed
"tracts." Here, "tracts" denotes the 3d-curves recovered from Diffusion Weighted
Imaging via tractography algorithms.
subset is Poisson with parameter equal to λ integrated over the subset. This
means the expected number of points is exactly that integral. We use this process
to model tract endpoint locations on the cortical surface. Our domain is the
connectivity space of the cortex (the product of the gray/white matter interface
with itself) which is the set of all pairs of possible endpoints, and each point in
this space corresponds with a pair of tract endpoints.
In our context of connectomics, this framework produces two different struc-
tures that are both analogous to connectivity. The first is the usual region-to-
region connectivity, which is produced by measuring the expected tract count on
a subset of the domain which is itself composed of all pairs of points in two sub-
sets of the cortex. These cortical subsets are not necessarily disjoint. The second
representation of connectivity is the intensity function λ. While the first is an
aggregation of the second, it is important to separate them. The intensity func-
tion has pointwise intensities which are not comparable with the second. The
second representation generalizes the first over the choice of regions. Varying
segmentation choices for the first representation can be compared in the context
of the second.
A more formal definition of the framework is as follows: Let Ω be union
of two disjoint subspaces each diffeomorphic to the 2-sphere representing the
cortical surface for a particular hemisphere, and assume that tracts randomly
and independently intersect with this surface at exactly two points. Further
consider the space Ω × Ω, which is the space of pairs of points on the cortical
surface, in our case the space of all possible pairs of endpoints of tracts. We
model connectivity as a function λ : Ω × Ω → R+ such that for any regions
E1, E2 ⊂ Ω the number of observed tracts having one end point in E1 and the
other in E2 is Poisson distributed with parameter
(cid:90)(cid:90)
C(E1, E2) =
λ(x, y)dxdy.
E1,E2
(1)
This is exactly a Poisson point process of tract endpoints over Ω × Ω with λ
is its intensity function. Note that for E1, E2 with non-trivial intersections this
double counts the number of actual tracts observed in that intersection. If a tract
has endpoint (x, y) ∈ E1 ∩ E2, then clearly there exists a tract (y, x) ∈ E1 ∩ E2.
We define λ(x, y) for any particular (x, y) as the pointwise connectivity be-
tween x and y. Though it is not technically required, we assume this pro-
cess (characterized by λ) to be smooth everywhere and symmetric (λ(x, y) =
λ(y, x)).. We further assume tracts are generated independently; more complex
point process models relax this assumption, and a more general form of this
model could be considered, in which λ is vector valued, and the Poisson dis-
tribution may be substituted for any distribution parameterizable by integral
terms. These models are unfortunately computationally intractable with current
estimation methods, but are relevant to the proceeding discussion.
We define a regional connectivity as the expected number of tracts between
any two regions. In the Poisson case this is exactly C(E1, E2). For any parti-
i Ei = Ω, this forms a connectivity matrix similar to
tion (parcellation) P =(cid:83)
the traditional connectomes (which provide connectivity for the P × P discrete
space). From empirical data, a valid estimator is simply counting the number of
tracts present between each pair of regions, as current methods dictate. Thus,
continuous connectivity models of the form proposed here generalizes graphs
generated by finite partitions of Ω (including overlapping paritions) for certain
classes of edge weight distributions.
An important summary statistic of traditional graphs is the degree sequence,
defined as the collection of sums of each edge for each node (this counts each
edge weight exactly twice, though for different terms in the sequence). A sim-
ilar construct exists for the continuous connectome which we call the marginal
connectivity, given by
(cid:90)
M (x) =
λ(x, y)dy.
Ω
M (x) itself is defined over Ω. Using the assumption that λ is continuous, it
can be shown fairly easily that M (x) is also continuous. Though not explored in
this paper, another natural extension are marginal connectivities from a particu-
E λ(x, y)dy,
lar region. Choosing any region E, we can define them as ME(x) =(cid:82)
which is the total connectivity from point x to the region E.
2.2 Estimation and Asymptotic Estimator Distribution
A sufficient statistic for Poisson process models is the intensity function λ(x, y).
Estimation of the function is non-trivial, and has been the subject of much
study in the spatial statistics community [5]. We choose to use a non-parametric
Kernel Density Estimation (KDE) approach due to an efficient closed (truncated
harmonic) form for estimation. We first inflate each surface to a sphere and
register them using a spherical registration (See section 3.1), assuming each
hemisphere is disjoint with the other. Our domain is then Ω × Ω the product of
spheres (S1 ∪ S2) × (S1 ∪ S2).
The unit normalized spherical heat kernel is a natural choice of kernel for S2.
We use its truncated spherical harmonic representation [3], defined as follows for
any two unit vectors p and q on the 2-sphere:
H(cid:88)
h
Kσ(p, q) =
2h + 1
4π
exp{−h(h + 1)σ}P 0
h (p · q)
h is the hth degree associated Legendre polynomial of order 0. Note that
Here, P 0
the non-zero order polynomials have coefficient zero due to the radial symmetry
of the spherical heat kernel [3]. However, since we are estimating a function on
Ω × Ω, we use the product of two heat kernels as our KDE kernel κ. For any
two points p and q, the kernel value associated to a end point pair (x, y) is
κ((p, q)(x, y)) = Kσ(x, p)Kσ(y, q).
tion. This is exactly λ(p, q) = κ((p, q)D) =(cid:80)
We then apply this kernel to our data in order to recover the intensity func-
(xi,yi)∈D Kσ(xi, p)Kσ(yi, q) where
λ denotes our estimation of λ, the true intensity. Since storing the exact func-
tional form is expensive, in practice we evaluate this at each vertex of the cortical
surface triangulated mesh.
It should be noted that while the end result of this process is an array of
values corresponding to a triangulated mesh of the cortical surface (which can
be inflated into a sphere), this is not equivalent to a high resolution discrete con-
nectome except in the roughest terms. Each point here represents a pointwise
evaluation of the density function, and the difference between adjacent mesh
vertices is bounded (since the domain is compact, we can show that the true
intensity function must be absolutely continuous if it is continuous at all). The
discrete connectome, no matter how small the region, represents an areal con-
nectivity value, and in most cases caries no geometric information, much less a
smoothness condition.
The product kernel complicates the asymptotic analysis of this estimator.
Unfortunately there is no true closed form for the heat kernel on the sphere [10],
and approximations include terms that are not consistent with traditional kernel
bandwidth analysis (which usually is univariate, and unbounded). On the other
hand, it can be shown that the heat kernel using these approximate forms will
be consistent with extant first and second moments [8].
3 Empirical Results
3.1 Pre-processing
Our data are comprised of 731 subjects from the Human Connectome Project
S900 release [16]. We used the minimally preprocessed T1-weighted (T1w) and
diffusion weighted (DWI) images rigidly aligned to MNI 152 space. Briefly, the
preprocessing of these images included gradient nonlinearity correction (T1w,
DWI), motion correction (DWI), eddy current correction (DWI), and linear
alignment (T1w, DWI). We used the HCP Pipeline (version 3.13.1) FreeSurfer
protocol to run an optimized version of the recon-all pipeline that computes
surface meshes in a higher resolution (0.7mm isotropic) space.
Tractography was conducted using the DWI in 1.25mm isotropic MNI 152
space. Probabilistic streamline tractography was performed using Dipy's Local-
Tracking module [7]. To model the fiber distribution at each voxel, Dipy's im-
plementation of constrained spherical deconvolution (CSD) [15] was used with a
spherical harmonics order of 8. Tractography streamlines were seeded at 2 ran-
dom locations in each voxel labeled as likely white matter via the segmentation
maps generated by FSL's FAST. Streamline tracking followed directions ran-
domly in proportion to the orientation function at each sample point at 0.5mm
steps, starting bidirectionaly from each seed point. Streamlines were only re-
tained if longer than 5mm and both ends terminated in voxels likely correspond-
ing to grey matter according to Dipy's implementation of ACT[14]. An additional
tractography was computed in the same manner, but replacing the CSD fit for
a diffusion tensor (DTI) fit.
Fig. 1. Here we see the estimated density of sampled marginal connectivity functions.
The top figure is the direct plot, and the bottom figure is the log-log plot of the same.
We then fit 5 separate scales of continuous connectivity models, varying σ
as to capture differing scales of spatial patterns. We use each σ = {0.01,0.005,
0.001, 0.0005, 0.0001}, for both DTI and CSD . We then integrate numerically
each resulting intensity function for each subject, forming two marginal connec-
tivity functions for each subject. A paired t-test was then performed at each of
the 20484 vertex across subjects (the surface is subsampled for intensity function
computation). Multiple comparison correction was applied before inference us-
ing the Bonferroni correction for 20484 hypotheses. The continuous assumption
makes this correction overly conservative; the test points are no longer inde-
Fig. 2. In this plot we see p-values for the CSD-DTI test. Blue denotes CSD > DTI
by as significant margin, while Red denotes DTI > CSD by a significant margin. Gray
areas denote regions that were unable to reject the null hypothesis of "no difference".
From top to bottom we have σ = {0.01,0.005, 0.001, 0.0005, 0.0001}, and from left
to right we have the left hemisphere and right hemisphere.
pendent due to the smoothness of the signal. Test statistics were evaluated at
α = 0.05. We also estimate continuous one dimensional densities for samples
from the average marginal connectivity function's distribution of values.
3.2 Degree Sequences
Much sensation in the past 20 years has focused on complex networks and the
distribution of their degree sequences [1]. We here display estimates for a similar
measure for sampled marginal connectivity functions. Multiple studies have pro-
vided evidence of small-worldness or scale-free properties in connectomes, both
functional and structural [6,2], sometimes making conflicting claims. While we
here make no rigorous claim about the fit of one model or another, we do note
the dissimilarity to strictly power law distributions, and their similarity to ge-
ometrically constrained communication networks, particularly ad hoc computer
networks [9]. (This is qualitative comparison only, and not backed by quanti-
tative observation of ad hoc networks). The plots in Fig 1 show clear second
modes, and their log plots are slightly non-linear.
3.3 DTI-CSD Difference
As shown in Figure 2, there are clear differences recovered by the continuous
connectivity model between the CSD and DTI tractography algorithms. This not
only illustrates use of a continuous model, but is important for understanding the
different trade offs between tractography algorithms. In general DTI is much less
flexible than probabilistic CSD, but also fit orders of magnitude faster. Since each
connectome is normalized, significant areas are proportionally more explored by
a particular algorithm. The authors would also like to reiterate the value and
importance of multiple test correction in large hypothesis set situations such as
this, due to the large number of sample points tested.
DTI appears to concentrate on motor and somatosensory cortices and their
corresponding tracts, while the CSD model appears to find more tracts in the
temporal lobe. This is not to say that CSD is missing the motor tracts or that
DTI does not have any temporal lobe tracts, but that the relative concentrations
are shifted. It is also interesting to note that several small regions break this
trend, particularly for DTI concentrations in the inferior temporal and temporal-
frotal regions for σ ≤ 0.001. While this work is preliminary, this demonstrates
some of the advantages of continuous connectivity models; while region specific
models may not be able to resolve such small differences, particularly those
surrounded by significant regions of the other density, the continuum model
allows very local differences to be detected, as well as large, non-convex regions
(such as those in the parietal/frontal regions for DTI, with σ = 0.01).
This also illustrates one unfortunate issue with the continuum model: Visual-
ization of the full connectivity is difficult (a four dimensional continuum usually
embedded in six dimensions and that is not a 4-sphere). In future work we plan
on addressing this issue and developing methods to visualize portions at a time.
However, we also believe this speaks to the possible issues when visualizing the
discrete networks.
4 Conclusion
In the present work we have described a parcellation free connectivity model.
We further used this model to explore degree sequence equivalents for spatial
continuum graphs, and to investigate pointwise differences in these functions for
two different tractography methods. We believe that parcellation based networks
are critical to the exploration of the cortical landscape, but that the development
of more general methods such as the one presented here is also vital to expanding
the set of testable questions in neuroscience, and improving the answers provided
by neuroimaging.
References
1. Barab´asi, A.L.: Scale-free networks: a decade and beyond. science 325(5939), 412 --
413 (2009)
2. Bullmore, E., Sporns, O.: Complex brain networks: graph theoretical analysis of
structural and functional systems. Nature Reviews Neuroscience 10(3), 186 -- 198
(2009)
3. Chung, M.K.: Heat kernel smoothing on unit sphere. In: Biomedical Imaging: Nano
to Macro, 2006. 3rd IEEE International Symposium on. pp. 992 -- 995. IEEE (2006)
4. Crossley, N.A., Mechelli, A., Scott, J., Carletti, F., Fox, P.T., McGuire, P., Bull-
more, E.T.: The hubs of the human connectome are generally implicated in the
anatomy of brain disorders. Brain 137(8), 2382 -- 2395 (2014)
5. Diggle, P.: A kernel method for smoothing point process data. Applied Statistics
pp. 138 -- 147 (1985)
6. Eguiluz, V.M., Chialvo, D.R., Cecchi, G.A., Baliki, M., Apkarian, A.V.: Scale-free
brain functional networks. Physical review letters 94(1), 018102 (2005)
7. Garyfallidis, E., et al.: Dipy, a library for the analysis of diffusion MRI data. Front.
Neuroinform 8(8) (2014)
8. Hall, P., Watson, G., Cabrera, J.: Kernel density estimation with spherical data.
Biometrika 74(4), 751 -- 762 (1987)
9. Hekmat, R., Van Mieghem, P.: Connectivity in wireless ad-hoc networks with a
log-normal radio model. Mobile networks and applications 11(3), 351 -- 360 (2006)
10. Lafferty, J., Lebanon, G.: Diffusion kernels on statistical manifolds. Journal of
Machine Learning Research 6(Jan), 129 -- 163 (2005)
11. Moyer, D., Gutman, B., Faskowitz, J., Thompson, P.M.: A continuous model of
cortical connectivity. In: MICCAI 2016. Springer (2016)
12. de Reus, M.A., Van den Heuvel, M.P.: The parcellation-based connectome: limita-
tions and extensions. NeuroImage 80, 397 -- 404 (2013)
13. Satterthwaite, T.D., Davatzikos, C.: Towards an individualized delineation of func-
tional neuroanatomy. Neuron 87(3), 471 -- 473 (2015)
14. Smith, R.E., et al.: Anatomically-constrained tractography: improved diffusion
MRI streamlines tractography through effective use of anatomical information.
NeuroImage 62(3), 1924 -- 1938 (2012)
15. Tournier, J.D., Yeh, C.H., Calamante, F., Cho, K.H., Connelly, A., Lin, C.P.: Re-
solving crossing fibres using constrained spherical deconvolution: validation using
diffusion-weighted imaging phantom data. NeuroImage 42(2), 617 -- 625 (2008)
16. Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil,
K., Consortium, W.M.H., et al.: The wu-minn human connectome project: an
overview. Neuroimage 80, 62 -- 79 (2013)
17. Wang, J., et al.: Parcellation-dependent small-world brain functional networks: A
resting-state fMRI study. Human Brain Mapping 30(5), 1511 -- 1523 (2009)
18. Zalesky, A., et al.: Whole-brain anatomical networks: does the choice of nodes
matter? NeuroImage 50(3), 970 -- 983 (2010)
|
0810.0029 | 3 | 0810 | 2010-03-19T23:12:26 | Topological Effects of Synaptic Time Dependent Plasticity | [
"q-bio.NC",
"cond-mat.dis-nn",
"math-ph",
"math-ph",
"q-bio.TO"
] | We show that the local Spike Timing-Dependent Plasticity (STDP) rule has the effect of regulating the trans-synaptic weights of loops of any length within a simulated network of neurons. We show that depending on STDP's polarity, functional loops are formed or eliminated in networks driven to normal spiking conditions by random, partially correlated inputs, where functional loops comprise weights that exceed a non-zero threshold. We further prove that STDP is a form of loop-regulating plasticity for the case of a linear network comprising random weights drawn from certain distributions. Thus a notable local synaptic learning rule makes a specific prediction about synapses in the brain in which standard STDP is present: that under normal spiking conditions, they should participate in predominantly feed-forward connections at all scales. Our model implies that any deviations from this prediction would require a substantial modification to the hypothesized role for standard STDP. Given its widespread occurrence in the brain, we predict that STDP could also regulate long range synaptic loops among individual neurons across all brain scales, up to, and including, the scale of global brain network topology. | q-bio.NC | q-bio | A Theory of Loop Formation and Elimination By Spike
Timing-Dependent Plasticity
James Kozloski and Guillermo A. Cecchi
Computational Biology Center, IBM Research Division,
IBM T.J. Watson Research Center, Yorktown Heights, NY 10598
(Dated: October 25, 2018)
Abstract
(FULL PAPER IN FRONTIERS IN NEURAL CIRCUITS, 2010) We show that the local Spike
Timing-Dependent Plasticity (STDP) rule has the effect of regulating the trans-synaptic weights
of loops of any length within a simulated network of neurons. We show that depending on STDP's
polarity, functional loops are formed or eliminated in networks driven to normal spiking conditions
by random, partially correlated inputs, where functional loops comprise weights that exceed a non-
zero threshold. We further prove that STDP is a form of loop-regulating plasticity for the case of a
linear network comprising random weights drawn from certain distributions. Thus a notable local
synaptic learning rule makes a specific prediction about synapses in the brain in which standard
STDP is present: that under normal spiking conditions, they should participate in predominantly
feed-forward connections at all scales. Our model implies that any deviations from this prediction
would require a substantial modification to the hypothesized role for standard STDP. Given its
widespread occurrence in the brain, we predict that STDP could also regulate long range synaptic
loops among individual neurons across all brain scales, up to, and including, the scale of global
brain network topology.
0
1
0
2
r
a
M
9
1
]
.
C
N
o
i
b
-
q
[
3
v
9
2
0
0
.
0
1
8
0
:
v
i
X
r
a
1
I.
INTRODUCTION
Connections between individual neurons in the brain are constrained first by the spa-
tial distribution of axons and dendrites within the neuropil [1][2]. Global brain networks
comprise dense connections within tissues, the gross structures in which these tissues are
embedded, and the bidirectional long-range projections joining these structures. The topol-
ogy of these networks is not yet fully specified at the level of microcircuitry, however [24].
One theoretical constraint on this level of organization, the "no strong loops hypothesis,"
considered only developmentally determined area to area connectivity patterns to imple-
ment its specific neuron to neuron network topological constraint [3]. While local synaptic
modifications are known to directly shape the pattern of connectivity in local neural tissue
and thus local microcircuit topology [4], our understanding of global brain network topology
still derives largely from this developmentally patterned, area to area connectivity. Further-
more, measuring simultaneously the relative strengths of specific microcircuit connections
remains technically challenging, and virtually impossible for even medium sized (100 − 200
neurons, 0.05 − 0.1 mm) microcircuits. For these reasons, it is not yet known how large
scale, long range microcircuit topology and the computation it supports emerges through
synaptic modifications in the brain.
We wondered whether a synaptic modification commonly observed in local circuit prepa-
rations and widely hypothesized to shape local dynamics in brain structures, STDP [5],
could be analyzed to yield an understanding of what topology it predicts for microcircuits
of any scale. The STDP model is a departure from traditional Hebbian models of learning,
which state that neurons that fire action potentials together will have their interconnections
strengthened. Instead, STDP takes into account the particular temporal order of pre- and
post-synaptic neuronal firing [6], such that the rule modifies synapses anti-symmetrically,
depending on whether the pre- or post-synaptic neuron fires first (Fig. 1). The basic ques-
tion we then aimed to answer is: what is the influence of this anti-symmetry on brain
microcircuit topology?
Consider first if a pre-synaptic "trigger" neuron causes a post-synaptic, first-order "fol-
lower" neuron to fire. If this follower makes a direct feedback connection onto the trigger, the
feedback connection will be weakened, since the spike generated by the follower will arrive
at the follower-trigger synapse immediately after the trigger neuron's backward propagating
2
action potential (Fig. 1). The principle that STDP is suitable for eliminating strong recur-
rent connections between two neurons was originally proposed by Abbott and Nelson [7].
Here we expand on the principle with the observation that it holds for all polysynaptic loops
connecting triggers and followers: if some nth-order follower's action potential produces in
the original trigger a subthreshold potential after the trigger has fired, the functional loop
will be broken by spike-timing dependent synaptic weakening of the feedback connection.
With this intuition, we set out to prove analytically and by means of numerical simulation
that network topology, and specifically the occurrence of functional loops in highly connected
networks, is directly and necessarily regulated by STDP.
This theory paper provides clear predictions about STDP's effect on neural circuit topol-
ogy. The proof and simulations dictate strong constraints on local and long range micro-
circuit connectivity. We propose that if these constraints are not obeyed by real neural
circuits, the hypothesis that standard STDP shapes the structure and function of real ner-
vous systems must be revised. Our approach suggests that similar analyses of other learning
rules may impose similar constraints on neural circuit topology and that the hypothetical
significance of these rules may similarly be tested.
II. RESULTS
A. A proof of STDP as a form of loop-regulating plasticity
First, we represent STDP acting on a weight w associated with the connection between
two neurons and their output variables x(t) to y(t), in the adiabatic approximation (i.e.
small learning rate), as:
(cid:90) ∞
−∞
where Cxy(t) =(cid:82) x(t(cid:48)−t)y(t(cid:48)) is the correlator, and S(t) is the anti-symmetric STDP update
∆wxy ∝
Cxy(t)S(t)dt
function, S(t < 0) = exp(λt), S(t > 0) = − exp(−λt). Consider this function operating over
connections within a linear network driven by uncorrelated Gaussian inputs, ξ, such that
(1)
x(t) = W x(t) + ξ(t), where x is a vector of activities with components denoted by xi, the
weight connection matrix has components Wij, and the input satisfies (cid:104)ξi(t)ξT
j (t + τ )(cid:105)t =
σ2δ(τ )δij. We show (see Appendix A 1) that the learning rule defined in Eq. 1 results in
an update for the network weight matrix of the form ∆W = ∆W (W, τ, C0) where τ is the
3
time constant of the STDP's exponential, and C0 is the instantaneous correlator C(0). This
update rule influences global network topology in a very specific way.
To formalize our original intuition analytically, consider a linear network with only ex-
citatory connections, such that the dynamics may be expressed as x = W x = (−I + A)x,
where Aij ≥ 0 is the network connectivity matrix (comprising the off-diagonal elements of
the weight matrix and zeros on the diagonal), and −I represents a self-decay term. Next,
we introduce a "loopiness" measure that estimates the strength of all loops of all sizes that
k tr (cid:2)Ak(cid:3). The function tr [·] stands for the trace operation;
occur in the network, El =(cid:80)∞
1
k=1
this operation, when acting on the k-exponentiation of the adjacency matrix (comprising
ones and zeros, where the nonzero entry aij represents a connection from network node i
to network node j), counts the total number of closed paths of length exactly equal to k,
i.e. k-loops [25]. When applied to the network connectivity matrix A, the operation counts
loops weighted by the product of the synaptic strengths of the looping connections, resulting
in a slightly different, but still useful, measure of loopiness.
It is possible, however, to reduce this measure without actually regulating topology by
simply reducing the weights of all connections. A topological
therefore include a penalty to the weights' vanishing; we choose − 1
the transpose operation) which, for weighted graphs, measures the sum of the squares of all
loopiness measure should
2tr (cid:2)AAT(cid:3), (T stands for
network weights (and for a binary graph, counts the number of links). We then define the
total topological loopiness as:
E =
∞(cid:88)
k=1
1
k
tr (cid:2)(A)k(cid:3) − 1
2
tr (cid:2)AAT(cid:3)
(2)
We showed computationally that for every weight matrix examined, each drawn from certain
random distributions (see Appendix A 1), the change in this energy as a function of the
evolution of the network under STDP, ∆E ∼ tr (cid:2)∂AE∆AT(cid:3), is strictly semi-negative, and
therefore STDP necessarily regulates this measure, resulting in a decrease in topological
loopiness. We therefore use the term STDP and "loop-regulating plasticity" interchangeably
throughout.
4
B. Loop-regulating plasticity in a network of simulated neurons
What are the effects of this form of plasticity on network topology (and specifically
on the number of functional loops) in nonlinear networks, such as those found in neural
microcircuits? Because our proof of STDP as a form of loop-regulating plasticity applies
only to linear networks or nonlinear networks that may be linearized, we aimed to show,
using simulation, that the same principle extends to a biologically relevant, nonlinear regime.
We replicated the simulation of Song and Abbott [8], extending it in three ways (see IV A).
First, we created a network of 100 neurons, each receiving excitatory synapses from all
other 99 "intra-network" input sources and from 401 randomly spiking "extra-network"
input sources selected at random from 2,500 homogeneous Poisson processes. All excitatory
synapses underwent STDP. Second, we provided 250 inhibitory synapses to each neuron,
sampled from 1,250 spiking sources; the inhibitory inputs modeled fast local inhibition to the
network using inhomogeneous Poisson processes with rates modulated by the instantaneous
aggregate firing rate of the network. Third, we explored four different forms of STDP update
[9] and observed robust loop-regulating plasticity for each; the results presented here used
the STDP update rule of Gutig et al. [10].
We initialized our network with maximum extra-network weights, and intra-network
weights at half maximum. This caused the network to spike vigorously when extra-network
inputs became active, but spike rates were limited by the fast local-inhibition. After 20 sec-
onds of simulated network activity, STDP had a profound effect on topological loopiness as
defined in Eq. 2, measured over loops of length 2 ≤ k ≤ 100 for convenience (Fig. 2A). We
counted the number of closed, functional loops of varying length using tr (cid:2)(cid:100)A(cid:101)k(cid:3), where (cid:100)A(cid:101)
was constructed by applying a sliding threshold to the network connectivity matrix (Fig.
2B). We compared this quantity to the same, measured for a randomized network, con-
structed by randomly reassigning weights from the learned weight distribution to synapses
in the network (see IV B). These results are representative of all loop lengths measured
(2 ≤ n ≤ 100) and show that as the weight threshold grows, the number of closed, func-
tional loops in the STDP-learned network decreases more than in the randomized network.
This form of loop-regulating plasticity can therefore be described as loop-eliminating.
5
C. The effect of synaptic delays on loop-regulating plasticity
We wondered what effect synaptic delays would have on this result, since we expected
follower feedback spikes to cause less anti-loop learning as they fell further from the zero
time difference maxima in the STDP update function. We also wondered if the decrease in
the number of functional loops compared to a randomized network also applied to unique
loops, in which no neuron is traversed more than once. We therefore sampled the number
of unique, functional loops through networks simulated with synaptic delays from 0.1 to 4.0
milliseconds. We constructed one million random paths of length k − 1 for each loop length
2 ≤ k ≤ 25, and for the learned and randomized networks (see IV B). We searched for each
path across all networks studied, and if the path and the kth link completing the functional
loop existed in the network, we counted it for that network (see IV B). The result is similar
to that for closed loops, and, as expected, longer synaptic delays resulted in an exponential
decrease in the number of loops as a function of loop length that deviated less from the same
function for randomized networks, indicating weaker loop-regulating plasticity (Fig. 2C).
D. Network in-hubs, out-hubs, and loop-regulating plasticity
Next, we asked if other topological measures of the STDP-learned networks may be cor-
related with our observation of STDP's effect on loopiness, since many different topological
properties might coincide with or support this effect. For example, one means to create
networks poor in loops is to ensure that nodes in the network are either "out-hubs" or "in-
hubs," but not both [11]. An out-hub in a network of neurons has many strong postsynaptic
connections but few strong presynaptic connections, and an in-hub has many strong presy-
naptic connections but few strong postsynaptic connections. We applied a sliding threshold
to the network connectivity matrix learned by STDP, and examined the manifold, colored
according to each applied threshold, which correlated in-degree versus out-degree for each
neuron in our network. This showed a clear inverse relationship between in- and out-degrees
that varied in form with weight threshold (Fig. 3A). In contrast, by examining the in-
degree from extra-network inputs, we found a positive correlation (Fig. 3B), indicating that
out-hubs were more likely to be in-hubs within the larger extra-network topology, and that
in-hubs in our network were more likely to receive only the weakest extra-network inputs.
6
E. Reverse STDP restored loops after loop-eliminating plasticity
Beyond these standard topological analyses, we also examined biological properties of the
network. We measured total synaptic input as a function of total synaptic output for all
neurons in the STDP-learned network. In the same experiment, we asked if reversing the
polarity of the standard STDP function might undo the effects of loop-regulating plastic-
ity that results from standard STDP, since under this "reverse" condition, follower spikes
would cause strengthening of closed-loop feedback connections. This reversal of polarity is
biologically relevant, since it occurs at the synaptic interface between major brain struc-
tures such as neocortex and striatum [12], arises specifically at synapses between certain
cell types, and is controlled by cholinergic and adrenergic neuromodulation, for example in
the neocortical microcircuit [13]. We found the same inverse relationship between in-degree
versus out-degree for each neuron in our network (Fig. 4A, green markers), as well as an in-
verse relationship between total synaptic input and output following 1.5 seconds of standard
STDP (Fig. 4B, green markers). These effects contributed to a reduction in the number
of closed loops (Fig. 4C, depicted as in Fig. 2B), and each of these relationships could
be largely abolished by 3 to 5 additional seconds of reverse STDP (Fig. 4, red markers),
in contrast to 3 to 5 additional seconds of standard STDP (Fig. 4, blue markers), which
strengthened them. We also found the same positive correlation between in-degree from
extra-network inputs and out-degree within the network (Fig. 4D) and between total extra-
network synaptic input and total intra-network synaptic output (Fig. 4E). This effect was
also largely abolished by 3 to 5 seconds of reverse loop-regulating plasticity, but reinforced
by 3 to 5 seconds of standard loop-regulating plasticity.
F. Dynamical effects of loop-regulating plasticity
What are the consequences of this form of network plasticity beyond topology? In the case
of a linear network, reducing the number of loops implies more stable dynamics. Consider
the stability of the unforced system (cid:126)x(t) − W (cid:126)x(t) = 0; the eigenvalues λ of W = −I + A
can be expressed as:
(3)
∞(cid:88)
tr (cid:2)Ak(cid:3)
1
k
− N(cid:88)
log λi =
i=1
k=1
7
[14], which emphasizes the contribution of loops to system instability. Such a simple ob-
servation, however, does not make clear predictions about the effects of loop-regulating
plasticity on nonlinear neural circuit function. We were surprised to find that raster plots
of network spiking activity, when sorted according to certain topological metrics (e.g., the
sum of extra-network input weights, the sum of intra-network output weights, in-degrees,
out-degree) consistently revealed network events that originate with weak synchronization
among out-hubs, followed by strong synchronization among in-hubs (Fig. 5A, top), across
8 independent simulations of the phenomenon. This effect was altered by randomizing the
intra-network weights, such that synchronization events became stronger, more frequently
global, and more frequent among out-hubs alone (Fig. 5A, bottom). Peri-event time his-
tograms constructed across 8 independent simulations reveal this same effect (Fig. 5B, left
panels), with synchronization arising strongly among in-hubs after weak out-hub activation
in the STDP-learned network, and globally in the randomized network (see IV B for a de-
scription of how network events were detected). In the STDP-learned network, both in-hubs
and out-hubs sustain spike rates ranging from 4−9 Hz that are not correlated with in-degree,
whereas in the randomized network, spike rates range more broadly (3 − 16 Hz) and are
highly correlated with in-degree (Fig. 5B, right panels). We examined the summed network
peri-event time histograms for the STDP-learned network and for networks that underwent
randomization of their intra-network weights, their extra-network weights, or both (Fig.
5C, top), across 8 simulations for each condition. The resulting distributions, as well as
a pooled distribution of times from all randomized networks each differed from each other
based on paired Kolmogorov-Smirnov tests (P ≈ 0). To examine what properties of these
distributions distinguished them, we measured kurtosis and skew for each distribution from
each simulation, and compared the distributions of kurtosis and skew measures between
each group. Kurtosis differed significantly between the STDP-learned network topology
and the randomized topologies (intra-network randomized, extra-network randomized, both
randomized, Fig. 5C, bottom). Results from unpaired t-tests of the distribution of skew
and kurtosis measurements between 4 simulation conditions (calculated separately for each
of the 8 simulations) are shown in table inset (Fig. 5C, bottom). These effects indicate
that the effect of standard loop-regulating plasticity is to generate network topologies that
support network events with greater spread and sharper peaks in time.
8
III. DISCUSSION
Based on our simulations and analytical results, we propose that standard STDP must
produce a network topology in real neural tissues that is conspicuously poor in both closed
and unique loops, and that it will segregate neurons into out- and in-hubs to achieve this.
Such a prediction can easily be tested by analyzing correlations between the number of
functional input connections and the number of functional output connections made by
neurons recorded during multi-patch clamp experiments in a structure in which STDP has
been observed (e.g., [4], [15]). Our theory predicts this correlation should be negative.
The network that emerges in such tissues will organize its relationship to inputs from
other structures in an orderly fashion, making local out-hubs the primary target for long
range inputs, and thus establishing a feed forward relationship between the network and its
pool of inputs. In a larger system, we anticipate that local in-hubs would become long range
outputs. This prediction may also be tested by correlating the local topological relationships
of a neuron with its identified role as either an input, output, or interneuron within that
structure.
We also make a clear prediction for the effect of synaptic delays on modulating the
topological effects of STDP. Correlations between these delays and functional connectivity
data from multi-patch clamp recordings from connected neurons undergoing STDP can also
be measured to determine if synaptic delays predict the strength of reciprocal connections.
Furthermore, we observe that, at interfaces between brain structures where STDP is reversed
by neuromodulation, circuit dynamics can be predicted from the expected change in network
topology. For example, changes in STDP at the cortico-striatal synapse [16] resulting in
reverse STDP would favor the emergence of strong cortico-striatal-thalamocortical loops
resulting in oscillations in this circuit.
Interestingly, the depletion of loops and the separation of nodes into out-hubs and in-hubs
has been recently reported in a variety of complex biological systems, including functional
networks at the level of spatio-temporal resolution of fMRI, and the neural network of
C. Elegans [11], suggesting a general principle of organization and dynamical stability for
entire classes of functional networks. These observations suggest quantitative measurements
of topology in vertebrate microcircuits could produce similarly interesting results.
It has been observed in local circuit preparations that a bias exists among layer 5 pyra-
9
midal neurons of rat neocortex towards strong reciprocal connectivity [4] [15], and towards
looping motifs among triplets of this neuronal class [15]. Furthermore, cyclic connections
are strongest among those neurons connected by the strongest synaptic weights. These same
neurons also exhibits STDP at the excitatory synapses that join them [5]. Given our analy-
sis, it is now clear that these observations contradict each other; specifically, we have shown
that standard STDP under normal spiking conditions with random uncorrelated inputs is
loop-eliminating. Therefore, other mechanisms and constraints than those we have analyzed
must be at play.
Consider the case of networks that have recently spiked at abnormal rates, either due
to increased excitability within the network (e.g., due to injury, epilepsy, etc.) or due to
otherwise elevated extra-network inputs. If the majority of post-synaptic potentials imme-
diately cause action potentials, standard STDP may have the effect of strengthening loops.
Also, a network driven by highly and specifically correlated inputs may spike in temporal
patterns conducive to loop-strengthening by STDP (a hypothesis we are currently studying).
Finally, as we have shown, a spiking network that has recently experienced a reversal of the
polarity of STDP will also show an increase in the number of loops observed. Clearly more
experiments and observation would be required in order to confirm or rule out each of these
mechanisms.
We observe that network activity propagates smoothly through the feed-forward topology
generated by STDP (Fig. 5A, top panel) without segregating neurons by average spike-rates
(Fig. 5B, upper right panel). The effect on global brain function of such properties would
include stable average firing rates shared among all neurons, regardless of their topological
position, and robust signal propagation, similar to "synfire chains" [17, 18]. Finally, our
theory holds that the reversal of STDP's polarity represents a local switch for the modi-
fication of both global brain network topology and global brain dynamics. Thus sources
of modulation [19] that accomplish this reversal locally are in fact regulating global brain
function by means of this switch.
10
IV. METHODS
A. Simulation
The simulation methods of Song and Abbot [8] were used to simulate each neuron in our
100 neuron network. We observed the reported topological results in each simulation after
10 seconds of network activity. In some cases, we performed additional, longer simulations
of network activity to explore the convergence and stability of these topological measures
under various conditions. Briefly, each neuron model was integrate-and-fire, with membrane
dt = Vrest − V + gexc(t)(Eexc − V ) + ginh(t)(Einh − V ),
potential determined as in [8], by τm
with τm = 20ms, Vrest = −60mV, Eexc = 0mV, Einh = −70mV, Vthresh = −54mV, and
Vreset = −60mV.
dV
The synaptic conductances gexc and ginh were modified by the arrival of a presynaptic
spike, as in [8], such that gexc(t) → gexc(t) + ¯ga, and ginh(t) → gexc(t) + ¯ginh.
dt = −ginh, with
absence of a spike, these quantities decay by τexc
τexc = τin = 5.0, ¯gin = 0.015, 0 ≤ ¯ga ≤ ¯gmax, and ¯gmax = 0.01. We initialized all elements ¯ga
to different values for intra-network (¯ga = 0.005) and extra-network (¯ga = 0.01) inputs.
dt = −gexc, and τinh
In the
dgexc
dginh
For extra-network inputs, excitatory homogeneous Poisson spike trains were generated
at a rate of 20 Hz. Inhibitory, inhomogeneous Poisson spike trains that model fast local
inhibition were generated at a rate rmin ≤ rinh ≤ rmax, where rmin = 5 Hz, and rmax = 1000
Hz. On each time step, dt = 0.1 ms, rinh is incremented by an amount proportional to the
fraction, γ, of network neurons that spiked during that timestep, and decays with a time
drinh
dt = − [rinh + (rmax − rmin)γ]. After this update, if rinh exceeds
constant, τr, such that τr
rmax, rinh → rmax.
For all simulations we report here, the STDP update rule for a synapse from neuron j
to neuron i was ¯ga(i, j) = ¯ga(i, j) + ¯ga(i, j)µM (i) for synaptic potentiation, and ¯ga(i, j) =
¯ga(i, j) + (¯gmax − ¯ga(i, j))µP a(i, j), for synaptic depression, µ = 0.1 (¯ga is maintained in the
interval [¯gmin, ¯gmax]). As in [8], M (i) and P (i, j) decay exponentially, such that τ− dM
dt =
−M (i) and τ+
dt = −Pa, τ+ = τ− = 20. Also as in [8], M (i) is decremented by A− every
time a neuron i generates an action potential, A− = 0.00035, and Pa(i, j) is incremented
dPa
by A+ every time a synapse onto neuron i from neuron j receives an action potential,
A+ = 0.00035. This update rule effectively implements the asymmetric function of STDP
11
(see Fig. 1).
B. Analysis
To randomize our networks for analysis (Figs. 2B, 2C, 4C, 5), we created a random
sequence of indices ranging uniquely from 1 to n, where n was the number of off diagonal
elements in our network's weight matrix. We used these indices to shuffle uniquely the
positions of all off diagonal elements in the matrix, thus preserving the network's learned
weight matrix, while destroying its learned topology.
We chose to sample unique loops (Fig. 2C) in the networks rather than enumerating them,
since for long loops (k > 20) the number of possible paths to search would exceed 1020. We
therefore constructed one million random paths of length k − 1 for each loop length. We
term these paths "unique" because we sampled units within each without substitution (i.e.,
no sub-loops were sampled in which a single node is traversed more than once). However, we
allowed that each path could be represented more than once in the one million constructed
paths. In fact, for the shortest paths constructed (k ≤ 3) this was necessarily the case since
the total number of possible unique paths is less than one million.
We sampled unique loops from adjacency matrices constructed such that the weight
threshold produced a matrix that was precisely half-full (for 100 neurons, a matrix with 5, 000
ones and 5, 000 zeros). For 4.0, 2.0, 1.0, 0.5, and 0.1 millisecond delays, these thresholds were
0.0032, 0.0030, 0.0033, 0.0037, and 0.0046. In this way, we controlled across experiments
for varying weight distributions, and sampled the same number of links for loops across
all experiments. Counts were compared against the randomized network constructed from
5, 000 links.
To detect network events (Fig. 5B, 5C), we employed the method described by Thivierge
and Cisek [20]. Briefly, for each simulation in which network events were detected, we
generated spike trains at 1 msec resolution for each neuron, equal in duration to the time
series analyzed and comprising the same number of spikes as observed for that neuron. We
constructed network spike time histograms of these spike trains across all neurons using a
bin width of 10 msec. We then determined a threshold for the network equal to a count
which 5% of these bins exceeded. Thresholds were determined 1000 times for each network
and each simulation, and the mean of these 1000 values used as the threshold above which
12
network events were detected in network spike time histograms for each simulation.
13
Appendix A: Supplementary Information
1. Update rule for the weight matrix
The classical definition of Hebbian learning for the weight w connecting two dynamical
variables x(t) and y(t) can be written, in its simplest form, as:
∆W = ηCxy
(cid:90) ∞
Cxy =
−∞
x(t)y(t)dt
(A1)
(A2)
Where η is the learning constant, which for exposition's sake will be set to 1. It is important,
however, to keep in mind that in order to write Eqs. A1-A2 we are assuming an adiabatic
approximation, i.e. the learning is small enough that the system can be considered to be in
steady-state for the purpose of computing the correlation.
A natural extension of Eqs. A1-A2 is to introduce time, i.e. to consider delayed as well
as instantaneous correlations:
∆W ∼
Cxy(t)S(t)dt
(cid:90) ∞
(cid:90) ∞
−∞
Cxy(t) =
−∞
(cid:48) − t)y(t
(cid:48)
x(t
(cid:48)
)dt
(A3)
(A4)
(A5)
(A6)
It
is assumed that
the time-dependent weight
function vanishes
for
long delays,
limt→±∞ S(t) = 0; the classical learning rule is recovered when S(t) = δ(t).
If, as ex-
perimental results describing STDP strongly suggest [5], the weight function displays strict
temporal anti-symmetry, i.e. S(t) = −S(−t), then
(cid:90) ∞
(cid:90) 0
∆W ∼
(cid:90) ∞
−∞
∆W ∼
Cxy(t)S(t)dt +
Cxy(t)S(t)dt
0
[Cxy(t) − Cxy(−t)] S(t)dt
A multi-dimensional linear system driven by uncorrelated input can be described as:
0
x(t) = W x(t) + ξ(t)
(A7)
where each unit is independently subject to Gaussian white noise ξ(t), a vector whose
components satisfy (cid:104)ξi(t)ξj(s)(cid:105) = σ2δijδ(t − s). The lagged correlator is related to the
zero-lagged correlator by [21]:
14
where for notational convenience we name C0 = (cid:82) xT (t)x(t)dt, i.e. the correlator at zero
C0eW T t
C(t) =
t > 0
(A8)
lag, by construction a symmetric matrix. Hence the expression for the learning update is:
∆W ∼
C0eW T t − eW tC0
S(t)dt
(A9)
eWtC0
(cid:90) ∞
(cid:104)
0
t < 0
(cid:105)
The temporal behavior of the weight function has been approximated by a piece-wise
exponential form:
S(t) =
+et/τ
0
−e−t/τ
t < 0
t = 0
t > 0
(A10)
where τ is STDP's time-constant, i.e. it expresses the window over which the plastic changes
due to temporal coincidence are significant. Assuming that the network connections are only
excitatory, and expressing without further loss of generality W = −I + A, we derive the
synaptic weight update ∆W = ∆A as follows:
(cid:90) ∞
∆A = −
(cid:90) ∞
0
−t/τ eW tdt =
e
(cid:90) ∞
0
0
S(t)eW tC0dt +
S(t)C0eW T tdt
−t/τ I+W tdt = − [W − 1/τ I]
e
−1
(cid:90) ∞
0
Given that
We obtain
∆A = − [W − 1/τ I]
−1 C0 + C0
Leading finally to
∆A ∼
(cid:20)
(cid:21)−1
I − τ
1 + τ
A
C0 − C0
(cid:2)W T − 1/τ I(cid:3)−1
(cid:21)−1
(cid:20)
I − τ
1 + τ
AT
(A11)
(A12)
(A13)
(A14)
after dropping the multiplying constant τ /(1 + τ ). From this expression it is possible to
derive that the weight update is anti-symmetric, and that a perfectly symmetric system
would not be modified, as C0 would commute with A (see below, Eq. A20). Of course,
any small initial asymmetry will eventually be blown up. We can also see that STDP's
time constant also introduces the same multiplying factor τ /(1 + τ ) for A, which can be
absorbed by a renormalization; we will assume therefore τ /(1 + τ ) → 1 for the remaining of
the exposition. Consistently, the limiting behavior of Eq. A14 implies ∆A(τ → 0) = 0.
15
2. Minimization of loops and dynamics
Now we can estimate the effect of the synaptic time-dependent plasticity expressed by
Eq. A14 on the topology of the network. For this, we will postulate a penalty or energy
function for what we will call "loopiness" of the network. A measure of the number of loops
occurring in the network can be obtained by summing the trace of the exponentiation of the
k tr (cid:2)Ak(cid:3) /k. This loop density can be simply minimized by
network connectivity matrix,(cid:80)
this effect; an obvious measure of the strength of the connections in a network is tr (cid:2)AAT(cid:3),
making the connections vanish, so we need to introduce a regularization penalty to avoid
which in a binary graph would be equivalent to the total number of links. We postulate
then the following "loopiness" energy:
tr (cid:2)Ak(cid:3) − 1
tr (cid:2)AAT(cid:3)
E =
(cid:88)
k
1
k
2
The change in this energy upon small changes ∆A is expressed as ∆E ∼ tr (cid:2)∂AE∆AT(cid:3);
it can be easily verified that ∂AE =(cid:0)I − AT(cid:1)−1 − A, and therefore:
−1(cid:3)
∆E ∼ −tr [K1] − tr [K2]
K1 = (I − AT )
−1(cid:2)(I − A)
−1C0 − C0(I − AT )
(A15)
K2 = A(cid:2)C0(I − AT )
−1 − (I − A)
−1C0
(cid:3)
(A16)
We will demonstrate in what follows that the traces of K1 and K2 are strictly semi-positive
under the synaptic changes elicited by STDP (i.e. Eq. A14), and therefore the loopiness
energy can only decrease over time. We will consider firs that the correlation matrix can be
approximated by the identity, i.e. C0 ≈ I. This case can be solved analytically; we will show
numerically that the same conclusions hold for the generic case of any C0 that results from
the dynamical system determined by A, drawn from a Gaussian or Poisson distribution.
Let us consider tr [K1], rewritten as:
tr (cid:2)(I − AT )
and which is of the from
For any matrix P ,
−2(cid:3)
−1(I − A)
−1 − (I − AT )
tr (cid:2)(P P T − P 2)(cid:3)
tr (cid:2)(P T − P )(P T − P )T(cid:3) = tr (cid:2)RRT(cid:3) ≥ 0
16
(A17)
(A18)
which upon expanding leads to
tr (cid:2)P P T − P 2(cid:3) ≥ 0
(A19)
and ensures the positivity of tr [K1], under the assumption C0 ≈ I. Moreover, we can show
that this is valid for an arbitrary C0. Under the assumption of stability and homogeneous
Gaussian noise, the correlation and weight matrices are related by the Lyapunov equation
[21, 22]:
W C0 + C0W T = −QQT
(A20)
where QQT is the generalized temperature tensor of the noise, whose components QiQj =
σ2
ij are the corresponding noise variances. For the case we are considering, QQT = I, a
symmetric system W = W T has the solution C0 = −W −1/2. A formal solution for the
general case is [23]
Assuming homogeneous noise, this reduces in our case to
Following the derivation in A18-A19, the full expression for tr [K1] can be written as:
(A21)
(A22)
(A23)
(A24)
(cid:3)
(A25)
C0 =
eW T teW tdt
(cid:90) ∞
(cid:90) ∞
0
−2IteAT teAtdt
e
C0 =
0
(cid:3)
tr [K1] = tr (cid:2)RRT C0
eA(cid:105)
= tr (cid:2)(eAR)(eAR)T(cid:3) ≥ 0
Given that
(cid:104)
tr
RRT eAT
It follows through A22 that tr [K1] ≥ 0
NEW STARTS HERE @@@@@@@@
Similarly, the second term in Eq. A16, tr [K2], can be rewritten as
tr (cid:2)AT (I − A)
−1C0
(cid:3) − tr (cid:2)A(I − A)
(cid:3)
−1C0
tr (cid:2)(AT − A)(I − A)
(cid:3)
−1C0
which can be reduced to
Replacing A(I − A)−1 by (I − A)−1 − I, the term can be transformed to
tr [C0] − tr (cid:2)(I − AT )(I − A)
−1C0
(cid:3) = tr [C0] − tr (cid:2)W T W
−1C0
17
Assuming that QQT = I in Eq. A20, and pre-multiplying by W −1, we obtain
−1C0W T = −W
W
−1C0] = −tr [W
−1 − C0
−1] − tr [C0]
tr [W T W
from which we derive:
tr [K2] = 2tr [C0] + tr [W
−1]
(A26)
(A27)
(A28)
Now we can use the formal solution of the Lyapunov equation, Eq. A21, and the fact that
for a stable matrix such as W (i.e. all its eigenvalues have negative real components) we
can write
−1 = −
W
eW tdt
(cid:90) ∞
and modify Eq. A28 accordingly:
tr [K2] = 2
0
tr [eW teW T t]dt −
(cid:90) ∞
0
(cid:90) ∞
0
tr [eW t]dt
(A29)
Given that for any matrix, the eigenvalues satisfy λ(eW ) = eλ(W ), we get
(cid:90) ∞
N(cid:88)
(cid:90) ∞
N(cid:88)
tr [K2] = 2
e2Re(λk)tdt −
0
k=1
0
k=1
eλktdt
(A30)
where λk are the N eigenvalues of W . Calling λk = −µk + iγk, µ > 0, the first term in the
r.h.s. above is
(cid:90) ∞
2
0
N(cid:88)
k=1
(cid:90) ∞
N(cid:88)
0
k=1
−2µktdt =
e
−µktdt
e
Now we can compare both terms in the r.h.s. of Eq. A30 for each k and t: e−µkt ≥ e−µkt+iγkt,
which leads directly to tr [K2] ≥ 0 and completes the proof of the semi-negativity of the
changes in the energy function (Eq. A16).
Interestingly, this result is related to a property of M -matrices. For an M -matrix: (1)
the off-diagonal elements are semi-negative, Mij ≤ 0 ∀i (cid:54)= j, and (2) is positive stable,
Re[λi(M )] > 0 ∀i. It can be shown that for any M -matrix of dimension N (Th. 5.7.23 in
[23]) tr (cid:2)M T M−1(cid:3) ≤ tr [I] = N . Choosing −W as an M -matrix, the theorem leads to a
similar result for the non-negativity of tr[K2] when C0 is close to the identity.
We have assumed throughout that the system is in a regime of dynamical stability, and
presented a case for the stabilizing effect of STDP by linking loops and eigenvalues in Eq. 3.
It follows that loopiness minimization (Eq. 2) is equivalent to maximization of stability (as
18
defined by the l.h.s. of Eq. 3), constrained by the total matrix weight. We can further
effect on U = −(cid:80)
understand this by explicitly expanding to first order the update equation A14 to see the
i log λi. Assuming again that QQT = I, the solution to the Lyapunov
equation (Eq. A21) can be approximated by a series expansion in powers of A [23]; to first
approximation C0 (cid:39) I + 1
δU (cid:39) 1
more stable.
2tr [A δA + δA A], and in turn δU (cid:39)(cid:0)tr [A2] − tr (cid:2)AAT(cid:3)(cid:1) ≤ 0, making the system
2(A + AT ), leading to ∆A ∼ A − AT . Through Eq. 3 we obtain
UP TO HERE @@@@
Acknowledgments
We would like to acknowledge the helpful contributions of Lucas Monz´on, University of
Colorado at Boulder, and Gustavo Stolovitzky, IBM Research, to the mathematical analysis.
[1] Braitenberg V. and Schuz A. Cortex: Statistics and geometry of neuronal connectivity (2nd
ed.). Berlin: Springer-Verlag, 1998.
[2] Stepanyants A., Hirsch J.A., Martinez L.M., Kisvrday Z.F., Ferecsk A.S., and Chklovskii D.B.
Local potential connectivity in cat primary visual cortex. Cerebral Cortex, 18(1):13 -- 28, 2007.
[3] Crick F. and Koch C. Constraints on cortical and thalamic projections: the no-strong-loops
hypothesis. Nature, 391:245 -- 250, 1998.
[4] Le Be J-V. and Markram H. Spontaneous and evoked synaptic rewiring in the neonatal
neocortex. Proc. Natl. Acad. Sci. USA, 103:13214 -- 13219, 2006.
[5] Markram H., Lubke J., Frotscher M., and Sakmann B. Regulation of synaptic efficacy by
coincidence of postsynaptic APs and EPSPs. Science, 275:213 -- 215, 1997.
[6] Morrison A., Diesmann M., and Gerstner W. Phenomenological models of synaptic plasticity
based on spike timing. Biol. Cybern., 98:459 -- 478, 2008.
[7] Abbott L.F. and Nelson S. B. Synaptic plasticity: taming the beast. Nature Neuroscience,
3(11s):1178 -- 1183, 2000.
[8] Song S., Miller K.D., and Abbott L.F. Competitive hebbian learning through spike-timing-
dependent synaptic plasticity. Nat. Neurosci., 3:919 -- 926, 2000.
19
[9] Burkitt A.N., Meffin H., and Grayden D.B. Spike-timing-dependent plasticity: The relation-
ship to rate-based learning for models with weight dynamics determined by a stable fixed
point. Neural Comput., 16:885 -- 940, 2004.
[10] Gutig R., Aharonov R., Rotter S., and Sompolinsky H. Learning input correlations through
nonlinear temporally asymmetric hebbian plasticity. J. Neurosci., 23:3697 -- 3714, 2003.
[11] Ma'ayan A., Cecchi G.A., Wagner J., Rao A.R., Iyengar R., and Stolovitzky G. Ordered
cyclic motifs contributes to dynamic stability in biological and engineered networks. Proc.
Natl. Acad. Sci. USA, 105(49):19235 -- 40, 2008.
[12] Fino E., Deniau J-M., and Venance L. Cell-specific spike-timing-dependent plasticity
in gabaergic and cholinergic interneurons in corticostriatal rat brain slices.
J. Physiol.,
586(1):265 -- 282, 2008.
[13] Seol G.H., Ziburkus J., Huang S., Song L., Kim I.T., and Takamiya K. Neuromodulators
control the polarity of spike-timing-dependent synaptic plasticity. Cell, 55:919 -- 929, 2007.
[14] Prasolov V.V. Problems and theorems in linear algebra. American Mathematical Society,
1994.
[15] Song S., Sjostrom P.J., Reigl M., Nelson S., and Chklovskii D.B. Highly nonrandom features
of synaptic connectivity in local cortical circuits. PLoS biology, 3(3):507 -- 509, 2005.
[16] Fino E., Glowinski J., and Venance L. Bidirectional activity-dependent plasticity at cortico-
striatal synapses. J. Neuroscience, 25(49):11279 -- 11287, 2005.
[17] Abeles M. Corticonics: Neural Circuits of the Cerebral Cortex. Cambridge, 1991.
[18] Hosaka R., Araki O., and Ikeguchi T. Stdp provides the substrate for igniting synfire chains
by spatiotemporal input patterns. Neural Computation, 20:415 -- 435, 2008.
[19] Pawlak V. and Kerr J.N.D. Dopamine receptor activation is required for corticostriatal spike-
timing dependent plasticity. J. Neuroscience, 28(10):2435 -- 2446, 2008.
[20] Thivierge J.P. and Cisek P. Nonperiodic synchronization in heterogeneous networks of spiking
neurons. J. Neuroscience, 28(32):7968 -- 7978, 2008.
[21] Risken H. The Fokker Planck Equation: Methods of Solutions and Applications (2nd Ed.).
Springer, 1996.
[22] DelSole T. Stochastic models of shear-flow turbulence with enstrophy transfer to subgrid
scales. J. Atmosph. Sci., 56:3692 -- 3703, 1999.
[23] Horn R.A. and Johnson C.R.J. Topics in Matrix Analysis. Cambridge University Press, 1991.
20
[24] We use microcircuitry to refer to neural circuitry observed at the level of neurons and synapses
and not necessarily in reference to a restricted spatial extent of these neurons (e.g., a "col-
umn"). For this reason, we view every brain connection as part of some microcircuit topology.
[25] Note that paths that traverse the same network node more than once are also counted.
21
FIG. 1: Schematic of the topological effect of STDP. Feedback connections in an initial topology
(left) from first (1) and second (2) order "follower" neurons (light blue) to a "trigger" neuron
(green) create loops of length k = 2 and k = 3. These connections are selectively penalized by
the STDP learning rule (lower middle, red). The plot (middle) depicts this rule, with the time
difference between follower (black) and trigger (green) action potentials on the x-axis, and the
expected synaptic modification on the y-axis. When spikes successfully propagate through the
loopy network they feed back to the trigger, arriving at the follower-trigger synapse immediately
after the trigger neuron fired and resulting in synaptic depression (red). Through repeated spike
propagation events, STDP results in a completely feed forward learned topology (right) to the
output neuron (dark blue).
22
121212INITIAL TOPOLOGYLEARNED TOPOLOGYLOOP-REGULATING STDP∆t∆w FIG. 2: Global topological effects of STDP. A) A monotonic decrease in the loopiness measure (see
Equation 2, first term) over time is observed in a simulated network of 100 neurons undergoing
STDP. Simultaneously, STDP results in a net increase in the weightedness of the network (inset,
see Equation 2, second term). Shown here and in B) is the average of 8 separate simulations of 20
seconds of network activity; error bars are standard deviation. B) Number of closed loops of length
5, 3, and 2, decreases as a function of weight threshold for network connections. Dotted lines show
counts for randomized networks with same number of total connections. C) Number of unique loops
sampled from five networks with varying synaptic delays following 10 seconds of simulated activity,
and from a random network with 5, 000 connections. Number of loops is shown as a function of
loop length. Loops were sampled across different learned networks while maintaining the number
of network connections at 5, 000 by varying the weight threshold (from 0.003 to 0.0046) for each
delay. Greater synaptic delays decreases the loop-eliminating topological effect of standard STDP.
23
00.0020.0040.0060.0080.011001051010Adjacency weight thresholdNumber of closed loopsFewer Closed Loops Than Randomlen=5len=3len=2 randomSTDP051015202510-2100102104106Fewer Unique Loops, Varying DelayLoop lengthNumber of unique loops random4.0 ms2.01.00.50.1051015200.050.10.150.2Time (secs)Decrease In Network Loopiness100n=11ktr(AT)k100n=112trAAT05101520121314151617181920Time (secs)Increase In Network WeightednessABC FIG. 3: Local topological effects of STDP. A) Inverse relationship of in-degree versus out-degree of
intra-network connections for each neuron in a network after 10 seconds of STDP across multiple
weight thresholds for network connections. Colors in left and right panels correspond to a weight
threshold used to construct the network over which degrees were measured. The color key can
be read from the left panels' vertical axes and corresponding color found along each manifold.
B) Correlated extra-network in-degree and intra-network out-degree indicate an opposite effect of
STDP on extra-network inputs.
24
05010005010000.0050.010.015koutkinAdjacency weight threshold020040060005010000.0050.010.015koutkinAdjacency weight threshold020406080100020406080100kinkoutIntra-network In-Degree, Intra-network Out-Degree0100200300400500020406080100kinkoutExtra-network In-Degree, Intra-network Out-DegreeAB FIG. 4: Effect of reverse STDP. A) In-degree versus out-degree of intra-network connections fol-
lowing different durations and polarities of STDP, shows a strong inverse relationship for standard
STDP. Adjacency weight threshold was 0.005. Green markers correspond to the network after 1
second of standard STDP, followed by 3 or 5 seconds of standard STDP (blue markers) or reverse
STDP (red markers). B) Total synaptic input versus output for intra-network connections shows a
similar inverse relationship for standard STDP. C) Number of closed loops of length 5, 3, and 2, is
decreased by standard STDP, and restored with reverse STDP (plotted as in Fig. 2B). D) In-degree
of extra-network inputs versus out-degree of intra-network outputs, plotted as in A, shows a strong
positive correlation for standard STDP. Adjacency weight threshold was 0.007 for extra-network
inputs and 0.005 for intra-network outputs. E) Total synaptic extra-network input versus total
synaptic intra-network output, plotted as in B, shows a similar positive correlation. In each panel
the topological effects of STDP are reversed by reverse STDP.
25
00.20.40.60.810.10.20.30.40.50.60.70.80.91Σ winΣ wout0204060801000102030405060708090100kinkoutIntra-network Inputs, Intra-network Outputs 2.830.10.20.30.40.50.60.70.80.912002500102030405060708090100Extra-netw1:STDP 1 s2:STDP 3 s3:STDP 5 s2:RSTDP 3 s3:RSTDP 5 s00.0050.011001051010len=5len=3len=200.0050.011001051010len=5len=3len=200.0050.011001051010Number of closed loopslen=5len=3len=200.0050.011001051010Adjacency weight thresholdlen=5len=3len=2ADCΣ woutkout1234BE FIG. 5: Dynamical effects of STDP. A) Raster plot of the spiking activity for a network after
STDP (top), and for a surrogate network where intra-network weights were reassigned randomly
to network connections, thus destroying STDP-learned topology (bottom). Each point corresponds
to a spike for each neuron. Each neuron was assigned a rank according to the sum of its extra-
network input weights, with the lowest rank corresponding to the highest sum. B) Peri-event time
histograms for each neuron in the STDP network (top left) and its surrogate (bottom left), pooled
across 8 separate simulations (bin width, 2msec). Histograms show different network propaga-
tion properties. Spike counts and extra-network weights for the same networks do not co-vary
in the STDP-learned topology (top right), but are highly correlated for the surrogate (bottom
right). C) Peri-stimulus time histograms summed across all simulations and all neurons for the
STDP network (blue) and three surrogates, in which the intra-network connections (red), extra-
network connections (green) or both (magenta) were randomized (top). Skewness versus kurtosis
26
of these histograms averaged across 8 separate simulations each (bottom, error bars show stan-
24682.62.833.23.4Spikes/secΣ win, extranet0510152.62.833.23.4Spikes/secΣ win, extranet1010.51111.51212.51302040608010020406080100Peri-event time (msec)Neuron Rank, Σ win, extranetNon-randomized Network1010.51111.51212.51302040608010020406080100Peri-event time (msec)Neuron Rank, Σ win, extranetIntra-network-randomized Network-15-10-505101520-20-15-10-505101520-20ABt (sec)t (sec)Non-randomized NetworkIntra-network-randomized Network |
1605.05291 | 1 | 1605 | 2016-05-17T19:14:03 | Multistep Model for Predicting Upper-Limb 3D Isometric Force Application from Pre-Movement Electrocorticographic Features | [
"q-bio.NC",
"cs.HC"
] | Neural correlates of movement planning onset and direction may be present in human electrocorticography in the signal dynamics of both motor and non-motor cortical regions. We use a three-stage model of jPCA reduced-rank hidden Markov model (jPCA-RR-HMM), regularized shrunken-centroid discriminant analysis (RDA), and LASSO regression to extract direction-sensitive planning information and movement onset in an upper-limb 3D isometric force task in a human subject. This mode achieves a relatively high true positive force-onset prediction rate of 60% within 250ms, and an above-chance 36% accuracy (17% chance) in predicting one of six planned 3D directions of isometric force using pre-movement signals. We also find direction-distinguishing information up to 400ms before force onset in the pre-movement signals, captured by electrodes placed over the limb-ipsilateral dorsal premotor regions. This approach can contribute to more accurate decoding of higher-level movement goals, at earlier timescales, and inform sensor placement. Our results also contribute to further understanding of the spatiotemporal features of human motor planning. | q-bio.NC | q-bio | Multistep Model for Predicting Upper-Limb 3D Isometric Force
Application from Pre-Movement Electrocorticographic Features*
Jing Wu, Student Member, IEEE, Benjamin R. Shuman, Bingni W. Brunton, Katherine M. Steele,
Jared D. Olson, Rajesh P.N. Rao, Member, IEEE, and Jeffrey G. Ojemann
Abstract-Neural correlates of movement planning onset and
direction may be present in human electrocorticography in the
signal dynamics of both motor and non-motor cortical regions.
We use a three-stage model of jPCA reduced-rank hidden
Markov model
(jPCA-RR-HMM), regularized shrunken-
centroid discriminant analysis (RDA), and LASSO regression
to extract direction-sensitive planning
information and
movement onset in an upper-limb 3D isometric force task in a
human subject. This mode achieves a relatively high true
positive force-onset prediction rate of 60% within 250ms, and
an above-chance 36% accuracy (17% chance) in predicting one
of six planned 3D directions of isometric force using pre-
movement signals. We also
find direction-distinguishing
information up to 400ms before force onset in the pre-
movement signals, captured by electrodes placed over the limb-
ipsilateral dorsal premotor regions. This approach can
contribute to more accurate decoding of higher-level movement
goals, at earlier timescales, and inform sensor placement. Our
results also contribute to further understanding of the
spatiotemporal features of human motor planning.
I. INTRODUCTION
Decoding human arm movement has been a major goal of
motor brain-computer interfaces (BCIs) for the development
upper-limb neuroprostheses. Many previous decoding
approaches have related recorded sensorimotor signals with
specific virtual or robotic movements [1]–[3], allowing a
potential BCI to reconstruct intended movements from a
combination of basic signal changes. For example, primary
motor cortex signals during individual finger movements
could be combined to reconstruct multi-axis virtual cursor
movements [4], [5], or pinch or grasp movements [6]–[8].
Alternatively, more
of
movements, or movement plans, may be present in recorded
representations
abstract
*Research supported by grants from NSF EEC-1028725, NINDS
5R01NS065186, NIH 2K12HD001097, NIH 5U10NS086525, and the WRF
Fund for Innovation in Neuroengineering.
J. Wu is with the department of Bioengineering, University of
Washington, Seattle, WA 98105 USA. (phone: 858-729-8669; e-mail:
[email protected]).
B.R. Shuman is with the department of Mechanical Engineering,
(e-mail:
University of Washington, Seattle, WA 98105 USA
[email protected])
B.W. Brunton is with the department of Biology, University of
Washington, Seattle, WA 98105 USA. (e-mail: [email protected]).
K.M. Steele is with the department of Mechanical Engineering,
(email:
University of Washington, Seattle, WA 98105 USA
[email protected])
J.D. Olson is with the department of Rehabilitation Medicine, University
of Washington, Seattle, WA 98105 USA. (email: [email protected])
R.P.N. Rao is with the department of Computer Science, University of
Washington, Seattle, WA 98105 USA. (e-mail: [email protected]).
J.G. Ojemann is with the department of Neurosurgery, University of
Washington, Seattle, WA 98105 USA. (e-mail: [email protected]).
brain signals prior to the onset of movement. The presence of
these signals may especially be true in areas outside of the
sensorimotor cortex. For instance, studies in non-human
primates and human neuroimaging have established that,
during an upper-extremity movement in a reach task [9]–
[11], the parietal and premotor cortices play a key role in the
representation of target trajectories and executing planned
movements in the desired spatial coordinates.
implanted for medically
An effective control signal for a BCI must maximize the
information available, including preparation prior to the onset
of movement. Patients with indwelling electrocorticography
(ECoG) electrodes
intractable
epilepsy provide recordings with centimeter resolution
combined with millisecond sampling over a large brain
region. These signals offer an opportunity to detect correlates
of the entire movement preparation and execution process.
Previous work have demonstrated the capacity of ECoG
signals to resolve the timing and interaction between regions
during finger movements [8], upper-limb movements [1], [2],
as well as premotor responses that precede the sensory and
motor cortex activation up to 100ms in advance [12]. One
previous work has also shown some success in detecting
ECoG representations of onset and direction of hand-arm
movements [13].
However, the temporal and spatial distributions of the
recorded signal in motor preparatory regions are not well-
established in human ECoG. It is yet unclear which non-
motor regions contain ECoG signals with direction-
distinguishing preparatory information, whether these signals
are localized to just the contralateral hemisphere or both
hemispheres, and how they are localized in time. It is also
unknown whether correlates of preparatory activity are
limited to reach trajectories, or also can generalize to other
directional tasks such as isometric force application towards
planned directions.
These unknown factors are particularly relevant to the
engineering of BCI, as signals conveying information about
the planning of a movement may allow for better resolution
of differing task goals, inform locations for sensor placement,
and enable decoding of these goals earlier than signals in the
sensorimotor cortex.
In this work, we investigated the recorded ECoG
activities of a patient with fortuitous bilateral implantation of
electrodes in the parietal and arm-associated sensorimotor
cortices. The patient performed a right arm isometric force
task toward designated directions. We constructed a three-
stage model to examine the localization and timing of
direction-discriminable information, allowing for the parallel
prediction of isometric force onset and direction.
II. METHODS
A. Experimental Task and Data Acquisition
The patient is a 26-year-old right-handed male with
intractable seizures thought to be related to a medial parietal
lesion; invasive monitoring of bilateral parietal regions was
recommended. The subject received pre-operative anatomical
MRI, and was implanted with two frontoparietal 2×8
platinum electrode grids and two 1×8 strips (Ad-Tech,
Racine, WA, USA), with 3mm electrode-diameters
embedded in silastic with center-to-center distances of
10mm. Subsequently, X-ray CT imaging captured the
location of the electrodes. These electrodes were sampled at
1220.7 Hz on a 128-channel Tucker-Davis Technology
(Alachua, FL, USA) System 3 RZ5D Neurophysiology Base
Processor with associated PZ5 NeuroDigitizer. The subject
had normal motor function pre-operatively and throughout
the experiments and inpatient stay. The patient gave written,
informed consent through a protocol approved by the
University of Washington Institutional Review Board.
The patient was comfortably positioned in front of one
screen during the task, and was visually cued to apply
isometric force in one of 6 directions (up, down, left, right,
forward, back) to an affixed AMTI force transducer handle
(Watertown, MA, USA) held in the right hand. Each trial
consisted of a hold with approximately 15 lbs (67 N) of force
and 2 seconds in duration, with an inter-trial interval of 3
seconds. 10 trials of each of 6 directions were recorded for a
total of 60 isometric force trials.
B. Data Preprocessing and Feature Extraction
using
Statistical
Parametric
Pre-operative MRI was co-registered with postoperative
Mapping
CT
(http://www.fil.ion.ucl.ac.uk/spm), with
the pial surface
reconstructed with FreeSurfer (http://freesurfer.net) and
custom mapping and projection code [14] implemented in
MATLAB (Natick, MA, USA). The two frontoparietal
hemisphere 2×8 grid recordings were visually inspected for
data quality, and channels with high level of artifacts and low
impedance were rejected (1 in the contralateral L grid and 2
in R). The remaining 29 recorded channels were filtered with
a common average reference filter siʹ=si−J−1∑jsj for J non-
rejected channels within each separate lateral grid. The
filtered signals were then analyzed for time-frequency
content by continuous wavelet transform with the non-
analytic Morlet wavelet defined by Ψ(sω) = π−¼e[−(sω−ω0)^2]/2,
where ω0=6 and s reflects log scaling factors to obtain ⅛-
octave resolution across pseudo-frequencies 2−200 Hz. We
used the absolute magnitude of the wavelet coefficient time
series at each scaling, and also extracted the lowpassed ECoG
local motor potential (LMP) for each channel as outlined in
[15] as an additional time series. All time series were binned
by 50ms and normalized to unit variance.
Recorded force transducer voltages were converted to
force using manufacturer sensitivity calibration data;
Cartesian forces were transformed to spherical coordinates to
obtain task force direction and magnitude. Subsequently, the
initiation and end of recorded isometric force in each trial
were carefully hand-labeled (examples shown with dotted
force traces and vertical onset lines in Figure 3).
C. Pre-Movement Classification of Force Direction
jPCA-RR-HMM analysis: We constructed a reduced-rank
hidden Markov model (RR-HMM) [16] from the normalized
binned time series using jPCA, a dimensionality reduction
method designed to find orthonormal bases that also exhibit a
rotational structure, with the expectation of a rhythmic
repeating behavior from the signal dynamics [17]. To find
these repeating mean trajectory templates, filtered time series
were split into 2-second segments center-aligned to the
initiation of movement, reflecting a
time period of
[−1000ms…+1000ms] for each movement trial. The high-
dimensional time series for each channel × frequency xw,t,c for
[t-1000, t-950, …t0…t+1000] were averaged for each of the 6
movement direction conditions d = [1…6]. jPCA was used
on these mean timeseries to obtain a set of jPC coefficients
that describe a 10-dimensional space to maximize rotational
dynamics in these mean ECoG signal trajectories. The
original trial segments xw,t,c,d for [t-1000…t+1000] were then
projected through these jPCs, such that each trial segment is
now reduced to a 10D space as xʹ[jPC1,2],t,d for [t-1000…t+1000].
All time points were then labeled as belonging to one of
12 clusters agnostic
to condition, using unsupervised
hierarchical clustering of all xʹ time points with Ward's
method, which recursively locates clusters that minimize
within-cluster distance variance in the dimensionally-reduced
Euclidean space. This method generated a list of observed
states st,d for t = [−1000ms … +1000ms], where all observed s
are categorical class labels, s ϵ [1…12], and d is one of six
directions of movement.
The pre-movement cluster-time-series were then isolated
and used as the sole training data to construct the HMM.
Cluster labels immediately preceding movement st,d for t =
[−500ms…0ms] were grouped by movement direction
condition, and used as observed sequences to train six hidden
Markov models Hd, one for each force movement direction
with d = [1…6]. The Baum-Welch algorithm was used with
12 observed states and 8 hidden states to train each Hd.
Direction predictions were made with
leave-one-out
validation, where each HMM generated from training st,d was
used to calculate the posterior state probability of observing
the sequences in the testing st,d. The direction associated with
the Hd giving the highest posterior probability (smallest
negative log-probability) was used as the classified direction.
Regularized discriminant analysis: The normalized,
binned time series were split into 1-second chunks aligned to
the initiation of movement segments, reflecting a time period
of
the start
corresponding to 1 second before movement initiation. Each
feature xw,t,c at wavelet scale w, timepoint t, and channel c is
used as a feature for the training of a multiclass regularized
discriminant analysis model with 10-fold validation, where
the posterior probability that observation x belongs to class k
follows
[−1000ms…0ms]
trial, with
for each
Shrunken centroids regularization as described in [18]
attempts to reduce overfitting bias by (a) introducing a
modified covariance matrix to stabilize the cross-validated
sample covariance Σ and (b) applying a threshold to a
modified correlation matrix to prune down the number of
useful predictors, as described by
where,
This is accomplished by iteratively searching through
parameters γ and δ, where γ modifies the sample covariance
matrix Σ, and δ thresholds the correlation matrix C and
reduces the number of predictors by thresholding them from
the posterior probability evaluation [18].
D. Movement Onset and Force Magnitude Prediction
(Lasso)
regression
The normalized binned time series was fit to the
magnitude of the force using least absolute shrinkage and
selection operator
[19]. Two-fold
validation was used wherein the whole time series of the each
half of the recorded movement session rw,c, with channel ×
frequency wavelet coefficients through time was used to fit
linear regression coefficients β where predicted ŷt = βw,c,t rw,c,t
+ β0. This generates βw,c,t as a function of a regularization
parameter λ, which thresholds the number of nonzero β
coefficients. λ was chosen such that the mean-squared-error is
the lowest in validation, which resulted in 132 nonzero β
coefficients, which were used to predict the forces in the
other half of the recorded session. The predicted force time
series ŷt were then smoothed using a 1st-order 2 Hz, −10dB
Butterworth lowpass filter and numerically differentiated to
obtain dŷt/dt. Peak analysis was used to find the local maxima
of dŷt/dt using a fixed threshold, and the local maxima points
represent potential times of force onset.
III. RESULTS
A. Force Direction Classification Using jPCA-RR-HMM
An average HMM classification accuracy of 35.6% was
obtained using pre-movement clustered timepoints 500ms
prior to the onset of movement. Accuracies were computed
using
force
application directions using jPCA-projected, hierarchically
clustered pre-movement ECoG. This performance is signifi-
leave-one-out cross-validation between 6
Figure 1. Confusion matrix of the model classification accuracies using
jPCA-RR-HMM, in 59 valid isometric force application trials in 6
directions using pre-movement ECoG recordings.
Regularized features in time and space
107
7
Ch 26
Ch 5
Ch 17
-400
-300
-200
-100
200
180
160
140
120
100
80
60
40
20
)
z
H
(
y
c
n
e
u
q
e
r
f
o
d
u
e
s
p
t
l
e
e
v
a
W
-500
)
s
s
e
l
t
i
n
u
(
l
e
u
a
v
6
5
4
3
2
1
0
0
Time before movement (ms)
Figure 2. Left: Blue = recorded electrodes; Yellow = recorded electrodes
corresponding to regularized features in Right plot; Gray = rejected for
artifacts. Right: Time-space plot of salient features from regularization.
Each block refers to one feature, most salient (high δ) features in yellow.
cantly higher than chance of 17%. Figure 1 shows the
confusion matrix of classification in the 6 directions, where
the main diagonal shows the correct classifications. Five of
six directions showed accuracies above chance, with the most
accurate direction being isometric movement back toward the
torso, which also had very low false-prediction rates.
B. Time-Space Localization of Directional Information
In a second analysis, we used regularized discriminant
analysis directly on the wavelet features of recorded pre-
movement ECoG signals
to obtain a cross-validated
directional classification accuracy of 27%. While this is not
as high as the outcome obtained with the jPCA-RR-HMM
analysis, RDA regularization and shrinkage operates on the
original feature space, and the remaining 27 features included
in computing the posterior class probability must exceed the
regularization threshold set by δ as outlined in equation (5).
These remaining wavelet coefficient features demonstrably
contain
directionally-discriminable
information, and are
time and space.
Interestingly, the spatiotemporal distribution of these features
as shown in Fig. 2 reveal that the earliest information about
movement is present in channel 26 (dorsal premotor cortex),
then passes on to channel 17, and finally to channel 5
(primary motor cortex).
C. Lasso Regression Onset and Force Prediction Results
cross-validated
localized
in
Cross-validated force predictions agree with overall force
magnitude (R2 = 0.42), but a strong predictive relationship
between peak force magnitude and ECoG signals magnitude
in this task was not found. However, the smoothed predicted
e
c
r
o
f
N
Figure 3. Force and onset timing predictions from Lasso regression.
Triangles: predicted force onset; dotted vertical lines: actual onset.
force time-series can be numerically derived to obtain an
accurate assessment of force onset, predicting onset to within
250ms at a 60% accuracy and within 500ms at an 80%
accuracy. The false positive predictive rate was 22%.
IV. DISCUSSION
A. Direction-Discriminating Dynamics
We showed that the combination of dimensionality
reduction and HMM allowed for RR-HMM states to capture
complex multidimensional dynamics in a relatively small
number of observed states. The performance of this technique
suggests that the dynamics between ECoG timepoints during
movement preparatory behavior may carry useful task-
relevant information. Though the predictive performance of
this technique is not sufficient alone to drive brain-computer
interfaces, it may be used in combination with techniques
such as Kalman filters that can benefit from earlier
probability updates to improve the inference of user intent.
B. Spatiotemporal Feature Localization
the
showed
We found that wavelet-based spectral characteristics of
ECoG
recordings can discriminate between different
directions of isometric arm movements significantly better
than chance in a regularized discriminant analysis, despite the
overlap between the recruited muscle groups and the
isometric task. We find that signals carrying direction-
discriminatory information occurred over the ipsilateral
premotor cortex up to 400ms before the onset of detected
force, in contrast to the signals from the contralateral
sensorimotor
strongest
discrimination ability starting roughly 100−150msec prior to
movement onset (Fig. 2). The timing of the information in
ipsilateral changes is particular interesting in potentially
offering a technique to screen for neural processes spanning
multiple cortical regions, and may give hints to the possible
timing of distributed trajectory planning processes.
C. Differences Between Movement Trajectory and
cortex which
Magnitude
While the ECoG signal can be used to both decode
direction and force onset in parallel, we found in this dataset
that high-performing regression features in predicting force
changes do not also perform well in discriminating between
directions, potentially suggesting separate usable ECoG
signatures for motor execution and planning.
D. Future Considerations
Further work is necessary to determine whether these
particular correlates of movement planning generalize across
tasks to other types of upper-limb movements, or generalizes
to other measures of force application such as non-task
muscle contraction. We also wish to examine the possibility
of learning effects with repeated practice.
ACKNOWLEDGMENT
We would
like
to
thank Lila Levinson for her
outstanding exploratory analyses on this dataset.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[13]
REFERENCES
[1]
T. Pistohl, T. Ball, A. Schulze-Bonhage, A. Aertsen, and C.
Mehring, "Prediction of arm movement trajectories from ECoG-
recordings in humans," J. Neurosci. Methods, vol. 167, no. 1, pp.
105–114, Jan. 2008.
N. R. Anderson, T. Blakely, G. Schalk, E. C. Leuthardt, and D. W.
Moran, "Electrocorticographic (ECoG) correlates of human arm
movements," Exp. Brain Res., vol. 223, no. 1, pp. 1–10, Sep. 2012.
[3] M. S. Fifer, S. Acharya, H. L. Benz, M. Mollazadeh, N. E. Crone,
[2]
and N. V. Thakor, "Towards Electrocorticographic Control of a
Dexterous Upper Limb Prosthesis," IEEE Pulse, vol. 3, no. 1, pp.
38–42, Jan. 2012.
J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G. Pfurtscheller, and
T. M. Vaughan, "Brain–computer interfaces for communication and
control," Clin. Neurophysiol., vol. 113, no. 6, pp. 767–791, Jun.
2002.
G. Schalk, K. J. Miller, N. R. Anderson, J. A. Wilson, M. D. Smyth,
J. G. Ojemann, D. W. Moran, J. R. Wolpaw, and E. C. Leuthardt,
"Two-dimensional movement control using electrocorticographic
signals in humans," J. Neural Eng., vol. 5, no. 1, p. 75, Mar. 2008.
P. Shenoy, K. J. Miller, J. G. Ojemann, and R. P. N. Rao, "Finger
Movement Classification for an Electrocorticographic BCI," in 3rd
International IEEE/EMBS Conference on Neural Engineering,
2007. CNE '07, 2007, pp. 192–195.
J. Kubánek, K. J. Miller, J. G. Ojemann, J. R. Wolpaw, and G.
Schalk, "Decoding flexion of individual fingers using
electrocorticographic signals in humans," J. Neural Eng., vol. 6, no.
6, p. 066001, Dec. 2009.
K. J. Miller, C. J. Honey, D. Hermes, R. P. Rao, M. denNijs, and J.
G. Ojemann, "Broadband changes in the cortical surface potential
track activation of functionally diverse neuronal populations,"
NeuroImage, vol. 85, Part 2, pp. 711–720, Jan. 2014.
L. R. Bremner and R. A. Andersen, "Coding of the Reach Vector in
Parietal Area 5d," Neuron, vol. 75, no. 2, pp. 342–351, Jul. 2012.
P. S. Archambault, S. Ferrari-Toniolo, R. Caminiti, and A.
Battaglia-Mayer, "Visually-guided correction of hand reaching
movements: The neurophysiological bases in the cerebral cortex,"
Vision Res., Sep. 2014.
S. Schaffelhofer, A. Agudelo-Toro, and H. Scherberger, "Decoding
a Wide Range of Hand Configurations from Macaque Motor,
Premotor, and Parietal Cortices," J. Neurosci., vol. 35, no. 3, pp.
1068–1081, Jan. 2015.
[12] H. Sun, T. M. Blakely, F. Darvas, J. D. Wander, L. A. Johnson, D.
K. Su, K. J. Miller, E. E. Fetz, and J. G. Ojemann, "Sequential
activation of premotor, primary somatosensory and primary motor
areas in humans during cued finger movements," Clin.
Neurophysiol., vol. 126, no. 11, pp. 2150–2161, Nov. 2015.
Z. Wang, A. Gunduz, P. Brunner, A. L. Ritaccio, Q. Ji, and G.
Schalk, "Decoding onset and direction of movements using
Electrocorticographic (ECoG) signals in humans," Front.
Neuroengineering, vol. 5, p. 15, 2012.
[14] D. Hermes, K. J. Miller, H. J. Noordmans, M. J. Vansteensel, and
N. F. Ramsey, "Automated electrocorticographic electrode
localization on individually rendered brain surfaces," J. Neurosci.
Methods, vol. 185, no. 2, pp. 293–298, Jan. 2010.
[15] G. Schalk, J. Kubánek, K. J. Miller, N. R. Anderson, E. C.
Leuthardt, J. G. Ojemann, D. Limbrick, D. Moran, L. A. Gerhardt,
and J. R. Wolpaw, "Decoding two-dimensional movement
trajectories using electrocorticographic signals in humans," J.
Neural Eng., vol. 4, no. 3, p. 264, Sep. 2007.
S. M. Siddiqi, B. Boots, and G. J. Gordon, "Reduced-Rank Hidden
Markov Models," ArXiv09100902 Cs, Oct. 2009.
[16]
[17] M. M. Churchland, J. P. Cunningham, M. T. Kaufman, J. D. Foster,
P. Nuyujukian, S. I. Ryu, and K. V. Shenoy, "Neural population
dynamics during reaching," Nature, vol. 487, no. 7405, pp. 51–56,
Jul. 2012.
[18] Y. Guo, T. Hastie, and R. Tibshirani, "Regularized linear
discriminant analysis and its application in microarrays,"
Biostatistics, vol. 8, no. 1, pp. 86–100, Jan. 2007.
[19] R. Tibshirani, "Regression Shrinkage and Selection via the Lasso,"
J. R. Stat. Soc. Ser. B Methodol., vol. 58, no. 1, pp. 267–288, 1996.
|
1012.3623 | 1 | 1012 | 2010-12-16T14:37:47 | Information capacity and transmission are maximized in balanced cortical networks with neuronal avalanches | [
"q-bio.NC",
"cond-mat.dis-nn",
"physics.bio-ph"
] | The repertoire of neural activity patterns that a cortical network can produce constrains the network's ability to transfer and process information. Here, we measured activity patterns obtained from multi-site local field potential (LFP) recordings in cortex cultures, urethane anesthetized rats, and awake macaque monkeys. First, we quantified the information capacity of the pattern repertoire of ongoing and stimulus-evoked activity using Shannon entropy. Next, we quantified the efficacy of information transmission between stimulus and response using mutual information. By systematically changing the ratio of excitation/inhibition (E/I) in vitro and in a network model, we discovered that both information capacity and information transmission are maximized at a particular intermediate E/I, at which ongoing activity emerges as neuronal avalanches. Next, we used our in vitro and model results to correctly predict in vivo information capacity and interactions between neuronal groups during ongoing activity. Close agreement between our experiments and model suggest that neuronal avalanches and peak information capacity arise due to criticality and are general properties of cortical networks with balanced E/I. | q-bio.NC | q-bio | Information capacity and transmission are maximized in
balanced cortical networks with neuronal avalanches
Woodrow L. Shew*1, Hongdian Yang (杨黉典)*1,2, Shan Yu (余山)1, Rajarshi Roy2,
Dietmar Plenz1
1 Section on Critical Brain Dynamics, Laboratory of Systems Neuroscience, National
Institutes of Mental Health, Bethesda, MD, USA 20892;
2 Institute for Physical Science and Technology, University of Maryland, College Park,
MD, USA 20742
* These authors contributed equally to this work
The repertoire of neural activity patterns that a cortical network can
produce constrains the network’s ability to transfer and process
information. Here, we measured activity patterns obtained from multi-site
local field potential (LFP) recordings in cortex cultures, urethane
anesthetized rats, and awake macaque monkeys. First, we quantified the
information capacity of the pattern repertoire of ongoing and stimulus-
evoked activity using Shannon entropy. Next, we quantified the efficacy of
information transmission between stimulus and response using mutual
information. By systematically changing the ratio of excitation/inhibition
(E/I) in vitro and in a network model, we discovered that both information
capacity and information transmission are maximized at a particular
intermediate E/I, at which ongoing activity emerges as neuronal
avalanches. Next, we used our in vitro and model results to correctly
predict in vivo information capacity and interactions between neuronal
groups during ongoing activity. Close agreement between our
experiments and model suggest that neuronal avalanches and peak
information capacity arise due to criticality and are general properties of
cortical networks with balanced E/I.
2
INTRODUCTION
In the cortex, populations of neurons continuously receive input from upstream neurons,
integrate it with their own ongoing activity, and generate output destined for downstream
neurons. Such cortical information processing and transmission is limited by the repertoire
of different activated configurations available to the population. The extent of this
repertoire may be quantified by its entropy H; in the context of information theory,
entropy characterizes the information capacity of the population (Shannon, 1948; Rieke et
al., 1997; Dayan and Abbott, 2001). Information capacity is important because, as the
name suggests, it defines upper limits on aspects of information processing. For example,
consider the information transmitted from input to output by a population that only has
two states in its repertoire (H≤1 bit). No matter how much information the input contains,
the information content of its output cannot exceed 1 bit. A network with low entropy
presents a bottleneck for information transmission in the cortex. Thus, it is important to
understand the mechanisms that modulate the entropy of cortical networks.
Cortical activity depends on the ratio of fast excitatory (E) to inhibitory (I) synaptic inputs
to neurons in the network. This E/I ratio remains fixed on average even during highly
fluctuating activity levels (Shu et al., 2003; Okun and Lampl, 2008). However, it is
currently not known whether a particular E/I ratio is advantageous for certain aspects of
information processing. The existence of such an optimal ratio is suggested by two
competing effects of E/I on entropy. First, a large E/I ratio, i.e. if excitation is
insufficiently restrained by inhibition, can cause very high correlations between neurons
(Dichter and Ayala, 1987). Since increased correlations decrease entropy (Rieke et al.,
1997; Dayan and Abbot, 2001), we anticipate that a sufficiently large E/I ratio limits
information transmission. This is consistent with findings that moderate levels of
correlation can play an important role in population coding (Pola et al., 2003; Jacobs et
al., 2009). At the other extreme, i.e. a small E/I ratio, weak excitatory drive is expected
to reduce correlations as well as the overall level of neural activity. Although reduced
correlations can lead to higher entropy, this increase may be counteracted by a concurrent
drop in activity. Sufficiently suppressed activity reduces entropy (Rieke et al., 1997;
Dayan and Abbot, 2001). Accordingly, we hypothesize that cortical entropy and
information transmission are maximized for an intermediate E/I ratio.
Here we tested our hypothesis experimentally in cortex cultures, anesthetized rats, and
awake monkeys and compared our results with predictions from a computational model.
We discovered an optimal intermediate E/I ratio distinguished by 1) maximal entropy and
2) maximal information transmission between input and network output. This finding was
based on analysis of both ongoing and stimulus-evoked population activity. Moreover, at
this optimal E/I ratio, ongoing activity emerges in the form of neuronal avalanches (Beggs
and Plenz, 2003) and interactions within the network are moderate. Agreement with our
model suggests that by maintaining this particular E/I ratio the cortex operates near
criticality and optimizes information processing.
MATERIAL AND METHODS
3
In vitro MEA recordings and pharmacology. Cortex-VTA (ventral tegmental area)
organotypic co-cultures were grown on planar integrated multi-electrode arrays (MEA;
for details see Shew et al., 2009). In brief, the MEA contained 60 recording electrodes
(8x8 grid with no corner electrodes, 30 μm diameter, 200 μm inter-electrode spacing).
LFP signals were sampled at 4 kHz and low-pass filtered (50 Hz cut off). Experiments
consisted of a 1 hour recording of ongoing activity, followed by 0.5 – 1 hr of stimulus-
evoked activity. Next, the antagonists were bath-applied and recordings of ongoing
activity followed by stimulus-evoked activity were repeated. We applied either 1) DNQX
(0.5-1 μM) + AP5 (10-20 μM) or 2) Picrotoxin (PTX, 5 μM). The stimuli consisted of
brief electric shocks delivered through a single electrode of the array located near layers
2/3 of the cortex culture. Each stimulus had a bipolar time course with amplitude –S for
50 μs in duration followed by an amplitude at +S/2 for 100 μs. We tested three similar
sets of stimuli, each with 10 different stimulus levels, in μA (S1=10, 20, 30, 50, 60, 80,
100, 120, 150, 200; S2=6, 12, 24, 50, 65, 80, 100, 125, 150, 200; or S3=6, 12, 24, 50, 74,
100, 150, 200, 300, 400.) Different amplitudes were applied in pseudorandom order at 5
s intervals.
MEA recordings in monkey. All procedures were in accordance with NIH guidelines and
were approved by the NIMH Animal Care and Use Committee. 96-channel MEA (10x10
grid with no corner electrodes, 400 μm separation, and 1.0 mm electrode length;
BlackRock Microsystem) were chronically implanted in the right arm representation
region of pre-motor cortex of two monkeys (Macaca mulatta, adults, one male, one
female). Ongoing activity was recorded for 30 min. Monkeys were awake, but not
engaged by any task or controlled sensory stimulation. Extracellular signals were sampled
at 30 kHz and filtered off-line (1 to 100 Hz; phase neutral, 4th order Butterworth). For
Figure 5B, a 4x4 subset of electrodes was analyzed, matching spatial dimensions of the
4x4 coarse resolution view of the in vitro data (see below). For Figure 5C, 4x4 patterns
based on a coarse-binned set of 8x8 electrodes spanning larger spatial area were analyzed.
MEA recordings in rats. Urethane anesthetized rats aged 15 – 25 days were studied. As
described previously (Gireesh and Plenz, 2008), an MEA was inserted into superficial
layers of barrel cortex (Michigan probe 8x4 electrodes, 200 μm interelectrode distance).
Ongoing activity was recorded for 20 – 30 min (4 kHz sampling rate, 1 – 3,000 Hz
bandwidth) and filtered between 1 – 100 Hz (phase neutral, 4th order Butterworth) to
obtain the LFP. In Figure 5B, the 8x4 MEA was coarse binned into a 4x4 pattern with 2
electrodes contributing to each bit. In Figure 5C, 4x4 patterns based on only 16
electrodes were also analyzed.
Population event detection. For each electrode, the standard deviation (SD) of the LFP
was calculated. Automated (Matlab) population event detection entailed first identifying
large (< -4SD in vitro, -3SD in vivo) negative LFP fluctuations. Second, the time stamp
ti, amplitude ai, and electrode ei of each negative peak (nLFP) occurring during these large
fluctuations was extracted. Next, consecutive nLFPs were assigned to the same
population event if they occurred sufficiently close (ti+1 – ti < τ) in time. The time
threshold τ was determined based on the inter-nLFP-interval distribution, as described
previously (Shew et al., 2009).
4
Event size distribution, neuronal avalanches, and κ. We defined the size s of a
population event as the sum of all ai that occur during the event. During one recording we
typically measured thousands of population events. We used a statistical measure called κ
to quantify the character of each recording. The measure κ was developed to provide a
network-level gauge of the E/I ratio, as demonstrated in a previous study (Shew et al.,
2009). Similar to the Kolmogorov statistic, κ is computed based on the cumulative
probability distribution of event sizes, F(β), which describes the probability of observing
an event with size less than β. More specifically, κ quantifies the similarity of the
measured F(β) and a theoretical reference distribution FNA(β). The reference distribution
was chosen based on empirical observations that cortical networks with unaltered E/I tend
to generate a probability distribution of population event sizes, P(s)~s-1.5, which defines
neuronal avalanches (Beggs and Plenz, 2003; Shew et al., 2009). Importantly, population
events are not considered to be neuronal avalanches unless they have this statistical
property. Thus, FNA(β)~β-0.5 is the cumulative distribution corresponding to neuronal
avalanches and κ assesses how close the observed distribution is to that of neuronal
avalanches. By definition, κ is 1 plus the sum of 10 differences between F(β) and FNA(β).
+=
k
1
1
10
(cid:229)
10
=
i
1
NA
F
(
b
)
i
-
F
(
b
)
i
.
(1)
Therefore, κ≈1 indicates a good match with the reference distribution and the existence of
neuronal avalanches. Positive (negative) deviations away from κ≈1 indicate a hyper-
(hypo-) excitable network and the absence of neuronal avalanches. The 10 points at
which the differences are computed are logarithmically spaced over the range of measured
event sizes. It was previously shown that κ is not sensitive to changes in the number of
differences (5 to 100 were tried; Shew et al., 2009).
Binary patterns and entropy H. Each population event is represented by an 8x8 binary
pattern with one bit per recording electrode. A bit is set to 1 if the corresponding
electrode is active during the event; otherwise it is set to 0. The entropy of this set of
patterns is defined as
-=
H
(cid:229)
n
=
i
1
p
i
log
2
p
i
,
(2)
where n is the number of unique binary patterns and pi is the probability that pattern i
occurs. In Figure 2B, the black curve is based on the 8x8 binary patterns which represent
the 60 electrodes of the MEA. We also studied entropy at different spatial resolutions by
coarse-binning and for different spatial extents by using sub-regions of the MEA. The
green curves in Figs. 2D, 3B, and 4A were obtained by reducing spatial resolution through
coarse-binning of 8x8 patterns into square 4x4 patterns. Each bit in the 4x4 pattern was
dependent on the state of 4 neighboring 2x2 electrode sets; if at least one electrode was
active the bit was set to 1. Reduced spatial extent was tested with16-bit patterns based on
only 16 electrodes arranged in a 4x4 square (Fig. 2D). As pointed out in the results, these
4x4 patterns also reduce potential undersampling bias when compared to 8x8 patterns.
For stimulus-evoked activity, binary patterns were defined based on LFP activity
5
measured during 500 ms following the stimulus. If the measured response at an electrode
exceeded -8SD of the noise, then the corresponding bit was set to 1, otherwise it was set
to 0. The stimulation electrode was always set to 0.
Note that the lack of corner electrodes on the MEA means that the corner bits of 8x8
patterns are always zero. This implies that the maximum entropy we could possibly
record for 8x8 patterns is 260 rather than 264. For coarse-binned 4x4 patterns, the
likelihood that the corner bits are active is slightly lower (about 25% lower). These
effects are present for all E/I ratios examined. Therefore, they may affect the absolute
values of entropy measurements, but they are not important for our conclusions, which are
primarily based on changes in entropy. This is further confirmed by the robustness of our
results to selecting 4x4 subregions from the center of the MEA for which corner
electrodes are not missing (Fig. 2D).
The calculation of entropy entails estimating the occurrence probability for each pattern.
Therefore, H generally will depend on number N of observed patterns unless N is so large
that the probability of each pattern is well represented by the samples recorded. H will be
underestimated for sufficiently small N but becomes independent of N for sufficiently large
N. To estimate potential ‘undersampling bias’ we computed corrected values following
the quadratic extrapolation method (e.g. Magri et al., 2009). First, we randomly selected
a fraction f of samples from the full set of N patterns. We recomputed the entropy for
fractions f=0.1 to 1 in steps of 0.1. We repeated this 10 times for each f. Next, we fit the
average H versus f data with the following function:
fH
(
)
=
H
0
-
a
fN
-
b
fN
(
) 2
(3)
The fit parameter H0 is the estimated corrected value reported in the results section.
Mutual information MI. From a set of N binary patterns, we defined a participation
vector qi (length N) for each recording site i. qi(j) =1 or 0 indicated that site i was active
or inactive during event j. The interaction between site i and site j was quantified by the
mutual information (Rieke et al., 1997; Dayan and Abbott, 2001) of qi and qj defined as
qqMI
;
(
i
j
)
=
(cid:229) (cid:229)
a
b
1,0
1,0
qp
(
i
=
qa
,
j
=
b
log)
2
(cid:230)
(cid:231)
(cid:231)
Ł
qp
(
qp
(
i
=
qa
,
j
i
=
qpa
(
)
=
b
=
j
)
b
)
(cid:246)
(cid:247)
(cid:247)
ł
,
(4)
where p(x) is the probability of x, p(x,y) is the joint probability x and y. MI quantifies (in
bits) the information shared by the two sites and provides similar information as a cross
correlation (Supplementary Fig. S1). The MI values reported in Results were averages
over all pairs of sites,
MI
=
2
MM
(
(cid:229)(cid:229)-
M
M
1
i qqMI
(
;
=
+=
i
j
i
1
1
)
.
j
-
)1
(5)
For the in vitro experiments, we also used mutual information in a different way to
6
quantify the efficacy of information transmission between stimulus and response. Here we
computed MI(S;R) = H(R) - H(RS). H(R) is the entropy of the full set of response
patterns for all stimuli. H(RS) is the conditional entropy, i.e. the response entropy for
single stimuli, averaged across the different stimuli (Rieke et al., 1997; Dayan and Abbott,
2001).
Likelihood of participation L Likelihood of participation Li for site i was defined as the
fraction of patterns in which the site participated:
=
L
i
1
N
(cid:229)
N
=
j
1
jq
)(
i
.
(6)
The average likelihood of participation L for all M sites is discussed in the main text and
defined as
=
L
1
NM
(cid:229)(cid:229)
M
N
=
=
i
j
1
1
jq
)(
i
.
(7)
Data shuffling to destroy interactions. For the purpose of understanding how the entropy
changes due to interactions between sites, we created surrogate data sets by shuffling the
events in which sites participated. The 1s and 0s were randomly reordered in each
participation vector qi such that interactions between sites were destroyed, but Li and N
remained fixed.
Model. The model consisted of M=16 binary sites (1=active, 0=inactive). Each site was
intended to model the activity of a large group of neurons like the nLFP recorded at an
electrode in the experiments. The strength of interactions between site i and site j was
modeled as an activation probability; pij was the probability that site i would be activated
at time t+1 if site j was active at time t. Therefore, if a set J(t) of sites were active at time
(cid:213)
-=
-
p
p
1
1(
)
t then site i would be active with probability
. Increasing
iJ
ij
t
)(
tJj
)(
(decreasing) the E/I ratio was modeled by increasing (decreasing) the average pij in the
range between 0.1/M and 1.5/M (all pij were initially drawn from a uniform distribution on
[0,1] and then all were scaled down by dividing by a constant). Population events were
modeled by activating a single initial site (like an electrical shock applied at a single
electrode or a spontaneous activation in the experiments) and recording the resulting
activations that propagated through the network. These dynamics were defined by
],)(
[
]
[
(cid:213)
-Q=
Q=+
t
)1
1
tJj
)(
where si(t) was the state of site i at time t, Θ[x]=0 (1) for x<0 (>0), and ζi(t) was a
random number drawn from a uniform distribution on [0,1] at each update of each site.
The states of all sites were updated simultaneously at each time step. Each population
event in the model was represented with a 16 bit binary pattern (1= the site was active at
least once during the response to the stimulus, otherwise 0). By simulating 1000
population events (always initiated at the same site), we generated a set of patterns for
which the entropy was computed. From the event size distribution of network events we
ts
(
i
z
i
z
i
t
)(
p
ij
)
-
1(
-
p
iJ
t
)(
-
(8)
7
computed κ (see Shew et al., 2009). The event size was defined as the sum of all
activations from all sites during the population event. The range of average pij values
studied with the model resulted in a range of κ=0.6 to 1.6.
Statistical Analysis. For determining the statistical significance of differences in entropy
for different drug conditions and differences in κ for different drug conditions, we first
used a one-way ANOVA to establish that at least one drug category was different from at
least one other. Next we performed a post-hoc test of significant pairwise differences
between the drug categories using a t-test with the Bonferroni correction for multiple
comparisons. The same procedure was used to assess significance of differences in H and
MI for different categories of κ.
RESULTS
In all of our experiments, multi-electrode array (MEA) recordings of the local field
potential (LFP, Fig. 1A) were used to obtain patterns of cortical population activity. We
defined a recording site as ‘active’ if it presented a large, negative deflection in the LFP
(Fig. 1A, green). We have previously demonstrated that such negative LFP deflections
correlate with increased firing rates of the local neuronal population for each of the
experimental preparations studied here: superficial layers of organotypic cultures (Shew et
al., 2009), urethane anesthetized rat (Gireesh and Plenz, 2008) and awake monkeys
(Petermann et al., 2009). We define a ‘population event’ as a set of electrodes which
were active together within a short time (Methods). In our analysis, each population
event was represented by a binary spatial pattern with one bit per recording site and 1 or 0
indicating an active or inactive site respectively (Fig. 1B; top). For each one hour
recording in vitro (n=47) or 30 min recording in vivo (n=4 monkey, n=6 rat), we typically
observed 103 to 104 population events.
First, we systematically explored a range of E/I conditions in cortex slice cultures. A
reduced E/I ratio was obtained by bath-application of antagonists of AMPA- and NMDA-
glutamate receptor mediated synaptic transmission (DNQX, 0.5-1 μM; AP5, 10-20 μM).
This resulted in population events that were typically small in spatial extent (Figs. 1B and
1C, left). Conversely, an increased E/I ratio was obtained with an antagonist of fast
GABAA-receptor mediated synaptic inhibition (Picrotoxin; PTX, 5 μM), which led to
stereotyped, spatially extended population events (Figs. 1B and 1C, right). In contrast,
unperturbed E/I (Figs. 1B and 1C, middle) typically yielded a diverse pattern repertoire.
The raster plots in Figure 1B (bottom) display examples of 100 consecutive population
events recorded under the three different E/I conditions. Figure 1C displays example
probability distributions of population event sizes for the three E/I conditions. We
performed 11 recordings with reduced AP5/DNQX, 26 with no drug, and 9 with PTX.
For each recording, we measured both ongoing activity and stimulus-evoked activity. For
all recordings, we assessed the information capacity by computing the Shannon entropy of
the full set of recorded binary patterns (Shannon, 1948; Rieke et al., 1997; Dayan and
Abbott, 2001; see also Methods).
8
Figure 1: Measuring the neural
activation pattern repertoire for a
range of E/I conditions. A Example
LFP recordings under conditions of
reduced E (left), unperturbed E/I
(middle), and suppressed reduced I
(right). Scale bars: 250 ms x 10 μV
(left, middle) and 250 ms x 100 μV
(right). Population events were
defined based on large negative
deflections (<-4SD, green). B (top)
Single examples of population events
represented as binary patterns:
1=active site, 0=inactive. (bottom)
Rasters including 100 consecutive
population events represented as
binary patterns; each row represents
one event, each column represents
one recording site. Left: reduced E.
Middle: unperturbed. Right: reduced
I. C Shape of event size distributions
reveal changes in E/I, which are
quantified with κ (see methods;
broken line: power law with exponent
of -1.5).
Peak information capacity of ongoing
activity for intermediate E/I and
neuronal avalanches.
Our first finding was that the entropy H
for ongoing activity peaks at an
intermediate E/I ratio. This was
demonstrated with two different
approaches. First, we compared entropy
to the three pharmacological categories:
AP5/DNQX, no drug, and PTX. We
found that under the unperturbed E/I
condition the average H was significantly
higher than either the reduced E/I
condition of the AP5/DNQX or the
increased E/I condition of PTX (Fig. 2A,
ANOVA, p<0.05). Second, we compared
entropy to a previously developed
statistical measure called κ, which characterizes E/I based on population dynamics of the
9
network (Shew et al., 2009; Methods). An advantage over the three pharmacology
categories is that κ is a graded measure, thus providing a continuous function of entropy
H versus E/I. A detailed definition of κ is given in Methods. Briefly, κ quantifies the
shape of the population event size distribution, which is sensitive to changes in E/I (Fig.
1C): κ <1 indicated reduced E/I and κ >1 indicated increased E/I (Fig. 2B). Indeed, κ
was significantly different for the two pharmacological manipulations compared to the no-
drug condition (Fig. 2B; p <0.05). When we plotted entropy versus κ (Fig. 2C), we
discovered a peaked function with maximum entropy occurring for κ≈1. This confirms
our finding of peak entropy for the no drug condition (Fig. 2A) and provides a more
refined view of the data; the peak occurred at κ*=1.16–0.12 (mean–SD, uncertainty
determined by rebinning the experimental data, see Supplementary Fig. S2). The
statistical significance of the peak in H was confirmed by comparing H for the ten
experiments with κ closest to 1 with the ten experiments with smallest κ and ten with
largest κ (p<0.05).
In addition to providing a graded measure of E/I, κ assesses the statistical character of
ongoing cortical population dynamics. Specifically, κ≈1 is the signature of neuronal
avalanches (Shew et al., 2009), a type of population dynamics defined by a power-law
population event size distribution with a power-law exponent near -1.5 (Beggs and Plenz,
2003; Stewart and Plenz, 2006; Gireesh and Plenz, 2008; Petermann et al., 2009; Shew et
al., 2009). The computation of κ entails first computing the difference between a
measured event size distribution and a theoretical reference distribution defined as a
power-law with exponent -1.5 (Fig. 1C, green dashed). Next, this difference is added to 1
(for historical reasons, Shew et al., 2009), resulting in κ=1 for an exact match with a -1.5
power law, i.e. neuronal avalanches. In this context, our findings indicate that entropy is
maximized under conditions which result in neuronal avalanches.
Next we tested the robustness of the peak in H with respect to changes in spatial and
temporal scales of recordings. First, as shown in Figure 2D (green), we found that the
peak in H remained close to κ = 1 (κ*=1.01–0.02), even when the original 8x8 patterns
were coarse-grained to obtain 4x4 patterns at half the spatial resolution (see Methods).
Second, the peak was also maintained when the spatial extent of the recorded area was
reduced by 75% (4x4 electrodes near center of the MEA; Fig. 2D). Finally, we confirmed
that the peak persisted for a restricted recording duration of 12 minutes rather than one
hour (Fig. 2D, purple). The robustness of our finding to shorter recording durations is
important since estimations of entropy depend on the number of samples recorded (see
below).
10
Figure 2: Ongoing activity - peak
information capacity at
intermediate E/I ratio specified by
κ≈1. A Information capacity (entropy
H) of the pattern repertoire is
maximized for when no drugs perturb
the E/I ratio. Significant differences
marked with * (p<0.05). Box plot
lines indicate lower quartile, median,
upper quartile; whiskers indicate
range of data, excluding outliers (+,
>1.5 times the interquartile range). B
The statistic κ, provides a graded
measure of E/I condition based on
network dynamics (methods). C
Entropy H peaks near κ≈1. Each point
represents one recording of ongoing
activity (n=47, 8x8 MEA, 1 hr, color
indicates drug condition; red=PTX,
blue=AP5/DNQX, black=no drug).
Line is the binned average of points.
D The peak in entropy H is robust to
changes in spatial resolution (green,
4x4 coarse-binned, 1 hr), spatial
extent (orange, 4x4 subregion, 1 hr)
and duration (purple, 4x4 coarse-
binned, 12 min) of recording. (black,
same data as in A). Error bars
indicate mean±s.e.m.
Peak information transmission between stimulus and response for intermediate E/I
and neuronal avalanches.
We now present measurements of stimulus-evoked activation patterns. A priori, one can
expect a different distribution of stimulus-evoked patterns compared to ongoing activity
and thus different entropy. Indeed studies suggest that ongoing activity is more diverse
than typical stimulus-evoked activity (Welicky, 2004; Luczak et al., 2009; Churchland,
2010). However, if the entropy of evoked patterns changes with E/I in the same way that
we found for ongoing activity, then evoked entropy may also peak near κ=1. This
possibility is in line with significant evidence that ongoing activity in the cortex is
intimately related to stimulus-evoked activity (Kenet et al., 2003; Ji and Wilson, 2007;
11
Han et al., 2008; Luczak et al., 2009). For instance, stimulus-evoked activity patterns
recur during ongoing activity, both at the population level (Kenet et al., 2003; Han et al.,
2008) and the level of spike sequences (Ji and Wilson, 2007). Therefore, our next aim
was to test whether our finding of peak entropy near κ=1 also holds for stimulus-evoked
activity.
Stimuli consisted of 10 different amplitude single bipolar shocks each applied 40 times in
randomized order though a single electrode of the MEA within cortical layers II/III
(Methods). A binary pattern was constructed to represent each response during the 20 –
500 ms after the stimulus. The evoked entropy H was calculated for the set of 400
stimulus-evoked activation patterns for each E/I. As found for ongoing activity, the
evoked entropy was highest near κ ≈1 for both fine and coarse spatial resolution (Fig. 3A;
black - 8x8, green - coarse-grained 4x4, p<0.05).
In the introduction, we gave a simple example in which information transmission from
input to output was limited due to low entropy. With our measurements of network
responses (i.e. output) to stimuli (i.e. input), we can directly test whether efficacy of
information transmission is optimized when entropy is maximized. This idea is concisely
summarized in the following equation:
MI(S;R) = H(R) - H(RS). Here, MI(S;R) is
the mutual information of stimulus and
response which quantifies the information
transmission (Rieke et al., 1997; Dayan and
Abbott, 2001). H(R) is the entropy of the full
set of response patterns for all stimuli, while
H(RS) is the conditional entropy, i.e. the
average entropy per stimulus (Rieke et al.,
1997; Dayan and Abbott, 2001). As shown
above, H(R) is maximized near κ ≈1. Since,
H(RS) is always positive, MI(S;R) is
bounded by H(R), and thus potentially also
peaks near κ =1. Indeed, we measured
MI(S;R) under different E/I conditions and
found that stimulus-response mutual
information was maximized near κ ≈1 (Fig.
3B; black - 8x8, green - coarse-grained 4x4,
p<0.05).
Figure 3: Stimulus-evoked activity -
peak information transmission at
intermediate E/I ratio specified by
κ≈1. A Single shock stimuli with 10
different amplitudes (10-200 μA) were
applied 40 times each using a single
electrode. The pattern repertoire of
stimulus-evoked activity has maximum
12
entropy near κ≈1. This holds for 8x8 response patterns (black line) as well as
coarse resolution 4x4 patterns (green line). Points correspond to 8x8 patterns:
light blue – AP5/DNQX, gray – no drug, pink – PTX. B The efficacy of
information transfer, i.e. mutual information of stimulus and response, also peaks
near κ≈1. The dashed line indicates the highest possible mutual information
given 10 stimulus levels. (black - 8x8; green – 4x4). Error bars indicate s.e.m.
Competition between activity rates and interactions explains peak in entropy.
To identify and quantify the mechanisms leading to the peak in entropy near k=1, we
analyzed in more detail the coarse-grained 4x4 patterns measured during ongoing activity
(Fig. 2D, green). A priori, the total number of unique patterns that are possible is 216,
implying a maximum H≤log2(216)=16 bits. This maximum would be reached if all 216
patterns occurred with equal probability. However, during a 1 hr recording, the network
did not generate all possible patterns, nor were different patterns equally likely, resulting in
H that was always below 16 bits. The peak in H was explained by three main factors that
changed with the E/I ratio: i) the number N of patterns observed during the recording, ii)
the likelihood L that sites participate in patterns, and iii) the strength of interactions
between sites. The first two effects are related to the rates of observed activity and
impose upper bounds on H: effect i requires H ≤ log2(N) (dash-dot line in Figs. 4A) and
effect ii limits H in a way that depends on L (dashed line Figs. 4A). Specifically, the
highest possible entropy for a given L can be computed by assuming that sites are
independent,
H
(
(cid:229)
-< M
=
i
1
L
i
log
2
L
i
-+
1(
L
i
log)
2
-
1(
L
i
)
)
,
(9)
where M is the number of recording sites and Li is the likelihood of participation for site i.
This formula is based on the fact that the entropy of two independent systems combined is
the sum of their individual entropies. Since a single site i is either active (with probability
-
--
-
Li) or inactive (with probability 1-Li), its entropy is
L
L
L
L
.
log
log)
1(
)
1(
i
i
i
i
2
2
Thus, adding the entropy of all sites, we obtain the formula above. When L<1/2,
increasing L increases the upper bound on H. When L>1/2, increasing L decreases the
upper bound on H. We found that L increased over the range of E/I conditions we studied
(Figs. 4C), while the number of patterns N did not show a systematic trend.
We turn now to effect iii. Increased interactions between sites always reduce H due to the
increased redundancy of the information at different sites (Schneidman et al., 2003). We
found that site-to-site interactions during ongoing activity increased with E/I (Figs. 4E),
and quantified this trend in two ways. First we computed mutual information (MI)
between the activity recorded from different pairs of sites (Figs. 4E; red). Note that above
we used mutual information in a different way, computed between stimulus and response
MI(R;S) to assess information transmission. Second, we estimated the effect of
interactions by computing the drop in entropy resulting from shuffling the data. The
shuffling procedure destroyed interactions by randomizing the set of population events in
13
which each site participated, while
keeping L and N fixed (Methods).
The entropy of the shuffled data for
the corresponding original κ value is
shown in figure 4A (black) and, as
expected, nearly reached the bounds
set by the combined effects i and ii.
The difference in entropy DH
between the measured and shuffled
data is due to interactions (Figs. 4E,
blue). DH has previously been used
to quantify redundancy (Dayan and
Abbott, 2001).
In summary, at low E/I, effects ii
and iii compete and effect ii wins,
i.e. activity rates drop sufficiently
low to cause low entropy even
though interactions are also low.
At high E/I, effects ii and iii
cooperate, i.e. both high activity
rates and strong interactions cause
low entropy. Entropy peaked at an
intermediate E/I ratio at which
interactions between sites were not
too low or too high (specified by
MI≈0.2) and activity was not too
depressed (L≈0.25).
Figure 4: Peak information capacity explained. A detailed analysis of in vitro
experimental results (left, Fig. 2b, green) and model results (right, Fig. 2c, blue)
A,B Upper bounds on entropy are set by 1) the average likelihood that sites
participate in patterns (dashed) and 2) the number of patterns observed (dash-
dot). When the effects of interactions are removed by shuffling (methods), the
entropy reaches these bounds (black), but the measured entropy (green) is
always lower due to interactions. C,D Rise in participation likelihood L as E/I
ratio is increased. This rise accounts for the bounds (dashed) shown in A,B. E,F
Rise in interactions between sites (mutual information, red) is proportional to the
loss in information capacity ΔH (blue). All error bars indicate s.e.m.
We remark that, if N were large enough (e.g. for longer recording duration), the upper
bound due to effect i would become irrelevant, in which case, we still expect H to peak
near κ=1 due to the combined effects of interactions (ΔH) and L. Nonetheless, the
persistence of the peak in H for shorter duration recordings may be more relevant for
14
cortex operations which occur on shorter time scales. We also tested the extent to which
our measurements are impacted by sample size following the methods developed by Magri
et al. (2009). The difference between our measured H and ‘corrected’ H was 0.06–0.06
(mean±SD) bits for 4x4 in vitro ongoing activity patterns and 0.22–0.18 bits for the 8x8
patterns. Thus, sample size effects are small compared to the variability from one
experiment to another (see error bars in Fig. 2.) We also point out that N, L, and MI are
not the only factors that could potentially influence H. For example, not every site was
equally likely to be active. Such spatial structure is expected to decrease entropy
compared to a spatially homogeneous system with all other properties held fixed. This
was not a major influence in our results.
Experimental results confirmed in a computational network-level model.
To gain further insight on our experiments, we compared our results to a network-level
simulation, which has been used previously to model neuronal avalanches (Haldeman and
Beggs, 2005; Kinouchi and Copelli, 2006; Shew et al., 2009). The model consisted of 16
binary sites. The state (1=active, 0=inactive) of each site was intended to represent a
population of neurons in the vicinity of a recording electrode (Methods). The propagation
of activity from one site to another was treated probabilistically; a connection matrix p
with entries pij specified the probability that site i would become activated due to site j
having been activated in the previous time step. Increases (decreases) in E/I were
modeled by increasing (decreasing) the average pij value through the range 0.006 to 0.1.
For each ‘E/I condition’, 1000 population events were simulated, beginning with a single
initially active site and the resulting patterns of activity were recorded. To facilitate
comparison with our experimental results we also parameterized each E/I condition of the
model using k, based on population event size distributions.
In good agreement with our experiments, we found that entropy reached a peak for k≈1
(Fig. 4B; green). Moreover, the explanation of peak entropy in terms of the competition
between activity rates and site-to- site interactions also held for the model. Just as in the
experiments, when the model data was shuffled to remove effects due to interactions, H
(Fig. 4B; black) approached the upper bounds set by the number of events (Fig. 4B; dash-
dot) and the likelihoods of participation (Fig. 4B; dashed). The model H results matched
the experimental values, because the underlying changes in L versus κ (Fig. 4D) and the
changes in site-to-site MI versus κ (Fig. 4F, red) were very similar to those measured
experimentally. This agreement is not trivial; the same values of entropy could in principle
be reached with different combinations of the underlying L and MI versus κ. For example,
a peak in H could result if L remained fixed at 0.5 and interactions were minimized at κ=1.
Site-to-site mutual information in the model reached slightly lower levels for high κ when
compared to experiments(Fig. 4F, red), which could be due to the lack of significant
structure in the model connectivity matrix p.
15
In vivo entropy matches in vitro
prediction.
Finally, we analyzed recordings of ongoing
activity from superficial cortical layers in
two awake monkeys (premotor cortex)
not engaged in any particular task and in
urethane-anesthetized rats (n=6, barrel
cortex) with no whisker stimulation. In
agreement with previous studies (Gireesh
and Plenz, 2008; Petermann et al., 2009),
we found that the ongoing activity was
organized as neuronal avalanches (Fig.
5A). More precisely, we found that
κ=1.02–0.02 for the monkeys and
κ=1.08–0.02 for the rats. Based on our in
vitro findings, these κ values suggest that
the in vivo networks are operating under
E/I conditions that maximize entropy and
information transmission. Although we
cannot fully test this idea without a full
range of κ in vivo, we can test whether the
in vivo values of H, L, and MI match with
those predicted from the in vitro results.
As shown in Figure 5B and summarized in
Table 1, we found good agreement with
these predictions. We found no
statistically significant difference between
the in vivo results and the prediction from
in vitro experiments with the same range
of κ (1.0<κ<1.1, p<0.05). Nonetheless,
the fact that entropy values in vivo were
slightly higher than the in vitro results,
may be due to the corresponding slightly
lower MI values.
Figure 5: In vivo properties
predicted from in vitro results. A
Population event size distributions
from ongoing activity in two awake
monkeys (blue) and an example rat
(green) are near a power-law with
exponent -1.5 (dashed line), i.e. they
exhibit neuronal avalanches and κ≈1.
B In line with in vitro and model
predictions for κ≈1, in vivo entropy
16
was high and mutual information between recording sites was moderate. (stars -
two recordings on different days from each monkey; squares - anesthetized rats,
n=6). The spatial extent of recorded area was approximately matched. C The
result holds even when the spatial scales and resolution differ by factor of 4.
We note that the in vivo values of MI, which are based on LFP measurements, coexist
with low values of pair-wise correlation r between spiking activity of units (mean±s.e.m.
r=0.03±0.01, Supplementary Fig. S3), in line with recent reports for awake monkeys
(Ecker et al., 2010) and anesthetized rats (Renart et al., 2010). The success of our
prediction requires matching the number of recording sites (16 here), but is robust to large
changes in spatial extent and resolution of recordings (Fig 5C). The prediction is also
robust to changes in the threshold used for generating binary activity patterns from
continuous LFP data (Supplementary Table S1).
Entropy, H
(bits)
Participation
likelihood, L
Site-to-site mutual
information, MI (bits)
In vitro predictions for
1.0<κ<1.1
Awake monkeys
κ=1.02–0.02
Anesthetized rats
κ=1.08–0.02
5.7±1.6
7.5–0.5
7.1–1.2
0.3±0.1
0.3–0.03
0.4–0.1
0.2±0.2
0.1–0.01
0.2–0.1
Table 1: In vivo results match in vitro predictions. Given the range of κ found in
the in vivo recordings (1<κ<1.1), our in vitro results provide the predictions of H,
L, and MI shown in the first row. The corresponding measurements from the
awake monkeys (second row) and anesthetized rats (third row) match the in vitro
predictions, i.e. they are not significantly different (p<0.05). Corresponding data
are shown in Fig. 5B. All numbers are mean±SD.
DISCUSSION
We employed in vitro and in vivo experiments as well as a computational model to study
the effects of the E/I ratio on entropy and information transmission in cortical networks.
We analyzed multisite measurements of LFP recorded during ongoing as well as stimulus-
evoked activity. We found that entropy and information transmission are maximized for
the particular E/I ratio specified by κ=1, which is the same E/I condition under which
neuronal avalanches emerge.
We emphasize that the relative changes in H as we altered E/I are the meaningful results
of our in vitro study; the absolute entropy values in bits depend upon arbitrary aspects of
the analysis and measurements, e.g. the number of electrodes in the MEA. Thus, we are
not suggesting that there is an absolute cap on the information that a cortical circuit can
17
represent at ~10 bits and it is not appropriate to compare our H values to those found in
other studies of population entropy measures (e.g. Quian Quiroga and Panzeri, 2009).
The important feature of our result is the peak in H near κ ≈1. We expect that any
measure of population entropy would also peak for the same intermediate E/I, specified by
κ ≈1.
Previous studies have separately addressed the topics of entropy maximization (Laughlin,
1981; Dong and Atick, 1995; Dan et al., 1996; Li, 1996; Rieke et al., 1997; Dayan and
Abbott, 2001; Garrigan et al., 2010), neuronal avalanches (Beggs and Plenz, 2003;
Haldeman and Beggs, 2005; Stewart and Plenz, 2006; Ramo et al., 2007; Gireesh and
Plenz, 2008; Tanaka et al., 2009; Petermann et al., 2009; Shew et al., 2009), and the
balance of E/I (van Vreeswijk and Sompolinsky, 1996; Shadlen and Newsome, 1998; Shu
et al., 2003; Okun and Lampl, 2008; Susillo and Abbott, 2009; Roudi and Latham, 2007),
but our work is the first to show how these ideas converge in cortical dynamics.
Significant evidence suggests that maximization of entropy is an organizing principle of
neural information processing systems. For example, single neurons in the blowfly visual
system have been shown to exhibit spike trains with maximized entropy, considering the
stimuli the fly encounters naturally (Laughlin, 1981). Applied at the level of neural
populations, the principle of maximized entropy has provided successful predictions of
receptive field properties in mammalian retina (Garrigan et al., 2010), lateral geniculate
nucleus (Dong and Atick, 1995; Dan et al., 1996), and visual cortex (Li, 1996). Our work
shows that the potential ability of a neural population in the cortex to achieve maximum
entropy and maximum information transmission depends on the E/I ratio. Thus, if such
properties are optimal for the organism, then the particular E/I ratio specified by κ=1 may
best facilitate this goal.
We note that our investigation is not directly related to ‘maximum entropy’ models (e.g.
Schneidman et al., 2006). In those studies, the aim was to use the maximum entropy
principle (Jaynes, 1957) to find the simplest model to describe an experimental data set;
entropy served as a modeling constraint. In contrast, here we compare the entropy across
different experiments, searching for conditions which result in maximum entropy; entropy
measurements are the results.
Several theory and modeling studies (including our own model) offer a deeper explanation
of why κ=1 and neuronal avalanches occurs under E/I conditions which maximize entropy
and information transmission (Beggs and Plenz, 2003; Haldeman and Beggs, 2005; Ramo
et al., 2007; Tanaka et al., 2009). Recall that neuronal avalanches and κ=1, by definition,
indicate a power-law event size distribution with exponent -3/2. This same property is
found in many dynamical systems that operate near ‘criticality’. Criticality refers to a
particular mode of operation balanced at the boundary between order and disorder (e.g.
Stanley, 1971; Jensen, 1998), akin to the balance of excitation and inhibition that we
explore in our experiments. In our model, criticality occurs when the average pij equals
1/M (M is the number of sites). When pc>1/M, activity propagation is widespread and
highly synchronous, like a seizure, while pc<1/M results in weakly interacting, mostly
independent neurons (Beggs and Plenz, 2003; Haldeman and Beggs, 2005; Kinouchi and
Copelli, 2006). The balanced propagation that occurs at criticality might be attributed to
18
interactions between excitatory and inhibitory neurons in the cortex. Using theory of
Boolean networks, Ramo et al. (2007) showed theoretically that entropy of the event size
distribution is maximized at criticality. Simulations of a model similar to our own showed
that the number of activation patterns that repeat is maximized at criticality (Haldeman
and Beggs, 2005). Tanaka et al. (2009) found that recurrent network models in which
information transmission is optimized also exhibit neuronal avalanches and repeating
activation patterns. Likewise, it has been shown that mutual information of input and
output in feed-forward network models is maximized near criticality (Beggs and Plenz,
2003) and mutual information between neurons in the same network is maximized at
criticality (Greenfield and Lecar, 2001). In line with these theory and model predictions,
our results are the first experimental demonstration of peak entropy and information
transmission in relation to criticality in the cortex.
Finally, a separate line of research has focused on the E/I ratio in cortical networks.
Models emphasize the importance of balanced E/I for explaining the variability observed in
spike trains (van Vreeswijk and Sompolinsky, 1996; Shadlen and Newsome, 1998), low
correlations between spiking units (Renart et al., 2010), and generating diverse population
activity patterns (Susillo and Abbott, 2009), which may play a role in memory (Roudi and
Latham, 2007). Moreover, in vivo experiments have shown that synaptic input received
by cortical neurons exhibits a fixed ratio of excitatory to inhibitory current amplitudes
(Shu et al., 2003; Okun and Lampl, 2008). Since we measure κ≈1 in vivo, it follows that
the ‘balanced E/I’ discussed in these previous studies may also correspond to the optimal
E/I that we identify here.
In summary, our results suggest that by operating at the E/I ratio specified by κ ≈1, the
cortex maintains a moderate level of network-level activity and interactions which
maximizes information capacity and transmission. This finding supports the hypotheses
that balanced E/I and criticality optimize information processing in the cortex.
REFERENCES
Beggs JM, Plenz D (2003) Neuronal avalanches in neocortical circuits J Neurosci
23:11167-11177.
Churchland MM et al. (2010) Stimulus onset quenches neural variability: a widespread
cortical phenomenon. Nat Neurosci 13, 369-378.
Dan Y, Atick JJ, Reid RC (1996) Efficient coding of natural scenes in the lateral
geniculate nucleus: experimental test of a computational theory. J Neurosci 10:3351-3362.
Dayan P, Abbott LF (2001) Theoretical neuroscience (MIT Press, Cambridge,
Massachussetts, USA).
Dichter M, Ayala G (1987) Cellular mechanisms of epilepsy: a status report. Science
237:157-164.
Dong DW, Atick JJ (1995) Temporal decorrelation: a theory of lagged and nonlagged
19
responses in the lateral geniculate nucleus. Network: Computation in Neural Systems
6:159-178.
Ecker AS et al. (2010) Decorrelated neuronal firing in cortical microcircuits. Science
327:584-587.
Fiser J, Chiu C, Weliky M (2004) Small modulation of ongoing cortical dynamics by
sensory input during natural vision. Nature 431, 573-578.
Garrigan P et al. (2010) Design of a trichromatic cone array. PLoS Comput Biol
6:e1000677.
Gireesh ED, Plenz D (2008) Neuronal avalanches organize as nested theta- and
beta/gamma-oscillations during development of cortical layer 2/3. Proc Nat Acad Sci USA
105:7576-7581.
Greenfield, E., and Lecar, H. (2001). Mutual information in a dilute, asymmetric neural
network model. Physical Review E 63, 1-10.
Haldeman C, Beggs JM (2005) Critical branching captures activity in living neural
networks and maximizes the number of metastable states. Phys Rev Lett 94:058101.
Han F, Caporale N, Dan Y (2008) Reverberation of recent visual experience in
spontaneous cortical waves. Neuron 60:321-327.
Ji D, Wilson MA (2007) Coordinated memory replay in the visual cortex and
hippocampus during sleep. Nat Neurosci 10:100-107.
Jacobs et al. (2009) Ruling out and ruling in neural codes. Proc. Nat. Acad. Sci. USA
106:5936-5941.
Jaynes ET (1957) Information theory and statistical mechanics. Phys Rev 106:62–79.
Jensen HJ (1998) Self-organized criticality: emergent complex behavior in physical and
biological systems. (Cambridge University Press, Cambridge, UK).
Kenet T, Bibitchkov D, Tsodyks M, Grinvald A, Arieli A (2003) Spontaneously emerging
cortical representations of visual attributes. Nature 425:954–956.
Laughlin S (1981) A simple coding procedure enhances a neuron’s information capacity.
Zeitschrift fur Naturforschung 36:910-912.
Li Z (1996) A theory of the visual motion coding in the primary visual cortex. Neural
Comput 8:705-730.
Luczak A, Barthó P, Harris KD (2009) Spontaneous events outline the realm of possible
sensory responses in neocortical populations. Neuron 62:413-425.
Magri C, Whittingstall K, Singh V, Logothetis NK, Panzeri S (2009) A toolbox for the
fast information analysis of multiple-site LFP, EEG and spike train recordings. BMC
20
Neurosci. 10:81.
Nauhaus I, Busse L, Carandini M, Ringach DL (2009) Stimulus contrast modulates
functional connectivity in visual cortex. Nat Neurosci 12:70-76.
Okun M, Lampl I (2008) Instantaneous correlation of excitation and inhibition during
ongoing and sensory-evoked activities. Nat Neurosci 11:535-537.
Petermann T et al. (2009) Spontaneous cortical activity in awake monkeys composed of
neuronal avalanches. Proc Nat Acad Sci USA 106:15921-15926.
Pola G, Thiele A, Hoffmann K-P, Panzeri S (2003) An exact method to quantify the
information transmitted by different mechanisms of correlational coding. Network:
Comput. Neur. Syst. 14:35-60.
Quian Quiroga R, Panzeri S (2009) Extracting information from neuronal populations:
information theory and decoding approaches. Nat Rev Neurosci 10:173-185.
Ramo P, Kauffman S, Kesselia J, Yli-Harja O (2007) Measures for information
propagation in Boolean networks. Physica D 227:100-104.
Renart A et al. (2010) The asynchronous state in cortical circuits. Science 327:587-590.
Rieke F, Warland D, de Ruyter van Stevenick R, & Bialek W (1997) Spikes (MIT Press,
Cambridge, Massachussetts, USA).
Roudi Y, Latham PE (2007) A balanced memory network. PLoS Comp Biol 3:1679-1700.
Schneidman E, Berry II MJ, Segev R, Bialek W (2006) Weak pairwise correlations imply
strongly correlated network states in a neural population. Nature 440:1007-1012.
Schneidman E, Bialek W, Berry II MJ (2003) Synergy, redundancy, and independence in
population codes. J Neurosci 23:11539 –11553.
Shadlen MN, Newsome WT (1998) The variable discharge of cortical neurons:
implications for connectivity, computation, and information coding. J Neurosci 18:3870-
96.
Shannon CE (1948) A mathematical theory of communication. Bell System Technical J
27:379-423, 623–656.
Shew WL, Yang H, Petermann T, Roy R, Plenz D (2009) Neuronal avalanches imply
maximum dynamic range in cortical networks at criticality. J Neurosci 29:15595-15600.
Shu Y, Hasenstaub A, McCormick DA (2003) Turning on and off recurrent balanced
cortical activity. Nature 423:288-293.
Stanley HE (1971) Introduction to Phase Transitions and Critical Phenomena (Oxford
21
University Press, New York, USA).
Stewart C, Plenz D (2006) Inverted-U profile of dopamine-NMDA-mediated spontaneous
avalanche recurrence in superficial layers of rat prefrontal cortex. J Neurosci 26:8148-
8159.
Sussillo D, Abbott LF (2009) Generating coherent patterns of activity from chaotic neural
networks. Neuron 27:544-557.
Tanaka T, Kaneko T, Aoyagi T (2009) Recurrent infomax generates cell assemblies,
neuronal avalanches, and simple cell-like selectivity. Neural Comput 21:1038-1067.
van Vreeswijk C, Sompolinsky H (1996) Chaos in neuronal networks with balanced
excitatory and inhibitory activity. Science 274:1724-1726.
Supplementary Material for 'Information capacity and transmission are maximized
in balanced cortical networks with neuronal avalanches' by Shew et al.
Relationship between mutual information and correlation coefficient
To quantify interactions between sites we used average pair-wise mutual information
(MI). A more traditional approach is to use average pair-wise correlation coefficients
(CC). Our reason for working with mutual information is two-fold. First, MI is less
sensitive to noise when interactions are very weak. Second, MI arises from information
theory and, thus, is a more natural fit with the study of entropy. Nonetheless, MI and CC
are closely related. In Fig S1 we compare both quantities for the in vitro data presented in
the main text. Also marked in Fig S1 are estimated theoretical bounds on the relationship
between MI and CC. The lower bound is reached for L=0.5, while the upper bound
corresponds to the extreme values of L=0 and L=1.
22
Figure S1 Relationship between mutual information (MI) and correlation coefficient
(CC). The black points compare MI and CC for all in vitro experiments. The dashed line
is an estimated lower bound, which is reached for L=0.5. The upper line is an estimated
upper bound, which is reached for very low or very high L. The estimated bounds were
obtained numerically for a single pair of binary vectors (10,000 events), each with the
same L (0.1 to 0.9) and CC (0.1 to 0.9). These bounds are only approximate, because in
the experiments, L is not the same from one electrode to another.
Binning sensitivity for H vs. k.
Here we show that the results shown in Fig 2C are robust to different choices of the bins
used to produce the average line. The variability of the peaks of these curves was used to
estimate the uncertainty in the conclusion that peak entropy occurs at k≈1.
Figure S2 Results robust to bin choices. The results shown in Fig 2C (black, green) for
H vs. k for 8x8 patterns (top) and coarse-binned 4x4 patterns (bottom) were recomputed
with different averaging bins. The different bin partitions are shown below the curves with
23
corresponding colors.
Pair-wise spike train cross correlation in the monkey
Here we report the spike count cross correlation values between unit activity recorded
simultaneously with the monkey LFP recordings discussed in the main text. The average
correlation between unit signals is significantly lower than that between the population
signals provided by the LFP. As shown in Fig S2, the mean, s.e.m., and distributions of
correlation coefficients were in good agreement with recent reports from awake monkeys
(Ecker et al., 2010) and anesthetized rats (Renart et al., 2010).
Spike sorting was performed with Plexon offline spike sorter (V2.8.8). 66 and 40 well
isolated units were found for monkey 1 and monkey 2 respectively. Three principal
components (PCA), peak-trough amplitude, and nonlinear energy were used as the sorting
features. We defined ‘well isolated’ as follows: in a 2-D projection of at least 2 of the
sorting features the unit must have a mean which is strongly different from the mean of
noise waveforms (p≤0.001, multivariate ANOVA). If more than one unit was recorded
from the same electrode, the difference between means of each unit was also required to
be significant at this strict level (p≤0.001, multivariate ANOVA).
To compute spike count cross correlations between each pair of units recorded during
ongoing activity we followed established methods (Renart et al., 2010). First, to obtain
spike count vectors, the spike time stamps of each unit were 1) binned with 1 ms temporal
resolution, 2) convolved with a Gaussian window with 50 ms width. The cross
correlation coefficient was computed between all pairs (2145 pairs for monkey 1, 780
pairs for monkey 2) of spike count vectors.
Figure S3 Histograms of pairwise correlation coefficients of unit activity in the
awake monkeys. As reported previously (Ecker et al., 2010; Renart et al., 2010) the
average across all pairs is near zero and positive. The histogram of all pairwise CC values
for monkey 1 (left) and monkey 2 (right) are shown. The mean±sem CC values were
0.050±0.002 and 0.015±0.001 for monkey 1 and monkey 2 respectively.
Robustness to event detection threshold
In line with previous studies, our in vivo monkey results were robust to changes in the
detection threshold for nLFPs. For thresholds -2.5, -3 and -3.5 SD we found no
significant changes as shown in Table S1 below (mean±s.e.m.).
24
Monkey 1
Day 1
Day 2
Monkey 2
Day1
Day 2
κ
H
MI
1.02±0.02
1.07±0.03
7.96±0.35
7.58±0.32
0.11±0.01
0.14±0.02
0.99±0.02
1.00±0.02
6.99±0.10
7.51±0.08
0.13±0.01
0.11±0.02
|
1909.05908 | 1 | 1909 | 2019-09-09T18:13:41 | Neurological Nature of Vision and Thought and Mechanisms of Perception Experiences | [
"q-bio.NC"
] | Understanding of the phenomena of vision and thought require clarification of the general mechanism of perception. So far, philosophical inquiries and scientific investigations have not been able to address clearly the mysteries surrounding them. The present work is an attempt to unravel the essences of these phenomenal based on the presumption of computational brain. Within this context, the natures of thought is clarified, and the basis of the experience of perception is established. And by drawing from the successes of the developed tactile vision substitution systems (TVSS), which render some measure of vision,in vision handicapped persons, early or congenital blinds, the true nature of vision as cutaneous sensations is also divulged. The mechanism of perception involves sensing of the stimuli, and autonomous engagement of brain neuronal complexity resolution patterns; that is the brain implicit embedded computational instructions. Upon commencement of the triggers, brain computations, which aso involve engaging body's biophysical feedback system, are performed; and the results are outputted as motor signals that render the realization of perception. However, this requires deployment of a perception medium; an interface. Given the nature of efferent signals, there must be a (known) bio-mechanical system interface, other than the body muscle and skeletal system, which performs the needed function: Considering the fact that the vocal system performs such task for verbalization of brain's synthesis of language, the possibility of its further role in the experience of thought and vision, in the form of mostly quiet (inaudible) recital of the related motor signals, is suggested. | q-bio.NC | q-bio | Journal of Neurology & Stroke
Neurological Nature of Vision and Thought and
Mechanisms of Perception Experiences
Short Communication
Retired LBNL (UCB) Scientist, USA
Volume 4 Issue 5 - 2016
*Corresponding author: Jahan N Schad, PhD., Retired
LBNL (UCB) Scientist, 376 Tharp Drive, Moraga, Ca
94556, California, USA, Tel: 925-376-4126; Email:
Received: March 28, 2016 Published: April 19, 2016
Abstract
Understandings of the phenomena of vision and thought require clarification of
the general mechanism of perception, -the experience prompted by the (brain)
efferent signals -as well as the clarification of the natures of the related afferent
signals, which drive the mechanism. So far, philosophical inquiries and scientific
investigations have not been able to address clearly the mysteries surrounding
them.
The present work is an attempt to unravel the essences of these phenomena
based on the presumption of computational functioning of the brain, a concept
supported by scientific consensus. Within this context, the nature of the thought
is clarified, and the basis of the experience of perception is established. And by
drawing from the successes of the tactile vision substitution system (TVSS) [1],
-- which renders a measure of vision perception in vision handicapped, early or
congenital blinds -- the true nature of the vision, as cutaneous sensation, is also
divulged.
The mechanism of perception --what renders it and where it occurs --involves
sensing of stimuli, and, or, the autonomous engagement of brain inherent
neuronal complexity resolution patterns; the implicit embedded computational
instruction (codes). Upon commencement of such triggers, --of which one may not
be necessarily aware --brain computations, which also involve engaging body's
biophysiological feedback system, are performed; and the results are outputted
as motor (efferent) signals that render perception. However, embedded in the
process of the development of the experience of perception, is the deployment
of a perception medium; an interface. Given the nature of the efferent signals,
there must be a (known) biomechanical system interface, --other than the
irrelevant body muscle and skeletal systems --which performs the needed
function: Considering the fact that the vocal system performs such task for the
verbalization of brain's synthesis of language expressions, the possibility of its
further role in the experiences of thought and vision, in the form of mostly quiet
(inaudible) recital of related signals, is suggested.
Keywords: Vision; Thought; Neuronal computation; Simulation; Utterance
interface; Language; Biolinguistic
Understanding of the phenomenon of Vision, which is to know
"how and where we see what we see," beyond the knowledge
of the biophysiological and optical aspects of the eyes, and the
brain modalities where the trigger signals are processed, has
remained a mystery. Neurosciences' knowledge of the central
nervous system, and the brain neurocomputational concepts,
suggest that brain neuronal code (computational patterns)
processing of eye-extracted afferent data (somehow) renders
vision perception. However, this still falls short of a complete and
convincing addressing of the above question; leaving the vision
concept vague, as it has always been. Further, and very detailed
understanding of the anatomy and physiology of the vision and
the related processes [2], are not likely to provide the answer.
We resolve this ambiguity by bringing to light the nature of the
vision afferent) signals, their processing in the brain, and where
the (brain) efferent vision signals are experienced.
Submit Manuscript http://medcraveonline.com
Hypothesis
Same difficulty for the though, as to "what it is and where
it happens," has also always held true; beyond platitude. The
philosophical addressing of mental processes and scientific
understandings of the brain and its functions has not helped
to resolve the puzzle either. A step toward the resolution of the
thought ambiguity can be found in the biologic theory of linguistics,
which according to Chomsky [3] entails the presence of neuronal
language construct in the brain, and the proposition of two related
interfaces: The first is "the thought system which provides a place
for the interpreted Internal (I) language mechanism synthesis
of the structured expression in the brain;" and the second is the
"vocal system, activated by the motorsensory neurons," which
renders language vocalizations; whether it is referential as in the
calls of animals, or verbalized as in humans. However, despite this
enlightening concept, the dilemma about the overall nature of the
thought and its system still remains.
J Neurol Stroke 2016, 4(5): 00152
2/3
perception is same as cutaneous perception, and that there should
not be any differences in the natures of their afferent signals.
system which includes a pulsating patch on the skin or tongue.
To address these mysteries, we focus on:
Taking note of the fact that such subjects never experience
a) The nature of the brain information processing schemes
vision, neither in waking hours nor in dreams [7], the experience
and the triggers which drives them; and
of such perceptions seemed inexplicable: On the face of it, the
patch should only create cutaneous perceptions. True that from
b) On the nature of the brain outputs and the need for
the patch location, massive and simultaneous amount of data
biological interfaces.
pulses are sent to the brain, however, this should only lead to
functioning of the
The presumption of computational
the development of some matter (object) perception, likely that
brain is based on the vast body of computational sciences and
of the patch itself. However, the repeated experimentally verified
neurosciences findings, in tandem with experimental works in
phenomenon can be explained in the context of brain neuronal
the area of information processing of the neurons [4]; and the
net computational procedure which can engage any available
successes of the brain inspired scientific neural networks in
neural circuitry (brain modality), including those of vision, for
developing some measure of human-like intelligence. Extending
the processing of large amount of afferent data, such as those
the general workings principal of the artificial neural nets to the
of cutaneous nature sensed from the TVSS patches. Given the
brain neuronal computations [5], is a very plausible assumption:
visual perception experiences of the blind experimental subjects,
In the scientific neural networks the resolution of complex
rendered by the brain post processed efferent signals, the
problems calls for increasing units (layers) of calculation nodes,
inference would be that the difference between the natures of
and verifiable [6] brain neuronal plasticity allows for engagement
the vision and cutaneous afferent signals must not be in kind but
of various available neuronal modalities (possibly in hierarchical
only in intensity, which dictates the vision details. And this claim
manner). Obviously brain neuronal net with the estimated
is evinced by the hazy visual-like perceptions of blind subjects
availability of many trillions of biological microprocessors is an
due to still insufficiency of the tactile afferent data. Based on such
unfathomable complex, intelligent parallel process computation
observations the disruptive discovery is that: the nature of vision
engine that
is evolutionary perfected and configured for
sustenance of life. Implications are its lightening speed, complexity
resolution potential by virtue of its constructs, and the engrained
As in the case of thought, perception experiences require
learned schemes (as neural patterns), for handling all relevant
an interface, a display venue for the related brain (output)
natural phenomena; including vision and thought that life entails.
efferent signals-- generated post computational processing of
The process of natural phenomena resolution in the brain
the environmental stimuli from the vision and tactile sensing: A
begins with the receipt of the sensed stimuli, and/or with the
known human biological systems interface must be serving this
autonomous deployment of brain
inherent computational
function: Considering the fact that the vocal system performs
neuronal patterns (implicit complexity resolution codes);
such a task for the verbalization of brain's synthesis of language
which is followed by the onset of necessary computations while
expressions efferent signals, the possibility of its versatility for
engaging body's biophysiological feedback system. Clearly, these
expression of the thought, the tactile and the vision perception
operations, due to the ever presence of triggers, are perpetual;
experiences in the form of quiet (inaudible) -- often unaware --
and the streaming outputs, as motor (efferent) signals, render
recital of related information, is suggested.
continuous perceptions of various phenomena. Among them
are the experience of thought, triggered by internal or external
environmental elements; and the experience of vision, mostly
A measure of validation can be found in the very comprehensive
by the external environmental triggers during waking hours;
and detailed experimental work on mirror neuron activities
of which one may or may not be aware. However, embedded in
performed by Keysers et al. [8,9]: They examined a phenomenon
the process of the experience of perception, is the availability
called "tactile sympathy." In these studies, the areas of motor
of a medium for it; an interface. Given the nature of the efferent
neurons activated in a group of subjects watching a movie of
signals, there must be a known biomechanical system interface,
a second group being very lightly touched on the skin were
other than body muscle and skeletal systems, which performs the
significantly similar to those who were actually being touched.
needed function.
Also the combined fMRI and TMS evidence of such sympathy,
The vocal system, mentioned earlier, is a proven candidate:
which is shown in the work of Alaert et al. [10], provides
This interface is responsible for the vocalization of the language,
additional credence to our hypothesis. We believe these results
and occasionally of thought; the latter true for almost all. And this
provide strong experimental support for the concept of the tactile
experience of switching from thought to talk, metaphorically a
nature of vision presented in this theory.
gearshift, discloses perhaps disruptively, the immense possibility
For further validation of theory, we suggest experiments in
which specific vocal motor neuron activities are monitored in
different groups of subjects, some normally sighted and others
As the to natures of vision efferent signals, we made a seemingly
with congenital blindness. Subjects would be monitored during
important and unexpected discovery by critically examining the
speech, thinking, writing and visual (or, in the case of blind
experimental results of the tactile vision substitution system
subjects, vibro-tactile) engagements. Vocal vibrations during
(TVSS), demonstrated in the initial work of Bach-y-Rita et al. [1],
conscious thinking would also be recorded. The results of such
Published in Nature: The work had established the development
experimentation would definitively either prove or refute the
of vision-like perception in blind subjects, when fitted with the
theories put forward in this work.
Citation: Schad JN (2016) Neurological Nature of Vision and Thought and Mechanisms of Perception Experiences. J Neurol Stroke 4(5): 00152. DOI:
10.15406/jnsk.2016.04.00152
Validation
of presence of dual mode to the vocal system, which allows for
expressions of audible and inaudible thoughts.
Neurological Nature of Vision and Thought and Mechanisms of Perception ExperiencesCopyright:©2016 Schad 3/3
References
1. Bach-y-Rita P, Collins CC, Saunders FA, White B, Scadden L (1969)
Vision substitution by tactile image projection. Nature 221(5184):
963-964.
2. Schwarz SH (1999) Visual Perception. McGraw Hill, New York, USA.
3. Chomsky N (2007) On Language. The New Press, New York, USA.
4. Kandel ER, Schwartz JH, Jessell TM (2010) Principles of Neurosciences.
McGraw-Hill, New York, USA.
5. Schad NJ (2016) Brain Neurological Constructs: The Neuronal
Computational Schemes for Resolution of Life's complexities. J Neurol
Neurophysiol 7(1): 356.
6. Edelman GM (1987) Neural Darwinism. The Theory of Neuronal Group
Selection. Basic Books, New York, USA.
7. Edison, Tommy (2013) Intangible Concepts to a Blind Person. Youtube
Recording.
8. Keysers C, Wicker B, Gazzola V, Anton JL, Fogassi L, et al. (2004) A
touching sight SII/PV activation during the observation and experience
of touch. Neuron 42(2): 335-346.
9. Keysers C, Gazzola V (2009) Expanding the mirror: Vicarious activity
for actions, emotions and sensations. Curr Opin Neurobiol 19(16):
666-671.
10. Alaerts K, Swinnen SP, Wenderoth N (2009) Is human primary motor
cortex activated by muscular or direction-dependent features of
observed movements? Cortex 45(10): 1145-1155.
Conclusion
Visual and cutaneous stimuli sensations are (computationally)
processed in the brain similarly; which are evinced by the
development of vision perception in blind subjects, congenial or
otherwise, fitted with tactile visual substitution (TVSS) systems. It
is the scarcity of normal tactile sense data in blinds, which limits
their proper perceptions of the environment. In case of normal
eyesight, retinal neurons figuratively extend themselves by virtue
of receiving rays of Photons which are environmentally modulated
for the physical reality of the object from which they are reflected.
Putting it simply, in the experience of vision we are being touched
by the external world, while in cutaneous experience we are
physically touching them.
Brain's computational operations are also constantly triggered
by beings, exposure to other life phenomena, which are resolved,
and streamed as efferent signals for perception. The experiences
of perceptions in response to senses stimuli must be realized at a
venue, a biomechanical interface which expresses (displays) the
corresponding post process brain efferent signals: Vocal system,
which serves language verbalization, is seemingly the only such
device which could offer this possibility. And this leads to the
presumption that vision, thought and tactile perceptions are but
mostly inaudible utterances at the vocal machinery.
Perhaps this discovery would be disturbing to the poetic
thought; on the face of it; however, knowing that seeing has more
to it than sweep of glance, the "Thought" would be more incensed
of its romantic implications!
Thanks are due to Professor Noam Chomsky for his pioneering
biolinguistic work and for the lecture that has served as an overall
inspiration for me. Further thanks go to Dr. Roya Noorishad, who
instigated my interest in neurosciences.
Acknowledgement
Citation: Schad JN (2016) Neurological Nature of Vision and Thought and Mechanisms of Perception Experiences. J Neurol Stroke 4(5): 00152. DOI:
10.15406/jnsk.2016.04.00152
Neurological Nature of Vision and Thought and Mechanisms of Perception ExperiencesCopyright:©2016 Schad |
1801.03880 | 1 | 1801 | 2018-01-11T17:18:42 | Measuring the Complexity of Consciousness | [
"q-bio.NC",
"physics.bio-ph"
] | The quest for a scientific description of consciousness has given rise to new theoretical and empirical paradigms for the investigation of phenomenological contents as well as clinical disorders of consciousness. An outstanding challenge in the field is to develop measures that uniquely quantify global brain states tied to consciousness. In particular, information-theoretic complexity measures such as integrated information have recently been proposed as measures of conscious awareness. This suggests a new framework to quantitatively classify states of consciousness. However, it has proven increasingly difficult to apply these complexity measures to realistic brain networks. In part, this is due to high computational costs incurred when implementing these measures on realistically large network dimensions. Nonetheless, complexity measures for quantifying states of consciousness are important for assisting clinical diagnosis and therapy. This article is meant to serve as a lookup table of measures of consciousness, with particular emphasis on clinical applicability of these measures. We consider both, principle-based complexity measures as well as empirical measures tested on patients. We address challenges facing these measures with regard to realistic brain networks, and where necessary, suggest possible resolutions. | q-bio.NC | q-bio | Measuring the Complexity of Consciousness
Xerxes D. Arsiwalla1
,
2
,
3 and Paul Verschure1
,
2
,
3
,
4
1 Institute for Bioengineering of Catalonia, Barcelona, Spain.
2 Barcelona Institute for Science and Technology, Barcelona, Spain.
3 Universitat Pompeu Fabra, Barcelona, Spain.
4 Instituci´o Catalana de Recerca i Estudis Avan¸cats (ICREA),
Barcelona, Spain.
{[email protected]}
Abstract. The quest for a scientific description of consciousness has
given rise to new theoretical and empirical paradigms for the investiga-
tion of phenomenological contents as well as clinical disorders of con-
sciousness. An outstanding challenge in the field is to develop measures
that uniquely quantify global brain states tied to consciousness. In partic-
ular, information-theoretic complexity measures such as integrated infor-
mation have recently been proposed as measures of conscious awareness.
This suggests a new framework to quantitatively classify states of con-
sciousness. However, it has proven increasingly difficult to apply these
complexity measures to realistic brain networks. In part, this is due to
high computational costs incurred when implementing these measures on
realistically large network dimensions. Nonetheless, complexity measures
for quantifying states of consciousness are important for assisting clinical
diagnosis and therapy. This article is meant to serve as a lookup table
of measures of consciousness, with particular emphasis on clinical appli-
cability of these measures. We consider both, principle-based complexity
measures as well as empirical measures tested on patients. We address
challenges facing these measures with regard to realistic brain networks,
and where necessary, suggest possible resolutions.
Keywords: Consciousness in the Clinic, Computational Neuroscience,
Complexity Measures.
1
Introduction
In patients with disorders of consciousness, such as coma, locked-in syndrome
or vegetative state, levels of consciousness are assessed in the clinic through
a battery of behavioral tests and neurophysiological recordings. In particular,
these methods are used to assess levels of wakefulness (arousal) and awareness
in patients [28], [27]. Such assessments have led to a two dimensional operational
definition of consciousness for clinical purposes. Assessments of awareness use
behavioral and neurophysiological (fMRI or EEG) protocols in order to gauge
how patients perform on various cognitive functions. Assessments of wakefulness
are based on metabolic markers (if reporting is not possible) such as glucose
8
1
0
2
n
a
J
1
1
]
.
C
N
o
i
b
-
q
[
1
v
0
8
8
3
0
.
1
0
8
1
:
v
i
X
r
a
uptake in the brain, captured using PET scans [17]. As such a clinically-oriented
definition of consciousness enables classification of closely associated states and
disorders of consciousness into clusters on a bivariate scale with awareness and
wakefulness on orthogonal axes. Under healthy conditions, these two levels are
almost linearly correlated, as in conscious wakefulness (high arousal and high
awareness) or in deep sleep (low arousal and low awareness). However, in patho-
logical states, wakefulness without awareness can be observed in the vegetative
state [28], while transiently reduced awareness is observed following seizures [16].
Patients in the minimally conscious state show intermittent and limited non-
reflexive and purposeful behavior [20], [19], whereas patients with hemi-spatial
neglect display reduced awareness of stimuli contralateral to the side where brain
damage has occurred [30]. Given the aforementioned scales for labeling states
and disorders of consciousness, the crucial question is how should one quantify
awareness and wakefulness from neurophysiological data? This is particularly
useful for non-communicative patients such as those in coma or states of min-
imal wakefulness. For this reason, several dynamical complexity measures have
been developed. In this article, we first describe theoretically-grounded com-
plexity measures and the challenges one faces when applying these measures to
realistic brain data. We then outline alternative empirical approaches to classify
states and disorders of consciousness. We end with a discussion on how these
two approaches might inform each other.
2 Measures of Integrated Information
Dynamical complexity measures are designed to capture both, network topol-
ogy as well as causal dynamics. The most prominent among these is integrated
information, denoted as Φ. This was first introduced in [37] and is defined as
the quantity of information generated by a network as a whole, over and above
that of its parts, taking into account the system's causal dynamical interactions.
This reflects the intuition going back to William James that conscious states
are integrated, yet diverse. Φ seeks to operationalize this intuition in terms of
complexity, stating that complexity arises from simultaneous integration and dif-
ferentiation of the network's structure as well as dynamics. Differentiation refers
to functional specialization of neural populations, while integration, as a com-
plementary design principle, results in distributed coordination among neural
populations. This interplay generates integrated yet diversified information be-
lieved to support cognitive and behavioral states. The earliest proposals defining
integrated information were made in [37], [36] and [34]. Since then, considerable
progress has been made towards the development of a normative theory as well
as applications of integrated information [12], [14], [35], [1], [29], [7], [8], [26], [33],
[5]. The core idea of integrated information as a whole versus parts quantity has
been formalized in several distinct information measures such as neural com-
plexity [37], causal density [32], Φ from integrated information theory: IIT 1.0,
2.0 & 3.0 [34], [12], [29], stochastic interaction [39], [11], stochastic integrated
information [14], [1], [9] and synergistic Φ [23], [22]. Table 1 summarizes these
measures along with corresponding information metrics upon which they have
been based.
Table 1. Theoretical complexity measures alongside their corresponding information
metrics.
Integrated Information Measures
Information Metrics
Neural Complexity
Causal Density
Mutual Information (MI)
Granger Causality (GC)
Stochastic Interaction
Kullback-Leibler Divergence (KLD)
IIT 1.0 & 2.0
Stochastic Integrated Information
KLD
MI or KLD
IIT 3.0
Synergistic Φ
Earth Mover's Distance
Synergistic Information
However, computing integrated information for large neurophysiological datasets
has been challenging due to both, computational difficulties and limits on do-
mains where these measures can be implemented. For instance, many of these
measures use the minimum information partition of the network. This involves
evaluating a large number of network configurations (more precisely, the Bell
number), which makes their computational cost extremely high for large net-
works. As for domains of applicability, the measure of [12] has been formu-
lated for discrete-state, deterministic, Markovian systems with the maximum
entropy distribution. On the other hand, the measure of [14] has been devised
to continuous-state, stochastic, non-Markovian systems and in principle, admits
dynamics with any empirical distribution (although in practice, it is easier to use
assuming Gaussian distributions). The formulation in [14] is based on mutual
information, whereas [12] uses a measure based on the Kullback-Leibler diver-
gence. Note however, that in some cases the measure of [14] can take negative
values and that complicates its interpretation. The Kullback-Leibler based defi-
nition computes the information generated during state transitions and remains
positive in the regime of stable dynamics. This gives it a natural interpretation
as an integrated information measure. Both measures [12], [14] make use of a
normalization scheme in their formulations. Normalization inadvertently intro-
duces ambiguities in computations. The normalization is actually used for the
purpose of determining the partition of the network that minimizes the inte-
grated information, but a normalization dependent choice of partition ends up
influencing the value and interpretation of Φ. An alternate measure based on the
Earth Mover's distance was proposed in [29]. This does away with the normaliza-
tion problem (though the current version is not formulated for continuous-state
variables). However, the formulation of [29] lies outside the scope of standard
information theory and is still difficult for performing computations on large
networks.
More recently, these issues have been addressed in [9], using a formulation
of stochastic integrated information based on the Kullback-Leibler divergence
between the conditional multivariate distribution on the set of network states
versus the corresponding factorized distribution over its parts, while implement-
ing the maximum information partition instead of the minimum information
partition. Using this formulation, Φ can be computed for large-scale networks
with linear stochastic dynamics, for both, attractor as well as non-stationary
states [9] (for network simulations see [2], [3], [10]). This work also demonstrated
the first computation of Φ for the resting-state human brain connectome. The
connectome network is estimated from cortical white matter tractography data,
comprising 998 voxels (nodes) with approximately 28,000 weighted symmetric
connections [24]. [9] show that the dynamics and topology of the healthy resting-
state brain generates greater information complexity than a (weight-preserving)
random rewiring of the same network. Even though this formulation of stochas-
tic integrated information was successfully implemented for the human cerebral
connectome, a network of 998 nodes and about 28,000 edges, it was limited to
linearized dynamics. This is well-defined in the vicinity of attractor states such
as the resting-state, however, it would be desirable to extend this formulation to
include non-linearities existing in brain dynamics.
3 Empirical Measures
Ideally, integrated information was intended as a measure of awareness, one that
could account for informational differences between states and also disorders
of consciousness. However, as described above, for realistic brain dynamics and
physiological data that task has in fact proven difficult. On the other hand, the
basic conceptualization of consciousness in terms of integration and differentia-
tion of causal information has motivated several empirical measures that seek to
classify consciousness-related disorders from patient data. For example, [13] in-
vestigated changes in conscious levels using Granger Causality (GC) as a causal
connectivity measure. Given two stationary time-series signals, Granger Causal-
ity measures the extent to which the past of one assists in predicting the future
of the other, over and above the extent to which the past of the latter already
predicts its own future [21], thus quantifying causal relations between two sig-
naling sources. This was tested using electroencephalographic (EEG) data from
subjects undergoing propofol-induced anesthesia, with signals source-localized
to the anterior and posterior cingulate cortices. [13] found a significant increases
in bidirectional GC in most subjects during loss of consciousness, especially in
the beta and gamma frequency ranges. Another useful measure of causal connec-
tivity is transfer entropy, which extends Granger causality to the non-Gaussian
case. However, so far this has only been implemented on neuronal cultures by
[40] and holds future potential as a clinically relevant measure. Yet another
measure that has already proven useful as a clinical classifier of conscious lev-
els is the Perturbational Complexity Index (PCI), which was introduced by [18]
and tested on TMS-evoked potentials measured with EEG. PCI is calculated
by perturbing the cortex with transcranial magnetic stimulation (TMS) in or-
der to engage distributed interactions in the brain and then compressing the
resulting spatiotemporal EEG responses to measure their algorithmic complex-
ity, based on the Lempel-Ziv compression. For a given segment of EEG data,
the Lempel-Ziv algorithm quantifies complexity by counting the number of dis-
tinct patterns in the data. For example, this can be proportional to the size
of a computer file after applying a data compression algorithm. Computing the
Lempel-Ziv compressibility requires binarizing the time-series data, based either
on event-related potentials or with respect to a given threshold. Using PCI, [18]
were able to discriminate levels of consciousness during wakefulness, sleep, and
anesthesia, as well as in patients who had emerged from coma and recovered a
minimal level of consciousness. Later, the Lempel-Ziv complexity was also used
by [31] on spontaneous high-density EEG data recorded from subjects under-
going propofol-induced anesthesia. Once again, a robust decline in complexity
was observed during anesthesia. These are complexity measures based on data
compression algorithms. A qualitative comparison between a data compression
measure inspired by PCI and Φ was made in [38]. While compression-based
measures do seem to capture certain aspects of Φ, the exact relationship be-
tween the two is not completely clear. Nonetheless, these empirical measures
have been useful for clinical purposes, in terms of broadly discriminating dis-
orders of consciousness. Another relevant complexity measure is the weighted
symbolic mutual information (wSMI), introduced by [25]. This is a measure of
global information sharing across brain areas. It evaluates the extent to which
two EEG channels present nonrandom joint fluctuations, suggesting that they
share common sources. This is done by first transforming continuous signals into
discrete symbols, and subsequently computing the joint probabilities of symbol
pairs between two EEG channels. Before computing the symbolic mutual infor-
mation between two time-series signals, a weighting is introduced to disregard
conjunctions of identical or opposite-sign symbols from the two signal trains as
that could potentially arise from common-source artifacts. In [25] wSMI was
estimated for 181 EEG recordings from awake but noncommunicating patients
diagnosed in various disorders of consciousness (including 143 from patients in
vegetative and minimally conscious states). This measure of information sharing
was found to systematically increases with consciousness. In particular, it was
able to distinguish patients in the vegetative state, minimally conscious state,
and fully conscious state. In Table 2 we summarize the above empirical mea-
sures along with their domains of application.
4 Discussion
The paradigm-shifting proposal that consciousness might be measurable in terms
of the information generated by causal dynamics of the brain as a whole, over
the sum of its parts, has led to precise quantitative formulations of information-
theoretic complexity measures. These measures seek to operationalize the intu-
ition that the complexity associated to consciousness arises from simultaneous
integration and differentiation of the brain's structural and dynamical hierar-
chies. However, progress in this direction has faced practical challenges such
Table 2. Empirical complexity measures alongside their tested domains of application.
Empirical Measures
Tested Application Domains
Granger Causality
Wakefulness vs propofol-induced
anesthesia using EEG
Perturbational Complexity Index Wakefulness, sleep, anesthesia, coma & minimal
Lempel-Ziv Complexity
Weighted Symbolic Mutual Information
consciousness using TMS-evoked EEG
Wakefulness vs propofol-induced
anesthesia using EEG
Vegetative, minimally conscious &
fully conscious states using EEG
as high computational cost upon scaling with network size. This is especially
true with regard to realistic neuroimaging or physiological datasets. Even in the
approach of [9], where both, the scaling and normalization problem have been
solved, the formulation is still applicable only to linear dynamical systems. A
possible way to extend this formulation to non-linear systems such as the brain
might be to first solve the Fokker-Planck equations for these systems (as prob-
ability distributions will no longer remain Gaussian) and subsequently estimate
entropies and conditional entropies numerically to compute Φ. Another solution
to the problem might be to construct statistical estimators for the covariance
matrices from data and then compute Φ.
In the meanwhile, for clinical purposes, it has been useful to consider empiri-
cal complexity measures, which serve as classifiers that very broadly discriminate
states of consciousness, such as between wakefulness and anesthesia or broadly
between disorders of consciousness. However, these measures do not strictly cor-
respond to integrated information. Some of them are based on signal compres-
sion, which does capture differentiation, though not directly integration. So far
these methods have been applied on the scale of EEG datasets. One has yet to
demonstrate their computational feasibility for larger datasets (which might only
be a matter of time though). All in all, bottom-up approaches suggest important
features that might help inform or constrain implementations of principle-based
approaches. However, the latter are indispensable for ultimately understanding
causal aspects of information generation and flow in the brain.
This article is intended as a lookup table spanning the landscape of both,
theoretically-motivated as well as empirically-based complexity measures used in
current consciousness research. Even though, for the purpose of this article, we
have treated complexity as a global correlate of consciousness, there are indica-
tions that multiple complexity types, based on cognitive and behavioral control,
might be important for a more precise classification of various states of con-
sciousness [15], [6]. This latter observation alludes to the need for an integrative
systems approach to consciousness research, one that is grounded in cognitive
architectures and helps understand control mechanisms underlying systems level
neural information processing [4].
Acknowledgments. This work has been supported by the European Research
Council's CDAC project: "The Role of Consciousness in Adaptive Behavior: A
Combined Empirical, Computational and Robot based Approach" (ERC-2013-
ADG 341196).
References
1. Arsiwalla, X.D., Verschure, P.F.M.J.: Integrated information for large complex net-
works. In: The 2013 International Joint Conference on Neural Networks (IJCNN).
pp. 1–7 (Aug 2013)
2. Arsiwalla, X.D., Betella, A., Bueno, E.M., Omedas, P., Zucca, R., Verschure, P.F.:
The dynamic connectome: A tool for large-scale 3d reconstruction of brain activity
in real-time. In: ECMS. pp. 865–869 (2013)
3. Arsiwalla, X.D., Dalmazzo, D., Zucca, R., Betella, A., Brandi, S., Martinez, E.,
Omedas, P., Verschure, P.: Connectomics to semantomics: Addressing the brain's
big data challenge. Procedia Computer Science 53, 48–55 (2015)
4. Arsiwalla, X.D., Herreros, I., Moulin-Frier, C., Sanchez, M., Verschure, P.F.: Is
Consciousness a Control Process?, pp. 233–238. IOS Press, Amsterdam (2016)
5. Arsiwalla, X.D., Mediano, P.A., Verschure, P.F.: Spectral modes of network dynam-
ics reveal increased informational complexity near criticality. Procedia Computer
Science 108, 119–128 (2017)
6. Arsiwalla, X.D., Moulin-Frier, C., Herreros, I., Sanchez-Fibla, M., Verschure, P.F.:
The morphospace of consciousness. arXiv preprint arXiv:1705.11190 (2017)
7. Arsiwalla, X.D., Verschure, P.: Computing Information Integration in Brain Net-
works, pp. 136–146. Springer International Publishing, Cham, Switzerland (2016)
8. Arsiwalla, X.D., Verschure, P.F.M.J.: High Integrated Information in Complex
Networks Near Criticality, pp. 184–191. Springer International Publishing, Cham,
Switzerland (2016)
9. Arsiwalla, X.D., Verschure, P.F.: The global dynamical complexity of the human
brain network. Applied Network Science 1(1), 16 (2016)
10. Arsiwalla, X.D., Zucca, R., Betella, A., Martinez, E., Dalmazzo, D., Omedas, P.,
Deco, G., Verschure, P.: Network dynamics with brainx3: A large-scale simulation
of the human brain network with real-time interaction. Frontiers in Neuroinfor-
matics 9(2) (2015)
11. Ay, N.: Information geometry on complexity and stochastic interaction. Entropy
17(4), 2432–2458 (2015)
12. Balduzzi, D., Tononi, G.: Integrated information in discrete dynamical systems:
motivation and theoretical framework. PLoS Comput Biol 4(6), e1000091 (2008)
13. Barrett, A.B., Murphy, M., Bruno, M.A., Noirhomme, Q., Boly, M., Laureys, S.,
Seth, A.K.: Granger causality analysis of steady-state electroencephalographic sig-
nals during propofol-induced anaesthesia. PloS one 7(1), e29072 (2012)
14. Barrett, A.B., Seth, A.K.: Practical measures of integrated information for time-
series data. PLoS Comput Biol 7(1), e1001052 (2011)
15. Bayne, T., Hohwy, J., Owen, A.M.: Are there levels of consciousness? Trends in
cognitive sciences 20(6), 405–413 (2016)
16. Blumenfeld, H.: Impaired consciousness in epilepsy. The Lancet Neurology 11(9),
814–826 (2012)
17. Bodart, O., Gosseries, O., Wannez, S., Thibaut, A., Annen, J., Boly, M., Rosanova,
M., Casali, A.G., Casarotto, S., Tononi, G., et al.: Measures of metabolism and
complexity in the brain of patients with disorders of consciousness. NeuroImage:
Clinical 14, 354–362 (2017)
18. Casali, A.G., Gosseries, O., Rosanova, M., Boly, M., Sarasso, S., Casali, K.R.,
Casarotto, S., Bruno, M.A., Laureys, S., Tononi, G., et al.: A theoretically based
index of consciousness independent of sensory processing and behavior. Science
translational medicine 5(198), 198ra105–198ra105 (2013)
19. Giacino, J.T.: The vegetative and minimally conscious states: consensus-based cri-
teria for establishing diagnosis and prognosis. NeuroRehabilitation 19(4), 293–298
(2004)
20. Giacino, J.T., Ashwal, S., Childs, N., Cranford, R., Jennett, B., Katz, D.I., Kelly,
J.P., Rosenberg, J.H., Whyte, J., Zafonte, R., et al.: The minimally conscious state
definition and diagnostic criteria. Neurology 58(3), 349–353 (2002)
21. Granger, C.W.: Investigating causal relations by econometric models and cross-
spectral methods. Econometrica: Journal of the Econometric Society pp. 424–438
(1969)
22. Griffith, V.: A principled infotheoretic\ phi-like measure. arXiv preprint
arXiv:1401.0978 (2014)
23. Griffith, V., Koch, C.: Quantifying
Synergistic Mutual
Information,
(2014),
159–190.
pp.
http://dx.doi.org/10.1007/978-3-642-53734-9_6
Springer Berlin Heidelberg, Berlin, Heidelberg
24. Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C.J., Wedeen, V.J.,
Sporns, O.: Mapping the Structural Core of Human Cerebral Cortex. PLoS Biology
6(7), 15 (2008)
25. King, J.R., Sitt, J.D., Faugeras, F., Rohaut, B., El Karoui, I., Cohen, L., Nac-
cache, L., Dehaene, S.: Information sharing in the brain indexes consciousness in
noncommunicative patients. Current Biology 23(19), 1914–1919 (2013)
26. Krohn, S., Ostwald, D.: Computing integrated information. arXiv preprint
arXiv:1610.03627 (2016)
27. Laureys, S.: The neural correlate of (un) awareness: lessons from the vegetative
state. Trends in cognitive sciences 9(12), 556–559 (2005)
28. Laureys, S., Owen, A.M., Schiff, N.D.: Brain function in coma, vegetative state,
and related disorders. The Lancet Neurology 3(9), 537–546 (2004)
29. Oizumi, M., Albantakis, L., Tononi, G.: From the phenomenology to the mech-
anisms of consciousness: integrated information theory 3.0. PLoS Comput Biol
10(5), e1003588 (2014)
30. Parton, A., Malhotra, P., Husain, M.: Hemispatial neglect. Journal of Neurology,
Neurosurgery & Psychiatry 75(1), 13–21 (2004)
31. Schartner, M., Seth, A., Noirhomme, Q., Boly, M., Bruno, M.A., Laureys, S., Bar-
rett, A.: Complexity of multi-dimensional spontaneous eeg decreases during propo-
fol induced general anaesthesia. PloS one 10(8), e0133532 (2015)
32. Seth, A.K.: Causal connectivity of evolved neural networks during behavior. Net-
work: Computation in Neural Systems 16(1), 35–54 (2005)
33. Tegmark, M.: Improved measures of
integrated information. arXiv preprint
arXiv:1601.02626 (2016)
34. Tononi, G.: An information integration theory of consciousness. BMC neuroscience
5(1), 42 (2004)
35. Tononi, G.: Integrated information theory of consciousness: an updated account.
Arch Ital Biol 150(2-3), 56–90 (2012)
36. Tononi, G., Sporns, O.: Measuring information integration. BMC neuroscience
4(1), 31 (2003)
37. Tononi, G., Sporns, O., Edelman, G.M.: A measure for brain complexity: relating
functional segregation and integration in the nervous system. Proceedings of the
National Academy of Sciences 91(11), 5033–5037 (1994)
38. Virmani, M., Nagaraj, N.: A compression-complexity measure of integrated infor-
mation. arXiv preprint arXiv:1608.08450 (2016)
39. Wennekers, T., Ay, N.: Stochastic interaction in associative nets. Neurocomputing
65, 387–392 (2005)
40. Wibral, M., Vicente, R., Lindner, M.: Transfer Entropy in Neuroscience, pp. 3–36.
Springer Berlin Heidelberg, Berlin, Heidelberg (2014)
|
1608.06548 | 3 | 1608 | 2016-10-27T15:31:08 | Grand Challenges for Global Brain Sciences | [
"q-bio.NC"
] | The next grand challenges for society and science are in the brain sciences. A collection of 60+ scientists from around the world, together with 10+ observers from national, private, and foundations, spent two days together discussing the top challenges that we could solve as a global community in the next decade. We eventually settled on three challenges, spanning anatomy, physiology, and medicine. Addressing all three challenges requires novel computational infrastructure. The group proposed the advent of The International Brain Station (TIBS), to address these challenges, and launch brain sciences to the next level of understanding. | q-bio.NC | q-bio | Grand Challenges for Global Brain Sciences
Global Brain Workshop 2016 Attendees*
The next grand challenges for science and society are in the brain sciences.
A collection of 60+ scientists from around the world, together with 15+
observers from national, private, and foundations, spent two days together
discussing the top challenges that we could solve as a global community in
the next decade. We settled on three challenges, spanning anatomy,
physiology, and medicine. Addressing all three challenges requires novel
computational
infrastructure. The group proposed the creation of The
International Brain Station (TIBS), to address these challenges, and launch
brain sciences to the next level of understanding.
Understanding the brain and curing its diseases are among the most exciting challenges of our
time. Consequently, national, transnational, and private parties are investing billions of dollars
(USD). To efficiently join forces, Global Brain Workshop 2016 was hosted at Johns Hopkins
University's Kavli Neuroscience Discovery Institute on April 7-8. A second workshop, Open Data
Ecosystem in Neuroscience took place July 25-26 in Washington, DC to continue the discussion
specifically about computational challenges and opportunities. A third conference, Coordinating
took place in New York City on September 19th in association with the
Global Brain Projects,
United Nations General Assembly. So vast are both the challenges and the opportunities that
global coordination is crucial.
To find ways of synergistically studying the brain, the kick-off workshop welcomed over 60
scientists, representing 12 different countries and a wide range of subdisciplines. They were
joined by 15 observers from various national and international
funding organizations.
Participants were engaged weeks before the conference and charged with coming up with
ambitious projects that are both feasible and internationally inclusive, on par with the
International Space Station (i.e., worthy of a global, decade-long effort). Over the course of 36
hours, scientists discussed, debated, and gathered feedback, ultimately proposing several
"grand challenges for global brain sciences" that were refined by working groups. The workshop
was covered in a media piece in Science
The group began with 60+ ideas, each forged independently by one of the scientific participants.
Each participant proposed a unique challenge that was designed to meet
the following
desiderata:
April 15, 2016.
1. Significant
: it will yield tangible societal, economic, and medical benefits to the world.
2. Feasible
: it can achieve major milestones within 10 years given existing funding
3.
opportunities.
: nations throughout the world can meaningfully contribute to and benefit from
Inclusive
each challenge, and the collection of challenges are collectively scientifically diverse.
the proposed ideas were similar to one another and others were
Interestingly, a lot of
complementary. This allowed the group to converge on three grand challenges for global brain
sciences, each depending on a common universal resource.
Challenge 1: What makes our brains unique?
Both within and across species, brain structure is known to exhibit significant variability across
many orders of magnitude in scale-including anatomy, biochemistry, connectivity,
development, and gene expression (ABCDE). It remains mysterious how and why the nervous
system tightly regulates certain properties, while allowing others to vary. Understanding the
design principles governing variability may hold the key to understanding intelligence and
subjective experience, as well as the influence of variability on health and function.
This grand challenge is a global project to coordinate the construction of comprehensive
multiscale maps of the ABCDE's of multiple brains from multiple species using multiple
cognitive and mental health disease models. Within a decade, we expect to have addressed
this challenge in brains including but not
limited to Drosophila, Zebrafish, Mouse, and
Marmoset, and to have developed tools to conduct massive neurocartographic analyses. The
result will be a state-of-the-art "Virtual NeuroZoo" with fully annotated data and analytic tools for
analysis and discovery. This virtual NeuroZoo can be utilized by neuroscientists and citizens
alike, both as a reference and for educational materials. By incorporating disease models, we
explicitly link this challenge with the third challenge.
Challenge 2: How does the brain solve complex computational problems?
Brains remain the most computationally advanced machines for a large array of cognitive
tasks-whether navigating hazardous terrain, translating languages, conducting surgery, or
recognizing emotional states-despite the fact that modern computers can utilize millions of
training samples, megawatts of power, and tons of hardware. While the ABCDEs establish the
"wetware" upon which our brains can solve such computations, to understand the mechanisms
we need to measure, manipulate, and model neural activity simultaneously across many
resolutions and scales-including wearables, embedded sensors, and
spatiotemporal
actuators-while animals are exhibiting complex ecological behaviors in naturalistic
environments.
This grand challenge is a global project to investigate a single naturalistic behavior that
is ecologically relevant across phylogenies, such as foraging, and measure brain and
body properties across spatial, temporal, and genetic scales. The challenge differs from
previous efforts in three key ways. First, it requires studying animals in complex and naturalistic
it requires coordinated attacks at many different scales by many
environments. Second,
investigators while the animals are performing the same complex behaviors. We
different
envision groups of 20-30 investigators all operating together to share data and experimental
design. Third, the richness of the mental repertoire of cognition suggests that deciphering its
codes will require many parallel
investigations to uncover different facets of brain function.
These experiments in turn will produce multiscale models of neural systems with the potential to
accomplish computational tasks that no current computer system can perform. Mechanistic
studies, guided by theoretical models, will help to ask how perturbations of those systems lead
to aberrant function, linking this challenge with the next one.
financial costs, and loss of productivity. Despite a growing awareness of
Challenge 3: How can we augment clinical decision-making to prevent
disease and restore brain function?
illnesses levy enormous burdens upon humanity: impairment,
Psychiatric and neurological
suffering,
the
challenges, clinicians consistently battle the lack of objective tests to guide clinical
treatments, prognosis). Compounding these
decision-making (e.g., diagnosis, selection of
limitations are societal stigmas regarding mental illness that increase the suffering of patients
and their families. The ABCDEs of neurobiological variability, when coupled with multiscale
mechanistic models of cognition, will provide new approaches to neurobiologically-informed
clinical decision making.
This grand challenge is a global project to transform clinical decision-making via
incorporating neural mechanisms of dysfunction. This will require collecting, organizing and
analyzing human and non-human anatomical and functional data. These data, and the tools
developed to explore and discover novel treatment therapies, will be the foundation upon which
the next decades of experiments and clinical decisions will be based. The distributed and
multimodal nature of these datasets further motivate the need for an all-purpose computational
platform, upon which models of disease can be developed, deployed, tested, and refined.
A Universal Resource
All three of the grand challenges for global brain sciences represent severe methodological
challenges, both technological and computational. The technological developments required for
each of the challenges are non-overlapping. In contrast, regardless of the nature of the scientific
questions or data modalities involved, each project will require computational capabilities
including collecting, storing, exploring, analyzing, modeling, and discovering data. Although
neuroscience has developed a large number of computational
tools to deal with existing
datasets, the datasets proposed here bring with them a whole suite of new challenges.
This resource would be a comprehensive computational platform, deployed in the cloud, that will
provide web services for all the current "pain points" in daily neuroscience practice associated
with big data. This resource will realize a new era of brain sciences, one in which the
bottlenecks to discovery transition away from data collection and processing to data enriching
exploring, and modeling. While science has always benefitted from standing on the shoulders of
. Today, essentially every
giants, this will enable science to stand on the shoulders of everyone
practicing neuroscientist's productivity is limited due to computational resources, access to data
or algorithms, or struggling with determining which data and algorithms are best suited to
answer the most pressing questions of our generation. This resource will create a future where
those limitations will feel as archaic as fitting the data with paper and pencil feels today. For
further details, see an upcoming NeuroView called "To the Cloud! A Grassroots Proposal to
Accelerate Brain Science Discovery".
Societal Considerations
Each nation affords different opportunities and restrictions, owing to ethical, policy, and cultural
considerations. Because these grand challenges are inherently inclusive, manifesting them will
require understanding and mitigating issues that arise in cross-cultural endeavors. Indeed,
addressing the vast diversity of partnerships in such an endeavor is a challenge in itself. We
therefore recommend the following. First, form a cultural sensitivity committee to consider
and investigate potentially sensitive issues. Second, bolstered by their research, establish
cross-cultural collaboration education materials, including written guidelines and videos,
which will be recommended to all participating scientists. Third, to deepen the understanding of
transnational collaborations, develop trainee exchange programs in which participating
trainees will spend six months to a year working and training in a foreign country. This will also
facilitate cross-cultural knowledge dissemination and fertilization. Fourth, require frequent
assessments to ensure maintenance of cultural sensitivities. These assessments will feedback
into the educational material and be used to modify the exchange programs.
Next Steps
Crucial to the success of this endeavor is a sequence of actionable steps that the community
can follow. Because we are not proposing any additional funding, realizing the eventual goals of
these grand challenges will rely on marshalling existing funds. Due to the incoming leadership
changes, both on national and transnational levels, quick action is of the essence. Therefore,
we have taken the following steps: We have created a webpage, http://brainx.io, containing a
bibliography of reports that resulted from this conference, as well as a list of all scientific
participants and observers who attended the original brainstorming meeting leading to this
document. We will also be monitoring comments on https://neurostars.org/
with the tag
"neurostorm" for further discussion. Finally, we will have an outpost at the NeuroData booth
(#4126) at the SfN meeting in San Diego to discuss these issues further. We encourage
anybody who feels inspired by this document to join the discussion, engage, and get in touch
with funders and other scientists with your ideas.
Acknowledgements
National Science Foundation (1637376) and the Kavli Foundation.
*Global Brain Workshop 2016 Attendees
Joshua T. Vogelstein1,27,28,29,30,31
●
Katrin Amunts7,8
●
Andreas Andreou30
●
Dora Angelaki32
●
Giorgio A. Ascoli33
●
Cori Bargmann34
●
Randal Burns28, 29
●
Corrado Cali11
●
Frances Chance35
●
George Church36
●
Hollis Cline37
●
Todd Coleman38
●
Stephanie de La Rochefoucauld39
●
● Winfried Denk40
Ana Belén Elgoyhen41
●
●
Ralph Etienne Cummings42
Alan Evans5
●
Kenneth Harris43
●
● Michael Hausser3
●
Sean Hill9
●
Samuel Inverso44
Chad Jackson45
●
Viren Jain46
●
●
Rob Kass47
●
Bobby Kasthuri13
Adam Kepecs15
●
Gregory Kiar1, 27
●
●
Dean M. Kleissas25
●
Konrad Kording6
●
Sandhya P. Koushika10
John Krakauer48
●
Story Landis49
●
●
Jeff Layton50
●
Qingming Luo51
Adam Marblestone52
●
David Markowitz26
●
●
Justin McArthur53
●
Brett Mensh2,4
● Michael P. Milham19
Partha Mitra15
●
Pedja Neskovic54
●
● Miguel Nicolelis55
●
Richard O'Brien56
Aude Oliva57
●
Gergo Orban58
●
●
Hanchuan Peng14
Eric Perlman27
●
● Marina Picciotto59
● Mu-Ming Poo17
●
●
●
●
●
●
Jean-Baptiste Poline18
Alexandre Pouget60
Sridhar Raghavachari61
Jane Roskams14
Alyssa Picchini Schaffer20
Terry Sejnowski62
Friedrich T. Sommer63
●
Nelson Spruston4
●
Larry Swanson64
●
Arthur Toga65
●
R. Jacob Vogelstein26
●
Anthony Zador15
●
●
Richard Huganir30,31
● Michael I. Miller1,27,31
1. Department of Biomedical Engineering, Institute for Computational Medicine, Johns Hopkins University, Baltimore, MD,
USA
Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, VA, USA
2. Optimize Science, Mill Valley, CA USA; UCSF Kavli Institute for Fundamental Neuroscience, San Francisco, CA, USA
3. Department of Physiology, University College London, London, UK
4.
5. Montreal Neurological Institute, McGill University,,Montreal, Quebec, Canada
6.
Physical Medicine and Rehabilitation, Physiology, and Applied Mathematics, and Biomedical Engineering, Northwestern
University, Chicago, IL, USA
Institute for Neuroscience and Medicine, INM-1, Research Centre Juelich, Germany, C. and O. Vogt Institute for Brain
Research, Forschungszentrum Jülich; University Hospital Duesseldorf, University Duesseldorf, Germany
7.
Blue Brain Project, EPFL, Campus Biotech, Geneva, Switzerland
8. Human Brain Project, EPFL, Geneva, Switzerland
9.
10. Department of Biological Sciences, Tata Institute of Fundamental Research, Navy Nagar, Colaba, Mumbai, India
11. Biological and Environmental Science and Engineering, KAUST,Thuwal, 23955-6900 Saudi Arabia
12. Cuban Neuroscience Center, 190 e / 25 and 27, Cubanacan, Playa. Havana. CP 11600; University of Electronic Science
and Technology of China, Shahe Campus:No.4, Section 2, North Jianshe Road, 610054, Chengdu, Sichuan, P.R.China
13. Argonne National Laboratory, Argonne, IL, USA
14. Allen Institute for Brain Science, Seattle, WA, USA
15. Cold Spring Harbor Laboratory, Cold Spring Harbor, NY, USA
16. Department of Psychological and Brain Sciences, Dartmouth College, Hanover, NH, USA
17.
Institute of Neuroscience, CAS Center for Brain Science, 320 Yue Yang Road Shanghai, 200031 P.R.China; Intelligence
Technology, Chinese Academy of Sciences, 319 Yueyang Road, Shanghai 200031, P.R.China
18. Henry H. Wheeler Jr. Brain Imaging Center, Helen Wills Neuroscience Institute, 188 Li Ka Shing Center for Biomedical
and Health Sciences, Henry H. Wheeler, Jr. Brain Imaging Center, Suite B107, University of California, Berkeley, CA
94720, USA
19. Center for the Developing Brain, Child Mind Institute, New York, NY; Nathan S. Kline Institute for Psychiatric Research,
Orangeburg, NY, USA
20. Simons Collaboration on the Global Brain, Simons Foundation, New York, NY, USA
21.
22. Department of Physiology, Keio University School of Medicine, 35 Shinanomachi, Shinjuku-ku, Tokyo,, Japan; RIKEN
Israel Brain Technologies, Hakfar Hayarok, Ramat Hasharon, Israel
Brain Science Institute, Laboratory for Marmoset Neural Architecture, 2-1 Hirosawa, Wako, Saitama, Japan
23. Mind Research Network, Department of Electrical and Computer Engineering, University of New Mexico, Albuquerque,
NM, USA
Intelligence Advanced Research Projects Activity (IARPA), Maryland Square Research Park, Riverdale Park, MD, USA
Institute for Data Intensive Engineering and Sciences, Johns Hopkins University, Baltimore,MD, USA
24. The Kavli Foundation, Oxnard, CA, USA
25. Johns Hopkins University Applied Physics Laboratory, Laurel, MD, USA
26.
27. Center for Imaging Science, Johns Hopkins University, Baltimore, MD, USA
28. Department of Computer Science, Johns Hopkins University, Baltimore, MD, USA
29.
30. Department of Neuroscience, Johns Hopkins University, Baltimore, MD, USA
31. Kavli Neuroscience Discovery Institute, Johns Hopkins University, Baltimore, MD, USA
32. Dept of Neuroscience, Baylor College of Medicine, Houston, TX, USA
33. Dept of Molecular Neuroscience, George Mason University, Fairfax, VA, USA
34. Howard Hughes Medical Institute, Rockefeller University, New York, NY, USA
35. Sandia National Laboratories, Albuquerque, NM, USA
36. Harvard Medical School, Harvard University, Boston, MA, USA
37. Department of Molecular and Cellular Neuroscience, The Scripps Research Institute, La Jolla, CA, USA
38. Department of Bioengineering, University of California, San Diego, CA, USA
International Brain Research Organization (IBRO), Paris, France
39.
40. Max Planck Institute of Neurobiology, Martinsried, Germany
41. Molecular Biology and Genetic Engineer Institute, CONICET, Argentina
42. Department on Electrical & Computer Engineering, Johns Hopkins University, Baltimore, MD, USA
43. Department of Quantative Neuroscience, University College London, London, England
44. Department of Genetics, Harvard University, Boston, MA, USA
45. U.S. Department of State, Washington D.C., USA
46. Google, Mountain View, CA, USA
47. Department of Statistics, Carnegie Mellon University, Pittsburgh, PA, USA
48. Department of Neurology, Johns Hopkins Hospital, Baltimore, MD, USA
49. National Institute of Neurological Disorders and Stroke (NINDS), National Institute of Health, Bethesda, MD, USA
50. Amazon Web Services, Atlanta, GA, USA
51. Britton Chance Center for Biomedical Photonics, Wuhan National Laboratory for Optoelectronics, Huazhong University of
Science & Technology, Wuhan, 430074, China
52. MIT Media Lab, Cambridge, MA, USA
53. Department of Neurology, Johns Hopkins Hospital, Baltimore, MD, USA
54. Office of Naval Research, Arlington, VA, USA
55. Duke Institute for Brain Sciences, Duke University, Durham, NC, USA
56. Department of Neurology, Duke University School of Medicine, Durham, NC, USA
57. National Science Foundation, Arlington, VA, USA
58. Department of Theoretical Physics, MTA Wigner Research Centre for Physics, Budapest, Hungary
59. Department of Psychiatry, Yale University, New Haven, CT, USA
60. Neuroscience Center, University of Geneva, Geneva, Swizterland
61. National Science Foundation, Arlington, VA, USA
62. Salk Institute for Biological Studies, La Jolla, CA, USA
63. Redwood Center for Theoretical Neuroscience, University of California, Berkley, CA, USA
64. Department of Biological Sciences, University of Southern California, Los Angeles, CA, USA
65.
Institute for Neuroimaging and Informatics, University of Southern California, Los Angeles, CA, USA
|
1604.07457 | 1 | 1604 | 2016-04-25T22:01:50 | Modeling the Contribution of Central Versus Peripheral Vision in Scene, Object, and Face Recognition | [
"q-bio.NC",
"cs.CV"
] | It is commonly believed that the central visual field is important for recognizing objects and faces, and the peripheral region is useful for scene recognition. However, the relative importance of central versus peripheral information for object, scene, and face recognition is unclear. In a behavioral study, Larson and Loschky (2009) investigated this question by measuring the scene recognition accuracy as a function of visual angle, and demonstrated that peripheral vision was indeed more useful in recognizing scenes than central vision. In this work, we modeled and replicated the result of Larson and Loschky (2009), using deep convolutional neural networks. Having fit the data for scenes, we used the model to predict future data for large-scale scene recognition as well as for objects and faces. Our results suggest that the relative order of importance of using central visual field information is face recognition>object recognition>scene recognition, and vice-versa for peripheral information. | q-bio.NC | q-bio | Modeling the Contribution of Central Versus Peripheral Vision in Scene, Object,
and Face Recognition
Panqu Wang ([email protected])
Department of Electrical and Engineering, University of California San Diego
9500 Gilman Dr 0407, La Jolla, CA 92093 USA
Garrison W. Cottrell ([email protected])
Department of Computer Science and Engineering, University of California San Diego
9500 Gilman Dr 0404, La Jolla, CA 92093 USA
6
1
0
2
r
p
A
5
2
]
.
C
N
o
i
b
-
q
[
1
v
7
5
4
7
0
.
4
0
6
1
:
v
i
X
r
a
Abstract
It is commonly believed that the central visual field (fovea and
parafovea) is important for recognizing objects and faces, and
the peripheral region is useful for scene recognition. However,
the relative importance of central versus peripheral informa-
tion for object, scene, and face recognition is unclear. Larson
and Loschky (2009) investigated this question in the context of
scene processing using experimental conditions where a cir-
cular region only reveals the central visual field and blocks
peripheral information ("Window"), and in a "Scotoma" con-
dition, where only the peripheral region is available. They
measured the scene recognition accuracy as a function of vi-
sual angle, and demonstrated that peripheral vision was indeed
more useful in recognizing scenes than central vision in terms
of achieving maximum recognition accuracy.
In this work,
we modeled and replicated the result of Larson and Loschky
(2009), using deep convolutional neural networks (CNNs).
Having fit the data for scenes, we used the model to predict
future data for large-scale scene recognition as well as for ob-
jects and faces. Our results suggest that the relative order of
importance of using central visual field information is face
recognition>object recognition>scene recognition, and vice-
versa for peripheral information. Furthermore, our results pre-
dict that central information is more efficient than peripheral
information on a per-pixel basis across all categories, which is
consistent with Larson and Loschky's data.
Keywords: face recognition; object recognition; scene recog-
nition; central and peripheral vision; deep neural networks
Introduction
Viewing a real-world scene occupies the entire visual field,
but the visual resolution across the visual field varies. The
fovea, a small region in the center of the visual field that sub-
tends approximately 1◦ of visual angle (Polyak, 1941), per-
ceives the highest visual resolution of 20 to 45 cycles/degree
(cpd) (Loschky, McConkie, Yang, & Miller, 2005). The
parafovea has a slightly lower visual resolution and extends
to about 4-5◦ eccentricity, where the highest density of rods
is found (Wandell, 1995). Beyond the parafovea is generally
considered to be peripheral vision (Holmes, Cohen, Haith, &
Morrison, 1977), which receives the lowest visual resolution.
Due to the high density and small receptive field of retinal
receptors,the central (foveal and parafoveal) vision encodes
information of higher spatial frequency and more detail; pe-
ripheral vision, on the contrary, encodes coarser and lower
spatial frequency information.
This retinotopic representation of the visual field is mapped
to visual cortical areas through a log-polar representation.
Recent studies have shown that orderly central and periph-
eral representations can be found not only in low-level to
mid-level visual areas (V1-V4), but also in higher-level re-
gions, where perception and recognition for faces or scenes
is engaged (Malach, Levy, & Hasson, 2002; Grill-Spector &
Malach, 2004). More specifically, Malach et al. (2002) pro-
posed that the need for visual resolution is a crucial factor
in organizing object areas in higher-level visual cortex: ob-
ject recognition that depends more on fine detail is associated
with central-biased representations, such as faces and words;
object recognition that depends more on large-scale integra-
tion is associated with peripheral-biased representations, such
as buildings and scenes. This hypothesis is supported by
fMRI evidence, which shows that the brain areas that are
more activated for faces (FFA; Kanwisher, McDermott, and
Chun (1997)) and words (VWFA; McCandliss, Cohen, and
Dehaene (2003)) sit in the eccentricity band expanded by
central visual-field bias, whereas buildings and scenes (PPA;
Epstein, Harris, Stanley, and Kanwisher (1999)) are asso-
ciated with peripheral bias. More recent studies even sug-
gest that the central-biased pathway for recognizing faces and
peripheral-biased pathway for recognizing scenes are segre-
gated by mid-fusiform sulcus (MFS) to enable fast parallel
processing (Gomez et al., 2015).
In the domain of behavioral research, studies have shown
that object perception performance is the best around 1◦-2◦
of fixation point and drops rapidly as eccentricity increases
(Henderson & Hollingworth, 1999; Nelson & Loftus, 1980).
For scene recognition, Larson and Loschky (2009) used a
"Window" and "Scotoma" design (see Figure 1), to test the
contributions of central versus peripheral vision to scene
recognition. The Window condition (top rows of the right-
hand columns of Figure 1) presents central information at
various visual angles to the subjects, while the Scotoma con-
dition (second row on the right) blocks it. Using images from
10 categories, subjects were required to verify the category
in each condition. The recognition accuracy as a function of
visual angle is shown in Figure 2. They found that foveal
vision is not accurate for scene perception, while peripheral
vision is, despite its much lower resolution. However, they
also found that central vision is more efficient, in the sense
that less area is needed to achieve equal accuracy. The visual
area is equal at 10.8◦, and the crossover point, where central
vision starts to perform better than peripheral, is to the left of
that point.
Despite the common belief that central vision is important
for face and object recognition, and peripheral vision is im-
portant for scene perception shown in studies above, a careful
examination of the contribution of central versus peripheral
vision in object, scene, and face recognition is needed. In this
work, we modeled the experiment of Larson and Loschky
(2009) using deep convolutional neural networks. Further-
more, we extended the modeling work to a greater range of
stimuli, and answer the following questions: How does the
model perform as the number of scene categories is scaled
up? Besides scenes, can the model predict the importance
of central vision versus peripheral information in object and
face recognition? What is the result compared to scenes?
In the following, we show that our modeling results match
the observations of Larson and Loschky (2009), and that it
scales up to over 200 scene categories. By running a sim-
ilar analysis for large-scale object and face recognition, our
model predicts that central vision is very important for face
recognition, important for object recognition, and less impor-
tant for scene recognition. Peripheral vision, however, serves
an important role for scene recognition, but is less impor-
tant for recognizing objects and faces. Furthermore, across
all conditions we tried, central vision is more efficient than
peripheral vision on a per-pixel basis (when equal areas are
presented), which is consistent with the result of Larson and
Loschky (2009).
Method
Image Preprocessing
To create foveated images, we preprocessed the images using
the Space Variant Imaging System1. To mimic human vision,
we set the parameter that specifies the eccentricity at which
resolution drops to half of the fovea to 2.3◦. Example images
and their preprocessed retinal versions are shown in the first
and second columns of Figure 1.
As in the experiments of Larson and Loschky (2009), we
used the Window and Scotoma paradigms as specified by van
Diepen, Wampers, and dYdewalle (1998) to process the input
stimulus. The idea of both paradigms is to evaluate the value
of missing information - if the missing information is needed,
then the perception process may be disrupted and recognition
performance may drop; if the missing information is not nec-
essary, then the processing remains normal.
Input images in our experiments are 256× 256 pixels, and
we assume that corresponds to 27◦ × 27◦ of visual angle,
the number in (Larson & Loschky, 2009).
In (Larson &
Loschky, 2009), they used four sets of radius conditions for
Windows and Scotomas: 1◦ represents the presence or ab-
sence of foveal vision; 5◦ represents the presence or absence
of central vision; 10.8◦ presents equal viewable area inside
the Windows or outside the Scotomas; 13.6◦ presents more
viewable area in the Windows than the Scotomas. In order
make the prediction of the model more accurate, we added
five additional radius conditions in all of our experiments:
1http://svi.cps.utexas.edu/software.shtml
Figure 1: Examples of images used in our experiment. First
column: original images. Second column: foveated images.
Third to last column: images processed through "Window"
and "Scotoma" conditions with different radii in degrees of
visual angle.
3◦,7◦,9◦,12◦, and 16◦. The example Window and Scotoma
images are shown in Figure 1.
Deep Convolutional Neural Networks (CNNs)
Deep CNNs are neural networks with many layers that stack
computations in a hierarchical way, repeatedly performing:
1) 2-dimensional convolutions over the stimulus generated
from previous layers using learned filters, which are con-
nected locally to a small subregion of the visual field; 2) a
pooling operation on local regions of the feature maps ob-
tained from convolution operation, which is used to reduce
the dimensionality and gain translational invariance; 3) non-
linearities to the upstream response, which is used to gener-
ate more discriminative features useful for the task. As layers
go higher, the receptive fields of filters are generally larger,
and the learned features go from low-level (edges, contours)
to high-level object-related representations (object parts and
shapes) (Zeiler & Fergus, 2014). Several fully-connected lay-
ers are usually added on top of these computations to learn
more abstract and task-related features.
We used deep CNNs in our experiments for two reasons.
First, deep CNNs are the best models in computer vision:
they achieve the state-of-the-art performance on many large-
scale computer vision tasks, such as image classification
(Krizhevsky, Sutskever, & Hinton, 2012; He, Zhang, Ren,
& Sun, 2015), object detection (Ren, He, Girshick, & Sun,
2015), and scene recognition (Zhou, Lapedriza, Xiao, Tor-
ralba, & Oliva, 2014). Thus, the models should achieve de-
cent performance in our experiments. Smaller networks or
other algorithms are not competent for our tasks. Second,
deep CNNs have been shown to be the best models of the vi-
sual cortex: they are able to explain a variety of neural data
in human and monkey IT (Yamins et al., 2014; Guc¸lu & van
Gerven, 2015; Wang, Malave, & Cipollini, 2015). As a result,
it is natural to use them in our work modeling a behavioral
study related to human vision.
Experiments
In this section, we first describe our model of the behavioral
study of Larson and Loschky (2009). We then introduce the
experiment for measuring the contribution of central versus
peripheral vision for large-scale scene, object, and face recog-
nition tasks.
Modeling Larson and Loschky (2009)
In Larson and Loschky (2009), scene recognition accuracy
was measured across 100 human subjects on 10 categories:
Beach, Desert, Forest, Mountain, River, Farm, Home, Mar-
ket, Pool, and Street. For each trial in the Windows and Sco-
tomas conditions, subjects were first presented a scene im-
age, and then were asked to press "yes" or "no" for the cue
(category name) presented on the screen. Their experimental
result is summarized in Figure 2. They showed that central
vision (5◦ window condition) performs less well than periph-
eral vision in terms of getting maximum recognition perfor-
mance. They further demonstrated the peripheral advantage
is due to more viewing areas in the Scotomas conditions, and
central vision is more privileged when given equal viewable
areas (10.8◦).
We obtained the stimuli of the above 10 categories from
the Places205 database (Zhou et al., 2014), which contains
205 scene categories and 2.5 million images. All input stim-
uli were preprocessed using the retina model described in the
above section. As 10 categories is small and can easily lead
to overfitting problems in training deep CNNs, we trained
our recognition model by performing fine-tuning (or transfer
learning) based on pretrained models. The model pretrained
on the Places205 database can be treated as a mature scene
recognition pathway, and fine-tuning can be thought as addi-
tional training for the task. To investigate whether different
network architectures, especially depth, have different impact
on the modeling result, we applied three different pre-trained
models, namely:
1. AlexNet (Krizhevsky et al., 2012): A network with 5 con-
volutional layers and 3 fully connected layers, about 60
million trainable parameters. Achieved 81.10% top-5 ac-
curacy on the Places205 validation set.
2. VGG-16 (Simonyan & Zisserman, 2014): A network with
13 convolutional layers and 3 fully connected layers, about
138 million trainable parameters. Achieved 85.41% top-5
accuracy on the Places205 validation set.
3. GoogLeNet (Szegedy et al., 2015): A network with 21
convolutional layers and 1 fully connected layer, about 6.8
million trainable parameters. Achieved 87.70% top-5 ac-
curacy on the Places205 validation set.
For all models, the fine-tuning process starts by keeping the
weights except for the last fully connected layer intact, and
Figure 2: Results for scene recognition accuracy as a func-
tion of viewing condition (Windows (w) and Scotomas (s))
and visual angle. Left: result of Larson and Loschky (2009).
Right: our modeling result.
initializing the weights of the last layer to be random with
zero mean and unit variance. To be compatible with the "yes"
or "no" condition in the behavioral experiment, we replaced
the last layer in the networks with a single logistic unit, and
trained the networks for each of the 10 object categories sep-
arately, using half of the training images from the target cat-
egory and half randomly selected from all other 9 categories.
As the last layer needs more learning, we set the learning rate
of the last layer to 0.001, and all previous layers to 1e−4. The
training set of the 10 scene categories contains a total num-
ber of 129,210 full resolution images, and we trained all net-
works using minibatch stochastic gradient descent with batch
size from 32 to 256, using the Caffe deep learning framework
(Jia et al., 2014) on NVIDIA Titan Black 6GB GPUs. All
networks were trained for a maximum number of 24,000 it-
erations to ensure convergence. Each test set contains 200
images (100 from target category and 100 from all other cat-
egories), and the label distribution is the same as the training
set.. Test images were preprocessed to meet each of the Win-
dows and Scotomas condition. We tested the performance of
the fine-tuned models on all conditions by reporting the mean
classification accuracy, which is shown in Figure 2.
From Figure 2, we can clearly see our result for all
three models qualitatively matches the result of Larson and
Loschky (2009). First, for Window and Scotoma conditions,
an increasing radius of visual angle (x axis) yields a mono-
tonic increase or decrease in classification accuracy (y axis).
The sharper increase from 1◦ to 5◦ in the behavioral study
may be due to the higher efficiency of human central vision.
Second, we replicated the fact that central vision (less than
5◦) is less useful than peripheral vision in terms achieving the
best scene recognition performance. Third, however, when
using equal viewable areas (10.8◦), central vision performs
better than peripheral, exhibiting higher efficiency. Fourth,
the critical radius (the crossover point where the two con-
ditions produce equal performance, see Figure 2b) is 8.26◦
(averaged across all models), which is within the 8.22◦-9.24◦
range reported by Larson and Loschky (2009). This suggests
our models are quite plausible.
Figure 3: Results for large-scale scene recognition accuracy
as a function of viewing condition (Windows (w) and Sco-
tomas (s)) and visual angle. Softmax output is used instead
of logistic unit, so chance is 0.005. Left: experiment using
original images. Right: experiment using foveated images.
Figure 4: Results for large-scale object recognition accuracy
as a function of viewing condition (Windows (w) and Sco-
tomas (s)) and visual angle. Softmax output is used instead
of logistic unit, so chance is 0.001. Left: experiment with
original images. Right: experiment with foveated images.
When comparing the performance across the three models
we use, we cannot find a notable difference in terms of perfor-
mance, though GoogLeNet usually performs slightly better,
indicating that depth of processing might be the key factor in
obtaining better performance.
Large-Scale Scene, Object, and Face Recognition
The above modeling work is based on a scene recognition
task using 10 categories.
In real life, however, there are a
much larger number of scene categories. Beyond scenes, gen-
eral object recognition and face recognition are the two most
important recognition tasks that are performed regularly. The
relative importance of central versus peripheral vision among
the three categories needs to be examined carefully. Using a
similar modeling approach, we describe our findings in large-
scale scene, object, and face recognition in the sections below.
Scene Recognition We used all 205 categories in the
Places205 dataset. The trained models of AlexNet, VGG-16,
and GoogLeNet are deployed to examine the recognition ac-
curacy on the Place205 validation set, which contains 20,500
images, in all Windows and Scotoma conditions. In addition,
we tested the models using images both processed and un-
processed by the retina model to examine the generalization
power of the learned features. The result is shown in Figure 3.
From Figure 3, we can see the general trend that we ob-
served in Figure 2 still holds: peripheral vision is more im-
portant than central vision, but central vision is more efficient.
All models behave similarly. However, we can see the perfor-
mance on images preprocessed through the retina model is
inferior. Apparently, since there are many more categories in
this experiment, the foveation has more of an effect. Recall
that the models are trained using images with full resolution;
missing the peripheral information may the cause learned fea-
tures to imperfectly generalize.
Object Recognition We ran our object recognition experi-
ment on the ILSVRC 2012 dataset (Russakovsky et al., 2015),
which contains 1000 object categories and over 1.2 million
training images. We used the pretrained models of AlexNet,
VGG-16, and GoogLeNet, which achieve top-5 accuracy of
80.13%, 88.44%, and 89.00%, respectively, on the ILSVRC
2012 validation set. Similar to scene recognition, we tested
all models under all Windows and Scotoma conditions, us-
ing original and foveated images. The results are shown in
Figure 4.
At the first glance of looking at Figure 4, we may draw the
conclusion that the result is the same as scene recognition:
central vision is still more important than peripheral vision.
However, when we compare the scene and object recogni-
tion results (shown in Figure 5), we can clearly see that cen-
tral information in object recognition is more important than
that in scene recognition: the accuracy of the Scotoma con-
ditions drops much faster for object recognition than scene
recognition as visual angle increases from 1◦ to 7◦, suggest-
ing that losing central vision causes a greater impairment for
object recognition performance. This is consistent with our
knowledge that central vision plays a more important role in
object recognition than scenes, as there are more high spa-
tial frequency details in objects than scenes. Another find-
ing from this experiment is that AlexNet (8 layers) performs
much worse than VGG-16 (16 layers) and GoogLeNet (23
layers), suggesting that depth is important to produce good
performance.
Face Recognition We performed the face recognition ex-
periment on the Labeled Faces in the Wild (LFW) dataset
(Huang, Ramesh, Berg, & Learned-Miller, 2007), which con-
tains 13,233 labeled images from 5,749 individuals. As there
is only 1 image for some identities, researchers usually pre-
train their network on larger datasets (not publicly available)
and test their models on the LFW dataset. In this experiment,
we tested three pretrained models, namely Lighten-A (10 lay-
ers; (Wu, He, & Sun, 2015)), Lighten-B (16 layers), and
VGG-Face (16 layers;(Parkhi, Vedaldi, & Zisserman, 2015)),
Figure 5: Comparison results for scene and object recognition
using the VGG-16 model. Losing central vision decreases
performance for object recognition more quickly than scene
recognition. Left: original images. Right: foveated images.
on the face verification task for the LFW dataset, where they
achieve accuracy of 90.33%, 92.37%, and 96.23%, respec-
tively. Face images were preprocessed so that they occupy
the entire visual field (Figure 1). Same as the previous ex-
periments, we tested all models using Windows and Scotoma
conditions, with original and foveated images. Results are
shown in Figure 6.
We see very different performance in Figure 6 compared
to object and scene recognition. First, central information
is obviously much more important than peripheral informa-
tion for face recognition, given the accuracy at 5◦ is much
higher for the Window condition than the Scotoma condition
for Lighten models, and very similar with each other for the
VGG model. This is consistent with our intuition that face
recognition is a fine-grained discrimination process. Second,
the Window performance grows much more slowly after 7◦,
suggesting the more peripheral region provides little addi-
tional information for recognizing faces, unlike objects and
scenes, which needs lots of peripheral information to obtain
the maximal accuracy. Third, the foveated images produce
nearly identical results as the original image, demonstrating
that face recognition only involves central vision, and the
blurred peripheral vision is not needed.
Finally, as central vision appears to be more efficient (on a
per-pixel basis) than peripheral vision in all experiments we
tried, we tested the relative efficiency of the central vision
over peripheral vision by measuring the recognition accuracy
as a function of viewable area. The result is shown in Fig-
ure 7.
From Figure 7, we can clearly see that the recognition ac-
curacy of central vision is always superior than peripheral
vision for all tasks. However, central vision is even more
efficient when recognizing faces than recognizing objects or
scenes, as viewable areas over 50% of the whole image can
only provide a limited boost for face recognition, while sig-
nificantly improving the accuracy of object and scene recog-
nition. Contrarywise, peripheral information provides little to
no help for face recognition, unless over 90% of the image is
Figure 6: Results for large-scale face recognition accuracy
as a function of viewing condition (Windows (w) and Sco-
tomas (s)) and visual angle. Left: experiment with origi-
nal images. Right: experiment with foveated images. For
Lighten-A and Lighten-B models, the visual angle only ex-
pands to 9.5◦, as the input image is smaller (144× 144) than
for the VGG model (256×256). The accuracy for face verifi-
cation task is measured as the true positive rate at Equal Error
Rate (EER) point on the ROC curve. Chance is 0.5.
presented, but the accuracy still suffers due to the loss of cen-
tral vision. However, peripheral information is important for
object and scene recognition (and more important for scene
recognition, as shown in Figure 5).
These large-scale scene, object, and face recognition mod-
eling results suggest there is an order of relative importance
of central versus peripheral vision in those tasks: peripheral
vision is most important for scene recognition, less impor-
tant for object recognition, and basically not helpful for face
recognition. Central vision, however, plays a crucial role in
face recognition, is important for object recognition, and is
less important for scene recognition.
Conclusion
In this paper, we modeled the contribution of central ver-
sus peripheral visual information for scene, object, and face
recognition, using deep CNNs. We first modeled the behav-
ioral study of Larson and Loschky (2009), and replicated their
findings of the importance of peripheral vision in scene recog-
nition. In addition, by running a large-scale scene, object, and
face recognition simulation, our models make testable predic-
tions for the relative order of importance for central versus
peripheral vision for those tasks.
Acknowledgments
This work was supported by NSF grants IIS-1219252 and
SMA 1041755 to GWC. PW was supported by a fellowship
from Hewlett-Packard.
References
Epstein, R., Harris, A., Stanley, D., & Kanwisher, N. (1999).
The parahippocampal place area: Recognition, navigation,
or encoding? Neuron, 23(1), 115–125.
of central versus peripheral vision to scene gist recognition.
Journal of Vision, 9(10), 6.
Loschky, L., McConkie, G., Yang, J., & Miller, M. (2005).
The limits of visual resolution in natural scene viewing. Vi-
sual Cognition, 12(6), 1057–1092.
Malach, R., Levy, I., & Hasson, U.
(2002). The topogra-
phy of high-order human object areas. Trends in cognitive
sciences, 6(4), 176–184.
McCandliss, B. D., Cohen, L., & Dehaene, S. (2003). The
visual word form area: expertise for reading in the fusiform
gyrus. Trends in cognitive sciences, 7(7), 293–299.
Nelson, W. W., & Loftus, G. R. (1980). The functional vi-
sual field during picture viewing. Journal of Experimental
Psychology: Human Learning and Memory, 6(4), 391.
Parkhi, O. M., Vedaldi, A., & Zisserman, A. (2015). Deep
face recognition. In British machine vision conference.
Polyak, S. L. (1941). The retina.
Ren, S., He, K., Girshick, R., & Sun, J.
(2015). Faster r-
cnn: Towards real-time object detection with region pro-
In Advances in neural information pro-
posal networks.
cessing systems (pp. 91–99).
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,
Ma, S., . . . Fei-Fei, L.
ImageNet Large
Scale Visual Recognition Challenge. International Journal
of Computer Vision (IJCV), 1-42. doi: 10.1007/s11263-
015-0816-y
(2015, April).
Simonyan, K., & Zisserman, A. (2014). Very deep convo-
lutional networks for large-scale image recognition. arXiv
preprint arXiv:1409.1556.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,
(2015, June). Going
Anguelov, D., . . . Rabinovich, A.
deeper with convolutions..
van Diepen, P. M., Wampers, M., & dYdewalle, G. (1998).
Functional division of the visual field: Moving masks and
moving windows. Eye guidance in reading and scene per-
ception, 337–355.
Wandell, B. A. (1995). Foundations of vision. Sinauer Asso-
ciates.
Wang, P., Malave, V., & Cipollini, B.
(2015). Encoding
voxels with deep learning. The Journal of Neuroscience,
35(48), 15769–15771.
Wu, X., He, R., & Sun, Z. (2015). A lightened cnn for deep
face representation. arXiv preprint arXiv:1511.02683.
Yamins, D. L., Hong, H., Cadieu, C. F., Solomon, E. A., Seib-
ert, D., & DiCarlo, J. J. (2014). Performance-optimized hi-
erarchical models predict neural responses in higher visual
cortex. Proceedings of the National Academy of Sciences,
111(23), 8619–8624.
Zeiler, M. D., & Fergus, R. (2014). Visualizing and under-
standing convolutional networks. In Computer vision–eccv
2014 (pp. 818–833). Springer.
Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., & Oliva, A.
(2014). Learning deep features for scene recognition using
In Advances in neural information pro-
places database.
cessing systems (pp. 487–495).
Figure 7: Accuracy for object (left), scene (middle) and face
(right) recognition as a function of the percentage of view-
able area presented under Window (blue) and Scotoma (red)
conditions, using original (solid line) and foveated images
(dashed line).
Gomez, J., Pestilli, F., Witthoft, N., Golarai, G., Liberman,
A., Poltoratski, S., . . . Grill-Spector, K.
(2015). Func-
tionally defined white matter reveals segregated pathways
in human ventral temporal cortex associated with category-
specific processing. Neuron, 85(1), 216–227.
Grill-Spector, K., & Malach, R. (2004). The human visual
cortex. Annu. Rev. Neurosci., 27, 649–677.
Guc¸lu, U., & van Gerven, M. A. (2015). Deep neural net-
works reveal a gradient in the complexity of neural repre-
sentations across the ventral stream. The Journal of Neuro-
science, 35(27), 10005–10014.
He, K., Zhang, X., Ren, S., & Sun, J.
(2015). Deep
residual learning for image recognition. arXiv preprint
arXiv:1512.03385.
Henderson, J. M., & Hollingworth, A. (1999). The role of fix-
ation position in detecting scene changes across saccades.
Psychological Science, 10(5), 438–443.
Holmes, D. L., Cohen, K. M., Haith, M. M., & Morrison,
F. J. (1977). Peripheral visual processing. Perception &
Psychophysics, 22(6), 571–577.
Huang, G. B., Ramesh, M., Berg, T., & Learned-Miller, E.
(2007). Labeled faces in the wild: A database for study-
ing face recognition in unconstrained environments (Tech.
Rep. No. 07-49). University of Massachusetts, Amherst.
Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J.,
Girshick, R., . . . Darrell, T. (2014). Caffe: Convolutional
architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093.
Kanwisher, N., McDermott, J., & Chun, M. M. (1997). The
fusiform face area: a module in human extrastriate cor-
tex specialized for face perception. The Journal of Neu-
roscience, 17(11), 4302–4311.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Im-
agenet classification with deep convolutional neural net-
works. In Advances in neural information processing sys-
tems (pp. 1097–1105).
Larson, A. M., & Loschky, L. C. (2009). The contributions
|
1702.08617 | 1 | 1702 | 2017-02-28T03:03:40 | Spontaneous Activity in the Visual Cortex is Organized by Visual Streams | [
"q-bio.NC"
] | Large-scale functional networks have been extensively studied using resting state functional magnetic resonance imaging. However, the pattern, organization, and function of fine-scale network activity remain largely unknown. Here we characterized the spontaneously emerging visual cortical activity by applying independent component analysis to resting state fMRI signals exclusively within the visual cortex. In this sub-system scale, we observed about 50 spatially independent components that were reproducible within and across subjects, and analyzed their spatial patterns and temporal relationships to reveal the intrinsic parcellation and organization of the visual cortex. We found that the visual cortical parcels were aligned with the steepest gradient of cortical myelination, and organized into functional modules segregated along the dorsal/ventral pathways and foveal/peripheral early visual areas. In contrast, cortical retinotopy, folding, and cytoarchitecture impose limited constraints to the organization of resting state activity. From these findings, we conclude that spontaneous activity patterns in the visual cortex are primarily organized by visual streams, likely reflecting feedback network interactions. | q-bio.NC | q-bio | Spontaneous Activity in the Visual Cortex is Organized by Visual Streams
Kun-Han Lu2,3, Jun Young Jeong2, Haiguang Wen2,3, Zhongming Liu*1,2,3
1Weldon School of Biomedical Engineering
2School of Electrical and Computer Engineering
3Purdue Institute for Integrative Neuroscience
Purdue University, West Lafayette, IN, USA
*Correspondence
Zhongming Liu, PhD
Assistant Professor of Biomedical Engineering
Assistant Professor of Electrical and Computer Engineering
College of Engineering, Purdue University
206 S. Martin Jischke Dr.
West Lafayette, IN 47907, USA
Phone: +1 765 496 1872
Fax: +1 765 496 1459
Email: [email protected]
1
Abstract
Large-scale functional networks have been extensively studied using resting state functional magnetic
resonance imaging. However, the pattern, organization, and function of fine-scale network activity remain
largely unknown. Here we characterized the spontaneously emerging visual cortical activity by applying
independent component analysis to resting state fMRI signals exclusively within the visual cortex. In this
sub-system scale, we observed about 50 spatially independent components that were reproducible within
and across subjects, and analyzed their spatial patterns and temporal relationships to reveal the intrinsic
parcellation and organization of the visual cortex. We found that the visual cortical parcels were aligned
with the steepest gradient of cortical myelination, and organized into functional modules segregated along
the dorsal/ventral pathways and foveal/peripheral early visual areas. In contrast, cortical retinotopy,
folding, and cytoarchitecture impose limited constraints to the organization of resting state activity. From
these findings, we conclude that spontaneous activity patterns in the visual cortex are primarily organized
by visual streams, likely reflecting feedback network interactions.
Keywords:
Fine-scale networks; independent component analysis; functional parcellation; visual streams.
2
Introduction
Resting state functional magnetic resonance imaging (rs-fMRI) has been widely explored to map
resting state networks (RSNs) that collectively report the brain's intrinsic functional organization (Fox
and Raichle, 2007). These networks consist of temporally correlated regions (Biswal et al., 1995; Van
Dijk et al., 2010), arise from structural connections (Honey et al., 2007; Van Den Heuvel et al., 2009),
resemble and predict task activations (Smith et al., 2009; Cole et al., 2014; Tavor et al., 2016), and
distinguish individual subjects (Finn et al., 2015) or diseases (Fox and Greicius, 2010). Although brain
activity spans a variety of spatial scales (Yoshimura et al., 2005; Doucet et al., 2011; Hutchison et al.,
2013), RSNs have been mostly characterized in the whole brain, where their functions are empirically and
coarsely defined in terms of motor (Biswal et al., 1995), vision (Yeo et al., 2011), default-mode (Greicius
et al., 2003), attention (Fox et al., 2006), salience (Seeley et al., 2007), and so on. In finer spatial scales,
patterns of spontaneous activity and connectivity remain unclear, but may bear more specific functional
roles related to perception, behavior, or cognition (Kenet et al., 2003; Yoshimura et al., 2005; Wang et al.,
2013; Long et al., 2014; Wilf et al., 2015; Lewis et al., 2016).
In this regard, the visual cortex is a rich and ideal benchmark. It has been characterized in terms
of cellular architectonics (Amunts et al., 2000; Glasser and Van Essen, 2011), structural connections
(Felleman and Van Essen, 1991; Salin and Bullier, 1995), cortical folding (Fischl et al., 2008; Benson et
al., 2012), functional pathways (Hubel and Wiesel, 1962; Ungerleider and Haxby, 1994), and neural
coding (Ohiorhenuan et al., 2010; Guclu and van Gerven, 2015). In recent studies, correlation patterns of
spontaneous activity in the visual cortex have been examined and compared with retinotopy (Heinzle et
al., 2011; Yeo et al., 2011; Jo et al., 2012; Butt et al., 2013; de Zwart et al., 2013; Gravel et al., 2014;
Raemaekers et al., 2014; Arcaro et al., 2015; Bock et al., 2015; Striem-Amit et al., 2015; Wilf et al., 2015;
Dawson et al., 2016; Genc et al., 2016; Lewis et al., 2016). Since much of the visual cortex has visual-
field maps – topographic representations of the polar angle and eccentricity (Wandell et al., 2007), it is
reasonable to initially hypothesize that resting state activity is also intrinsically organized by retinotopy.
3
Support for this hypothesis comes in part from the finding that in early visual areas (e.g. V1-V3),
correlations in spontaneous activity are generally higher between locations with similar eccentricity
representations (Heinzle et al., 2011; Yeo et al., 2011; Jo et al., 2012; Gravel et al., 2014; Arcaro et al.,
2015; Striem-Amit et al., 2015; Dawson et al., 2016; Genc et al., 2016; Lewis et al., 2016). However,
such eccentricity-dependent functional connectivity has rarely been reported beyond early visual areas
(Baldassano et al., 2013; Striem-Amit et al., 2015); it may be relatively weak (Wilf et al., 2015) or
observable only after regressing out large-scale activity (Raemaekers et al., 2014); and it may also be
confounded by the decay of local connectivity over cortical distance (Butt et al., 2013; Dawson et al.,
2016). Further against this hypothesis, spontaneous activity preserves little or less polar-angle dependence
in intra-hemispheric correlations (Gravel et al., 2014; Bock et al., 2015; Wilf et al., 2015), and shows
strong correlations between bilateral V1 locations despite their lack of common receptive fields or direct
connections (Jo et al., 2012; de Zwart et al., 2013). As such, the retinotopic organization of fine-scale
resting activity in the visual cortex is questionable (Butt et al., 2013; Wilf et al., 2015). The previously
reported eccentricity-dependent spontaneous activity and connectivity may be attributed to alternative
representations that partly overlap with the eccentricity representation in early visual areas, as opposed to
retinotopy per se.
In fact, cortical representations of the peripheral and central visual fields partly overlap with the
magnocellular and parvocellular streams (Schiller et al., 1990; Nassi and Callaway, 2009), and extend
onto the dorsal and ventral pathways for visual action and perception, respectively (Goodale and Milner,
1992; Ungerleider and Haxby, 1994). Along these pathways, feedforward neuronal circuits convey and
integrate not only visual positions, but also increasingly complex visual or conceptual features (Martin,
2007; Hasson et al., 2008; Yamins and DiCarlo, 2016). Top-down feedback connections are not or less
retinotopically organized than are feedforward pathways (Salin and Bullier, 1995). The complex interplay
between feedforward and feedback processes is essential for natural vision (Rao and Ballard, 1999;
Gilbert and Li, 2013), but remains largely unclear in a stimulus-free resting state. Thus, spontaneously
4
emerging networks in the visual cortex may not readily fit the retinotopic organization, or arguably any
other presumable organizations.
What is needed is a data-driven analysis of resting-state activity in a finer spatial scale, unbiased
by any presumed areal definition or organizational hypothesis. For this purpose, independent component
analysis (ICA) is well suited but has not been applied to finer spatial scales, to our knowledge, despite its
wide application to whole-brain fMRI signals (Damoiseaux et al., 2006). Unlike the correlation analysis
(Baldassano et al., 2012; Genc et al., 2016), ICA also has the advantages of being multivariate and data-
driven, thereby bypassing the potential bias from any narrowly-focused hypothesis (Calhoun et al., 2009).
Here, we explored a new application of ICA for mapping cortical visual areas and networks based on rs-
fMRI signals within the visual cortex. In this sub-system scale, the fine-grained activity patterns derived
from ICA were systematically characterized, interpreted, and evaluated for their test-retest reproducibility
and individual variations. Moreover, they were also compared against cortical folding (Destrieux et al.,
2010), retinotopy (Abdollahi et al., 2014), cytoarchitecture (Van Essen et al., 2012a), myeloarchitecture
(Glasser et al., 2014), and the latest multimodal cortical parcellation (Glasser et al., 2016). Clustering
analysis further reveals that spontaneously emerging network patterns in the human visual cortex are not
retinotopically organized; instead, they are temporally clustered into three functional modules: the dorsal
pathway, the ventral pathway, and the foveal and peripheral sub-divisions of early visual areas.
Materials and Methods
Subjects and Data
We used the rs-fMRI data released from the Human Connectome Project (HCP) (Van Essen et
al., 2013). Briefly, we randomly selected 201 independent healthy subjects; for each subject, we used the
5
data from two resting state sessions (session 1 for test and session 2 for retest); each session was 14
minutes and 33 seconds with the eyes open and fixated.
As elaborated elsewhere (Van Essen et al., 2012b), data were acquired in a 3-tesla MRI system
with a 32-channel head coil (Skyra, Siemens, Germany). The rs-fMRI data we used were acquired with a
single-shot, multiband-accelerated, gradient-recalled echo-planar imaging with nominally 2mm isotropic
spatial resolution and 0.72s temporal resolution, and left-to-right phase-encoding. In addition, structural
images with T1 and T2-weighted contrast were both acquired with 0.7mm isotropic resolution.
As elaborated elsewhere (Glasser et al., 2013), the structural images were non-linearly registered
to the Montreal Neurological Institute (MNI) template, where the images were combined and segmented
to generate cortical surfaces. The fMRI images were corrected for slice timing and motion, aligned to
structural images, normalized to the MNI space, projected onto the cortical surfaces, and co-registered
across subjects. In addition to the minimal preprocessing described above, we removed the slow trend in
the fMRI time series by regressing out a fourth-order polynomial function, and subtracted the mean and
standardized the signal variance. Note that spatial smoothing was not performed to minimize spurious
correlations in neighboring voxels.
Independent Component Analysis
We applied the ICA to the rs-fMRI signals within a mask of the visual cortex (Fig. 1.A), defined
by a system-level functional parcellation of the human cortex (Yeo et al., 2011). The fMRI time series
within the mask was temporally standardized and concatenated across subjects. Infomax ICA (Bell and
Sejnowski, 1995) was used to decompose the concatenated rs-fMRI data into 59 spatially independent
components within the visual cortex. Here, the number of independent components, 59, was determined
by maximizing the Laplace approximation of the posterior probability of the ICA model order (Beckmann
and Smith, 2004). The spatial pattern of each component was converted to a z-score map by dividing the
IC weight at each voxel by the standard deviation of voxel-wise residual noise that could not be explained
6
by the ICA model (Beckmann and Smith, 2004). All color scale for displaying the spatial IC maps in this
study are z-scores.
Reproducibility test
Furthermore, we evaluated the test-retest reproducibility of the ICA results. In a group level, the
ICA applied to the data in one session (i.e. session 1) was also applied to the data in a repeated session
(i.e. session 2); both sessions were from the same group of subjects, and the data was concatenated in the
same order across subjects. We calculated the absolute value of the spatial correlation (r) between every
component from session 1 and every component from session 2, and paired these components across
sessions into distinct pairs to maximize the sum of their absolute spatial correlations. More specifically,
we used an iterative procedure toward the optimal pairing: we began with identifying a pair of ICs (i.e.
one from session1, and the other from session2) with the highest correlation; then we paired these two
ICs, and excluded them from subsequent pairing, which continued until all ICs were paired. Note that the
absolute spatial correlation was used since a reproducible IC could show the same spatial distribution
despite opposite polarity. A threshold (r ≥ 0.4) was used to identify reproducible components for further
interpretation (Fig. 1.C).
We also explored the potential confounding effects of head motion on fine-scale independent
components. For each subject, we regressed out the time series of six motion correction parameters prior
to group-level ICA. We compared the ICs obtained without and with the above motion correction, and
further identified and excluded those ICs that appeared inconsistent solely due to this preprocessing step.
Modularity analysis
For all the ICs that were reproducible and unaffected by head motion correction, we computed the
temporal correlations between different components. Such correlations were first calculated based on
component time series from each subject, and then averaged across subjects. This procedure prevented the
7
resulting correlations from being dominated or biased by inter-subject variations. To evaluate the
statistical significance of the between-component correlation, we converted the correlation coefficient to
the z-score using the Fisher's r-to-z transform separately for each subject, and then applied a one-sample
t-test (with dof=200, significance level at 0.01, and Bonferroni correction) to the z-scores from all the
subjects.
We also applied the Louvain modularity analysis (Blondel et al., 2008) to the cross-component
correlation matrix averaged across subjects. It assigned individual ICs to different modules, such that the
temporal correlations were higher within modules but lower between modules. The modularity analysis,
including the determination of the number of modules, was conducted with the algorithm (a Matlab
function: modularity_louvain_und_sign) implemented in Brain Connectivity Toolbox (Rubinov and
Sporns, 2010). The modularity index, Q, which quantified the goodness of modularity partitions (ranged
from 0 to 1), was obtained. To evaluate the statistical significance of Q, we randomly shuffled the values
within the correlation matrix 10,000 times and computed Q for each permutation given the same module
assignment. This generated a null distribution of Q, against which the p value was computed for the Q
value without permutation. To visualize the modular organization on the cortical surface, we represented
each IC by a sphere located at the peak location in the component map, and color-coded every IC by its
module membership.
Computation of cortical distance
To examine the effect of cortical distance on temporal correlations between components, we first
identified the peak location in each component map (hereafter referred to as the component centroid). For
bilaterally-distributed components, two centroids were identified: one at each hemisphere. We computed
the geodesic cortical distances along the cortical surface between different component centroids by using
HCP's "-surface-geodesic-distance" function. The cortical distances were computed for each subject and
for each hemisphere, and then averaged across subjects.
8
The relationship between the temporal correlations and the corresponding cortical distances was
examined and fitted by a rational function based on least-squares estimations (the "lsqcurvefit" function
in Matlab). Separately for each hemisphere, we defined the spatial affinity between components as the
reciprocal of the cortical distances between the corresponding centroids. Then we applied the Louvain
modularity analysis to the spatial-affinity matrix in the same way as for the modularity analysis of
functional connectivity. We further compared such anatomically-defined modules with the functional
modules obtained on the basis of the temporal correlations between components.
Intrinsic functional parcellation of the visual cortex
From all the ICs, we created a group-level intrinsic functional parcellation of the visual cortex
with an increasing level of granularity. Specifically, we defined a feature vector for each voxel in the
visual cortex. This feature entails the weights by which the individual time series of different ICs were
linearly combined to explain the fMRI signal observed at each voxel. Then we grouped the cortical voxels
into distinct parcels by applying the k-means clustering to the corresponding feature vectors using a
correlation-based "distance" and 1,000 replications with random initialization. The number of clusters (k)
was empirically set to 10, 20, 30, 40. The parcels were color-coded from 0 to 1 in an ascending order
according to the averaged "distance" within each parcel. We preferred this k-means clustering analysis to
a "winner-take-all" alternative, in which each voxel was assigned to only one IC with the greatest weight
(among all the ICs) at the given voxel. This was because single voxel time series were not necessarily
represented by only one IC, but instead often by a few ICs.
To facilitate interpretation, the ICA components and the parcellation derived from them were
compared against conventional visual areas or cortical parcellation based on various structural and/or
functional properties, including myeloarchitecture (Glasser et al., 2014), cytoarchitecture (Eickhoff et al.,
2005), cortical folding (Destrieux et al., 2010), retinotopic mapping (Abdollahi et al., 2014), and
multimodal parcellation (Glasser et al., 2016).
9
Dual-regression and individual-level parcellation
Following group ICA, we also used dual regression (Filippini et al. 2009) against each subject's
fMRI data to characterize subject-specific ICA maps (Tavor et al. 2016). Briefly, we first applied multiple
regression to the spatial domain, using the group-level ICA spatial maps as a set of spatial regressors to
obtain individual time series that was associated with each group-level spatial map based on the subject-
specific fMRI data; after normalizing these individual-level time series to a zero mean and a unitary
variance, we applied multiple regression to the time domain, by using the normalized individual time
series as temporal regressors to obtain the subject-specific ICA spatial maps. From the individual-level
ICA maps, we also used the k-means clustering analysis as mentioned above to create subject-specific
parcellation of the visual cortex, and compared it against the group-level parcellation.
Results
Intrinsic activity patterns within the visual cortex
Here, we explored a data-driven analysis of spontaneous activity confined to the human visual
cortex. In this finer sub-system scale, our goal was to characterize and map intrinsic activity patterns in
order to delineate cortical visual areas and networks independent of any task context or any presumed
organizational hypothesis. Towards this goal, we used a cortical mask (Fig. 1.A), based on a systems-
level parcellation of the entire cortex (Yeo et al., 2011), to only select rs-fMRI activity within the human
visual cortex for group ICA. The selected rs-fMRI data from 201 healthy human subjects in the Human
Connectome Project (HCP) were temporally standardized within each subject and concatenated across
subjects. The concatenated data were then decomposed into 59 spatially independent components (ICs).
Repeating this analysis with data from a different resting-state session for the same subjects allowed us to
evaluate the test-retest reproducibility of every component. For example, Fig. 1.B displays three typical
ICs that were spatially consistent (or correlated) between the two repeated sessions. By pairing the ICs
across sessions to maximize the sum of the absolute pairwise correlation coefficients, we identified 50
10
unique pairs of reproducible components, which showed higher spatial correlations within pairs (r>0.4,
mean±S.D = 0.78±0.14) than across pairs (mean±S.D = 0.04±0.05) (Fig. 1.C). In addition, we found that
IC51 and IC55 yielded high spatial correlations with IC36 (r = 0.578) and IC41 (r = 0.63), respectively.
But they were not paired, because IC36 and IC41 were better paired with other components while the
pairing algorithm did not allow any duplication. We included IC51 and IC55 as reproducible components
in subsequent analyses. These results suggest that the ICA-derived spontaneous activity patterns are
robust and reproducible in a sub-system spatial scale.
Fig. 2 shows the spatial maps of all 52 reproducible components in a descending order of their
test-retest reproducibility. Among them, 92% showed focal patterns, and fewer ICs were distributed (IC#:
34, 35, 49, 50); 30% showed bilateral distributions (IC#: 4, 8, 17, 19, 23, 28, 30, 31, 33, 34, 35, 36, 46,
49, 50); 36% showed clearly anti-correlated patterns with well localized positivity and negativity (IC#: 2,
3, 5, 6, 7, 8, 13, 15, 17, 21, 27, 29, 30, 31, 34, 42, 49, 50). Some of these components not only aligned
with existing anatomical borders (e.g. IC36 and IC50 aligned with V1/V2 border), but also aligned with
regions with known functional properties (e.g. IC41 and IC55 both matched well with the region MT).
Next, we attempted to identify the components that might be susceptible to artifacts related to
head motion. We compared the ICs obtained without and with motion correction (i.e. regressing out head
motion correction parameters from voxel time series). Among all the components shown in Fig. 2, three
ICs (IC34, IC49, IC50) with distributed patterns did not match to any of the ICs (r < 0.15) obtained after
motion correction (Supplementary Fig. 1A). Thus we attributed these ICs to head motion, and further
excluded them for subsequent analyses. All other 49 ICs in Fig. 2 were one-to-one matched to the ICs
after head motion correction, showing high spatial correlations for all matched pairs (Supplementary Fig.
1.B). In the following sections, we further segregated and interpreted these 49 ICs by comparing them to
existing visual areas or networks.
Comparing discrete ICA components with existing visual areas
11
Since the ICA-derived activity patterns mostly showed discrete regions with well-defined borders
(Fig. 2), we further compared such discrete ICs with the visual areas defined with a recently published
multi-modal parcellation (MMP) (Glasser et al., 2016). For the primary visual area (V1), four components
were found to be sharply confined to V1 (Fig. 3.A). IC8 matched the bilateral foveal representations in
V1; IC36 also showed bilateral distributions and corresponded to more peripheral representations; IC26
and IC38 showed unilateral distributions, corresponding to the most peripheral part of the right and left
visual fields, respectively; these components did not overlap each other and all aligned with the V1
border. Therefore, V1 consists of multiple intrinsic functional sub-divisions apparently organized
according to eccentricity representations, being largely symmetric not only between the left and right
hemispheres, but also between the upper and lower sides of the calcarine sulcus. Unlike V1, V2 or V3 did
not confine any component within itself. Instead, multiple components spanned across V2 and V3 along
either the dorsal (IC18, IC30) or ventral (IC15 and IC9) direction (Fig. 3.B). Compared to those
components within V1, the V2/V3 components were less bilaterally symmetric; none of them included
regions in both dorsal and ventral pathways.
Beyond those in early visual areas (V1/V2/V3), other discrete components were all anatomically
split by the dorsal-ventral division. In the dorsal pathway, some components matched well with existing
visual areas (Fig. 4A), including the middle temporal (MT: IC41 on the right hemisphere, IC55 one the
left hemisphere), caudal area of inferior parietal cortex (PGp: IC19), dorsal visual transitional area (DVT:
IC51 on the right hemisphere, IC39 on the left hemisphere) and parieto-occipital sulcus area 2(POS2:
IC4). Some other components were distributed across multiple visual areas, including the third visual area
and the area intraparietal 0 (V3B/IP0: IC44 on the right hemisphere, IC20 on the left hemisphere), third
visual areas and the fourth visual area (V3A/V3B/V3CD/V4: IC31), the dorsal visual transitional area and
the sixth visual area (DVT/V6A: IC28), and the third visual area, the sixth visual area and the seventh
visual area (V3A/V6A/V7: IC23); about half of these dorsal components were bilaterally symmetric.
Along the ventral pathway, the majority of the components covered multiple visual areas (Fig. 4B),
including ventral-medial visual areas (VMV1/VMV2/VMV3: IC42 on the right hemisphere, IC22 on the
12
left hemisphere), para-hippocampal areas (PHA2/PHA3: IC10 on the right hemisphere, IC16 on the left
hemisphere), fourth visual areas and the area lateral occipital 2 (V4t/LO2: IC33), the ventral visual
complex and the fusiform face complex (VVC/FFC: IC40) and VMV2/PHA2/PHA3 (IC37), except for
VMV1 (IC45); about two third of these ventral components were bilaterally symmetric; the rest of them
were unilateral.
Functional modularity in the visual cortex
For all discrete components, we further evaluated their modular organization based on their
temporal correlations. The between-component correlation matrix was calculated for every subject, and
then averaged across subjects. The resulting group-level correlation matrix was re-organized into four
functional modules based on the Louvain modularity analysis (Blondel et al., 2008; Rubinov and Sporns,
2010). Components within the same module were strongly and positively correlated, whereas components
from different modules were weakly or negatively correlated (Fig. 5B, left). Visualizing the distributions
of these functional modules on the cortical surface revealed their anatomical segregation (Fig. 5A). The
first module included components, each of which was represented by one or two spheres for unilateral (or
bilateral) components, over the foveal representations of early visual areas (V1, V2, and V3), whereas the
second module was mostly distributed over the peripheral representations of early visual areas. The third
module was distributed along the dorsal pathway, and the fourth module was distributed along the ventral
pathway (Fig. 5.A). The modularity (Q=0.4966) was statistically significant (p < 0.0001, non-parametric
permutation test). The between-component correlations within every module were consistently high for
most of the subjects, yielding high t statistics, especially for the module over the foveal early visual areas
and the module over the ventral visual areas. The t statistics corresponding to the correlations across
different modules were generally low and not significant. These results lend support to the notion that
intrinsic networks within the visual cortex are organized into functional modules: the dorsal pathway, the
ventral pathway, as well as the foveal and peripheral parts of the early visual areas.
13
Effects of cortical distance on functional connectivity and modularity
The observation that the functional modules were anatomically clustered led us to ask whether the
functional relationships between individual ICs were entirely attributable to their cortical distances. To
address this question, we tried to model the temporal correlations between components as a function of
the cortical-surface distance between component centroids, separately evaluated for each hemisphere. The
scatter-plots of correlation (r) vs. distance (d) revealed reciprocal relationships, which were modeled as
r=1.432/d and r=1.521/d for left and right hemispheres, respectively, based on least-squares estimation
(Fig. 6, left). It suggests that fine-scale functional connectivity within a cortical hemisphere depends, at
least in part, on cortical distance, in line with findings from previous studies (Dawson et al., 2016; Genc
et al., 2016).
Despite this relationship, we further asked whether the modular organization based on functional
connectivity (Fig. 5) could be a result of spatial affinity, described as the reciprocal of the distance along
the cortical surface. The matrix of spatial affinity was computed for each hemisphere, and reorganized
into modules using the same modularity analysis as used for functional connectivity (Fig. 6, right). The
left hemisphere contained two anatomical modules: one covering the lateral cortical surface, and the other
covering the medial surface. The right hemisphere exhibited similar anatomical modules as the left
hemisphere, except that the medial cortical surface was sub-divided into two modules in the right
hemisphere. The modularity (Q) of the spatial affinity matrices was 0.2039 and 0.2053, which were
statistically significant (p<0.01) but notably lower than the modularity of the functional connectivity
(Q=0.4966). More importantly, the functional and anatomical modules, on the basis of temporal
correlations and spatial affinity respectively, showed different spatial distributions, as comparatively
shown in Fig. 5 and Fig. 6. Thus, spontaneous functional connectivity in the visual cortex revealed a
stronger and different modular organization and distribution than what were attainable based on spatial
affinity alone.
14
Functional parcellation of the visual cortex
Following ICA, we applied the k-means clustering to the ICA weighting vectors at individual
locations within the visual cortex, yielding an automated intrinsic parcellation of the visual cortex with a
varying level of granularity with the number of clusters (k) being 10, 20, 30, 40. We found that as the
number of clusters increased, coarser parcels were progressively sub-divided into finer parcels (Fig. 7).
We settled at k=40, which roughly matched the expected number of visual areas (Glasser et al., 2016),
and generated a set of well-defined and bilaterally symmetric parcels (Fig. 8). This parcellation based on
spontaneous activity was further compared against existing parcellations of the visual cortex, based on the
whole-brain multimodal images (i.e. MMP) (Glasser et al., 2016), visual-field maps (Abdollahi et al.,
2014), cortical folding patterns (Destrieux et al., 2010), cortical cytoarchitecture (Eickhoff et al., 2005),
and cortical myelination (Glasser et al., 2014). It was found that none of the existing parcellations
precisely agreed with the fully-automated and data-driven parcellation reported here. In particular, our
parcellation did not match with those based on cortical retinotopy, cytoarchitecture, and folding, but
matched relatively better with the cortical myelination and the MMP. In our parcellation, the outer
contour of cortical parcels that covered early visual areas tended to align well with the steep gradients of
cortical myelination. Our reported cortical parcels that covered the high-level visual areas tended to align
reasonably well with the corresponding parcels in MMP; the alignment was not one to one. Compared to
MMP, our parcellation was coarser in high-level visual areas, but finer in low-level visual areas.
Parcellation in the level of single subjects
In addition to the group-level analysis, we also applied dual regression to the dataset, in an attempt
to obtain the corresponding ICA patterns from individual subjects. The individual-level ICA patterns were
generally consistent with the group-level ICA patterns and comparable across subjects (Fig. 9.A). On the
basis of individual-specific ICs, the visual cortex could be parcellated (k = 40) for each subject, yielding
cortical parcels apparently noisy but generally similar to the group-level parcellation (Fig. 9.B).
15
Discussion
We characterized the network patterns emerging from spontaneous resting-state activity within
the human visual cortex. On the basis of such patterns and their interactions, we delineated the intrinsic
functional parcellation and organization of the visual cortex. Here we report that fine-scale intrinsic visual
cortical networks are not organized by the patterns of cortical retinotopy, folding, or cytoarchitecture, but
align with the gradient of cortical myelination, and are segregated into functional modules specific to the
ventral and dorsal visual streams.
Whole-brain vs. fine-scale functional networks
The majority of resting-state fMRI literature focuses on intrinsic functional networks in the whole
brain scale (Smith et al., 2013). Large-scale networks are supported by long-range structural connections
(Yeo et al., 2011), and reflect coarsely defined functions (Smith et al., 2009). However, in a finer scale
within the visual cortex, bi-directional structural connections co-exist over short distances (Felleman and
Van Essen, 1991; Salin and Bullier, 1995), forming the network basis of vision (Rao and Ballard, 1999).
The patterns and dynamics of fine-scale intrinsic networks may indicate how visual information is coded
in spontaneous brain activity (Kenet et al., 2003), and offer a more specific clue on the functional role of
spontaneous activity in shaping perception or behavior (Wilf et al., 2015).
To explore the topographic organization of fine-scale visual networks, it is necessary to confine
the analysis to the functional connectivity profile within the visual system. Otherwise, in the whole-brain
scale, the connectivity profile between a seed location in the visual cortex and the rest of the brain, as a
function of locations, includes mostly the remote locations that are not or non-specifically associated with
the seed location. As a result, the difference in the whole-brain connectivity profile of two distinct seed
locations becomes subtle, even if they may actually interact with different sets of vison-related brain
16
locations. Focusing on a fine scale improves the sensitivity to differentiate the topographic difference in
functional connectivity of specific interest to vision. In line with this notion, previous studies have shown
that fine-scale functional networks may be obscured by large-scale network activity (Raemaekers et al.,
2014), and thus appear to exhibit a coarse topographic organization (Nir et al., 2006; Yeo et al., 2011).
Within the visual cortex, we used data-driven ICA to explore the multivariate voxel patterns and
dynamics, instead of bivariate correlations between voxels or areas, as in previous studies (Heinzle et al.,
2011; Yeo et al., 2011; Butt et al., 2013; de Zwart et al., 2013; Gravel et al., 2014; Raemaekers et al.,
2014; Arcaro et al., 2015; Bock et al., 2015; Striem-Amit et al., 2015; Wilf et al., 2015; Dawson et al.,
2016; Genc et al., 2016). In the whole-brain scale, ICA and seed-based correlation analyses have been
shown to reveal similar network patterns (Van Dijk et al., 2010). However, structural connections are
much denser at a reduced spatial scale or distance (Bassett and Bullmore, 2006), giving rise to more
complex patterns of functional interactions. This makes ICA a more preferable method for fine-scale
network mapping.
Fine-scale visual networks are not retinotopically organized
Our data suggest that resting state networks within the visual cortex are robust in group (Fig. 1)
and individual (Fig. 9) levels, and well-organized in space (Figs. 2 through 4) and time (Figs. 5, 6). Here,
the results obtained with data-driven ICA extend the previous findings obtained by correlational analyses
of rs-fMRI signals in the visual cortex.
However, unlike some prior studies (Heinzle et al., 2011; Gravel et al., 2014; Raemaekers et al.,
2014), we did not find any evidence for the retinotopic organization of resting state activity beyond early
visual areas. Perhaps, the exception was only in V1, where activity patterns were found to agree with
eccentricity representations (Fig. 3A), consistent with previous findings (Yeo et al., 2011; Arcaro et al.,
2015; Wilf et al., 2015). Beyond V1, spontaneous activity patterns were independent of either the
eccentricity or the polar angle (Fig. 3). Even in V1, resting state activity was correlated between the left
17
and right hemispheres (Fig. 3), although the two hemispheres correspond to different hemi-fields in the
visual space. Note that the left and right V1 areas have little or no callosal connections (Tootell et al.,
1998) to directly support their synchronization. The inter-hemispherical V1 correlation is most likely due
to a common input to both hemispheres. The topography of this common input seems retinotopically non-
specific, at least in terms of the polar angle.
Arguably, the eccentricity-dependent intrinsic activity patterns in V1 may be coincidental, and
reflect the relative distributions of magnocellular (M) and parvocellular (P) projections that happen to
vary with eccentricity. Previous studies have shown that P cells, relative to M cells, over-represent the
central vision but under-represent the periphery in lateral geniculate nuclei (LGN) (Connolly and Van
Essen, 1984; Schiller et al., 1990) and V1 (Baseler and Sutter, 1997; Azzopardi et al., 1999). The ratio
between P and M projections to/from V1 notably decreases with eccentricity (Baseler and Sutter, 1997).
Note that P and M pathways convey distinct visual attributes, but share the same retinotopic maps (Nassi
and Callaway, 2009; Denison et al., 2014). Although the representation of the M-to-P ratio seems similar
as the eccentricity representation in V1, the M-P pathways bear a different organization specific to visual
streams as opposed to visual locations. Caution should be exercised before interpreting an eccentricity-
dependent pattern alone as evidence for the retinotopic organization.
Resting-state activity reflects feedback visual-network interactions
We further speculate that the common input to V1, which drives visual-stream specific resting-
state activity, arises from top-down modulations through feedback connections. In the visual hierarchy
(Felleman and Van Essen, 1991), the population receptive field becomes larger and less specific from
lower to higher visual areas (Wandell et al., 2007), to progressively converge information in the visual
space through feedforward connections. While the feedforward connections are retinotopically organized,
the feedback corticocortical connections are not so (Salin and Bullier, 1995). Through feedback, the top-
down modulations transfer information about a large or even the whole visual field to cortical locations
18
with specific receptive fields (Salin and Bullier, 1995), driving network activity patterns away from being
retinotopically specific. It is conceivable that the organization of feedback connections plays a defining
role to resting state networks within the visual cortex. This is because when feedforward pathways are not
driven by fluctuating external inputs, feedback pathways are modulated by the brain's intrinsic activity.
In addition, feedback connections in the visual system are generally separated by the ventral and
dorsal pathways (Salin and Bullier, 1995; Gilbert and Li, 2013), functionally specialized for recognition
and action, respectively (Goodale and Milner, 1992; Ungerleider and Haxby, 1994). Such a visual-stream-
specific organization also applies to feedback connections from V1 to LGN (Briggs and Usrey, 2009). It
lends support for our interpretation that feedback projections serve the structural network basis of the
observed modular organization of intrinsic fine-scale functional networks distributed along the ventral
and dorsal pathways (Fig. 5). As the ventral and dorsal pathways become intricate in early visual areas,
fine-scale network patterns in V1/V2/V3 constitute the third functional module (Fig. 5).
Intrinsic functional parcellation of the visual cortex
On the basis of fine-scale resting-state activity patterns, the functional parcellation of the visual
cortex appeared to be notably different from other parcellations based on cortical folding, retinotopy, and
cytoarchitecture. The discrepancy with the cortical folding is perhaps reasonable, because the relationship
between cortical morphology and functional organization is elusive and indirect, despite a developmental
linkage likely between them as proposed elsewhere (Benson et al., 2012; Ronan and Fletcher, 2015). The
discrepancy with visual field maps is perhaps also understandable, for the reasons elaborated in previous
sections. In addition, the parcellation based on functional connectivity is applicable to all locations in the
visual cortex, whereas the visual field is not mapped at all visual areas. For the apparently different
parcellations based on regional cellular composition and inter-regional functional connectivity suggests a
lack of one-to-one relationships between cell types and functional networks.
19
Our results show that the functional networks and parcels seem to align with the gradient of
myelination (Fig. 8). We speculate that spontaneous activity shapes myelin density. It has been shown
that electrical activity may promote myelination(Gibson et al., 2014), and that functional organization and
cortical myelination may co-vary given plasticity (Fields, 2015; Hunt et al., 2016). As such, greater
myelin density may imply greater functional specificity but less plasticity: early sensory areas are more
functionally specific with greater myelination or less plasticity, whereas higher-order or multi-sensory
areas are less functionally specific with less myelin density and less plasticity (Glasser et al., 2014).
Acknowledgments
This work was supported in part by NIH R01MH104402 (Z Liu) and Purdue University. Authors have no
conflict of interest.
20
Figure Caption
21
Figure 1. Visual cortical mask and reproducibility of fine-scale network patterns. A. The visual
cortex mask (in blue) is illustrated on inflated and flattened cortical surfaces. Reference lines mark the
occipital-parietal sulcus (red dash line), the calcarine sulcus (purple dash line), a rough ventral-dorsal
division (yellow dash line), from where the ventral and dorsal pathways are along the green and red
arrows. B. Three examples of reproducible ICs that exhibit high correlations (r) in their spatial patterns.
The color scale represents the z-score and is ranged from -0.1 to 0.1. C. Shown in the left is the spatial
correlation matrix between the ICs obtained from two repeated resting-state sessions. The diagonal
elements correspond to uniquely paired ICs; the off-diagonal elements are between unpaired ICs. The red
line represents a correlation threshold (r=0.4), by which a pair of ICs was considered reproducible
between sessions. Shown in the middle is the p-values associated with the spatial correlations. Shown in
the right panel is the discrete histograms of the correlations for the paired (red) and unpaired (blue) ICs.
Figure 2. 52 reproducible independent components with spatial correlation greater than 0.4. A. The
ICs were numbered in a descending order of their test-retest reproducibility. IC51 and IC55 were not
22
optimally matched by our algorithm, but they yielded spatial correlations of 0.58 and 0.63 with IC39 and
IC41 respectively. Therefore, they are also considered to be reproducible. B. The remaining ICs that have
less consistency across the two datasets.
Figure 3. Discrete ICs within the early visual areas (V1/V2/V3). A. Four ICs within V1. B. Four ICs
within V2 and V3. The green lines are the borders of the multi-modal parcellation (Glasser et al., 2016),
where V1/V2/V3 are defined.
23
Figure 4. Discrete ICs along the dorsal pathway and ventral pathway. A. The left panel shows the
ICs that match well with existing lateral visual areas in the dorsal pathway. The right panel shows the ICs
that match well with existing medial visual areas in the dorsal pathway. B. Example ICs that match well
with existing medial visual areas in the ventral pathway such as VMV1-3, FFC and VVC. The green lines
mark the existing visual areal borders according to the MMP.
24
Figure 5. Functional modules in the visual cortex. A. Each component is represented with a sphere,
colored coded by its modular membership, and placed at the peak location of its spatial map. B. The
matrix on the left shows the temporal correlations (r) between ICs, organized into four modules (in (1)
foveal early visual cortex, (2) peripheral visual cortex, (3) dorsal pathway, (4) ventral pathway). The
matrix on the right shows the t-statistics associated with the temporal correlations (z) between ICs,
organized in the same order as the matrix on the left.
25
Figure 6. The effect of cortical distance on functional connectivity and modularity. A. Left panel:
The scatter-plots of correlation (r) vs. distance (d), which were modeled as r=1.432/d based on least-
squares estimation, in the left hemisphere. Right panel: The matrix shows the spatial-affinity between
component centroids within the left hemisphere. Middle panel: Each component is represented with a
sphere, colored coded by its modular membership, and placed at the peak location of its spatial map. B.
Left panel: The scatter-plots of correlation (r) vs. distance (d), which were modeled as r=1.521/d based on
least-squares estimation, in the right hemisphere. Right panel: The matrix shows the spatial-affinity
between component centroids within the right hemisphere. Middle panel: Each component is represented
with a sphere, colored coded by its modular membership, and placed at the peak location of its spatial
map.
26
Figure 7. Functional parcellation of the visual cortex with varying number of parcels (k=10, 20, 30,
40). The parcels are color-coded from 0 to 1, sorted in an ascending order according to the averaged sum
of distances within each parcel.
27
Figure 8. Comparing our parcellation (k = 40) with existing visual cortex parcellations. Our
functional parcellation is compared against (1) whole-brain multimodal images (i.e. MMP) (Glasser et al.,
2016); (2) visual-field maps (Abdollahi et al., 2014); (3) visual-field maps (Abdollahi et al., 2014); (4)
cortical folding patterns (Destrieux et al., 2010); (5) cortical myelination (Glasser et al., 2014).
28
Figure 9. Independent components and parcellations from three individual subjects obtained
through dual regression. A. Five example ICs from three subjects. B. The individual-level parcellations
obtained from three example subjects.
Supplementary Figure 1. 54 independent components obtained from motion-controlled fMRI
dataset with spatial correlation greater than 0.15. A. The 54 ICs are spatially consistent with those
observed in Fig. 2. B. Spatial cross correlation between ICs with and without controlling for motion
effects.
29
References
Abdollahi RO, Kolster H, Glasser MF, Robinson EC, Coalson TS, Dierker D, Jenkinson M, Van Essen
DC, Orban GA (2014) Correspondences between retinotopic areas and myelin maps in human
visual cortex. Neuroimage 99:509-524.
Amunts K, Malikovic A, Mohlberg H, Schormann T, Zilles K (2000) Brodmann's areas 17 and 18
brought into stereotaxic space-where and how variable? Neuroimage 11:66-84.
Arcaro MJ, Honey CJ, Mruczek RE, Kastner S, Hasson U (2015) Widespread correlation patterns of
fMRI signal across visual cortex reflect eccentricity organization. Elife 4.
Azzopardi P, Jones KE, Cowey A (1999) Uneven mapping of magnocellular and parvocellular
projections from the lateral geniculate nucleus to the striate cortex in the macaque monkey.
Vision research 39:2179-2189.
Baldassano C, Beck DM, Fei-Fei L (2013) Differential connectivity within the Parahippocampal Place
Baldassano C, Iordan MC, Beck DM, Fei-Fei L (2012) Voxel-level functional connectivity using spatial
Baseler H, Sutter E (1997) M and P components of the VEP and their visual field distribution. Vision
Area. Neuroimage 75:228-237.
regularization. Neuroimage 63:1099-1106.
research 37:675-690.
Bassett DS, Bullmore E (2006) Small-world brain networks. Neuroscientist 12:512-523.
Beckmann CF, Smith SM (2004) Probabilistic independent component analysis for functional magnetic
resonance imaging. IEEE Trans Med Imaging 23:137-152.
Bell AJ, Sejnowski TJ (1995) An information-maximization approach to blind separation and blind
deconvolution. Neural Comput 7:1129-1159.
Benson NC, Butt OH, Datta R, Radoeva PD, Brainard DH, Aguirre GK (2012) The retinotopic
organization of striate cortex is well predicted by surface topology. Current biology : CB
22:2081-2085.
Biswal B, Zerrin Yetkin F, Haughton VM, Hyde JS (1995) Functional connectivity in the motor cortex of
resting human brain using echo-planar mri. Magnetic resonance in medicine 34:537-541.
Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E (2008) Fast unfolding of communities in large
networks. Journal of Statistical Mechanics-Theory and Experiment 2008:P10008.
Bock AS, Binda P, Benson NC, Bridge H, Watkins KE, Fine I (2015) Resting-State Retinotopic
Organization in the Absence of Retinal Input and Visual Experience. The Journal of neuroscience
: the official journal of the Society for Neuroscience 35:12366-12382.
Briggs F, Usrey WM (2009) Parallel processing in the corticogeniculate pathway of the macaque monkey.
Neuron 62:135-146.
Butt OH, Benson NC, Datta R, Aguirre GK (2013) The fine-scale functional correlation of striate cortex
in sighted and blind people. The Journal of neuroscience : the official journal of the Society for
Neuroscience 33:16209-16219.
Calhoun VD, Liu J, Adali T (2009) A review of group ICA for fMRI data and ICA for joint inference of
imaging, genetic, and ERP data. Neuroimage 45:S163-172.
Cole MW, Bassett DS, Power JD, Braver TS, Petersen SE (2014) Intrinsic and task-evoked network
architectures of the human brain. Neuron 83:238-251.
Connolly M, Van Essen D (1984) The representation of the visual field in parvicellular and magnocellular
layers of the lateral geniculate nucleus in the macaque monkey. J Comp Neurol 226:544-564.
Damoiseaux JS, Rombouts SA, Barkhof F, Scheltens P, Stam CJ, Smith SM, Beckmann CF (2006)
Consistent resting-state networks across healthy subjects. Proceedings of the National Academy
of Sciences of the United States of America 103:13848-13853.
Dawson DA, Lam J, Lewis LB, Carbonell F, Mendola JD, Shmuel A (2016) Partial Correlation-Based
Retinotopically Organized Resting-State Functional Connectivity Within and Between Areas of
the Visual Cortex Reflects More Than Cortical Distance. Brain connectivity 6:57-75.
30
de Zwart JA, van Gelderen P, Liu Z, Duyn JH (2013) Independent sources of spontaneous BOLD
fluctuation along the visual pathway. Brain topography 26:525-537.
Denison RN, Vu AT, Yacoub E, Feinberg DA, Silver MA (2014) Functional mapping of the
magnocellular and parvocellular subdivisions of human LGN. Neuroimage 102 Pt 2:358-369.
Destrieux C, Fischl B, Dale A, Halgren E (2010) Automatic parcellation of human cortical gyri and sulci
using standard anatomical nomenclature. Neuroimage 53:1-15.
Doucet G, Naveau M, Petit L, Delcroix N, Zago L, Crivello F, Jobard G, Tzourio-Mazoyer N, Mazoyer
B, Mellet E, Joliot M (2011) Brain activity at rest: a multiscale hierarchical functional
organization. Journal of neurophysiology 105:2753-2763.
Eickhoff SB, Stephan KE, Mohlberg H, Grefkes C, Fink GR, Amunts K, Zilles K (2005) A new SPM
toolbox for combining probabilistic cytoarchitectonic maps and functional imaging data.
Neuroimage 25:1325-1335.
Felleman DJ, Van Essen DC (1991) Distributed hierarchical processing in the primate cerebral cortex.
Fields RD (2015) A new mechanism of nervous system plasticity: activity-dependent myelination. Nature
Cerebral cortex 1:1-47.
reviews Neuroscience 16:756-767.
Finn ES, Shen X, Scheinost D, Rosenberg MD, Huang J, Chun MM, Papademetris X, Constable RT
(2015) Functional connectome fingerprinting: identifying individuals using patterns of brain
connectivity. Nature neuroscience 18:1664-1671.
Fischl B, Rajendran N, Busa E, Augustinack J, Hinds O, Yeo BT, Mohlberg H, Amunts K, Zilles K
(2008) Cortical folding patterns and predicting cytoarchitecture. Cerebral cortex 18:1973-1980.
Fox MD, Raichle ME (2007) Spontaneous fluctuations in brain activity observed with functional
magnetic resonance imaging. Nature reviews Neuroscience 8:700-711.
Fox MD, Greicius M (2010) Clinical applications of resting state functional connectivity. Frontiers in
systems neuroscience 4:19.
Fox MD, Corbetta M, Snyder AZ, Vincent JL, Raichle ME (2006) Spontaneous neuronal activity
distinguishes human dorsal and ventral attention systems. Proceedings of the National Academy
of Sciences 103:10046-10051.
Genc E, Scholvinck ML, Bergmann J, Singer W, Kohler A (2016) Functional Connectivity Patterns of
Visual Cortex Reflect its Anatomical Organization. Cerebral cortex 26:3719-3731.
Gibson EM, Purger D, Mount CW, Goldstein AK, Lin GL, Wood LS, Inema I, Miller SE, Bieri G,
Zuchero JB, Barres BA, Woo PJ, Vogel H, Monje M (2014) Neuronal activity promotes
oligodendrogenesis and adaptive myelination in the mammalian brain. Science 344:1252304.
Gilbert CD, Li W (2013) Top-down influences on visual processing. Nature reviews Neuroscience
14:350-363.
Glasser MF, Van Essen DC (2011) Mapping human cortical areas in vivo based on myelin content as
revealed by T1- and T2-weighted MRI. The Journal of neuroscience : the official journal of the
Society for Neuroscience 31:11597-11616.
Glasser MF, Goyal MS, Preuss TM, Raichle ME, Van Essen DC (2014) Trends and properties of human
cerebral cortex: correlations with cortical myelin content. Neuroimage 93 Pt 2:165-175.
Glasser MF, Coalson TS, Robinson EC, Hacker CD, Harwell J, Yacoub E, Ugurbil K, Andersson J,
Beckmann CF, Jenkinson M, Smith SM, Van Essen DC (2016) A multi-modal parcellation of
human cerebral cortex. Nature 536:171-178.
Glasser MF, Sotiropoulos SN, Wilson JA, Coalson TS, Fischl B, Andersson JL, Xu J, Jbabdi S, Webster
M, Polimeni JR, Van Essen DC, Jenkinson M, Consortium WU-MH (2013) The minimal
preprocessing pipelines for the Human Connectome Project. Neuroimage 80:105-124.
Goodale MA, Milner AD (1992) Separate visual pathways for perception and action. Trends Neurosci
15:20-25.
Gravel N, Harvey B, Nordhjem B, Haak KV, Dumoulin SO, Renken R, Curcic-Blake B, Cornelissen FW
(2014) Cortical connective field estimates from resting state fMRI activity. Front Neurosci 8:339.
31
Greicius MD, Krasnow B, Reiss AL, Menon V (2003) Functional connectivity in the resting brain: a
network analysis of the default mode hypothesis. Proceedings of the National Academy of
Sciences 100:253-258.
Guclu U, van Gerven MA (2015) Deep Neural Networks Reveal a Gradient in the Complexity of Neural
Representations across the Ventral Stream. The Journal of neuroscience : the official journal of
the Society for Neuroscience 35:10005-10014.
Hasson U, Yang E, Vallines I, Heeger DJ, Rubin N (2008) A hierarchy of temporal receptive windows in
human cortex. The Journal of neuroscience : the official journal of the Society for Neuroscience
28:2539-2550.
Heinzle J, Kahnt T, Haynes JD (2011) Topographically specific functional connectivity between visual
field maps in the human brain. Neuroimage 56:1426-1436.
Honey CJ, Kotter R, Breakspear M, Sporns O (2007) Network structure of cerebral cortex shapes
functional connectivity on multiple time scales. Proceedings of the National Academy of
Sciences of the United States of America 104:10240-10245.
Hubel DH, Wiesel TN (1962) Receptive fields, binocular interaction and functional architecture in the
cat's visual cortex. J Physiol 160:106-154.
Hunt BA, Tewarie PK, Mougin OE, Geades N, Jones DK, Singh KD, Morris PG, Gowland PA, Brookes
MJ (2016) Relationships between cortical myeloarchitecture and electrophysiological networks.
Proceedings of the National Academy of Sciences 113:13510-13515.
Hutchison RM, Womelsdorf T, Allen EA, Bandettini PA, Calhoun VD, Corbetta M, Della Penna S, Duyn
JH, Glover GH, Gonzalez-Castillo J (2013) Dynamic functional connectivity: promise, issues,
and interpretations. Neuroimage 80:360-378.
Jo HJ, Saad ZS, Gotts SJ, Martin A, Cox RW (2012) Quantifying agreement between anatomical and
functional interhemispheric correspondences in the resting brain. PloS one 7:e48847.
Kenet T, Bibitchkov D, Tsodyks M, Grinvald A, Arieli A (2003) Spontaneously emerging cortical
representations of visual attributes. Nature 425:954-956.
Lewis CM, Bosman CA, Womelsdorf T, Fries P (2016) Stimulus-induced visual cortical networks are
recapitulated by spontaneous local and interareal synchronization. Proceedings of the National
Academy of Sciences of the United States of America 113:E606-615.
Long X, Goltz D, Margulies DS, Nierhaus T, Villringer A (2014) Functional connectivity-based
parcellation of the human sensorimotor cortex. The European journal of neuroscience 39:1332-
1342.
Martin A (2007) The representation of object concepts in the brain. Annu Rev Psychol 58:25-45.
Nassi JJ, Callaway EM (2009) Parallel processing strategies of the primate visual system. Nature reviews
Neuroscience 10:360-372.
Nir Y, Hasson U, Levy I, Yeshurun Y, Malach R (2006) Widespread functional connectivity and fMRI
fluctuations in human visual cortex in the absence of visual stimulation. Neuroimage 30:1313-
1324.
Ohiorhenuan IE, Mechler F, Purpura KP, Schmid AM, Hu Q, Victor JD (2010) Sparse coding and high-
order correlations in fine-scale cortical networks. Nature 466:617-621.
Raemaekers M, Schellekens W, van Wezel RJ, Petridou N, Kristo G, Ramsey NF (2014) Patterns of
resting state connectivity in human primary visual cortical areas: a 7T fMRI study. Neuroimage
84:911-921.
Rao RPN, Ballard DH (1999) Predictive coding in the visual cortex: a functional interpretation of some
extra-classical receptive-field effects. Nature neuroscience 2:79-87.
Ronan L, Fletcher PC (2015) From genes to folds: a review of cortical gyrification theory. Brain structure
Rubinov M, Sporns O (2010) Complex network measures of brain connectivity: uses and interpretations.
& function 220:2475-2483.
Neuroimage 52:1059-1069.
Physiol Rev 75:107-154.
Salin PA, Bullier J (1995) Corticocortical connections in the visual system: structure and function.
32
Schiller PH, Logothetis NK, Charles ER (1990) Functions of the colour-opponent and broad-band
channels of the visual system. Nature 343:68-70.
Seeley WW, Menon V, Schatzberg AF, Keller J, Glover GH, Kenna H, Reiss AL, Greicius MD (2007)
Dissociable intrinsic connectivity networks for salience processing and executive control. The
Journal of neuroscience 27:2349-2356.
Smith SM, Fox PT, Miller KL, Glahn DC, Fox PM, Mackay CE, Filippini N, Watkins KE, Toro R, Laird
AR, Beckmann CF (2009) Correspondence of the brain's functional architecture during activation
and rest. Proceedings of the National Academy of Sciences of the United States of America
106:13040-13045.
Smith SM, Vidaurre D, Beckmann CF, Glasser MF, Jenkinson M, Miller KL, Nichols TE, Robinson EC,
Salimi-Khorshidi G, Woolrich MW, Barch DM, Ugurbil K, Van Essen DC (2013) Functional
connectomics from resting-state fMRI. Trends in cognitive sciences 17:666-682.
Striem-Amit E, Ovadia-Caro S, Caramazza A, Margulies DS, Villringer A, Amedi A (2015) Functional
connectivity of visual cortex in the blind follows retinotopic organization principles. Brain : a
journal of neurology 138:1679-1695.
Tavor I, Parker Jones O, Mars RB, Smith SM, Behrens TE, Jbabdi S (2016) Task-free MRI predicts
individual differences in brain activity during task performance. Science 352:216-220.
Tootell RB, Mendola JD, Hadjikhani NK, Liu AK, Dale AM (1998) The representation of the ipsilateral
visual field in human cerebral cortex. Proceedings of the National Academy of Sciences of the
United States of America 95:818-824.
Ungerleider LG, Haxby JV (1994) 'What' and 'where' in the human brain. Curr Opin Neurobiol 4:157-
Van Den Heuvel MP, Mandl RC, Kahn RS, Pol H, Hilleke E (2009) Functionally linked resting-state
networks reflect the underlying structural connectivity architecture of the human brain. Human
brain mapping 30:3127-3141.
Van Dijk KR, Hedden T, Venkataraman A, Evans KC, Lazar SW, Buckner RL (2010) Intrinsic functional
connectivity as a tool for human connectomics: theory, properties, and optimization. Journal of
neurophysiology 103:297-321.
Van Essen DC, Glasser MF, Dierker DL, Harwell J, Coalson T (2012a) Parcellations and hemispheric
asymmetries of human cerebral cortex analyzed on surface-based atlases. Cerebral cortex
22:2241-2262.
Van Essen DC, Smith SM, Barch DM, Behrens TE, Yacoub E, Ugurbil K, Consortium WU-MH (2013)
The WU-Minn Human Connectome Project: an overview. Neuroimage 80:62-79.
Van Essen DC et al. (2012b) The Human Connectome Project: a data acquisition perspective.
Neuroimage 62:2222-2231.
165.
Wandell BA, Dumoulin SO, Brewer AA (2007) Visual field maps in human cortex. Neuron 56:366-383.
Wang Z, Chen LM, Negyessy L, Friedman RM, Mishra A, Gore JC, Roe AW (2013) The relationship of
anatomical and functional connectivity to resting-state connectivity in primate somatosensory
cortex. Neuron 78:1116-1126.
Wilf M, Strappini F, Golan T, Hahamy A, Harel M, Malach R (2015) Spontaneously Emerging Patterns
in Human Visual Cortex Reflect Responses to Naturalistic Sensory Stimuli. Cerebral cortex.
Yamins DL, DiCarlo JJ (2016) Using goal-driven deep learning models to understand sensory cortex.
Nature neuroscience 19:356-365.
Yeo BT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead M, Roffman JL, Smoller JW,
Zollei L, Polimeni JR, Fischl B, Liu H, Buckner RL (2011) The organization of the human
cerebral cortex estimated by intrinsic functional connectivity. Journal of neurophysiology
106:1125-1165.
Yoshimura Y, Dantzker JL, Callaway EM (2005) Excitatory cortical neurons form fine-scale functional
networks. Nature 433:868-873.
33
|
1902.06250 | 1 | 1902 | 2019-02-17T12:32:50 | Afferent Fiber Activity-Induced Cytoplasmic Calcium Signaling in Parvalbumin-Positive Inhibitory Interneurons of the Spinal Cord Dorsal Horn | [
"q-bio.NC",
"q-bio.TO"
] | Neuronal calcium (Ca2+) signaling represents a molecular trigger for diverse central nervous system adaptations and maladaptions. The altered function of dorsal spinal inhibitory interneurons is strongly implicated in the mechanisms underlying central sensitization in chronic pain. Surprisingly little is known, however, about the characteristics and consequences of Ca2+ signaling in these cells, including whether and how they are changed following a peripheral insult or injury and how such alterations might influence maladaptive pain plasticity. As a first step towards clarifying the precise role of Ca2+ signaling in dorsal spinal inhibitory neurons for central sensitization, we established methods for characterizing Ca2+ signals in genetically defined populations of these cells. In particular, we employed recombinant adeno-associated viral vectors to deliver subcellularly targeted, genetically encoded Ca2+ indicators into parvalbumin-positive spinal inhibitory neurons. Using wide-field microscopy, we observed both spontaneous and afferent fiber activity triggered Ca2+ signals in these cells. We propose that these methods may be adapted in future studies for the precise characterization and manipulation of Ca2+ signaling in diverse spinal inhibitory neuron subtypes, thereby enabling the clarification of its role in the mechanisms underlying pain chronicity and opening the door for possibly novel treatment directions. | q-bio.NC | q-bio | Title
Afferent Fiber Activity-Induced Cytoplasmic Calcium Signaling in Parvalbumin-Positive
Inhibitory Interneurons of the Spinal Cord Dorsal Horn
Authors
Anna M. Hagenston*, Sara Ben Ayed, Hilmar Bading
Department of Neurobiology
Interdisciplinary Center for Neurosciences (IZN)
Heidelberg University
INF 366
69120 Heidelberg
Germany
*Corresponding author Anna M. Hagenston: [email protected]
Sara Ben Ayed: [email protected]
Hilmar Bading: [email protected]
1
Abstract
Neuronal calcium (Ca2+) signaling represents a molecular trigger for diverse central nervous
system adaptations and maladaptions. The altered function of dorsal spinal inhibitory
interneurons is strongly implicated in the mechanisms underlying central sensitization in
chronic pain. Surprisingly little is known, however, about the characteristics and
consequences of Ca2+ signaling in these cells, including whether and how they are changed
following a peripheral insult or injury and how such alterations might influence maladaptive
pain plasticity. As a first step towards clarifying the precise role of Ca2+ signaling in dorsal
spinal inhibitory neurons for central sensitization, we established methods for characterizing
Ca2+ signals in genetically defined populations of these cells. In particular, we employed
recombinant adeno-associated viral vectors to deliver subcellularly targeted, genetically
encoded Ca2+ indicators into parvalbumin-positive spinal inhibitory neurons. Using wide-field
microscopy, we observed both spontaneous and afferent fiber activity triggered Ca2+ signals
in these cells. We propose that these methods may be adapted in future studies for the
precise characterization and manipulation of Ca2+ signaling in diverse spinal inhibitory
neuron subtypes, thereby enabling the clarification of its role in the mechanisms underlying
pain chronicity and opening the door for possibly novel treatment directions.
Keywords
inhibitory neuron, dorsal spinal cord, calcium, central sensitization, GCaMP3, parvalbumin
Main text
Changes in intracellular calcium (Ca2+) levels in neurons are intimately linked to
mechanisms underlying adaptive plastic changes within the central nervous system. Ca2+
rises influence neuronal properties by diverse means, including modulation of ion channel
activity or localization, regulation of neurotransmitter release, activation of Ca2+-dependent
2
enzymes, and induction or downregulation of gene expression. Dysregulation of intracellular
Ca2+ levels and their stimulus-evoked changes are similarly implicated in several central
nervous system pathologies; aberrant intraneuronal Ca2+ signaling is involved in processes
as broad-ranging as excitotoxic neuronal death following stroke (1), impaired synaptic
plasticity and memory formation in aging (2), and synapse loss in Alzheimer's disease (3),
but also peripheral and central sensitization in chronic pain (4). Spinal central sensitization,
defined as the aberrant amplification of noxious signal transmission in the spinal cord, is
characterized both by the enhanced excitation of projection neurons and by a loss of
inhibitory tone. Numerous mechanisms have been proposed to explain altered spinal
inhibitory tone in central sensitization, including reduced excitation or excitability of inhibitory
interneurons, degeneration of inhibitory neurons and/or their processes, reduced numbers
of GABA- and glycinergic synaptic terminals, and diminished inhibitory neurotransmitter
release (5, 6). Nonetheless, the precise molecular triggers underlying these alterations have
not been elucidated. In light of its importance for regulating diverse neuronal functions,
dysregulated intracellular Ca2+ signaling clearly represents a possible proximal trigger for
altered inhibitory neuron function in central sensitization. Surprisingly little is known,
however, about the characteristics and consequences of Ca2+ signaling in dorsal spinal
inhibitory neurons, including whether it is altered in chronic pain states. We therefore aimed
to establish methods for measuring Ca2+ signals in defined inhibitory neuron types and
subcellular compartments. Here we present data demonstrating the use of these methods
for characterizing cytoplasmic Ca2+ responses of parvalbumin (PV)-positive inhibitory
neurons in the dorsal spinal cord to stimulation of sensory fiber afferents.
We used the genetically encoded calcium indicator (GECI) GCaMP3, linked to a nuclear
export signal (NES), to characterize cytoplasmic Ca2+ signals. While the maximum response
amplitude and kinetics of GCaMP3 are inferior to those of, for example, GCaMP6f (7), this
particular GECI has the advantage that its relatively high baseline fluorescence allows for
3
the easy identification of transduced cells. Thus, GCaMP3 is our preferred GECI for wide-
field Ca2+ imaging in acute tissue preparations. We used recombinant adeno-associated
viral (rAAV)-mediated gene delivery to express cytoplasmically-localized GCaMP3
(GCaMP3.NES) in PV-positive neurons of the spinal cord dorsal horn of PV-Cre mice.
Specifically, we generated rAAV vectors containing a doubly floxed inverted open reading
frame for GCaMP3.NES, which is expressed in a Cre recombinase-dependent manner
under control of the human EF1α promoter (Fig. 1A). Initial tests carried out in primary
hippocampal cultures co-infected (or not) with a second rAAV driving the expression of myc-
tagged Cre recombinase -- detected immunocytochemically -- showed no GCaMP3.NES
expression in the absence of Cre recombinase (data not shown). We injected rAAV vectors
into the spinal parenchyma at spinal level L3/L4 and, after an incubation period of 3-4 weeks
to allow for stable transgene expression, prepared acute transverse slices from the L3-L5
spinal cord with attached dorsal roots for Ca2+ imaging (Fig. 1B). Consistent with numerous
studies demonstrating that PV-positive inhibitory neurons are localized in lamina III and
make projections into lamina II of the rodent spinal cord dorsal horn (8, 9), we observed
GCaMP3.NES fluorescence in multiple cell bodies and processes within laminae II-III of the
dorsal spinal cord (Fig. 1C).
Aβ, Aẟ, and C afferent fibers, which carry information related to innocuous touch, fast pain,
and slow pain, respectively, were electrically stimulated with varying frequency and
amplitude through a suction electrode attached to the dorsal root. The chosen stimulation
frequencies (4, 20, and 100 Hz) were roughly tuned to the maximum following frequencies
of each fiber class (10, 11), and a broad range of stimulation intensities (0.01-2.0 mA) was
employed in order to progressively activate each fiber class (12). As we have previously
reported for excitatory neurons in laminae I-II recorded under similar conditions (13), PV-
positive inhibitory neurons produced graded Ca2+ responses to afferent fiber stimulation of
increasing frequency and intensity (Fig. 1D,E; n=3-9 slices from 3-7 mice). PV-positive
4
inhibitory neurons in lamina III predominantly receive synaptic input from Aβ sensory
afferents (8). In keeping with this feature, we observed no sharp rise in response amplitudes
above Aẟ (~0.1 mA) or C (~0.5 mA) stimulation intensities for 20 Hz and 4 Hz stimulation
trains, respectively. In most recorded slices (n=13/16 slices from 9 mice), PV-positive cell
bodies and processes exhibited spontaneous Ca2+ rises in the absence of synaptic input
from the periphery (Fig. 1D), suggestive of their involvement in a basally active and recurrent
spinal network. We have not observed such spontaneous activity in dorsal spinal excitatory
neurons (data not shown).
The data we present here demonstrate that PV-positive inhibitory neurons in the dorsal
spinal cord exhibit both afferent activity-evoked and spontaneous cytoplasmic Ca2+ rises.
Following peripheral nerve injury, the number of synapses formed by these cells onto
excitatory interneurons decreases (9). Consequently, innocuous tactile information
impinging on laminae III and IV gains access to nociceptive circuits in laminae I/II, and
mechanical allodynia arises (9). A possible molecular mechanism underlying inhibitory
synapse loss is the impaired activity- and Ca2+-dependent expression of genes involved in
inhibitory synapse maintenance (14). Thus, an examination of afferent activity-evoked Ca2+
transients following nerve injury -- embedded in a study employing gain- and loss-of-function
tools specifically in PV-positive dorsal spinal neurons -- may provide important mechanistic
insight into spinal disinhibition in neuropathic pain.
We limited ourselves here to the description of Ca2+ signaling in one inhibitory neuron type
and subcellular compartment. Given the expanding repertoire of available Cre mouse lines
for defined populations of inhibitory neurons, the broad range of proven subcellular targeting
signals for GECIs, and the growing number of genetically encoded calcium signaling
inhibitors (e.g., CaMBP4, PV.NLS (13, 15)), we envision that our methods can be easily
adapted for studies aimed at gaining a precise understanding of the characteristics and
5
consequences of spinal inhibitory neuronal Ca2+ in central sensitization and pain
chronification.
Extended Materials and Methods
Mice
Mice expressing the Cre recombinase in parvalbumin (PV)-positive inhibitory neurons (PV-
Cre mice) were bred from founders originally obtained from Jackson Laboratories (JAX
Stock #008069 (16)) and kindly provided by R. Kuner (Heidelberg University). All
experimental procedures were approved by the local animal care and use committee
(Regierungspräsidium Karlsruhe, Referat 35, Karlsruhe, Germany), project G-272/14, and
carried out in accordance with ARRIVE guidelines and the EU Directive 2010/63/EU for
animal experiments. Mice had access to food and water ad libitum and were housed on a
12 h light-dark cycle.
Production and intra-spinal injection of rAAV vectors
The coding sequence for GCaMP3.NES, targeted to the cytoplasm using a nuclear export
signal (NES), was subcloned using standard procedures between the NheI and AscI
restriction sites of the Cre recombinase-dependent AAV expression plasmid, pAAV-EF1a-
double floxed-EYFP-WPRE-HGHpA, a gift from Karl Deisseroth (Addgene plasmid # 20296;
http://n2t.net/addgene:20296; RRID:Addgene_20296). Recombinant serotype 1/2 adeno-
associated viral vectors (rAAV) were produced and purified using previously described
methods (17, 18) and injected into the lumbar spinal cord of adult PV-Cre mice. Briefly, mice
were anesthetized with an intraperitoneal injection of sleep mix containing 0.5 mg/kg
medetomidine, 5.0 mg/kg midazolam, and 0.5 mg/kg fentanyl. When the mice were no
longer responsive to a foot pinch, an ~1 cm incision was made in the skin over the vertebral
column, and the tissue between thoracic vertebrae T12 and T13 -- corresponding to spinal
lumbar segments L3/L4 -- was dissected to reveal the dorsal spinal cord. The dura mater
6
was opened to improve needle penetration, and a 35 G beveled needle mounted in a 10 μl
Hamilton syringe (World Precision Instruments) was slowly advanced into the spinal
parenchyma. 500 nl of a 2:1 mixture of rAAV vector suspended in PBS and 20 % mannitol
containing ~1011 virions was then injected at a rate of 100 nl/min using a microprocessor-
controlled minipump (World Precision Instruments). The needle was left in place for 5 min
after the injection was complete to allow the viral infusion to diffuse away from the injection
site. The wound was subsequently covered with gelatin foam (Gelita Medical) and the skin
closed with sutures. Animals received carprofen (5 mg/kg s.c.) prior to the first incision and
buprenorphine (0.1 mg/kg s.c.) 1 h after subcutaneous administration of wake mix containing
0.75 mg/kg atipamezole, 0.5 mg/kg flumazenil and 1.2 mg/kg naloxone to minimize post-
surgical pain. After the procedure, mice were returned their home cage, which was placed
on a heating plate for 24 h, and provided with wet food pellets. As we have demonstrated in
the past (13), this viral delivery method generally results in transgene expression spreading
from L2 to the border of S1 of the mouse spinal cord with a maximum in L3-L5, the area of
the spinal cord that innervates the hindpaw.
Ca2+ imaging and dorsal root stimulation
Three to four weeks following rAAV injection surgery, acute spinal cord slices with attached
dorsal roots from spinal cord segments L3-L5 of rAAV-injected mice were prepared as
described previously (19). Briefly, mice were anesthetized with sodium pentobarbital (320
mg/kg, i.p.) and perfused with freshly prepared, ice-cold NMDG slicing solution containing,
in mM, 93 HCl, 2.5 KCl, 1.2 NaH2PO4, 3 sodium pyruvate, 20 HEPES, 30 NaHCO3, 25 D-
glucose, 5 L-ascorbic acid, 10 N-acetyl-L-cysteine, 2 thiourea, 10 MgSO4, 0.5 CaCl2,
adjusted to a final osmolarity of 300-305 mOsm and pH of 7.35, and bubbled with 95 % O2/5
% CO2. Afterwards, the spinal cord was isolated, the dura mater carefully removed, and the
ventral roots resected. The spinal cord was then mounted into a groove within a block of 1.5
7
% agarose in PBS and secured with cyanoacrylate adhesive to the stage of a Microm
HM650V (Thermo Scientific) vibratome containing NMDG slicing solution bubbled with 95
% O2/5 % CO2 and cooled to 1-4 °C. Slices (~330 µm) were prepared using double-edge
coated blades (Science Services, washed version, #7200-WA) and then placed into NMDG
slicing solution bubbled with 95 % O2/5 % CO2 and warmed to 34 °C for 10 min. Slices were
subsequently transferred to extracellular recording solution containing, in mM, 125 NaCl, 10
D-glucose, 25 NaHCO3, 2.5 KCl, 1.25 NaH2PO4, 1.2 MgSO4, 2 CaCl2, adjusted to a final
osmolarity of 300-305 mOsm and pH 7.35, bubbled with 95 % O2/5 % CO2 and warmed to
34 °C, and then allowed to cool to room temperature for a recovery period of ≥ 1 h. For Ca2+
imaging, slices were transferred to a recording chamber (PM-1, Warner Instruments,
Hamden, CT, USA or PC-R, Siskiyou, OR, USA) perfused (~3 ml/min) with 95 % O2/5 %
CO2-bubbled extracellular recording solution warmed with an in-line heater to 34 °C. Slices
were secured with a platinum ring with nylon strings. Dorsal roots were sucked into a glass
stimulation electrode, the tip of which had been dipped into Vaseline to improve the electrical
seal around the root, and stimulated with 1 s trains of electrical pulses (0.1 ms pulse width,
100 Hz and 20 Hz stimuli; 1 ms pulse width, 4 Hz stimuli) using a stimulus isolator (World
Precision Instruments, A365) controlled by Clampex 10.3 Software (Molecular Devices).
Ca2+ responses from GCaMP3.NES-labeled cells and their processes in laminae II-III were
imaged as described in detail previously (20). GCaMP3.NES was excited using a CoolLED
light source (480±10 nm), and fluorescence emission (530±20 nm) acquired at a rate of 2
Hz with a cooled camera (Photometrics Coolsnap HQ, Roper Scientific) with 2x2 binning
through a 20x water-immersion objective (XLM PlanFluor 0.95W, Olympus) on an upright
microscope (Olympus BX51WI) connected to a software interface (Metafluor, Universal
Imaging Systems and Molecular Devices). Overview images were taken with a 4x air
objective (PlanCN, Olympus).
8
Data analysis
Image sequences were imported into ImageJ (Fiji RRID:SCR_002285) and, when
necessary, processed with the "Template Matching" plugin to correct for slice movement
during the recording. ROIs were manually drawn around single cells and larger regions
encompassing nearly the entire field of view within laminae II/III containing fluorescently
labeled cells and processes, and fluorescence intensity changes measured over time.
Further data analysis was carried out using Igor Pro (WaveMetrics, RRID:SCR_000325).
Fluorescence intensity changes were expressed as the percent change with respect to
baseline in the 15 s prior to the onset of the stimulation train (% ΔF/F) and quantified using
the peak amplitude above baseline. Data were plotted as mean ± SEM using GraphPad
Prism 6 (RRID:SCR_002798).
Post-imaging analysis of GCaMP3.NES expression
Following Ca2+ imaging, acute spinal cord slices were fixed in 4 % paraformaldehyde in PBS
for 2 h than them placed in 30 % sucrose in PBS containing 0.04 % Thimerosal (Sigma-
Aldrich) for a minimum of 24 h. Tissue slices were embedded in tissue freezing medium
(Leica), and 20 μm sections cut using a cryotome (CM 1950, Ag Protect, Leica) and mounted
directly onto SuperFrost Plus slides (Thermo Scientific). Slides were coverslipped with
Mowiol 4-88 (Calbiochem) medium containing 2 µg/ml Hoechst 33258 (Serva, Cat # 15090)
as a nuclear counterstain. Images of the dorsal horn were obtained with an upright wide-
field microscope (DM IRBE, Leica) and 20x objective and processed using Adobe
Photoshop CS4.
List of abbreviations
Ca2+: calcium
GECI: genetically encoded calcium indicator
9
NES: nuclear export signal
PV: parvalbumin
rAAV: recombinant adeno-associated virus
Declarations
Ethics approval
This study was approved by the local animal care and use committee (Regierungspräsidium
Karlsruhe, Referat 35, Karlsruhe, Germany), project G-272/14. All animal experiments were
designed and carried out in accordance with ARRIVE guidelines and the EU Directive
2010/63/EU for animal experiments.
Availability of data and materials
The data sets analyzed during the current study are available from the corresponding author
upon reasonable request.
Consent for publication
Not applicable
Competing interests
The authors declare that they have no competing interests.
Funding
This study was funded by the German Research Foundation (SFB1158, grant A05).
Acknowledgements
10
We would like to thank Rohini Kuner (Heidelberg University) for providing PV-Cre mice for
use in this study.
Authors' contributions
AMH and HB conceived the project, designed the study, and wrote the paper. AMH and SK
performed the experiments. AMH and HB analyzed and interpreted the data. All authors
read and approved the final manuscript.
References
1.
Hardingham GE, Bading H. Synaptic versus extrasynaptic NMDA receptor signalling:
implications for neurodegenerative disorders. Nat Rev Neurosci. 2010;11:682-96.
2.
Kumar A, Bodhinathan K, Foster TC. Susceptibility to Calcium Dysregulation during
Brain Aging. Front Aging Neurosci. 2009;1:2.
3.
Bezprozvanny I, Mattson MP. Neuronal calcium mishandling and the pathogenesis
of Alzheimer's disease. Trends Neurosci. 2008;31(9):454-63.
4.
Hagenston AM, Simonetti M. Neuronal calcium signaling in chronic pain. Cell Tissue
Res. 2014;357(2):407-26.
5.
Alles SRA, Smith PA. Etiology and Pharmacology of Neuropathic Pain. Pharmacol
Rev. 2018;70(2):315-47.
6.
Zeilhofer HU, Wildner H, Yevenes GE. Fast synaptic inhibition in spinal sensory
processing and pain control. Physiol Rev. 2012;92(1):193-235.
7.
Chen TW, Wardill TJ, Sun Y, Pulver SR, Renninger SL, Baohan A, et al.
Ultrasensitive
fluorescent
proteins
for
imaging
neuronal
activity. Nature.
2013;499(7458):295-300.
8.
Todd AJ. Identifying functional populations among the interneurons in laminae I-III of
the spinal dorsal horn. Mol Pain. 2017;13:1744806917693003.
11
9.
Petitjean H, Pawlowski SA, Fraine SL, Sharif B, Hamad D, Fatima T, et al. Dorsal
Horn Parvalbumin Neurons Are Gate-Keepers of Touch-Evoked Pain after Nerve Injury. Cell
Rep. 2015;13(6):1246-57.
10. Daniele CA, MacDermott AB. Low-threshold primary afferent drive onto GABAergic
interneurons in the superficial dorsal horn of the mouse. J Neurosci. 2009;29(3):686-95.
11. Waddell PJ, Lawson SN. Electrophysiological properties of subpopulations of rat
dorsal root ganglion neurons in vitro. Neuroscience. 1990;36(3):811-22.
12. Sdrulla AD, Xu Q, He SQ, Tiwari V, Yang F, Zhang C, et al. Electrical stimulation of
low-threshold afferent fibers induces a prolonged synaptic depression in lamina II dorsal
horn neurons to high-threshold afferent inputs in mice. Pain. 2015;156(6):1008-17.
13. Simonetti M, Hagenston AM, Vardeh D, Freitag HE, Mauceri D, Lu J, et al. Nuclear
calcium signaling in spinal neurons drives a genomic program required for persistent
inflammatory pain. Neuron. 2013;77(1):43-57.
14. Sun X, Lin Y. Npas4: Linking Neuronal Activity to Memory. Trends Neurosci.
2016;39(4):264-75.
15. Mauceri D, Hagenston AM, Schramm K, Weiss U, Bading H. Nuclear Calcium
Buffering Capacity Shapes Neuronal Architecture. J Biol Chem. 2015;290(38):23039-49.
16. Hippenmeyer S, Vrieseling E, Sigrist M, Portmann T, Laengle C, Ladle DR, et al. A
developmental switch in the response of DRG neurons to ETS transcription factor signaling.
PLoS Biol. 2005;3(5):e159.
17. McClure C, Cole KL, Wulff P, Klugmann M, Murray AJ. Production and titering of
recombinant adeno-associated viral vectors. J Vis Exp. 2011(57):e3348.
18.
Zhang SJ, Steijaert MN, Lau D, Schutz G, Delucinge-Vivier C, Descombes P, et al.
Decoding NMDA receptor signaling: identification of genomic programs specifying neuronal
survival and death. Neuron. 2007;53(4):549-62.
12
19.
Luo C, Seeburg PH, Sprengel R, Kuner R. Activity-dependent potentiation of calcium
signals in spinal sensory networks in inflammatory pain states. Pain. 2008;140(2):358-67.
20. Bengtson CP, Freitag HE, Weislogel JM, Bading H. Nuclear calcium sensors reveal
that repetition of trains of synaptic stimuli boosts nuclear calcium signaling in CA1 pyramidal
neurons. Biophys J. 2010;99(12):4066-77.
13
Figures
Figure 1. Afferent activity-triggered and spontaneous Ca2+ signaling in dorsal spinal
inhibitory interneurons. (A) We generated serotype 1/2 rAAV vectors containing a doubly
floxed inverted open reading frame coding for the cytoplasmically-localized GECI,
GCaMP3.NES, expressed in a Cre recombinase-dependent manner under control of the
EF1α promoter. (B) rAAV vectors were injected into the spinal parenchyma of adult PV-Cre
mice. Acute slices with attached dorsal roots from the L3-L5 spinal cord were prepared 3-4
weeks later. Sensory afferent activity was evoked by suction electrode stimulation of the
attached dorsal root. Recordings of Ca2+ responses from PV-positive neurons were made
in laminae II-III of the spinal cord dorsal horn. (C) Analysis of GCaMP3.NES expression in
the spinal cord dorsal horn. GCaMP3.NES was observed in the extranuclear soma and
processes of cells residing in lamina III and extending processes into lamina II. The upper
border of lamina I is marked with a dashed line. Scale bar, 100 μm. (D) Top, representative
image from a Ca2+ imaging experiment in which multiple GCaMP3.NES-positive neurons
14
could be visualized. Bottom, changes in cytoplasmic Ca2+ levels were visualized using the
normalized fluorescence intensity in regions of interest laid over individual neurons (red,
green, and blue traces; corresponding the colored arrows in the image above) and the bulk
tissue (black traces; corresponding to the large white ellipse in the image above) during 100
Hz dorsal root stimulation (0.2-2.0 mA). Both stimulus-evoked responses and spontaneous
activity with highly variable amplitude and kinetics were observed. (E) Quantification of Ca2+
rises in the bulk tissue triggered by 1 s, 0.01-2.0 mA stimulation trains delivered at 4, 20,
and 100 Hz (n = 3-9 slices from 3-7 animals). The amplitudes of evoked Ca2+ transients
varied positively with both stimulus frequency and intensity.
15
|
1602.07389 | 3 | 1602 | 2017-07-07T20:22:20 | Capturing the dynamical repertoire of single neurons with generalized linear models | [
"q-bio.NC"
] | A key problem in computational neuroscience is to find simple, tractable models that are nevertheless flexible enough to capture the response properties of real neurons. Here we examine the capabilities of recurrent point process models known as Poisson generalized linear models (GLMs). These models are defined by a set of linear filters, a point nonlinearity, and conditionally Poisson spiking. They have desirable statistical properties for fitting and have been widely used to analyze spike trains from electrophysiological recordings. However, the dynamical repertoire of GLMs has not been systematically compared to that of real neurons. Here we show that GLMs can reproduce a comprehensive suite of canonical neural response behaviors, including tonic and phasic spiking, bursting, spike rate adaptation, type I and type II excitation, and two forms of bistability. GLMs can also capture stimulus-dependent changes in spike timing precision and reliability that mimic those observed in real neurons, and can exhibit varying degrees of stochasticity, from virtually deterministic responses to greater-than-Poisson variability. These results show that Poisson GLMs can exhibit a wide range of dynamic spiking behaviors found in real neurons, making them well suited for qualitative dynamical as well as quantitative statistical studies of single-neuron and population response properties. | q-bio.NC | q-bio | Capturing the Dynamical Repertoire of Single Neurons
with Generalized Linear Models
Alison I. Weber1 & Jonathan W. Pillow2,3
1Graduate Program in Neuroscience, University of Washington, Seattle, WA, USA.
2Princeton Neuroscience Institute, Princeton University, Princeton, NJ, USA.
3Dept. of Psychology, Princeton University, Princeton, NJ, USA.
Keywords: point process; generalized linear model (GLM); Izhikevich model; spike timing;
variability
Abstract
A key problem in computational neuroscience is to find simple, tractable models that are never-
theless flexible enough to capture the response properties of real neurons. Here we examine
the capabilities of recurrent point process models known as Poisson generalized linear models
(GLMs). These models are defined by a set of linear filters, a point nonlinearity, and condi-
tionally Poisson spiking. They have desirable statistical properties for fitting and have been
widely used to analyze spike trains from electrophysiological recordings. However, the dynam-
ical repertoire of GLMs has not been systematically compared to that of real neurons. Here
we show that GLMs can reproduce a comprehensive suite of canonical neural response be-
haviors, including tonic and phasic spiking, bursting, spike rate adaptation, type I and type II
excitation, and two forms of bistability. GLMs can also capture stimulus-dependent changes in
spike timing precision and reliability that mimic those observed in real neurons, and can exhibit
varying degrees of stochasticity, from virtually deterministic responses to greater-than-Poisson
variability. These results show that Poisson GLMs can exhibit a wide range of dynamic spiking
behaviors found in real neurons, making them well suited for qualitative dynamical as well as
quantitative statistical studies of single-neuron and population response properties.
7
1
0
2
l
u
J
7
]
.
C
N
o
i
b
-
q
[
3
v
9
8
3
7
0
.
2
0
6
1
:
v
i
X
r
a
1 Introduction
Understanding the dynamical and computational properties of neurons is a fundamental chal-
lenge in cellular and systems neuroscience. A wide variety of single-neuron models have been
proposed to account for neural response properties. These models can be arranged along a
complexity axis ranging from detailed, interpretable, biophysically accurate models to simple,
tractable, reduced functional models. Detailed Hodgkin-Huxley style models, which sit at one
end of this continuum, provide a biophysically detailed account of the conductances, currents,
and channel kinetics governing neural response properties [21]. These models can account
for the vast dynamical repertoire of real neurons, but they are often unwieldy for theoretical
analyses of neural coding and computation. This motivates the need for simplified models of
neural spike responses that are tractable enough for mathematical, computational, and statis-
tical analyses.
A variety of simplified dynamical models have been proposed to serve the need for math-
ematically tractable models, including the integrate-and-fire model, Fitzhugh-Nagumo, Morris-
Lecar, and Izhikevich models [9, 33, 32, 22, 4]. Generally, these models aim to reduce the
biophysically detailed descriptions of realistic neurons to systems of differential equations with
fewer variables and/or simplified dynamics. The one-dimensional integrate-and-fire model is
arguably the simplest of these, and the simplest to analyze mathematically, but it fails to cap-
ture many of the response properties of real neurons. The two-dimensional Izhikevich model,
by contrast, was specifically formulated to retain the rich dynamical repertoire of more com-
plex, biophysically realistic models [23].
An alternative to a mathematical notion of simplicity is the statistical property of being
tractable for fitting from intracellular or extracellular physiological recordings. One well-known
statistical model that satisfies this desideratum is the recurrent linear-nonlinear Poisson model,
commonly referred to in the neuroscience literature as the generalized linear model (GLM)
[48, 40]. GLMs are closely related to generalized integrate-and-fire models such as the spike-
response model, which has linear dynamics but incorporates spike-dependent feedback to
capture the nonlinear effects of spiking on neural membrane potential and subsequent spike
generation [13, 26, 24, 39, 16]. In fact, a variant of the spike response model that incorporates
noise into the spike threshold is mathematically equivalent to the models we study here [15,
12, 25, 14]. GLMs are popular for characterizing neural responses in reverse-correlation or
white-noise experiments, due to the tractability of likelihood-based fitting methods. Recent
work has shown that GLMs can capture the detailed statistics of spiking in single and multi-
neuron recordings from a variety of brain areas [40, 2, 5, 50, 31, 41].
While several studies have shown that GLMs can successfully recapitulate various re-
sponse properties of biological or simulated neurons, here, we provide a more systematic
study of the dynamical repertoire of the GLM . We study this issue by fitting GLMs to data
from simulated neurons exhibiting a number of complex response properties. We show that
GLMs can reproduce a remarkably rich set of dynamical behaviors, including tonic and phasic
spiking, bursting, spike rate adaptation, type I and type II excitation, and two different forms of
bistability. Furthermore, GLMs can exhibit stimulus-dependent degrees of spike timing preci-
sion and reliability [29], and mimic a recently reported form of greater-than-Poisson variability
[17].
2
2 Models of dynamical behaviors
2.1 Izhikevich model
First, we will examine whether generalized linear models can reproduce a suite of canonical
spiking behaviors exhibited by the well-known Izhikevich model [22, 23]. The Izhikevich model
is a biophysically-inspired model of intracellular membrane potential defined by a two-variable
system of ordinary differential equations governing membrane potential v(t) and a recovery
variable u(t):
v = 0.04v2 + 5v + 140 − u + I(t)
u = a(bv − u)
with spiking and voltage-reset governed by the boundary condition:
if v(t) ≥ 30, "spike" and set
(cid:26) v(t+) = c
u(t+) = u(t) + d,
(1)
(2)
(3)
where I(t) is injected current, t+ denotes the next time step after t, and parameters (a, b, c, d)
determine the model's dynamics. Different settings of these parameters lead to qualitatively
different spiking behaviors, as shown in [23]. We focus on this model because of its demon-
strated ability to produce a wide range of response properties exhibited by real neurons. (See
Table 1 for parameter values used in this study and Methods for simulation details.)
2.2 Generalized linear model (GLM)
The GLM is a regression model typically used to characterize the relationship between exter-
nal or internal covariates and a set of recorded spike trains.
In systems neuroscience, the
label "GLM" often refers to an autoregressive point process model, a model in which linear
functions of stimulus and spike history are nonlinearly transformed to produce the spike rate
or conditional intensity of a Poisson process [48, 40].
The GLM is parametrized by a stimulus filter (cid:126)k, which describes how the neuron integrates
an external stimulus, a post-spike filter (cid:126)h, which captures the influence of spike history on the
current probability of spiking, and a scalar µ that determines the baseline spike rate. (See
Figure 1.) The outputs of these filters are summed and passed through a nonlinear function f
to determine the conditional intensity λ(t):
λ(t) = f ((cid:126)k · (cid:126)x(t) + (cid:126)h · (cid:126)yhist (t) + µ),
(4)
where (cid:126)x(t) is the (vectorized) spatio-temporal stimulus, (cid:126)yhist(t) is a vector representing spike
history at time t, and f is a nonlinear function that ensures the spike rate is non-negative.
Spikes are generated according to a conditionally Poisson process [37, 6], so spike count y(t)
in a time bin of size ∆ is distributed according to a Poisson distribution:
P (y(t)λ(t)) =
1
y(t)!
∆λ(t)
e−∆λ(t).
(5)
(cid:17)y(t)
(cid:16)
3
Figure 1: Schematic of the generalized linear model. A: The stimulus filter (cid:126)k operates
linearly on the stimulus x(t), is combined with input from the post-spike filter (cid:126)h and mean input
level µ. This combined linear signal passes through a point nonlinearity f (·), whose output
drives spiking via a conditionally Poisson process. B: An equivalent view of the GLM, which
emphasizes the dependencies between a particular time window of stimulus and spike history
and conditional intensity λ(t), which governs the probability of a spike in the current time bin
(dark gray box).
In this study we set f to be exponential, although similar properties can be obtained with other
nonlinearities such as the soft-rectification function.
Unlike classical deterministic models like Hodgkin-Huxley and integrate-and-fire, the GLM
is fundamentally stochastic due to the assumption of conditionally Poisson spiking. However,
this stochasticity is helpful for fitting purposes because it assigns graded probabilities to firing
events and allows for likelihood-based methods for parameter fitting [35, 39].
In fact, the
Poisson GLM comes with a well-known guarantee that the log-likelihood function is concave for
suitable choices of nonlinearity f [34]. This means we can be assured of approaching a global
optimum of the likelihood function via gradient ascent, for any set of stimuli and spike trains
(barring any numerical issues that may complicate achieving the actual maximum for certain
datasets, cf. [53]). This guarantee does not hold for stochastic formulations of most nonlinear
biophysical models, including the Izhikevich model. Moreover, despite its stochasticity, the
GLM can produce highly precise and repeatable spike trains in certain parameter settings, as
we will demonstrate below.
3 GLMs capture a wide array of complex dynamical behav-
iors
We fit GLMs to data simulated from Izhikevich neurons set up to exhibit a range of different
qualitative response behaviors.
In the following, we describe these behaviors in detail, be-
ginning with simpler behaviors, such as tonic spiking and bursting (which have already been
demonstrated in previous work, e.g., [15, 25]) in order to build intuition for the GLM's basic
capabilities, and then move on to more complex behaviors (such as bistability) and questions
of spike timing reliability and precision.
4
generalized linear model (GLM)equivalent diagramABstimulusspike traintimestimulus filternonlinearitypost-spike filterstochasticspiking+ 3.1 Tonic spiking
We first examined an Izhikevich neuron tuned to exhibit tonic spiking (Figure 2A-B; see Table
1 for parameters). When presented with a step input current, the Izhikevich neuron responds
with a few high-frequency spikes and then settles into a regular firing pattern that persists for
the duration of the step (Figure 2B). This response pattern resembles that of a deterministic
Hodgkin-Huxley or integrate-and-fire neuron, albeit with an added transient burst of spikes at
stimulus onset.
We simulated the Izhikevich neuron's response to a series of step currents and used the
resulting training data to perform maximum-likelihood fitting of the GLM parameters {(cid:126)k,(cid:126)h, µ}.
The estimated stimulus filter (cid:126)k is biphasic (Figure 2C), resulting in a large transient response
to stimulus onset, and has a positive integral, ensuring a sustained positive response to a
Figure 2: Tonic spiking behavior. A: A step current stimulus. B: Voltage response of the
simulated Izhikevich neuron. C: The fitted GLM stimulus filter (cid:126)k has a biphasic shape that
gives the model a vigorous response to stimulus onset and a net positive response to a sus-
tained input. D: The fitted GLM post-spike filter (cid:126)h has a negative lobe that imposes strong
refractoriness on a timescale of ≈50 ms. E: Stimulus (blue) and post-spike (red) filter outputs
during simulated response of the fitted GLM to the stimulus shown above on a single trial. F:
The summed filter outputs are passed through an exponential nonlinearity to determine the
conditional intensity λ(t), shown here for a single trial. G: Spike train of the Izhikevich neuron
(black) and simulated repeats of the GLM (gray). GLM spike responses are slightly different on
each trial, due to the stochasticity of spike generation, but reproduce Izhikevich model spike
times with high precision.
5
Izhikevich neuron responseA!EFGinputmVfilter outputsresponsesGLM repeatsconditional!intensitytime (ms)stimulusC!time (ms)(sp/s)Izhikevich = -35fitted GLM parametersBDstimulus filterpost-spike filter outputoutput02004006001201010106-200002000-80-30200510-100-500-0.500.511.5050100-1000-750-500-2500 current step (Figure 2E, blue trace). The estimated post-spike filter (cid:126)h (Figure 2D), by contrast,
has a large negative lobe that provides recurrent inhibition after every spike, enforcing a strong
relative refractory period. The stimulus filter and post-spike filter output (shown together for
a single trial in Figure 2E) are summed together and exponentiated to obtain the conditional
intensity λ(t) (Figure 2F), also known as the instantaneous spike rate.
For this stimulus, the intensity rises very quickly once (cid:126)h decays, which occurs approxi-
mately 50 ms after the previous spike. Note that the output of the stimulus filter is identical
on each trial, whereas the the output of the post-spike filter varies from trial to trial because
of variability in the exact timing of spikes. However, because the rising phase of the condi-
tional intensity is so rapid, spiking is virtually certain within a small time window sitting at a
fixed latency after the previous spike time. The combination of strong excitatory drive from the
stimulus filter and strong suppressive drive from the post-spike filter produces precisely timed
spikes across trials, allowing the GLM to closely match the deterministic firing pattern of the
Izhikevich neuron (Figure 2G).
3.2 Bursting
We next examined multi-spike bursting, a more complex temporal response pattern that re-
quires dependencies beyond the most recent interspike interval. Once again, we simulated
responses from an Izhikevich neuron tuned to exhibit tonic bursting (Figure 3A-B) and used
the resulting data to fit GLM parameters (Figure 3C-D). The estimated stimulus filter (cid:126)k is bipha-
sic with a larger positive than negative lobe, which drives rapid spiking at stimulus onset and
generates sustained drive during an elevated stimulus (Figure 3E, blue trace). The post-spike
filter (cid:126)h has an immediate negative component that creates a relative refractory period after
each spike, and an even more negative mode after a latency of ≈40 ms; the accumulation
of these negative components over multiple spikes gives rise to a sustained suppression of
activity between bursts.
The GLM captures the bursting behavior of the Izhikevich neuron with high precision, in-
cluding the fact that the first burst after stimulus onset contains a different pattern of spikes
than subsequent bursts. This difference arises from precise interactions between the stimulus
and post-spike filter outputs. During the first burst, fast spiking arises from an interplay be-
tween monotonically increasing stimulus filter output (Figure 3E, blue) and tonic decrements
induced by the post-spike filter after each spike (Figure 3E, red). After each spike, the post-
spike filter reduces the conditional intensity by a fixed decrement, but this decrement is soon
overwhelmed by the rising wave of input from the stimulus filter, which creates a rapid rise and
fall of the conditional intensity time-locked to each Izhikevich neuron spike time. The pattern
continues until accumulated contributions from the delayed negative lobes of post-spike filter
overwhelm those from the stimulus filter and the burst terminates. Subsequent bursts are gov-
erned by a somewhat different interplay between stimulus and post-spike filter outputs: bursts
sit on a rising phase of the conditional intensity due to the removal of suppression from the
previous burst. This rise is more gradual than the drive induced by stimulus onset, and results
in bursts with longer inter-spike intervals and fewer spikes per burst, but the resulting spike
pattern is nonetheless captured with high precision and reliability from trial to trial.
6
Figure 3: Bursting behavior. A: Step current stimulus. B: Voltage response of Izhikevich neu-
ron. C: Fitted GLM stimulus filter. D: Fitted GLM post-spike filter, which creates refractoriness
on short timescales (within each burst) due to instantaneous depolarization following a spike.
The large negative lobe ≈25-50 ms after a spike terminates bursting and strongly suppresses
firing between bursts. E: Stimulus (blue) and post-spike (red) filter outputs for simulated re-
sponse of the GLM to step current shown above on a single trial. F: Output of the nonlinearity
(conditional intensity) λ(t) on a single trial. G: Spike train of Izhikevich neuron (black) and
simulated repeats of the fitted GLM (gray).
3.3 Bistability
Bistability refers to the phenomenon in which there are multiple stable response modes for
a single input condition. A common form of bistability observed in real neurons is the ability
to inhabit either a tonically active state or a silent state for a given level of current injection.
The Izhikevich model can exhibit this form of bistability, wherein a brief positive current pulse
is sufficient to kick it between states:
the neuron can inhabit a silent state in the absence
of stimulation, but a brief positive current pulse kicks it into a tonically active state, and an
appropriately-timed positive pulse kicks it back to the silent state (Figure 4A-B).
We fit a GLM to spike trains simulated from such a bistable Izhikevich neuron and found
that the fitted GLM can capture bistable behavior of the original model with high accuracy
(Figure 4C-G). When stimulated during the silent state, the GLM emits a spike due to the
positive output of the stimulus filter (Figure 4E, blue), and tonic firing ensues due to a positive
lobe in the post-spike filter that causes self-excitation at a fixed latency of approximately 10 ms
after the previous spike (Figure 4E). The GLM returns from the active state to the silent state
when a positive stimulus pulse synchronizes negative lobes of the stimulus and post-spike
7
3Izhikevich neuron responseA!EFGinputmVfilter outputsresponsesGLM repeatsconditional!intensitytime (ms)stimulusC!time (ms)(sp/s)Izhikevich = -121fitted GLM parametersBDstimulus filterpost-spike filter outputoutput050-200-150-100-500-100-500-0.100.10.20.30.4010-80-3020-1000-5000500106010100100200300400100 filters. Only the combination of these two negative drives is strong enough to shut off spiking;
without the negative drive created by previous spikes (appearing in the post-spike filter at a
latency of 15 ms after a spike), suppression from the negative lobe of the stimulus filter is
not strong enough to prevent spiking. As with the tonic spiking neuron discussed above, the
interaction between the stimulus and post-spike filters generates rapid rises in the conditional
intensity (Figure 4F), leading to precisely timed spikes that mimic those of the deterministic
Izhikevich neuron (Figure 4G).
This Izhikevich neuron (with the same parameters) can also exhibit a second form of bista-
bility, in which the return to the silent state from the active state is induced by a negative
instead of a positive current pulse. We performed a similar fitting exercise and found that the
GLM is also able to reproduce this behavior. Firing is initiated and maintained by a similar
mechanism as the first form of bistability, but firing offset occurs due to the fact that a negative
stimulus pulse creates immediate negative output from the stimulus filter, which suppresses
firing during the time when a spike would have occurred due to spike-history filter input. Tonic
firing is extinguished more rapidly in this second form of bistability than the first (see Figure 6
Figure 4: Bistable responses. A: Stimulus consisting of two brief positive current pulses.
B: Voltage response of Izhikevich neuron, which exhibits bistability. The first pulse initiates a
tonic spiking mode and the second pulse (precisely timed to the phase of the spike response)
terminates it, returning to a quiescent mode. C: Fitted GLM stimulus filter, which provides a
biphasic impulse response. D: Fitted post-spike filter, which imposes a refractory period of
≈4 ms and gives increased probability of firing ≈5-10 ms after each spike. E: GLM stimulus
(blue) and post-spike (red) filter outputs on a single trial. F: Output of the nonlinearity gives the
conditional intensity for a single trial. G: Spike train of Izhikevich neuron (black) and simulated
repeats of the fitted GLM (gray).
8
Izhikevich neuron responseA!EFGinputmVfilter outputsresponsesGLM repeatsconditional!intensitytime (ms)stimulusC!time (ms)(sp/s)Izhikevich = -204fitted GLM parametersBDstimulus filterpost-spike filter outputoutput0601010103-60-50-40-80-3020-400-200050100-100-500-0.0200.0201020-400-2000 below).
3.4 Type I and type II firing
Neurons have been classified as exhibiting either type I or type II dynamics based on the
shape of their firing rate vs.
intensity (F-I) curve. Type I neurons can fire at arbitrarily low
rates for low levels of injected current, whereas type II neurons have discontinuous F-I curves
that arise from an abrupt transition from silence to a finite non-zero firing rate as the level of
injected current increases [20]. We simulated Izhikevich neurons that exhibit each of these
response types using published parameters.
Inputs consisted of 500 ms current steps of
varying amplitude. The resulting F-I curves for the Izhikevich type I and type II neurons are
shown in black in Figure 5A and Figure 5B, respectively. We fit GLMs using data from each
Izhikevich neuron and found that the fitted GLMs capture the two response types with high
temporal precision. The corresponding F-I curves are shown in gray in Figure 5 and accurately
mimic the behaviors of the Izhikevich neuron. Similar F-I curves have been demonstrated
previously in [15] and [31].
The only discrepancy between the Izhikevich and GLM neurons occurs for the type II cell
at input amplitudes near the Izhikevich neuron's threshold. On some trials when the input
amplitude falls below this threshold, the GLM jumps into a a tonic firing state for the duration
Figure 5: Type I and type II firing curves. A: Top: Example responses of a GLM exhibiting
type I firing behavior. The spike rate increases continuously from zero in response to current
steps of increasing amplitude. Bottom: F-I curve for a type I Izhikevich neuron (black) and
corresponding GLM (gray). For the GLM, responses are plotted for five repetitions of each
input amplitude. B: Similar plots for type II firing behavior, characterized by a discontinuous
jump from zero to a finite spike rate in responses to current steps of increasing amplitude.
9
15202530010203040input amplitudefiring rate (Hz) !! IzhikevichGLM0.20.40.6010203040input amplitudefiring rate (Hz)ABType IType II of the stimulus. Similarly, on some trials when the input amplitude falls above this threshold,
the GLM fails to initiate firing. This is unsurprising given the stochastic nature of the GLM.
Importantly, the GLM never fires at a low rate, but rather abruptly transitions from no firing to
firing at a baseline level of ≈25 Hz, reflecting type II behavior.
3.5 Additional behaviors
We fit GLMs to every dynamical behavior considered in [23] with the exception of purely sub-
threshold behaviors, since GLM fitting uses spike trains and does not consider sub-threshold
responses. The full suite of behaviors is shown in Figure 6, with responses of the Izhikevich
neurons in black and spike responses of the GLM in gray. This list includes tonic and phasic
spiking, tonic and phasic bursting, mixed mode firing, spike frequency adaptation, type I and
type II excitability, two different forms of bistability, and several others that depend primarily
on the shape of stimulus filter. Several additional behaviors that can be captured by a GLM
are not depicted in Figure 6 as they can be achieved by a trivial manipulation of the stimulus
filter; for example, inhibition-induced bursting can be achieved by simply flipping the sign of
the stimulus filter for the bursting neuron shown in Figure 6C. Previous work has shown the
Izhikevich neuron to be capable of producing 18 distinct spiking behaviors [23], and we found
that all can also be produced by a GLM.
3.6 Systematic variation of filter amplitudes
We next considered what happens to the behaviors produced by a GLM as some aspect of the
filters is systematically varied. To do so, we created stimulus and post-spike filters composed
by linear combinations of two basis filters, and then systematically varied the amplitude of one
basis filter while holding the other fixed. (See Methods for details.) Figure 7 shows the phase
space of qualitative spiking behaviors obtained at different points in this 2D filter space.
When the stimulus filter has a strongly negative component (center panel, bottom), a pos-
itive stimulus pulse does not produce enough driving force to cause the neuron to spike at all
(quiescent). As the amplitude of this component of the stimulus filter is increased, the neuron
receives stronger and stronger input and is driven first to spike once or twice (phasic spiking),
and eventually to emit a burst of spikes (phasic bursting). The stimulus filter largely drives
changes between these behaviors, with the additional detail that a strongly negative post-
spike filter component is able to inhibit a burst that would otherwise occur (upper left corner
of "phasic spiking" region). As the stimulus filter component becomes still more positive and
the stimulus filter transitions from being biphasic to more monophasic, it produces a positive
driving force for the duration of the stimulus step, rather than just the onset. This causes the
neuron to fire for the duration of the step.
Importantly, the post-spike filter here determines the nature of this sustained firing. A
post-spike filter that is purely negative beginning at short timescales (top, left side) mimics a
relative refractory period, inhibiting additional spikes for a short window following each elicited
spike and resulting in tonic spiking. If, on the other hand, the post-spike filter is only weakly
negative at short timescales while being more strongly negative at longer timescales, this
creates multiple timescales in the neuron's response (top, right side). At short timescales,
10
Figure 6: Suite of dynamical behaviors of Izhikevich and GLM neurons. Each panel,
top to bottom: stimulus (blue), Izhikevich neuron response (black), GLM responses on five
trials (gray), stimulus filter (left, blue), and post-spike filter (right, red). Black line in each plot
indicates a 50 ms scale bar for the stimulus and spike response. (Differing timescales reflect
timescales used for each behavior in original Izhikevich paper [23]). Stimulus filter and post-
spike filter plots all have 100 ms duration.
there is little inhibition from each spike (beyond the absolute refractory period), so additional
spikes may occur. Over longer timescales, inhibition is accumulated over multiple spikes,
which eventually shuts off spiking. After a brief window of no spikes, the inhibition is relaxed
and spiking commences again until enough inhibition is accumulated to shut spiking off. This
cycle results in tonic bursting for the duration of the step.
In the extreme case where the
post-spike filter is actually biphasic (top, far right), each spike promotes additional spikes on
short timescales, leading to highly regular timing of spikes within bursts (Figure 7C). This set
11
ABCDEFGHIJKLMNOPtonic spikingphasic spikingtonic burstingphasic burstingmixed modetype Itype IIspike latencyresonatorintegratorrebound spikerebound burstvariabilitybistability Ibistability II50 msspike frequencyadaptationthreshold Figure 7: Changes in a single component of each filter can produce a variety of behav-
iors. Center: Amplitudes for a single component of the stimulus filter (ordinate) or post-spike
filter (abscissa) were varied. Responses were simulated for 15 trials of a step stimulus, and
the most common behavior produced is indicated by color. Small panels show example filters
at each extreme of the range tested. A-D: Example responses (gray) to step stimulus (blue)
for GLMs with filters indicated by corresponding letter in center panel.
of behaviors could be achieved by simply sweeping over the amplitude of a single basis vector
in each filter. Incorporating shifts to the basis vectors or additional basis vectors would likely
be necessary to achieve more complex behaviors, such as bistability.
Although we have drawn clear borders at the transition between behaviors, these transi-
tions in fact occur gradually. Near the border between phasic spiking and phasic bursting, for
example, there will be some trials where a single spike is produced and other trials where a
burst is elicited. We have indicated the behavior that is produced most frequently here for sim-
plicity. The transition from tonic bursting to tonic spiking also occurs gradually, with the near
perfectly regular bursting breaking down into more irregular firing until no apparent bursts are
produced. If the post-spike filter is made even more negative than the range explored in this fig-
ure, the timing of tonic spiking becomes near perfectly regular as well. This is easily explained
by the fact that as the post-spike filter becomes more and more negative, it imposes stronger
refractoriness on the cell, which results in more regular spike timing. As the post-spike filter
component amplitude is changed, there is therefore a gradual change from precisely timed
bursts, to irregular firing, to precisely timed tonic spiking. In the following section, we further
explore questions of spike timing precision in the GLM.
3.7 Generalization to new stimuli
The behaviors shows in Figure 6 were all fit using stimuli that probe only a small range of the
possible behaviors of the neuron. For example, many were probed using only a single step
height. A natural question that arises is therefore: how well will these fitted GLMs generalize to
predict the Izhikevich neuron responses to new stimuli? We examined this question for three
canonical Izhikevich neurons from our study (Figure 8). We first generated responses from a
12
component 2 amplitudecomponent 1 amplitudeABCDABDquiescentphasic spikingphasic burstingtonicburstingtonicspiking50 ms50 ms200 msC-1-0.500.51-1.5-1-0.50 Figure 8: GLMs have lim-
ited ability to genearlize to
new stimuli. A: Inputs used
to generate responses from
Izhikevich neurons.
Step
amplitudes were 7, 14, &
28; 0.3, 0.6, & 1.2; and
5, 10, & 20 for B, C, &
D, respectively.
Standard
deviation of noise was 7.2
for all neuron types.
B:
Top: Responses of a regu-
lar spiking Izhikevich neuron
to the above stimuli. Mid-
dle: Spike responses of the
Izhikevich neuron (red) and
GLM (black) fit on responses
to only the middle step size
(indicated by gray box). Bot-
tom: Spike responses of the
Izhikevich neuron (red) and
GLM (black) fit on responses
to all three step sizes (indi-
cated by gray box). C: Re-
sponses of a phasic bursting
Izhikevich neuron. All pan-
els as in B. D: Responses
of a tonic bursting Izhikevich
neuron. All panels as in B &
C.
13
020040002004000200400024602004000246-100-50050ABCD-100-50050024602004000246020040002004000200400002004000200400020040000200400time (ms)mVresponsesresponsesmVresponsesresponsesmVresponsesresponsesinput-100-50050024602004000246020040002004000200400 regular spiking Izhikevich neuron using three step heights and one noise stimulus (Figure 8A;
B, top). We then simulated responses to these stimuli using the GLM fit only to the intermediate
step height (Figure 8B, middle). (This is the same fit as Figure 2.) The GLM responses nearly
perfectly capture the Izhikevich responses for the original stimulus, and the GLM maintains
regular firing patterns for the other step heights. However, the GLM's firing rate is too high
for the small step and too low for the large step. While the GLM accurately captures some
firing events for the noise stimulus, the firing rate is overall too high. We next refit a GLM
to responses from all three step heights, using the same set of basis vectors as the original
fit (Figure 8, bottom). The responses to all three step heights are captured nearly perfectly.
Additionally, the response to the noise stimulus is much more similar to that of the Izhikevich
neuron. Although there is one firing event that occurs at a delay, the regular spiking GLM fit on
an enriched stimulus set is better able to generalize to the noise stimulus.
We performed this same test for neurons showing phasic bursting (Figure 8C) and tonic
bursting (Figure 8D). (Note that although some responses of the phasic bursting Izhikevich
neuron are not actually phasic, we retain the naming convention given to this set of parameters
in the original paper.) For phasic bursting, the GLM fit on additional step sizes improves the
accuracy of responses to the smallest and largest steps, while decreasing the accuracy of
responses for the original step of intermediate size. There is marked improvement in the
accuracy of responses to the noise stimulus, with many firing events being accurately captured
and the GLM no longer exhibiting runaway excitation. For tonic bursting, the refit GLM retains
bursting behavior but fails to even capture responses for the steps on which it was trained.
As noted above, when refitting we used the same set of basis vectors as the initial fits for fair
comparison. It is possible that by increasing the number of basis vectors used or tuning their
properties that better fits to all stimuli might be achieved.
Taken together, these results show that while GLMs might retain some characteristic re-
sponse features (such as bursting) when probed with new stimuli, they often have limited ability
to generalize beyond stimuli on which they are directly fit.
4 Spiking precision and reliability
A noteworthy feature of the spike trains of the GLM neurons considered above is their high
degree of spike timing precision and reliability across trials. This precision arises from the fact
that the conditional intensity (or instantaneous spike rate) rises abruptly at spike times (due to
filter outputs passing through a rapidly accelerating exponential nonlinearity), and decreases
immediately after each spike due to suppressive effects of the post-spike filter. By contrast, a
Poisson GLM without recurrent feedback, more commonly known as a linear-nonlinear Pois-
son (LNP) cascade model, cannot produce temporally precise spike responses to a constant
stimulus because its output is constrained to be a Poisson process.
Real neurons, however, seem to be capable of both response modes: they emit precisely-
timed spikes in some settings and highly variable spike trains in others. A seminal paper by
Mainen & Sejnowski illustrated this duality by showing that spike responses to a constant DC
current exhibit substantial trial-to-trial variability, whereas responses to a rapidly fluctuating
injected current are precise and repeatable across trials [29]. Deterministic models like the
Izhikevich model cannot, of course, mimic this property because their spikes are perfectly
14
Figure 9: Stimulus-dependent spike timing reliability. A: Top: Weak step stimulus. Bottom:
Spike train responses of tonic-spiking GLM on 15 repeated trials. Although the first spike is
precise and reliable, subsequent spikes have irregular timing from trial to trial. B: Top: Rapidly
fluctuating stimulus. Bottom: Spike response of same GLM neuron on 15 repeated trials,
exhibiting a high degree of precision and reliability. (Compare to Figure 1 of [29].)
reproducible for any stimulus. (A stochastic version of the Izhikevich model with an appropriate
level of injected noise could likely overcome this shortcoming, however. See [44] for a similar
case in a Hodgkin-Huxley neuron.) Here we show that the GLM naturally reproduces the
same form of stimulus-dependent changes in precision and reliability observed in real neurons.
Figure 9 shows that a single GLM (with parameters identical to those fit to the tonic-spiking
Izhikevich neuron, shown in Figure 2) produces irregular spiking in response to a constant
stimulus with low-to-intermediate amplitude, and precisely-timed, reliable spikes in response
to a stimulus with large, rapid fluctuations.
5 GLMs can produce super-Poisson variability
We have shown that GLM neurons can reproduce the high degree of spike timing precision
found in real neurons stimulated with injected currents. However, a variety of studies have
reported that neurons exhibit overdispersed responses, or greater-than-Poisson spike count
variability in response to repeated presentations of a sensory stimulus [47, 46, 45, 17]. A
prominent recent study from Goris, Movshon, & Simoncelli showed that the degree of overdis-
persion grows with mean spike count, so that the Fano factor (variance-to-mean ratio) is an
increasing function of spike rate [17]. They proposed a doubly stochastic model to account for
this phenomenon, in which the rate of a Poisson process is modulated by a slowly fluctuating
stochastic gain variable g. For each trial, g is drawn from a gamma distribution with mean 1
and variance σ2
g. (See Methods for details.)
We sought to determine if a GLM with spike-history dependence can also account for
the mean-dependent overdispersion found in neural responses. To test this possibility, we
simulated spike trains from the doubly stochastic model of Goris et al for three different settings
15
0369020040060080051015time (ms)constant inputfluctuating input-50050100020040060080051015time (ms)ABstimulus amplitudetrial g with the same mean spike rate (100 spikes/s), and fit a GLM
of the over-dispersion factor σ2
g. We then simulated responses from each
to the spike trains associated with each value of σ2
GLM to 500 ms pulses at a number of different input intensities, with each point in Figure 10
corresponding to a different intensity.
We found that the GLM can indeed match the qualitative behavior of the Goris et al model,
giving approximately Poisson responses at low spike rates and increasingly overdispersed re-
sponses at higher rates (Figure 10C). To match the data from larger values of overdispersion
g (darker curves in Figure 10C), the GLM relies on increasing amounts of self-excitation
factor σ2
from the post-spike filter, but exhibits no changes in stimulus filter (Figure 10A-B). The filters
here do not include an absolute refractory period, as the original model does not incorporate
one. However, similar results can be achieved when a refractory period is enforced in the train-
ing data. This will result in post-spike filters with strongly negative lobes on short timescales,
which impose refractoriness, but otherwise similar filters to those in Figure 10. Unlike models
with purely suppressive spike history effects, which capture effects due to refractoriness and
reduce variability in firing (e.g., [3, 49]), here we show that allowing spike history effects to be
excitatory can result in increased variability. While the former might be suitable for early sen-
sory areas, such as the retina, the latter better captures the super-Poisson variability observed
in higher visual areas.
Intuitively, the GLM generates overdispersed spike counts because of dependencies the
spike-history filter induces between early and late spikes during a trial:
if the GLM neuron
generates a larger-than-average number of spikes early in a trial, the positive post-spike filter
produces a higher conditional intensity (and hence more spiking) later in the trial; conversely, if
a neuron emits fewer-than-average spikes early in a trial, the conditional intensity will be lower
later in the trial (yielding less spiking). The Goris et al model can be seen to capture similar
dependencies between early and late spikes via the stochastic gain variable g, which is con-
stant during a trial but independent across trials. Thus, it is reasonable to view g in the Goris et
Figure 10: GLMs can produce super-Poisson variability. A: Stimulus filters for GLMs trained
on three different levels of variability: high (dark blue), medium (medium), and low (light) super-
Poisson variability. Spike count mean was identical in the three cases: 100 spikes/s. B: Post-
spike filters for high (dark red), medium (medium), and low (light) super-Poisson variability. C:
Spike count variance versus mean for three levels of variability. This relationship is strikingly
similar to that observed in many cortical neurons.
16
0204060801000100200300400spike count meanPoisson spikingspike count variance-50-25000.10.2020040000.02time (ms) = -0.84 = -1.11 = -0.86time (ms)Astimulus filtersspike-history filtersBC al model as a proxy for the accumulated self-excitation from spike-history filter outputs under
a GLM. We note, however, that attempts to drive the GLM to higher levels of overdispersion
(e.g., Fano factors significantly > 3) often resulted in runaway self-excitation, indicating that
GLMs may require additional mechanisms to maintain stability in order to produce highly over-
dispersed responses through recurrent excitation alone [11, 19]. An alternative mechanism
for generating over-dispersed responses with GLMs is through the addition of latent stochastic
inputs [42], an avenue we have not explored here.
6 Discussion
We have shown that recurrent Poisson GLMs can capture an extensive set of behaviors exhib-
ited by biological neurons, including tonic and phasic spiking, bursting, spike frequency adap-
tation, type I and type II behavior, and bistability. GLMs can also reproduce widely varying
levels of response stochasticity, ranging from precisely timed spikes with negligible trial-to-trial
variability, to substantially super-Poisson spike count variability. We have also shown that, like
real neurons, GLMs can exhibit irregular firing in response to a constant stimulus, but precise
and repeatable firing patterns in response to a temporally varying stimulus. Thus, generalized
linear models are able to capture a rich array of spiking behaviors like many dynamical models,
while remaining tractable to fit to neural data.
6.1 Relationship to previous work
As mentioned above, GLMs have strong connections to a number of other models. It is partic-
ularly worth noting the connection between GLMs and generalized integrate-and-fire models,
such as the spike response model (SRM) extensively studied by Gerstner and colleagues
[15, 13]. These models draw on much earlier work which incorporated a variable threshold
that depends on spiking history [51, 10]. The SRM includes a membrane filter (analogous to
the stimulus filter here) and both a spike afterpotential and moving threshold (which can be
combined and are analogous to the spike history filter here). In its simplest formulation, the
SRM is a deterministic model. Although the threshold for spiking can shift as a function of
spike history, a spike will occur precisely at each threshold crossing.
Extensions of the model have incorporated so-called "escape noise," where spiking no
longer occurs deterministically at threshold crossings, but rather the probability of spiking de-
pends on the distance of the membrane voltage to threshold [15, 25]. This variant of the SRM
is in fact a GLM, and it is therefore worthwhile to consider how previous work investigating
the SRM with escape noise relates to our results here. Early work demonstrated that such a
model was capable of producing responses with high temporal precision, including both tonic
spiking as well as tonic bursting [15], though demonstrating the range of behaviors that could
be produced by the SRM was not the focus of this study. Additional work demonstrated that the
model could produce highly repeatable spike trains to a noisy stimulus (similar to Figure 9B,
though no comparison of irregular firing in response to a constant stimulus was shown) [25].
Other studies have shown that the SRM can capture the detailed statistics of neural responses
[31, 41]. Further, many of these studies show that the spike responses model can be used
to capture not only the relationship between an external stimulus and a neuron's response,
17
but also to faithfully capture the relationship between intracellularly recorded neural responses
and injected current [25, 41].
6.2 Limitations
Despite their many advantages, GLMs have several limitations that bear further discussion.
First, GLMs often do not generalize well across stimulus distributions; models fit with a partic-
ular set of stimuli often do not accurately predict responses to stimuli with markedly different
statistics (e.g., stimuli with large changes in mean or variance, or white noise vs. naturalistic
stimuli) [18].
Secondly, GLMs often lack clear interpretability in terms of underlying mechanisms. This
stands in contrast to dynamical models designed to capture specific biophysical variables and
processes. In the two-dimensional Izhikevich model, for example, one variable (v) represents
the neuron's membrane potential, and the other variable (u) can be understood as a membrane
recovery variable, which reflects K+ channel activation and Na+ channel inactivation. Despite
the fact the GLM filters do not represent specific biophysical variables, in some cases they can
still provide insight into underlying biological processes. For example, recent work provides an
interpretation of the GLM as a synaptic conductance based model with linear sub-threshold
dynamics [27]. Work on the SRM has shown that by dividing the effects of spike history into a
dynamic threshold and a spike afterpotential, one can in fact measure their separate contribu-
tions with intracellular recordings of a neuron's subthreshold voltage; the spike afterpotential
can be observed directly in this voltage trace, while the effects on thereshold can be estimated
indirectly by noting the absence of firing [31, 41, 14].
A third known limitation of GLMs is that they lack the flexibility to capture some nonlinear
response properties of real spike trains. For example, as point neuron models, GLMs do not
reflect the fact that neurons often receive spatially segregated inputs on the dendritic tree, and
these inputs can be processed separately and combined nonlinearly [28]. Some extensions of
the GLM that incorporate nonlinear inputs and multiple subunits [8, 43, 36, 7, 30, 1, 52] may
begin to address this issue, but certainly fall short of capturing the full complexity of dendritic
processing. For the range of dynamical behaviors considered here, however, we did not find
these extensions to be necessary.
For all results shown, we used a GLM with an exponential nonlinearity. To test the depen-
dence of our results on the form of the nonlinearity, we also fit GLMs to several of the behaviors
with a "soft-rectifying" nonlinearity given by f (x) = log (1 + exp (x)). This function grows only
linearly for large input values, but still has an exponential decay on its left tail and remains
in the family of nonlinearities (convex and log-concave) for which the GLM log-likelihood is
provably concave [34]. For the behaviors tested (tonic spiking, tonic bursting, phasic spiking,
and phasic bursting), our results were similar to those with an exponential nonlinearity, though
generally not as temporally precise. This increased precision is likely due to the fact that an
exponential nonlinearity rises more steeply than a linear-rectifying function, causing the condi-
tional intensity to accelerate more rapidly from a low-probability to a high-probability of spiking
regime. Past studies have found that responses of both retinal ganglion cells and neocortical
pyramidal neurons are well described by a GLM with exponential nonlinearity [40, 25].
It is worth noting that for many of the dynamic behaviors studied here, the GLM parameters
18
were not strongly constrained by the training data. (See "Sensitivity to changes in parameter
values" in Appendix.) Slight changes in the model parameters did not produce noticeable
changes in response, at least for the stereotyped range of input currents and output spike
patterns considered. The filter parameters were therefore only weakly identifiable, which cor-
responds to a likelihood function with a very gradual falloff along certain directions in parameter
space. This uncertainty potentially complicates interpretation of the filters in terms of functions
performed by the underlying biophysical mechanism. Conversely, it reveals that the suite of
behaviors considered by Izhikevich and others can be achieved by a range of different GLMs,
and that a richer set of input-output patterns is needed to identify a unique set of GLM param-
eters.
Conclusion
The GLM has the ability to mimic a wide range of biophysically realistic behaviors exhibited by
real neurons. Although it is clear there are some forms of nonlinear behavior it cannot produce,
such as frequency-doubled responses of cat Y cells or V1 complex cells to a contrast-reversing
grating, our work provides an existence proof for its ability to exhibit an important range of re-
sponse types considered previously only in biophysics and applied math modeling literature.
Moreover, by considering response stochasticity as another dimension along which real neu-
rons vary, we have shown that that the GLM can generate response characteristics ranging
from quasi-deterministic to greater-than-Poisson variability. The GLM therefore provides a
flexible yet powerful tool for studying the dynamics of real neurons and the computations they
carry out.
Appendix
MATLAB code used to generate example responses from Izhikevich neurons and to fit GLMs
to these responses is available in a Github repository (https://github.com/aiweber/
GLM_and_Izhikevich).
Izhikevich model simulations
To generate training data for fitting the GLM, we simulated responses from an Izhikevich model
[22] with parameters set to published values given for each behavior in [23] (Table 1; parameter
values can be found at http://www.izhikevich.org/publications/izhikevich.m). For each behavior,
we generated approximately 20 seconds of training data using the forward Euler method with
fixed time step size (dt) given in Table 1. It should be noted that in some cases, published pa-
rameter values did not produce the desired qualitative behavior. In these cases, we tuned the
simulation parameters to achieve the desired behavior. Parameters marked with an asterisk in
Table 1 indicate those that differ from published values for the corresponding behavior in [23].
Additionally, some behaviors of the Izhikevich neuron are not robust to small changes in stim-
ulus timing, stimulus amplitude, or time step of integration. In particular, we found bistability to
19
be highly dependent on the precise stimulus timing (onset and duration), stimulus amplitude,
and integration window. We tuned these values by hand to produce the desired behavior.
GLM fitting and simulations
A generalized linear model for a single neuron attributes features of a spike train to both stimu-
lus dependence and spike history. Stimulus dependence is captured by a stimulus filter k, and
spike-history dependence is captured by a post-spike filter h. k and h are represented with a
raised cosine basis to reduce to the dimensionality necessary to fit and ensure smoothness of
the filters. Basis vectors are of the form:
1
2
cos(a log[t + c] − φj) +
bj(t) =
(6)
for t such that a log(t + c) ∈ [φj − π, φj + π] and 0 elsewhere. The parameter c determines the
extent to which peaks of the basis vectors are linearly spaced, with larger values of c resulting
in more linear spacing. We typically used 6 such basis vectors to fit a 100 ms stimulus filter
k and 8 basis vectors to fit a 150 ms post-spike filter h, for a total of 15 parameters (including
1
2
a
mixed mode
type I
type II
spike latency
neuron type
tonic spiking
phasic spiking
tonic bursting
phasic bursting
b
0.02
0.2
0.02 0.25
0.02
0.2
0.02 0.25
0.2
0.02
spike frequency adaptation 0.01
0.2
-0.1
0.02
0.26
0.2
0.02
0.2
0.26
0.1*
0.02
-0.1
0.03 0.25
0.03 0.25
0.03 0.25
1.5
1.5
rebound spike
rebound burst
bistability I
bistability II
resonator
integrator
threshold variability
1
1
c
-65
-65
-50
-55
-55
-65
-55
-65
-65
-60
-66*
-60
-52
-60
-60
-60
d
6
6
2
0.05
4
5*
6
0
6
-1
6
4
0
4
0
0
I
14
0.5
10*
0.6
10
20*
25
0.5
3.49*
0.3
27.4
-5
-5
2.3
30*
40
dt (ms)
0.1
0.1
0.1
0.1
0.1
0.1
1
1
0.1
0.5
0.5
0.1
0.1
1
0.05
0.05
Table 1: Parameters of the Izhikevich neuron for dynamic behaviors shown in Figures 2-6, 8-
9, & 11. Parameters marked with * indicate parameters that differ from those used in [23].
Additionally, only a single form of bistability (bistability I) was presented in [23].
20
In some cases, as few as 7 or as many as
one for µ that determines baseline firing rate).
26 parameters were used to fit an individual Izhikevich neuron's behavior.
In general, the
fewest number of basis vectors required to reproduce a given behavior were used, though it
is likely that by altering specific features of the basis vectors (e.g., their spacing), even fewer
parameters would suffice.
We fit the model parameters (weights on the basis functions for k, weights on the basis
functions of h, and µ) by maximizing the log-likelihood:
log λ(t) − ∆
L(θ) =
(cid:88)
(cid:88)
λ(t)
(7)
t=spike
t
where ∆ is the time resolution of y(t). We used MATLAB's fminunc function, part of the
MATLAB optimization toolbox, to find the global maximum of the likelihood function.
We simulated the GLM response in time bins of the same size as the corresponding Izhike-
vich neuron and computed the single-bin probability of a spike as
P (y(t) ≥ 1λ(t)) = 1 − P (y(t) = 0λ(t)) = 1 − exp(∆λ(t)),
(8)
where ∆ is the time bin size, so that the probability of 0 or 1 spikes in a bin sums to 1 (resulting
in a Bernoulli approximation to the Poisson process), disallowing spike counts greater than 1
in a single bin.
Systematic variation of filter amplitudes
In order to more carefully examine the transitions between different behaviors as the stimulus
and post-spike filter change, we systematically varied the amplitude of individual filter compo-
nents and observed the behavior produced. Each filter was parameterized with 2 components.
The amplitude of one was fixed while the amplitude of the other was varied. For the stimulus
filter, the amplitude of the second component was varied (-1.5 to +0.25). This allowed us to
transition from monophasic to biphasic filters. The amplitude of the first component was set to
be positive (+1), creating an "ON" filter appropriate for a positive step stimulus. For the post-
spike filter, the amplitude of the first component was varied (-1 to +1). The amplitude of the
second component was set to be negative (-3), ensuring that spiking would be suppressed on
longer timescales. For the post-spike filter we also imposed an absolute refractory period of 5
ms. Finally, we included a negative baseline drive (µ = -1) to suppress spontaneous spiking
so that the baseline firing rate was zero.
We simulated responses to 25 identical step stimuli for each set of filters and then classi-
fied the behaviors as quiescent, phasic spiking, phasic bursting, tonic spiking, or tonic bursting.
The most commonly observed behavior over the 25 repetitions is depicted in Figure 7. Re-
sponses were classified in the following way. If no spikes were elicited in the first 200 ms of
stimulus presentation and fewer than 5 spikes were elicited during the final 10 seconds of stim-
ulus presentation, the behavior was classified as quiescent. If at least one spike was elicited
in the first 200 ms following stimulus onset and fewer than 5 spikes were elicited during the
final 10 seconds of stimulus presentation, the behavior was classified as phasic. Phasic firing
patterns were further classified into phasic spiking if only 1 or 2 spikes were elicited in the
first 200 ms, and phasic bursting if 3 or more spikes were elicited in the first 200 ms. The
21
remaining responses were classified as either tonic spiking or tonic bursting in the following
manner.
Inter-spike interval distributions were fit with a Gaussian mixture distribution using
MATLAB's gmdistribution function. We fit both a single Gaussian distribution as well as
a mixture of two Gaussians and then compared the Akaike information criterion (AIC) values
to determine whether the ISI distribution was better fit as a unimodal distribution or a bimodal
If 0.9 · AICunimodal < AICbimodal, the spike
distribution, with a lower AIC indicating better fit.
train was classified as tonic spiking; otherwise, it was classified as tonic bursting. (We added
the 0.9 factor to create a more stringent standard for what is classified as bursting activity so
that the responses that fall into this category are strongly bimodal distributions that would be
readily identified as bursting. Slightly altering the value of this factor, or eliminating it entirely,
gives the same qualitative results, but merely shifts the boundary in Figure 7 between the tonic
spiking and tonic bursting regions.)
Figure 11: Sensitivity to changes in parameter values for a regular spiking (left) and tonic
bursting (right) GLM. A: Fit coefficient of each eigenvector of Hessian matrix of likelihood,
normalized by corresponding eigenvalue. Eigenvectors are in order of decreasing eigenvalues
(not necessarily decreasing z-scored eigenvalues). B: Stimulus filter (top, blue) and spike
history filter (bottom, red), along with two most constrained eigenvectors. These correspond
to the largest (dark gray) and second largest (light gray) eigenvalues. Eigenvectors are scaled
to size comparable with filters. C: Same as B, for least constrained eigenvectors. D-F: Same
as A-D for tonic bursting neuron.
22
DEF246810121410ï1100101102eigenvector0255075100-0.200.20.40.6most constrained0255075100-300-200-1000100time (ms)0255075100least constrained0255075100time (ms)tonic burstingz-scoredcoefficientfilter
amplitudefilter
amplitudeAB246810121410ï810ï4
100 eigenvector0255075100-1012most constrained050100150-1000-5000time (ms)0255075100least constrained050100150time (ms)regular spikingCfilter
amplitudez-scoredcoefficientfilter
amplitude Sensitivity to changes in parameter values
We wished to investigate how well constrained different features of our fit GLMs were. To
do so, we calculated the eigendecomposition of the Hessian matrix of the likelihood function:
H = QΛQ−1. The Hessian matrix provides a local quadratic approximation to the likelihood
function, with eigenvectors qi pointing along the principal axes and length of these axes propor-
tional to 1√
. Thus, larger magnitude eigenvalues indicate greater curvature (i.e., shorter axes)
λi
and better constrained directions, while smaller magnitude eigenvalues indicate lower curva-
ture and more poorly constrained directions. Results of this analysis are shown in Figure 11
for both a regular spiking neuron (left) and tonic bursting neuron (right). For both neurons, the
least constrained directions correspond to eigenvectors similar in shape to the best-fit filters or
the absolute refractory period. As such, perturbations in these directions do not result in large
changes in the behavior. Perturbations along eigenvectors corresponding to the most con-
strained directions, on the other hand, would result in significant changes to the filter shape.
The difference in scale between Panel A and Panel D indicates that overall, parameters for the
tonic bursting neuron are more constrained than those for the regular spiking neuron.
Doubly stochastic model with super-Poisson variability
In Figure 10, we used a negative binomial model to generate spike trains with greater-than-
Poisson variability [38, 17]. The negative binomial distribution can be conceived as a doubly-
stochastic model in which the rate of a Poisson process is modulated by an iid gamma random
variable on each trial. Following [17], we modeled responses with a stochastic gain variable g
with mean 1 and variance σ2
g that obeys a gamma distribution:
P (gr, s) =
gr−1 exp
1
srΓ(r)
(cid:17)
(cid:16)−g
s
,
(9)
where s = σ2
represents the gamma function.
g denotes the scale parameter, r = 1/σ2
g is the shape parameter, and Γ(·)
The spike count conditioned on g and a stimulus S for each trial then obeys a Poisson
distribution:
P (yg, S) =
(∆gf (S))y
y!
exp (−∆gf (S)) ,
(10)
where ∆ is the time bin size, g is the gain, and f (S) is the tuning curve that specifies the
mean response to stimulus S.
g = 0, the gain g is deterministically equal to 1
and the spike count is Poisson with mean and variance equal to ∆f (S). For responses with
g > 0, however, responses are overdispersed relative to the Poisson distribution and have
σ2
mean ∆f (S) and variance ∆f (S)(1 + σ2
In the limit σ2
For the results shown in Figure 10, we simulated data from the negative binomial dis-
g ∈
tribution with a single mean rate f (S) (100 spikes/s) at three different gain variances σ2
{.0125, 0.02, 0.05}. Spike counts were drawn iid across trials, with spike times distributed uni-
formly within each trial to generate spike trains suitable for GLM fitting. We used these spike
trains to fit a GLM to the data associated with each value of σ2
g, with an assumed constant
input current for each trial.
g∆f (S)).
23
Acknowledgments
We would like to thank Adrienne Fairhall, Sara Solla, and James Fitzgerald for helpful com-
ments and discussions.
References
[1] Ahrens, M. B., Linden, J. F., and Sahani, M. (2008). Nonlinearities and contextual influ-
ences in auditory cortical responses modeled with multilinear spectrotemporal methods. J
Neurosci, 28(8):1929–1942.
[2] Babadi, B., Casti, A., Xiao, Y., Kaplan, E., and Paninski, L. (2010). A generalized linear
model of the impact of direct and indirect inputs to the lateral geniculate nucleus. Journal of
Vision, 10(10):22.
[3] Berry, M. and Meister, M. (1998). Refractoriness and neural precision. Journal of Neuro-
science, 18:2200–2211.
[4] Brette, R. and Gerstner, W. (2005). Adaptive exponential integrate-and-fire model as an
effective description of neuronal activity. Journal of neurophysiology, 94(5):3637–3642.
[5] Calabrese, A., Schumacher, J. W., Schneider, D. M., Paninski, L., and Woolley, S. M. N.
(2011). A generalized linear model for estimating spectrotemporal receptive fields from
responses to natural sounds. PLoS One, 6(1):e16104.
[6] Cox, D. and Isham, V. (1980). Point Processes. Chapman & Hall/CRC Monographs on
Statistics & Applied Probability. Taylor & Francis.
[7] Cui, Y., Liu, L. D., Khawaja, F. A., Pack, C. C., and Butts, D. A. (2013). Diverse sup-
pressive influences in area mt and selectivity to complex motion features. The Journal of
Neuroscience, 33(42):16715–16728.
[8] Fitzgerald, J. D., Rowekamp, R. J., Sincich, L. C., and Sharpee, T. O. (2011). Second order
dimensionality reduction using minimum and maximum mutual information models. PLoS
Comput Biol, 7(10):e1002249.
[9] FitzHugh, R. (1961).
Impulses and physiological states in theoretical models of nerve
membrane. Biophysical journal, 1(6):445.
[10] Geisler, C. D. and Goldberg, J. M. (1966). A Stochastic Model of the Repetitive Activity of
Neurons. Biophysical Journal, 6(1):53–69.
[11] Gerhard, F., Deger, M., and Truccolo, W. (2017). On the stability and dynamics of stochas-
tic spiking neuron models: Nonlinear hawkes process and point process glms. PLoS com-
putational biology, 13(2):e1005390.
[12] Gerstner, W. (1995). Time structure of the activity in neural network models. Physical
Review E, 51(1):738–758.
24
[13] Gerstner, W. (2001). A framework for spiking neuron models: The spike response model.
In Moss, F. and Gielen, S., editors, The Handbook of Biological Physics, volume 4, pages
469–516.
[14] Gerstner, W., Kistler, W. M., Naud, R., and Paninski, L. (2014). Neuronal Dynamics:
From Single Neurons to Networks and Models of Cognition. Cambridge University Press,
New York, NY, USA.
[15] Gerstner, W. and van Hemmen, J. (1992). Associative memory in a network of 'spiking'
neurons. Network: Computation in Neural Systems, 3(2):139–164.
[16] Gerstner, W., van Hemmen, J. L., and Cowan, J. D. (1996). What matters in neuronal
locking? Neural computation, 8:1653–1676.
[17] Goris, R. L. T., Movshon, J. A., and Simoncelli, E. P. (2014). Partitioning neuronal vari-
ability. Nat Neurosci, 17(6):858–865.
[18] Heitman, A., Greschner, M., Field, G., Li, P., Ahn, D., Sher, A., Litke, A., and Chichilnisky,
In
E. (2014). Representation and reconstruction of natural scenes in the primate retina.
Computational and Systems Neuroscience (CoSyNe) Abstracts, pages 134 – 135.
[19] Hocker, D. and Park, I. M. (2017). Multistep inference for generalized linear spiking mod-
els curbs runaway excitation. In 8th International IEEE EMBS Conference On Neural Engi-
neering.
[20] Hodgkin, A. (1948). The local electric changes associated with repetitive action in a non-
medullated axon. The Journal of physiology, 107(2):165–181.
[21] Hodgkin, A. L. and Huxley, A. F. (1952). A quantitative description of membrane current
and its application to conduction and excitation in nerve. J. Physiol., 117(4):500–544.
[22] Izhikevich, E. M. (2003). Simple model of spiking neurons.
IEEE Trans Neural Netw,
14(6):1569–1572.
[23] Izhikevich, E. M. (2004). Which model to use for cortical spiking neurons? IEEE Trans
Neural Netw, 15(5):1063–1070.
[24] Jolivet, R., Lewis, T., and Gerstner, W. (2003). The spike response model: a framework to
predict neuronal spike trains. Springer Lecture notes in computer science, 2714:846–853.
[25] Jolivet, R., Rauch, A., Lüscher, H. R., and Gerstner, W. (2006). Predicting spike timing
of neocortical pyramidal neurons by simple threshold models. Journal of Computational
Neuroscience, 21(1):35–49.
[26] Keat, J., Reinagel, P., Reid, R., and Meister, M. (2001). Predicting every spike: a model
for the responses of visual neurons. Neuron, 30:803–817.
25
[27] Latimer, K. W., Chichilnisky, E. J., Rieke, F., and Pillow, J. W. (2014).
Inferring synap-
tic conductances from spike trains with a biophysically inspired point process model.
In
Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., and Weinberger, K., editors, Ad-
vances in Neural Information Processing Systems 27, pages 954–962. Curran Associates,
Inc.
[28] London, M. and Häusser, M. (2005). Dendritic Computation. Annual Review of Neuro-
science, 28(1):503–532.
[29] Mainen, Z. and Sejnowski, T. (1995). Reliability of spike timing in neocortical neurons.
Science, 268:1503–1506.
[30] McFarland, J. M., Cui, Y., and Butts, D. A. (2013). Inferring nonlinear neuronal computa-
tion based on physiologically plausible inputs. PLoS Comput Biol, 9(7):e1003143+.
[31] Mensi, S., Naud, R., Pozzorini, C., Avermann, M., Petersen, C. C. H., and Gerstner, W.
(2012). Parameter extraction and classification of three cortical neuron types reveals two
distinct adaptation mechanisms. Journal of Neurophysiology, 107:1756–1775.
[32] Morris, C. and Lecar, H. (1981). Voltage oscillations in the barnacle giant muscle fiber.
Biophysical journal, 35(July):193–213.
[33] Nagumo, J., Arimoto, S., and Yoshizawa, S. (1962). An Active Pulse Transmission Line
Simulating Nerve Axon. Proceedings of the IRE, 117(m V):2061–2070.
[34] Paninski, L. (2004). Maximum likelihood estimation of cascade point-process neural en-
coding models. Network: Computation in Neural Systems, 15:243–262.
[35] Paninski, L., Pillow, J. W., and Simoncelli, E. P. (2004). Maximum likelihood estimation of
a stochastic integrate-and-fire neural model. Neural Computation, 16:2533–2561.
[36] Park, I. M., Archer, E. W., Priebe, N., and Pillow, J. W. (2013). Spectral methods for neural
In Advances in Neural Information
characterization using generalized quadratic models.
Processing Systems 26, pages 2454–2462.
[37] Perkel, D. H., Gerstein, G. L., and Moore, G. P. (1967). Neuronal spike trains and stochas-
tic point processes i. the single spike train. Biophysical Journal, 7(4):391–418.
[38] Pillow, J. and Scott, J. (2012). Fully bayesian inference for neural models with negative-
In Bartlett, P., Pereira, F., Burges, C., Bottou, L., and Weinberger, K.,
binomial spiking.
editors, Advances in Neural Information Processing Systems 25, pages 1907–1915.
[39] Pillow, J. W., Paninski, L., Uzzell, V. J., Simoncelli, E. P., and Chichilnisky, E. J. (2005).
Prediction and decoding of retinal ganglion cell responses with a probabilistic spiking model.
The Journal of Neuroscience, 25:11003–11013.
[40] Pillow, J. W., Shlens, J., Paninski, L., Sher, A., Litke, A. M., and Chichilnisky, E. J. Simon-
celli, E. P. (2008). Spatio-temporal correlations and visual signaling in a complete neuronal
population. Nature, 454:995–999.
26
[41] Pozzorini, C., Naud, R., Mensi, S., and Gerstner, W. (2013). Temporal whitening by
power-law adaptation in neocortical neurons. Nature Neuroscience, 16(7):942–948.
[42] Rabinowitz, N. C., Goris, R. L., Cohen, M., and Simoncelli, E. (2015). Attention stabilizes
the shared gain of v4 populations. eLife.
[43] Rajan, K., Marre, O., and Tkacik, G. (2013). Learning quadratic receptive fields from
neural responses to natural stimuli. Neural Computation, 25(7):1661–1692.
[44] Schneidman, E., Freedman, B., and Segev, I. (1998).
Ion channel stochasticity may
be critical in determining the reliability and precision of spike timing. Neural computation,
10(7):1679–703.
[45] Shadlen, M. and Newsome, W. (1998). The variable discharge of cortical neurons: im-
plications for connectivity, computation, and information coding. Journal of Neuroscience,
18:3870–3896.
[46] Tolhurst, D. J., Movshon, J. A., and Dean, A. F. (1983). The statistical reliability of signals
in single neurons in cat and monkey visual cortex. Vision Res, 23(8):775–785.
[47] Tomko, G. J. and Crapper, D. R. (1974). Neuronal variability: non-stationary responses
to identical visual stimuli. Brain research, 79(3):405–418.
[48] Truccolo, W., Eden, U. T., Fellows, M. R., Donoghue, J. P., and Brown, E. N. (2005). A
point process framework for relating neural spiking activity to spiking history, neural ensem-
ble and extrinsic covariate effects. J. Neurophysiol, 93(2):1074–1089.
[49] Uzzell, V. J. (2004). Precision of Spike Trains in Primate Retinal Ganglion Cells. Journal
of Neurophysiology, 92(2):780–789.
[50] Weber, F., Machens, C. K., and Borst, A. (2012). Disentangling the functional conse-
quences of the connectivity between optic-flow processing neurons. Nature neuroscience,
15(3):441–448.
[51] Weiss, T. F. (1966). A model of the peripheral auditory system. Kybernetik, 3(4):153–175.
[52] Williamson, R. S., Sahani, M., and Pillow, J. W. (2015). The equivalence of information-
theoretic and likelihood-based methods for neural dimensionality reduction. PLoS Comput
Biol, 11(4):e1004141.
[53] Zhao, M. and Iyengar, S. (2010). Nonconvergence in logistic and poisson models for
neural spiking. Neural Computation, 22(5):1231–1244. PMID: 20100077.
27
|
1008.1410 | 1 | 1008 | 2010-08-08T15:13:12 | Neuronal Response Clamp | [
"q-bio.NC"
] | Since the first recordings made of evoked action potentials it has become apparent that the responses of individual neurons to ongoing physiologically relevant input, are highly variable. This variability is manifested in non-stationary behavior of practically every observable neuronal response feature. Here we introduce the Neuronal Response Clamp, a closed-loop technique enabling full control over two important single neuron activity variables: response probability and stimulus-spike latency. The technique is applicable over extended durations (up to several hours), and is effective even on the background of ongoing neuronal network activity. The Response Clamp technique is a powerful tool, extending the voltage-clamp and dynamic-clamp approaches to the neuron's functional level, namely - its spiking behavior. | q-bio.NC | q-bio | Neuronal Response Clamp
immediate
November 20, 2018
Avner Wallach1,2, Danny Eytan1,3, Asaf Gal1,4, Christoph Zrenner1,
Ron Meir1,2 and Shimon Marom1
1 Network Biology Research Laboratories, Lorry Lokey Interdisciplinary Center
for Life Sciences and Engineering, Technion, Haifa, Israel
2 Faculty of Electrical Engineering, Technion, Haifa, Israel
3 Rambam Medical Center, Haifa, Israel
4 The Interdisciplinary Center for Neural Computation (ICNC), The Hebrew Uni-
versity, Jerusalem, Israel
0
1
0
2
g
u
A
8
]
.
C
N
o
i
b
-
q
[
1
v
0
1
4
1
.
8
0
0
1
:
v
i
X
r
a
1
Abstract
Since the first recordings made of evoked action potentials it has be-
come apparent that the responses of individual neurons to ongoing
physiologically relevant input, are highly variable. This variability is
manifested in non-stationary behavior of practically every observable
neuronal response feature. Here we introduce the Neuronal Response
Clamp, a closed-loop technique enabling full control over two important
single neuron activity variables: response probability and stimulus-
spike latency. The technique is applicable over extended durations (up
to several hours), and is effective even on the background of ongoing
neuronal network activity. The Response Clamp technique is a power-
ful tool, extending the voltage-clamp and dynamic-clamp approaches
to the neuron's functional level, namely - its spiking behavior.
Keywords: Excitability ; Neuron ; Clamp
2
Introduction
The responses of single neurons and neuronal populations to ongoing
stimuli within a physiological range of frequencies are highly variable;
this variability is manifested in both the ability to evoke an action
potential and the latency between stimulus and response. Response
variability was already obvious a century ago, when Adrian and Zot-
terman described a "lack of regularity" in the "all-or-none" impulse
response in their classical paper (Adrian & Zotterman, 1926), the first
documentation of recorded action potentials. Nowadays, response vari-
ability is generally accepted to be an important phenomenon (Stein,
1965; Mainen & Sejnowski, 1995; Arieli, Sterkin, Grinvald, & Aertsen,
1996; Reich, Victor, Knight, Ozaki, & Kaplan, 1997; Arsiero, Luscher,
Lundstrom, & Giugliano, 2007; Soteropoulos & Baker, 2009) that re-
flects the immensity of involved cellular level machineries, covering any
observable cell-physiology timescale (Marom, 2010). Analyses of the
sources of response variability, as well as its impact on neuronal func-
tionality, is practically intractable due to the non-linearity involved in
the reciprocal relations between response variability and its underlying
sources.
It would be highly desirable to develop a method for controlling
the reliability of neuronal responses to input. By studying the sig-
nals required to achieve this control one may hopefully shed light on
the input-output relationships of the neuron as well as lead to a bet-
ter understaning of the underlying sources of response variability. In
this context, analogy to the voltage-clamp method (Hodgkin, Huxley,
& Katz, 1952) immediately comes to mind. There, Hodgkin, Hux-
ley and Katz studied the dependency of conductances on voltage by
"clamping" the membrane potential and analyzing the required control
signals (currents). By doing so, they broke the circular relationships
between membrane voltage and voltage-dependent conductances; in-
deed, one cannot imagine neuronal physiology without the insights
that the voltage-clamp method provided throughout the years.
3
In this study we present and demonstrate the concept of Neuronal
Response Clamp. The idea is to "clamp" a defined feature of the neu-
ronal spiking response. This is achieved by implementing a control cir-
cuit that measures pre-defined neuronal response characteristics, com-
pares them to desired values, and corrects errors by changing stimu-
lation features. Specifically, we devised a closed-loop real-time system
that implements what is known as a Proportional-Integral-Derivative
(PID) controller in system engineering (Levine et al., 1996), to com-
pute the "error" signal between desired and actual neuronal responses,
and correct the error by manipulating the stimulation frequency. Here
we show that this system fully restrains the variability of neuronal spik-
ing probability, and provides complete control over the precise latency
between stimulus and response, both in synaptically isolated neurons,
and within the context of active networks.
4
Materials and Methods
Cell preparation
Cortical neurons were obtained from newborn rats (Sprague-Dawley)
within 24 hours after birth using mechanical and enzymatic procedures
described in earlier studies (Marom & Shahaf, 2002). The neurons
were plated directly onto substrate-integrated multi electrode arrays
and allowed to develop functionally and structurally mature networks
over a time period of 2-3 weeks. The number of neurons in a typi-
cal network is on the order of 10,000. The preparations were bathed
in MEM supplemented with heat-inactivated horse serum (5%), glu-
tamine (0.5 mM), glucose (20 mM), and gentamycin (10 g/ml), and
maintained in an atmosphere of 37◦C, 5% CO2 and 95% air in an incu-
bator as well as during the recording phases. An array of Ti/Au/TiN
extracellular electrodes, 30µm in diameter, and spaced either 500 µm
or 200 µm from each other (MultiChannelSystems, Reutlingen, Ger-
many) were used. The insulation layer (silicon nitride) was pre-treated
with polyethyleneimine (Sigma, 0.01% in 0.1M Borate buffer solu-
tion). To completely block synaptic transmission in the network, 20
µM APV (amino-5-phosphonovaleric acid), 10 µM CNQX (6-cyano-7-
nitroquinoxaline-2,3-dione), and 5 µM Bicuculline were added to the
bathing solution.
Measurements and Stimulation
A commercial amplifier (MEA-1060-inv-BC, MCS, Reutlingen, Ger-
many) with frequency limits of 150-3,000Hz and a gain of x1024 was
used. Rectangular 200 µs biphasic 20-50 µA current or 600-800 mV
stimulation through extracellular electrodes was performed using a
dedicated stimulus generator (MCS, Reutlingen, Germany). Data was
digitized using data acquisition board (PD2-MF-64-3M/12H, UEI, Walpole,
MA, USA). Each channel was sampled at a frequency of 16 Ksample/s
(for spike probability experiments) or 96 Ksamples/s (for spike latency
5
experiments). Action potentials were detected online by threshold
crossing. All spike times and shapes, as well as 15 ms voltage traces
from all electrodes after each stimulus, were recorded for analyses.
Data processing and closed-loop stimulation was performed using a
Simulink (The Mathworks, Natick, MA, USA) based xPC target ap-
plication.
Computing spike probability on-line
Let sn be an indicator function, so that sn = 1 if the neuron generated
a spike on the nth stimulus and sn = 0 otherwise. We define π(t) as
the probability of the neuron to emit a spike if it is stimulated at time
t. We can estimate this probability using all past responses {si}n
i=1 at
times {ti}n
i=1, by integrating them with an exponential kernel,
(cid:101)Pn = e− tn−t0
τ +
n(cid:88)
i=1
si(1 − e− ti−ti−1
τ
)e− tn−ti
τ
,
where τ is the kernel's time-constant (10-30 sec in the experiments
presented). To compute this on-line, we used the recursive formula:
(cid:101)Pn = sn(1 − e− tn−tn−1
τ
) + (cid:101)Pn−1e− tn−tn−1
τ
.
PID Controller
A Proportional-Integral-Derivative (PID) controller was realized on the
xPC target system. The input to the controller is the error signal,
en = R∗
n − Rn
where R∗
n and Rn are the desired and actual response to the nth stim-
ulus, respectively. The output of the controller is generally composed
of three components,
yn = gP en + gI
ei + gD(en − en−1)
n(cid:88)
i=0
6
where gP , gI and gD are the proportional, integral and derivative gains,
respectively (in the experiments presented in this work, gP was 25-
30Hz, gI was 0.125-0.56Hz and gD was 0-125Hz). Finally, the instan-
taneous stimulation frequency (i.e. the reciprocal of the ISI) consisted
of the controller's output and some 'baseline' frequency:
fn = yn + fbaseline,
where fbaseline was 2Hz in latency experiments and 6.67Hz in proba-
bility experiments. In our system, the maximal stimulation rate was
40 Hz. In the case of failed spikes when latency was controlled, the
controller maintained the same stimulation frequency that produced a
spike at a previous stimulus.
7
Results
The experiments described in this work were conducted in networks of
rat cortical neurons, plated in-vitro on a substrate that caters to ex-
tracellular recording and stimulation of electrical activity, at the single
cell level. Evoked spiking activity of a given individual neuron in such
a network, reflects both the direct effect of external stimulation, as
well as the effects of synaptic activations by other neurons (Marom &
Shahaf, 2002). We first examine the capacity of the Neuronal Response
Clamp technique to control the variability of neuronal responses of neu-
rons that are isolated from the network by pharmacologically blocking
all major types of synaptic inputs (APV [amino-5-phosphonovaleric
acid; 20 µM], CNQX [6-cyano-7-nitroquinoxaline-2,3-dione; 10 µM],
and Bicuculline [5 µM]).
Under complete blockade of synaptic transmission, all spontaneous
activity dies out and the network becomes quiescent (Marom & Shahaf,
2002). At this state, applying a single, short (sub-millisecond) pulse
between two electrodes, generates a single action potential in a subset
of neurons. The dynamics of the single neuron response may thus be in-
terrogated using a long sequence of stimuli. We refer to this stimulation
scenario as an open loop protocol, since the (predefined) stimulation
sequence is not affected by the evoked neuronal responses. In the open
loop protocol, the nature of the evoked responses leads to a distinction
between two qualitatively different stimulation regimes: When stim-
ulation rate is low (below 5-7 Hz), response probability is practically
one, i.e. the neuron responds almost to each and every stimulus, and
the time-delay (latency) between the stimulus and the evoked spike
is quite stable (Figure 1 middle panel). At higher stimulation rates,
both response probability and response latency gradually become un-
stable, exhibiting considerable, seemingly erratic fluctuations (Figure
1 bottom). Figure 2 depicts the dynamics of the two features (referred
to collectively as the neural response) at three stimulation rates (1, 5,
and 20 Hz) over extended timescales. Note the slow drift of response
8
latency in the 5 Hz stimulation experiment, probably reflecting slow
modulations of membrane excitability.
Thus, two features of neural activity, response latency and response
probability, emerge as natural candidates for stabilization using the Re-
sponse Clamp technique. These two features are markers of the neu-
ronal excitability status, reflecting the balance between exciting and
restoring membrane conductances. In order to clamp these two fea-
tures, continuous, online estimation of their values is required. Unlike
response latency, which can be measured individually for each of the
evoked action-potentials, response probability must be estimated based
on recent spiking history (i.e. using several responses). We calculate
this latter measure by convolving the spiking history with an expo-
nential kernel, which means that recent responses are given a greater
weight (explained in Materials and Methods).
The two response features mentioned above possess a useful qual-
ity, from the point of view of control: they exhibit a simple, monotonic
relationship with the instantaneous stimulation rate.
Increasing the
rate (i.e. by shortening the inter-stimulus-interval) leads to predictable
outcomes, namely a decrease in spike probability and an increase in
spike latency. Therefore, negative feedback may be used in order to
produce a desired response, either constant or time-varying. This is
achieved in the following manner (Figure 3): A response feature to be
clamped is chosen (either response probability or response latency).
Following each stimulus, a clamp error is computed by comparing the
evoked response with its desired value. This error signal is then fed
into a PID controller which accordingly alters the stimulation rate in
order to decrease the error towards zero. The PID controller consists of
three components: the proportional component (P) supplies the neces-
sary negative feedback; the integrated component (I) counteracts slow
changes in the system's responsiveness; and the derivative component
(D) contributes dissipation needed to damp oscillations. For details on
controller construction and implementation see Methods.
We first compared the neuronal response in open-loop, where stim-
9
ulation rate is maintained constant, to the neuronal response under
closed-loop conditions in synaptically isolated neurons for both re-
sponse features, over a wide range of stimulation rates and response
levels. In contrast to the open loop response, the response under closed-
loop conditions quickly converges to the desired value (Figure 4); the
clamp may be maintained for up to several hours (see below). Thus,
Response Clamp counteracts the slow trends observed in the spike re-
sponse latency, and restrains the huge fluctuations of the spike response
probability. The response clamp technique is reliable and robust: Fig-
ure 5 summarizes the results of 131 experimental blocks, showing that
the standard deviation in response probability is markedly decreased
under closed loop, compared to open loop stimulation regimes. The
robustness of the technique is reflected in the fact that no fine-tuning
of the PID gain parameters was required in order to achieve successful
clamp.
We were interested to determine whether it is the specific pattern
of stimulation, evoked by the controller is sufficient, in and by itself to
restrain response variability. We therefore played-back a stimulation
pattern produced by the controller under Response Clamp conditions,
to stimulate the same neuron again in an open-loop manner. Although
the "replayed" input sequences were identical, stability quickly deteri-
orated and the response probability diverged from the desired value
(Figure 6). Thus, feedback is essential in order to achieve response
stabilization: the exact temporal pattern of the stimulation series does
not, in and by itself, cause stable neuronal response; rather, the instan-
taneous state of the neuron must be continuously monitored and taken
into account while computing the control signal. Note that the result-
ing response rate (i.e. the product of stimulation rate and response
probability), while not constant in either the open loop or under Re-
sponse Clamp, is significantly more stable in the latter condition. Also,
note the duration of the experiment presented in Figure 6, extending
over 3 hours. The control signal in this long experiment, as well as
in similarly long experiments (results not shown), exhibits complex
10
dynamics that expose the myriad underlying adaptive processes.
In all the above mentioned experiments the response was clamped to
a constant value. However, the neuronal response can also be clamped
to a desired time-varying pattern. To demonstrate this capacity, we
performed an experiment where the desired response latency is grad-
ually ramped up and down to three different values (Figure 7). We
observed that in the time-varying clamp, the rate at which the de-
sired response may be changed is limited by the maximal and minimal
stimulation rates; if the controller surpasses these values, its output
saturates and the stimulation loop is effectively broken.
In all the experiments described so far, neurons were pharmacolog-
ically isolated from the network input. It was not at all obvious that
the ability of the response clamp to restrain response variability would
be conserved once ongoing input from the network is allowed. We thus
repeated the experiments described above without blocking synaptic
transmission, i.e. when the neuron is embedded in an active network.
We found that the Response Clamp technique was equally effective
under these conditions (Figure 8). Moreover, analyzing spikes simulta-
neously recorded from many different neurons in the network (beyond
the controlled neuron), revealed that the responses of these other neu-
rons fall into two broad categories: Some neurons were influenced by
the dynamic control over the target (controlled) neuron, whereas oth-
ers seem to be completely unaffected. These results suggest that the
Response Clamp technique may be used as a tool to study single neu-
ron dynamics even when isolation of the neuron from its context is
impossible or undesirable.
11
Discussion
In this study we show that evoked neuronal spiking patterns may be
controlled using a simple feedback system. The Response Clamp design
was applied to control either the time delay between stimulus and an
evoked spike, or the probability of evoked spikes. Control of these
response variables was shown to be applicable both when the target
neuron is synaptically isolated from the rest of the network, and on
the background of ongoing synaptic input.
Historically, the concept of feedback control proved effective in the
analyses of excitability. Its use, however, was largely focused on the
study of specific membrane conductances under voltage-clamp and
patch-clamp conditions (Hodgkin et al., 1952; Neher, Sakmann, &
Steinbach, 1978). The approach was further advanced by the dynamic-
clamp experimental design to unfold the impacts of specific conduc-
tances on the dynamics of the overall excitability (Sharp, O'Neil, Ab-
bott, & Marder, 1993). In that context, the methodology presented
here is a natural step up in the ladder of organization levels. It enables
control of neuronal response patterns at the macroscopic level, without
monitoring underlying microscopic variables.
One should bear in mind that in the original voltage-clamp stud-
ies of action potential generation the variable under control (namely,
membrane potential) determines the reaction rates of all the relevant
processes, so that the closed loop behavior becomes linear and time in-
variant. In contrast, in the dynamic clamp method only a few compo-
nents of the system are being controlled, with the hope of illuminating
the role of these components in the overall behavior. In this context
the method suggested here resembles the dynamic clamp in the sense
that only the processes that depend on the responsiveness of the sys-
tem are being clamped. Hopefully, this may aid in the identification of
these processes and the evaluation of their contribution to the response
dynamics.
Two technical difficulties in the design of the Response Clamp must
12
be noted. First, unlike the membrane potential which may be moni-
tored continuously, the neuron must be externally perturbed in order
to estimate its state (i.e. responsiveness to stimulation). Therefore,
the controller is always "one step behind" the system being controlled.
Since these perturbations themselves affect the neuron's responsive-
ness, a classic "observer effect" arises. This problem becomes more
pronounced in the case of neural response probability, where the state
is estimated using past spiking history. Future improvements of the
state estimation process might incorporate compensation for the effects
of external stimulation.
A second limitation in implementing the Response Clamp technique
relates to the response features to be controlled. The choice of response
latency and response probability stems from the need for a predictable
reaction to changes in stimulation rate. This does not necessarily re-
quire a monotonic relationship; a successful clamp may be realized
as long as the short term outcomes of modifying the stimulation are
predictable.
The Response Clamp technique offers the potential of facilitating
full characterization of the input-output relationships of the neuron.
This challenge is practically intractable in open loop due to the nonlin-
earity of the system and the cumulative effect of underlying processes
spanning a wide range of timescales. By manipulating well-defined
features of the neuronal responsiveness, one may hope to control state-
dependent dynamics, thus enabling the identification of the role of such
processes in the overall behavior of the system.
13
Acknowledgments
The authors thank Erez Braun, Naama Brenner, Danni Dagan, Steve
Goldstein, Effraim Wallach and Noam Ziv for their useful comments
and suggestions.
14
References
Adrian, E. D., & Zotterman, Y. (1926). The impulses produced by sensory nerve
endings: Part 3. impulses set up by touch and pressure. J Physiol, 61 (4),
465 -- 483.
Arieli, A., Sterkin, A., Grinvald, A., & Aertsen, A. (1996). Dynamics of ongoing
activity: explanation of the large variability in evoked cortical responses.
Science, 273 (5283), 1868.
Arsiero, M., Luscher, H., Lundstrom, B., & Giugliano, M. (2007). The impact of
input fluctuations on the frequency-current relationships of layer 5 pyramidal
neurons in the rat medial prefrontal cortex. Journal of Neuroscience, 27 (12),
3274.
Hodgkin, A., Huxley, A., & Katz, B.(1952). Measurement of current-voltage rela-
tions in the membrane of the giant axon of Loligo. The Journal of physiology,
116 (4), 424.
Levine, W., et al.(1996). The Control Handbook. CRC press Boca Raton, FL.
Mainen, Z., & Sejnowski, T. (1995). Reliability of spike timing in neocortical
neurons. Science, 268 (5216), 1503.
Marom, S. (2010). Neural timescales or lack thereof. Progress in Neurobiology,
90 (1), 16 - 28.
Marom, S., & Shahaf, G. (2002). Development, learning and memory in large
lessons beyond anatomy. Quarterly
random networks of cortical neurons:
Reviews of Biophysics, 35 (01), 63 -- 87.
Neher, E., Sakmann, B., & Steinbach, J.(1978). The extracellular patch clamp: a
method for resolving currents through individual open channels in biological
membranes. Pflugers Archiv European Journal of Physiology, 375 (2), 219 --
228.
Reich, D., Victor, J., Knight, B., Ozaki, T., & Kaplan, E. (1997). Response
variability and timing precision of neuronal spike trains in vivo. Journal of
neurophysiology, 77 (5), 2836.
Sharp, A., O'Neil, M., Abbott, L., & Marder, E. (1993). Dynamic clamp:
computer-generated conductances in real neurons. Journal of neurophysi-
ology, 69 (3), 992.
Soteropoulos, D., & Baker, S.(2009). Quantifying Neural Coding of Event Timing.
Journal of Neurophysiology, 101 (1), 402.
Stein, R.(1965). A theoretical analysis of neuronal variability. Biophysical Journal,
5 (2), 173 -- 194.
15
Figures
16
Figure 1: Evoked responses of an isolated neuron in-vitro to periodic stimulation in
open loop. The top panel (a) shows a single extracellular voltage trace recorded for
20 ms, beginning at the onset of stimulation. The stimulus artifact lasts for 1-2 ms
(shaded area). Response latency is the time elapsed from the onset of stimulation
to the detected peak of the action potential; in our system, response latency is in
the range of 3-10 ms. (b) Traces (every 10th, shown top to bottom) in response
to a low (1 Hz) stimulation rate. Spiking under this condition is highly regular
and the latency is constant. (c) Traces (every 10th, shown top to bottom) of the
same neuron when stimulation rate is high (20 Hz). Both response probability and
latency rapidly become irregular under this condition.
17
Figure 2: Response latency (a) and response probability (b) for three open loop
stimulation rates. At extremely low stimulation rate (1 Hz, depicted yellow),
response is highly reliable (i.e. response probability is one) and response latency is
practically constant. At higher stimulation rate (5 Hz, depicted purple), response
probability is also one, but slow trends in the response latency appear, probably
reflecting slow processes of inactivation. At high stimulation rate (20 Hz, depicted
blue) both response probability and response latency are highly unstable. The
response probability was computed using 2 s bins.
18
Figure 3: General scheme of a PID controller, designed for clamping neuronal re-
sponses (stimulus-spike time delay, or response probability). The error signal is
calculated (feedback), and subjected to three different transformations that addi-
tively dictate the nature of stimulation needed for clamping the response. This is
standard control algorithm was implemented within a Simulink environment.
19
Figure 4: Demonstration of the neuronal Response Clamp. In each of the panels,
two experiments are shown (depicted by two different colors). Top row shows
examples of controlling response probability (a) and stimulus-spike time delay
(b). Two clamped values are demonstrated in each of these panels. Note the
stability of the clamped response. In contrast, when the average stimulation rates
used for the clamping procedure of the two top panels are applied under open loop
conditions (c and d), the response develops marked fluctuations. This is especially
apparent when response probability (c) is considered.
20
Figure 5: The Response Clamp consistently decreases the variability of response
probability. Neuronal response probability was clamped to different values ranging
from 1 to 0.25, for 10 to 15 minutes. The response's standard deviation during
this period is compared with that of the same neuron, when the mean rate used
during the clamp period is re-applied in an open loop design (see the examples
depicted in Figure 4). Data from 131 such experimental sessions conducted on
3 different neurons (each in a different culture, depicted by different colors) are
presented. The data are normalized to the maximal standard deviation acquired
for each neuron. Histograms of the data are aligned to the relevant axes.
21
Figure 6: Demonstration of the importance of feedback in the stabilization of
neuronal response probability. Stimulation rate (a), response probability (b) and
response rate (c) at two stimulation scenarios are depicted. First, the neuron's
response probability was clamped to a constant value (0.67) for three hours (blue).
Then, the stimulation pattern generated by the controller was replayed in open
loop (purple). Stability of both response probability and response rate is obtained
only in the presence of feedback.
22
Figure 7: Demonstration of time varying Response Clamp. Response latency
was clamped, while the desired value (depicted blue) was gradually increased and
decreased alternatingly. Latency is normalized so that the baseline level (i.e. the
latency at an extremely low stimulation rate, in the example presented here it is
6 ms) is 0 and maximal latency detected (10 ms in this example) is set to 1. Note
that at extremly high latency values the behavior becomes irregular and the clamp
is effectively lost.
23
Figure 8: Time varying Response Clamp of a neuron with background synaptic
input.
(a) A neuron that responds (blue) directly to external stimulation was
controlled to follow a sine-wave response probability pattern (12 minutes period,
range between 0.75 and 0.25, depicted light green). The figure shows the response
probability of another, directly stimulated yet uncontrolled neuron (purple). The
response of such neurons, while not follwing the desired response closely like the
controlled neuron does, is modulated by the clamp. The figure also presents the
response probability of yet another neuron, which is not affected directly by the
stimulation, but is activated by the synaptic inputs it recieves from the network
(yellow). The activity of such neurons seems to be unaffected by the clamp. This
is also apparent by looking at the spike rate histogram of 20 other neurons in
the network (b, 10 s bins), demonstrating the presence of uninterrupted on-going
background activity in the network.
24
|
1803.01236 | 1 | 1803 | 2018-03-03T20:45:33 | Vocal effort modulates the motor planning of short speech structures | [
"q-bio.NC",
"physics.bio-ph"
] | Speech requires programming the sequence of vocal gestures that produce the sounds of words. Here we explored the timing of this program by asking our participants to pronounce, as quickly as possible, a sequence of consonant-consonant-vowel (CCV) structures appearing on screen. We measured the delay between visual presentation and voice onset. In the case of plosive consonants, produced by sharp and well defined movements of the vocal tract, we found that delays are positively correlated with the duration of the transition between consonants. We then used a battery of statistical tests and mathematical vocal models to show that delays reflect the motor planning of CCVs and transitions are proxy indicators of the vocal effort needed to produce them. These results support that the effort required to produce the sequence of movements of a vocal gesture modulates the onset of the motor plan. | q-bio.NC | q-bio | Vocal effort modulates the motor planning of short speech structures
Alan Taitz1, Diego E. Shalom2 and Marcos A. Trevisan1,2
1 Physics Institute of Buenos Aires (IFIBA) CONICET, Buenos Aires, Argentina,
2 Department of Physics, Universidad de Buenos Aires, Buenos Aires 1428EGA, Argentina.
Corresponding author:
[email protected] (MAT)
March 2, 2018
Speech requires programming the sequence of vocal gestures that produce the sounds of words. Here
we explored the timing of this program by asking our participants to pronounce, as quickly as
possible, a sequence of consonant-consonant-vowel (CCV) structures appearing on screen. We
measured the delay between visual presentation and voice onset. In the case of plosive consonants,
produced by sharp and well defined movements of the vocal tract, we found that delays are positively
correlated with the duration of the transition between consonants. We then used a battery of statistical
tests and mathematical vocal models to show that delays reflect the motor planning of CCVs and
transitions are proxy indicators of the vocal effort needed to produce them. These results support that
the effort required to produce the sequence of movements of a vocal gesture modulates the onset of the
motor plan.
I. INTRODUCTION
Voluntary movements need preparation before execution [1]. This preparatory phase received
extensive attention from the scientific community, whose investigations across species [2,3] helped
unveiling the instructions to the effectors before tasks as locomotion [4], arm reaching [5] and vocal
production [6,7]. In humans, the readiness potential (RP) is considered a universal signature of
planning of volitional acts that has been largely investigated with electroencephalographic
techniques [8–10].
Speech requires programming the instructions to drive two acoustically uncoupled effectors: the
vocal folds and vocal tract [11]. The folds are a pair of opposed membranes located at the larynx,
which leads to a series of configurable cavities ending at the mouth, known as vocal tract. When the
larynx is adducted and lung pressure reaches a threshold, the folds oscillate colliding with each
other [12]. The resulting periodic interruptions of airflow produce a pressure wave that propagates
along the vocal tract. When the tract is open, the sound radiated from the mouth corresponds to a
vowel [13]. Consonants are produced in different manners: fricatives are vortex sounds produced by
the turbulent passage of air through a constriction in the vocal tract [14]. When lips are narrowed,
an [f] is produced; when the tip of the tongue approaches the upper teeth, we hear [s]. Plosives are
generated by completely occluding the vocal tract at a given point [15]. When lips are released, we
hear [b]; releasing the tip or the body of the tongue produces a [d] or a [g], respectively. From a
mechanical point of view, these families of sounds represent vocal gestures of different effort:
vowels are produced by smooth and opened tract configurations, fricatives by narrow constrictions
and plosives by sharp closures.
Recent advances give a cause for great optimism in describing vocal motor control in the language
of dynamical systems [5]. As a matter of fact, the avian vocal system presents strong analogies with
the human case [16], and birdsong has been fully described in dynamical terms [7,17], leading to
devices that generate song controlled by a few physiological variables [18].
What are the dynamical variables driving the human vocal system? The production of speech
requires coordination of the larynx, lips, tongue and jaw to produce a gesture, and then control of
'the strength with which the associated gesture (e.g., lip closure) "attempts" to shape vocal tract
movements at a given point in time' [19]. These two variables, the gesture and its strength, have
been interpreted in dynamical terms [19]. In this theoretical description, the vocal program is
executed by coupled neural oscillators; their relative phases represent the sequence of movements
of the articulators that define a gesture, and their amplitude represents its strength.
Here we hypothesized that the vocal program is sensitive to the strength of the vocal gestures, and
that this dependence is revealed in the timing between the instruction and the vocal onset. To
explore this hypothesis, we created and analyzed a database of speech structures produced under a
simple reaction-time paradigm.
II. MATERIALS AND METHODS
A. Participants
Forty five native Spanish speakers (23 females, age range 20-28, mean age 22), undergraduate and
graduate students at the University of Buenos Aires, with normal or corrected-to-normal vision and
no speech impairments, completed the experiment. All the participants signed a written consent
form. All the experiments described in this paper were approved by the ethics committee Comité de
Ética del Centro de Educación Médica e Investigaciones Clínicas 'Norberto Quirno' (CEMIC)
qualified by the Department of Health and Human Services (HHS, USA): IRb00001745-IORG
0001315.
B. Stimuli and tasks
Participants sat in a silent room, 0.7 m away from a 19' monitor. They were asked to pronounce a
sequence of consonant-consonant-vowel (CCV) structures, as soon as they appeared at the center of
the screen. The CCVs were recorded with a commercial microphone at approximately 0.2 m of the
mouth.
After pronouncing a CCV, the participant was instructed to press a key. A blank screen preceded the
following CCV, which appeared after a random time between 0.5 and 2 s. The procedure was
repeated until completing 3 rounds of the complete CCV set. All the presentations were
randomized.
The experiment was coded in Python, using Pyaudio and PsychoPy [20] libraries to ensure robust
audio-video synchronization (30 ms maximum offset between visual presentation and audio
recording).
CCVs were formed by combinations of fricatives and plosives, followed by the vowel [a]. We used
the most common Spanish fricatives [21,22] [f, s, x], pronounced as in face, stand and loch, and the
most common Spanish plosives [b, p, d, t, g, k] [21], pronounced as in bay, pay, die, tie, gray and
cray. Plosives come in voiced-unvoiced pairs [b, p], [d, t] and [g, k]. The tract anatomy is identical
for each pair, but folds are active only for the first components. Since plosives are recognized
through the sound produced at the occlusion release, unvoiced ones leave no acoustical traces
during occlusion. On the contrary, the folds' fundamental frequency serves to spectrally mark vocal
onset during occlusion in voiced plosives. For this reason, CCVs starting with plosives were
reduced to the voiced ones [b, d, g].
We presented the 6 possible combinations of fricatives (fja, fsa, jfa, jsa, sfa, sja), the 9 possible
plosive-fricative combinations (bfa, bja, bsa, dfa, dja, dsa, gfa, gja, gsa), a selection of 11 fricative-
plosive combinations (fga, fba, fpa, jba, jca, jga, jda, sba, sca, sga, sta) and 11 combinations of
plosives (bca, bda, bga, bta, dba, dca, dga, dta, gba, gda, gta). This set of 37 CCVs allowed
keeping a short (mean 20 minutes), single-session experiment per participant.
The number of recorded samples was 4995 = 45 subjects × 3 trials × (6 fricative-fricative + 9
plosive-fricative +11 fricative-plosive + 11 plosive-plosive) combinations.
C. Data selection
Stage 1.
Participants were instructed to pronounce the CCVs clearly, as soon as possible after visual
presentation, and they were not allowed to correct their utterances. These requirements produced a
considerable amount of errors during the experiment. We discarded audio files that a. contained
more than one attempt to produce the CCV and b. did not match with the presented CCV. This left
us with 3296 CCVs.
Stage 2.
For each audio file that passed stage 1, we measured three timing quantities: the delay Δ between
visual presentation and voice onset, the transition between consonants and duration T of the CCV.
Audio files were inspected with Praat [23].
Delays Δ were measured directly from the spectrogram. Transitions are defined as the time during
which the vocal tract changes its configuration from the first to the second consonant. Since vocal
gestures producing the same speech content are not necessarily unique [24], we extracted by
inspection the most robust features of each CCV across trials and participants, and therefore only
selected the samples that strictly met the following spectral signatures for consonantal transitions:
a. Fricative-fricative combinations. As noisy turbulent sounds, the spectral signature of a fricative
is a dark spot, such as the ones shown in the left panel of Fig. 1A for [f] and [s]. Transitions are
defined by the continuous passage from the first to the second constriction, therefore presenting
a mixture of spectral components from both consonants.
b. Plosive-fricative combinations. Transitions are characterized by the interval that goes from
releasing the occlusion of the plosive to the formation of a constriction to generate the fricative.
Spectrally, the transition is described by the voiced sound structure produced after the plosive,
and before the formation of the purely noisy fricative spot (Fig. 1B, left).
c. Fricative-plosive combinations. During these transitions, the vocal tract evolves from a
constriction at one point to an occlusion at another one. Transitions are spectrally defined as the
interval between the abrupt end of the fricative and the release of the plosive into the vowel [a],
as shown in the left panel of Fig. 1C.
d. Plosive-plosive combinations. Vocal tract passes from an occlusion at a given point of the tract
to another one at a different location. Transitions, as shown in the representative case of the left
panel of Fig. 1D, go from the release of the first plosive characterized by a voiced sound
structure, to the release of the second one within the vowel [a].
A pool of 1958 CCVs was successfully classified into these spectral categories. Four subjects were
discarded for presenting systematical spectral differences from the described categories.
Stage 3.
We finally discarded data that passed stage 2 for which delays Δ and transitions '=/T were greater
than two standard deviations from the speaker mean values. A final timing dataset from 1519 CCVs
from N=41 participants was used to perform the analyses presented in this work.
D. CCV frequency
None of the CCVs are Spanish words. We counted the intra-word appearances of each CCV in a
large Spanish corpus [25], and assigned the appearances per million words as the frequency of each
CCV. Since the range of frequency values was large (from 0 to roughly 105 apmw), in Fig. 2 we
show a discretization of this range in low, medium or high frequency values.
E. Vocal model
The main ingredients of the folds' dynamics are caught by the following equation of motion, known
as the flapping model [26]:
m z=−k (z) z−b( z, z) z+ ps
δ+2Γ z
z0+δ /2+z+Γ z
(1)
In Fig. 3A we sketch its elements. The variable z represents the transverse displacement of the folds
from the pre-phonatory position z0. The tissue is described by nonlinear elastic k and dissipative b
functions. The last term represents the energy transferred to the folds by lung pressure ps, which
depends on the parameters δ and Γ that define the folds' configuration. Dynamical analysis of Eq. 1
revealed that pressure ps beyond a threshold creates oscillations through a Hopf bifurcation [12] and
that the threshold increases with glottal abduction z0 [27]. Oscillations make the folds collide,
interrupting the airflow at the vocal tract entrance and producing pressure waves p that propagate
along a vocal tract. The tract cross-section A(x,t) is shown in Fig. 3B, where x represents the
distance from the glottis (x=0) to the mouth (x=L). During a general sequence of vowels and
plosive consonants, A(x,t) was described by Story [28]:
A (x ,t)=π/4 [Ω(x)+q1(t)ϕ1(x)+q2(t)ϕ2( x)]2∏
n
k=1
[1−ck(x)m(t−t k)]
(2)
The first factor represents the vowel substrate, with empirical functions Ω, φ1 and φ2 obtained from
an orthogonal decomposition calculated from MRI data [29]. Any sequence of vowels is generated
by the evolution of q1 and q2. Plosive consonants are generated through the bell-shaped functions ck
and m. These functions reach the value 1 around x=xk and t=0 respectively, producing a specific
occlusion A(xk,tk)=0. The spatial functions ck represent the anatomy of the occlusion for each
consonant, and the time function m represents its activation-deactivation.
F. CCV synthesis
Eqs. 1 and 2 were solved with MATLAB. Numerical integration of Eq. 2 was performed through a
standard Runge-Kutta algorithm at a sampling rate of 44.1 kHz. A wave-reflection
model [30,31] was used to simulate propagation of sound along the vocal tract, which is
approximated by a series of N=44 tubes of cross-section Ai, 1≤i≤N. Sound propagation is solved by
splitting an incoming sound wave pi into a reflected and a transmitted wave at each interface, with
reflection and transmission coefficients ri,j=(Ai-Aj)/(Ai+Aj) and ti,j=1-ri,j (j=i±1) [32]. The sound
radiated from the mouth is proportional to the pressure at the last tube, p44. The resulting time series
was converted to wav audio format. To synthesize CCVs for voiced plosives, we proceeded as
follows:
1. Subglottal pressure ps was set at a normal speech value such that Eq. 1 generated oscillations of
the folds with a pitch around 100 Hz, from t=0 to voice offset at t=Tf. The rest of the parameters
can be found elsewhere [11].
2. Linear functions q1(t)=4t/T and q2(t)=t/T [26] were used in Eq. 2 to drive the tract from a neutral
tract (q1,q2)=(0,0) at t=0 to the shape of the vowel [a] (q1,q2)=(4,1) at t=T, and maintained there
until t=Tf.
3. While the tract evolves towards the vowel [a], two occlusions take place at t1 and t2, activated
through Gaussian functions m of width 0.14 s [28], separated by τ=t2-t1. We used the mean value
of τ across CCVs and participants. The spatial functions ck reach the value 1 at xkL for [b],
xk9/10L for [d] and xk7/10L for [g]. Functional forms and parameter values for the functions
ck and m are reported elsewhere [28,33].
III. RESULTS
We investigated the vocal planning prior to the utterance of short speech structures. For that sake,
we recorded 45 speakers while pronouncing a set of consonant-consonant-vowel structures (CCV)
as soon as they appeared on a computer screen. We used combinations of fricatives [f, s, x] and
plosives [b, p, d, t, g, k] followed by the vowel [a]. Examples of these structures are fsa, gfa, fga
and dca, which are also representatives of fricative-fricative, plosive-fricative, fricative-plosive and
plosive-plosive combinations respectively.
For each audio file, we measured three timing variables: the delay between visual presentation and
phonation onset (Δ), the duration of the CCV (T) and the duration of the transition between the
consonants (τ). In the left panels of Fig. 1 we show these timing variables for representative cases of
the fricative-fricative (A), fricative-plosive (B), plosive-fricative (C) and plosive-plosive (D) types.
After selecting the audio files that met the spectral criteria for each family type (see section II C),
we collected timing data from 1519 CCVs produced by 41 subjects.
How are these timing variables related? To compare dynamics across participants with different
speech rates, we normalized the transitions to the duration T of the CCV, τ'=τ/T. In the right panels
of Fig. 1 we plot Δ vs. τ' for the different families of consonantal combinations. No correlations
were found between delays and transitions for fricatives (Fig. 1A, r=-0.27, p=0.6) or combinations
of fricatives and plosives (Fig. 1B, r=-0.29 p=0.38 and Fig. 1C, r=0.15, p=0.7). Interestingly,
however, a positive correlation emerged for combinations of plosive consonants (Fig. 1D, r=0.68,
p<0.02). We next explored how these timing variables relate to relevant motor and planning
variables.
FIG. 1. Timing variables for the different consonantal transitions.
Left panels show onset delays Δ and consonantal transitions τ measured from sound and
spectrogram, for representative cases of the fricative-fricative (A), fricative-plosive (B) plosive-
fricative (C) and plosive-plosive (D) types. In the right panels we show that vocal onset delays Δ are
not significantly correlated with the normalized transitions τ'=τ/T for the 6 fricative-fricative (A),
11 fricative-plosive (B) and 9 plosive-fricative (C) combinations. On the contrary, timing variables
are positively correlated for the 11 plosive-plosive combinations (D).
A. Delays and motor planning
Vocal onset delays represent the integration of reading, memory retrieval and vocal planning tasks,
followed by articulatory movements produced before phonation. We first analyzed the effects of
each of these contributions.
Reading. Reading task involve visualization and recognition of three common Spanish phonemes of
identical visual size. Given the short length of the strings, we assumed that this effect is constant
across CCVs.
Memory retrieval. We explored two memory variables. Long term effects were investigated by
computing the frequency of occurrence of the CCVs in Spanish, shown in Fig. 2 (see details in
section IV D). Since our participants completed three rounds of the experiment, trial number was
used to investigate short term memory effects. We submitted timing data to ANOVA with frequency
and trial as independent variables (Table I). The analysis revealed effects of frequency on Δ and τ'
for all the consonant combinations, except for the case of plosive-plosive type. This merely reflects
that these are very unusual combinations in Spanish, and therefore no long term memory effects
were detected. Trial number, on the contrary, has a significant effect on Δ and no effect on τ', for
any consonantal combination. A follow up of this analysis revealed that this effect on Δ was
accounted for by an increase in Δ for trial 1 (mean 0.94±0.01 s) than for trials 2 and 3 (mean
0.85±0.01 s), (t=8.13, df=1176, p<10-14). We therefore considered trial 1 as a training set and used
timing data from trials 2 and 3 for the following analyses.
Articulatory kinematics. Since Δ was measured from visual presentation to vocal onset, articulatory
kinematics that normally precede phonation [34,35] are included within this time period (e.g. lip
closure before folds' activation for an initial [b]). To study the effect of movements prior to sound
onset, submitted our data to ANOVA using the first consonant of the CCV as independent variable,
restricting the study to initial plosives [b], [d] and [g]. The production of these consonants before
phonation require articulation of the lips, tongue tip and tongue body movements respectively. This
analysis revealed no significant effects of initial plosives on Δ (p=0.323, F=1.13, df=2), which
allowed us to assume that the period of articulation previous to phonation is constant across CCVs.
Due to the nature of the experiments performed here, vocal planning cannot be explored directly.
Instead, we analyzed the different tasks performed by the participants from CCV visualization to
phonation. We found that, for plosive-plosive combinations, the tasks of reading, memory and
prephonatory movements can be assumed constant across CCVs. We therefore assumed that
variations of Δ across CCVs account for differences in the vocal planning.
FIG. 2. Frequency of occurrence of consonants in CCVs.
Rows and columns correspond to the first and the second CCV consonants respectively. Colors
represent low (blue), medium (light blue) and high (yellow) frequency of occurence values of the
these combinations in Spanish. Occurrences of fricative-fricative and fricative-plosive combinations
present high variability, while only one plosive-plosive combination presents a high occurrence
level.
B. Consonantal transitions and vocal effort
Under the simplistic hypothesis that effortless gestures are produced faster than harder ones, we
could roughly assume that transitions τ' reflect the effort needed to produce a CCV.
We first note that, consistently with this hypothesis, the right panels Fig. 1 show that the shorter
transitions correspond to fricatives (τ'=0.124±0.002), which require relatively low efforts to
produce constrictions in the tract; longer transitions were found in the production of plosive-
fricative (τ'=0.163±0.003) or fricative-plosive combinations ('=0.280±0.003), and these are in turn
shorter than transitions between plosives ('=0.389±0.003), which require the largest efforts to
completely occlude the vocal tract twice.
To further formalize this notion of effort, we propose a physically-based model that represents the
effort E exerted by the tract and the folds during speech. The expression reads:
E=E tract+ Efolds=∫
∫
A (x ,t )− AΩ(x)dxdt +E 0
(3)
T
L
0
0
The shape of the vocal tract is mathematically described by its cross-section A(x,t), with x the
distance from the glottis (x=0) to the mouth (x=L), as shown in the sketch of Fig. 3B. Building on a
previous mass-spring model [27], the first term of Eq. 3 represents the elastic force exerted by the
departures of the tract from a relaxed configuration AΩ(x) along an utterance of duration T.
During a CCV of plosive consonants, two occlusions separated by occur while the tract is
evolving towards the vowel. This is described in Eq. 2 (section II E), where plosives are spatially
described by functions ck(x) that model the extension and location of the occlusion, and temporally
described by functions m(t) that control the activations of closures. Using the parameters reported
for consonantal activations [19,28], efforts computed using each CCV transition (in a range that
goes from DGA=0.33±0.03 s to GTA=0.45±0.04 s, shown in Fig. 3C) display the same organization
as the efforts computed for all the CCVs using the population mean (=0.40±0.04 s) for all CCVs.
To build confidence on the pertinence of these temporal parameters, we used Eqs. 1 and 2 to
generate synthetic CCVs with the activations reported (see section II F) and fixed mean value,
obtaining audio files with the same spectral features and speech content than the experimental
records (Supplementary Materials).
Tract effort does not explicitly depend on , but instead on the specific consonants forming the CCV
and their order of appearance. The dependence on the particular CCV consonants stem from the
anatomical differences between them. For instance, the function ck(x) that models the extension of
occlusion is wider for [d] than for [b], accounting for the greater extent where the constriction area
is zero when it is performed with the tongue than with the lips; a greater effort is then needed to
produce a [d] than a [b].
The dependence of the effort with the order of appearance of the consonants is due to the fact that
consonants occur while the vocal tract is evolving towards the vowel [a]. For instance, bda requires
less effort than dba. This asymmetry is explained because in dba, the lips closure for [b] occurs
closer to the [a], for which the mouth is fully opened, requiring a larger vocal tract deformation that
corresponds to a larger effort. Transitions ' systematically reflect this asymmetry: 'DGA<'GDA (t=-
2.9, df=49.8, p<0.01); 'BGA<'GBA (t=-2.13, df=86.9, p<0.05) and 'BDA<'DBA (t=-0.71, df=136.5,
p=0.48).
The second term in Eq. 3 refers to laryngeal efforts. Vibration of the vocal folds has an on/off
function in speech. Within our set of CCVs, vibration is either active during the whole vocalization
(for the subset bda, dba, bga, gba, dga, gda) or inactivated during the second (unvoiced) plosive
(subset bca, bta, dca, gta, dta). This second group involves a devoicing effort by glottal abduction.
We model Efolds in Eq. 3 as a constant that takes the value E0=0 for the first group and E0>0 for the
second one. The distribution of CCVs along E varies with E0. For instance, a sufficiently large value
of E0 separates one group from the other, eventually producing a bimodal distribution. We set the
value of E0 that matched the experimental distribution of CCVs along ' (vertical axis of Fig. 3C) to
the theoretical distribution along E (horizontal axis of Fig. 3C). We obtained values of E0 between
7% and 11% of the mean vocal tract effort (for distributions between 4-8 bins).
The effort model of Eq. 3 represents a compromise between mathematical simplicity and vocal
biomechanics, accounting for both tract reconfigurations and control of folds, which are the main
gestures necessary to generate speech. Although effort does not depend on ' within the range of
measured values, Fig. 3C shows that both variables are significantly correlated (R=0.84, p<0.001),
reflecting that consonantal transitions are indeed shorter when the global gesture requires less effort.
This implies that ' is a good proxy for the effort needed to produce the CCVs.
FIG. 3. Elements of the articulatory model used to calculate vocal effort.
A. Vocal tract movements, described by its cross-section A(x,t), require an effort proportional to its
deformations with respect to a relaxed configuration. B. Elements of the vocal folds model of Eq. 1.
Abduction effort is needed to separate the folds, increasing z0 in Eq. 1 to deactivate oscillations
during unvoiced plosives. This is modeled as a constant E0 for the group (bca, bta, dca, gta, dta)
shown in green. C. The transitions measured on the recorded CCVs are positively correlated with
vocal effort E computed with Eq. 3. The distributions of CCVs with E and are shown along the
horizontal and vertical axis respectively, for E0=9% of the mean tract effort.
Taken together, these results show that the timing variables measured directly from speech
recordings are related to planning and motor variables. For combinations of plosive consonants, we
found that differences in Δ account for differences in the motor planning and that ' are proportional
to the effort needed to produce the global CCV gesture.
IV. CONCLUSIONS AND DISCUSSION
Here we hypothesized that the vocal program is sensitive to the strength of the vocal gestures, and
that this dependence is revealed in the time required to activate vocal production. To explore this,
we created and analyzed a database of speech structures produced under a simple reaction-time
paradigm. Our participants were instructed to pronounce a set of consonant-consonant-vowel
structures (CCV) as soon as they appeared on screen. We measured two timing quantities from the
recorded audio files: the delay between visual presentation and vocal onset, and the duration of the
transition between consonants, normalized to the duration of the CCV.
Using a battery of tools that includes statistical tests and nonlinear modeling of the vocal system,
we found that 1. these variables are positively correlated for CCVs formed with plosive consonants.
Then, we show the plausibility that 1. variations in delays account for variations in vocal planning
and 2. consonantal transitions reflect the vocal effort required to produce the CCVs.
Overall, these results support the idea that the time required to plan short speech structures is
proportional to the effort needed to produce them. The fact that no correlations were found for the
rest of the consonantal combinations can be due to multiple factors. In particular, the memory
effects detected on these combinations suggests a probable interaction between motor planning and
memory tasks.
Our findings are indeed compatible with electroencephalographic measurements of the readiness
potential (RP), defined as the buildup of electrical potential beginning about one second before cued
or self-paced movement onset. This universal signature of planning and initiation of volitional
acts [8] has been largely explored, and previous works reported that the level of force and speed of
a motor task can be detected from these cortical potentials [10]. Moreover, differences in RP onset
time account not only for initial movements but also for a sequence of movements [36].
Evidence is accumulating towards the vision of motor cortex as a dynamical system that generates
and control movements [5]. The particular case of the vocal motor program has been extensively
studied and successfully described in dynamical terms for songbirds [7]. Although strong analogies
connect the avian and human vocal systems, descriptions of the human motor program in terms of
dynamical systems received considerably less attention. One exception is the approach proposed by
Goldstein and collaborators [19], whose theoretical model represents the vocal motor program as a
set of coupled neural oscillators controlling the vocal articulators. The relative phases of the
oscillators and their amplitudes configure a vocal gesture and its strength.
The model of vocal effort presented here presents a plausible quantification of this notion of vocal
strength, supporting the hypothesis that the timing of the motor program is sensitive to the strength
of the vocal gestures.
It is well established that brain motor areas activate not only for voice production, but also to assist
other non-motor tasks as speech perception [37]. More interestingly perhaps, these motor maps are
also used in the case of purely mental operations such the inner speech, i.e. the voice that "narrates
the words we read or the conversations we imagine" [38]. In fact, as happens with speech itself,
inner speech is associated with an efference copy with detailed auditory properties [38]. In light of
these results, we plan to capitalize on the simple experimental paradigm presented here to explore
vocal motor signatures during silent lecture.
Acknowledgements
The research reported in this work was partially funded by the Concejo Nacional de Investigaciones
Científicas y Técnicas, CONICET; The University of Buenos Aires UBA and NIH through R01-
DC-012859.
References
[1] M. S. Graziano, Neuron 71, 387 (2011).
[2] W. A. Friedman, L. M. Jones, N. P. Cramer, E. E. Kwegyir-Afful, H. P. Zeigler, and A. Keller, J
[3]
[4]
Neurophysiol 95, 1274 (2006).
A. Riehle and J. Requin, J. Neurophysiol. 6, (1989).
G. S. Stent, W. B. Kristan, W. O. Friesen, C. a Ort, M. Poon, and R. L. Calabrese, Science 200, 1348
(1978).
[5] M. M. Churchland, J. P. Cunningham, M. T. Kaufman, J. D. Foster, P. Nuyujukian, S. I. Ryu, and K.
V Shenoy, Nature 487, 51 (2012).
[6]
R. G. Alonso, M. a. Trevisan, A. Amador, F. Goller, and G. B. Mindlin, Front. Comput. Neurosci. 9, 1
(2015).
[7]
A. Amador, Y. S. Perl, G. B. Mindlin, and D. Margoliash, Nature 495, 59 (2013).
[8]
[9]
A. Schurger, J. D. Sitt, and S. Dehaene, Proc. Natl. Acad. Sci. 109, E2904 (2012).
V. Siemionow, G. H. Yue, V. K. Ranganathan, J. Z. Liu, and V. Sahgal, Exp. Brain Res. 133, 303
(2000).
[10] M. Jochumsen, I. K. Niazi, N. Mrachacz-Kersting, D. Farina, and K. Dremstrup, J. Neural Eng. 10,
(2013).
[11] R. Laje, T. Gardner, and G. B. Mindlin, Phys. Rev. E - Stat. Nonlinear, Soft Matter Phys. 64,
056201/1 (2001).
[12]
[13]
J. C. Lucero and L. L. Koenig, J. Acoust. Soc. Am. 117, 1362 (2005).
I. R. Titze, J. Acoust. Soc. Am. 83, 1536 (1988).
[14] M. H. Krane, J. Acoust. Soc. Am. 118, 410 (2005).
[15] B. H. Story and K. Bunton, J. Speech. Lang. Hear. Res. 53, 1514 (2010).
[16] H. P. Zeigler and P. Marler, Neuroscience of Birdsong (Cambridge Univ Pr, 2008).
[17] M. Trevisan, G. Mindlin, and F. Goller, Phys. Rev. Lett. 96, 1 (2006).
[18]
[19]
[20]
[21]
E. M. Arneodo, Y. S. Perl, F. Goller, and G. B. Mindlin, PLoS Comput. Biol. 8, e1002546 (2012).
L. Goldstein, D. Byrd, and E. Saltzman, in Action to Laguage via Mirror Neuron Syst. (Cambridge
University Press, 2006), pp. 215–249.
J. W. Peirce, J. Neurosci. Methods 162, 8 (2007).
J. I. Hualde, A. Olarrea, and E. O'Rourke, The Handbook of Hispanic Linguistics (2012).
[22] A. M. Borzone de Manrique and M. I. Massone, J. Acoust. Soc. Am. 69, 1145 (1981).
[23]
P. Boersma and D. Weenink, (2013).
[24] M. F. Assaneo, M. A. Trevisan, and G. B. Mindlin, PLoS One 8, e80373 (2013).
[25]
F. Cuetos, M. Glez-Nosti, A. Barbón, and M. Brysbaert, Psicológica 32, 133 (2011).
[26] M. F. Assaneo, J. I. Nichols, and M. A. Trevisan, PLoS One 6, e28317 (2011).
[27] R. Kirchner, An Effort Based Approach to Consonant Lenition (Routledge, 2013).
[28] B. H. Story, J. Acoust. Soc. Am. 117, 3231 (2005).
[29] B. H. Story and I. R. Titze, J. Phon. 26, 223 (1998).
[30] M. M. Sondhi, J. Acoust. Soc. Am. 55, 1070 (1974).
[31]
J. Flanagan, J. Allen, and M. Hasegawa-Johnson, Speech Analysis, Synthesis and Perception (2008).
[32] B. H. Story, Physiologically-Based Speech Simulation Using an Enhanced Wave-Reflection Model of
the Vocal Tract, University of Iowa, 1995.
[33] M. F. Assaneo, J. Sitt, G. Varoquaux, M. Sigman, L. Cohen, and M. A. Trevisan, Neuroimage 141, 31
(2016).
[34] K. E. Bouchard, N. Mesgarani, K. Johnson, and E. F. Chang, Nature 495, 327 (2013).
[35] D. F. Conant, K. E. Bouchard, M. K. Leonard, and E. F. Chang, J. Neurosci. 2382 (2018).
[36] M. Simonetta, M. Clanet, and O. Rascol, Electroencephalogr. Clin. Neurophysiol. Evoked Potentials
81, 129 (1991).
[37]
[38]
J. M. Correia, B. M. B. Jansma, and M. Bonte, J. Neurosci. 35, 15015 (2015).
T. J. Whitford, B. N. Jack, D. Pearson, O. Griffiths, D. Luque, A. W. F. Harris, K. M. Spencer, and
M. E. Le Pelley, Elife 1 (2017).
Tables
TABLE I. ANOVA analysis with frequency of appearance of the CCVs in Spanish and trial number
as independent factors. We also performed the tests using the absolute frequency values, with
equivalent results.
Frequency (low, medium, high)
Trial number
p<10-5, F=9.54, df=2
p=0.004, F=8.37, df=2
p<10-5, F=42.6, df=2
p=0.4, F=0.89, df=2
fricative-fricative
fricative-plosive
p=6×10-4, F=7.51, df=2
p=0.4, F=0.93, df=2
plosive-fricative
p=0.002, F=9.93, df=1
p=0.12, F=2.1, df=2
Δ
'
plosive-plosive
p=0.62, F=0.24, df=1
p=0.59, F=0.54, df=2
Supplementary Materials
supplementary.zip contains the pool recorded of CCVs in wav format from our experiment, marked
with the timing variables (in Praat format).
|
1607.01706 | 3 | 1607 | 2017-01-10T04:18:23 | Optimal Trajectories of Brain State Transitions | [
"q-bio.NC"
] | The complexity of neural dynamics stems in part from the complexity of the underlying anatomy. Yet how the organization of white matter architecture constrains how the brain transitions from one cognitive state to another remains unknown. Here we address this question from a computational perspective by defining a brain state as a pattern of activity across brain regions. Drawing on recent advances in network control theory, we model the underlying mechanisms of brain state transitions as elicited by the collective control of region sets. Specifically, we examine how the brain moves from a specified initial state (characterized by high activity in the default mode) to a specified target state (characterized by high activity in primary sensorimotor cortex) in finite time. Across all state transitions, we observe that the supramarginal gyrus and the inferior parietal lobule consistently acted as efficient, low energy control hubs, consistent with their strong anatomical connections to key input areas of sensorimotor cortex. Importantly, both these and other regions in the fronto-parietal, cingulo-opercular, and attention systems are poised to affect a broad array of state transitions that cannot easily be classified by traditional notions of control common in the engineering literature. This theoretical versatility comes with a vulnerability to injury. In patients with mild traumatic brain injury, we observe a loss of specificity in putative control processes, suggesting greater susceptibility to damage-induced noise in neurophysiological activity. These results offer fundamentally new insights into the mechanisms driving brain state transitions in healthy cognition and their alteration following injury. | q-bio.NC | q-bio | Optimal Trajectories of Brain State Transitions
Shi Gua,b, Richard F. Betzelb, Marcelo G. Mattarc, Matthew Cieslakd, Philip R. Deliod,e, Scott T. Graftond, Fabio
Pasqualettif, Danielle S. Bassettb,g,∗
aApplied Mathematics and Computational Science,
University of Pennsylvania, Philadelphia, PA, 19104 USA
bDepartment of Bioengineering,
University of Pennsylvania, Philadelphia, PA, 19104 USA
cPrinceton Neuroscience Institute, Princeton University, Princeton NJ 08544
dDepartment of Psychological and Brain Sciences,
University of California, Santa Barbara, CA, 93106 USA
eNeurology Associates of Santa Barbara,
Santa Barbara, CA, 93105 USA
fDepartment of Mechanical Engineering,
University of California, Riverside, CA, 92521 USA
g Department of Electrical & Systems Engineering,
University of Pennsylvania, Philadelphia, PA, 19104 USA
7
1
0
2
The complexity of neural dynamics stems in part from the complexity of the underlying anatomy. Yet how white matter
structure constrains how the brain transitions from one cognitive state to another remains unknown. Here we address
this question by drawing on recent advances in network control theory to model the underlying mechanisms of brain state
transitions as elicited by the collective control of region sets. We find that previously identified attention and executive control
systems are poised to affect a broad array of state transitions that cannot easily be classified by traditional engineering-based
notions of control. This theoretical versatility comes with a vulnerability to injury. In patients with mild traumatic brain
injury, we observe a loss of specificity in putative control processes, suggesting greater susceptibility to neurophysiological
noise. These results offer fundamental insights into the mechanisms driving brain state transitions in healthy cognition and
their alteration following injury.
Keywords: network neuroscience, control theory, traumatic brain injury, cognitive control, diffusion imaging
Abstract
n
a
J
0
1
]
.
C
N
o
i
b
-
q
[
3
v
6
0
7
1
0
.
7
0
6
1
:
v
i
X
r
a
Introduction
The human brain is a complex dynamical system that transitions smoothly and continuously through states that directly
support cognitive function (Deco et al., 2011). Intuitively, these trajectories can map out the mental states that our brain
may pass through as we go about the activities of daily living. In a mathematical sense, these transitions can be thought of as
trajectories through an underlying state space (Shenoy et al., 2011; Freeman, 1994; Gu et al., 2016). While an understanding
of these trajectories is critical for our understanding of cognition and its alteration following brain injury, fundamental and
therefore generalizable mechanisms explaining how the brain moves through states have remained elusive.
One key challenge hampering progress is the complexity of these trajectories, which stems in part from the architectural
complexity of the underlying anatomy (Hermundstad et al., 2011, 2013a, 2014a). Different components (neurons, cortical
columns, brain areas) are linked with one another in complex spatial patterns that enable diverse neural functions (Rajan
et al., 2016; Fiete et al., 2010; Levy et al., 2001). These structural interactions can be represented as a graph or network,
where component parts form the nodes of the network, and where anatomical links form the edges between nodes (Bullmore
& Sporns, 2009). The architecture of these networks displays heterogenous features that play a role in neural function
(Medaglia et al., 2015), development (Di Martino et al., 2014), disease (Braun et al., 2015), and sensitivity to rehabilitation
(Weiss et al., 2011). Despite these recent discoveries, how architectural features constrain neural dynamics in any of these
phenomena is far from understood.
One simple and intuitive way to formulate questions about how neural dynamics are constrained by brain network
architecture is to define a state of the brain by the 1 × N vector representing magnitudes of neural activity across N brain
regions, and to further define brain network architecture by the N × N adjacency matrix representing the number of white
matter streamlines linking brain regions (Gu et al., 2015). Building on these two definitions, we can ask how the organization
of the white matter architecture constrains the possible states in which the brain can or does exist (Durstewitz & Deco, 2008;
Hansen et al., 2015). Moreover, building on decades of cognitive neuroscience research that have carefully delineated the
∗Corresponding author
Email address: [email protected] (Danielle S. Bassett)
Preprint submitted to NeuroImage
October 26, 2018
Figure 1: Conceptual Schematic. (A) Diffusion imaging data can be used to estimate connectivity from one voxel to any other voxel via
diffusion tractography algorithms. (B) From the tractography, we construct a weighted network in which N = 234 brain regions are connected by
the quantitative anisotropy along the tracts linking them (see Methods). (C) We study the optimal control problem in which the brain starts from
an initial state (red) at time t = 0 and uses multi-point control (control of multiple regions; blue) to arrive at a target state (yellow) at time t = T .
role of regional activation in cognitive functions (Gazzaniga, 2013; Szameitat et al., 2011; Alavash et al., 2015), we can then
map brain states to cognitive processes, and extend our question to: how does the organization of white matter architecture
constrain cognitive states (Hermundstad et al., 2013a, 2014b), and the processes that enable us to move between those
cognitive states (Cocchi et al., 2013)?
To address these questions, we draw on recent advances in network control theory (Pasqualetti et al., 2014) to develop
a biologically-informed mathematical model of brain dynamics from which we can infer how the topology of white matter
architecture constrains how the brain may affect (or control ) transitions between brain states. Within this model, we
examine finite-time transitions (from initial to target state) that are elicited via the collective control of many regions,
consistent with the collective dynamics observed to support cognition (Salvador et al., 2005; Meunier et al., 2009; Power
et al., 2011; Yeo et al., 2011) and action (Bassett et al., 2011b, 2013, 2015). A natural choice for an initial state is the
brain's well-known baseline condition, a state characterized by high activity in the precuneus, posterior cingulate, medial and
lateral temporal, and superior frontal cortex (Raichle, 2015; Raichle & Snyder, 2007; Raichle et al., 2001). While potential
transitions from this default mode are myriad, we focus this first study on examining transitions into target states of high
activity in sensorimotor cortex: specifically the extended visual, auditory, and motor cortices. These states represent the
simplest and most fundamental targets to transition from the default mode: for example, transitioning from the default mode
to visual states might represent an immediate response to a surprising stimuli. Similarly, the transition from the default
mode to motor states might represent the simple transition from rest to action. Moreover, these transitions are of particular
interest in many clinical disorders including stroke (Carter et al., 2012) and traumatic brain injury (Nudo, 2006; Lee et al.,
2011) where the cognitive functions performed by these target areas are often altered, significantly effecting quality of life
(Kalpinski et al., 2013).
Using network control theory, we examine the optimal trajectories from an initial state (composed of high activity in the
default mode system) to target states (composed of high activity in sensorimotor systems) with finite time and limited energy.
In this optimal control context, we investigate the role of white matter connectivity between brain regions in constraining
2
00.050.100.150.200.250.300.35RegionRegion0TABCInital StateTarget StateControl RegionsInputConnectivity Strength12341001234100 dynamic state transitions by asking three interrelated questions. First, we ask which brain regions are theoretically predicted
to be most energetically efficient in eliciting state transitions. Second, we ask whether these state transitions are best elicited
by one of three well-known control strategies commonly utilized in mechanical systems (Gu et al., 2015). Third, we ask how
specific each region's role is in these state transitions, and we compare this specificity between a group of healthy adults and a
group of patients with mild traumatic brain injury. In particular, the inclusion of this clinical cohort enables us to determine
whether widespread injury leads to a decrement in the healthy network control profiles, thus requiring greater energy for
the same functions, or an enhancement of the healthy network control profiles at the cost of a more fragile system, overly
sensitive to external perturbations. Together, these studies offer initial insights into how structural network characteristics
constrain transitions between brain states, and predict their alteration following brain injury.
To address these questions, we build structural brain networks from diffusion spectrum imaging (DSI) data acquired from
48 healthy adults and 11 individuals with mild traumatic brain injury (Fig. 1A). We perform diffusion tractography on these
images to estimate the quantitative anisotropy along the streamlines linking N = 234 large-scale cortical and subcortical
regions extracted from the Lausanne atlas (Cammoun et al., 2012; Daducci et al., 2012). We summarize these estimates
in a weighted adjacency matrix whose entries reflect the number of streamlines connecting different regions (Fig. 1B). We
then define a model of brain state dynamics informed by the weighted adjacency matrix, and we use this model to perform
a systematic study of the controllability of the system. This construction enables us to examine how structural network
differences between brain regions impact their putative roles in controlling transitions between cognitive states (Fig. 1C).
Materials and Methods
Data Acquisition and Brain Network Construction
Diffusion spectrum images (DSI) were acquired from 59 human adults with 72 scans in total, among which 61 scans were
acquired from 48 healthy subjects (mean age 22.6± 5.1 years, 24 female, 2 left handed) and 11 were acquired from individuals
with mild traumatic brain injury (Cieslak & Grafton, 2014)(mean age 33.8 ± 13.3 years, 4 female, handedness unclear). All
participants volunteered with informed written consent in accordance with the Institutional Review Board/Human Subjects
Committee, University of California, Santa Barbara. Deterministic fiber tracking using a modified FACT algorithm was
performed until 100, 000 streamlines were reconstructed for each individual. Consistent with previous work (Bassett et al.,
2010, 2011a; Hermundstad et al., 2013b, 2014a; Klimm et al., 2014; Gu et al., 2015; Muldoon et al., 2016a,b; Sizemore et al.,
2015), we defined structural brain networks from the streamlines linking N = 234 large-scale cortical and subcortical regions
extracted from the Lausanne atlas (Hagmann et al., 2008). We summarize these estimates in a weighted adjacency matrix A
whose entries Aij reflect the structural connectivity (quantitative anisotropy) between region i and region j (Fig. 1A). See
SI for further details.
Network Control Theory
Next, we consider the general question of how the brain moves between different states, where a state is defined as a
pattern of activity across brain regions or voxels. In particular, we are interested in studying how the activity in individual
brain regions affects the trajectory of the brain as it transitions between states; here, we define a trajectory as a set of states
ordered in time. To address this question, we follow (Gu et al., 2015; Muldoon et al., 2016a; Betzel et al., 2016) by adopting
notions from the emerging field of network control theory, which offers a theoretical framework for describing the role of
network nodes in the control of a dynamical networked system.
Network control theory is predicated on the choice of both a structural network representation for the system, and
a prescribed model of node dynamics.
In the context of the human brain, a natural choice for the structural network
representation is the graph on N brain regions whose ijth edge represents the QA between node i and node j. The choice
for the model of node dynamics is perhaps less constrained, as many models are available to the investigator. These models
range in complexity from simple linear models of neural dynamics with few parameters to nonlinear neural mass models with
hundreds of parameters (Gu et al., 2015; Muldoon et al., 2016a).
In choosing a model of neural dynamics to employ, we consider multiple factors. First, although the evolution of neural
activity acts as a collection of nonlinear dynamic processes, prior studies have demonstrated the possibility of predicting a
significant amount of variance in neural dynamics as measured by fMRI through simplified linear models (Gal´an, 2008; Honey
et al., 2009; Gu et al., 2015). On the basis of this literature, we employ a simplified noise-free linear continuous-time and
time-invariant network model
(1)
where x : R≥0 → RN describes the state of brain regions over time, and A ∈ RN×N is a symmetric and weighted adjacency
matrix. The diagonal elements of the matrix A satisfy Aii = 0. The input matrix BK identifies the control nodes K in the
brain, where K = {k1, . . . , km} and
x(t) = Ax(t) + Bu(t),
(2)
and ei denotes the i-th canonical vector of dimension N . The input uK : R≥0 → Rm denotes the control strategy. Intuitively,
this model enables us to frame questions related to brain state trajectories in a formal mathematics. Moreover, it allows us
BK = [ek1,··· , ek2]
3
to capitalize on recent advances in network control theory (Pasqualetti et al., 2014) to inform our understanding of internal
cognitive control (Gu et al., 2015; Betzel et al., 2016) and to inform the development of optimal external neuromodulation
using brain stimulation (Muldoon et al., 2016b).
Optimal Control Trajectories
Given the above-defined model of neural dynamics, as well as the structural network representation extracted from
diffusion imaging data, we can now formally address the question of how the activity in individual brain regions affects the
trajectory of the brain as it transitions between states.
We begin by defining an optimization problem to identify the trajectory between a specified pair of brain states that
minimizes a given cost function. We define a cost function by the weighted sum of the energy cost of the transition and
the integrated squared distance between the transition states and the target state. We choose this dual-term cost function
for two reasons. First, theoretically, the energy cost term constrains the range of the time-dependent control energy u(t).
In practice, this means that the brain cannot use an infinite amount of energy to perform the task (i.e., elicit the state
transition), a constraint that is consistent with the natural energetic restrictions implicit in the nature of all biological
systems but particularly neural systems (Niven & Laughlin, 2008; Laughlin et al., 1998; Attwell & Laughlin, 2001; Laughlin,
2001). Second, the term of the integrated distance term provides a direct constraint on the trajectory. Mathematically,
this constraint penalizes trajectories that traverse states that are far away from the target state, based on the intuition that
optimal transitions between states should possess reasonable lengths rather than being characterized by a random walk in
state space. Together, these two terms in the cost function enable us to define an optimal control model from which we
expect to find trajectories (from a given initial state to a specified target state) characterized by a balance between energy
cost and trajectory length.
In the context of the optimization problem defined above, we wish to determine the trajectory from an initial state x0 to a
target state xT . To do so, it suffices to solve the variational problem with the constraints from Equation 1 and the boundary
conditions for x(t), i.e. x(0) is the initial state and x(T ) is the target state. Note that here, the variational problem does
not refer to the Bayesian variational inference, which tries to approximate an intractable posterior distribution. Instead, we
use the term in the more traditional sense, and address the variational problem to infer a control input function u(t) to
minimize the cost functional defined in Equation [4] with the boundary constraints. Mathematically, the variational problem
is formulated as
(cid:0)(xT − x(t))T (xT − x(t)) + ρu(t)T u(t)(cid:1) dt,
(cid:90) T
min
u
s.t.
0
x(t) = Ax(t) + Bu(t),
x(0) = x0,
x(T ) = xT ,
where T is the control horizon, ρ ∈ R>0, and (xT − x(t)) is the distance between the state at time t and the target state.
To compute an optimal control u∗ that induces a transition from the initial state x0 to the target state xT , we define the
Hamiltonian as
(4)
From the Pontryagin minimum principle (Boltyanskii et al., 1960), if u∗ is an optimal solution to the minimization problem
with corresponding state trajectory x∗, then there exists p∗ such that
H(p, x, u, t) = xT x + ρuT u + pT (Ax + Bu).
which reduces to
Next, we denote
) + AT p∗
= − p∗,
∂H
∂x
∂H
∂u
(cid:21)
(cid:20) x∗
p∗
= −2(xT − x∗
= 2ρu∗
+ BT p∗
(cid:20) A −(2ρ)−1BBT
=
−2I
2xT
(cid:20)0
(cid:21)
I
,
+
(cid:21)
= 0.
p∗
−AT
(cid:21)
(cid:21)(cid:20)x∗
(cid:20) A −(2ρ)−1BBT
(cid:20)x∗
(cid:21)
(cid:21)
(cid:20)0
−AT
−2I
p∗
,
2xT ,
I
A =
x =
b =
then Eqn [7] can be written as
x = Ax + b,
4
(3)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
from which we can derive that
x + A−1 b = eAtc,
where c is a constant to be fixed from the boundary conditions. Let
t = 0, T with the corresponding x0 and xT , we have
(cid:35)
(cid:34) b1
b2
= A−1 b, e−AT =
(cid:20)E11 E12
(cid:21)
E21 E22
(12)
and plug in
(cid:20)x(0)
(cid:20)x(T )
p(0)
(cid:21)
(cid:21)
p(T )
+
+
(cid:35)
(cid:35)
(cid:34) b1
(cid:34) b1
b2
b2
=
=
b =
(cid:20) c1
(cid:21)
(cid:20)E11 E12
c2
,
(cid:21)−1(cid:20) c1
(cid:21)
.
E21 E22
c2
Note that from Equation[13], we can solve for c1, where
c1 = x(0) +
b1.
Finally, with c1 on hand from Eqn[14], we can compute p(T ), where
b1 − E12
p(T ) = E−1
12 (c1 − E11
b2 − E11x(T )),
with which we can finally get c2, where
c2 = E21x(T ) + E22p(T ) + E21
b1 + E22
b2
(13)
(14)
(15)
(16)
(17)
(20)
(21)
(cid:90) T
(cid:90) T
0
and further the u(t) and x(t) from Equation 12.
Note that the formulae we derive here are the closed form solutions to the optimization objective, and therefore a numerical
solver is not needed.
Statistics of Optimal Control Trajectories
After calculating the optimal trajectories between initial and final states, we next sought to address the question of whether
these trajectories differed in their energetic and spatial requirements for different choices of control strategies, and between
individual's whose brains were healthy and normally functioning, and individuals who had experienced a mild traumatic
brain injury and had presented with complaints of mild cognitive impairment. To address this question, we computed the
energy cost of a trajectory, integrated over time T , as
E(K, x0, xT ) =
u2K,x0,xT
dt,
(18)
and the spatial cost of a trajectory, integrated over time T , as
(19)
where uK,x0,xT is the associated control input and xK,x0,xT is the controlled trajectory with the given control set K, initial
state x0 and the target state xT . We treat this energy as a simple statistic that can be compared across trajectories and
subject groups, as an indirect measure from which we may infer optimality of cognitive function.
0
x2K,x0,xT
dt,
S(K, x0, xT ) =
Control Efficiency
The control efficiency is defined for each region to quantify its efficiency in affecting the transition from the default
mode state to the three target states. Mathematically, suppose we have N randomly chosen control sets, each indexed by
K1, . . . ,KN , for the target states xj
T , j = 1, 2, 3, we calculate the corresponding optimal trajectory with respect to Kk and
denote the energy cost of the trajectory as E(Kk, x0, xj
T is then defined
as
T ). The tiered value of the control set Kk for target xj
N(cid:88)
l=1
tkj =
1(E(Kl, x0, xj
T ) > E(Kk, x0, xj
T ))
where lower energy costs imply higher tiered values. The control efficiency for node i in task j is then
or intuitively, the average of these tiered values.
ζij =
(cid:80)N
(cid:80)N
k=1 1(i ∈ Kk) · tkj
k=1 1(i ∈ Kk)
.
5
Network Communicability to the Target State
G = exp(D−1/2AD−1/2), where D is the diagonal matrix with the diagonal element Dii =(cid:80)
communicability to all of the target regions, i.e. GTi =(cid:80)
For a given weighted network A, the network communicability G quantifies the extent of indirect connectivity among
nodes. Here we adopt the generalized definition in (Crofts & Higham, 2009) and define the network communicability as
i Aij. For a given target state
xT , denote the set of active regions as IxT , the communicability to the target states (GTi) is then defined as the sum of
Gij. Further, the normalized network communicability to
the target regions (Ci) is then defined as
.
(22)
j∈IxT
GTi(cid:80)
j GTj
Ci =
All results reported in this study are based on the normalized network communicability.
Energetic Impact of Brain Regions on Control Trajectories
To quantify the robustness of controllability of a node when it is removed from the control set consisting of all nodes, we
iteratively remove nodes from the network and compute the energetic impact of each region on the optimal trajectory as the
resulting increase in the log value of the energy cost. Intuitively, regions with high energetic impact are those whose removal
from the network causes the greatest increase in the energy required for the state transition. Mathematically, denote K0 as
the control set of all nodes and Ki as the control set without node [K]i, the energetic impact of node i for target xj
T is defined
as
Iij = log
E(Ki, x0, xj
T )
E(K0, x0, xj
T )
(23)
which intuitively measures robustness controllability.
Results
To begin, we set the initial state of the brain to be an activation pattern consistent with those empirically observed in
the brain's baseline condition. More specifically, we set the initial state such that the regions of the default mode network
had activity magnitudes equal to 1 ("on"), while all other regions had activity magnitudes equal to 0 ("off"). Furthermore,
we examined 3 distinct target states such that regions of the (i) auditory, (ii) extended visual, or (iii) motor systems had
activity magnitudes equal to 1 ("on"), while all other regions had activity magnitudes equal to 0 ("off"). In this context,
we sought to understand characteristics of the transitions between initial and target states that could be performed with
minimal energy, minimal time, and along short trajectories in state space by multiple control regions (multi-point control;
see Fig. 1C and Methods). We note that mathematically, we measure time in arbitrary units, at each of which control energy
can be utilized by a brain region. Intuitively, we operationalize time as consistent with the temporal scale at which brain
regions can alter their activity magnitudes to affect state transitions.
Characteristics of Optimal Control Trajectories
We first study the three state transitions from the default mode to (i) auditory, (ii) extended visual, and (iii) motor states
(Fig. 2A). We take a hypothesis-driven approach and define the "control set" to be composed of dorsal and ventral attention
(Posner & Petersen, 1989), fronto-parietal, and cingulo-opercular cognitive control regions (Gu et al., 2015). That is, this
set of 87 regions will utilize control energy using a multi-point control strategy, thereby changing the time-varying activity
magnitudes of all brain regions (Fig. 2B). The optimal trajectories display multiple peaks in the distance from the target
state as a function of time, and are altered very little by whether the target state is the auditory, extended visual, or motor
system (Fig. 2C). Because the optimal trajectory is determined via a balance of control energy and trajectory distance (see
Methods), it stands to reason that the time-dependent energy utilized by the control set is inversely related to the distance
between the current state and the target state. When little control energy is utilized, the current state can drift far from the
target state, while when a larger magnitude of control energy is utilized, the current state moves closer to the target state
(Fig. 2D).
It is important to note that these general characteristics of the optimal control trajectories are dependent on our choice
of the control set (which here we guide with biologically motivated hypotheses), as well as on a penalty on the time required
for the transition (ρ in Equation[3]; see Methods). In the supplement, we examine the effect of alternative choices for both
the control set and ρ. First, we find that when the control set includes every node in the network, the distance to the target
state decreases monotonically to zero along the trajectory (Fig. S1A). Second, we consider the effect of the penalty term on
control energy, ρ. For the results presented here, we fix ρ to be equal to 1. However, in the supplement, we explore a wide
range of ρ values, and show that when ρ is small, the optimal control trajectory is largely driven by a minimization of the
integrated squared distance to the target. In contrast, when ρ is large, the optimal control trajectory is largely driven by
the magnitude of the utilized energy (Fig. S1B). Importantly, we did not perform a full sweep of ρ from 0 to infinity because
very small values of ρ cause numeric instabilities in the calculations.
6
Figure 2: Optimal Control Trajectories. (A) We study 3 distinct types of state transitions in which the initial state is characterized by high
activity in the default mode system, and the target states are characterized by high activity in auditory (blue), extended visual (green), or motor
(red) systems. (B) The activation profiles of all N = 234 brain regions as a function of time along the optimal control trajectory, illustrating that
activity magnitudes vary by region and by time. Activation can be either positive or negative and the exact range of values will depend on the
initial state, the target state, and the control set. Regions are listed in the following order: initial state, target state, controllers and others. (C)
The average distance from the current state x(t) to the target state x(T ) as a function of time for the trajectories from the default mode system
to the auditory, visual, and motor systems, illustrating behavior in the large state space. (D) The average control energy utilized by the control
set as a function of time for the trajectories from the default mode system to the auditory, visual, and motor systems. The similarity of the curves
observed in panels (C) and (D) is driven largely by the fact that they share the same control set. See Fig. S2(B) for additional information on the
range of these control energy values along the trajectories. Colors representing target states are identical in panels (A), (C), and (D).
7
BACDDefault ModeVisual ModeSomatosensory ModeAuditory ModeEnergyEnergyEnergyTime (a.u.)2004006008001000Region50100150200-10010Time (a.u.)2004006008001000Region50100150200-10010Time (a.u.)2004006008001000Region50100150200-10010Time (a.u.)05001000Trajectory Distance, xt-xf0102030405060Time (a.u.)05001000Control Energy, u0200400600800State Value, xState Value, xState Value, x Structurally-Driven Task Preference for Control Regions
We next ask whether certain brain regions are located at specific points in the structural network that make them
predisposed to play consistent and important roles in driving optimal control trajectories. To answer this question, we choose
control sets of the same size as the brain's hypothesized cognitive control set; recall that in the previous section, we defined
the brain's cognitive control set to consist of the 87 nodes of the dorsal and ventral attention, fronto-parietal, and cingulo-
opercular systems following (Gu et al., 2015). Here, we choose the 87 regions of these new control sets uniformly at random
from the set of all nodes. Using these "random" control sets, we computed the optimal control trajectory for each of the
three state transitions and for each subject separately. Then, we rank the random control sets in descending order according
to the energy cost of the trajectory and we assign every region participating in an r-ranked control set with rank-value r.
Next, we define the control efficiency of a brain region to be the sum of its rank values in all of the random control sets it
belongs divided by the total number of sets it belongs to. Intuitively, a region with a high control efficiency is one that exerts
control with little energy utilization. Importantly, it must decrease activation in the initial state, and increase activation in
the target state, a pair of capabilities that depend on the pattern of connections emanating from the region.
In general, we observe that a region's preference for being an optimal controller (exerting control with little energy
utilization) is positively correlated with its network communicability to the regions of high activity in the target state
(Spearman correlation r = 0.27, p < 4.8 × 10−4; see Fig. 3A). We recall that network communicability is a measurement of
the strength of a connection from one region to another that accounts for walks of all lengths (see Methods). Interestingly,
we observed this same correlation between control efficiency and network communicability across optimal control trajectories
for all three state transitions, from the default mode to the auditory (r = 0.36, p = 1.4 × 10−8), extended visual (r = 0.51,
p = 1.1× 10−16), or motor (r = 0.42, p = 2.1× 10−11) systems (Fig. 3B-D). Together, these results indicate that regions that
are close (in terms of walk lengths) to regions of high activity in the target state are efficient controllers for that specific state
transition. Note that these regions are not purely target areas, likely due to the fact that they must also decrease activation
in the initial state.
The general role that network proximity to the target state plays for control regions ensures that regions that are
proximate to all three target states (auditory, extended visual, and motor) will be consistent controllers, while regions that
are proximate to only one of the target states will be task-specific controllers. To better understand the anatomy of efficient
controllers, we transformed control efficiency values to z-scores and defined an efficient control hub to be any region whose
associated p-value was less than 0.025. Across all three state transitions, we found that the supramarginal gyrus specifically,
and the inferior parietal lobule more generally, consistently acted as efficient control hubs. The consistent control role of
these regions is likely due to the fact that these areas are structurally interconnected with ventral premotor cortex, a key
input to primary sensorimotor areas (Kandel et al., 2000). The areas that are more specific to the three state transitions
include medial parietal cortex (motor transition), orbitofrontal and inferior temporal cortex (visual transition), and superior
temporal cortex (auditory transition).
Regional Roles in Control Tasks
The analyses outlined above are built on the assumption that the brain uses fronto-parietal, cingulo-opercular, and
attention systems to affect cognitive control, which we define as the ability to move the brain from an initial state (e.g., the
default mode systems) to a specified final state (e.g., activation of extended visual, auditory, or motor cortex). However, one
might naturally ask whether these regions of the brain could have been predicted a priori to be effective controllers based on
traditional engineering-based notions of control. In the control theory literature, particularly the literature devoted to the
subfield of network controllability, there exist several controllability notions, including average, modal, and boundary control
(Pasqualetti et al., 2014). Average controllability identifies brain areas that can theoretically steer the system into many
different states, or patterns of neurophysiological activity magnitudes across brain regions. Modal controllability identifies
brain areas that can theoretically steer the system into difficult-to-reach states. Boundary controllability identifies brain
areas that can theoretically steer the system into states where different cognitive systems are either coupled or decoupled.
See the SI for mathematical definitions and (Gu et al., 2015) for prior studies in human neuroimaging.
We calculated average, modal, and boundary control values for each node in the network. We observe that while cognitive
control regions cover a broad swath of frontal and parietal cortex, including medial frontal cortex and anterior cingulate
(Fig. 4A), the number of these regions that intersect with the strongest 87 average, modal, or boundary control hubs was on
average approximately 50 (Fig. 4B). These results suggest that the control capabilities of the human brain's cognitive control
regions may not be perfectly aligned with control notions previously developed in the field of mechanical engineering, provided
that the model assumptions and data quality are appropriate (see Methodological Considerations). Instead, cognitive control
regions in the human brain may have distinct capabilities necessary for the specific transitions required by the brain under
the constraints imposed by neuroanatomy and neurophysiology.
To more directly test this possibility, we examined the average distance (Fig. 4C) and energy (Fig. 4D) for transitions
from the default mode to the auditory, extended visual, and sensorimotor states that are driven by average, modal, and
boundary control hubs, or by regions of fronto-parietal, cingulo-opercular, and attention systems. We observed that both
the trajectory cost and the energy cost differ by control strategy and by target state. We quantify this observation using
a 2-way ANOVA with both the control strategy and target state as categorical factors. Using the trajectory cost as the
dependent variable, we observed a significant main effect of control strategy (F = 78.74, p = 4.65× 10−41), a significant main
8
Figure 3: Structurally-Driven Task Preference for Control Regions. (A) Top Regions with high control efficiency (see Eqn 21) across
all 3 state transitions: from the default mode to auditory, extended visual, and motor systems. Bottom Scatterplot of the control efficiency with
the average network communicability to all 3 target regions (Spearman correlation r = 0.27, p < 4.8 × 10−4). (B -- D) Top Regions with high
control efficiency for the transition from default mode to (B) motor, (C) extended visual, and (D) auditory (r = 0.36, p = 1.4 × 10−8) targets
(top). Bottom Scatter plot of control efficiency versus normalized network communicability with regions that are active in the target state: motor
(r = 0.42, p = 2.1 × 10−11), extended visual (r = 0.51, p = 1.1 × 10−16), and auditory (r = 0.36, p = 1.4 × 10−8). Values of control efficiency in
all four panels are averaged of subjects.
effect of target state (F = 29.24, p = 1.12 × 10−12), and a significant interaction between control strategy and target state
(F = 11.36, p = 7.6× 10−12). Similarly, using the energy cost as the dependent variable, we observed a significant main effect
of control strategy (F = 67.94, p = 2.48 × 10−36), a significant main effect of target state (F = 39.18, p = 1.99 × 10−16),
and a significant interaction between control strategy and target state (F = 10.93, p = 2.18 × 10−11). Collapsing over target
states and performing post-hoc testing, we observed that cognitive control regions displayed a similar average trajectory cost
to average control hubs, but a lower average trajectory cost than modal and boundary control hubs (p < 0.05 uncorrected).
Furthermore, cognitive control regions possessed a higher average energy cost than the average and modal control hubs, but
a lower average energy cost than the boundary control hubs. These results interestingly suggest that the human's cognitive
control regions, as defined by decades of research in cognitive neuroscience, may affect state transitions using neither the
shortest distances nor the lowest energies possible, provided that the model assumptions and data quality are appropriate
(see Methodological Considerations). This is likely due to the fact that cognitive control regions must affect a broad array
of state transitions that cannot easily be classified into average, modal, and boundary control strategies.
Specificity of Control in Health and Following Injury
The unique role of brain regions in affecting control strategies may bring with it vulnerability to injury. When a brain
network is injured, regional control roles may be significantly altered, potentially increasing susceptibility to underlying
abnormalities in neuronal dynamics. To characterize this vulnerability, we determine the degree to which a single brain
region impacts putative control processes and we ask whether that specificity is maintained or altered following brain injury.
We measure specificity by iteratively removing nodes from the control set, and we compute the energetic impact of each
region on the optimal trajectory as the resulting increase in the log value of the energy cost (see Fig. 5A and Eqn 23 in
Methods).
Intuitively, regions with high energetic impact are those whose removal from the network causes the greatest
increase in the energy required for the state transition. Across all subjects and all tasks, we observe that the regions with
the highest energy impact are the supramarginal gyrus specifically, and the inferior parietal lobule more generally, the same
regions that emerged as consistent and efficient controllers in Fig. 2A.
Next we determined whether energetic impact -- our proxy for regional specificity of control roles -- is altered in individuals
with mild traumatic brain injury (mTBI). Intuitively, if all regions of a brain have high energetic impact, this indicates that
each region is performing a different control role which is destroyed by removal of the node. By contrast, if all regions of a brain
have low average energetic impact, this indicates that each region is performing a similar control role that is not destroyed
by removal of a node. We observed that individuals with mTBI displayed anatomically similar patterns of energetic impact
on control trajectories as regions are removed from the network (Fig. 5B). However, the average magnitude and variability
of the energetic impact differed significantly between the two groups, with individuals having experienced mTBI displaying
9
SomatosensoryControl E(cid:31)ciencyControl E(cid:31)ciency Control E(cid:31)ciencty Control E(cid:31)ciencyRank of Communicability with TargetsRank of Communicability with TargetsRank of Communicability with TargetsRank of Communicability with TargetsVisualABCDCombinationAuditory050100150200250×10-33.84.04.24.44.64.85050100150200250×10-33.844.24.44.64.85050100150200250×10-33.844.24.44.64.85050100150200250×10-33.844.24.44.64.85 Figure 4: Regional Roles in Control Tasks. (A) Cognitive control regions cover a broad swath of frontal and parietal cortex, including medial
frontal cortex and anterior cingulate, and are defined as regions included in fronto-parietal, cingulo-opercular, and attention systems (Gu et al.,
2015). (B) The number of these regions overlapping with the strongest 87 average, modal and boundary control hubs is approximately 50. Different
choices of control strategies result in variation in both (C) trajectory cost and (D) energy cost. Here, HC refers to cognitive control regions, AC
refers to average control hubs, MC refers to modal control hubs, and BC refers to boundary control hubs.
significantly lower values of average magnitude of energetic impact (permutation test: p = 5.0 × 10−6) and lower values of
the average standard deviation of energetic impact (p = 2.0 × 10−6). We note that common graphic metrics including the
degree, path length, clustering coefficient, modularity, local efficiency, global efficiency, and density were not significantly
different between the two groups, suggesting that this effect is specific to control (see Supplement). These results indicate
that mTBI patients display a loss of specificity in the putative control roles of brain regions, suggesting greater susceptibility
to damage-induced noise in neurophysiological processes, or to external drivers in the form of stimulation.
Discussion
Here we ask whether structural connectivity forms a fundamental constraint on how the brain may move between diverse
cognitive states. To address this question, we capitalize on recent advances in network control theory to identify and
characterize optimal trajectories from an initial state (composed of high activity in the default mode system) to target states
(composed of high activity in sensorimotor systems) with finite time, limited energy, and multi-point control. Using structural
brain networks estimated from diffusion imaging data acquired in a large cohort of 48 healthy individuals and 11 patients
with mild traumatic brain injury, we show that these optimal control trajectories are characterized by continuous changes
in regional activity across the brain. We show that the regions critical for eliciting these state transitions differ depending
on the target state, but that heteromodal association hubs -- predominantly in the supramarginal gyrus specifically, and the
inferior parietal lobule more generally -- are consistently recruited for all three transitions. Finally, we study the sensitivity
of optimal control trajectories to the removal of nodes from the network, and we demonstrate that brain networks from
individuals with mTBI display maladaptive control capabilities suggestive of a limited dynamic range of states available to
the system. Together, these results offer initial insights into how structural network differences between individuals impact
their potential to control transitions between cognitive states.
Role of Structural Connectivity in Shaping Brain Functional Patterns
A growing body of literature on the relationship between brain structure and function has demonstrated that the brain's
network of anatomical connections constrains the range of spontaneous (Deco et al., 2011) and task-related (Hermundstad
et al., 2013b) fluctuations in brain activity. Evidence for such structural underpinnings comes from two distinct lines of
research. On one hand, empirical studies have demonstrated that structural insults in the form of lesions result in acute
reorganization of the brain's pattern of functional coupling (Johnston et al., 2008; O'Reilly et al., 2013). These observations
are further buttressed by simulation studies in which structural connectivity has been used to constrain interactions among
dynamic elements in biophysical models of brain activity (Honey et al., 2007, 2009; Adachi et al., 2011) and models of network
communication (Goni et al., 2014; Abdelnour et al., 2014; Misi´c et al., 2015). Though this forward modeling approach has
10
Control StrategiesControl StrategiesControl StrategiesControl StrategiesControl StrategiesControl StrategiesControl StrategiesNumber of Common NoddesFronto-parietal, cingulo-opercular,and attention systemsDMN --> SMDMN --> VSDMN --> ADDMN --> ADDMN --> VSDMN --> SMABCD253035404550ACACMCBCHCACMCBCHCACMCBCHCACMCBCHCACMCBCHCACMCBCHCMCBC468101214567891011121346810121416186810121416182010111213141516171881012141618202224Log Energy CostLog Energy CostLog Energy CostLog Spatial CostLog Spatial CostLog Spatial Cost01 Figure 5: Specificity of Control in Health and Following Injury. (A) Theoretically, the brain is fully controllable when every region is a
control point, but may not be fully controllable when fewer regions are used to affect control. (B) The regions with the highest values of energetic
impact on control trajectories upon removal from the network, on average across subjects and tasks, were the supramarginal gyrus specifically, and
the inferior parietal lobule more generally. In general, the healthy group and the mTBI group displayed similar anatomical patterns of energetic
impact. (C) Magnitude and standard derivation of energetic impact averaged over regions and tasks; boxplots indicate variation over subjects.
Even after removing the single outlier in the healthy group, patients with mTBI displayed significantly lower values of average magnitude of
energetic impact (permutation test: p = 1.1 × 10−5) and lower values of the average standard deviation of energetic impact (p = 2.0 × 10−6) than
healthy controls.
11
Controlled RegionsControlled RegionsControlled RegionsUncontrolled RegionsThe Fully Controlled BrainThe Almost Fully Controlled BrainInitial StateTarget StateHealthy GroupHealthy GroupTBI GroupHealthy GroupTBI GroupE0E1∆EABDCMean of Energetic ImpactStandard Deviation of Energetic Impact0.0017mTBI Group0.68910.00170.7180Energetic ImpactEnergetic Impact0.120.140.160.180.20.22p = 5.0 x 10-60.30.350.40.450.50.55p = 2.0 x 10-6 proven fruitful in predicting observed patterns of functional connectivity, the precise mapping of brain structure to function
remains unclear.
The present study builds on this body of work, using a dynamical model of how brain activity propagates over a network in
order to gain insight into what features of that network facilitate easy transitions from a baseline (default mode) state to states
where the brain's primary sensorimotor systems are activated. In contrast to previous simulation studies that have focused
on network features that influence the passive spread of activity over time, this present study directly engages the question of
how those same features enable the state of the system to be controlled. We use this model to demonstrate that brain regions
are differentially-suited for particular control tasks, roles that can be predicted on the basis of how well-connected they are to
regions in the target state. Regions that are close (in terms of walk lengths) to regions of high activity in the target state are
efficient controllers for that specific state transition. It follows, then, that a brain region's capacity to dynamically influence
a network depends not only on its pattern of connectivity, but also the repertoire of states that the system visits. In other
words, a region that maintains many connections (both direct and indirect), but never to regions that are "active" in target
states, may exert less influence than a region that maintains few connections, but whose connections are distributed among
regions that are "active" in many target states. We further demonstrate that this mapping of brain structure to specific
functions is altered in individuals with mTBI, suggesting that injury may alter control profiles of individual brain regions.
Our finding that the inferior parietal lobule forms a consistently effective control region, across all three target states,
is particularly interesting when considered in the context of prior literature on this region's structural and functional roles.
In particular, the inferior parietal lobule represents the superior portion of the temporoparietal junction, a multimodal area
associated with functions as wide ranging as calculation, finger gnosis, left/right orientation, and writing (Rusconi et al.,
2009). Focal damage to this area leads to wide-spread cognitive disfunction (e.g., Gerstmann syndrome) as a consequence of
the unique confluence of white matter pathways underlying this region (Rushworth et al., 2006). The diverse white matter
projections emanating from this area may support its putative role in effectively controlling brain function. Indeed, recent
evidence suggests that the right temporoparietal junction links two antagonistic brain networks processing external versus
internal information: a midcingulate-motor-insular network associated with attention, and a parietal network associated with
social cognition and memory retrieval (Bzdok et al., 2013). These data support the notion that the right temporoparietal
junction controls our attention to salient external events (Corbetta & Shulman, 2002), perhaps with early input from the
right fronto-insular cortex thought to drive switching between central-executive and default-mode networks (Sridharan et al.,
2008).
Single versus Multipoint Control
An important feature of our model lies in the delineation of a control set, a group of brain regions that can affect distributed
control. The focus on multiple points of control throughout the system is one that has important theoretical motivations
and empirical correlates. Prior computational models demonstrate that while the brain is theoretically controllable via input
to a single control point, the energy and time required for that control is such that the brain is practically uncontrollable
(Gu et al., 2015). These data argue for an assessment of multi-point control as a better proxy of control strategies that
the brain might utilize. Indeed, such an argument is consistent with empirical observations that stimulation (or even drug
manipulations) focused on single brain regions are less effective in treating psychiatric disease than interventions that target
multiple brain regions (Sommer et al., 2012; Tortella et al., 2014). A prime empirical example of multi-point control is
cognitive behavioral therapy, which offers a spatio-temporal pattern of activations that enhances cognitive function and
decreases psychiatric symptoms across diagnostic categories (Lett et al., 2014; Radhu et al., 2012; Cima et al., 2014). Other
potential multi-point control mechanisms include grid stimulation across multiple electrodes, suggested in the control of
medically refractory epilepsy (Ching et al., 2012).
Diversity in Human Brain Control Strategies
By studying multi-point control, we were able to directly assess whether the optimal control trajectories elicited from
fronto-parietal, cingulo-opercular, and attention systems displayed similar distance and energy to trajectories to those ob-
tained using regions selected for engineering-based notions of control (Pasqualetti et al., 2014). Specifically, we compare and
contrast the performance of human cognitive control regions to average, modal, and boundary controllers Gu et al. (2015).
One might naturally ask whether and how each of these engineering-based notions of controllers is an appropriate theoretical
quantity to consider in the context of neurobiology. Average controllers are those theoretically able to push the system
from any arbitrary initial state to any easily-reachable state, nearby on the energy landscape. Modal controllers are those
that are optimally placed to move the system from any arbitrary initial state to any difficult-to-reach state, far away on the
energy landscape. Boundary controllers are those that are optimally placed to integrate or segregate network communities
in the system. Common assumptions that underlie each of these control strategies are that (i) controllers can be identified
independently from the initial and final states, and (ii) all states in the energy landscape are accessible to the system. The
common constraint on each of these strategies is that expended energy must be minimized for a region to be referred to as a
controller. In the human brain, it is not well-known whether these assumptions are met, or whether the energetic constraint
is sufficient to predict control functions.
Interestingly, we observe that the optimal trajectories elicited by canonically defined cognitive control regions do not
show similar energy requirements or trajectory distances to any of these previously described control types. There are several
12
potential reasons for this observation: (i) false negatives in the diffusion imaging data impacting on the observed network
profiles, (ii) assumptions of the linear model, and (iii) bona fide differences between mechanical controllers and biological
controllers. Prior work demonstrating robustness of controllability profiles across large cohorts and different diffusion imaging
acquisition protocols provide initial evidence that the first explanation is unlikely to fully explain our findings (Gu et al., 2015).
Regarding the second explanation -- assumptions of the linear model -- it is interesting to note that recent evidence suggests
that the average, modal, and boundary controllability profiles identified by the linear model provide excellent predictions
for the behavior of nonlinear models (Muldoon et al., 2016b). Evidence supporting the third interpretation -- that these
observed differences are bona fide differences between mechanical controllers and biological controllers -- is provided by the
fact that cognitive control regions must affect a broad array of state transitions that do not easily fit into prior classifications.
These transitions include switching behavior (Hansen et al., 2015), inter-state competition (Cocchi et al., 2013), distributed
rather than centralized control (Eisenreich et al., 2016), and push-pull control (Khambhati et al., 2016), which may each offer
differential advantages for neural computations (Durstewitz & Deco, 2008).
Maladaptive Control in Traumatic Brain Injury
Finally, our assessment of patients with mild traumatic brain injury enabled us to determine whether widespread injury
leads to a decrement in the healthy network control profiles, thus theoretically requiring greater energy for the same functions,
or an enhancement of the healthy network control profiles at the cost of expected sensitivity to external perturbations.
Our data provide initial evidence for maladaptive control of the latter sort in patients with mild traumatic brain injury.
Understanding the impact of brain injury on cognitive processes, including the ability to switch between cognitive states,
is a major goal in clinical neuroscience. Indeed, traumatic brain injury is a common source of brain dysfunction, affecting
more than 200,000 individuals per year in the United States alone.
Injuries -- often caused by motor vehicle and sports
accidents -- result in damage to neuronal axons, including long-distance white matter fiber bundles (Johnson et al., 2013) as
well as u-fibers and deep white matter tracks with multiple crossings. The pattern of injury can be multi-focal and variable
across individuals (Kinnunen et al., 2011; Sidaros et al., 2008; Hellyer et al., 2013), challenging comprehensive predictors and
generalizable interventions.
Recent evidence suggests that injury-induced, widespread damage to white matter tracts critically impacts large-scale
network organization in the human brain, as measured by diffusion imaging tractography (Kinnunen et al., 2011; Fagerholm
et al., 2015). Moreover, this damage is associated with fundamental changes in cognitive function (Sharp et al., 2014),
including information processing speed, executive function, and associative memory (Fagerholm et al., 2015). Each of these
cognitive deficits intuitively depends on the ability to transition from one cognitive state to another; yet an understanding
of structural drivers of these transitions and their potential alteration in mTBI has remained elusive. Here we demonstrate
a loss of specificity in putative control processes in mTBI, suggesting that the unique roles of individual brain regions in
supporting cognitive state transitions are damaged. It is intuitively plausible that this decrement in regional specificity of
control leads to broad changes in functional dynamics, particularly in the system's susceptibility to damage-induced noise
in neurophysiological processes (Garrett et al., 2013). Indeed, the observed decrements in energetic impact might further
provide a direct structural mechanism for the decreased signal variability observed in mTBI using electrophysiological imaging
(Raja Beharelle et al., 2012; Nenadovic et al., 2008). More generally, these findings highlight the fact that the healthy brain
might display a degree of controllability that is either decremented or enhanced in injury and disease, suggesting the possibility
of a U-shaped curve reminiscent of similar curves observed in other brain network phenotypes (Collin & van den Heuvel,
2013; Cools & D'Esposito, 2011).
Methodological Considerations
A few methodological points are worthy of additional consideration. First, in this study we examined structural brain
networks derived from diffusion imaging data and associated tractography algorithms. These algorithms remain in their
relative infancy, and can still report spurious tracts or fail to report existing tracts (Thomas et al., 2014; Reveley et al.,
2015; Pestilli et al., 2014). Despite the evolving nature of diffusion protocols and tractography algorithms, preliminary data
provide initial evidence that consistent controllability profiles can be robustly observed across large cohorts and different
diffusion imaging acquisition protocols (Gu et al., 2015). Formal validation in axonal tracing studies in monkeys and other
mammals (Jbabdi et al., 2013) remains the gold standard for these types of data. However, it is important to note that initial
work supports the notion that much of the structure present in DSI connectivity matrices recapitulates known projections
observed in tract tracing studies of the macaque Hagmann et al. (2008).
Second, following (Gu et al., 2015; Betzel et al., 2016; Muldoon et al., 2016b), we employ a linear dynamical model,
consistent with prior empirical studies demonstrating their ability to predict features of resting state fMRI data (Gal´an,
2008; Honey et al., 2009). This choice is to some degree predicated on the well-developed theoretical and analytical results
in the engineering and physics literatures examining the relationship between control and network topology (Liu et al., 2011;
Muller & Schuppert, 2011; Yan et al., 2012). Moreover, it is plausible that even these results using simple linear models may
offer important intuitions for controlling nonlinear models of brain function. Indeed, theoretical work over the last several
decades has demonstrated the utility of describing non-linear systems in terms of a linear approximation in the neighborhood
of the system's equilibrium points (Luenberger, 1979). Very recent evidence has extended these intuitions to neuroimaging
data, demonstrating that the average, modal, and boundary controllability profiles identified by the linear model can be used
13
to predict the behavior of nonlinear models in the form of Wilson-Cowan oscillators, which are commonly used to understand
the dynamics of cortical columns (Muldoon et al., 2016b).
Future directions
An interesting hypothesis generated by the current framework is that control capabilities may be altered dimensionally
across traditionally separated diagnostic groups that display dysconnectivity in network hubs, as measured by regions of
high eigenvector centrality. Such a hypothesis builds on the now seminal dysconnection hypothesis in schizophrenia (Stephan
et al., 2009), and expands it to include an explicit dynamical control component. Indeed, mounting evidence suggests that
the overload or failure of brain network hubs may be a common neurophysiological mechanism of a range of neurological
disorders including Alzheimer's disease, multiple sclerosis, traumatic brain injury and epilepsy (Stam, 2014). Alterations in
these hubs can also be used to predict the progression of psychiatric disorders such as schizophrenia (Collin et al., 2015). In
both neurological and psychiatric disorders, these changes to network hubs may alter the control capabilities of the individual,
challenging the normal executive functions required for daily living. It is also intuitively plausible that normal variation in
hub architecture may play a role in individual differences in control capabilities in healthy individuals, impacting on the
speed with which they transition between cognitive states. These topics will form important provender for future work.
Acknowledgements:
D.S.B., S.G., and R.F.B. acknowledge support from the John D. and Catherine T. MacArthur Foundation, the Alfred P.
Sloan Foundation, the Army Research Laboratory and the Army Research Office through contract numbers W911NF-10-2-
0022 and W911NF-14-1-0679,the National Institute of Mental Health (2-R01-DC-009209-11), the National Institute of Child
Health and Human Development (1R01HD086888-01), the Office of Naval Research, and the National Science Foundation
(BCS-1441502, BCS-1430087, and CAREER PHY-1554488). S.T.G. and P.R.D. acknowledge support from a Head Health
Challenge grant from General Electric and the National Football League. F.P. acknowledges support from the National
Science Foundation award #BCS 1430280. The content is solely the responsibility of the authors and does not necessarily
represent the official views of any of the funding agencies.
Author Contributions
S.G. performed the analysis. M.C. preprocessed the data. S.T.G. and P.R.D. acquired the data. S.G., R.F.B., F.P.,
D.S.B. developed the project. S.G., R.F.B., F.P., D.S.B. wrote the paper.
Competing Interests
The authors declare that they have no competing financial interests.
References
References
Abdelnour, F., Voss, H. U., & Raj, A. (2014). Network diffusion accurately models the relationship between structural and
functional brain connectivity networks. Neuroimage, 90 , 335 -- 347.
Adachi, Y., Osada, T., Sporns, O., Watanabe, T., Matsui, T., Miyamoto, K., & Miyashita, Y. (2011). Functional connectivity
between anatomically unconnected areas is shaped by collective network-level effects in the macaque cortex. Cerebral cortex ,
(p. bhr234).
Alavash, M., Hilgetag, C. C., Thiel, C. M., & Giessing, C. (2015). Persistency and flexibility of complex brain networks
underlie dual-task interference. Human brain mapping, 36 , 3542 -- 3562.
Attwell, D., & Laughlin, S. B. (2001). An energy budget for signaling in the grey matter of the brain. J Cereb Blood Flow
Metab, 21 , 1133 -- 1145.
Bassett, D. S., Brown, J. A., Deshpande, V., Carlson, J. M., & Grafton, S. T. (2011a). Conserved and variable architecture
of human white matter connectivity. Neuroimage, 54 , 1262 -- 1279.
Bassett, D. S., Greenfield, D. L., Meyer-Lindenberg, A., Weinberger, D. R., Moore, S. W., & Bullmore, E. T. (2010).
Efficient physical embedding of topologically complex information processing networks in brains and computer circuits.
PLoS Comput Biol , 6 , e1000748.
14
Bassett, D. S., Wymbs, N. F., Porter, M. A., Mucha, P. J., Carlson, J. M., & Grafton, S. T. (2011b). Dynamic reconfiguration
of human brain networks during learning. Proc Natl Acad Sci U S A, 108 , 7641 -- 7646.
Bassett, D. S., Wymbs, N. F., Rombach, M. P., Porter, M. A., Mucha, P. J., & Grafton, S. T. (2013). Task-based core-
periphery organization of human brain dynamics. PLoS Comput Biol , 9 , e1003171.
Bassett, D. S., Yang, M., Wymbs, N. F., & Grafton, S. T. (2015). Learning-induced autonomy of sensorimotor systems. Nat
Neurosci , 18 , 744 -- 751.
Betzel, R. F., Gu, S., Medaglia, J. D., Pasqualetti, F., & Bassett, D. S. (2016). Optimally controlling the human connectome:
the role of network topology. arXiv preprint arXiv:1603.05261 , .
Boltyanskii, V. G., Gamkrelidze, R. V., & Pontryagin, L. S. (1960). The theory of optimal processes. I. The maximum
principle. Technical Report DTIC Document.
Braun, U., Muldoon, S. F., & Bassett, D. S. (2015). On human brain networks in health and disease. eLS , .
Bullmore, E., & Sporns, O. (2009). Complex brain networks: graph theoretical analysis of structural and functional systems.
Nature Reviews Neuroscience, 10 , 186 -- 198.
Bzdok, D., Langner, R., Schilbach, L., Jakobs, O., Roski, C., Caspers, S., Laird, A. R., Fox, P. T., Zilles, K., & Eickhoff,
S. B. (2013). Characterization of the temporo-parietal junction by combining data-driven parcellation, complementary
connectivity analyses, and functional decoding. Neuroimage, 81 , 381 -- 392.
Cammoun, L., Gigandet, X., Meskaldji, D., Thiran, J. P., Sporns, O., Do, K. Q., Maeder, P., Meuli, R., & Hagmann,
P. (2012). Mapping the human connectome at multiple scales with diffusion spectrum MRI. J Neurosci Methods, 203 ,
386 -- 397.
Carter, A. R., Patel, K. R., Astafiev, S. V., Snyder, A. Z., Rengachary, J., Strube, M. J., Pope, A., Shimony, J. S., Lang, C. E.,
Shulman, G. L., & Corbetta, M. (2012). Upstream dysfunction of somatomotor functional connectivity after corticospinal
damage in stroke. Neurorehabil Neural Repair , 26 , 7 -- 19.
Ching, S., Brown, E. N., & Kramer, M. A. (2012). Distributed control in a mean-field cortical network model: implications
for seizure suppression. Phys Rev E Stat Nonlin Soft Matter Phys, 86 , 021920.
Cieslak, M., & Grafton, S. (2014). Local termination pattern analysis: a tool for comparing white matter morphology. Brain
imaging and behavior , 8 , 292 -- 299.
Cima, R. F., Andersson, G., Schmidt, C. J., & Henry, J. A. (2014). Cognitive-behavioral treatments for tinnitus: a review
of the literature. J Am Acad Audiol , 25 , 29 -- 61.
Cocchi, L., Zalesky, A., Fornito, A., & Mattingley, J. B. (2013). Dynamic cooperation and competition between brain systems
during cognitive control. Trends in cognitive sciences, 17 , 493 -- 501.
Collin, G., & van den Heuvel, M. P. (2013). The ontogeny of the human connectome: development and dynamic changes of
brain connectivity across the life span. Neuroscientist, 19 , 616 -- 628.
Collin, G., de Nijs, J., Hulshoff Pol, H. E., Cahn, W., & van den Heuvel, M. P. (2015). Connectome organization is related
to longitudinal changes in general functioning, symptoms and IQ in chronic schizophrenia. Schizophr Res, S0920 -- 9964 ,
00141 -- 00143.
Cools, R., & D'Esposito, M. (2011). Inverted-U-shaped dopamine actions on human working memory and cognitive control.
Biol Psychiatry, 69 , e113 -- e125.
Corbetta, M., & Shulman, G. L. (2002). Control of goal-directed and stimulus-driven attention in the brain. Nat Rev
Neurosci , 3 , 201 -- 215.
Crofts, J. J., & Higham, D. J. (2009). A weighted communicability measure applied to complex brain networks. Journal of
the Royal Society Interface, (pp. rsif -- 2008).
Daducci, A., Gerhard, S., Griffa, A., Lemkaddem, A., Cammoun, L., Gigandet, X., Meuli, R., Hagmann, P., & Thiran,
J. P. (2012). The connectome mapper: an open-source processing pipeline to map connectomes with MRI. PLoS One, 7 ,
e48121.
Deco, G., Jirsa, V. K., & McIntosh, A. R. (2011). Emerging concepts for the dynamical organization of resting-state activity
in the brain. Nat Rev Neurosci , 12 , 43 -- 56.
15
Di Martino, A., Fair, D. A., Kelly, C., Satterthwaite, T. D., Castellanos, F. X., Thomason, M. E., Craddock, R. C., Luna,
B., Leventhal, B. L., Zuo, X.-N. et al. (2014). Unraveling the miswired connectome: a developmental perspective. Neuron,
83 , 1335 -- 1353.
Durstewitz, D., & Deco, G. (2008). Computational significance of transient dynamics in cortical networks. Eur J Neurosci ,
27 , 217 -- 227.
Eisenreich, B., Akaishi, R., & Hayden, B. (2016). Control without controllers: Towards a distributed neuroscience of executive
control. bioRxiv , 1101 , 077685.
Fagerholm, E. D., Hellyer, P. J., Scott, G., Leech, R., & Sharp, D. J. (2015). Disconnection of network hubs and cognitive
impairment after traumatic brain injury. Brain, 138 , 1696 -- 1709.
Fiete, I. R., Senn, W., Wang, C. Z., & Hahnloser, R. H. R. (2010). Spike-time-dependent plasticity and heterosynaptic
competition organize networks to produce long scale-free sequences of neural activity. Neuron, 65 , 563 -- 576.
Freeman, W. J. (1994). Characterization of state transitions in spatially distributed, chaotic, nonlinear, dynamical systems
in cerebral cortex. Integr Physiol Behav Sci , 29 , 294 -- 306.
Gal´an, R. F. (2008). On how network architecture determines the dominant patterns of spontaneous neural activity. PLoS
One, 3 , e2148.
Garrett, D. D., Samanez-Larkin, G. R., MacDonald, S. W., Lindenberger, U., McIntosh, A. R., & Grady, C. L. (2013).
Moment-to-moment brain signal variability: a next frontier in human brain mapping? Neurosci Biobehav Rev , 37 ,
610 -- 624.
Gazzaniga, M. S. (Ed.) (2013). The cognitive neurosciences. MIT Press.
Goni, J., van den Heuvel, M. P., Avena-Koenigsberger, A., de Mendizabal, N. V., Betzel, R. F., Griffa, A., Hagmann, P.,
Corominas-Murtra, B., Thiran, J.-P., & Sporns, O. (2014). Resting-brain functional connectivity predicted by analytic
measures of network communication. Proceedings of the National Academy of Sciences, 111 , 833 -- 838.
Gu, S., Cieslak, M., Baird, B., Muldoon, S. F., Grafton, S. T., Pasqualetti, F., & Bassett, D. S. (2016). The energy landscape
of neurophysiological activity implicit in brain network structure. Submitted , .
Gu, S., Pasqualetti, F., Cieslak, M., Telesford, Q. K., Alfred, B. Y., Kahn, A. E., Medaglia, J. D., Vettel, J. M., Miller,
M. B., Grafton, S. T. et al. (2015). Controllability of structural brain networks. Nature communications, 6 .
Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C. J., Wedeen, V. J., & Sporns, O. (2008). Mapping the
structural core of human cerebral cortex. PLoS Biol , 6 , e159.
Hansen, E. C., Battaglia, D., Spiegler, A., Deco, G., & Jirsa, V. K. (2015). Functional connectivity dynamics: modeling the
switching behavior of the resting state. Neuroimage, 105 , 525 -- 535.
Hellyer, P. J., Leech, R., Ham, T. E., Bonnelle, V., & Sharp, D. J. (2013). Individual prediction of white matter injury
following traumatic brain injury. Ann Neurol , 73 , 489 -- 499.
Hermundstad, A. M., Bassett, D. S., Brown, K. S., Aminoff, E. M., Clewett, D., Freeman, S., Frithsen, A., Johnson, A.,
Tipper, C. M., Miller, M. B. et al. (2013a). Structural foundations of resting-state and task-based functional connectivity
in the human brain. Proceedings of the National Academy of Sciences, 110 , 6169 -- 6174.
Hermundstad, A. M., Bassett, D. S., Brown, K. S., Aminoff, E. M., Clewett, D., Freeman, S., Frithsen, A., Johnson, A.,
Tipper, C. M., Miller, M. B. et al. (2013b). Structural foundations of resting-state and task-based functional connectivity
in the human brain. Proceedings of the National Academy of Sciences, 110 , 6169 -- 6174.
Hermundstad, A. M., Brown, K. S., Bassett, D. S., Aminoff, E. M., Frithsen, A., Johnson, A., Tipper, C. M., Miller, M. B.,
Grafton, S. T., & Carlson, J. M. (2014a). Structurally-constrained relationships between cognitive states in the human
brain. PLoS Comput Biol , 10 , e1003591.
Hermundstad, A. M., Brown, K. S., Bassett, D. S., Aminoff, E. M., Frithsen, A., Johnson, A., Tipper, C. M., Miller, M. B.,
Grafton, S. T., & Carlson, J. M. (2014b). Structurally-constrained relationships between cognitive states in the human
brain. PLoS Comput Biol , 10 , e1003591.
Hermundstad, A. M., Brown, K. S., Bassett, D. S., & Carlson, J. M. (2011). Learning, memory, and the role of neural
network architecture. PLoS Comput Biol , 7 , e1002063.
16
Honey, C., Sporns, O., Cammoun, L., Gigandet, X., Thiran, J.-P., Meuli, R., & Hagmann, P. (2009). Predicting human
resting-state functional connectivity from structural connectivity. Proceedings of the National Academy of Sciences, 106 ,
2035 -- 2040.
Honey, C. J., Kotter, R., Breakspear, M., & Sporns, O. (2007). Network structure of cerebral cortex shapes functional
connectivity on multiple time scales. Proceedings of the National Academy of Sciences, 104 , 10240 -- 10245.
Jbabdi, S., Sotiropoulos, S. N., & Behrens, T. E. (2013). The topographic connectome. Current opinion in neurobiology, 23 ,
207 -- 215.
Johnson, V. E., Stewart, W., & Smith, D. H. (2013). Axonal pathology in traumatic brain injury. Exp Neurol , 246 , 35 -- 43.
Johnston, J. M., Vaishnavi, S. N., Smyth, M. D., Zhang, D., He, B. J., Zempel, J. M., Shimony, J. S., Snyder, A. Z., & Raichle,
M. E. (2008). Loss of resting interhemispheric functional connectivity after complete section of the corpus callosum. The
Journal of neuroscience, 28 , 6453 -- 6458.
Kalpinski, R. J., Williamson, M. L., Elliott, T. R., Berry, J. W., Underhill, A. T., & Fine, P. R. (2013). Modeling the
prospective relationships of impairment, injury severity, and participation to quality of life following traumatic brain
injury. Biomed Res Int, 2013 , 102570.
Kandel, E. R., Schwartz, J. H., Jessell, T. M. et al. (2000). Principles of neural science volume 4. McGraw-hill New York.
Khambhati, A., Davis, K., Lucas, T., Litt, B., & Bassett, D. S. (2016). Virtual cortical resection reveals push-pull network
control preceding seizure evolution. Submitted , .
Kinnunen, K. M., Greenwood, R., Powell, J. H., Leech, R., Hawkins, P. C., Bonnelle, V., Patel, M. C., Counsell, S. J., &
Sharp, D. J. (2011). White matter damage and cognitive impairment after traumatic brain injury. Brain, 134 , 449 -- 463.
Klimm, F., Bassett, D. S., Carlson, J. M., & Mucha, P. J. (2014). Resolving structural variability in network models and the
brain. PLOS Comput Biol , 10 , e1003491.
Laughlin, S. B. (2001). Efficiency and complexity in neural coding. Novartis Found Symp, 239 , 177 -- 187.
Laughlin, S. B., de Ruyter van Steveninck, R. R., & Anderson, J. C. (1998). The metabolic cost of neural information. Nat
Neurosci , 1 , 36 -- 41.
Lee, S., Ueno, M., & Yamashita, T. (2011). Axonal remodeling for motor recovery after traumatic brain injury requires
downregulation of gamma-aminobutyric acid signaling. Cell Death Dis, 2 , e133.
Lett, T. A., Voineskos, A. N., Kennedy, J. L., Levine, B., & Daskalakis, Z. J. (2014). Treating working memory deficits in
schizophrenia: a review of the neurobiology. Biol Psychiatry, 75 , 361 -- 370.
Levy, N., Horn, D., Meilijson, I., & Ruppin, E. (2001). Distributed synchrony in a cell assembly of spiking neurons. Neural
Netw , 14 , 815 -- 824.
Liu, Y.-Y., Slotine, J.-J., & Barab´asi, A.-L. (2011). Controllability of complex networks. Nature, 473 , 167 -- 173.
Luenberger, D. (1979). Introduction to dynamic systems: theory, models, and applications, .
Medaglia, J. D., Lynall, M. E., & Bassett, D. S. (2015). Cognitive network neuroscience. J Cogn Neurosci , 27 , 1471 -- 1491.
Meunier, D., Achard, S., Morcom, A., & Bullmore, E. (2009). Age-related changes in modular organization of human brain
functional networks. Neuroimage, 44 , 715 -- 723.
Misi´c, B., Betzel, R. F., Nematzadeh, A., Goni, J., Griffa, A., Hagmann, P., Flammini, A., Ahn, Y.-Y., & Sporns, O. (2015).
Cooperative and competitive spreading dynamics on the human connectome. Neuron, 86 , 1518 -- 1529.
Muldoon, S. F., Bridgeford, E. W., & Bassett, D. S. (2016a). Small-world propensity and weighted brain networks. Scientific
reports, 6 .
Muldoon, S. F., Pasqualetti, F., Gu, S., Cieslak, M., Grafton, S. T., Vettel, J. M., & Bassett, D. S. (2016b). Stimulation-based
control of dynamic brain networks. arXiv preprint arXiv:1601.00987 , .
Muller, F.-J., & Schuppert, A. (2011). Few inputs can reprogram biological networks. Nature, 478 , E4 -- E4.
Nenadovic, V., Hutchison, J. S., Dominguez, L. G., Otsubo, H., Gray, M. P., Sharma, R., Belkas, J., & Perez Velazquez,
J. L. (2008). Fluctuations in cortical synchronization in pediatric traumatic brain injury. J Neurotrauma, 25 , 615 -- 627.
17
Niven, J. E., & Laughlin, S. B. (2008). Energy limitation as a selective pressure on the evolution of sensory systems. J Exp
Biol , 211 , 1792 -- 1804.
Nudo, R. J. (2006). Mechanisms for recovery of motor function following cortical damage. Curr Opin Neurobiol , 16 , 638 -- 644.
O'Reilly, J. X., Croxson, P. L., Jbabdi, S., Sallet, J., Noonan, M. P., Mars, R. B., Browning, P. G., Wilson, C. R., Mitchell,
A. S., Miller, K. L. et al. (2013). Causal effect of disconnection lesions on interhemispheric functional connectivity in rhesus
monkeys. Proceedings of the National Academy of Sciences, 110 , 13982 -- 13987.
Pasqualetti, F., Zampieri, S., & Bullo, F. (2014). Controllability metrics, limitations and algorithms for complex networks.
Control of Network Systems, IEEE Transactions on, 1 , 40 -- 52.
Pestilli, F., Yeatman, J. D., Rokem, A., Kay, K. N., & Wandell, B. A. (2014). Evaluation and statistical inference for human
connectomes. Nature methods, 11 , 1058 -- 1063.
Posner, M. I., & Petersen, S. E. (1989). The attention system of the human brain. Technical Report DTIC Document.
Power, J. D., Cohen, A. L., Nelson, S. M., Wig, G. S., Barnes, K. A., Church, J. A., Vogel, A. C., Laumann, T. O., Miezin,
F. M., Schlaggar, B. L., & Petersen, S. E. (2011). Functional network organization of the human brain. Neuron, 72 ,
665 -- 678.
Radhu, N., Daskalakis, Z. J., Guglietti, C. L., Farzan, F., Barr, M. S., Arpin-Cribbie, C. A., Fitzgerald, P. B., & Ritvo, P.
(2012). Cognitive behavioral therapy-related increases in cortical inhibition in problematic perfectionists. Brain Stimul ,
5 , 44 -- 54.
Raichle, M. E. (2015). The brain's default mode network. Annu Rev Neurosci , 38 , 433 -- 447.
Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., & Shulman, G. L. (2001). A default mode
of brain function. Proc Natl Acad Sci U S A, 98 , 676 -- 682.
Raichle, M. E., & Snyder, A. Z. (2007). A default mode of brain function: a brief history of an evolving idea. Neuroimage,
37 , 1083 -- 1090.
Raja Beharelle, A., Kovacevic, N., McIntosh, A. R., & Levine, B. (2012). Brain signal variability relates to stability of
behavior after recovery from diffuse brain injury. Neuroimage, 60 , 1528 -- 1537.
Rajan, K., Harvey, C. D., & Tank, D. W. (2016). Recurrent network models of sequence generation and memory. Neuron,
90 , 128 -- 142.
Reveley, C., Seth, A. K., Pierpaoli, C., Silva, A. C., Yu, D., Saunders, R. C., Leopold, D. A., & Frank, Q. Y. (2015).
Superficial white matter fiber systems impede detection of long-range cortical connections in diffusion mr tractography.
Proceedings of the National Academy of Sciences, 112 , E2820 -- E2828.
Rusconi, E., Pinel, P., Eger, E., LeBihan, D., Thirion, B., Dehaene, S., & Kleinschmidt, A. (2009). A disconnection account
of Gerstmann syndrome: functional neuroanatomy evidence. Ann Neurol , 66 , 654 -- 662.
Rushworth, M. F., Behrens, T. E., & Johansen-Berg, H. (2006). Connection patterns distinguish 3 regions of human parietal
cortex. Cereb Cortex , 16 , 1418 -- 1430.
Salvador, R., Suckling, J., Schwarzbauer, C., & Bullmore, E. (2005). Undirected graphs of frequency-dependent functional
connectivity in whole brain networks. Philos Trans R Soc Lond B Biol Sci , 360 , 937 -- 946.
Sharp, D. J., Scott, G., & Leech, R. (2014). Network dysfunction after traumatic brain injury. Nat Rev Neurol , 10 , 156 -- 166.
Shenoy, K. V., Kaufman, M. T., Sahani, M., & Churchland, M. M. (2011). A dynamical systems view of motor preparation:
implications for neural prosthetic system design. Prog Brain Res, 192 , 33 -- 58.
Sidaros, A., Engberg, A. W., Sidaros, K., Liptrot, M. G., Herning, M., Petersen, P., Paulson, O. B., Jernigan, T. L., &
Rostrup, E. (2008). Diffusion tensor imaging during recovery from severe traumatic brain injury and relation to clinical
outcome: a longitudinal study.
Sizemore, A., Giusti, C., & Bassett, D. (2015). Classification of weighted networks through mesoscale homological features.
arXiv preprint arXiv:1512.06457 , .
Sommer, I. E., Slotema, C. W., Daskalakis, Z. J., Derks, E. M., Blom, J. D., & van der Gaag, M. (2012). The treatment of
hallucinations in schizophrenia spectrum disorders. Schizophr Bull , 38 , 704 -- 714.
18
Sridharan, D., Levitin, D. J., & Menon, V. (2008). A critical role for the right fronto-insular cortex in switching between
central-executive and default-mode networks. Proc Natl Acad Sci U S A, 105 , 12569 -- 125674.
Stam, C. J. (2014). Modern network science of neurological disorders. Nat Rev Neurosci , 15 , 683 -- 695.
Stephan, K. E., Friston, K. J., & Frith, C. D. (2009). Dysconnection in schizophrenia: from abnormal synaptic plasticity to
failures of self-monitoring. Schizophr Bull , 35 , 509 -- 527.
Szameitat, A. J., Schubert, T., & Muller, H. J. (2011). How to test for dual-task-specific effects in brain imaging studiesan
evaluation of potential analysis methods. Neuroimage, 54 , 1765 -- 1773.
Thomas, C., Frank, Q. Y., Irfanoglu, M. O., Modi, P., Saleem, K. S., Leopold, D. A., & Pierpaoli, C. (2014). Anatomical
accuracy of brain connections derived from diffusion mri tractography is inherently limited. Proceedings of the National
Academy of Sciences, 111 , 16574 -- 16579.
Tortella, G., Selingardi, P. M., Moreno, M. L., Veronezi, B. P., & Brunoni, A. R. (2014). Does non-invasive brain stimulation
improve cognition in major depressive disorder? a systematic review. CNS Neurol Disord Drug Targets, 13 , 1759 -- 1769.
Weiss, S. A., Bassett, D. S., Rubinstein, D., Holroyd, T., Apud, J., Dickinson, D., & Coppola, R. (2011). Functional brain
network characterization and adaptivity during task practice in healthy volunteers and people with schizophrenia. Frontiers
in human neuroscience, 5 , 81.
Yan, G., Ren, J., Lai, Y.-C., Lai, C.-H., & Li, B. (2012). Controlling complex networks: How much energy is needed?
Physical review letters, 108 , 218703.
Yeo, B. T., Krienen, F. M., Sepulcre, J., Sabuncu, M. R., Lashkari, D., Hollinshead, M., Roffman, J. L., Smoller, J. W.,
Zollei, L., Polimeni, J. R., Fischl, B., Liu, H., & Buckner, R. L. (2011). The organization of the human cerebral cortex
estimated by intrinsic functional connectivity. J Neurophysiol , 106 , 1125 -- 1165.
19
|
1706.10297 | 1 | 1706 | 2017-06-30T17:59:14 | Exploring the Human Connectome Topology in Group Studies | [
"q-bio.NC",
"cs.HC"
] | Visually comparing brain networks, or connectomes, is an essential task in the field of neuroscience. Especially relevant to the field of clinical neuroscience, group studies that examine differences between populations or changes over time within a population enable neuroscientists to reason about effective diagnoses and treatments for a range of neuropsychiatric disorders. In this paper, we specifically explore how visual analytics tools can be used to facilitate various clinical neuroscience tasks, in which observation and analysis of meaningful patterns in the connectome can support patient diagnosis and treatment. We conduct a survey of visualization tasks that enable clinical neuroscience activities, and further explore how existing connectome visualization tools support or fail to support these tasks. Based on our investigation of these tasks, we introduce a novel visualization tool, NeuroCave, to support group studies analyses. We discuss how our design decisions (the use of immersive visualization, the use of hierarchical clustering and dimensionality reduction techniques, and the choice of visual encodings) are motivated by these tasks. We evaluate NeuroCave through two use cases that illustrate the utility of interactive connectome visualization in clinical neuroscience contexts. In the first use case, we study sex differences using functional connectomes and discover hidden connectome patterns associated with well-known cognitive differences in spatial and verbal abilities. In the second use case, we show how the utility of visualizing the brain in different topological space coupled with clustering information can reveal the brain's intrinsic structure. | q-bio.NC | q-bio | Exploring the Human Connectome Topology in Group Studies
Johnson J.G. Keiriz, Liang Zhan, Morris Chukhman, Olu Ajilore, Alex D. Leow, and Angus G. Forbes
7
1
0
2
n
u
J
0
3
]
.
C
N
o
i
b
-
q
[
1
v
7
9
2
0
1
.
6
0
7
1
:
v
i
X
r
a
Fig. 1. A screen capture of the NeuroCave visualization tool showing the average resting state functional connectome of subjects
between the age of 20 and 30 years old for females in the clustering space (left) versus males in the anatomical space (right). Our tool
facilitates the simultaneous exploration of multiple connectome datasets in a variety of configurations, enabling researchers to make
meaningful comparisons between them and to reason about their differences.
Abstract-Visually comparing brain networks, or connectomes, is an essential task in the field of neuroscience. Especially relevant to
the field of clinical neuroscience, group studies that examine differences between populations or changes over time within a population
enable neuroscientists to reason about effective diagnoses and treatments for a range of neuropsychiatric disorders. In this paper, we
specifically explore how visual analytics tools can be used to facilitate various clinical neuroscience tasks, in which observation and
analysis of meaningful patterns in the connectome can support patient diagnosis and treatment. We conduct a survey of visualization
tasks that enable clinical neuroscience activities, and further explore how existing connectome visualization tools support or fail to
support these tasks. Based on our investigation of these tasks, we introduce a novel visualization tool, NeuroCave, to support group
studies analyses. We discuss how our design decisions (the use of immersive visualization, the use of hierarchical clustering and
dimensionality reduction techniques, and the choice of visual encodings) are motivated by these tasks. We evaluate NeuroCave
through two use cases that illustrate the utility of interactive connectome visualization in clinical neuroscience contexts. In the first use
case, we study sex differences using functional connectomes and discover hidden connectome patterns associated with well-known
cognitive differences in spatial and verbal abilities. In the second use case, we show how the utility of visualizing the brain in different
topological space coupled with clustering information can reveal the brain's intrinsic structure.
Index Terms-Brain networks, visual comparison, intrinsic topology, connectome visualization
1 INTRODUCTION
Over the past decade, the study of the human brain has progressed
through advancements in the magnetic resonance imaging (MRI) and
• Johnson J.G. Keiriz, Morris Chukhman, Angus G. Forbes, Olu Ajilore, and
Alex D. Leow are with University of Illinois at Chicago. E-mail: {jgadel2,
chukhman, aforbes}@uic.edu; {oajilore, aleow}@psych.uic.edu
• Liang Zhan is with University of Wisconsin–Stout. Email:
[email protected]
Manuscript received xx xxx. 201x; accepted xx xxx. 201x. Date of Publication
xx xxx. 201x; date of current version xx xxx. 201x. For information on
obtaining reprints of this article, please send e-mail to: [email protected].
Digital Object Identifier: xx.xxxx/TVCG.201x.xxxxxxx
other neuroimaging technology. Those advancements have allowed neu-
roscientists to non-invasively probe the brain's structural and functional
inter-regional connectivity and derive the human brain connectome [54].
High resolution MRI scans with submillimeter voxel size coupled with
advanced non-linear registration algorithms allows the creation of brain
label maps [36]. Those maps are created by registering brain MRI scans
with a pre-segmented atlas by a highly experienced neuroscientist. Dif-
fusion Weighted Imaging (DWI), also called diffusion MRI (dMRI),
allows us to reconstruct the brain's white matter fiber tracts through
a post-processing procedure called tractography. Counting the fibers
interconnecting each pair of regions derived from the label map gener-
ates a brain structural connectome. Functional MRI (fMRI) measures
the blood-oxygen-level dependent (BOLD) signal which represents the
activation level of the different brain regions due to the execution of
Fig. 2. Different methods of visualization in connectomics. Top row, left to right: connectome: adjacency matrix (self prepared), node-link [62],
edge-bundling [8], connectogram [32]. Bottom row, left to right: tractography: heatmap [41], linked views [33], abstraction [21].
specific tasks. The functional connectome is then generated by com-
puting the correlation coefficient of each pair of BOLD signal at the
different brain regions.
Connectomes are modeled as graphs which allows researchers to
use graph theory mathematics to analyze them. Graph metrics can
provide insights about the network topological properties such as: its
functional integration, which is the network ability to combine infor-
mation from its various parts; the clustering or segregation properties,
which quantify the existence of groups (clusters or modules); and the
network small-world properties, which describe the balance between
the functional integration and local clustering. A more extensive review
of different graph metrics used in the field of connectomics can be
found in [48]. Several important characteristics were derived for the
healthy brain network such as small-worldness [1, 49], clustering and
modularity [42], and rich-club configuration [58]. The functional con-
nectome enables neuroscientists to determine the default mode network
(DMN), or default state network, and was found to be the active brain
interacting regions when the subject is at wakeful rest and not involved
in a task [10].
Group studies are used to study the effects of the onset of neurolog-
ical and neuropsychological diseases, as well as age or injury on the
human brain, and how it alters the connectome. In a group study, con-
nectomes are generated for both healthy and disease populations. Graph
metrics are then computed for the resultant networks, and then statisti-
cally significant variations in the metrics values can be determined. For
a specific disease, researchers will often form a hypothesis based on
previous findings in the literature and then attempt to correlate it to the
alterations found in the collected graph data metrics [5]. Although the
use of graph metrics can provide a detailed aspect of the brain network,
it does not deliver potentially useful spatial information, which could
help the researcher to contextualize and interpret the results of a graph
analysis. This mandates an efficient visualization for the connectome
under study in order to help in the interpretation and diagnosis [45].
Moreover, visualization that supports effective comparison becomes
crucial in group studies in which alterations in the disease group are
more easily understood when analyzed in relation to healthy subjects.
The contributions in this paper are as follow:
• We delineate visual analytics tasks for clinical neuroscientists, espe-
cially as related to group studies.
• We introduce a taxonomy of these tasks, providing a comprehensive
survey of existing visualization software that supports connectome
analysis.
• We present NeuroCave, a novel, web-based, immersive visual ana-
lytics system that facilitates the visual inspection of structural and
functional connectome datasets. With NeuroCave, brain researchers
can interact with the connectome in any coordinate system or topo-
logical space, as well as group brain regions into different modules
on demand. A default side-by-side layout enables simultaneous, syn-
chronized manipulation in 3D that facilitates comparison tasks across
different subjects or diagnostic groups, or longitudinally within the
same subject.
• We provide the opportunity for researchers to engage in immersive
connectome analytics sessions while wearing portable VR headsets
(e.g., Oculus Rift) or on stereoscopic displays (e.g., CAVE systems
or 3D video walls).
• We reduce visual clutter inherent in dense networks (both in 2D and
in 3D views), by introducing a real-time, hardware accelerated edge
bundling algorithm, and combining it with a user need-based edge
selection strategy.
• We present two real-world use cases that demonstrate how our visu-
alization system can be used to identify patterns and support compar-
ison tasks in order to understand differences between connectome
datasets.
2 RELATED WORK
The term connectomics was first coined by Sporns et al. [54] to de-
scribe the wiring diagram of the anatomical connectivity of the human
brain. The interconnectivity found between parcellated brain regions as
described above is considered the macroscale connectome. Meso- and
micro- scale connectomes are at the local neuronal circuits and single
neuronal cells levels respectively, and they currently require an invasive
high resolution scans of dissected brains using microtome machinery.
Friston defined functional connectivity as the temporal correlation be-
tween spatially remote neurophysiological events [22]. This paper
focuses on visualization tasks relevant to macroscale functional and
structural connectomes, derived from functional and diffusion-weighted
MRI respectively.
2.1 Visualization in connectomics
Initial efforts into visualizing connectomics data have focused on trac-
tography data. The outcome of tractography computation is a set of
lines (fibers) spanning the brain white matter area representing the
axonal tracts of neuron cells. Volume rendering is used to overlay
streamlines on top of the anatomical brain image [17, 44, 51, 52, 66].
Often, a color code is used to identify the direction of the plotted fibers
(see Fig. 2). The main task of such visualizations is to allow the user
to select and explore specific fiber bundles such as the Corpus Callo-
sum interconnecting the left and right hemispheres. Due to the huge
number of generated fibers, on the order of 105, a range of techniques
are used to facilitate the navigation of the different fiber bundles. Some
techniques make use of 2D representations of predefined fiber tracts
that are synchronized together with a more conventional 3D visualiza-
tion, such as 2D hierarchical tree-like grapha [33], low-dimensional 2D
embedding representations [34], and 2D coronal, sagittal and axial pro-
jections of the fiber bundles [15] (see Fig. 2). Three dimensional visual
abstraction of the fiber bundles is used in the recent work of Everts et
al. [21] (see Fig. 2). Heatmaps are also used in some techniques, where
they are, for example, overlaid on top of anatomical isosurfaces [41]
(see Fig. 2). Segmented neurons and axons in mesoscale connectomes
are also visualized using streamlines [2, 6, 25].
Two dimensional adjacency matrices can provide a good overview
representation for large connectome datasets (see Fig. 2). Recent work
enhances the readability and flexibility of the adjacency matrix [3,
38], but have not been incorporated into software platforms that are
readily available for neuroscientists. Moreover, the use of adjacency
matrices can hinder users in performing some visual analysis tasks,
such as detecting graph alterations in group studies [23, 35]. The
most widely used representation for connectome visualization are 3D
node-link diagrams, in which nodes are positioned relative to their
corresponding anatomical locations, and in which links represent the
structural or functional connectivity between the nodes. Many examples
of visualization applications are presented in Table 1. An inherent
problem of node-link diagrams is the difficulty of effectively displaying
the total available links in dense networks. In dense networks, visual
clutter due to a preponderance of edge crossings can adversely affect
the accuracy and completion time for trend tasks (assessing change
in edge weight of a node's connections), connectivity tasks (assessing
the connectivity of common neighbors) and region identification tasks
(identifying the region with the most changes) [3]. One technique to
reduce such clutter is to combine visually compatible edges into bundles
according to a compatibility metric, hence the name edge bundling, first
introduced by Holten in [28]. Recently, some connectome visualization
projects have utilized edge bundling. Two dimensional edge bundling
is used by McGraw [40] and Yang et al. [63], while 3D edge bundling
is used in [8,9] for representing functional connectivity that shows high
levels of common interconnections (see Fig. 2). NeuroCave also utilizes
edge bundling in order to reduce visual clutter in large connectome
datasets that may have 2500 or more interconnected nodes (see Fig. 8).
The connectogram [32], a recent connectome visualization technique,
is a node-link diagram in a circular layout (see Fig. 8). Names of brain
regions are positioned along the perimeter of a circle, which is split into
two halves according to their hemispheric affiliation. Each hemisphere
is further divided into the different brain lobes. The inner sphere
contains multiple colored rings, each representing a specific metric.
The regions are interconnected within the circle using curved lines. This
technique prevents some of the clutter that is found in other network
visualizations. However, it is harder to correlate anatomical structures
with connectivity using a circular diagram like the connectogram, as
the multiple rings may make it difficult for a user to make sense of the
data [12]. Table 1 contains a list of different visualization tools in the
field of connectomics and their capabilities. Although previous efforts
have surveyed visualization in connectomics, such as Margulies et
al. [39] and Pfister et al. [45], our survey is the first to comprehensively
explore recent connectomics visualization software in terms of their
ability to support group studies analysis tasks. Moreover, we catalog
these software tools by the primary connectome dataset types they
support.
2.2 Connectome Analysis Tasks
Extending the user task definitions presented in Alper et al. [3] through
a thorough investigation of connectome visualization projects and sur-
veying neuroscientists that work with group studies, we have identified
visual analytics tasks for clinical neuroscientists:
T1 Identify regions responsible for specific cognitive functions and
study their interactions with other regions.
T2 Compare individual networks to the mean or group average con-
nectome. In group studies, individual variations as well as joint
network characteristics are studied in order to identify commonali-
ties or differences.
T3 Identify the effect of structural connectivity on the functional activity
of the brain. Comparing both structural and functional at the same
time to reveal the complex mappings between them [11, 30, 31].
T4 Identify individual or group changes occurring on the structural
or functional connectivities due to the onset of disease [50] or
aging [20] as well as gender differences. Moreover, researchers can
assess its restoration in drug studies [55].
T5 Identify dynamic changes in structural and functional connectivity
over time for both within subject and between subjects [16].
T6 Identify structural re-routing occurring after a brain injury or dam-
age with rehabilitation training and its effect on functional connec-
tivity. Similarly, in case of neurosurgery, it is important to identify
affected structural pathways and predict the corresponding loss in
motor and cognitive abilities after the procedure.
Although each of the the visualization software tools listed in Table 1
may partially address many of these tasks, none provides a visualization
that can directly facilitate T2–T5, involving various types of compari-
son between datasets, since they all lack the ability to simultaneously
load and synchronize a comparative visualization of multiple connec-
tomes. The user needs to open multiple instances of the application
and usually requires multiple monitors in order to visually be able
to compare the structural and functional connectomes of the same
subject or two subjects belonging to different groups. Clearly, all
user actions will not be synchronized which makes it even harder for
assessing visual differences. Applications implemented in scripting
languages such as R and MATLAB provide the user with the flexibility
to customize views. However, this requires additional efforts as well
as programming expertise. Introducing a novel side-by-side layout,
NeuroCave helps neuroscientists and researchers efficiently execute
T1–T5, which involve comparative analyses, and to simultaneously
spot changes occurring within and across subjects. It is important to
state that NeuroCave does not target tractography-related usages, such
as the sixth task (T6), which, although an important area of connec-
tomics visualization, is usually a requirement by neurosurgeons rather
than clinical neuroscientists (who are the intended audience for our
visualization system).
2.3 Web-based VR
While most commonly used visualization tools are dedicated desktop
applications, web-based implementations, such as Slice:Drop [24] or
BrainBrowser [53], free the user from being attached to a specific
operating system [46]. NeuroCave is also a web-based application
and runs in any modern browser. The use of stereoscopic techniques
can provide a more immersive way to explore brain imaging data [47],
Hanel et al. find that healthcare professionals perceive the increased
dimensionality provided by stereoscopy as beneficial for understanding
depth in the displayed scenery [26]. Moreover, Ware and Mitchell find
that the use of stereographic visualizations reduces the perception error
rate in graph perception for large graphs with more than 500 nodes [60],
and Alper et al. [4] observed that, when coupled with highlighting,
stereoscopic representations of 3D graphs outperformed their non-
immersive counterpart. NeuroCave lets the user move between desktop
and VR environments for interactively exploring 3D connectomes.
2.4 Connectome Topology
The techniques of linear and nonlinear dimensionality reduction have
been widely used in the field of brain imaging, specially for processing
Table 1. Neuroimaging connectomic software. This table indicates the visual analysis tasks (Section 2.2) supported by each of these software tools.
Software (V)olume / (S)urface / (G)raph Supported Tasks Description Website Structural DTI Studio S T1, T6 Generates and visualizes tractography and allows user-defined ROI. https://www.mristudio.org/ TrackVis V T1, T6 Generates and visualizes tractography and allows user-defined ROI. http://www.trackvis.org/ Camino V/S T1, T6 Generates and visualizes tractography and allows user-defined ROI. http://camino.cs.ucl.ac.uk/ DoDTI V T1, T6 MATLAB-based tool to generate and visualize tractography. http://neuroimage.yonsei.ac.kr/dodti/ ExploreDTI V/S T1, T6 Generates tractography and allows user-defined ROI. http://www.exploredti.com/ BrainVISA V/S T1. T6 Generates and visualizes tractography data with Python interface. http://brainvisa.info/web/index.html OpenWalnut V/S T1, T6 Visualization of tractography data. http://www.openwalnut.org/ FiberNavigator2 V/S T1, T6 Visualization of tractography data. https://code.google.com/p/fibernavigator2/ TractoR V T1, T6 Plugin to the R programming language to perform tractography and visualize results. http://www.tractor-mri.org.uk/ DSI Studio V/S/G T1, T2, T3, T4, T6 Generates and visualizes tractography and allows user-defined ROI. Generates and visualizes connectomes. http://dsi-studio.labsolver.org/ Brainnetome V/S/G T1, T6 Generates and visualizes tractography and allows user-defined ROI. Generates connectomes. http://diffusion.brainnetome.org/en/latest/ NPerspective S T1 Google-maps like 2D representation of the neural pathways in three projections: Sagittal, Coronal and Axial. http://graphics.cs.brown.edu/research/sciviz/newbraininteraction/ 3D Slicer V/S T1, T6 Generate and visualize tractography data. www.slicer.org Connectome Explorer V/G T1 Visualizes mesoscale connectome generated using MRI scans of dissected brain slices. http://people.seas.harvard.edu/~jbeyer/connectome_explorer.html Functional InstaCorr (SUMA/AFNI) V/S T1, T2, T3, T4 Analysis and visualization of fMRI data. http://afni.nimh.nih.gov/afni VidView S/G T1, T2, T3, T4 Viewer for fMRI connectomic data (connexel). https://github.com/NeuroanatomyAndConnectivity/vidview Brainbundler S/G T1, T2, T3, T4 Visualizes functional connectivity as glyphs on the brain surface. Applies edge bundling to connectome data. https://github.com/NeuroanatomyAndConnectivity/brainbundler/ Fubraconnec S/G T1, T2, T3, T4 Visualization of functional connectome using synchronized anatomy, node-link and circular views. https://code.google.com/p/fubraconnex/ VAMCA V/S T1, T2, T3, T4 MATLAB-based toolbox for analysis and visualization. http://www.nitrc.org/projects/vamca REST G T1, T2, T3, T4 MATLAB-based resting-state functional connectivity analysis and visualization tool. http://restfmri.net/forum/REST CONN V/S/G T1, T2, T3, T4 MATLAB-based for fMRI connectome analysis and visualization. https://www.nitrc.org/projects/conn/ Multimodal Connectome Viewer V/S/G T1, T2, T3, T4 Anatomic and functional connectivity graph visualization. Analysis enabled through Python interface. http://cmtk.org/viewer/ Connectome Workbench V/S/G T1, T2, T3, T4 Investigates data acquired through Human Connectome Project. http://www.humanconnectome.org/software/ BrainNet Viewer V/S/G T1 MATLAB-based brain surface and graph visualization. http://www.nitrc.org/projects/bnv/ Brain Connectivity Toolbox G T2, T3, T4, T5 MATLAB-based list of graph theory analysis functions with basic matrix visualization. https://sites.google.com/site/bctnet/visualization Visual Connectome S/G T1, T2, T3, T4 MATLAB-based structural and functional connectome analysis and visualization. http://code.google.com/p/visualconnectome/ MNET G T1, T2, T3, T4 MATLAB-based dMRI and fMRI processing tool that generates and visualizes tractography data as well as structural and functional connectomes. http://neuroimage.yonsei.ac.kr/mnet/ BRAVIZ V/S T1, T2, T3, T4, T6 MATLAB-based. Overlays dMRI metrics on top of tractography data. http://diego0020.github.io/braviz/ Online BrainBrowser V/S T1, T2 Visualizes structural connectomic data. https://brainbrowser.cbrain.mcgill.ca/ XTK V/S NA Framework for building Web-based neuroimaging viewers. https://github.com/xtk/X Slicedrop V/S T1 Based on XTK, visualizes tractography data. http://slicedrop.com/ Other Connectogram G T1 Standardized circular schematic of multimodal data. http://circos.ca/tutorials/lessons/recipes/cortical_maps/ geometrical properties of platonic solids. In brief, a platonic solid is a
regular, convex polyhedron constructed by congruent regular polygonal
faces with the same number of faces meeting at each vertex. Five
platonic solids exist: tetrahedron, cube, octahedron, dodecahedron and
icosahedron, with four, six, eight, twelve, and twenty faces, respectively.
Based on how many clusters are generated, a suitable platonic solid
is chosen such that its number of faces is greater than the number of
these clusters. The glyphs of each cluster are then equally distributed
(according to the sunflower algorithm [59]) over half of the spherical
cap covering the corresponding face of of a platonic solid embedded in
a sphere (see Fig. 3). This enables the user to "enter" into the geometry
(i.e. into the "NeuroCave") via one of the unpopulated face(s), provid-
ing a more immersive experience of the data.
Node visualization. We utilize two different glyphs (spheres and
cubes) to differentiate between left and right hemisphere affiliation.
Nodes can be colored according to lobar or modular information. Con-
trolling nodal transparency is also possible according to their color
scheme. Three modes exist: opaque, semi-transparent or totally trans-
parent (invisible). Glyphs size is also user adjustable.
Edge Visualization. As mentioned in Section 2, edge clutter can be a
problem in the visualization of dense node-link diagrams. Our approach
to overcome this problem is based on two steps. First, we provide the
option to hide all edges by default (i.e., to show only the nodes). The
user can then select a root node, and all connected edges stemming from
this node will be displayed. Second, to overcome the clutter occurring
from edge crossings, we use the force directed edge bundling (FDEB)
algorithm to group edges going in the same direction [29]. FDEB is an
iterative algorithm that consists of cycles. In each cycle, we subdivide
an edge into a specified number of points (we chose 6 cycles, and we
double the number of points each cycle, ending up with 64 subdivision
points plus the two original points of the edge). After the subdivision,
we iteratively move each subdivision point in an update step to a new
position determined by modeling the forces among the points. Each
point is affected by the sum of spring and electrostatic forces (Fs and
Fe). For an edge i, at a subdivision point pi j, Fs is defined as:
Fs = kp((cid:107)pi( j−1) − pi j(cid:107) +(cid:107)pi j − pi( j+1)(cid:107))
(1)
where kp is a spring constant and pi( j−1) and pi( j+1) are the neighbor-
ing points of pi j, the (cid:107).(cid:107) is the length. The electrostatic force Fe is
defined as
Fe = ∑
m∈E
1
(cid:107)pi j − pm j(cid:107)
(2)
where E is a set of compatible edges with the current edge, pm j is the
corresponding subdivision point to point pi j in the edges belonging to
the set E. The compatibility metrics used to define the set E are defined
in details in Holten et al. [29]. Each cycle contains a prespecified
number of update steps.
Our original Javascript implementation of NeuroCave turned out to
be too slow for large numbers of edges (more than 500), preventing
a real-time experience. Therefore, we used (and enhanced) a WebGL
texture-based implementation suggested by Wu et al. [61]. The FDEB
algorithm is parallelizable since the division and update operations are
performed on each point independently. The texture-based method
stores the subdivision points in a 2D GPU texture, where each row
represents the 3D coordinates of points belonging to the same edge.
Since write operations are unavailable to GPU textures in WebGL, a
ping-pong algorithm 4 is used to write results to a framebuffer object
(FBO). Two shaders are utilized: the first performs the subdivision
operation, and the second executes the update steps. In Wu et al.'s
implementation, the limitation on the texture size limits is not addressed.
GPU textures possess a limitation on their sizes. Hence, a large number
of edges can not be fit in one texture. We enhanced the algorithm using
tiling. Since the total number of points of each edge after all cycles
will be known ahead (64 subdivision + 2 end points), we tile the edges
when the maximum number of possible rows per texture is achieved.
4https://www.khronos.org/opengl/wiki/Memory_Model.
Fig. 3. The figure shows the proposed method of distributing clustered
nodes. The nodes are distributed on half the spherical cap covering a
platonic solid face. The chosen platonic solid must have a number of
faces that is equal to or larger than the number of clusters (for 8 clusters
a dodecahedron was chosen).
fMRI data [13, 57]. Recently, nonlinear dimensionality reduction tech-
niques, such as isomap and t-distributed stochastic neighbor embedding
(t-SNE), have been applied onto the human connectome in order to find
the brain's intrinsic geometry [64]. Using such embedding, new 3D
topologies for the brain were found other than the regular anatomical
one. Such topologies depends on the internodal graph shortest path
length computed using Dijkstra's algorithm [19]. Nodes possessing
efficient connectivity to the rest of the brain were found to be positioned
near the origin of those new topologies inferring their importance. Neu-
roCave facilitates finding topological alterations in group studies. This
helps researchers in studying the connectome topology of healthy sub-
jects, such as the DMN, and comparing it with diseased groups. In
DMN, important hub nodes, such as the precuneus (see Fig. 8), possess
a strong connectivity with the rest of the brain and are located near the
center of the brain's intrinsic geometry [64]. A deviation from such
topology in disease groups need to be studied for illnesses associated
with DMN alterations such as Alzheimer's disease which is known to
disrupt it [10].
3 THE NeuroCave SOFTWARE SYSTEM
NeuroCave is implemented as a web-based application. The Javascript
WebGL-based graphics library three.js 1 was used for the 3D rendering
and visualization functionalities. It runs on all major web browsers,
and is thus platform independent. The default view is formed of two
side-by-side 3D rendering views (see Fig. 1). Each view allows the
interactive visualization of a connectome as a node-link diagram.
Group Visualization. The application loads subjects data from a study
folder. The folder should contains all adjacency matrices as well as
the corresponding topological and clustering information of the sub-
jects within the study. An indexing file states the subject ID and its
corresponding data files. Each study requires a predefined Atlas that
provides numerical labels and their corresponding anatomical names
to each node. NeuroCave directly supports three Atlases: FSL-based
which consists of 82 labels of FreeSurfer 2, brain hierarchical Atlas
(BHA) made of 2514 labels [18] and Harvard-Oxford Atlas made of
177 labels.3 However, additional Atlases can be easily created follow-
ing the pre-existing ones as example.
Topology Visualization. The application layouts the nodes according
to the provided topological information. Topologies can be the anatom-
ical positioning or an applied transformation in some abstract space
as explained in Section 2. The available topologies are automatically
identified by the application.
Clustering Visualization. Clustering information is input as a vector
of integer values. Each value represents a different module or cluster.
Since there are no prespecified positions for clusters, we exploit the
1http://threejs.org.
2https://surfer.nmr.mgh.harvard.edu/.
3http://neuro.imm.dtu.dk/wiki/Harvard-Oxford_Atlas.
Fig. 4. Edge bundling (left) versus no edge bundling (right) in clustering space displaying four clusters. Notice the clutter caused by the edges in the
right panel versus the edge bundling case. Also, notice the edge color gradient that depends on the nodes color and their relative strengths: the
orange sphere possesses a greater strength compared to the blue nodes it is connected to, while it possesses a lesser strength compared to the
green nodes it is connected to. NeuroCave enables users to toggle between edge bundling and edge coloring modes on demand.
Our texture-based implementation can bundle the closet 1000 edges to
the selected node at interactive rates on a desktop computer (Intel Core
i7, 3.4 GHz CPU, Nvidia GTX 1070 GPU card and 32GB RAM).
Each edge is colored according to a two color gradient chosen ac-
cording to the interconnected nodes colors. The gradient is skewed
towards the node possessing the higher nodal strength (the sum of
weights of links connected to the node) value (see Fig. 4). For an edge
made of n points the color, Cout, at point i is
Cout = C2 + (C1 −C2)R
(3)
2))/P1P2
R = (r2(p2 − 0.5) + r(0.5− p2
(4)
where C1 and C2 are the nodes, r = i/(n − 1), P1 = S1/S1 + S2),
P2 = 1− P1, S1 and S2 are the nodes strengths. This allows the user to
recognize the strength of the selected node with respect to its intercon-
nected neighbors which helps identifying strong and important nodes
and hubs as well as highlight the reason for modular changes when
occurred in group studies as will be shown in Section 4.
Virtual Reality. In addition to the standard 3D manipulations of pan-
ning, rotating, and zooming, NeuroCave supports the Oculus Rift and
the Oculus Rift Touch controllers. The Touch controllers are a pair of
VR input devices that track each hand, enabling an effective gesture-
based manipulation of the VR environment. The user selects the pre-
view area to be explored in VR and then uses the thumbsticks of the
Touch devices to navigate the visualized connectome through panning,
rotating, and zooming. Nodal selection is enabled via a two step proce-
dure: first, pressing the grip button lets the user point at and highlight a
node; second, pressing the index button selects the highlighted node.
4 RESULTS AND DISCUSSION
In this section, we present two use cases that demonstrate how clinical
neuroscientists can perform the tasks described in Section 2.
4.1 Use Case 1
Our first use case investigates the sex-specific resting-state functional
connectomes in the F1000 repository, a large 986 subjects publicly-
available resting-state fMRI connectome dataset.5 The following post-
processing was performed: first, to eliminate the potential confounding
effect of age, we only included subjects between 20 to 30 years old
(319 females at 23.25 ± 2.26 years of age and 233 males at 23.19
± 2.35). We then constructed hierarchical modularity in the form of
4-level dendrograms (16 clusters at the most local level) [65]. Previ-
ously, rigorous permutation testing established sex differences already
statistically significant at the first level (2 clusters; p = 0.0378) for the
following six regions: posterior part of left and right frontal pole (near
the fronto-temporal junction), left precentral gyrus, right hippocampus,
right superior temporal gyrus posterior division (sTG-p) and the right
inferior frontal gyrus pars opercularis (POP) (for more detail about the
statistical analysis, refer to [65]).
In order to interpret these sex differences, male and female group-
averaged connectomes were visualized in NeuroCave. We highlighted
the connectivity of the aforementioned six regions in both the anatomi-
cal and clustering (1st hierarchical level) spaces and applied a threshold
to restrict edges to values with absolute correlation values for fMRI
BOLD signals larger than 0.1 (see Fig. 5, Fig. 6). From the clustering
space, affiliation patterns for the right hippocampus and sTG-p differ in
that in women they are clustered with other frontal and temporal ROIs
as part of the default mode network (DMN), while in males they are
clustered with other parietal and occipital non-DMN ROIs. Notably,
the affiliation differences are opposite for the left and right frontal poles,
the left precentral gyrus and the right POP, such that they are more
clustered along with several non-DMN parietal and occipital ROIs in
the average female connectome (see Fig. 5).
The left and right fronto-temporal junctions are part of the larger
language system, with the right POP (functionally coupled with its
homologous area on the left that forms the Broca's language area)
linked to the processing of semantic information [27], and the superior
temporal gyrus involved in the comprehension of language (as well as
in the perception of emotions in facial stimuli [7]). Thus, the observed
connectivity differences are related to well-known sex differences in
language and emotion/affect processing, as well as differences in self-
referential/autobiographical information retrieval. By contrast, the
5https://www.nitrc.org/projects/fcon_1000/.
Fig. 5. Left and right frontal poles (red and green rings encircling the sphere and cube glyphs), left precentral gyrus (yellow ring) and the right inferior
frontal gyrus (blue ring) nodes selected in anatomical (left panel) and clustering (right panel) spaces for female and male average connectomes. The
color code represents the two clusters in the anatomical space, while in the clustering space, it represents the lobe affiliation. Notice the change in
group affiliation of the four nodes between the female and male connectomes (left panel).
Fig. 6. Right hippocampus (red ring) and superior temporal gyrus (green ring) nodes selected in anatomical (left panel) and clustering (right panel)
spaces for female and male average connectomes. Color code is similar to the above figure. Notice the change in group affiliation of the four nodes
between the female and male connectomes (left panel).
Fig. 7. Right hippocampus (left panel) and left hippocampus (right panel) nodes selected in the anatomical space for female and male average
connectomes. Note the tendency in females for both left and right hippocampus to be functionally connected to the contralateral frontal lobe (but not
in males). Note that such differences are not previously known to our clinicians and neuroscientists. Here, the color code represents connectome's
hierarchical modularity, represented as a dendrogram, at the most global level (2 modules).
Fig. 8. Connectivity emerging from the anterior (red ring) and posterior (yellow ring) parts of the precuneus in anatomical (left panel) and isomap
"intrinsic" (middle panel) spaces. Right panel: Connectivity emerging from the posterior part of the precuneus visualized in the intrinsic space. The
color code represents the modular structure of the connectome consisting of 4 modules. Note that the orange community contains the default mode
network. The inset plot shows the residual geodesics for the first 10 dimensions of the isomap dimensionality reduction algorithm.
hippocampus is known to play an important role in the formation
of new memory, retrieval of declarative long-term memory, and the
management and processing of spatial and spatiotemporal working
memory. The modular affinity between the right hippocampus and
other non-DMN regions in the parietal and occipital lobes in males may
thus be related to their established advantage in spatial tasks, including
spatial visualization, perception and mental rotation [37] (visual system
is heavily composed of the occipital lobe responsible for first-level
visual processing, while part of the parietal lobe is instrumental for
visuospatial skills).
Lastly, closer inspections of the hippocampal functional connec-
tivity between women and men further revealed differences that are
not known previously to our clinicians (Fig 7). Indeed, in the female
connectome there is a strong tendency for both left and right hippocam-
pus to functionally communicate with the contralateral frontal lobe, a
phenomenon that is not visually present in the male connectome. The
discovery of such subtle differences is dependent upon an iteratively
explorative visualization process, only possible with the comprehensive
suite of tools implemented in NeuroCave.
This use case demonstrates the effectiveness of NeuroCave is sup-
porting tasks T3 and T4, enabling neuroscientists to better understand
neurological gender differences in connectome datasets and to observe
how these differences relate to various psychological studies.
4.2 Use Case 2
Our second use case explores a resting-state fMRI high-resolution
dataset consisting of 2514 regions-of-interest publicly available at NI-
TRC.6 Most functional connectome studies treat the negative corre-
lation entries (anticorrelated BOLD signals) in the network by either
taking the absolute value or clamping to zero (heuristically chosen)
which affects all network metrics' computations. Instead, we applied a
recently proposed probabilistic framework which more rigorously ac-
counts for negative edges [65]. For an N×N functional connectome, the
framework estimates the probability that an edge, ei j, is positive or neg-
ative using the connectomes of a group of subjects (i and j varies from
1 to N). The edge positivity EPi j and edge negativity ENi j form a com-
plementary pair, since EPi j + ENi j = 1, and thus can be jointly coded
using the angle of a unit-length vector: θi j = arctan((cid:112)ENi j/EPi j)
which varies from 0 to 90 degrees. Using dissimilarity graph embed-
ding, each node, i, is then embedded in an n-dimensional space at the
coordinates (θi1,θi2, ...θiN )T . In order to reduce the dimensionality of
the resultant topology, we applied the nonlinear dimensionality reduc-
tion isomap algorithm [56] that aims to preserve geodesic distances in
6http://www.nitrc.org/frs/?group_id=964.
a lower-dimensional space (i.e., the "intrinsic space"). Separately, we
determined the community structure by maximizing the Q-modularity
metric [43].
From the inset plot in Fig. 8, it is clear that the degree of isometric
embedding levels off after the fifth dimension, thus suggesting that the
intrinsic topology of the resting-state functional connectome has a di-
mension of five, a novel finding that merits further research. To enable
3D visualization of the transformed topology, we retained the first three
dimensions of isomap and visualized the modular structure of the brain
in both the anatomical space as well as this novel intrinsic topological
space (Fig 8). As expected, the nodes assigned to the same community
are positioned close to one another in this novel topology. To illustrate
how neuroscientists explore this complex topological space and gain
further insight into the brain, we selected two nodes that belong to the
anterior and posterior part of the precuneus. Although the nodes are
anatomically close to each other, they are known to be functionally
distinct (thus belong to different modules). Indeed, the anterior part of
the precuneus is an important region of the DMN known to be respon-
sible for self-referential imagery (thinking about self) and is involved
in autobiographical tasks and self-consciousness, thus activated during
"resting consciousness" [14]. As we can see in Fig. 8, in this intrinsic
space the anterior part of the precuneus, while assigned to the orange
module (DMN), exhibits diverse connections with various regions of
the brain: the sensori-motor module (blue) and the frontoparietal exec-
utive or task-positive system (red). By contrast, the posterior precuneus
is part of the visual system (green) and has a relatively restricted pattern
(compared with its more anterior counterpart) of connectivity with the
rest of the brain. Notably, such connectivity differences only become
visually apparent when neuroscientists visualize in this novel space.
This use case highlights how our visualization system is effective at
supporting tasks T1 and T2, enabling neuroscientists to explore high
density connectomics data comprising a large number of ROIs in order
to identify and further understand the specific functionality of different
brain regions. Moreover, the side-by-side visualization enables users to
reason about the relationships between the anatomical and the intrinsic
topology, facilitating further insight into how the same brain region can
take part in different tasks.
5 CONCLUSION AND FUTURE WORK
In this paper, we presented NeuroCave, a novel VR-compatible visual-
ization environment for exploring and analyzing the human connectome.
We also performed a task taxonomy for neuroscientists and researchers
in the field of connectomics. NeuroCave facilitates comparison tasks of
two connectomes, an activity that clinical neuroscientists often require
for the analysis of group studies. Our system includes various visu-
alization enhancements, such as edge-bundling to reduce edge clutter
and edge gradient coloring to help identify nodal strengths. More-
over, it enables users to explore connectomic data in an immersive VR
environment, which can help to improve their perception of the data
under study. The GPU is extensively used in NeuroCave for hardware
acceleration of both computation and visualization tasks. We demon-
strated the effectiveness of the system using two real-world use cases
in which neuroscientists were able to use NeuroCave to effectively
perform research and analysis tasks.
However, many challenges in visualizing the brain connectome
remain. Future work will adapt our system to support temporally-
varying dynamic connectome datasets (T5). We also will continue to
work with clinical neuroscientists to integrate additional analytic tools
to assist with diagnosis and treatment of patients. Finally, our side-by-
side visualization framework can be extended to other domains that
require group study analysis, such as cancer biology, where researchers
are interested in comparing experiment vs. control networks or disease
vs. normal gene networks in order to better understanding cancer and
other disease genomes.
REFERENCES
[1] S. Achard, R. Salvador, B. Whitcher, J. Suckling, and E. Bullmore. A
resilient, low-frequency, small-world human brain functional network
with highly connected association cortical hubs. Journal of Neuroscience,
26(1):63–72, 2006.
[2] A. K. Ai-Awami, J. Beyer, D. Haehn, N. Kasthuri, J. W. Lichtman, H. Pfis-
ter, and M. Hadwiger. NeuroBlocks–Visual tracking of segmentation
and proofreading for large connectomics projects. IEEE Transactions on
Visualization and Computer Graphics, 22(1):738–746, 2016.
[3] B. Alper, B. Bach, N. Henry Riche, T. Isenberg, and J.-D. Fekete. Weighted
graph comparison techniques for brain connectivity analysis. In Proceed-
ings of the SIGCHI Conference on Human Factors in Computing Systems,
pp. 483–492, 2013.
[4] B. Alper, T. Hollerer, J. Kuchera-Morin, and A. G. Forbes. Stereoscopic
highlighting: 2d graph visualization on stereo displays. IEEE Transactions
on Visualization and Computer Graphics, 17(12):2325–2333, 2011.
[5] D. A. Angulo, C. Schneider, J. H. Oliver, N. Charpak, and J. T. Hernandez.
A multi-facetted visual analytics tool for exploratory analysis of human
brain and function datasets. Frontiers in Neuroinformatics, 10, 2016.
[6] J. Beyer, A. Al-Awami, N. Kasthuri, J. W. Lichtman, H. Pfister, and
M. Hadwiger. ConnectomeExplorer: Query-guided visual analysis of
large volumetric neuroscience data. IEEE Transactions on Visualization
and Computer Graphics, 19(12):2868–2877, 2013.
[7] E. D. Bigler, S. Mortensen, E. S. Neeley, S. Ozonoff, L. Krasny, M. John-
son, J. Lu, S. L. Provencal, W. McMahon, and J. E. Lainhart. Superior
temporal gyrus, language function, and autism. Developmental Neuropsy-
chology, 31(2):217–238, 2007.
[8] J. Bottger, A. Schafer, G. Lohmann, A. Villringer, and D. S. Margulies.
Three-dimensional mean-shift edge bundling for the visualization of func-
tional connectivity in the brain. IEEE Transactions on Visualization and
Computer Graphics, 20(3):471–480, 2014.
[9] J. Bottger, R. Schurade, E. Jakobsen, A. Schaefer, and D. S. Margulies.
Connexel visualization: A software implementation of glyphs and edge-
bundling for dense connectivity data using brainGL. Frontiers in Neuro-
science, 8:15, 2014.
[10] R. L. Buckner, J. R. Andrews-Hanna, and D. L. Schacter. The brain's
default network. Annals of the New York Academy of Sciences, 1124(1):1–
38, 2008.
[11] E. Bullmore and O. Sporns. Complex brain networks: Graph theoretical
analysis of structural and functional systems. Nature Reviews Neuro-
science, 10(3):186–198, 2009.
[12] M. Burch and D. Weiskopf. On the benefits and drawbacks of radial
diagrams. In Handbook of Human-Centric Visualization, pp. 429–451.
2014.
[13] V. D. Calhoun, J. Liu, and T. Adalı. A review of group ica for fmri data
and ica for joint inference of imaging, genetic, and erp data. Neuroimage,
45(1):S163–S172, 2009.
[14] A. E. Cavanna. The precuneus and consciousness. CNS Spectrums,
12(07):545–552, 2007.
[15] W. Chen, Z. Ding, S. Zhang, A. MacKay-Brandt, S. Correia, H. Qu,
J. A. Crow, D. F. Tate, Z. Yan, and Q. Peng. A novel interface for
interactive exploration of DTI fibers. IEEE Transactions on Visualization
and Computer Graphics, 15(6):1433–1440, 2009.
[16] N. A. Crossley, T. R. Marques, H. Taylor, C. Chaddock, F. DellAcqua,
A. A. Reinders, V. Mondelli, M. DiForti, A. Simmons, A. S. David, et al.
Connectomic correlates of response to treatment in first-episode psychosis.
Brain, p. aww297, 2016.
[17] M. Da Silva, S. Zhang, C. Demiralp, and D. H. Laidlaw. Visualizing
diffusion tensor volume differences. In IEEE Visualization 01 Proceedings,
Work in Progress, 2001.
[18] I. Diez, P. Bonifazi, I. n. Escudero, B. Mateos, M. A. Munoz, S. Stramaglia,
and J. M. Cortes. A novel brain partition highlights the modular skeleton
shared by structure and function. Scientific Reports, 5, 2015.
[19] E. W. Dijkstra. A note on two problems in connexion with graphs. Nu-
merische Mathematik, 1(1):269–271, 1959.
[20] N. U. Dosenbach, B. Nardos, A. L. Cohen, D. A. Fair, J. D. Power,
J. A. Church, S. M. Nelson, G. S. Wig, A. C. Vogel, C. N. Lessov-
Schlaggar, et al. Prediction of individual brain maturity using fMRI.
Science, 329(5997):1358–1361, 2010.
[21] M. H. Everts, E. Begue, H. Bekker, J. B. Roerdink, and T. Isenberg.
Exploration of the brains white matter structure through visual abstrac-
tion and multi-scale local fiber tract contraction. IEEE Transactions on
Visualization and Computer Graphics, 21(7):808–821, 2015.
[22] K. J. Friston. Functional and effective connectivity in neuroimaging: A
synthesis. Human Brain Mapping, 2(1-2):56–78, 1994.
[23] M. Ghoniem, J.-D. Fekete, and P. Castagliola. On the readability of graphs
using node-link and matrix-based representations: A controlled experiment
and statistical analysis. Information Visualization, 4(2):114–135, 2005.
[24] D. Haehn. Slice:Drop: Collaborative medical imaging in the browser. In
ACM SIGGRAPH 2013 Computer Animation Festival. ACM, 2013.
[25] D. Haehn, S. Knowles-Barley, M. Roberts, J. Beyer, N. Kasthuri, J. W.
Lichtman, and H. Pfister. Design and evaluation of interactive proofreading
tools for connectomics. IEEE transactions on visualization and computer
graphics, 20(12):2466–2475, 2014.
[26] C. Hanel, P. Pieperhoff, B. Hentschel, K. Amunts, and T. Kuhlen. Interac-
tive 3D visualization of structural changes in the brain of a person with
corticobasal syndrome. Frontiers in Neuroinformatics, 8(42), 2014.
[27] S. Heim, K. Alter, A. K. Ischebeck, K. Amunts, S. B. Eickhoff,
H. Mohlberg, K. Zilles, D. Y. von Cramon, and A. D. Friederici. The role
of the left Brodmann's areas 44 and 45 in reading words and pseudowords.
Cognitive Brain Research, 25(3):982–993, 2005.
[28] D. Holten. Hierarchical edge bundles: Visualization of adjacency relations
in hierarchical data. IEEE Transactions on Visualization and Computer
Graphics, 12(5):741–748, 2006.
[29] D. Holten and J. J. Van Wijk. Force-directed edge bundling for graph
visualization. In Proceedings of the 11th Eurographics / IEEE - VGTC
Conference on Visualization, EuroVis'09, pp. 983–998. Wiley Online
Library, 2009.
[30] C. Honey, O. Sporns, L. Cammoun, X. Gigandet, J.-P. Thiran, R. Meuli,
and P. Hagmann. Predicting human resting-state functional connectivity
from structural connectivity. Proceedings of the National Academy of
Sciences, 106(6):2035–2040, 2009.
[31] C. J. Honey, R. Kotter, M. Breakspear, and O. Sporns. Network structure
of cerebral cortex shapes functional connectivity on multiple time scales.
Proceedings of the National Academy of Sciences, 104(24):10240–10245,
2007.
[32] A. Irimia, M. C. Chambers, C. M. Torgerson, and J. D. Van Horn. Circular
representation of human cortical networks for subject and population-level
connectomic visualization. Neuroimage, 60(2):1340–1351, 2012.
[33] R. Jianu, C. Demiralp, and D. Laidlaw. Exploring 3d dti fiber tracts
with linked 2d representations. IEEE Transactions on Visualization and
Computer Graphics, 15(6):1449–1456, 2009.
[34] R. Jianu, C. Demiralp, and D. H. Laidlaw. Exploring brain connectivity
with two-dimensional neural maps. IEEE Transactions on Visualization
and Computer Graphics, 18(6):978–987, 2012.
[35] R. Keller, C. M. Eckert, and P. J. Clarkson. Matrices or node-link diagrams:
which visual representation is better for visualising connectivity models?
Information Visualization, 5(1):62–76, 2006.
[36] A. Klein, J. Andersson, B. A. Ardekani, J. Ashburner, B. Avants, M.-
C. Chiang, G. E. Christensen, D. L. Collins, J. Gee, P. Hellier, et al.
Evaluation of 14 nonlinear deformation algorithms applied to human brain
MRI registration. Neuroimage, 46(3):786–802, 2009.
[37] M. C. Linn and A. C. Petersen. Emergence and characterization of sex
differences in spatial ability: A meta-analysis. Child Development, pp.
1479–1498, 1985.
approach for interactively visualizing large graphs. In Big Data (Big Data),
2015 IEEE International Conference on, pp. 2501–2508, 2015.
[62] M. Xia, J. Wang, and Y. He. Brainnet viewer: a network visualization tool
for human brain connectomics. PloS one, 8(7):e68910, 2013.
[63] X. Yang, L. Shi, M. Daianu, H. Tong, Q. Liu, and P. Thompson. Blockwise
human brain network visual comparison using NodeTrix representation.
IEEE Transactions on Visualization and Computer Graphics, 23(1):181–
190, 2017.
[64] A. Q. Ye, O. A. Ajilore, G. Conte, J. GadElkarim, G. Thomas-Ramos,
L. Zhan, S. Yang, A. Kumar, R. L. Magin, A. G. Forbes, and A. D. Leow.
The intrinsic geometry of the human brain connectome. Brain Informatics,
2(4):197–210, 2015.
[65] L. Zhan, L. M. Jenkins, O. Wolfson, J. Jonaris, K. N. GadElkarim, P. M.
Thompson, O. Ajilore, K. Moo, and A. L. Chung. The importance of
being negative: A serious treatment of non-trivial edges in brain functional
connectome. arXiv preprint arXiv:1609.01384v2, 2016.
[66] S. Zhang, C. Demiralp, and D. H. Laidlaw. Visualizing diffusion tensor
mr images using streamtubes and streamsurfaces. IEEE Transactions on
Visualization and Computer Graphics, 9(4):454–462, 2003.
[38] C. Ma, R. V. Kenyon, A. G. Forbes, T. Berger-Wolf, B. J. Slater, and
D. A. Llano. Visualizing dynamic brain networks using an animated
dual-representation. In Proceedings of the Eurographics Conference on
Visualization (EuroVis), pp. 73–77, 2015.
[39] D. S. Margulies, J. Bottger, A. Watanabe, and K. J. Gorgolewski. Visualiz-
ing the human connectome. NeuroImage, 80:445–461, 2013.
[40] T. McGraw. Graph-based visualization of neuronal connectivity using
matrix block partitioning and edge bundling. In International Symposium
on Visual Computing, pp. 3–13, 2015.
[41] T. McGraw and M. Nadar. Stochastic DT-MRI connectivity mapping on
the GPU. IEEE Transactions on Visualization and Computer Graphics,
13(6), 2007.
[42] D. Meunier, R. Lambiotte, and E. T. Bullmore. Modular and hierarchically
modular organization of brain networks. Frontiers in Neuroscience, 4:200,
2010.
[43] M. E. Newman. Modularity and community structure in networks. Pro-
ceedings of the National Academy of Sciences, 103(23):8577–8582, 2006.
[44] V. Petrovic, J. Fallon, and F. Kuester. Visualizing whole-brain DTI tractog-
raphy with GPU-based tuboids and LoD management. IEEE Transactions
on Visualization and Computer Graphics, 13(6):1488–1495, 2007.
[45] H. Pfister, V. Kaynig, C. P. Botha, S. Bruckner, V. J. Dercksen, H.-C.
Hege, and J. B. Roerdink. Visualization in connectomics. In Scientific
Visualization, pp. 221–245. 2014.
[46] C. Pieloth, J. M. Pizarro, T. Knosche, B. Maess, and M. Fuchs. An online
system for neuroelectromagnetic source imaging. In IEEE International
Conference on Intelligent Data Acquisition and Advanced Computing
Systems (IDAACS), vol. 1, pp. 270–274, 2013.
[47] G. M. Rojas, M. G´alvez, N. Vega Potler, R. C. Craddock, D. S. Margulies,
F. X. Castellanos, and M. P. Milham. Stereoscopic three-dimensional
visualization applied to multimodal brain images: Clinical applications
and a functional connectivity atlas. Frontiers in Neuroscience, 8:328,
2014.
[48] M. Rubinov and O. Sporns. Complex network measures of brain connec-
tivity: Uses and interpretations. Neuroimage, 52(3):1059–1069, 2010.
[49] R. Salvador, J. Suckling, M. R. Coleman, J. D. Pickard, D. Menon, and
E. Bullmore. Neurophysiological architecture of functional magnetic
resonance images of human brain. Cerebral Cortex, 15(9):1332–1342,
2005.
[50] E. J. Sanz-Arigita, M. M. Schoonheim, J. S. Damoiseaux, S. A. Rom-
bouts, E. Maris, F. Barkhof, P. Scheltens, and C. J. Stam. Loss of small-
worldnetworks in alzheimer's disease: graph analysis of fmri resting-state
functional connectivity. PloS One, 5(11):e13788, 2010.
[51] T. Schultz, H. Theisel, and H.-P. Seidel. Topological visualization of brain
diffusion mri data. IEEE Transactions on Visualization and Computer
Graphics, 13(6), 2007.
[52] A. Sherbondy, D. Akers, R. Mackenzie, R. Dougherty, and B. Wandell.
Exploration of the brains white matter pathways with dynamic queries.
IEEE Transactions on Visualization and Computer Graphics, 11(4):419–
430, 2005.
[53] T. Sherif, N. Kassis, M.- ´E. Rousseau, R. Adalat, and A. C. Evans. Brain-
browser: distributed, web-based neurological data visualization. Frontiers
in Neuroinformatics, 8:89, 2015.
[54] O. Sporns, G. Tononi, and R. Kotter. The human connectome: A structural
description of the human brain. PLoS Computational Biology, 1(4):e42,
2005.
[55] R. Tadayonnejad, O. Ajilore, B. J. Mickey, N. A. Crane, D. T. Hsu, A. Ku-
mar, J.-K. Zubieta, and S. A. Langenecker. Pharmacological modulation
of pulvinar resting-state regional oscillations and network dynamics in
major depression. Psychiatry Research: Neuroimaging, 252:10–18, 2016.
[56] J. B. Tenenbaum, V. De Silva, and J. C. Langford. A global ge-
ometric framework for nonlinear dimensionality reduction. Science,
290(5500):2319–2323, 2000.
[57] B. Thirion and O. Faugeras. Nonlinear dimension reduction of fMRI
data: The Laplacian embedding approach. In Proceedings of the IEEE
International Symposium on Biomedical Imaging, pp. 372–375, 2004.
[58] M. P. Van Den Heuvel and O. Sporns. Rich-club organization of the human
connectome. Journal of Neuroscience, 31(44):15775–15786, 2011.
[59] H. Vogel. A better way to construct the sunflower head. Mathematical
biosciences, 44(3-4):179–189, 1979.
[60] C. Ware and P. Mitchell. Visualizing graphs in three dimensions. ACM
Transactions on Applied Perception (TAP), 5(1):2, 2008.
[61] J. Wu, L. Yu, and H. Yu. Texture-based edge bundling: A web-based
|
1812.06234 | 1 | 1812 | 2018-12-15T05:15:40 | Integrating Artificial Intelligence with Real-time Intracranial EEG Monitoring to Automate Interictal Identification of Seizure Onset Zones in Focal Epilepsy | [
"q-bio.NC",
"cs.AI",
"q-bio.QM"
] | An ability to map seizure-generating brain tissue, i.e., the seizure onset zone (SOZ), without recording actual seizures could reduce the duration of invasive EEG monitoring for patients with drug-resistant epilepsy. A widely-adopted practice in the literature is to compare the incidence (events/time) of putative pathological electrophysiological biomarkers associated with epileptic brain tissue with the SOZ determined from spontaneous seizures recorded with intracranial EEG, primarily using a single biomarker. Clinical translation of the previous efforts suffers from their inability to generalize across multiple patients because of (a) the inter-patient variability and (b) the temporal variability in the epileptogenic activity. Here, we report an artificial intelligence-based approach for combining multiple interictal electrophysiological biomarkers and their temporal characteristics as a way of accounting for the above barriers and show that it can reliably identify seizure onset zones in a study cohort of 82 patients who underwent evaluation for drug-resistant epilepsy. Our investigation provides evidence that utilizing the complementary information provided by multiple electrophysiological biomarkers and their temporal characteristics can significantly improve the localization potential compared to previously published single-biomarker incidence-based approaches, resulting in an average area under ROC curve (AUC) value of 0.73 in a cohort of 82 patients. Our results also suggest that recording durations between ninety minutes and two hours are sufficient to localize SOZs with accuracies that may prove clinically relevant. The successful validation of our approach on a large cohort of 82 patients warrants future investigation on the feasibility of utilizing intra-operative EEG monitoring and artificial intelligence to localize epileptogenic brain tissue. | q-bio.NC | q-bio | Integrating Artificial Intelligence with Real-time Intracranial EEG Monitoring to Automate
Interictal Identification of Seizure Onset Zones in Focal Epilepsy
Authors
Yogatheesan Varatharajah1, Brent Berry2,5, Jan Cimbalnik3, Vaclav Kremen2,3, Jamie Van Gompel4,
Matt Stead2, Benjamin Brinkmann2,5, Ravishankar Iyer1, and Gregory Worrell2
1 Electrical and Computer Engineering, University of Illinois, Urbana, IL 61801, USA.
2 Mayo Systems Electrophysiology Laboratory, Department of Neurology, Mayo Clinic, Rochester, MN 55905, USA
3 Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague, 166 36 Prague 6, Czech Republic
4 Department of Neurosurgery, Mayo Clinic, Rochester MN, 55905, USA
5 Department of Physiology & Biomedical Engineering, Mayo Clinic, Rochester MN, 55905, USA
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Abstract
An ability to map seizure-generating brain tissue, i.e., the seizure onset zone (SOZ), without recording
actual seizures could reduce the duration of invasive EEG monitoring for patients with drug-resistant
epilepsy. A widely-adopted practice in the literature is to compare the incidence (events/time) of
putative pathological electrophysiological biomarkers associated with epileptic brain tissue with the
SOZ determined from spontaneous seizures recorded with intracranial EEG, primarily using a single
biomarker. Clinical translation of the previous efforts suffers from their inability to generalize across
multiple patients because of (a) the inter-patient variability and (b) the temporal variability in the
epileptogenic activity. Here, we report an artificial intelligence-based approach for combining multiple
interictal electrophysiological biomarkers and their temporal characteristics as a way of accounting for
the above barriers and show that it can reliably identify seizure onset zones in a study cohort of 82
patients who underwent evaluation for drug-resistant epilepsy. Our investigation provides evidence
that utilizing the complementary information provided by multiple electrophysiological biomarkers
and their temporal characteristics can significantly improve the localization potential compared to
previously published single-biomarker incidence-based approaches, resulting in an average area under
ROC curve (AUC) value of 0.73 in a cohort of 82 patients. Our results also suggest that recording
durations between ninety minutes and two hours are sufficient to localize SOZs with accuracies that
may prove clinically relevant. The successful validation of our approach on a large cohort of 82
patients warrants future investigation on the feasibility of utilizing intra-operative EEG monitoring and
artificial intelligence to localize epileptogenic brain tissue. Broadly, our study demonstrates the use of
artificial intelligence coupled with careful feature engineering in augmenting clinical decision making.
Keywords
Seizure Onset Zone, Epilepsy Surgery, Artificial Intelligence in Neurological Applications, High-
Frequency Oscillation, Interictal Epileptiform Discharge, Phase-Amplitude Coupling, and Support
Vector Machine.
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Introduction
Epilepsy is one of the most prevalent and disabling neurologic diseases. It is characterized by the
occurrence of unprovoked seizures and affects ~1% of the world's population [Leonardi 2002]. Many
patients with epilepsy achieve seizure control with medication, but approximately one-third of people
with epilepsy continue to have seizures despite taking medications [Kwan 2000]. In such cases, one
treatment option is surgical resection of the brain tissue responsible for seizures, but this option
depends critically on accurate localization of the pathological brain tissue, which is referred to as the
seizure onset zone (SOZ). Clinical SOZ localization requires implanting of electrodes for intracranial
EEG (iEEG) that is recorded over several days to allow sufficient time for spontaneous seizures to
occur [Lüders 2006]. The electrodes that are in the SOZ are identified based on visual inspection of
the iEEG captured at the time of seizures, and some tissue around these electrodes is removed during
a surgical procedure. Despite being the current gold standard for mapping of the epileptic brain in a
clinical setting, this manual procedure is time-consuming, costly, and associated with potential
morbidity [Van Gompel 2008, Wellmer 2012]. Recently, the use of interictal (non-seizure) iEEG data
for the identification of SOZs and for the possibility of replacing multi-day ICU monitoring to record
habitual seizures has received notable interest [van 't Klooster 2015].
A common practice undertaken in the literature investigating interictal SOZ localization is to
compare the incidence rates (events/time) of putative pathological electrophysiological events
associated with epileptic brain tissue (known as electrophysiological biomarkers of epilepsy) detected
in iEEG recorded from individual electrodes against the gold-standard SOZ electrodes determined
from spontaneous seizures. Among the potential electrophysiological biomarkers, high-frequency
oscillations (HFOs) [Bragin 1999, Bragin 2002, Worrell 2011, Zijlmans 2012] and interictal
epileptiform discharges (IEDs) [Staley 2011] have been the most widely investigated. Phase-
amplitude coupling (PAC) and other forms of cross-frequency coupling (CFC) have more recently
been investigated as promising clinical biomarkers for epilepsy [Papdelis 2016]. HFOs are local field
potentials that reflect short-term synchronization of neuronal activity, and they are widely believed to
be clinically useful for localization of epileptic brain [Bragin 1999, Worrell 2004, Jirsch 2006, Staba
2002]. Furthermore, there is an extensive literature investigating IEDs as interictal markers of seizure
onset zones, but it has met with limited success [Lüders 2006, Marsh 2010, Hauf 2012]. PAC (a
measure of cross-frequency coupling) [Jensen 2007] as an adjunct to ictal (seizure) biomarkers was
shown to be useful for SOZ localization [Edakawa 2016], and more recently, PAC has been evaluated
as an interictal marker for determining SOZs [Amiri 2016]. Most the existing studies have utilized a
simple counting of the above biomarkers (detected either manually or using software) in fixed
durations to classify the electrodes that are in the SOZ [Cimbalnik 2017]. Although some recent
approaches have utilized clustering methods [Liu 2016] and dimensionality reduction methods [Weiss
2016] as preprocessing steps in identifying pathologic interictal HFOs, the determination of SOZs was
still performed using simple counting of HFOs. Furthermore, these approaches have predominantly
utilized a single biomarker to identify SOZs and have not considered the inter-patient variability nor
the temporal dynamics of the epileptic activity [Cimbalnik 2017]. As a result, they have not been able
to generalize across multiple patients and their overall accuracies have been insufficient to bring them
into clinical practice [Holler 2015, Nonoda 2016, Sinha 2017].
The reasons for inter-patient variability might include electrode placement, false-positive
detections of biomarkers, signal artifacts, the varied etiology of focal epilepsy, or the fact that some
biomarkers are incident in both physiologic and pathologic states [Matsumoto 2013, Ben-Ari 2007,
Cimbalnik 2016]. Thus, utilizing a single biomarker to identify SOZs of patients with potentially
heterogeneous epileptogenic mechanisms may result in unsatisfactory accuracy for some individuals.
We hypothesize that it may be possible to reduce inter-patient variability by combining the
complementary values contained within different electrophysiological biomarkers and thereby
improve SOZ localization potential. However, despite the growing interest in each of the above
biomarkers, the extent to which they provide independent predictive value for epileptogenic tissue
localization remains unclear. From a signal-processing perspective, IED represents a relatively distinct
electrophysiological phenomenon compared to HFO and PAC. However, temporal correlations of
HFO events and PAC may be observed when short HFOs co-occur with IEDs [Weiss 2016]. Apart
from
that each biomarker will constitute specific
that specific
instance,
it
is possible
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
electrophysiological information about the epileptogenicity of brain tissue and might add predictive
value when used in unison with other biomarkers. However, the potential clinical utility of combining
electrophysiological biomarkers has received relatively little investigation [Gnatkovsky 2014].
In addition, there is evidence that behavioral states play a role in altering the temporal patterns
of epileptiform activity in the brain [Staba 2002, Worrell 2008, Amiri 2016]. As a result, the
occurrence of the biomarkers exhibits temporally varying rates when long EEG recordings with mixed
behavioral states are considered [Pearce 2013]. Hence, the common practice of simply counting HFOs
or IEDs for a fixed duration and using an average rate to determine the SOZ is likely suboptimal. We
recently proposed a temporal filtering-based unsupervised approach to utilize temporal characteristics
of spectral power features to determine SOZs interictally [Varatharajah 2017]. However, more
sophisticated models are needed to effectively utilize multiple electrophysiological biomarkers and
their temporal characteristics to accurately determine SOZs. Modern artificial intelligence (AI) based
methods facilitate (a) the ability to learn high-dimensional decision functions from labeled training
data and (b) the flexibility to define customized features representing domain knowledge [Russell
1995]. We believe that these properties can be useful in harnessing multiple electrophysiological
biomarkers and their temporal characteristics for the interictal classification of SOZs. However, AI-
based approaches have been underexplored in the SOZ classification literature, mainly because of the
unavailability of large-scale iEEG datasets collected during epilepsy-surgery evaluation. Fortunately,
the availability of continuous iEEG recordings collected from a large cohort of 82 patients and
approximately 5000 electrodes gives us a unique opportunity to assess the potential utility of AI-based
approaches in this study. Although this dataset is still not large enough to automatically extract class-
specific electrophysiological patterns using deep-learning approaches, comparable performance can
be realized on this dataset using careful feature engineering and appropriate model selection.
In that context, the aim of this study is to develop an AI-based analytic framework that utilizes
multiple interictal electrophysiological biomarkers (e.g., HFO, IED, and PAC) and their temporal
characteristics for interictal electrode classification and mapping of SOZs. To that end, we developed a
support vector machine (SVM) based classification model utilizing customized features (based on the
above biomarkers) extracted from 120-minute interictal iEEG recordings of 82 patients with drug-
resistant epilepsy to interictally identify the electrodes representing their seizure onset zones. This
approach achieved an average AUC (area under ROC curve) of 0.73 when the HFO, IED, and PAC
biomarkers were used jointly, which is 14% better than the AUC achieved by a conventional
biomarker incidence-based approach using all three biomarkers, and 4 -- 13% better than that of an
SVM-based model that used any one of the biomarkers. This result indicates that exploiting the
temporal variations in biomarker activity can improve localization of epileptic brain and that the
biomarkers utilized in this study have complementary predictive values. Our analysis of individual
patients reveals that the AUCs improve or remain unchanged for more than 65% of the patients when
the composite, rather than any single biomarker, is utilized, supporting the hypothesis that combining
multiple biomarkers can provide more generalizability than can individual biomarkers. Development
of this technology also enabled us to develop an understanding of the recording durations required for
interictal localization of SOZs. By analyzing iEEG segments of different durations (10 -- 120 minutes),
we show that longer iEEG segments provide better accuracy than do short segments, and that the
improvements become statistically insignificant for durations beyond 90 minutes. These promising
results warrant further investigation on the feasibility of using intra-operative mapping to localize
epileptic brain.
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Methods
Experimental Setup
Data used in this study were recorded from patients undergoing evaluation for epilepsy surgery at the
Mayo Clinic, Rochester, MN. The Mayo Clinic Institutional Review Board approved this study, and
all subjects provided informed consent. Subjects underwent intracranial depth electrode implantation
as part of their evaluation for epilepsy surgery whenever noninvasive studies could not adequately
localize the SOZ. To provide an unbiased dataset for analysis, we took 2 hours of continuous iEEG
data (during some period between 12:00 and 3:00 AM) on the night after surgery.
Subjects
Data from 82 subjects (48 males and 34 females, with an average age of 31) with focal epilepsy were
investigated by post hoc analysis. All subjects were implanted with intracranial depth arrays, grids,
and/or strips; see supplementary Table 1 for details. Subjects underwent multiple days of iEEG and
video monitoring to record their habitual seizures.
Electrodes and Anatomical Localization
Depth electrode arrays (from AD-Tech Medical Inc., Racine, WI) were 4- and 8-contact electrode
arrays consisting of a 1.3-mm-diameter polyurethane shaft with platinum/iridium
(Pt/Ir)
macroelectrode contacts. Each contact was 2.3 mm long, with 10-mm or 5-mm center-to-center
spacing (with a surface area of 9.4 mm2 and an impedance of 200 -- 500 Ohms). Grid and strip
electrodes had 2.5-mm-diameter exposed surfaces and 1-cm center-to-center spacing of adjacent
contacts. Anatomical localization of electrodes was achieved using post-implant CT data co-registered
to the patient's MRI using normalized mutual information [Ashburner 2008]. Electrode coordinates
were then automatically labeled by the SPM Anatomy toolbox, with an estimated accuracy of 0.5 mm
[Tzourio-Mazoyer, et al., 2002].
Signal Recordings
All iEEG data were acquired with a common reference using a Neuralynx Cheetah electrophysiology
system. (It had a 9-kHz antialiasing analog filter, and was digitized at a 32-kHz sampling rate, filtered
by a low-pass, zero-phase-shift, 1-kHz, low-pass Bartlett-Hanning window, and down-sampled to 5
kHz.)
Clinical SOZ Localization
The SOZ electrodes and time of seizure onset were determined by visually identifying the electrodes
with the earliest iEEG seizure discharges. Seizure onset times and zones were determined by visual
identification of a clear electrographic seizure discharge, followed by looking back at earlier iEEG
recordings for the earliest electroencephalographic change contiguously associated with the visually
definitive seizure discharge. The same approach has been used previously to identify neocortical SOZs
[Worrell 2004] and medial temporal lobe seizures [Worrell 2008]. The identified SOZ electrodes were
used as the gold standard to test and validate our analyses.
Data
Continuous 2-hour interictal segments of iEEG data, sufficiently separated from seizures, were chosen
for all 82 patients to represent a monitoring duration that could be achieved during surgery. A total of
4966 electrodes were implanted across the 82 subjects, and 911 of them were identified to be in SOZs
via ictal localization performed by clinical epileptologists caring for the patients.
Data Preprocessing
Prior to analysis, continuous scalp and intracranial EEG recordings were reviewed using a custom
MATLAB viewer [Brinkmann 2009]. Electrode channels and time segments containing significant
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
artifacts or seizures were not included in subsequent analyses. All iEEG recordings were filtered to
remove 60-Hz power-line artifacts.
Overall Analytic Scheme
Selected 2-hour iEEG recordings were divided into non-overlapping 3-second epochs. A 3-second
epoch length was chosen to accommodate at least a single transient electrophysiologic event (in the
form of a PAC, HFO, or IED) that could be associated with the SOZ. The HFO [Cimbalnik 2017],
IED [Barkmeier 2012], and PAC [Amiri 2016] biomarkers were extracted using previously published
detectors to measure their presence in each 3-second epoch. Then, a clustering procedure was
performed to assign a binary observation of normal or abnormal to each channel. This procedure was
performed separately for each patient, and the biomarker measures extracted in a 3-second recording
of all the channels of a specific patient were considered. These channels were clustered into two
groups based on their similarities with respect to each biomarker, and the cluster with the larger
average biomarker rate was considered the abnormal cluster. This step was performed so that
biomarker detections that had strong magnitudes and showed strong spatial correlation were retained,
and electrodes with noisy detections were minimized. At the end of that procedure, every 3-second
recording of a channel was associated with three binary values (one for each biomarker) representing
the presence of the HFO, IED, and PAC biomarkers. We refer to those binary values as observations.
Since there are 2400 3-second epochs in a 2-hour period, the total number of observations made in a
channel was 3 * 2400 = 7200. Since that number of features is relatively large compared to the
number of channels available in our study, we reduced the number of observations by applying a
straightforward summation approach. Binary observations made within a 10-minute window (200
epochs) were counted to arrive at a measure of the local rate of the biomarker incidence for each 10-
minute window of the 2-hour recording. Although the 10-minute window length may appear to have
been chosen arbitrarily, it is large enough to have less noisy local biomarker incidence rates and yet
not so large as to mask the temporal variations in interictal biomarker activity. This method reduces
the number of observations for a channel to 36 (3 local biomarker rates × 12 windows). These
observations, made across a 2-hour period of a channel, were used to infer whether that channel
belonged to an SOZ under a supervised learning setting using a support vector machine (SVM)
classifier. The whole process is illustrated as a flow diagram in Figure 1.
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Figure 1: The overall analytic scheme of the SOZ detection algorithm utilized in this study. A 2-
hour data segment is analyzed for each patient. PAC, HFO, and IED biomarkers are extracted in 3-
second epochs, and a clustering method is used to group channels based on similarities with respect to
the biomarkers. These groupings are converted to binary (0, 1) observations and counted within a 10-
minute window to obtain local biomarker incidence rates. These local biomarker incidence rates for
the three biomarkers within all the 10-minute windows of a channel's 2-hour recording are utilized as
the features of that channel in a machine learning setting. A support vector machine (SVM) classifier,
which was trained and tested using labeled training data, is used to predict whether an electrode is in
an SOZ.
Detection of Interictal Electrophysiological Biomarkers
The PAC measure was calculated by correlating instantaneous phase of the low-frequency signal with
the corresponding amplitude of a high-frequency signal for a given set of low- and high-frequency
bands (Figure 2a). In this implementation, low- and high-frequency contents in the signal were
extracted using MORLET wavelet filters, and all frequency bands were correlated against all others to
create a so-called PAC-gram (Figure 2b). Based on the observed high PAC content and the existing
literature [Weiss 2015], 0.1 -- 30 Hz was chosen as the low-frequency (modulating) signal, and 65 -- 115
Hz was chosen as the high-frequency (modulated) signal in the rest of the analysis. HFOs were
detected using a Hilbert transform-based method [Kucewicz & Berry 2015, Pail 2017, Cimbalnik
2016]. The data segments were bandpass-filtered for every 1-Hz band step from 50 to 500 Hz. Then,
the filtered-data frequency bands were normalized (z-score), and the segments in which the signal
amplitudes were three standard deviations above the mean for a duration of one complete cycle of a
respective high frequency (in 65 -- 500 Hz) were marked as HFOs (see Figure 2c) [Matsumoto 2013].
IEDs were extracted using a previously validated spike-detection algorithm [Barkmeier 2012]. A
detection threshold of four standard deviations (of differential amplitude) around the mean was used
to mark IEDs in this algorithm (see Figure 2d). The HFO, IED, and PAC detected events were stored
in a database.
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Figure 2: Phase amplitude coupling (PAC), high-frequency oscillations (HFO), and interictal
epileptiform discharge (IED) detection. (A) Detailed illustration of the PAC feature extraction
algorithm. Low (0.1 -- 30 Hz) and high (65 -- 115 Hz) frequency components are filtered out from the
raw signal. The phase of the slow wave is correlated with the high-frequency amplitude envelope to
measure coupling. (B) A PAC-gram representing the average interictal PAC measured between
different frequency bands. Highlighted portion indicates the low- and high-frequency bands utilized in
the rest of our analysis. (C) Pictorial illustration of HFO detection. Oscillations that have an amplitude
of three standard deviations above the mean and lasting for more than one complete cycle in low-
gamma (30 -- 60 Hz), high-gamma (60 -- 100 Hz), and ripple (100 -- 150 Hz) bands are detected. (D) An
illustration of detected IEDs. Differential amplitude is standardized, and a threshold of four standard
deviations around the mean was used to mark IEDs.
Prediction of SOZ Electrodes Using a Support Vector Machine Classifier
The biomarkers extracted from a 2-hour recording of a channel were converted to a 36-dimensional
feature vector as shown in Figure 1. The features represent the local biomarker incidence rates (with a
separate rate for each biomarker) within each 10-minute window of the 2-hour recording. These
features were standardized to eliminate any differences in scale. We took two different approaches to
perform cross-validation. First, we performed 10-fold cross validation. The dataset, including
standardized features of 4966 electrodes (including 911 SOZ electrodes), was divided into a 60%
training set and a 40% testing set, keeping the same proportion of SOZ and NSOZ (non-SOZ)
electrodes in both sets. Second, we performed leave-1-out cross-validation. For every subject in the
dataset, we used the data from the rest of the subjects as the training data and the respective subject's
data as the testing data. In each of the cross-validation iteration, training and testing datasets were
generated using one of the cross-validation approaches. An SVM classifier was trained on the training
set, whose hyper-parameters (described below) we optimized by performing a grid search with a
tenfold cross-validation within the training set. The classifier trained on the best-performing hyper-
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
parameters was used to predict the labels of the channels in the testing set. By comparing those
predictions against the ground-truth labels of the testing set channels, we calculated the metrics of
model fitness. This process was repeated 10 times with different combinations of training and testing
sets to obtain metrics of generalized performance. This process is illustrated in Figure 3.
Figure 3: A flow diagram illustrating the prediction framework. The input is the whole dataset,
including the 36 features extracted from 82 subjects (4966 channels) and their gold-standard labels
assigned by clinical epileptologists. (A) 10-fold cross-validation: first, the dataset is shuffled to
randomize the order of channels in it. After randomization, the dataset is partitioned into two sets, a
60% training set and a 40% testing set, keeping the same proportion of NSOZ and SOZ channels in
both sets. (C) Inner CV loop: A set of optimal hyper-parameters is selected for the SVM classifier
based on a tenfold cross-validation within the training set. (D) Goodness-of-fit metrics: The classifier
learned in the previous step is tested on the testing dataset, and measures of its performance are
generated. This whole procedure is repeated 10 times (i.e., 10-fold CV) or 82 times (i.e., leave-1-out
CV) to produce generalized performance metrics, eliminating any bias introduced by a specific split
of training and testing sets.
A support vector machine is a binary classifier that finds the maximum margin hyper-plane that
separates the two classes in the data [Boser 1992]. The data being classified are denoted by 𝑋 ∈ ℝ𝑁×𝑃
(𝑁 channels and 𝑃 features), and the data from channel 𝑖 are denoted by 𝑋(𝑖) ∈ ℝ𝑃. The class labels
for all the channels are denoted by 𝑌 ∈ {−1,1}𝑁 (where −1 and 1 are numerical labels for the two
classes), and the class label for channel 𝑖 is denoted by 𝑌(𝑖) ∈ {−1,1}. The optimization problem to
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
find the optimal hyper-plane (described by weights 𝑊 ∈ ℝ𝑃 and intercept term 𝑏 ∈ ℝ) is shown in
Eq. 1.
min
W,b
1
2
‖𝑊‖2 (1)
subject to 𝑌(𝑖)(𝑊 𝑇𝑋(𝑖) + 𝑏) ≥ 1, 𝑖 ∈ {1, … , 𝑁}
Once the optimal hyper-plane [𝑊𝑜𝑝𝑡, 𝑏𝑜𝑝𝑡] is found, the predicted class label for channel 𝑖 is obtained
𝑇 𝑋(𝑖) + 𝑏𝑜𝑝𝑡. This formulation assumes that the data have a clear separation
as the sign of 𝑊𝑜𝑝𝑡
between the two classes. When that is not the case, slack variables and a tolerance parameter (box-
constraint) can be introduced to obtain separating hyper-planes that tolerate small misclassification
errors [Cortes 1995].
Dual formulation of SVM has received considerable interest because it enables use of different kernel
transformations of the original feature space without altering the optimization task and because of its
advantages in complexity when the data are high-dimensional [Boser 1992]. Specifically, this allows
for the features to be transformed from the original feature space to a kernel space. With this
transformation, the cases in which the original data are not linearly separable may be solved because
transformation of the data to higher dimensions may introduce linear separation in the transformed
domain. Linear, radial basis function (RBF), and polynomial kernels are widely used kernels in this
context.
Goodness of fit of the SVM classifier is evaluated by predicting the classes of the test dataset by
using the classifier that was trained on the training dataset and comparing the predictions against the
true class labels of the test dataset. This comparison is performed using standard performance metrics
such as receiver operating characteristics (ROC) curve analysis, area under ROC curve (AUC),
sensitivity, specificity, accuracy, precision, recall, and F1-score. Because of heterogeneity in the data,
choosing one partition of training and test datasets is not sufficient to credibly evaluate the
performance of a classifier. A common practice to obviate the effect of heterogeneity in the data is to
perform several iterations of training-testing cross-validation of the dataset. One run of this procedure
is carried out by choosing a subset of the dataset as training data, and testing on the rest of the dataset.
This approach allows the calculation of generalizable performance metrics for the analyzed classifier.
Results
Nonlinear Classification Boundary between SOZ and NSOZ Electrodes
Understanding the nature of the separation between the two classes in feature space is important to
achieve the maximum classification performance in binary classification. If the separation is linear, a
linear classifier should be sufficient (and preferable due to the Occam's razor principle) to achieve the
maximum attainable classification performance. On the other hand, when the separation is nonlinear,
linear classifiers perform poorly compared to nonlinear classifiers. However, when the feature space
is high-dimensional, visualizing the boundary between classes can be difficult. An option is to use
linear and nonlinear classifiers to classify the two classes and plot the histograms of likelihood
probabilities predicted by the classifier to understand the degree of separation achieved by linear and
nonlinear boundaries [Cherkassky 2010]. We performed this analysis for our dataset by using an SVM
classifier with linear and RBF kernels. Furthermore, we used the 10-fold cross-validation approach to
perform this analysis because the individual AUCs obtained using the leave-1-out cross-validation
approach were highly variable across patients. We trained two SVM classifiers with linear and RBF
kernels, respectively, using the framework shown in Figure 3 with 60% of all the electrodes and all
the biomarkers. These classifiers were used to predict the class labels for the rest of the electrodes
(40%). We then compared the distribution of the likelihood probabilities generated by the two SVM
classifiers against the true class labels. Figures 4a and 4b show the histograms of the likelihood
probabilities obtained using an SVM classifier with linear and RBF kernels, respectively, for SOZ and
NSOZ electrodes in the testing set. The nonlinear boundary achieved through an SVM classifier with
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
an RBF kernel clearly has a better separation between the two classes, as can also be seen in Figure
4c. To quantify this observation, we plotted ROC curves (which are shown in Figure 4d) for the
predictions obtained using linear and RBF kernels. The linear-SVM classifier obtained an AUC value
of 0.57, while the RBF-SVM classifier obtained an AUC of 0.79 when all the biomarkers were
utilized. Hence, we conclude that the boundary between SOZ and NSOZ electrodes is nonlinear in the
feature space in which the features are derived as described in Figure 1, and the rest of our analyses
focus on the results obtained by the SVM classifier with an RBF kernel.
Figure 4: Results obtained using a support vector machine with interictal electrophysiological
biomarkers to classify seizure onset zone (SOZ) electrodes. (A) and (B) show the probability
densities of the likelihoods predicted by an SVM classifier for SOZ and NSOZ electrodes in the
testing set for linear and RBF kernels, respectively. The RBF kernel results in less overlap between the
SOZ and NSOZ probability densities. (C) Boxplot showing the range of likelihood probabilities
obtained for SOZ and NSOZ electrodes when all biomarkers were used as features in a n SVM
classifier with an RBF kernel. (D) A comparison between the ROC curves when an SVM classifier
was used with an RBF kernel for individual biomarkers and their combination, and when it was used
with a linear kernel with a combination of all biomarkers. (E) A comparison between the AUCs
obtained using conventional unsupervised methods that use overall rates of biomarker incidence to
predict SOZ electrodes, and those obtained using our SVM-based supervised approach.
A Supervised Learning Approach Improves the SOZ Localization Accuracy
This paper proposes a supervised-learning-based approach that uses an SVM classifier to predict
electrodes in an SOZ by using interictal iEEG data. In order to understand whether our supervised
approach is better than simply using biomarker rates, we implemented biomarker-rate-based SOZ
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
electrode classification for HFO, IED, and PAC separately and in combination. (We simply added the
individual biomarker incidence rates to obtain an overall biomarker incidence rate.) Figure 4e shows a
comparison between the ROC curves obtained for the unsupervised biomarker incidence rate-based
approach and the supervised approach that uses an SVM classifier with an RBF kernel when all
biomarkers were utilized. Predictions using the unsupervised approach were performed on the same
testing set electrodes that were used in the supervised approach. The SVM-based supervised approach
outperformed the unsupervised approaches with a 17 -- 23% gain in the AUC value. Notably, the
performance of the SVM classifier with a linear kernel was comparable to that of the unsupervised
approach, with an AUC value of 0.57, as seen in Figure 4d. This highlights the ability to significantly
improve the correct classification of previously unseen SOZ electrodes by utilizing the right machine
learning method (in this case, SVM with an RBF kernel) to learn the characteristics of SOZ
electrodes.
Goodness-of-Fit Metrics for SOZ Electrode Classification
Other goodness-of-fit metrics, as specified previously, were calculated for the different combinations
of biomarkers and methods and are listed in Table 1. For all the metrics other than AUC, the
likelihood probabilities assigned by the classifier were applied with a threshold to classify SOZ and
NSOZ electrodes. In order to compare the different approaches, this threshold was chosen using a
common criterion, i.e., the false positive rate is approximately 25%. AUCs obtained for the training
set are reported in supplementary Table 3 along with optimal hyper-parameters used in each of the
cross-validations.
Table 1: Cross-validated goodness-of-fit metrics for SOZ determination. Here we list the
goodness-of-fit metrics (AUC, sensitivity, specificity, accuracy, precision, recall, and F1-score)
obtained for the test dataset, for the different combinations of biomarkers and analytic techniques
shown in Figure 4. The average values and standard deviations were computed using a) a tenfold
stratified cross-validation and b) a leave-one-out cross-validation.
Bio-
marker
10-fold CV
All
All
Method
AUC
Sensitivity
Specificity
Accuracy
Precision
(%)
(%)
(%)
(%)
Recall
(%)
F1-score
(%)
SVM-LIN
0.56(0.03)
32.20(4.17)
75.09(0.02)
67.23(0.78)
22.42(2.31)
32.2(4.17)
26.43(3.01)
SVM-RBF
0.79(0.01)
70.36(1.78)
75.09(0.00)
74.22(0.33)
38.79(0.60)
70.36(1.78)
50.01(0.95)
HFO
SVM-RBF
0.68(0.01)
53.71(1.70)
75.16(0.09)
71.23(0.29)
32.66(0.66)
53.71(1.70)
40.62(1.00)
IED
PAC
ALL
HFO
IED
PAC
SVM-RBF
0.68(0.01)
55.11(2.53)
75.07(0.03)
71.41(0.45)
33.14(1.01)
55.11(2.53)
41.39(1.50)
SVM-RBF
0.73(0.01)
60.63(2.74)
75.09(0.02)
72.44(0.51)
35.31(1.03)
60.63(2.74)
44.63(1.57)
RATE
0.58(0.01)
35.91(2.03)
75.09(0.02)
67.91(0.37)
24.43(1.03)
35.91(2.03)
29.07(1.40)
RATE
0.56(0.01)
32.39(2.20)
76.31(0.62)
68.26(0.77)
23.48(1.46)
32.39(2.20)
27.22(1.74)
RATE
0.58(0.01)
35.19(1.42)
75.18(0.09)
67.85(0.27)
24.14(0.74)
35.19(1.42)
28.63(0.99)
RATE
0.62(0.01)
43.76(2.25)
75.27(0.12)
69.49(0.42)
28.41(1.03)
43.76(2.25)
34.45(1.46)
Leave one (subject) out CV
SVM-RBF
0.73(0.02)
57.45(2.82)
79.49(0.57)
73.3(0.90)
38.71(2.45)
57.45(2.82)
43.49(1.99)
SVM-RBF
0.63(0.01)
35.53(2.89)
84.16(0.86)
73.10(1.00)
34.63(2.72)
35.53(2.89)
35.09(2.11)
SVM-RBF
0.60(0.01)
33.50(2.22)
80.77(0.66)
69.47(0.89)
30.55(2.42)
33.50(2.22)
29.16(1.55)
SVM-RBF
0.69(0.01)
47.70(2.81)
81.03(0.63)
72.63(0.93)
36.23(2.39)
47.70(2.81)
39.06(1.94)
ALL
HFO
IED
PAC
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Combining Multiple Electrophysiological Biomarkers Improves the Localization Accuracy
The classification framework depicted in Figure 3 was utilized with the features relevant to
HFO, IED, and PAC biomarkers separately to reveal the predictive ability of individual biomarkers.
Then the individual classification performances were compared against the performance obtained
when all the biomarkers were used together. Figure 4d shows the ROC curves obtained for the
different runs of the classification framework. While the PAC biomarker had the best predictive ability
individually (AUC: 0.74 -- 10-fold CV, 0.69 -- leave-1-out CV), the classification obtained using all
the biomarkers together performed better than any of the individual biomarkers, providing an AUC of
0.79 with the 10-fold CV approach and an AUC of 0.73 with the leave-1-out CV approach. These
findings support the idea that the different interictal electrophysiological biomarkers used in this study
possess complementary information that can be harnessed to achieve a superior performance in
predicting the electrodes in an SOZ.
Figure 5: Improvements in patient-specific SOZ classification achieved by means of combining
multiple biomarkers as opposed to utilizing a single biomarker in the SVM framework. (A), (B),
and (C): Improvements obtained in AUCs for patient-specific SOZ classification when the
combination of multiple biomarkers was utilized compared to when HFO, IED, or PAC, respectively,
was utilized alone. (D) Histogram densities of the patient-specific AUCs for the prediction of SOZ
electrodes using HFO, IED, PAC, and their composite.
To determine whether the combination of multiple biomarkers can reduce the inter-patient
variability, we analyzed the improvements in SOZ electrode classification potential in each individual
by calculating the AUC for each individual separately. We predicted the SOZ electrodes of each
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
individual patient using HFO features, IED features, and PAC features separately and then their
composite, within the SVM-based framework described previously using a leave-one-out cross-
validation approach (see supplementary Table 2). Figures 5a -- 5c illustrate the respective improvements
in the AUCs of individual patients attained when the combination of the three biomarkers was utilized
instead of HFO, IED, or PAC by itself. Our analysis indicates that the AUCs of individual patients
improved or remained unchanged for more than 65% of the patients when the composite was utilized
compared to any individual biomarker. Then we plotted the histograms of the AUCs of individual
patients for each biomarker and their composite separately, and approximated the densities by using
kernel density estimation. Figure 5d shows that the histogram density of AUCs of individual patients
becomes skewed towards higher AUC values when the composite is utilized to predict SOZ
electrodes. This indicates that the utilization of multiple biomarkers with complementary information
reduces the overall variability across patients in the ability to classify SOZ electrodes -- variability that
is apparent with any single biomarker.
Recording Durations Between 90 and 120 Minutes May be Sufficient for Interictal SOZ Localization
The ability to localize seizure-generating brain tissue is a cornerstone of clinical epileptology.
We investigated how the duration of interictal iEEG recording impacted the the localization of the
SOZ (as illustrated in Figure 6a). Here we applied our AI-based framework on a range of recording
durations between 10 and 120 minutes. Figure 6 shows the mean ROC curves obtained using a tenfold
cross-validation for different durations when an SVM classifier with an RBF kernel was utilized with
all the biomarkers. To quantify the different runs, we plotted the AUC metric against the recording
length used for SOZ electrode classification prediction. That is shown in Figure 6c, where the error
bars indicate the standard deviations of the AUCs based on a tenfold cross-validation. Statistical
significance tests using two-tailed paired t-tests indicate that the AUCs obtained using 90-, 100-, and
110-minute recordings are not statistically very different from the AUCs obtained using a 120-minute
recording. This finding indicates that recording durations between 90 and 120 minutes may be
sufficient for interictal SOZ identification with clinically relevant accuracies.
Figure 6: Evaluation of the length of recordings and interictal SOZ localization. (A) ROC curves
obtained when shorter interictal segments of durations ranging from 20 to 120 minutes were utilized
for analysis. (B) AUC values obtained with short interictal segments. Longer interictal segments result
in better AUC values; however, the AUCs obtained using segments longer than 90 minutes are
statistically indifferent (based on 2-sided paired t-tests between AUCs).
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
20406080100120Recording duration (minutes)0.60.650.70.750.80.85Area Under ROC Curve (AUC)00.20.40.60.81False positive rate00.20.40.60.81True positive rate20 minutes40 minutes60 minutes80 minutes100 minutes120 minutes(A)(B)NSNSNSNS: P<0.05: Not significant Discussion
Main Contribution of the Study
The current study describes a machine learning method for classification of SOZ and NSOZ
electrodes using multiple electrophysiological biomarkers extracted from interictal iEEG data
collected in a clinical setting. Our study, to our knowledge, is the first to utilize the complementary
information provided by multiple electrophysiological biomarkers and their temporal characteristics
as a way of reducing variability across patients to improve interictal SOZ localization. Using 2-hour
wide-bandwidth intracranial EEG recordings of ~5000 electrodes from 82 patients, our study provides
a large-scale evaluation of an artificial intelligence-based approach for interictal SOZ localization and
shows that when used in concert with multiple interictal biomarkers, it can outperform single
biomarker approaches. The ability to perform real-time feature processing and SOZ determination
with clinically relevant accuracies, with a maximum monitoring duration of 2 hours, supports the
feasibility of SOZ determination using interictal intracranial EEG data (see supplementary Table 2 for
specific computation times for each task included in interictal SOZ classification). In addition to the
above specific contributions, our study also exemplifies the role of artificial intelligence in
augmenting clinical workflows.
The Value of Combining Multiple Biomarkers
Interictal SOZ localization techniques have been widely discussed, with the main focus being
on the search for and validation of a single biomarker that can be used in all patients [Bragin 1999,
Jacobs 2010, Jacobs 2008, Worrell 2011, Engel 2013]. Conventional methods have focused on HFO
biomarker detection algorithms and on HFOs themselves [Worrell 2012, Burnos 2014, Balach 2014].
However, the generalizability of such single biomarkers has been insufficient for clinical practice
[Nonoda 2016, Sinha 2017, Holler 2015, Cimbalnik 2017], primarily because of inter-patient
variability, and it appears that one biomarker may not be sufficient to identify SOZs in all patients.
While there have been multiple attempts to automate SOZ localization [Liu 2016, Graef 2013, Gritsch
2011], very little work has attempted to improve localization potential by means of combining
multiple biomarkers. This exploratory study shows
interictal
electrophysiological biomarkers within a rigorous, supervised machine learning setting can be more
accurate in performing interictal SOZ localization than can utilization of a single biomarker,
essentially by reducing inter-patient variability. This is evident from our individual patient-based
analysis (Figure 5), in which we show that SOZ electrode classification AUCs improve or remain
unchanged for more than 65% of patients when the combination of the three biomarkers is utilized
instead of single biomarkers. This finding indicates that combining multiple biomarkers reduces the
variance in SOZ electrode classification and therefore achieves better generalizability than single-
biomarker-based approaches. Figure 5 also shows that combining multiple biomarkers reduces the
accuracy in SOZ electrode classification for some patients when compared with any of the individual
biomarkers. The reason for this could be that there are more disagreements within the different
biomarkers than agreements with respect to SOZ electrodes. The situations in which combining
multiple biomarkers is detrimental can be explored and tested by comparing the individual predictions
provided by the biomarkers.
that combining multiple
Exploiting the Temporal Variability in Epileptic Activity to Improve Classification Potential
We showed in this study that utilizing a machine learning approach that uses local rates of
biomarkers within 10-minute subintervals for classifying SOZ electrodes improves the AUC by 19%
compared to traditional unsupervised approaches that primarily utilize the overall rates of biomarkers.
We believe that the improvement is due to the inability of the overall-rate-based approaches to
account for the temporal variations in epileptic activity. Recent studies have reported that the rate of
epileptiform activity changes significantly between different behavioral states [Worrell 2008, Amiri
2016]. Since our study uses night segments with mixed behavioral states, the ability to differentiate
SOZ electrodes from NSOZ electrodes using rates of epileptiform activity varies depending on the
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
portion of the segment. Hence, looking at the overall rates of epileptiform activity might average out
the variations between subintervals and hence degrade performance, and result in accuracy lower than
the maximum attainable. We also show that a nonlinear classification technique provides significant
improvements in AUC compared to a linear classification technique and that the performance
provided by the latter is similar to that of commonly used unsupervised approaches. Although a linear
classifier considers each subinterval for classification, it is impractical to assign a particular
subinterval a higher weight because the exact subintervals in which the epileptiform activity is highly
discriminative may not be the same across different patients. Therefore, when trained across multiple
patients, the linear classifier perceives each subinterval as equally important and assigns all of them
equal weights. As a result, its performance is similar to that of the approaches that use overall
epileptiform activity rates to classify SOZ electrodes. On the other hand, an RBF kernel measures
similarities (Euclidean distances) between a channel and selected SOZ and NSOZ channels (known as
support vectors) with respect to their local rates of epileptiform activities. Therefore, regardless of the
position of the subinterval that is highly discriminative between SOZ and NSOZ channels, that
discriminative ability will be reflected in the overall distance between those channels.
That is pictorially illustrated in Figures 7a -- 7d using an example. Figure 7a shows that there is
a large variation in the local rates of PAC (in 10-minute windows) across channels of a selected
patient. Figure 7b shows the overall PAC rates for each channel, and shows that classifying SOZ
electrodes simply by thresholding the overall rates results in poor sensitivity and specificity. Figure 7c
shows that transforming the local rates to a kernel space using a linear kernel does not produce any
major changes compared to the previous approach with regard to efficiency (in terms of sensitivity
and specificity), whereas Figure 7d shows that transforming the local rates using an RBF kernel
produces a more favorable transformation because it provides improved sensitivity and specificity. As
shown in Figures 7c and 7d, kernel distances between feature vectors of individual channels and
average feature vectors of the SOZ channels were calculated using linear and RBF kernels. This
example elucidates the utility of a nonlinear machine learning classification approach and its
relationship to the underlying physiology. It is also noteworthy that a nonlinear classification
approach is suitable in this case due to the manner in which the features of channels were derived, and
that linear classifiers may be sufficient for a different derivation of the features.
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Figure 7: Temporal variability in the rates of epileptiform activity (with respect to PAC) and its
relation to nonlinear classification. This figure illustrates that the RBF kernel is more specific in
capturing the similarities between electrodes with regards to their epileptiform activity patterns. (A)
Normalized local PAC rates in different 10-minute intervals of SOZ and NSOZ channels of a selected
patient. (B) Overall PAC rates (obtained by summing local rates) for the channels and their poor
ability to classify SOZ electrodes. (SS is sensitivity, SP is specificity, and a dashed line indicates
application of a threshold). (C) Features after application of a transformation using the linear kernel.
This transformation does not provide any notable improvement compared to the summation approach.
(D) Features after application of a nonlinear transformation using the RBF kernel. It is evident that
this transformation provides better discrimination between SOZ and NSOZ channels.
Identifying SOZs During Long and Short Recordings of Night Segments
In this study, we showed that localizing an SOZ using multiple features identified in 120
minutes of mixed behavioral state data provided accuracy similar to that obtained with shorter
segments (see Figure 6). There is a clear relationship between the utility of this platform and the time
of the recording used in analysis. Segments of 120 minutes were arbitrarily chosen to be the
maximum amount of time a neurologist could use during an operating room recording to identify the
SOZ. Interestingly, results as measured by the AUC did not significantly differ for recording
durations between 90 and 120 minutes. It appears that, given satisfactory recording conditions, less
than one hour may be all that is needed to achieve high SOZ identification accuracy with this
platform. The relationship between the sleep-wake cycle and epileptiform discharges is well known
[Sammaritano 1991, Staba 2002, Malow 1998]. Investigation of PAC, HFO, and IED using both
short-term and long-term iEEG data shows an increment of the HFO rate with the non-rapid eye
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
movement (NREM) sleep stage, especially in subjects with temporal lobe epilepsy [von Ellenrieder
2017, Amiri 2016, Staba 2004] [Engel 2009]. In addition, recent work has shown that the PAC
localization potential increases in slow-wave sleep [Amiri 2016]. Our results suggest that obtaining
sleep recordings of a sufficient duration in the clinical routine can be beneficial to quantitative SOZ
localization.
Contributions Towards Advancing the Current State of Clinical Decision-Making
We demonstrated that our approach can identify electrodes in an SOZ with an accuracy of
approximately 80% (AUC) on a large cohort of 82 patients, using interictal iEEG recordings of
durations less than two hours. The idea of utilizing artificial intelligence to augment clinical
workflows has become central in the era of "big data." Studies have shown the utility of deep-
learning-based approaches in augmenting clinical diagnosis of skin cancer and diabetic retinopathy,
primarily using imaging measurements [Esteva 2017, Gulshan 2016]. These studies have benefited
considerably from the availability of a) huge image databases and b) substantially validated artificial
neural-network-based classification models. Our study, on the other hand, embodies an alternative
approach that utilizes feature engineering as a remedial option for the unavailability of large datasets
at the scale of the currently available imaging datasets (with millions of medical images). We believe
that the insights drawn from this study will be particularly useful for tasks that lack an abundance of
labeled training samples, which inevitably is the case for most clinical problems.
Study Limitations
Our approach in its current implementation does not possess the ability to differentiate
pathological and physiological electrophysiological events. For instance, HFOs are also associated
with normal physiological function, and how to distinguish physiological HFOs from pathological
HFOs is an active research area [Matsumoto 2013]. Recent studies have shown that utilizing only the
pathological events results in increased sensitivity and specificity in determining SOZs interictally
[Weiss 2016]. Hence, the ability to differentiate pathological events and then utilize them in
determining SOZs may further improve our localization accuracy.
As shown recently [Amiri 2016], there is a clear connection between behavioral/sleep state
and each of the biomarkers implemented in this study. We also showed that the temporal variations in
epileptiform activity due to changes in behavioral states influences the machine learning paradigm
utilized in this work. Therefore, accurate annotations of behavioral states considered together with
interictal electrophysiologic biomarkers may further improve classification of epileptic and normal
brain tissue. However, the patients in this cohort did not have scalp EEGs as required for accurate
behavioral state classifications. New methods that can classify sleep stages based on intracranial
recordings could prove useful for future analyses [Kremen & Duque 2016].
A successful clinical translation of this approach would depend on the accuracy of this
approach in the data collected under operative settings. Prior studies evaluating IEDs [Wass 2001,
Schwartz 1997] and HFOs [Zijlmans 2012, Wu 2010] for their potential to localize epileptic brain
under intraoperative settings show promise. The PAC biomarker, however, despite being the best
individual predictor in interictal settings, has not been evaluated under intraoperative settings and its
specificity in such settings is still unclear. Hence, future studies evaluating multiple epilepsy patients
are required to accurately determine the clinical utility of this approach.
Another limitation of our approach is that we take clinical SOZ as the gold standard in
determining the accuracy of the interictal approach. This is a limitation because only a fraction of the
patients going through epilepsy patients achieve complete seizure freedom (i.e., ILAE outcome 1).
However, the goal of this study is to evaluate how accurately interictal biomarkers can localize the
ictal recording SOZ localization. We make the assumption that any clues regarding SOZs that were
not captured in ictal localization are less likely to be captured using interictal localization. Regardless,
we observed that there is a good agreement between ictal and interictal SOZ localization approaches
when the patients eventually had good outcomes (see supplementary Figure 1).
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Future Directions
Advances in neuroimaging methods have advanced non-invasive seizure localization
capabilities in epilepsy. Because of the holistic spatial view made possible by imaging techniques,
they can provide independent information about potential regions of epileptogenic brain that could be
used in concert with interictal electrophysiological biomarkers to further improve localization of
pathological brain tissue. Alternatively, one can see the utility of combining pathological event
classification, behavioral state classification, and multiple imaging modalities as another way of
accounting for inter-patient variability, in addition to utilizing multiple interictal electrophysiological
biomarkers. However, combining such different data types might require a more complex analytic
technique in order to effectively capture the complementary information provided by each data type.
We hypothesize that probabilistic graphical models provide an exceptional platform for handling such
complexities, as shown in a recent work that combined spatial and temporal relationships in EEGs
using a factor-graph-based model [Varatharajah 2017]. To that end, our future work will be directed
towards harnessing the utilities of probabilistic graphical models in combining the forenamed
multimodal data.
Data and software availability
The iEEG data and the software used in this study for AI-based interictal SOZ identification is
available for download at ftp://msel.mayo.edu/EEG_Data/ and ftp://msel.mayo.edu/JNE_bundle.zip
respectively.
Conclusions
Current methods for localizing seizure onset zones (SOZs) in drug-resistant epilepsy patients are
lengthy, costly, and associated with patient discomfort and potential complications. Furthermore,
among patients undergoing surgical resection for treatment of drug-resistant epilepsy, approximately
46% suffer seizure recurrence within 5 years. The current gold standard for localizing the tissue to be
surgically resected is visual review of ictal (seizure) recordings. In this paper, we reported an artificial
intelligence (AI) based approach for the prediction of SOZ electrodes using only non-seizure data.
This technique was validated using interictal iEEG data clinically collected from 82 patients with
drug-resistant epilepsy. The approach uses three pathological electrophysiological transients reported
to be interictal (non-seizure) biomarkers of epileptic brain tissue: high-frequency oscillations (HFOs),
interictal epileptiform discharges (IEDs), and phase-amplitude coupling (PAC). We utilize the
complementary information provided by these 3 biomarkers and their temporal dynamics in a support
vector machine classification (SVM) paradigm to reduce variability across patients and eventually to
achieve an average area under ROC curve (AUC) measure of 0.73 in correctly classifying SOZ
electrodes interictally. Furthermore, our results suggest that recording durations of approximately two
hours are sufficient to localize the SOZ. Some potential applications of this technology are intra-
operative electrode placement for prolonged monitoring, electrode placement for electrical stimulation
devices [Fisher 2014] to treat seizures, and guiding of the margins of resection around epileptic
structural lesions seen on MRI (magnetic resonance imaging). The potential to reduce the number of
patients requiring long-term (multiple-day) iEEG monitoring is also interesting, but will require
additional study. We postulate that in the future, combining pathological event classification,
behavioral state classification, and multiple imaging modalities with interictal electrophysiological
biomarkers could further improve interictal localization potential. We also believe that the insights
drawn from this AI-based study will be particularly useful for clinical problems that require real-time
data collection and decision making.
References
Amiri, M., Frauscher, B., & Gotman, J. Phase-amplitude coupling is elevated in deep sleep and in the
in Human Neuroscience, 10, 387.
onset
http://doi.org/10.3389/fnhum.2016.00387 (2016).
seizures. Frontiers
zone of
focal
epileptic
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Ashburner, J., Barnes, G., Chen, C., Daunizeau, J., Flandin, G., Friston, K., Gitelman, D., Kiebel, S.,
Kilner, J., Litvak, V. and Moran, R., 2008. SPM8 manual. Functional Imaging Laboratory, Institute of
Neurology, 41.
Balach, J., et al. Comparison of algorithms for detection of high frequency oscillations in intracranial
EEG. Proc. 2014 IEEE International Symposium on Medical Measurements and Applications (MeMeA)
1 -- 4. http://doi.org/10.1109/MeMeA.2014.6860107 (2014).
Barkmeier, D. T., et al. High inter-reviewer variability of spike detection on intracranial EEG addressed
by an automated multi-channel algorithm. Clinical Neurophysiology, 123(6), 1088 -- 1095.
http://doi.org/10.1016/j.clinph.2011.09.023 (2012).
Ben-Ari, Y., & Bragin, A. Physiologic and pathologic oscillations. Trends in Neurosciences, 30(7), 307 --
308. http://doi.org/10.1016/j.tins.2007.05.008 (2007).
9(2),
137 -- 142.
Blakely, T., Miller, K. J., Rao, R. P. N., Holmes, M. D., & Ojemann, J. G. Localization and classification
of phonemes using high spatial resolution electrocorticography (ECoG) grids. Proc. 2008 30th Annual
International Conference of the IEEE Engineering in Medicine and Biology Society, 4964 -- 4967.
http://doi.org/10.1109/IEMBS.2008.4650328 (2008).
Boser, B. E., Guyon, I. M. & Vapnik, V. N. A training algorithm for optimal margin classifiers. Proc.
Fifth Annual Workshop on Computational Learning Theory, 144 -- 152. (1992).
Bragin, A., Engel, J., Wilson, C. L., Fried, I., & Buzski, G. High-frequency oscillations in human brain.
Hippocampus,
http://doi.org/10.1002/(SICI)1098-1063(1999)9:2<137::AID-
HIPO5>3.0.CO;2-0 (1999).
Bragin, A., Wilson, C. L., Staba, R. J., Reddick, M., Fried, I., & Engel, J. Interictal high-frequency
oscillations (80-500 Hz) in the human epileptic brain: Entorhinal cortex. Ann Neurol, 52, 407 -- 415. (2002).
Brinkmann, B. H., Bower, M. R., Stengel, K. A., Worrell, G. A., & Stead, M. Large-scale
electrophysiology: Acquisition, compression, encryption, and storage of big data. Journal of Neuroscience
Methods, 180(1), 185 -- 192. http://doi.org/10.1016/j.jneumeth.2009.03.022 (2009).
Brodie, M. J., & Kwan, P. Staged approach to epilepsy management. Neurology, 58(8 Suppl 5), S2 -- S8.
http://doi.org/10.1212/WNL.58.8_SUPPL_5.S2 (2002).
Burkholder, D. B., Sulc, V., Hoffman, E. M., Cascino, G. D., Britton, J. W., So, E. L., Marsh, W. R.,
Meyer, F. B., Van Gompel, J. J., Giannini, C., Wass, C. T., Watson, R. E., & Worrell G. A. Interictal scalp
electroencephalography and intraoperative electrocorticography in magnetic resonance imaging-negative
temporal lobe epilepsy surgery. JAMA Neurology, 71(6), 702 -- 709. (2014).
Burnos, S., et al. Human intracranial high frequency oscillations (HFOs) detected by automatic time-
frequency analysis. PLoS ONE, 9(4), e94381. http://doi.org/10.1371/journal.pone.0094381 (2014).
Canolty, R. T., Cadieu, C. F., Koepsell, K., Knight, R. T., & Carmena, J. M. Multivariate phase -- amplitude
cross-frequency coupling in neurophysiological signals. IEEE Transactions on Biomedical Engineering,
59(1), 8 -- 11. http://doi.org/10.1109/TBME.2011.2172439 (2012).
Canolty, R. T., et al. High gamma power is phase-locked to theta oscillations in human neocortex.
Science, 313(5793), 1626 -- 1628. http://doi.org/10.1126/science.1128115 (2006).
Cherkassky, V. and Dhar, S., 2010, July. Simple Method for Interpretation of High-Dimensional
Nonlinear SVM Classification Models. In DMIN (pp. 267-272).
Cimbalnik, J., Kucewicz, M. T., & Worrell, G. Interictal high-frequency oscillations in focal human
epilepsy.
175 -- 181.
http://doi.org/10.1097/WCO.0000000000000302 (2016).
Neurology,
Opinion
Current
in
29(2),
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
&
80(7),
of
Neurology,
Psychiatry,
Neurosurgery
Colgin, L. L. Rhythms of the hippocampal network. Nature Reviews Neuroscience, 17(4), 239 -- 249.
http://doi.org/10.1038/nrn.2016.21 (2016).
Cortes, C. & Vapnik, V. Support-vector networks. Machine Learning, 20(3), 273 -- 297. (1995).
Crisler, S., Morrissey, M. J., Anch, A. M., & Barnett, D. W. Sleep-stage scoring in the rat using a support
vector machine. Journal of Neuroscience Methods, 168(2), 524 -- 534. (2008).
D'Alessandro, M., et al. Epileptic seizure prediction using hybrid feature selection over multiple
intracranial EEG electrode contacts: A report of four patients. IEEE Transactions on Biomedical
Engineering, 50(5), 603 -- 615. http://doi.org/10.1109/TBME.2003.810706 (2003).
DiLorenzo, D. J., Mangubat, E. Z., Rossi, M. A., & Byrne, R. W. Chronic unlimited recording
electrocorticography-guided resective epilepsy surgery: Technology-enabled enhanced fidelity in seizure
focus localization with improved surgical efficacy. Journal of Neurosurgery, 120(6), 1402 -- 1414.
http://doi.org/10.3171/2014.1.JNS131592 (2014).
Edakawa, K., et al. Detection of epileptic seizures using phase-amplitude coupling in intracranial
electroencephalography. Scientific Reports, 6(1), 25422. http://doi.org/10.1038/srep25422 (2016).
Elsharkawy, A. E., et al. Long-term outcome of lesional posterior cortical epilepsy surgery in adults.
Journal
773 -- 780.
http://doi.org/10.1136/jnnp.2008.164145 (2009).
Engel Jr, J., Bragin, A., Staba, R., & Mody, I. High-frequency oscillations: What is normal and what is
not? Epilepsia, 50(4), 598 -- 604. http://doi.org/10.1111/j.1528-1167.2008.01917.x (2009).
Engel, J., et al. Epilepsy biomarkers. Epilepsia, 54(Suppl. 4), 61 -- 69. http://doi.org/10.1111/epi.12299
(2013).
Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. Dermatologist-level
classification of skin cancer with deep neural networks. Nature, 542(7639), 115 -- 118. (2017).
Fisher, R. S. & Velasco, A. L. Electrical brain stimulation for epilepsy. Nat Rev Neurol, 10, 261 -- 270.
(2014).
Gnatkovsky, V., de Curtis, M., Pastori, C., Cardinale, F., Lo Russo, G., Mai, R., Nobili, L., Sartori, I.,
Tassi, L., & Francione, S. Biomarkers of epileptogenic zone defined by quantified stereo-EEG analysis.
Epilepsia, 55(2), 296 -- 305. (2014).
Graef, A., et al. Automatic ictal HFO detection for determination of initial seizure spread. Proc. 2013 35th
Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2096 -- 2099.
https://doi.org/10.1109/EMBC.2013.6609946 (2013).
Gritsch, G., et al. Automatic detection of the seizure onset zone based on ictal EEG. Proc. 2011 Annual
International Conference of the IEEE Engineering in Medicine and Biology Society, 3901 -- 3904.
http://doi.org/10.1109/IEMBS.2011.6090969 (2011).
Gump, W. C., Skjei, K. L., & Karkare, S. N. Seizure control after subtotal lesional resection.
Neurosurgical Focus, 34(6), E1. http://doi.org/10.3171/2013.3.FOCUS1348 (2013).
Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Wu, D., Narayanaswamy, A., Venugopalan, S.,
Widner, K., Madams, T., Cuadros, J., & Kim, R. Development and validation of a deep learning algorithm
for detection of diabetic retinopathy in retinal fundus photographs. JAMA, 316(22), 2402 -- 2410. (2016).
Hauf, M., et al. Localizing seizure-onset zones in presurgical evaluation of drug-resistant epilepsy by
electroencephalography/fMRI: Effectiveness of alternative thresholding strategies. American Journal of
Neuroradiology, 33(9), 1818 -- 1824. http://doi.org/10.3174/ajnr.A3052 (2012).
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Höller, Y., et al. High-frequency oscillations in epilepsy and surgical outcome: A meta-analysis. Frontiers
in Human Neuroscience, 9, 574. http://doi.org/10.3389/fnhum.2015.00574 (2015).
Jacobs, J., LeVan, P., Chander, R., Hall, J., Dubeau, F., & Gotman, J. Interictal high-frequency
oscillations (80-500 Hz) are an indicator of seizure onset areas independent of spikes in the human
epileptic brain. Epilepsia, 49(11), 1893 -- 1907. http://doi.org/10.1111/j.1528-1167.2008.01656.x (2008).
Jacobs, J., et al. High-frequency electroencephalographic oscillations correlate with outcome of epilepsy
surgery. Annals of Neurology, 67(2), 209 -- 220. http://doi.org/10.1002/ana.21847 (2010).
Jenks, G. F. The data model concept in statistical mapping. International Yearbook of Cartography, 7,
186 -- 190. (1967).
Jensen, O., & Colgin, L. L. Cross-frequency coupling between neuronal oscillations. Trends in Cognitive
Sciences, 11(7), 267 -- 269. http://doi.org/10.1016/j.tics.2007.05.003 (2007).
Jirsch, J. D., et al. High-frequency oscillations during human focal seizures. Brain, 129(6), 1593 -- 1608.
http://doi.org/10.1093/brain/awl085 (2006).
Karoly, P. J., et al. Interictal spikes and epileptic seizures: Their relationship and underlying rhythmicity.
Brain, 139(4), 1066 -- 1078. http://doi.org/10.1093/brain/aww019 (2016).
Kremen, V., et al. Behavioral state classification in epileptic brain using intracranial electrophysiology.
Journal of Neural Engineering, 14(2), 26001. http://doi.org/10.1088/1741-2552/aa5688 (2017).
Kucewicz, M. T., et al. Dissecting gamma frequency activity during human memory processing. Brain,
140(5), 1337 -- 1350. http://doi.org/10.1093/brain/awx043 (2017).
Kwan, P., & Brodie, M. J. Early identification of refractory epilepsy. The New England Journal of
Medicine, 342(5), 314 -- 319. http://doi.org/10.1056/NEJM200002033420503 (2000).
Lee, S. K. Surgical approaches in nonlesional neocortical epilepsy. Journal of Epilepsy Research, 1(2),
47 -- 51. http://doi.org/10.14581/jer.11009 (2011).
Leonardi, M., & Ustun, T. B. The global burden of epilepsy. Epilepsia, 43(s6), 21 -- 25. (2002).
Liu, S., et al. Exploring the time -- frequency content of high frequency oscillations for automated
identification of seizure onset zone in epilepsy. Journal of Neural Engineering, 13(2), 26026.
http://doi.org/10.1088/1741-2560/13/2/026026 (2016).
Lüders, H. O., Najm, I., Nair, D., Widdess-Walsh, P., & Bingman, W. The epileptogenic zone: General
principles. Epileptic Disorders,
http://www.jle.com/en/revues/epd/e-
docs/the_epileptogenic_zone_general_principles_270796/article.phtml (2006).
8(Suppl.
2),
S1 -- S9.
Luther, N., Rubens, E., Sethi, N., Kandula, P., Labar, D. R., Harden, C., Perrine, K., Christos, P. J.,
Iorgulescu, J. B., Lancman, G., Schaul, N. S., Kolesnik, D. V., Nouri, S., Dawson, A., Tsiouris, A. J., &
Schwartz, T. H. The value of intraoperative electrocorticography in surgical decision making for temporal
lobe epilepsy with normal MRI. Epilepsia, 52(5), 941 -- 948. (2011).
Malow, B. A., Lin, X., Kushwaha, R., & Aldrich, M. S. Interictal spiking increases with sleep depth in
temporal lobe epilepsy. Epilepsia, 39(12), 1309 -- 1316. http://doi.org/10.1111/j.1528-1157.1998.tb01329.x
(1998).
Marsh, E. D., et al. Interictal EEG spikes identify the region of electrographic seizure onset in some, but
not all, pediatric epilepsy patients. Epilepsia, 51(4), 592 -- 601. http://doi.org/10.1111/j.1528-
1167.2009.02306.x (2010).
Matsumoto, A., et al. Pathological and physiological high-frequency oscillations in focal human epilepsy.
Journal of Neurophysiology, 110(8), 1958 -- 1964. http://doi.org/10.1152/jn.00341.2013 (2013).
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Miller, K. J., Zanos, S., Fetz, E. E., den Nijs, M., & Ojemann, J. G. Decoupling the cortical power
spectrum reveals real-time representation of individual finger movements in humans. Journal of
Neuroscience, 29(10), 3132 -- 3137. http://doi.org/10.1523/JNEUROSCI.5506-08.2009 (2009).
Nair, C., Prabhakar, B., & Shah, D. On entropy for mixtures of discrete and continuous variables.
Retrieved from http://arxiv.org/abs/cs/0607075 (2006).
Noe, K., et al. Long-term outcomes after nonlesional extratemporal lobe epilepsy surgery. JAMA
Neurology, 70(8), 1003 -- 1008. http://doi.org/10.1001/jamaneurol.2013.209 (2013).
Nonoda, Y., et al. Interictal high-frequency oscillations generated by seizure onset and eloquent areas may
be differentially coupled with different slow waves. Clinical Neurophysiology, 127(6), 2489 -- 2499.
http://doi.org/10.1016/j.clinph.2016.03.022 (2016).
Pail, M., Řehulka, P., Cimbálník, J., Doležalová, I., Chrastina, J., & Brázdil, M. Frequency-independent
in epileptic and non-epileptic regions. Clinical
characteristics of high-frequency oscillations
Neurophysiology, 128(1), 106 -- 114. http://doi.org/10.1016/j.clinph.2016.10.011 (2017).
Papadelis, C.,
simultaneous
magnetoencephalography and electroencephalography as biomarker of pediatric epilepsy. Journal of
Visualized Experiments, 118, e54883. http://doi.org/10.3791/54883 (2016).
Pearce, A., Wulsin, D., Blanco, J. A., Krieger, A., Litt, B., & Stacey, W. C. Temporal changes of
neocortical high frequency oscillations in epilepsy. J Neurophysiol, 110(5), 1167 -- 1179. (2013).
Refaeilzadeh, P., Tang, L., & Liu, H. Cross-validation. In Encyclopedia of Database Systems, L. Liu & M.
T. Özsu, Eds. (pp. 532 -- 538). Springer US. (2009).
Russell, S. J. and Norvig, P. Artificial Intelligence: A Modern Approach. Prentice-Hall, Englewood
Cliffs. (1995).
Sammaritano, M., Gigli, G. L., & Gotman, J. Interictal spiking during wakefulness and sleep and the
localization of foci in temporal lobe epilepsy. Neurology, 41(2(Pt 1)), 290 -- 297. Retrieved from
http://www.ncbi.nlm.nih.gov/pubmed/1992379 (1991).
frequency oscillations detected with
et al.
Interictal high
Schwartz, T.H., Bazil, C.W., Walczak, T.S., Chan, S., Pedley, T.A., and Goodman, R.R. (1997). The
predictive value of intraoperative electrocorticography in resections for limbic epilepsy associated with
mesial temporal sclerosis. Neurosurgery 40, 302-9; discussion 309-11.
Sinha, N., et al. Predicting neurosurgical outcomes in focal epilepsy patients using computational
modelling. Brain, 140(2), 319 -- 332. http://doi.org/10.1093/brain/aww299 (2017).
Staba, R. J., Wilson, C. L., Bragin, A., Fried, I., & Engel, J. Quantitative analysis of high-frequency
oscillations (80-500 Hz) recorded in human epileptic hippocampus and entorhinal cortex. Journal of
Neurophysiology, 88(4), 1743 -- 1752. https://doi.org/10.1152/jn.2002.88.4.1743 (2002).
Staba, R. J., Wilson, C. L., Bragin, A., Jhung, D., Fried, I., & Engel, J. High-frequency oscillations
recorded in human medial temporal lobe during sleep. Annals of Neurology, 56(1), 108 -- 115.
http://doi.org/10.1002/ana.20164 (2004).
Staley, K. J., & Dudek, F. E. Interictal spikes and epileptogenesis. Epilepsy Currents, 6(6), 199 -- 202.
http://doi.org/10.1111/j.1535-7511.2006.00145.x (2006).
Stead, M., Bower, M., Brinkmann, B. H., Lee, K., Marsh, W. R., Meyer, F. B., Litt, B., Van Gompel, J., &
Worrell, G. A. Microseizures and the spatiotemporal scales of human partial epilepsy. Brain, 133(9),
2789 -- 2797. (2010).
Tellez-Zenteno, J. F. Long-term seizure outcomes following epilepsy surgery: A systematic review and
meta-analysis. Brain, 128(5), 1188 -- 1198. http://doi.org/10.1093/brain/awh449 (2005).
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Tzallas, A. T., Tsipouras, M. G., & Fotiadis, D. I., Automatic seizure detection based on time-frequency
analysis and artificial neural networks. Computational
Intelligence and Neuroscience, 2007,
80510. https://www.hindawi.com/journals/cin/2007/080510/abs/ (2007).
Van Gompel, J. J., Worrell, G. A., Bell, M. L., Patrick, T. A., Cascino, G. D., Raffel, C., Marsh, W. R., &
Meyer, F. B. Intracranial electroencephalography with subdural grid electrodes: Techniques,
complications, and outcomes. Neurosurgery, 63(3), 498 -- 505; discussion 505 -- 506. (2008).
Van Gompel, J. J., Rubio, J., Cascino, G. D., Worrell, G. A., & Meyer, F. B. Electrocorticography-guided
resection of temporal cavernoma: Is electrocorticography warranted and does it alter the surgical
approach? J Neurosurg, 110(6), 1179 -- 1185. (2009).
van 't Klooster, M. A., Leijten, F. S., Huiskamp, G., Ronner, H. E., Baayen, J. C., van Rijen, P. C.,
Eijkemans, M. J., Braun, K. P., Zijlmans, M., & HFO study group. High frequency oscillations in the
intra-operative ECoG to guide epilepsy surgery ("The HFO Trial"): Study protocol for a randomized
controlled trial. Trials, 16, 422. (2015).
Vapnik, V. Estimation of Dependences Based on Empirical Data [2nd Ed.]: Empirical Inference Science
Afterword of 2006. New York, NY: Springer. https://doi.org/10.1007/0-387-34239-7 (2006).
Varatharajah, Y., Berry, B.M., Kalbarczyk, Z.T., Brinkmann, B.H., Worrell, G.A. and Iyer, R.K., 2017,
May. Inter-ictal Seizure Onset Zone localization using unsupervised clustering and Bayesian Filtering.
In Neural Engineering (NER), 2017 8th International IEEE/EMBS Conference on (pp. 533-539). IEEE.
Varatharajah, Y., Chong, M. J., Saboo, K., Berry, B., Brinkmann, B., Worrell, G., & Iyer, R. EEG-
GRAPH: A factor-graph-based model for capturing spatial, temporal, and observational relationships in
electroencephalograms. Advances
in Neural Information Processing Systems, 30, 5377 -- 5386.
https://papers.nips.cc/book/advances-in-neural-information-processing-systems-30-2017 (2017).
von Ellenrieder, N., Dubeau, F., Gotman, J., & Frauscher, B. Physiological and pathological high-
frequency oscillations have distinct sleep-homeostatic properties. NeuroImage: Clinical, 14, 566 -- 573.
http://doi.org/10.1016/j.nicl.2017.02.018 (2017).
Wass, C.T., Grady, R.E., Fessler, A.J., Cascino, G.D., Lozada, L., Bechtle, P.S., Marsh, W.R.,
Sharbrough, F.W., and Schroeder, D.R. (2001). The effects of remifentanil on epileptiform discharges
during intraoperative electrocorticography in patients undergoing epilepsy surgery. Epilepsia 42, 1340-
1344.
Weiss, S. A., et al. Ictal high frequency oscillations distinguish two types of seizure territories in humans.
Brain, 136(12), 3796 -- 3808. http://doi.org/10.1093/brain/awt276 (2013).
Weiss, S. A., et al. Seizure localization using ictal phase-locked high gamma: A retrospective surgical
outcome study. Neurology, 84(23), 2320 -- 2328. http://doi.org/10.1212/WNL.0000000000001656 (2015).
Weiss, S.A., Alvarado‐Rojas, C., Bragin, A., Behnke, E., Fields, T., Fried, I., Engel, J., & Staba, R. Ictal
onset patterns of local field potentials, high frequency oscillations, and unit activity in human mesial
temporal lobe epilepsy. Epilepsia, 57(1), 111 -- 121. (2016).
Wellmer, J., von der Groeben, F., Klarmann, U., Weber, C., Elger, C. E., Urbach, H., Clusmann, H., von
Lehe, M. Risks and benefits of invasive epilepsy surgery workup with implanted subdural and depth
electrodes. Epilepsia, 53, 1322 -- 1332. (2012).
Worrell, G. A., et al. High-frequency oscillations and seizure generation in neocortical epilepsy. Brain,
127(7), 1496 -- 1506. http://doi.org/10.1093/brain/awh149 (2004).
Worrell, G. A., et al. Recording and analysis techniques for high-frequency oscillations. Progress in
Neurobiology, 98(3), 265 -- 278. http://doi.org/10.1016/j.pneurobio.2012.02.006 (2012).
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
Worrell, G., & Gotman, J. High-frequency oscillations and other electrophysiological biomarkers of
epilepsy: Clinical studies. Biomarkers in Medicine, 5(5), 557 -- 566. http://doi.org/10.2217/bmm.11.74
(2011).
Wu, J. Y., et al. Removing interictal fast ripples on electrocorticography linked with seizure freedom in
children. Neurology, 75(19), 1686 -- 1694. http://doi.org/10.1212/WNL.0b013e3181fc27d0 (2010).
Ylinen, A., et al. Sharp wave-associated high-frequency oscillation (200 Hz) in the intact hippocampus:
intracellular mechanisms. The
Network
30 -- 46.
http://www.jneurosci.org/content/15/1/30.long (1995).
Zijlmans, M., et al. High-frequency oscillations as a new biomarker in epilepsy. Annals of Neurology,
71(2), 169 -- 178. http://doi.org/10.1002/ana.22548 (2012).
of Neuroscience,
and
Journal
15(1),
This manuscript has been published in the Journal of Neural Engineering. DOI: 10.1088/1741-2552/aac960
|
1501.00727 | 3 | 1501 | 2015-01-12T10:07:48 | Graph Theoretical Analysis Reveals: Women's Brains are Better Connected than Men's | [
"q-bio.NC"
] | Deep graph-theoretic ideas in the context with the graph of the World Wide Web led to the definition of Google's PageRank and the subsequent rise of the most-popular search engine to date. Brain graphs, or connectomes, are being widely explored today. We believe that non-trivial graph theoretic concepts, similarly as it happened in the case of the World Wide web, will lead to discoveries enlightening the structural and also the functional details of the animal and human brains. When scientists examine large networks of tens or hundreds of millions of vertices, only fast algorithms can be applied because of the size constraints. In the case of diffusion MRI-based structural human brain imaging, the effective vertex number of the connectomes, or brain graphs derived from the data is on the scale of several hundred today. That size facilitates applying strict mathematical graph algorithms even for some hard-to-compute (or NP-hard) quantities like vertex cover or balanced minimum cut. In the present work we have examined brain graphs, computed from the data of the Human Connectome Project, recorded from male and female subjects between ages 22 and 35. Significant differences were found between the male and female structural brain graphs: we show that the average female connectome has more edges, is a better expander graph, has larger minimal bisection width, and has more spanning trees than the average male connectome. Since the average female brain weights less than the brain of males, these properties show that the female brain is more "well-connected" or perhaps, more "efficient" in a sense than the brain of males. | q-bio.NC | q-bio |
Graph Theoretical Analysis Reveals: Women's Brains
are Better Connected than Men's
Bal´azs Szalkaia, B´alint Vargaa, Vince Grolmusza,b,∗
aPIT Bioinformatics Group, Eotvos University, H-1117 Budapest, Hungary
bUratim Ltd., H-1118 Budapest, Hungary
Abstract
Deep graph-theoretic ideas in the context with the graph of the World Wide Web
led to the definition of Google's PageRank and the subsequent rise of the most-
popular search engine to date. Brain graphs, or connectomes, are being widely
explored today. We believe that non-trivial graph theoretic concepts, similarly
as it happened in the case of the World Wide web, will lead to discoveries
enlightening the structural and also the functional details of the animal and
human brains. When scientists examine large networks of tens or hundreds
of millions of vertices, only fast algorithms can be applied because of the size
constraints. In the case of diffusion MRI-based structural human brain imaging,
the effective vertex number of the connectomes, or brain graphs derived from
the data is on the scale of several hundred today. That size facilitates applying
strict mathematical graph algorithms even for some hard-to-compute (or NP-
hard) quantities like vertex cover or balanced minimum cut.
In the present work we have examined brain graphs, computed from the data
of the Human Connectome Project, recorded from male and female subjects
between ages 22 and 35. Significant differences were found between the male
and female structural brain graphs: we show that the average female connectome
has more edges, is a better expander graph, has larger minimal bisection width,
and has more spanning trees than the average male connectome. Since the
average female brain weights less than the brain of males, these properties show
that the female brain is more "well-connected" or perhaps, more "efficient" in
a sense than the brain of males. It is known that the female brain has a larger
white matter/gray matter ratio than the brain of males; this observation is in
line with our findings concerning the number of edges, since the white matter
consists of myelinated axons, which, in turn, correspond to the connections
in the brain graph. We have also found that the minimum bisection width,
normalized with the edge number, is also significantly larger in the right and the
left hemispheres in females: therefore, that structural difference is independent
from the difference in the number of edges.
∗Corresponding author
Email addresses: [email protected] (Bal´azs Szalkai), [email protected] (B´alint
Varga), [email protected] (Vince Grolmusz)
1
1. Introduction
In the last several years hundreds of publications appeared describing or
analyzing structural or functional networks of the brain, frequently referred to
as "connectome" [1, 2]. Some of these publications analyzed data from healthy
humans [3, 4, 5, 6], and some compared the connectome of the healthy brain
with diseased one [7, 8, 9, 10, 11].
So far, the analyses of the connectomes mostly used tools developed for
very large networks, such as the graph of the World Wide Web (with billions of
vertices), or protein-protein interaction networks (with tens or hundreds of thou-
sands of vertices), and because of the huge size of original networks, these meth-
ods used only very fast algorithms and frequently just primary degree statistics
and graph-edge counting between pre-defined regions or lobes of the brain [12].
In the present work we demonstrate that deep and more intricate graph the-
oretic parameters could also be computed by using, among other tools, contem-
porary integer programming approaches for connectomes with several hundred
vertices.
With these mathematical tools we show statistically significant differences in
some graph properties of the connectomes, computed from MRI imaging data of
male and female brains. We will not try to associate behavioral patterns of males
and females with the discovered structural differences [12] (see also the debate
that article has generated: [13, 14, 15]), because we do not have behavioral data
of the subjects of the imaging study, and, additionally, we cannot describe high-
level functional properties implied by those structural differences. However,
we clearly demonstrate that deep graph-theoretic parameters show "better"
connections in a certain sense in female connectomes than in male ones.
The study of [12] analyzed the 95-vertex graphs of 949 subjects aged be-
tween 8 and 22 years, using basic statistics for the numbers of edges running
either between or within different lobes of the brain (the parameters deduced
were called hemispheric connectivity ratio, modularity, transitivity and partic-
ipation coefficients, see [12] for the definitions). It was found that males have
significantly more intra-hemispheric edges than females, while females have sig-
nificantly more inter-hemispheric edges than males.
2. Results and Discussion
We have analyzed the connectomes of 96 subjects, 52 females and 44
males, each with 83, 129 and 234 node resolutions, and each graphs with
five different weight functions. We considered the connectomes as graphs with
weighted edges, and performed graph-theoretic analyses with computing some
polynomial-time computable and also some NP-hard graph parameters on the
individual graphs, and then compared the results statistically for the male and
the female group.
2
We have found that female connectomes have more edges, larger (normal-
ized) minimum bisection widths, larger minimum-vertex covers and more span-
ning trees than the male connectomes.
In order to describe the parameters, which differ significantly among male
and female connectomes, we need to place them in the context of their graph
theoretical definitions.
2.1. Edge number and edge weights
We have found significantly higher number of edges (counted with 5 types
of weights and also without any weights) in both hemispheres and also in the
whole brain in females, in all resolutions. This finding is surprising, since we
used the same parcellation and the same tractography and the same graph-
construction methods for female and male brains, and because it is proven that
females have, on average, less-weighting brains than males [16]. For example, in
the 234-vertex resolution, the average number of (unweighted) edges in female
connectomes is 1826, in males 1742, with p = 0.00063 (see the Appendix for
tables with the results). The work of [12] reported similar findings in inter-
hemispheric connections only.
It is known that there are statistical differences in the size and the weight
of the female and the male cerebra [16]. It was also published [17] that female
brains statistically have a higher white matter/gray matter ratio than male
brains. We argue that this observation is in line with the quantitative differences
in the fibers and edges in the connectomes of the sexes.
In a simplified view, the edges of the braingraph correspond to the fibers of
the myelinated axons in the white matter, while the nodes of the graph to areas
of the gray matter. Therefore, since females have a higher white matter/gray
matter ratio than males by [17] that fact implies that the number of detected
fibers by the tractography step of the processing is relatively higher in females
than in males, and this higher number of fibers imply higher number of edges
in female connectomes.
2.2. Minimum cut and balanced minimum cut
Suppose the nodes, or the vertices, of a graph are partitioned into two,
disjoint, non-empty sets, say X and Y ; their union is the whole vertex-set of
the graph. The X, Y cut is the set of all edges connecting vertices of X with the
vertices of Y (Figure 1A). The size of the cut is the number of edges in the cut.
In graph theory, the size of the minimum cut is an interesting quantity. The
minimum cut between vertices a and b is the minimum cut, taken for all X and
Y , where vertex a is in X and b is in Y . This quantity gives the "bottleneck", in
a sense, between those two nodes (c.f., Menger theorems and Ford-Fulkerson's
Min-Cut-Max-Flow theorem [18, 19]). The minimum cut in a graph is defined
to be the cut with the fewest edges for all non-empty sets X and Y , partitioning
the vertices.
Clearly, for non-negative weights, the size of the minimum cut in a non-
connected graph is 0. Very frequently, however, in connected graphs, the mini-
mum cut is determined by just the smallest degree node: that node is the only
3
element of set X and all the other vertices of the graph are in Y (Figure 1B).
Because of this phenomenon, the minimum cut is frequently queried for the
"balanced" case, when the size (i.e., the number of vertices) of X and Y needs
to be equal (or, more exactly, may differ by at most one if the number of the
vertices of the graph is odd), see Figure 1C. This problem is referred to as the
balanced minimum cut or the minimum bisection problem. If the minimum bi-
section is small that means that there exist a partition of the vertices into two
sets of equal size that are connected with only a few edges. If the minimum
bisection is large then the two half-sets in every possible bisections of the graph
are connected by many edges.
Therefore, the balanced minimum cut of a graph is independent of the par-
ticular labeling of the nodes. The number of all the balanced cuts in a graph
with n vertices is greater than
1
n + 1
2n,
that is, for n = 250, this number is very close to the number of atoms in
the visible universe [20]. Consequently, one cannot practically compute the
minimum bisecton width by reviewing all the bisectons in a graph of that size.
Moreover, the complexity of computing this quantity is known to be NP-hard
[21] in general, but with contemporary integral programming approaches, for
the graph-sizes we are dealing with, the exact values are computable.
In computer engineering, an important measure of the quality of an inter-
connection network is its minimum bisection width [22]: the higher the width
is the better the network.
For the whole brain graph, as it is anticipated, we have found that the
minimum balanced cut is almost exactly represents the edges crossing the corpus
callosum, connecting the two cerebral hemispheres.
We show that within both hemispheres, the minimum bisection size of female
connectomes are significantly larger than the minimum bisection size of the
males. Much more importantly, we show that this remains true if we normalize
with the sum of all edge-weights: that is, this phenomenon cannot be due to the
higher number of edges or the greater edge weights in the female brain: it is an
intrinsic property of the female brain graph in our data analyzed.
For example, in the 234-vertex resolution, in the left hemisphere, the nor-
malized balanced minimum cut in females, on the average, is 0.09416, in the
males 0.07896, p = 0.00153 (see the Appendix for tables with the results).
We think that this finding is one of the main results of the present work:
even if the significant difference in the weighted edge numbers are due to some
artifacts in the data acquisition/processing workflow, the normalized balanced
minimum cut size seems to be independent from those processes.
2.3. Eigengap and the expander property
Expander graphs and the expander-property of graphs are one of the most
interesting area of graph theory: they are closely related to the convergence
rate and the ergodicity of Markov chains, and have applications in the design
4
Figure 1: Panel A: An X-Y cut. The cut-edges are colored black. Panel B: An un-balanced
minimum cut. Panel C: A balanced cut. Panel D: The wheel graph.
of communication- and sorting networks and methods for de-randomizing algo-
rithms [23]. A graph is an ε-expander, if every – not too small and not too
large – vertex-set S of the graph has at least εS outgoing edges (see [23] for
the exact definition).
Random walks on good expander graphs converge very fast to the limit
distribution: this means that good expander graphs, in a certain sense, are
"intrinsically better" connected than bad expanders.
It is known that large
eigengap of the walk transition matrix of the graph implies good expansion
property [23].
We have found that women's connectomes have significantly larger eigengap,
and, consequently, they are better expander graphs than the connectomes of
men. For example, in the 83-node resolution, in the left hemisphere and in the
unweighted graph, the average female connectome's eigengap is 0.306 while in
the case of men it is 0.272, with p = 0.00458.
2.4. The number of spanning forests
A tree in graph theory is a connected, cycle-free graph. Any tree on n
vertices has the same number of edges: n − 1. Trees, and tree-based structures
are common in science: phylogenetic trees, hierarchical clusters, data-storage
on hard-disks, or a computational model called decision trees all apply graph-
theoretic trees. A spanning tree is a minimal subgraph of a connected graph that
is still connected. Some graphs have no spanning trees at all: only connected
5
graphs have spanning trees. A tree has only one spanning tree:
itself. Any
connected graph on n vertices has a minimum of n− 1 and a maximum of n(n−
1)/2 edges [24]. A connected graph with few edges still may have exponentially
many different spanning trees: e.g., the n-vertex wheel on Figure 1D has at least
2n−1 spanning trees (for n ≥ 4). Cayley's famous theorem, and its celebrated
proof with Prufer codes [25] shows that the number of spanning trees of the
complete graph on n vertices is nn−2.
If a graph is not connected, then it contains more than one connected com-
ponents. Each connected component has at least one spanning tree, and the
whole graph has at least one spanning forest, comprising of the spanning trees
of the components. The number of spanning forests is clearly the product of
the numbers of the spanning trees of the components.
For graphs in general, one can compute the number of their spanning forests
by Kirchoff's matrix tree theorem [26, 27] using the eigenvalues of the Laplacian
matrix [27] of the graph.
We show that female connectomes have significantly higher number of span-
ning trees than the connectomes of males. For example, in the 129-vertex resolu-
tion, in the left hemisphere, the logarithm of the number of the spanning forests
in the unweighted case are 162.01 in females, 158.88 in males with p = 0.013.
3. Materials and Methods
3.1. Data source and graph computation:
The dataset applied is a subset of the Human Connectome Project [28]
anonymized 500 Subjects Release:
(http://www.humanconnectome.org/documentation/S500) of healthy subjects
between 22 and 35 years of age. Data was downloaded in October, 2014. The
Connectome Mapper Toolkit [29] (http://cmtk.org) was applied for brain tissue
segmentation into grey and white matter, partitioning, tractography and the
construction of the graphs from the fibers identified in the tractography step.
The Connectome Mapper Toolkit [29] default partitioning was used (computed
by the FreeSurfer, and based on the Desikan-Killiany anatomical atlas) into 83,
129 and 234 cortical and sub-cortical structures (as the brainstem and deep-
grey nuclei), referred to as "Regions of Interest", ROIs, (see Figure 4 in [29]).
Tractography was performed by choosing the deterministic streamline method
[29] with randomized seeding.
The graphs were constructed as follows: the nodes correspond to the ROIs
in the specific resolution. Two nodes were connected by an edge if there exists
at least one fiber (determined by the tractography step) connecting the ROIs,
corresponding to the nodes. More than one fibers, connecting the same nodes,
may give rise to the weight of that edge, depending on the weighting method.
Loops were deleted from the graph.
The weights of the edges are assigned by several methods, taking into account
the lengths and the multiplicities of the fibers, connecting the nodes:
• Unweighted: Each edge has weight 1.
6
• FiberN: The number of fibers traced along the edge: this number is larger
than one if more than one fibers connect two cortical or sub-cortical areas,
corresponding to the two endpoints of the edge.
• FAMean: The arithmetic mean of the fractional anisotropies [30] of the
fibers, belonging to the edge.
• FiberLengthMean: The average length of the fibers, connecting the two
endpoints of the edge.
• FiberNDivLength: The number of fibers belonging to the edge, divided
by their average length. This quantity is related to the simple electrical
model of the nerve fibers: by modeling the fibers as electrical resistors
with resistances proportional to the average fiber length, this quantity
is precisely the conductance between the two regions of interest. Addi-
tionally, FiberNDivLength can be observed as a reliability measure of the
edge:
longer fibers are less reliable than the shorter ones, due to possi-
ble error accumulation in the tractography algorithm that constructs the
fibers from the anisotropy data. Multiple fibers connecting the same two
ROIs, corresponding to the endpoints, add to the reliability of the edge,
because of the independently tractographed connections.
By generalized adjacency matrix we mean a matrix of size n × n where n
is the number of nodes (or vertices) in the graph, whose rows and columns
correspond to the nodes, and whose each element is either zero if there is no
edge between the two nodes, or equals to the weight of the edge connecting the
two nodes. By the generalized degree of a node we mean the sum of the weights
of the edges adjacent to that node. Note that the generalized degree of the node
v is exactly the sum of the elements in the row (or column) of the generalized
adjacency matrix corresponding to v. By generalized Laplacian matrix we mean
the matrix D − A, where D is a diagonal matrix containing the generalized
degrees, and A is the generalized adjacency matrix.
3.2. Graph parameters:
We calculated various graph parameters for each brain graph and weight
function. These parameters included:
• Number of edges (Sum). The weighted version of this quantity is the sum
of the weights of the edges.
• Normalized largest eigenvalue (AdjLMaxDivD): The largest eigenvalue of
the generalized adjacency matrix, divided by the average degree. Divid-
ing by the average degree of vertices was necessary because the largest
eigenvalue is bounded by the average- and maximum degrees, and thus is
considered by some a kind of "average degree" itself [24]. This means that
a denser graph may have a bigger λmax largest eigenvalue solely because
of a larger average degree. We note that the average degree is already
defined by the sum of weights.
7
• Eigengap of the transition matrix (PGEigengap): The transition matrix
PG is obtained by dividing all the rows of the generalized adjacency matrix
by the generalized degree of the corresponding node. When performing a
random walk on the graph, for nodes i and j, the corresponding matrix
element describes the probability of transitioning to node j, supposing
that we are at node i. The eigengap of a matrix is the difference of
the largest and the second largest eigenvalue. It is characteristic to the
expander properties of the graph: the larger the gap, the better expander
is the graph (see [23] for the exact statements and proofs).
• Hoffman's bound (HoffmanBound): The expression
1 +
λmax
λmin ,
where λmax and λmin denote the largest and smallest eigenvalues of the
adjacency matrix. It is a lower bound for the chromatic number of the
graph. The chromatic number is generally higher for denser graphs, as
the addition of an edge may make a previously valid coloring invalid.
• Logarithm of number of spanning forests (LogAbsSpanningForestN): The
number of the spanning trees in a connected graph can be calculated
from the spectrum of its Laplacian [26, 27]. Denser graphs tend to have
more spanning trees, as the addition of an edge introduces zero or more
new spanning trees.
If a graph is not connected, then the number of
spanning forests is the product of the numbers of the spanning trees of
the components. The parameter LogAbsSpanningForestN equals to the
logarithm of the number of spanning forests in the unweighted case. In
the case of other weight functions, if we define the weight of a tree by the
product of the weights of its edges, then this parameter equals to the sum
of the logarithms of the weights of the spanning trees in the forests.
• Balanced
divided
by
the
cut,
minimum
of
edges (MinCutBalDivSum): The task is to partition the graph into two
sets whose size may differ from each other by at most 1, so that the num-
ber of edges crossing the cut is minimal. This is the "balanced minimum
cut" problem, or sometimes called the "minimum bisection width" prob-
lem. For the whole brain graph, our expectation was that the minimum
cut corresponds to the boundary of the two hemispheres, which was indeed
proven when we analyzed the results.
number
• Minimum cost spanning tree (MinSpanningForest), calculated with
Kruskal's algorithm.
• Minimum weighted vertex cover (MinVertexCover): Each vertex should
have a (possibly fractional) weight assigned such that, for each edge, the
sum of the weights of its two endpoints is at least 1. This is the fractional
relaxation of the NP-hard vertex-cover problem [31]. The minimum of the
sum of all vertex-weights is computable by a linear programming approach.
8
• Minimum vertex cover (MinVertexCoverBinary): Same as above, but
each weight must be 0 or 1. In other words, a minimum size set of vertices
is selected such that each edge is covered by at least one of the selected ver-
tices. This NP-hard graph-parameter is computed only for the unweighted
case. The exact values are computed by an integer programming solver
SCIP (http://scip.zib.de), [32, 33].
The above 9 parameters were computed for all three resolutions and for the
left and the right hemispheres and also for the whole connectome, with all 5
weight functions (with the following exceptions: MinVertexCoverBinary was
computed only for the unweighted case, and the MinSpanningForest was not
computed for the unweighted case).
3.3. Statistical analysis
Since each connectome was computed in multiple resolutions (in 83, 129 and
234 nodes), we had three graphs for each brain. In addition, the parameters
were calculated separately for the connectome within the left and right hemi-
spheres as well, not only the whole graph, since we intended to examine whether
statistically significant differences can be attributed to the left or right hemi-
spheres. Each subjects' brain was corresponded to 9 graphs (3 resolutions, each
in the left and the right hemispheres, plus the whole cortex with sub-cortical
areas) and for each graph we calculated 9 parameters, each (with the excep-
tions noted above) with 5 different edge weights. This means that we assigned
7 · 5 · 3 + 1 · 3 + 4 · 3 = 120 attributes to each resolution of the 96 brains, that
is, 360 attributes to each brain.
The statistical null hypothesis [34] of ours was that the graph parameters do
not differ between the male and the female groups. As the first approach, we
have used ANOVA (Analysis of variance) [35] to assign p-values for all parame-
ters in each hemispheres and in each resolutions and in each weight-assignments.
Our very large number of attributes may lead to false negatives, i.e., to "type
II" statistical errors: in other words, it may happen that an attribute, with a
very small p-value may appear "at random", simply because we tested a lot of
attributes. In order to deal with "type II" statistical errors, we followed the
route described below.
We divided the population randomly into two sets by the parity of the sum
of the digits in their ID. The first set was used for making hypotheses and the
second set for testing these hypotheses. This was necessary to avoid type II
errors resulting from multiple testing correction. If we made hypotheses for all
the numerical parameters, then the Holm-Bonferroni correction [36] we used
would have unnecessarily increased the p-values. Thus we needed to filter the
hypotheses first, and that is why we needed the first set. Testing on the first
set allowed us to reduce the number of hypotheses and test only a few of them
on the second set.
The hypotheses were filtered by performing ANOVA (Analysis of variance)
[35] on the first set. Only those hypotheses were selected to qualify for the
second round where the p-value was less than 1%. The selected hypotheses
9
were then tested for the second set as well, and the resulting p-value corrected
with the Holm-Bonferroni correction method [36] with a significance level of 5%.
In Table 1 those hypotheses rejected were highlighted in bold, meaning that
all the corresponding graph parameters differ significantly in sex groups at a
combined significance level of 5%.
We also highlighted (in italic) those p-values which were individually less
than the threshold, meaning that these hypotheses can individually be rejected
at a level of 5%, but it is very likely that not all of these graph parameters are
significantly different between the sexes.
4. Conclusions:
We have computed 83-, 129- and 234-vertex-graphs from the diffusion MRI
images of the 96 subjects of 52 females and 44 males, between the age of 22
and 35. We have found, after a careful statistical analysis, significant differ-
ences between some graph theoretical parameters of the male and female brain
graphs. Our findings show that the female brain graphs have generally more
edges (counted with and without weights), have larger normalized minimum bi-
section widths and have more spanning trees (counted with and without weights)
than the connectomes of males (Table 1). Additionally, with weaker statistical
validity, some spectral properties and the minimum vertex cover also differ in
the connectomes of different sexes (each with p < 0.02).
5. Data availability:
The unprocessed and pre-processed MRI data is available at the Human
Connectome Project's website:
http://www.humanconnectome.org/documentation/S500 [28].
5.1. Table 1
Scale
129
83
234
129
83
83
129
234
83
129
83
83
234
234
129
Property
Right MinCutBalDivSum FAMean
All LogSpanningForestN FiberNDivLength
All PGEigengap FiberNDivLength
All PGEigengap FiberNDivLength
Left MinCutBalDivSum FiberN
Right MinCutBalDivSum FAMean
Left PGEigengap FiberNDivLength
All PGEigengap FiberN
All Sum Unweighted
Left MinCutBalDivSum FiberN
All LogSpanningForestN FiberN
Right Sum FAMean
All Sum Unweighted
Left PGEigengap FiberNDivLength
All Sum Unweighted
10
p (1st)
0.00807
0.00003
0.00321
0.00792
0.00403
0.00496
0.00223
0.00826
0.00025
0.00001
0.00001
0.00028
0.00063
0.00013
0.00026
p (2nd)
0.00003
0.00004
0.00007
0.00011
0.00011
0.00015
0.00015
0.00022
0.00022
0.00023
0.00028
0.00029
0.00032
0.00038
0.00042
p (corrected)
0.00401
0.00451
0.00798
0.01303
0.01300
0.01744
0.01797
0.02517
0.02504
0.02563
0.03084
0.03224
0.03512
0.04171
0.04563
234
129
83
129
234
83
234
234
129
234
83
83
234
83
83
129
83
129
234
234
129
83
129
129
129
129
129
234
83
234
83
129
83
234
234
234
234
83
83
129
83
234
83
234
234
129
83
234
234
129
All Sum FAMean
All LogSpanningForestN FiberN
All Sum FAMean
Right Sum FAMean
Right PGEigengap FiberNDivLength
Left Sum Unweighted
Right Sum FAMean
Left Sum Unweighted
All Sum FAMean
Left MinCutBalDivSum FiberN
Left LogSpanningForestN FiberNDivLength
All LogSpanningForestN Unweighted
Left MinCutBalDivSum FiberLengthMean
All LogSpanningForestN FAMean
Right Sum Unweighted
Left MinCutBalDivSum Unweighted
Left MinCutBalDivSum Unweighted
Left PGEigengap FiberN
All LogSpanningForestN FAMean
Left PGEigengap FiberN
Right LogSpanningForestN FAMean
Left MinCutBalDivSum FiberNDivLength
All LogSpanningForestN FiberNDivLength
All LogSpanningForestN Unweighted
Right Sum Unweighted
Left PGEigengap FAMean
All LogSpanningForestN FAMean
Left Sum FAMean
Right LogSpanningForestN FAMean
Left PGEigengap FAMean
Left PGEigengap Unweighted
Left MinCutBalDivSum FiberLengthMean
Left Sum FAMean
Left MinCutBalDivSum Unweighted
Left PGEigengap FiberLengthMean
Right LogSpanningForestN FAMean
Left PGEigengap Unweighted
Left PGEigengap FAMean
Left LogSpanningForestN FiberN
Left Sum Unweighted
Left LogSpanningForestN FAMean
Right MinCutBalDivSum Unweighted
Right LogSpanningForestN FiberNDivLength
Left LogSpanningForestN FAMean
Right PGEigengap Unweighted
Left Sum FAMean
Left AdjLMaxDivD FiberN
Right Sum Unweighted
Right PGEigengap FiberN
All Sum FiberN
11
0.00014
0.00000
0.00029
0.00062
0.00041
0.00378
0.00085
0.00293
0.00015
0.00002
0.00343
0.00113
0.00411
0.00012
0.00019
0.00265
0.00206
0.00382
0.00043
0.00066
0.00143
0.00031
0.00000
0.00218
0.00068
0.00995
0.00019
0.00026
0.00067
0.00141
0.00458
0.00892
0.00056
0.00154
0.00554
0.00380
0.00176
0.00215
0.00012
0.00232
0.00082
0.00462
0.00022
0.000129
0.00095
0.00032
0.00501
0.00224
0.00009
0.00000
0.00047
0.00048
0.00050
0.00051
0.00053
0.00068
0.00084
0.00092
0.00097
0.00108
0.00116
0.00121
0.00123
0.00126
0.00128
0.00134
0.00136
0.00142
0.00150
0.00163
0.00170
0.00175
0.00177
0.00182
0.00186
0.00191
0.00211
0.00212
0.00239
0.00240
0.00243
0.00245
0.00279
0.00289
0.00295
0.00305
0.00338
0.00359
0.00395
0.00456
0.00496
0.00543
0.00587
0.00595
0.00626
0.00660
0.00804
0.00845
0.00910
0.00938
0.04988
0.05045
0.05260
0.05355
0.05414
0.06936
0.08454
0.09212
0.09650
0.10539
0.11274
0.11629
0.11646
0.11823
0.11891
0.12351
0.12370
0.12775
0.13343
0.14369
0.14769
0.15023
0.15009
0.15279
0.15417
0.15694
0.17093
0.16978
0.18842
0.18684
0.18738
0.18596
0.20893
0.21355
0.21516
0.21935
0.24029
0.25152
0.27269
0.31006
0.33212
0.35825
0.38180
0.38054
0.39459
0.40907
0.49040
0.50692
0.53671
0.54418
234
129
83
129
129
129
129
83
234
83
129
129
129
234
83
83
234
83
234
83
129
83
234
234
129
129
129
234
129
83
83
83
234
234
234
83
83
129
129
129
83
234
129
234
83
129
234
83
83
234
Right PGEigengap FAMean
Right PGEigengap FAMean
Right PGEigengap Unweighted
Right MinCutBalDivSum FiberN
Right MinCutBalDivSum Unweighted
Left LogSpanningForestN FAMean
Left LogSpanningForestN FiberN
All Sum FiberN
All Sum FiberN
Right LogSpanningForestN Unweighted
Left LogSpanningForestN FiberNDivLength
Right PGEigengap Unweighted
Right PGEigengap FiberN
All MinVertexCover FAMean
All HoffmanBound FAMean
All Sum FiberNDivLength
Right MinCutBalDivSum FiberN
Right LogSpanningForestN FiberN
Right MinCutBalDivSum FiberLengthMean
Right MinCutBalDivSum FiberNDivLength
Left MinCutBalDivSum FiberNDivLength
Right PGEigengap FAMean
All LogSpanningForestN FiberN
Right PGEigengap FiberLengthMean
Right MinCutBalDivSum FiberLengthMean
All Sum FiberNDivLength
Right LogSpanningForestN FiberNDivLength
All LogSpanningForestN FiberNDivLength
Right LogSpanningForestN FiberN
Right MinCutBalDivSum FiberN
Right HoffmanBound FiberNDivLength
Right PGEigengap FiberLengthMean
Left MinCutBalDivSum FiberNDivLength
Left MinVertexCover FAMean
All Sum FiberNDivLength
Right Sum FiberN
Right Sum FiberNDivLength
Left Sum FiberN
Right Sum FiberN
Left HoffmanBound Unweighted
Left Sum FiberN
Left Sum FiberN
Right Sum FiberNDivLength
Right Sum FiberN
Left Sum FiberNDivLength
Left Sum FiberNDivLength
Right Sum FiberNDivLength
Right HoffmanBound FAMean
All MinVertexCoverBinary Unweighted
Right LogSpanningForestN FiberNDivLength
12
0.00074
0.00296
0.00087
0.00563
0.00492
0.00106
0.00014
0.00000
0.00000
0.00541
0.00288
0.00242
0.00869
0.00289
0.00087
0.00002
0.00234
0.00083
0.00234
0.00072
0.00019
0.00112
0.00091
0.00367
0.00768
0.00008
0.00051
0.00106
0.00045
0.00346
0.005129
0.00949
0.00642
0.00107
0.00044
0.00000
0.00018
0.00000
0.00001
0.00848
0.00000
0.00040
0.00180
0.00012
0.00043
0.00100
0.00562
0.00587
0.00716
0.00940
0.00974
0.00981
0.01053
0.01101
0.01212
0.01218
0.01258
0.01290
0.01358
0.01438
0.01447
0.01676
0.01706
0.01713
0.02011
0.02117
0.02197
0.02539
0.02663
0.02854
0.02897
0.02948
0.03308
0.03369
0.04500
0.04728
0.04891
0.05095
0.05578
0.06284
0.06309
0.06515
0.06548
0.07139
0.07318
0.07799
0.07920
0.08380
0.08653
0.08944
0.09430
0.11447
0.12102
0.16411
0.16774
0.22542
0.23691
0.32069
0.38829
0.40996
0.55538
0.54933
0.57889
0.59432
0.64227
0.63359
0.64134
0.64480
0.66520
0.69010
0.67995
0.77084
0.76750
0.75373
0.86462
0.88929
0.90065
1.01567
1.03841
1.08446
1.07195
1.06119
1.15795
1.14542
1.48511
1.51293
1.51627
1.52842
1.61751
1.75951
1.70341
1.69395
1.63696
1.71336
1.68305
1.71586
1.66329
1.67598
1.64406
1.60984
1.60310
1.83157
1.81523
2.29752
2.18062
2.70502
2.60604
3.20692
3.49459
3.27971
83
83
83
234
129
83
234
Left HoffmanBound FiberN
All MinVertexCover FiberNDivLength
Right MinSpanningForest FiberLengthMean
Right MinSpanningForest FiberLengthMean
All MinVertexCover FiberN
All MinVertexCover FiberN
All MinVertexCover FiberN
0.00175
0.00036
0.00491
0.00601
0.00232
0.00244
0.00055
0.41913
0.46677
0.55239
0.55631
0.71406
0.84437
0.92958
2.93394
2.80065
2.76195
2.22523
2.14217
1.68874
0.92958
Table 1: The results and the statistical analysis of the graph-theoretical
evaluation of the sex differences in the 96 diffusion MRI images. The first
column gives the resolution in each hemisphere; the number of nodes in
the whole graph is 83, 129 and 234, respectively. The second column
describes the graph parameter computed:
its syntactics is as follows:
each parameter-name contains two separating " " symbols that define
three parts of the parameter-name. The first part describe the hemi-
sphere or the whole connectome with the words Left, Right or All. The
second part describes the parameter computed, and the third part the
weight function used (their definitions are given in section "Materials and
methods"). The third column contains the p-values of the first round,
the second column the p-values of the second round, and the third col-
umn the (very strict) Holm-Bonferroni correction of the p-value. With
p=0.05 all the first 12 rows describe significantly different graph theo-
retical properties between sexes. One-by-one, each row with italic third
column describe significant differences between sexes, with p=0.05. For
the details we refer to the section "Statistical analysis".
6. Acknowledgments
The authors declare no conflicts of interest.
References
[1] P. Hagmann, P. E. Grant, D. A. Fair, Mr connectomics: a conceptual
framework for studying the developing brain., Front Syst Neurosci 6 (2012)
43. doi:10.3389/fnsys.2012.00043.
URL http://dx.doi.org/10.3389/fnsys.2012.00043
[2] R. C. Craddock, M. P. Milham, S. M. LaConte, Predicting intrinsic brain
activity., Neuroimage 82 (2013) 127–136. doi:10.1016/j.neuroimage.
2013.05.072.
URL http://dx.doi.org/10.1016/j.neuroimage.2013.05.072
[3] G. Ball, P. Aljabar, S. Zebari, N. Tusor, T. Arichi, N. Merchant, E. C.
Robinson, E. Ogundipe, D. Rueckert, A. D. Edwards, S. J. Counsell, Rich-
club organization of the newborn human brain., Proc Natl Acad Sci U S A
111 (20) (2014) 7456–7461. doi:10.1073/pnas.1324118111.
URL http://dx.doi.org/10.1073/pnas.1324118111
13
[4] C. I. Bargmann, Beyond the connectome: how neuromodulators shape
neural circuits., Bioessays 34 (6) (2012) 458–465. doi:10.1002/bies.
201100185.
URL http://dx.doi.org/10.1002/bies.201100185
[5] D. Batalle, E. Munoz-Moreno, F. Figueras, N. Bargallo, E. Eixarch,
E. Gratacos, Normalization of similarity-based individual brain networks
from gray matter MRI and its association with neurodevelopment in in-
fants with intrauterine growth restriction., Neuroimage 83 (2013) 901–911.
doi:10.1016/j.neuroimage.2013.07.045.
URL http://dx.doi.org/10.1016/j.neuroimage.2013.07.045
[6] D. J. Graham, Routing in the brain., Front Comput Neurosci 8 (2014) 44.
doi:10.3389/fncom.2014.00044.
URL http://dx.doi.org/10.3389/fncom.2014.00044
[7] F. Agosta, S. Galantucci, P. Valsasina, E. Canu, A. Meani, A. Marcone,
G. Magnani, A. Falini, G. Comi, M. Filippi, Disrupted brain connectome
in semantic variant of primary progressive aphasia., Neurobiol Agingdoi:
10.1016/j.neurobiolaging.2014.05.017.
URL http://dx.doi.org/10.1016/j.neurobiolaging.2014.05.017
[8] A. F. Alexander-Bloch, P. T. Reiss, J. Rapoport, H. McAdams, J. N. Giedd,
E. T. Bullmore, N. Gogtay, Abnormal cortical growth in schizophrenia
targets normative modules of synchronized development., Biol Psychia-
trydoi:10.1016/j.biopsych.2014.02.010.
URL http://dx.doi.org/10.1016/j.biopsych.2014.02.010
[9] J. T. Baker, A. J. Holmes, G. A. Masters, B. T. T. Yeo, F. Krienen,
R. L. Buckner, D. Ongur, Disruption of cortical association networks in
schizophrenia and psychotic bipolar disorder., JAMA Psychiatry 71 (2)
(2014) 109–118. doi:10.1001/jamapsychiatry.2013.3469.
URL http://dx.doi.org/10.1001/jamapsychiatry.2013.3469
[10] P. Besson, V. Dinkelacker, R. Valabregue, L. Thivard, X. Leclerc,
M. Baulac, D. Sammler, O. Colliot, S. Leh´ericy, S. Samson, S. Dupont,
Structural connectivity differences in left and right temporal lobe epilepsy.,
Neuroimage 100C (2014) 135–144. doi:10.1016/j.neuroimage.2014.04.
071.
URL http://dx.doi.org/10.1016/j.neuroimage.2014.04.071
[11] L. Bonilha, T. Nesland, C. Rorden, P. Fillmore, R. P. Ratnayake,
J. Fridriksson, Mapping remote subcortical ramifications of injury after
ischemic strokes., Behav Neurol 2014 (2014) 215380. doi:10.1155/2014/
215380.
URL http://dx.doi.org/10.1155/2014/215380
14
[12] M. Ingalhalikar, A. Smith, D. Parker, T. D. Satterthwaite, M. A. Elliott,
K. Ruparel, H. Hakonarson, R. E. Gur, R. C. Gur, R. Verma, Sex differ-
ences in the structural connectome of the human brain., Proc Natl Acad
Sci U S A 111 (2) (2014) 823–828. doi:10.1073/pnas.1316909110.
URL http://dx.doi.org/10.1073/pnas.1316909110
[13] D. Joel, R. Tarrasch, On the mis-presentation and misinterpretation of
gender-related data: the case of Ingalhalikar's human connectome study.,
Proc Natl Acad Sci U S A 111 (6) (2014) E637. doi:10.1073/pnas.
1323319111.
URL http://dx.doi.org/10.1073/pnas.1323319111
[14] M. Ingalhalikar, A. Smith, D. Parker, T. D. Satterthwaite, M. A. Elliott,
K. Ruparel, H. Hakonarson, R. E. Gur, R. C. Gur, R. Verma, Reply to
Joel and Tarrasch: On misreading and shooting the messenger., Proc Natl
Acad Sci U S A 111 (6) (2014) E638.
[15] C. Fine, Neuroscience. his brain, her brain?, Science 346 (6212) (2014) 915–
916. doi:10.1126/science.1262061.
URL http://dx.doi.org/10.1126/science.1262061
[16] S. F. Witelson, H. Beresh, D. L. Kigar, Intelligence and brain size in 100
postmortem brains: sex, lateralization and age factors., Brain 129 (Pt 2)
(2006) 386–398. doi:10.1093/brain/awh696.
URL http://dx.doi.org/10.1093/brain/awh696
[17] Y. Taki, B. Thyreau, S. Kinomura, K. Sato, R. Goto, R. Kawashima,
H. Fukuda, Correlations among brain gray matter volumes, age, gender,
and hemisphere in healthy individuals., PLoS One 6 (7) (2011) e22734.
doi:10.1371/journal.pone.0022734.
URL http://dx.doi.org/10.1371/journal.pone.0022734
[18] E. L. Lawler, Combinatorial optimization: networks and matroids, Courier
Dover Publications, 1976.
[19] L. R. Ford, D. R. Fulkerson, Maximal flow through a network, Canadian
Journal of Mathematics 8 (3) (1956) 399–404.
[20] P. Ade, N. Aghanim, C. Armitage-Caplan, M. Arnaud, M. Ashdown,
F. Atrio-Barandela, J. Aumont, C. Baccigalupi, A. Banday, R. Barreiro,
et al., Planck 2013 results. XVI. Cosmological parameters, arXiv preprint
arXiv:1303.5076.
[21] M. R. Garey, D. S. Johnson, L. Stockmeyer, Some simplified NP-complete
graph problems, Theoretical computer science 1 (3) (1976) 237–267.
[22] R. E. Tarjan, Data structures and network algorithms, Vol. 44 of CBMS-
NSF Regional Conference Series in Applied Mathematics, Society for In-
dustrial Applied Mathematics, 1983.
15
[23] S. Hoory, N. Linial, A. Wigderson, Expander graphs and their applications,
Bulletin of the American Mathematical Society 43 (4) (2006) 439–561.
[24] L. Lov´asz, Combinatorial problems and exercises, 2nd Edition, American
Mathematical Society, 2007.
[25] H. Prufer, Neuer Beweis eines Satzes uber Permutationen, Arch. Math.
Phys 27 (1918) 742–744.
[26] G. Kirchhoff, uber die Auflosung der Gleichungen, auf welche man bei
der untersuchung der linearen verteilung galvanischer Strome gefuhrt wird,
Ann. Phys. Chem. 72.
[27] F. R. Chung, Spectral graph theory, Vol. 92, American Mathematical Soc.,
1997.
[28] J. A. McNab, B. L. Edlow, T. Witzel, S. Y. Huang, H. Bhat, K. Heberlein,
T. Feiweier, K. Liu, B. Keil, J. Cohen-Adad, M. D. Tisdall, R. D. Folkerth,
H. C. Kinney, L. L. Wald, The Human Connectome Project and beyond:
initial applications of 300 mT/m gradients., Neuroimage 80 (2013) 234–
245. doi:10.1016/j.neuroimage.2013.05.074.
URL http://dx.doi.org/10.1016/j.neuroimage.2013.05.074
[29] A. Daducci, S. Gerhard, A. Griffa, A. Lemkaddem, L. Cammoun, X. Gi-
gandet, R. Meuli, P. Hagmann, J.-P. Thiran, The connectome mapper: an
open-source processing pipeline to map connectomes with MRI., PLoS One
7 (12) (2012) e48121. doi:10.1371/journal.pone.0048121.
URL http://dx.doi.org/10.1371/journal.pone.0048121
[30] P. J. Basser, C. Pierpaoli, Microstructural and physiological features of
tissues elucidated by quantitative-diffusion-tensor mri. 1996., J Magn Reson
213 (2) (2011) 560–570. doi:10.1016/j.jmr.2011.09.022.
URL http://dx.doi.org/10.1016/j.jmr.2011.09.022
[31] D. S. Hochbaum, Approximation algorithms for the set covering and vertex
cover problems, SIAM Journal on Computing 11 (3) (1982) 555–556.
[32] T. Achterberg, T. Berthold, T. Koch, K. Wolter, Constraint integer pro-
gramming: A new approach to integrate CP and MIP, in: Integration of
AI and OR techniques in constraint programming for combinatorial opti-
mization problems, Springer, 2008, pp. 6–20.
[33] T. Achterberg, Scip: solving constraint integer programs, Mathematical
Programming Computation 1 (1) (2009) 1–41.
[34] P. G. Hoel, Introduction to mathematical statistics., 5th Edition, John
Wiley & Sons, Inc., New York, 1984.
[35] T. H. Wonnacott, R. J. Wonnacott, Introductory statistics, Vol. 19690,
Wiley New York, 1972.
16
[36] S. Holm, A simple sequentially rejective multiple test procedure, Scandi-
navian Journal of Statistics (1979) 65–70.
17
Appendix
In this appendix we list the graph-theoretic parameters computed for the
resolutions of 83, 129 and 234 vertex graphs. The tables contain their arith-
metic means in the male and female groups, and the corresponding p-values.
The values in these tables contain the values corresponded to round 1 (see the
"Statistical analysis" subsection in the main text).
The graph-parameters are defined in the caption of Table 1.
Significant differences (p < 0.01) are denoted with an asterisk in the last
column.
Scale 83, round 1
Property
All AdjLMaxDivD FAMean
All AdjLMaxDivD FiberLengthMean
All AdjLMaxDivD FiberN
All AdjLMaxDivD FiberNDivLength
All AdjLMaxDivD Unweighted
All HoffmanBound FAMean
All HoffmanBound FiberLengthMean
All HoffmanBound FiberN
All HoffmanBound FiberNDivLength
All HoffmanBound Unweighted
All LogSpanningForestN FAMean
All LogSpanningForestN FiberLengthMean
All LogSpanningForestN FiberN
All LogSpanningForestN FiberNDivLength
All LogSpanningForestN Unweighted
All MinCutBalDivSum FAMean
All MinCutBalDivSum FiberLengthMean
All MinCutBalDivSum FiberN
All MinCutBalDivSum FiberNDivLength
All MinCutBalDivSum Unweighted
All MinSpanningForest FAMean
All MinSpanningForest FiberLengthMean
All MinSpanningForest FiberN
All MinSpanningForest FiberNDivLength
All MinVertexCoverBinary Unweighted
All MinVertexCover FAMean
All MinVertexCover FiberLengthMean
All MinVertexCover FiberN
All MinVertexCover FiberNDivLength
All MinVertexCover Unweighted
All PGEigengap FAMean
All PGEigengap FiberLengthMean
All PGEigengap FiberN
All PGEigengap FiberNDivLength
All PGEigengap Unweighted
All Sum FAMean
All Sum FiberLengthMean
All Sum FiberN
All Sum FiberNDivLength
All Sum Unweighted
Left AdjLMaxDivD FAMean
Left AdjLMaxDivD FiberLengthMean
Left AdjLMaxDivD FiberN
Left AdjLMaxDivD FiberNDivLength
Left AdjLMaxDivD Unweighted
Left HoffmanBound FAMean
Left HoffmanBound FiberLengthMean
Left HoffmanBound FiberN
Female
1.36008
1.44214
2.02416
1.84476
1.26760
4.36096
3.21938
2.63525
2.51038
4.55192
110.69890
456.60084
397.53780
148.03174
191.66035
0.00793
0.03115
0.02924
0.02868
0.04001
19.78188
1096.37958
99.53846
3.65548
59.80769
18.73144
2014.06431
2427.21154
110.25657
40.90385
0.05403
0.04167
0.03156
0.03470
0.05214
222.01291
Male
1.37750
1.43602
2.10529
1.86864
1.26456
4.18564
3.26552
2.55573
2.40550
4.43931
101.82758
452.95875
389.79037
139.85355
187.85180
0.00474
0.02889
0.02711
0.02644
0.03721
18.63722
1112.97289
102.93333
3.66822
59.00000
18.10619
1955.70824
2315.20000
103.59777
41.00000
0.05071
0.03891
0.02829
0.03062
0.04740
201.02562
16845.33062
11261.65385
476.56342
567.07692
15792.24352
10237.13333
433.37987
539.80000
1.33644
1.40515
1.90607
1.71498
1.24027
4.55406
3.25098
2.71430
1.35216
1.38890
2.02087
1.77482
1.23523
4.38621
3.28435
2.61098
p-value
0.06806
0.72030
0.05606
0.41834
0.63251
0.00087
0.33136
0.03144
0.01815
0.04616
0.00012
0.18687
0.00001
0.00003
0.00113
0.14869
0.47008
0.34092
0.38768
0.28887
0.02232
0.10506
0.14280
0.93669
0.00716
0.01699
0.37460
0.00244
0.00036
0.32897
0.28914
0.43309
0.03885
0.01847
0.09708
0.00029
0.06219
0.00000
0.00002
0.00025
0.15767
0.32795
0.00501
0.07539
0.43598
0.01297
0.51250
0.00175
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
18
Left HoffmanBound FiberNDivLength
Left HoffmanBound Unweighted
Left LogSpanningForestN FAMean
Left LogSpanningForestN FiberLengthMean
Left LogSpanningForestN FiberN
Left LogSpanningForestN FiberNDivLength
Left LogSpanningForestN Unweighted
Left MinCutBalDivSum FAMean
Left MinCutBalDivSum FiberLengthMean
Left MinCutBalDivSum FiberN
Left MinCutBalDivSum FiberNDivLength
Left MinCutBalDivSum Unweighted
Left MinSpanningForest FAMean
Left MinSpanningForest FiberLengthMean
Left MinSpanningForest FiberN
Left MinSpanningForest FiberNDivLength
Left MinVertexCoverBinary Unweighted
Left MinVertexCover FAMean
Left MinVertexCover FiberLengthMean
Left MinVertexCover FiberN
Left MinVertexCover FiberNDivLength
Left MinVertexCover Unweighted
Left PGEigengap FAMean
Left PGEigengap FiberLengthMean
Left PGEigengap FiberN
Left PGEigengap FiberNDivLength
Left PGEigengap Unweighted
Left Sum FAMean
Left Sum FiberLengthMean
Left Sum FiberN
Left Sum FiberNDivLength
Left Sum Unweighted
Right AdjLMaxDivD FAMean
Right AdjLMaxDivD FiberLengthMean
Right AdjLMaxDivD FiberN
Right AdjLMaxDivD FiberNDivLength
Right AdjLMaxDivD Unweighted
Right HoffmanBound FAMean
Right HoffmanBound FiberLengthMean
Right HoffmanBound FiberN
Right HoffmanBound FiberNDivLength
Right HoffmanBound Unweighted
Right LogSpanningForestN FAMean
Right LogSpanningForestN FiberLengthMean
Right LogSpanningForestN FiberN
Right LogSpanningForestN FiberNDivLength
Right LogSpanningForestN Unweighted
Right MinCutBalDivSum FAMean
Right MinCutBalDivSum FiberLengthMean
Right MinCutBalDivSum FiberN
Right MinCutBalDivSum FiberNDivLength
Right MinCutBalDivSum Unweighted
Right MinSpanningForest FAMean
Right MinSpanningForest FiberLengthMean
Right MinSpanningForest FiberN
Right MinSpanningForest FiberNDivLength
Right MinVertexCoverBinary Unweighted
Right MinVertexCover FAMean
Right MinVertexCover FiberLengthMean
Right MinVertexCover FiberN
Right MinVertexCover FiberNDivLength
Right MinVertexCover Unweighted
Right PGEigengap FAMean
Right PGEigengap FiberLengthMean
Right PGEigengap FiberN
Right PGEigengap FiberNDivLength
Right PGEigengap Unweighted
Right Sum FAMean
2.66652
4.73205
53.30579
229.63370
199.27958
73.53683
95.46307
0.00687
0.23438
0.13337
0.11057
0.24513
9.57924
561.47024
51.23077
1.82447
30.23077
9.23616
2.59451
4.57434
48.82905
227.32675
195.25428
69.82889
93.39767
0.00320
0.21147
0.12011
0.09321
0.22019
9.06313
560.36391
53.73333
1.89521
29.73333
8.88642
1064.27185
1158.21154
1027.73430
1143.46667
54.26322
20.80769
0.33446
0.33383
0.16980
0.14486
0.30646
106.64056
8629.73791
5514.61538
233.06402
282.50000
1.32878
1.39672
2.00803
1.76990
1.25268
4.47438
3.33823
2.67311
2.62635
4.61480
52.25642
218.25106
190.62427
69.84080
90.24090
0.02476
0.24577
0.13346
0.10831
0.23713
10.30911
532.13580
50.76923
1.94340
29.07692
9.26572
51.17634
20.83333
0.29469
0.29287
0.15238
0.13413
0.27160
96.80731
8122.82646
5049.73333
213.49323
269.06667
1.34242
1.38478
2.09048
1.81343
1.24720
4.28666
3.39478
2.57701
2.48983
4.50726
48.14346
216.24411
187.02757
66.17446
88.51678
0.00851
0.22309
0.12050
0.09357
0.22022
9.79708
547.85331
52.53333
1.89232
28.73333
9.03965
934.26071
1169.63462
897.95882
1122.93333
53.57144
20.11538
0.32454
0.34029
0.17666
0.15245
0.29582
105.62164
51.50298
20.26667
0.28808
0.29461
0.15912
0.14034
0.26081
95.26436
0.13782
0.01379
0.00082
0.18765
0.00012
0.00343
0.01389
0.17151
0.01779
0.00403
0.00031
0.00206
0.04242
0.87722
0.26795
0.62729
0.09601
0.01371
0.35926
0.55321
0.02122
0.75017
0.00215
0.01329
0.01654
0.02837
0.00458
0.00056
0.13250
0.00000
0.00043
0.00378
0.14511
0.30191
0.05380
0.09784
0.29540
0.00587
0.29902
0.05411
0.00560
0.03806
0.00067
0.16431
0.00083
0.00022
0.00541
0.00496
0.02216
0.00346
0.00072
0.01629
0.10419
0.00491
0.26282
0.58863
0.15457
0.12382
0.23661
0.07986
0.10452
0.10527
0.00112
0.00949
0.02617
0.01613
0.00087
0.00028
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
19
Right Sum FiberLengthMean
Right Sum FiberN
Right Sum FiberNDivLength
Right Sum Unweighted
7644.90330
5378.03846
225.94776
261.30769
7086.91000
4884.66667
206.97587
248.26667
0.02974
0.00000
0.00018
0.00019
∗
∗
∗
Scale 129, round 1
Property
All AdjLMaxDivD FAMean
All AdjLMaxDivD FiberLengthMean
All AdjLMaxDivD FiberN
All AdjLMaxDivD FiberNDivLength
All AdjLMaxDivD Unweighted
All HoffmanBound FAMean
All HoffmanBound FiberLengthMean
All HoffmanBound FiberN
All HoffmanBound FiberNDivLength
All HoffmanBound Unweighted
All LogSpanningForestN FAMean
All LogSpanningForestN FiberLengthMean
All LogSpanningForestN FiberN
All LogSpanningForestN FiberNDivLength
All LogSpanningForestN Unweighted
All MinCutBalDivSum FAMean
All MinCutBalDivSum FiberLengthMean
All MinCutBalDivSum FiberN
All MinCutBalDivSum FiberNDivLength
All MinCutBalDivSum Unweighted
All MinSpanningForest FAMean
All MinSpanningForest FiberLengthMean
All MinSpanningForest FiberN
All MinSpanningForest FiberNDivLength
All MinVertexCoverBinary Unweighted
All MinVertexCover FAMean
All MinVertexCover FiberLengthMean
All MinVertexCover FiberN
All MinVertexCover FiberNDivLength
All MinVertexCover Unweighted
All PGEigengap FAMean
All PGEigengap FiberLengthMean
All PGEigengap FiberN
All PGEigengap FiberNDivLength
All PGEigengap Unweighted
All Sum FAMean
All Sum FiberLengthMean
All Sum FiberN
All Sum FiberNDivLength
All Sum Unweighted
Left AdjLMaxDivD FAMean
Left AdjLMaxDivD FiberLengthMean
Left AdjLMaxDivD FiberN
Left AdjLMaxDivD FiberNDivLength
Left AdjLMaxDivD Unweighted
Left HoffmanBound FAMean
Left HoffmanBound FiberLengthMean
Left HoffmanBound FiberN
Left HoffmanBound FiberNDivLength
Left HoffmanBound Unweighted
Left LogSpanningForestN FAMean
Left LogSpanningForestN FiberLengthMean
Left LogSpanningForestN FiberN
Left LogSpanningForestN FiberNDivLength
Left LogSpanningForestN Unweighted
Left MinCutBalDivSum FAMean
Left MinCutBalDivSum FiberLengthMean
Left MinCutBalDivSum FiberN
Left MinCutBalDivSum FiberNDivLength
Female
1.40519
1.50483
2.14552
2.09783
1.30028
4.40157
3.19684
2.50604
2.34647
4.62935
194.37749
739.78985
599.76631
210.52236
322.09324
0.00668
0.01706
0.02658
0.02495
0.02218
30.14746
1642.68263
140.23077
4.42401
96.46154
29.56250
3230.07900
2444.92308
120.18766
63.88462
0.03143
0.02427
0.02781
0.02880
0.03012
397.68878
30670.09535
12375.61538
548.61301
1020.80769
1.37823
1.43638
1.84672
1.77313
1.26380
4.57539
3.23550
2.80373
2.70077
4.75280
96.11000
373.09476
300.77613
105.01323
162.01302
0.00873
0.19822
0.12848
0.06926
Male
1.42604
1.50158
2.22254
2.04782
1.29097
4.29660
3.24689
2.48884
2.41938
4.51267
181.03525
732.55388
588.61699
200.75240
316.62672
0.00324
0.01607
0.02429
0.02258
0.02065
28.58509
1664.23693
140.93333
4.43795
96.26667
28.72424
3121.21684
2337.40000
116.22553
63.96667
0.02928
0.02260
0.02453
0.02498
0.02725
360.50850
28478.19852
11458.13333
510.71378
972.86667
1.39812
1.42179
1.92762
1.80979
1.25501
4.44885
3.25088
2.74220
2.64308
4.61941
89.25516
368.65582
295.83044
100.80980
158.88026
0.00273
0.17378
0.10467
0.05546
p-value
0.10040
0.87806
0.15242
0.32031
0.27278
0.02644
0.32568
0.64956
0.07720
0.01233
0.00019
0.09867
0.00000
0.00000
0.00218
0.05930
0.56293
0.26627
0.30029
0.30082
0.02073
0.07510
0.55077
0.92181
0.66793
0.02181
0.29100
0.00232
0.02502
0.35805
0.25524
0.43054
0.01902
0.00792
0.09661
0.00015
0.03582
0.00000
0.00008
0.00026
0.12792
0.36739
0.12247
0.33521
0.16858
0.01512
0.77158
0.14090
0.21782
0.00848
0.00106
0.08843
0.00014
0.00288
0.01336
0.05683
0.00892
0.00001
0.00019
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
20
Left MinCutBalDivSum Unweighted
Left MinSpanningForest FAMean
Left MinSpanningForest FiberLengthMean
Left MinSpanningForest FiberN
Left MinSpanningForest FiberNDivLength
Left MinVertexCoverBinary Unweighted
Left MinVertexCover FAMean
Left MinVertexCover FiberLengthMean
Left MinVertexCover FiberN
Left MinVertexCover FiberNDivLength
Left MinVertexCover Unweighted
Left PGEigengap FAMean
Left PGEigengap FiberLengthMean
Left PGEigengap FiberN
Left PGEigengap FiberNDivLength
Left PGEigengap Unweighted
Left Sum FAMean
Left Sum FiberLengthMean
Left Sum FiberN
Left Sum FiberNDivLength
Left Sum Unweighted
Right AdjLMaxDivD FAMean
Right AdjLMaxDivD FiberLengthMean
Right AdjLMaxDivD FiberN
Right AdjLMaxDivD FiberNDivLength
Right AdjLMaxDivD Unweighted
Right HoffmanBound FAMean
Right HoffmanBound FiberLengthMean
Right HoffmanBound FiberN
Right HoffmanBound FiberNDivLength
Right HoffmanBound Unweighted
Right LogSpanningForestN FAMean
Right LogSpanningForestN FiberLengthMean
Right LogSpanningForestN FiberN
Right LogSpanningForestN FiberNDivLength
Right LogSpanningForestN Unweighted
Right MinCutBalDivSum FAMean
Right MinCutBalDivSum FiberLengthMean
Right MinCutBalDivSum FiberN
Right MinCutBalDivSum FiberNDivLength
Right MinCutBalDivSum Unweighted
Right MinSpanningForest FAMean
Right MinSpanningForest FiberLengthMean
Right MinSpanningForest FiberN
Right MinSpanningForest FiberNDivLength
Right MinVertexCoverBinary Unweighted
Right MinVertexCover FAMean
Right MinVertexCover FiberLengthMean
Right MinVertexCover FiberN
Right MinVertexCover FiberNDivLength
Right MinVertexCover Unweighted
Right PGEigengap FAMean
Right PGEigengap FiberLengthMean
Right PGEigengap FiberN
Right PGEigengap FiberNDivLength
Right PGEigengap Unweighted
Right Sum FAMean
Right Sum FiberLengthMean
Right Sum FiberN
Right Sum FiberNDivLength
Right Sum Unweighted
0.19535
14.57467
828.34729
69.30769
2.16989
48.76923
14.65360
0.17339
13.88500
834.54850
72.20000
2.25626
48.86667
14.09857
1700.29684
1169.82692
1637.18742
1125.20000
58.76113
32.28846
0.22611
0.23241
0.12346
0.09689
0.20204
56.23736
32.30000
0.19656
0.20065
0.10569
0.08572
0.17516
197.41850
16079.40944
6071.96154
269.09760
519.53846
178.80563
14931.40760
5641.93333
251.40080
492.86667
1.35746
1.42015
2.05564
1.82146
1.26684
4.37886
3.32686
2.66511
2.68679
4.60861
93.41904
358.00491
291.08563
100.74383
154.36558
0.02361
0.20000
0.11452
0.06865
0.19180
15.61479
808.14079
70.46154
2.32813
47.34615
14.70648
1.36837
1.41129
2.19134
1.86716
1.25522
4.29574
3.36662
2.56838
2.59830
4.51407
87.28295
354.73456
285.72242
96.22891
151.96595
0.01005
0.17303
0.10111
0.06326
0.16911
14.88977
824.37649
68.93333
2.26810
47.00000
14.40974
1516.99670
1175.50000
1461.52391
1166.36667
59.59421
31.61538
0.22838
0.23840
0.12500
0.10075
0.20584
58.78162
31.73333
0.19627
0.19868
0.11049
0.09371
0.17429
190.48228
13952.01182
5935.73077
262.31420
477.38462
172.48988
13003.32443
5525.26667
246.32048
454.86667
0.00265
0.06189
0.36946
0.02902
0.53695
0.69355
0.01273
0.30481
0.06266
0.06303
0.88865
0.00995
0.02197
0.00382
0.00223
0.01081
0.00032
0.07487
0.00000
0.00100
0.00232
0.36353
0.54264
0.01338
0.20816
0.12057
0.20294
0.49418
0.01727
0.01992
0.08448
0.00143
0.14280
0.00045
0.00051
0.01158
0.00807
0.00768
0.00563
0.09375
0.00492
0.06537
0.03729
0.07096
0.46298
0.29760
0.13709
0.23679
0.68666
0.47843
0.20363
0.00296
0.01013
0.00869
0.03033
0.00242
0.00062
0.04620
0.00001
0.00180
0.00068
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
Scale 234, round 1
Property
All AdjLMaxDivD FAMean
All AdjLMaxDivD FiberLengthMean
Female
2.15050
2.35868
Male
2.14489
2.34695
p-value
0.86385
0.80876
21
All AdjLMaxDivD FiberN
All AdjLMaxDivD FiberNDivLength
All AdjLMaxDivD Unweighted
All HoffmanBound FAMean
All HoffmanBound FiberLengthMean
All HoffmanBound FiberN
All HoffmanBound FiberNDivLength
All HoffmanBound Unweighted
All LogSpanningForestN FAMean
All LogSpanningForestN FiberLengthMean
All LogSpanningForestN FiberN
All LogSpanningForestN FiberNDivLength
All LogSpanningForestN Unweighted
All MinCutBalDivSum FAMean
All MinCutBalDivSum FiberLengthMean
All MinCutBalDivSum FiberN
All MinCutBalDivSum FiberNDivLength
All MinCutBalDivSum Unweighted
All MinSpanningForest FAMean
All MinSpanningForest FiberLengthMean
All MinSpanningForest FiberN
All MinSpanningForest FiberNDivLength
All MinVertexCoverBinary Unweighted
All MinVertexCover FAMean
All MinVertexCover FiberLengthMean
All MinVertexCover FiberN
All MinVertexCover FiberNDivLength
All MinVertexCover Unweighted
All PGEigengap FAMean
All PGEigengap FiberLengthMean
All PGEigengap FiberN
All PGEigengap FiberNDivLength
All PGEigengap Unweighted
All Sum FAMean
All Sum FiberLengthMean
All Sum FiberN
All Sum FiberNDivLength
All Sum Unweighted
Left AdjLMaxDivD FAMean
Left AdjLMaxDivD FiberLengthMean
Left AdjLMaxDivD FiberN
Left AdjLMaxDivD FiberNDivLength
Left AdjLMaxDivD Unweighted
Left HoffmanBound FAMean
Left HoffmanBound FiberLengthMean
Left HoffmanBound FiberN
Left HoffmanBound FiberNDivLength
Left HoffmanBound Unweighted
Left LogSpanningForestN FAMean
Left LogSpanningForestN FiberLengthMean
Left LogSpanningForestN FiberN
Left LogSpanningForestN FiberNDivLength
Left LogSpanningForestN Unweighted
Left MinCutBalDivSum FAMean
Left MinCutBalDivSum FiberLengthMean
Left MinCutBalDivSum FiberN
Left MinCutBalDivSum FiberNDivLength
Left MinCutBalDivSum Unweighted
Left MinSpanningForest FAMean
Left MinSpanningForest FiberLengthMean
Left MinSpanningForest FiberN
Left MinSpanningForest FiberNDivLength
Left MinVertexCoverBinary Unweighted
Left MinVertexCover FAMean
Left MinVertexCover FiberLengthMean
Left MinVertexCover FiberN
Left MinVertexCover FiberNDivLength
Left MinVertexCover Unweighted
5.14838
5.17072
1.89062
3.63940
2.92490
2.23619
2.20178
3.73661
446.86116
2324.68381
1456.24015
149.01647
942.01654
0.00000
0.00769
0.02405
0.00000
0.00898
98.19730
5358.83904
481.46154
18.53246
276.15385
89.53747
8136.04292
2430.61538
129.82332
222.57692
0.01106
0.00860
0.01894
0.01773
0.00995
5.00652
4.78287
1.84578
3.62013
2.98466
2.26557
2.23871
3.72935
416.54482
2325.52712
1445.53700
138.16817
944.27877
0.00000
0.00723
0.02168
0.00000
0.00834
92.47667
5379.38212
479.20000
18.36575
280.33333
87.25805
7957.20990
2344.50000
126.64639
223.33333
0.01201
0.00960
0.01927
0.01767
0.01067
1033.36931
74747.99556
13609.34615
652.17760
2801.69231
961.08503
71461.78993
12823.40000
623.38731
2746.20000
2.14627
2.29338
4.03186
3.93717
1.86339
3.74670
2.94312
2.51168
2.44470
3.82814
212.18613
1159.44274
723.10349
70.44766
467.24325
0.00000
0.09355
0.07158
0.00000
0.09416
47.28302
2702.23206
244.11538
9.45842
137.19231
43.50481
4136.87086
1168.19231
63.94002
111.38462
2.14335
2.29214
4.16381
3.84897
1.81508
3.77335
2.99233
2.47461
2.45140
3.84621
197.08273
1165.33847
723.01322
65.77187
470.94213
0.00000
0.07667
0.05914
0.00000
0.07896
44.78250
2712.65026
244.46667
9.50259
140.00000
42.59720
4052.71473
1153.66667
64.04107
112.26667
0.35870
0.02543
0.06482
0.57408
0.17340
0.30055
0.13550
0.82472
0.03232
0.96824
0.36683
0.15229
0.83734
nan
0.57442
0.21132
0.45008
0.32475
0.00151
0.44199
0.45787
0.71037
0.12225
0.06974
0.48358
0.00056
0.02087
0.39844
0.54543
0.45409
0.89543
0.97772
0.59117
0.00297
0.18467
0.00000
0.00139
0.21290
0.93401
0.97718
0.29128
0.38654
0.04174
0.55549
0.25660
0.28318
0.85286
0.65499
0.04326
0.58696
0.98899
0.31060
0.52729
nan
0.00655
0.00062
nan
0.00153
0.00239
0.49327
0.89014
0.88229
0.06105
0.16942
0.55895
0.46021
0.92511
0.09259
∗
∗
∗
∗
∗
∗
∗
∗
∗
22
Left PGEigengap FAMean
Left PGEigengap FiberLengthMean
Left PGEigengap FiberN
Left PGEigengap FiberNDivLength
Left PGEigengap Unweighted
Left Sum FAMean
Left Sum FiberLengthMean
Left Sum FiberN
Left Sum FiberNDivLength
Left Sum Unweighted
Right AdjLMaxDivD FAMean
Right AdjLMaxDivD FiberLengthMean
Right AdjLMaxDivD FiberN
Right AdjLMaxDivD FiberNDivLength
Right AdjLMaxDivD Unweighted
Right HoffmanBound FAMean
Right HoffmanBound FiberLengthMean
Right HoffmanBound FiberN
Right HoffmanBound FiberNDivLength
Right HoffmanBound Unweighted
Right LogSpanningForestN FAMean
Right LogSpanningForestN FiberLengthMean
Right LogSpanningForestN FiberN
Right LogSpanningForestN FiberNDivLength
Right LogSpanningForestN Unweighted
Right MinCutBalDivSum FAMean
Right MinCutBalDivSum FiberLengthMean
Right MinCutBalDivSum FiberN
Right MinCutBalDivSum FiberNDivLength
Right MinCutBalDivSum Unweighted
Right MinSpanningForest FAMean
Right MinSpanningForest FiberLengthMean
Right MinSpanningForest FiberN
Right MinSpanningForest FiberNDivLength
Right MinVertexCoverBinary Unweighted
Right MinVertexCover FAMean
Right MinVertexCover FiberLengthMean
Right MinVertexCover FiberN
Right MinVertexCover FiberNDivLength
Right MinVertexCover Unweighted
Right PGEigengap FAMean
Right PGEigengap FiberLengthMean
Right PGEigengap FiberN
Right PGEigengap FiberNDivLength
Right PGEigengap Unweighted
Right Sum FAMean
Right Sum FiberLengthMean
Right Sum FiberN
Right Sum FiberNDivLength
Right Sum Unweighted
0.08402
0.08669
0.06812
0.05084
0.07190
0.07554
0.07722
0.05737
0.04481
0.06398
504.02280
38178.70022
6716.53846
322.55630
1401.80769
470.30921
36255.83071
6389.20000
311.23280
1380.33333
2.00996
2.15381
4.11898
3.79534
1.79189
3.63008
3.00591
2.40837
2.45857
3.71704
228.90719
1154.04516
724.05083
72.92465
467.61765
0.00050
0.10021
0.07599
0.00034
0.09573
50.98056
2655.83115
238.96154
9.28191
138.30769
45.80119
3994.00115
1144.80769
62.35579
111.09615
0.08312
0.08538
0.06631
0.05084
0.07102
2.02718
2.18170
4.41926
3.75488
1.77141
3.59884
3.02300
2.33314
2.38848
3.69299
215.28259
1148.91122
716.03208
68.45678
466.56728
0.00000
0.08439
0.06701
0.00000
0.08171
47.79220
2655.71544
236.00000
9.00082
140.00000
44.57707
3884.90036
1129.73333
61.50301
111.10000
0.07683
0.07887
0.06080
0.04854
0.06430
517.36095
35857.03890
6524.53846
312.50248
1368.00000
481.68012
34486.76733
6187.46667
299.09835
1339.06667
0.28777
0.29463
0.09675
0.18106
0.24844
0.01077
0.19037
0.00107
0.04079
0.39428
0.61502
0.41400
0.03397
0.70781
0.38704
0.45778
0.69490
0.00150
0.01602
0.50645
0.07936
0.63377
0.22608
0.30478
0.85195
0.19303
0.01271
0.00641
0.18042
0.01034
0.00435
0.99483
0.15420
0.18645
0.25603
0.07765
0.36802
0.41752
0.47854
0.99385
0.33378
0.40909
0.28067
0.52890
0.25554
0.00745
0.26347
0.00050
0.01170
0.20464
∗
∗
∗
∗
∗
∗
23
|
1910.08423 | 1 | 1910 | 2019-10-17T15:03:17 | Quantum effects in the brain: A review | [
"q-bio.NC",
"quant-ph"
] | In the mid-1990s it was proposed that quantum effects in proteins known as microtubules play a role in the nature of consciousness. The theory was largely dismissed due to the fact that quantum effects were thought unlikely to occur in biological systems, which are warm and wet and subject to decoherence. However, the development of quantum biology now suggests otherwise. Quantum effects have been implicated in photosynthesis, a process fundamental to life on earth. They are also possibly at play in other biological processes such as avian migration and olfaction. The microtubule mechanism of quantum consciousness has been joined by other theories of quantum cognition. It has been proposed that general anaesthetic, which switches off consciousness, does this through quantum means, measured by changes in electron spin. The tunnelling hypothesis developed in the context of olfaction has been applied to the action of neurotransmitters. A recent theory outlines how quantum entanglement between phosphorus nuclei might influence the firing of neurons. These, and other theories, have contributed to a growing field of research that investigates whether quantum effects might contribute to neural processing. This review aims to investigate the current state of this research and how fully the theory is supported by convincing experimental evidence. It also aims to clarify the biological sites of these proposed quantum effects and how progress made in the wider field of quantum biology might be relevant to the specific case of the brain. | q-bio.NC | q-bio | Quantum effects in the brain: A review
Betony Adams1 and Francesco Petruccione1
Quantum Research Group, School of Chemistry and Physics,
and National Institute for Theoretical Physics, University of KwaZulu-Natal,
Durban, 4001, South Africa.
(Dated: 21 October 2019)
In the mid-1990s it was proposed that quantum effects in proteins known as micro-
tubules play a role in the nature of consciousness. The theory was largely dismissed
due to the fact that quantum effects were thought unlikely to occur in biological
systems, which are warm and wet and subject to decoherence. However, the de-
velopment of quantum biology now suggests otherwise. Quantum effects have been
implicated in photosynthesis, a process fundamental to life on earth. They are also
possibly at play in other biological processes such as avian migration and olfaction.
The microtubule mechanism of quantum consciousness has been joined by other the-
ories of quantum cognition. It has been proposed that general anaesthetic, which
switches off consciousness, does this through quantum means, measured by changes
in electron spin. The tunnelling hypothesis developed in the context of olfaction has
been applied to the action of neurotransmitters. A recent theory outlines how quan-
tum entanglement between phosphorus nuclei might influence the firing of neurons.
These, and other theories, have contributed to a growing field of research that inves-
tigates whether quantum effects might contribute to neural processing. This review
aims to investigate the current state of this research and how fully the theory is
supported by convincing experimental evidence. It also aims to clarify the biological
sites of these proposed quantum effects and how progress made in the wider field of
quantum biology might be relevant to the specific case of the brain.
9
1
0
2
t
c
O
7
1
]
.
C
N
o
i
b
-
q
[
1
v
3
2
4
8
0
.
0
1
9
1
:
v
i
X
r
a
1
I.
INTRODUCTION
The idea that quantum physics might have something to do with explaining that most
mysterious organ, the brain, has generated scepticism among scientists. Just because quan-
tum theory and consciousness are both complex concepts does not mean they necessarily
inform each other. Indeed the contexts in which both occur have long been assumed to be
incompatible. The brain is a biological system, functioning at physiological temperatures
and subject to the myriad interactions by which living organisms survive. Quantum effects,
on the other hand, are conventionally limited to low temperature systems isolated from the
detrimental effects of environmental interaction. However, the mutual exclusion of biological
and quantum systems is no longer so absolute. Research in the field of quantum biology has
had some success in identifying how quantum processes might benefit living organisms1 -- 5.
On a fundamental level, it can be said that all biological systems are quantum mechanical,
being composed of atoms and thus subject to the quantum theory of atomic structure first
developed by Bohr and Rutherford at the beginning of the twentieth century6,7.
In the
field of quantum biology these are considered to be trivial quantum effects, what is more
interesting is whether quantum phenomena such as coherence, entanglement and tunnelling
might play a non-trivial role in enhancing the efficacy of biological processes. There is
mounting evidence that quantum coherence might contribute to the extraordinary efficiency
of photosynthesis. The avian compass has also been suggested to exploit quantum effects
with behavioural evidence supporting the hypothesis. Olfaction, enzyme catalysis and the
intricacies of DNA have all fallen under the scrutiny of researchers working in the field of
quantum biology1 -- 5. The basic biology of the brain, elevated though it is by the inexplicable
phenomenon of consciousness, is perhaps not, on a mechanistic level, so very different from
other processes that take place in the body. It is no wonder then that interest has grown
in the detailed physiological mechanisms that constitute the central nervous system. Being
that it is still not perfectly understood exactly how the brain and its related systems work,
quantum theory might have at least some contribution to make. There have been attempts
to tackle the hard problem of consciousness8,9, or apply quantum theory to human psychol-
ogy and cognition10 -- 12. This review, however, focuses on the various instances in which it
has been suspected that quantum effects play a role in the structural mechanisms by which
the brain performs its integral functions: the firing of nerves; the actions of anaesthesia,
2
neurotransmitters and other drugs; the sensory interpretation and organised signalling that
is central to the vast neural network that we identify as our self.
A. The classical brain
Neuroscience has made great strides in understanding how the brain works. However,
if our model of the brain were already fully realised there would be no need to investigate
whether quantum theory might offer any insights. The question of how the brain works is, to
state the very obvious, complicated. On one level it is a matter of matter, the network of cells
and signalling processes that constitute the central nervous and related systems. But there
is also the question of how this physiology gives rise to the phenomenon of consciousness.
While advances in imaging techniques have resolved some of the structure of the brain,
new discoveries are still being made. Recently, previously unheard of lymphatic vessels were
discovered in the meninges of the brain, suggesting a link between the central nervous system
and the immune system and prompting research into their role in neurodegeneration13,14.
Even less well understood is the relationship between structure and function and to what
extent functional connectivity, or an understanding of the mind, can be explained by the
basic anatomy of the brain15 -- 18.
1. Brain organisation
The central nervous system is made up of the brain and the spinal cord. The human
brain is a complex organisation of neural tissue, with approximately 86 billion neurons19.
While neurons are responsible for the electrical activity of the brain they are supported
by glial cells, which have a number of functions20. Brain matter consists of both grey and
white matter, where the former is mainly cell bodies and the latter is predominantly made
up of myelinated axons that allow for the transport of signals and the connectivity of the
different brain sections21. The brain is a network, dependent on the complex interaction of
its different constituent parts, thus the assignment of specific function to specific region is
a simplification. Nevertheless, for classification purposes it is often divided and subdivided
into a number of regions, the simplest of these divisions being the fore, mid and hindbrain.
3
FIG. 1. The basic structure of a nerve cell. Although a number of the organelles are missing
this simplified schematic is sufficient to gain some understanding of the location of the various
quantum effects discussed in this review. Of interest are the microtubules which occur in different
arrangements throughout the nerve cell and the mitochondria, which will be discussed in the context
of electron transfer. The myelin insulated axon has also been proposed as a wave guide for neural
photonics. The meeting of dendritic spine and axon terminal at the synaptic cleft is enlarged for
clearer understanding of the mechanism of neurotransmission and the transfer of neural signals.
Synaptic vesicles are also the site of the proposed uptake of Posner molecules during endocytosis.
2. Neural action
Nerve cells, the main constituents of the central nervous system, are elongated cells con-
sisting of cell body, dendrites and axon. A complex network of neurons throughout the
body allows for the propagation of information that is facilitated by the firing or not fir-
ing of neurons, the measure of a neuron's action potential. The biophysical mechanism of
action potential in neurons is conventionally understood through the work of Hodgkin and
4
Huxley22. In order for a neuron to fire the resting potential must be raised to the requisite
threshold potential. This is mediated by the gradient of electrically charged ions distributed
across the cell membrane. Communication between neurons is integral to their ability to
convey information. There is still some contention as to exactly how nerves communicate,
with evidence for both chemical and electrical signal transmission23. This review will focus
on the former, where neural communication is mediated by the release of neurotransmitters.
When an action potential propagating along a neuron reaches the axon terminal it triggers
the opening of voltage gated ion channels which in turn stimulate exocytosis, or the release
of neurotransmitters into the synaptic cleft24. These neurotransmitters diffuse across the
synaptic cleft and bind to special receptors on the dendritic spines, opening other ion chan-
nels.
Ions can now enter this nerve cell and change the membrane potential, potentially
generating an action potential in the post-synaptic nerve cell24. While this is a simplistic
model of the action of neurons it serves as a means to locate the quantum effects discussed in
the next section. Although neurons have a number of constituent parts, those important in
the context of this review are as follows. Microtubules, as the location of effects described in
theories of consciousness as well as coherent quantum transport. Mitochondria, as the site of
electron transport processes. The axon, as mediator of electrical signals and possible site of
biophoton transfer. A basic understanding of the synaptic mechanism will also be beneficial
to the discussion of both quantum effects in neurotransmission as well as the neural action
of Posner molecules.
3. Consciousness
While this review is preoccupied with the simpler question of structure, that is, the
quantum physics of certain physiological mechanisms in the brain, it does include some
theories as to how this physiology manifests as consciousness. Biology based attempts to
explain consciousness include identifying its neural correlates by means of neuroimaging
methods which study the changes in neural activity between conscious and unconscious
states as well as altered states of consciousness25 -- 30. A formal description of consciousness,
given the difficulty of quantifying its subjective experience, would likely borrow from complex
network theory as well as disciplines from physics and philosophy15,31 -- 35. The question is
still open as to whether quantum physics has something to add to the debate.
5
FIG. 2. The non trivial quantum effects discussed in this review and their neural context.
B. Non trivial quantum effects
In 1913, Bohr presented his model of the atom where, in contrast to classical orbits,
electrons occupied discrete energy levels. This followed quickly on the heels of Planck's ad-
vances in understanding blackbody radiation and Einstein's explanation of the photoelectric
effect which, along with Compton's work with X-rays, ushered in the new era of quantum
mechanics6,7. Fundamentally, all biology can be described as being quantum mechanical
in the same way that all matter is quantum mechanical. However, the aim of this review
is to investigate non trivial quantum effects in biological systems, to widen the scope of
quantum theory to include biological mechanisms and not merely the description of their
constituent atoms. Quantum weirdness is a well documented phenomenon. First Planck
and Einstein demonstrated that radiation, normally understood as behaving like a wave,
can also behave like a particle. De Broglie then suggested that matter, which seems dis-
crete, can sometimes show wave-like effects such as interference. Despite being famous for
its uncertainty, the formalisation of the theory has proved extremely successful in describing
the behaviour of microscopic systems. The mathematical framework of quantum mechanics
associates a physical system with a quantum state that contains all possible information
on the system. What is interesting is that within this framework, for two quantum states
describing a system, a linear combination of these states also describes the system. It is
this that gives rise to the uniquely quantum effect of superposition states, one of the non-
trivial effects discussed in this review. The concept of quantum coherence, which quantifies
6
the relationship between states in a superposition, is investigated in various biological con-
texts. As is entanglement, the non-classical correlation between different quantum states.
Entanglement is often discussed with reference to quantum spin, the property of elementary
particles that determines their behaviour in a magnetic field. Electron tunnelling is also
raised as a potential candidate for quantum effects in biological systems. An explanation of
tunnelling follows from the probabilistic description of quantum mechanics which allows the
possibility of a quantum particle passing through a classically forbidden potential barrier6,7.
This review is particularly focused on inelastic tunnelling, in which the tunnelling electrons
are coupled to vibrational modes in their biological context.
II. QUANTUM EFFECTS IN NEURAL PROCESSES
It is perhaps misleading to talk about quantum processes in the brain. Nerve cells extend
throughout the body. The quantum processes discussed in this review are not confined to
the brain but take place within neurons and at synapses. They are therefore implicated in
the biological functioning of the entire body. It might be more accurate to describe these
effects as quantum enhanced neural processing.
A. Orchestrated Objective Reduction
Orchestrated objective reduction (Orch OR), the application of quantum mechanical for-
malism to the question of consciousness, was proposed by Stuart Hameroff and Roger Pen-
rose in the 1990s9,36,37. In his 1989 book The Emperor's New Mind: Concerning Computers,
Minds and The Laws of Physics Penrose addresses the possibility that the laws of classical
physics are not sufficient to explain the phenomenon of consciousness, suggesting instead
that quantum physics might be integral to this explanation8. His hypothesis initially lacked
a biological context in which these quantum effects might occur. Hameroff, an anaesthesi-
ologist by training, had been previously interested in microtubules and suggested that they
could be a contender in which to situate a quantum model for consciousness. Hameroff and
Penrose have subsequently collaborated on developing and refining the theory of Orch OR,
by which quantum computations in microtubules influence neural firing and by extension
constitute the neural manifestation of consciousness9. Microtubules in general have elicited
7
interest in various researchers attempting to model quantum effects in the brain.
1. Microtubules
Microtubules are formed by the polymerisation of tubulin dimers, which consist of α and β
tubulin proteins. Tubulin dimers first form longitudinal protofilaments; 13 of these protofil-
aments then form a microtubule with diameter of approximately 25 nm9. Microtubules form
part of the cytoskeleton of eukaryotic and some prokaryotic cells and contribute to cellular
shape and structure. They have a variety of functions. They are integral to cell division,
forming the spindle apparatus that mediates the division of chromosomes into daughter
cells. Microtubules also act as tracks along which motor proteins move cellular constituents
within the cell9,38,39. Whereas microtubules are present in all eukaryotic cells the theory
of Orch OR is focused on microtubules in nerve cells and in particular those found in the
dendrites and cell body (soma) of these cells. This is because microtubules in axons and
non-neural cells have a radial, regular arrangement that is arguably less supportive of in-
formation processing. Microtubules in the dendrites and soma are less regularly arrayed,
forming what Hameroff and Penrose refer to as recursive networks well suited to learning9,40.
Microtubules in non-neural cells are also dynamically unstable, able to disassemble in var-
ious ways. Microtubules in dendrites and nerve cell bodies are prevented from disassembly
by microtubule associated proteins, rendering them more stable and able to encode the
long-term information necessary to the theory of Orch OR9,41,42. The specific composition
of tubulin has also lent strength to quantum models of neural processing due to the fact that
it is partially composed of chromophores such as tryptophan, arranged in a manner similar
to photosynthetic systems in plants and bacteria43,44, which have been surmised to support
coherent quantum effects.
2. The quantum model of Orch OR
Hameroff and Penrose hypothesise that quantum computations encode information in
microtubules and objective reduction is how this quantum information results in a classical
output. The details by which Penrose's conception of quantum gravity gives rise to the
objective reduction of the quantum wavefunction are beyond the scope of this review. In
8
this case Hameroff and Penrose's 2014 review of the theory is instructive9. The choice
of microtubules, and more specifically tubulin dimers, as the biological context in which
Orch OR takes place has been motivated by various reasons. It has been suggested that
the manifestation of consciousness is not axonal firing but rather the signal integration
that occurs in the dendrites and cell bodies of nerve cells. This is given some support by
the fact that gamma wave synchrony, which has been suggested to be the neural correlate
of consciousness, is generated by dendritic-somatic integration potentials9,45. As outlined
above, microtubule arrangement in the dendrites and cell bodies of nerve cells is suit-
able for information processing, making them a good contender for the biological site of
consciousness9,40,42. The localisation of Orch OR in microtubules also offers a way to model
a biological qubit, which is integral to the quantum nature of the theory. A qubit, the basic
unit of quantum information, is a two-state system that can exist in a superposition of both
states at the same time. Initially Hameroff and Penrose proposed that tubulin dimers might
exist in a superposition of mechanical conformations coupled to London force dipoles9,37,46.
More recent iterations of the theory locate the quantum description firmly in the constituent
aromatic rings (phenylalanine, tyrosine and tryptophan) that make up the tubulin proteins.
These have pi orbital electron clouds that demonstrate spatial delocalisation, giving rise to
London force electric dipoles, that can exist in superposition. While Orch OR was originally
formulated with electric dipoles in mind, the authors now propose magnetic dipoles related
to electron spin9.
3. Discussions around Orch OR
This review gives only a basic overview of the theory of Orch OR as a means to introduce
the idea of quantum models of the brain and their biological sites of action. There have
been a number of objections raised against Orch OR, the details of which are given in detail
in Hameroff and Penrose's 2014 review9. Those responses that deal with the structural
viability of quantum effects in the brain are briefly addressed in this section. The problem
of decoherence is an issue commonly cited against the application of quantum theory to
biological systems. Quantum mechanics is conventionally applied to isolated systems at low
9
temperatures whereas biological systems are typically described as being warm, wet and
messy. In 2000 Max Tegmark addressed the question of decoherence in the context of Orch
OR, concluding that calculated decoherence time scales of 10-13 -- 10-20 seconds are much too
short for quantum effects to play any role in cognitive processes47. Hagan et al. responded
to this by asserting that Tegmark had based his calculations on a model that did not closely
resemble the one they had proposed. After recalculation for a more accurate model they
conclude that decoherence times are closer to 10-5 -- 10-4 seconds48. With the development of
quantum biology the question of whether quantum mechanics can contribute any meaningful
insight into the functioning of biological organisms has shifted towards a positive answer.
In particular the study of photosynthesis has revealed that the efficacy of energy transfer
and charge separation might be enhanced by quantum effects49 -- 54. Thus the decoherence
argument against Orch OR perhaps holds less weight than other evidence contradicting the
theory.
There has been some argument over the lattice arrangement of tubulin dimers in mi-
crotubules, which can be of two types: A and B. The predictions of Orch OR favour type
A whereas actual mouse brain tissue points to the predominance of type B. Proponents of
Orch OR argue that this does not necessarily discount the theory as quantum effects may
only occur in that fraction of microtubules that are geometrically suitable9,55. Objections
by McKemmish et al.
include the proposed conformational switching of the tubulin pro-
teins that constitute the biological qubit. They note that the theory demands significant
changes in tubulin structure and that any processes that might drive these conformational
changes would be metabolically expensive56. More recent reformulations of Orch OR do
not require such extreme conformational switching and the different states necessary for the
qubit can be achieved through superposed electric or magnetic dipoles in aromatic rings9.
McKemmish et al. also note that the electrons in a single aromatic ring cannot exist in
superposition states, being completely delocalised. This can be incorporated by considering
the electron clouds of two or more rings9. Reimers et al. take issue with the possibility of
strong Frohlich condensation in microtubules57. They explain megahertz coherence demon-
strated in microtubules by Pokorn´y58 as being due to weak classical Frohlich condensation.
They go on to suggest that as Orch OR involves coherent strong Frohlich condensation
the theory is flawed57. Hameroff and Penrose respond by citing experimental evidence of
10
the discovery of gigahertz, megahertz and kilohertz resonance in single microtubules. They
argue that the experimental evidence argues strongly in the favour of Orch OR9,59,60. Other
criticisms have been aimed at the objective reduction part of the theory, and the way in
which this translates into consciousness. These have been addressed in detail by Hameroff
and Penrose9.
4. Microtubules as a site for quantum effects
While Orch OR theory is not without controversy, the suitability of microtubules as a
site for quantum effects has given rise to a number of related approaches. It seems probable
that neuronal microtubules are indeed implicated in consciousness and cognition. This is
supported by the fact that certain chemicals that influence both consciousness as well as cog-
nitive function, such as general anaesthetics and antidepressants, involve microtubules9,61,62.
The proposed quantum action and alternative biological sites of processes related to both
anaesthetics and antidepressants will be discussed in the sections B and C. Craddock et
al. suggest an additional way in which microtubules might play a quantum role in neural
processing. They take, as a comparative example, evidence of quantum beats in the light-
harvesting complexes of plants and bacteria. They then suggest that the tryptophan residues
present in the tubulin proteins that constitute microtubules are structurally and functionally
capable of supporting the possibility of quantum coherent energy transfer43. The details of
this research will be discussed further in the context of electron transfer processes in section
D.
B. Quantum models of general anaesthesia
One of the ways in which we might understand the mechanism of consciousness or, less
ambitiously, the details of neural processes, is by looking at chemicals that disrupt these
processes. As Luca Turin puts it, 'the only thing we are sure about consciousness, is that
it is soluble in chloroform'63. To this end the study of general anaesthetics has offered
some way of structuring research into quantum neural processing. There are a few theories
11
as to how quantum effects are implicated in the action of general anaesthetics. One of
these follows from the investigation of possible quantum effects in microtubules, initiated by
Hameroff and Penrose's Orch OR theory64. Another looks at spin changes in anaesthetised
fruit flies65. Both of these are preoccupied with the possibility that anaesthetic action
disrupts electronic activity, a theory that first made its appearance in the 1980s66 -- 69. There
is also some suggestion that the nuclear spin of anaesthetic molecules might influence their
efficacy70.
1. The action of anaesthetics in tubulin proteins
Craddock et al. argue that understanding general anaesthetics as having a network71 or
synaptic based effect72 does not explain how anaesthetics inhibit the cognitive abilities of sim-
ple single celled organisms such as slime moulds73,74. The fact that this cognition is linked to
cytoskeletal microtubules points to a possible site for the action of general anaesthetics64,75.
There is some uncertainty concerning the exact biological location targeted by anaesthetics.
The Meyer-Overton rule links the potency of anaesthetics to their increased lipid solubility.
Further research suggests that anaesthetics act in lipid-like hydrophobic regions in proteins64.
Anaesthetics bind to a number of membrane and cytoplasmic proteins76 -- 78, Craddock et al.
propose that tubulin, the protein subunit of microtubules, seems the most likely due to the
fact that gene expression after exposure to anaesthetic compounds is concentrated on mi-
crotubule dependent functions64,79. The quantum mechanism by which anaesthetics operate
in microtubules is based on a number of theoretical observations. Building on research that
tryptophan rings arranged favourably in tubulin proteins can act as quantum channels sup-
porting coherent electron dynamics in a manner similar to photosynthesis43,80,81, Craddock
et al. argue that anaesthetic gas molecules binding in these channels inhibit quantum effects
and disrupt coherent energy transfer and that this is responsible for the effects of general
anaesthetics on consciousness64,75. In a later paper Craddock et al. expand their theory to
the prediction of anaesthetic potency. They argue that anaesthetic and related gases change
collective terahertz dipole oscillations in a way that predicts the potency of their action82.
12
2. Quantum spin and general anaesthetics
In his early research relating to quantum biology, Turin hypothesised that the sense
of smell, understood classically as a lock-and-key mechanism, might instead be better ex-
plained by quantum theory. More recently he has addressed the common thread between the
very different molecules that act as general anaesthetics. Turin's theory is that anaesthetic
molecules perturb electron currents in their target proteins65. The molecules that comprise
the group of general anaesthetics range from the structurally simple noble gas xenon to the
much more complex molecule alfaxalone, a range that includes numerous other chemicals
without any apparent similarities. It is this structure-function anomaly that is motivation to
look at the underlying physics of anaesthetics65,83. Being that it is structurally the simplest
of the anaesthetics, more than one attempt has been made to understand the anaesthetic
action of xenon.
In a recent paper Li et al.
examine the differing anaesthetic effects of xenon isotopes.
Xenon has nine stable isotopes. Seven of these have zero nuclear spin but xenon 129 has a
nuclear spin of 1
2 and xenon 131 of 3
2
70. In their experiment Li et al. compare the loss of
righting reflex in mice, which correlates with consciousness, under the influence of the differ-
ent isotopes. In order to exclude electronic effects they also calculate the polarisabilities of
the different isotopes, finding them to be undifferentiated. Their results demonstrate that
xenon isotopes with non-zero nuclear spin have a lower anaesthetic effect. As the results
do not show a correlation between anaesthetic effect and atomic mass they conclude that
the differences must depend on the value of the nuclear spin. Although the details of the
mechanism are not clear the authors hypothesise that as spin half particles have been shown
to be better suited to quantum entanglement, perhaps entanglement promotes consciousness
in opposition to the effects of anaesthetics70. Their results recall research done by Fisher
et al.
into a possible mechanism by which neural entanglement might proceed84 -- 86. Of
particular interest here perhaps are the isotope-dependent effects of lithium, another drug
that alters conscious experience.
In order to investigate whether quantum physics plays any common role in their mech-
anism of action Turin et al. also focus on the physics of the xenon atom.
In 1989 IBM
13
used a scanning tunnelling microscope to manipulate 35 xenon atoms on a nickel surface
into spelling out the IBM logo; where the xenon atoms extend the conducting surface
of the nickel87.
In a similar manner, Turin et al. suggest that xenon atoms extend the
highest occupied molecular orbit (HOMO) of the protein they interact with, facilitating
electron transfer, and that this effect is common to the other molecules that act as gen-
eral anaesthetics65. They test this hypothesis with an electron spin resonance experiment
using anaesthetised fruit flies, looking for increases in spin caused by general anaesthetic.
Discounting free electron spin signals from melanin pigments the results of the experiment
lead them to conclude that general anaesthetics cause a change in spin. Further verification
for the hypothesis is offered by the fact that spin changes differ for anaesthetic resistant
flies. In addition to this, using density functional theory they demonstrate that chemicals
which display similar but less exaggerated central nervous system effects to anaesthetics,
also extend the highest occupied orbital momentum but to a lesser degree65.
While electron spin changes under the influence of general anaesthetics points towards
a possible quantum mechanism, there is still some scepticism as to whether the experimen-
tal evidence is sufficient proof of principle. Turin et al. themselves note that the spin changes
could be due to melanin. Although they checked their results against flies deficient in one
of three possible melanins, they concede that particularly neuromelanin, the grey in grey
matter, might play a role65. This is potentially interesting in the context of recent research
investigating the quantum role of neuromelanin facilitated electron transport in sections
of the brain88. Turin et al. also document the fact that the experiment was performed
under distinctly unphysiologic conditions.
In order for the flies to remain immobile and
not disturb the readings they were kept at 6 C. Although some temperature variation was
tested this was only between 2 and 10 C. In addition to this the flies were kept in anoxic
conditions to isolate the effects of anaesthetic gases from those of oxygen. The potential
role of oxygen in the experimental results points to the involvement of respiration, more
specifically the movement of electrons in the electron transport chain of mitochondria89.
Turin et al. cite a number of instances in which it has been shown that mitochondria are
involved in the action of anaesthetics65,90 -- 92.
In a more recent collaboration with Turin,
Gaitanidis et al.
invite speculation on an unpublished preliminary report of spontaneous
radiofrequency emission from fruits flies subjected to a magnetic field. They suggest that
14
these emissions originate from the nervous system due to the fact that they stop under
the influence of chloroform anaesthetic and that they are related to spin-polarised electron
currents in cells. The authors note that these cellular currents could reflect mitochondrial
metabolism but that the variable signal and reaction to anaesthetic suggests some more
complex biological activity connected to neuronal activity93.
3. The outlook on quantum models of general anaesthesia
While the comparison of electron dynamics in photosynthetic and microtubule proteins
is largely theoretical the authors suggest its experimental verification is possible via two-
dimensional electronic spectroscopy, a technique well established in the analysis of photo-
synthetic systems5. They propose that quantum beating in tubulin will be altered under the
influence of anaesthetics and that this is a measurable phenomenon64. While this has yet
to be attempted there has been some recent experimental evidence that supports the possi-
bility that quantum effects play a role in the action of anaesthetics. Using various types of
spectroscopy, Burdick et al. report that, in contrast to nonhalogenated ethers, halogenated
ethers interact with entangled photons of specific wavelength94.
C. Neurochemical binding and activation mechanisms
For the purposes of this review the term neurotransmitter is used with respect to a
variety of neurochemicals that fall into this category, binding to neuroreceptors and modu-
lating neural signalling. Neurotransmitters emitted by one nerve cell bind to receptors on
an adjacent nerve cell and facilitate the opening of ion channels. This is fundamental to
the generation of action potentials and disruption of this process is believed to contribute
to mental illnesses95,96. Although this review focuses on the action of neurotransmitters it
has been suggested by Vaziri et al. that the mechanism by which ion channels allow the
selective transmission of ions might also not be strictly classical, showing evidence of quan-
tum coherence97. The binding action of neurotransmitters is conventionally understood as
a lock and key mechanism whereby the shape of the neurotransmitter matches its specific
receptor98,99. The lock and key or docking mechanism is implicated in a number of biological
processes: neurotransmission; the action of enzymes; olfaction; DNA binding, all depend to
15
some extent on lock and key theory98,100 -- 102. Despite the success of the theory however, an
alternative view suggests that something more than molecular shape might be necessary to
explain olfaction and, more recently, neurotransmission.
1. Lock and key or quantum vibration
Neurotransmitters are a class of molecules that bind to G protein-coupled receptors
(GPCRs). GPCRs are a group of receptors that, upon detection of appropriate molecules
known as ligands, activate signalling pathways in eukaryotic cells. GPCRs play an impor-
tant role in medical innovation, as targets for drug action103. In addition to neurotrans-
mitters, there are a number of biological molecules that bind to GPCRs; these include
odourants104. Olfaction is classically understood as depending on the respective shapes of
the chemical/receptor pair. However, an alternative vibrational theory of olfaction was first
developed by Dyson as long ago as the 1930s105,106.
In his 1996 paper Turin suggested
that the vibrational frequency of a given odourant contributes to quantum tunnelling at
the receptor107. In 2011 Franco et al. presented experimental evidence in support of the
theory.
In particular they claimed that fruit flies could differentiate between deuterated
odourants108. There is also some evidence that more complex species such as lake white-
fish and the American cockroach can differentiate between isotopes of amino acids and
pheromones109 -- 111. Horsfield et al. document a number of experiments that test the the-
ory on both the behavioural and physiological level105. Deuterated odourants are a useful
means to test the vibrational theory of receptor activation due to the fact that replacing
hydrogen by deuterium adds vibrational modes, for instance the 2150 cm−1 vibration of the
carbon-deuterium bond105,112,113. However, there has also been some recent discussion as to
the fact that a differential olfactory receptor response between undeuterated and deuterated
odourants could in fact be due to a minute contaminant in one of the samples. Paoli et
al. conclude that although their results do not prove the vibrational theory of olfaction
they also do not disprove it, merely calling for caution in experimental approach113. While
there is little verification for a vibrational olfactory mechanism in mammals, a number
of the experiments support the theory for the case of insects.
It should be noted, in a
discussion of GPCR mechanisms, that insect olfactory receptors differ from mammalian G-
protein related olfactory receptors and thus experimental results may not be generalisable105.
16
While the theory remains controversial it has recently been re-examined in the context of
neurotransmission. A number of different neuroreceptors have been investigated. Adenosine
receptors are rhodopsin-related G-protein coupled receptors that bind the neuromodulator
adenosine, but can also bind a variety of other agonist and antagonist molecules114. The
stimulating action of caffeine, for instance, is due to its antagonistic binding to adenosine
receptors. Because GPCRs are such an important target for pharmaceutical intervention
the classification of molecules that act as agonists or antagonists is a well-developed re-
search field. Molecules can be classified using various molecular descriptors which include
information from molecule structure, topology and geometry to dipole moment, electric
polarisability and electrostatic potential114. Following on from the fact that adenosine re-
ceptors and olfactory receptors are both class A GPCRs, Chee and Oh present research that
tests whether vibrational frequency might also be an effective molecular descriptor. They
suggest that classifying ligands by vibrational frequencies is an effective way of discrimi-
nating agonist from antagonist114. Chee et al. then refine this research using a machine
learning approach and conclude that selected features of molecular vibration allow for ligand
classification of adenosine receptor agonism115.
Whereas Chee et al 's research investigates adenosine receptors Hoehn et al. focus in partic-
ular on the neurotransmitter serotonin and its receptors109,116. Although it has a number
of other functions in biological systems serotonin is perhaps most well known for the role it
plays in mood. A widely prescribed class of antidepressants, the SSRIs (selective serotonin
reuptake inhibitors) target this neurotransmitter117. In their research Hoehn et al. address
the possibility that serotonin neurotransmission might utilise vibration assisted inelastic
tunnelling effects. Using inelastic electron tunneling spectroscopy (IETS) theory that was
first developed to understand olfaction they investigate the tunnelling spectra of endogenous
and non-endogenous agonists that bind to the serotonin receptor. Non-endogenous agonists
of this receptor include LSD (lysergic acid dimethylamide), DOI (2,5-dimethoxy-4-Iodo-
amphetamine) and other psychedelic phenethylamines109. Their results suggest that the
serotonin molecule shares an inelastic electron tunnelling spectral peak with other agonists
that activate the serotonin receptor. Although the lock and key mechanism has been very
successful in modelling certain aspects of the action of signalling proteins one of the ways in
17
which it falls short is the prediction of agonist potency. In addition to the shared spectral
peak of related agonists Hoehn et al. also report that the intensity of this peak might be
used as a predictor of agonist potency109.
2. Experimental verification of the vibration assisted tunnelling hypothesis
Experimental verification of the vibrational theory of neuroreception follows the lead
of Franco et al.
in the context of olfaction, where the effects of deuterated odourants are
investigated108. Hoehn et al. report the results of an experiment to test the feasibility of
their theoretical approach by measure of receptor affinity and activation. Once again, ago-
nists of the serotonin receptor were chosen, specifically 2,5-dimethoxy-4-iodoamphetamine
(DOI) and N,N-dimethyllysergamide (DAM-57). However the experiment failed to confirm
the vibrational theory of neuroreception as they report that selective deuteration had no
influence on binding affinity or activation116.
In another study regarding the binding mechanism of the neurotransmitter histamine,
authors Krzan et al. report a significant distinction in the binding patterns of histamine
and its deuterated counterpart, as well as for other agonists of the histamine receptor118.
In particular they found that deuterating histamine increased its binding affinity. They
discuss this in the context of the vibrational theory of olfaction and GPCRs more broadly.
Instead of confirming the theory they propose an alternative reason for the fact that ex-
periments suggest that animals can differentiate between deuterated odourants, concluding
that it is a nuclear quantum effect governed by differences in the strength of hydrogen
bonds before and after deuteration118. In contrast to this, theoretical work comparing the
structure-activity/vibration-activity relationship of the histamine receptor and its various
ligands suggests that molecular vibration does play some role in ligand function.
In the
study 47 ligands that bind to histamine receptors were investigated using a computational
approach to molecular vibration. This led to the conclusion that the many varying agonists
and antagonists can be to some extent classified by their molecular vibrations119.
18
3. The functional mechanisms of GPCRs in general
The lack of experimental verification for vibrational quantum effects in the context of
central nervous system GPCRs prompts Hoehn et al. to suggest that either olfactory recep-
tors function differently from other GPCRs or the vibrational theory of receptor activation
is wrong. Gehrckens et al. address the latter criticism in a recent preprint, in which they
use density functional theory to investigate the electronic structure of rhodopsin104. They
choose to look at rhodopsin rather than olfactory GPCRs due to the fact that a high res-
olution structure of rhodopsin is available for research purposes. Rhodopsin also shares
important features with olfactory receptors and is considered the evolutionary ancestor of
GPCRs120,121. The main aim of their research is to demonstrate a mechanism for electron
transfer in olfactory receptors. They go on to identify a tryptophan donor and zinc accep-
tor that would indeed allow electron transfer in rhodopsin, although they also emphasise
the fact that this capacity for electron transfer does not necessarily play a role in the
functionality of rhodopsin. In the context of whether this effect might be generalisable to
GPCRs they conclude that an electron transfer mechanism might have been exploited by
the offshoots of rhodopsin, in particular olfactory receptors. They do however suggest that
in this sense olfactory receptors may be unique in that they are relatively non-specific, a
requirement made necessary by the novelty and variety of odourants that will bind to them.
Neurotransmission, they argue, depends on a binding that is much more specific. On the
subject of neurotransmitters, they hypothesise that the anomalous binding structure of the
antagonist metitepine to serotonin related receptors can be explained if the metitepine is in
the form of a radical, having gained an electron in binding to the receptor. This mechanism
is in line with the idea that GPCRs are electronic devices rather than simply lock-and-key104.
Both theoretical and experimental results do not offer any clear conclusions with respect to
the vibrational theory of neurotransmission for adenosine, serotonin and histamine recep-
tors and their related ligands. Further research is necessary to clarify the viability of the
approach and its specific relation to the different actions of neurotransmitter binding affinity
and activation capability115. Understanding the mechanism or mechanisms of ligand-GPCR
interaction is particularly important due to the fact that GPCRs are a major drug target
associated with one third of all pharmaceuticals103,115.
19
D. Alternative signalling processes and biophotons
It is perhaps interesting that GPCRs, in addition to binding with molecules such as
neurotransmitters and odourants, also bind to photons122. There has recently been some
suggestion that cellular communication and even possibly neural signalling might make use of
biophotons in addition to more well established mechanisms such as neurotransmitters123 -- 125.
Biophotons, are spontaneous ultra-weak photons in the near-UV to near-IR spectral range
produced by biological systems, in particular through oxidative processes in mitochondria125,126.
As noted in a recent paper on the subject, it is known that biophotons are produced in
brains125. Correlations have been found between biophoton intensity and neural activity
as well as oxidative dysfunction of neural cells in rat and mouse brains127 -- 129. There has
even been some attempt to explain human intelligence in the context of biophoton emis-
sion. Following on from a study that demonstrated that glutamate, a key neurotransmitter,
can mediate biophoton production130, Wang et al. examine the spectral characteristics of
biophoton emission in a range of species. They conclude that there is a correlation between
higher order intelligence and the spectral redshift of biophotons emitted by sample brain
slices. This redshift increases in the order of frog, mouse, chicken, pig, monkey, and human,
which reflects with some accuracy the phylogenetic tree, although they concede that there
is no completely objective means to measure intelligence across species131. Their argument
has met with some scepticism, with Salari et al. responding that the experimental results
insufficiently support the hypothesis and that without a mechanistic connection between
spectral shift and intelligence, their correlation is more likely to be coincidence132.
Biophotons have been implicated in cellular signalling in plants, bacteria and even kid-
ney cells. Following from this Sun et al. present experimental evidence that nerve cells can
also conduct biophotons, and that this effect can be inhibited by the application of a neural
conduction block anaesthetic124,133. In their attempt to model the mechanistic details as to
how biophotons are utilised in information transfer, Kumar et al. argue that neurons are well
suited to photonics125,134. They cite evidence that nerve cells contain possible photon sources
from mitochondrial respiration or lipid peroxidation, as well as photon detectors such as
centrosomes and chromophores125,135 -- 138. They then hypothesise that myelin-coated axons
of nerve cells act as waveguides for biophotons and that this might facilitate quantum effects
20
such as entanglement125. In order to demonstrate proof of principle the authors develop a
theoretical model of light guidance in axons, solving the three dimensional electromagnetic
field equations in the relevant context. They investigate a number of limiting factors and
potential pitfalls and conclude that axons are mechanistically viable as waveguides. They
go on to suggest various experiments in which the different aspects of their hypothesis might
be put to the test125,134.
1. Quantum effects in electron transfer
In a review that addresses the possibility of quantum neurobiology, Jedlicka suggests
that cognition, typified by complex signal processing and integration, can potentially be
thought of as occurring outside of strictly neural systems and that the quantum information
processing abilities demonstrated by plants and bacteria could be useful to the study of
cognition in higher animals10. Thus far, the theoretical and experimental advances made in
understanding the role of quantum effects in biological systems have been primarily to do
with photosynthesis, in particular light harvesting and energy/charge transfer processes5.
Many of the theories of quantum effects in the context of neural processes focus on electron
dynamics. As noted by Toole et al. coherent energy transfer in photosynthesis is less a
unique feature of photosynthetic systems than it is the result of the specific arrangement of
chromophores within a protein44.
Kurian et al. propose that biophoton production in mitochondria is absorbed and channelled
via resonant energy transfer by co-localised microtubules. They argue that neurodegener-
ative diseases related to compromised microtubule networks could result from ineffective
channelling of biophotons into signalling or dissipation. Experimental evidence shows that
microtubules undergo organisational changes after exposure to photons, particularly in
the absorption range of tryptophan and tyrosine139. As previously noted, a related paper
presents a computational approach to quantum coherent energy transfer in microtubules43.
The authors argue that similarly to the arrangements of chromophores in light harvesting
photosynthetic complexes, the tubulin proteins that constitute microtubules have appropri-
ate arrangements of chromophores such as tryptophan. They conclude that it is feasible
that tubulin proteins could support coherent energy transfer in microtubules43,80. In a very
21
recent related paper Celardo et al. extend the similarities between photosynthetic antenna
complexes and microtubules by demonstrating that tryptophans in microtubules can theo-
retically exhibit superradiant excitonic states81.
The possible role of tryptophans in electron transfer processes has also been previously
addressed in this review with respect to the action of neurotransmitters in binding to and
activating GPCRs. In their investigation of rhodopsin, Gehrckens et al. outline a mechanism
by which a tryptophan donor and zinc acceptor facilitate electron transfer104. Tryptophan
seems particularly interesting with respect to neural processes as it is the precursor to
serotonin, an important neurotransmitter. This connection prompts Tonello et al. to hy-
pothesise that the structure of consciousness emerges in a manner analogous to the serotonin
dependent movement of plants towards their source of energy140. In neural processing this is
extrapolated as biophoton harvesting by tryptophans and concomitant serotonin-mediated
neural communication and plasticity140 In a later paper Tonello et al. expand on this idea
with the hypothesis that the gastrointestinal-brain axis in higher animals might have evolved
from the root-branch axis of plants, and that light plays an important role in both systems141.
While these ideas remain for the moment largely theoretical, much of the theory of
electron transport and resonant energy transfer has focused on the involvement of micro-
tubules. Alternatively, the role of biophoton use and energy transfer in neural processes
could be furthered by a closer look at the electron transport chain of mitochondrial respi-
ration. As previously noted, mitochondria are already implicated in the action of general
anaesthetics65,90 -- 92. They are also the primary production site of biophotons. They might
serve as an alternative biological location in which quantum effects could be investigated.
Research suggests that the arrangement of chromophores in bacteria allows for quantum
coherence in their photosynthetic processes142. Similarities between the DNA of bacteria
and eukaryotic mitochondria has led to the theory that mitochondria are descended from
bacteria, though recent research suggests they are less directly related143 -- 146.
While the ancestry of mitochondria remains uncertain their similarity to bacteria could
arguably manifest in biological structures that support quantum effects. Both photosyn-
thesis and mitochondrial respiration make use of an electron transport chain, which powers
22
FIG. 3. Simplistic comparison of the electron transport chains of respiration and photosynthe-
sis. Electrons are transported through a series of complexes or photosystems, which results in the
creation of a proton gradient. This is used to drive the production of ATP. While the specific
complexes differ between chains they share some features relevant to the possibility of quantum
processing. The presence of chromophores, for instance, which arguably support coherent energy
transfer. While the interaction of light with the photosynthetic electron transport chain has been
studied in detail, analogous effects in mitochondria are not well understood. Oxidative processes in
the electron transport chain of mitochondria are thought to be the main source of biophotons. The
function or target of these photons, however, is less clear. It has been proposed that microtubules
close to mitochondria absorb and channel biophotons in a quantum coherent manner139. Mito-
chondria themselves also demonstrate the capacity for photon absorption, in particular complex
IV in the electron transport chain, which is the target of therapeutic application of infrared light
to treat disorders of the brain151,153,158.
23
the charge gradient necessary for ATP synthesis. A group of researchers from the Quantum
Biology and Computational Physics research group at the University of Southern Denmark
are already investigating proton-coupled electron transfer in the cytochrome bc1 complex
in photosynthetic bacteria and higher organism cellular respiratory systems. The group
has published a number of papers motivated by the fact that malfunctions in the bc1
complex lead to many different diseases. They also hope to better understand photosyn-
thesis in order to optimise energy conversion research147 -- 149. The reduction and oxidation
of ubiquinone is essential to the electron transfer reactions in cytochrome bc1 and the
concentration of ubiquinone is particularly high in the high energy consumption environ-
ment of the brain150,151. There is some evidence that transcranial photobiomodulation, the
application of red or near-infrared laser light to the cranial area, is an effective treatment
for various brain disorders152,153. It has also shown promise in treating depression154 and
when coupled with ubiquinone supplementation151. Research suggests that, among other
effects, photobiomodulation can improve attention, memory, executive function and even
rule-based learning155 -- 157. The mechanism of this effect is still not completely clear but
the site of the effect is most likely the mitochondrial electron transfer chain, in particular
the chromophores in the different complexes that constitute the chain151,153,158. There are
a number of these chromophores, including tryptophan, which plays a central role in the
various theories that espouse quantum effects in neural processing. Given that the role of
quantum effects in light harvesting, electron and charge transfer has contributed to a better
understanding of elements of the electron transfer chain in photosynthesis, this knowledge
might be fruitfully applied to the mitochondrial electron transport chain.
2. Magnetic field effects and the brain
The importance of tryptophans in energy transfer is evident in another well established
field of quantum biology: the avian compass. A leading theory of bird migration suggests
that birds utilise the earth's magnetic field as a guiding tool by means of the radical pair
mechanism159 -- 161. The radical pair mechanism can be summarised in three main steps.
First, a photon incident on a molecule causes electron transfer and pair formation. Second,
the radical pair, originally in singlet spin state, interconverts between singlet and triplet
state under the influence of the nuclear hyperfine and geomagnetic Zeeman effects. And
24
finally spin-dependent recombination leads to some signalling state that the bird interprets
as a spatial directive161.
It is generally accepted that the molecule in which this occurs
is the blue light activated flavoprotein cryptochrome162 -- 165. Theoretical research, backed
by spectroscopic evidence, suggests that light activated electron transport occurs between
flavin adenine dinucleotide (FAD) and tryptophan residues161.
Although it is well known that a number of species have a functioning magnetic sense,
humans have not yet been added to that list. In a recent paper, Wang et al. present the
results of an experiment that demonstrates the effects of earth strength magnetic fields on
the human brain166. The authors report that magnetic field changes result in a decrease in
amplitude of alpha frequency (8-13 Hz) brain waves, an effect normally associated with the
brain's processing of external stimuli. They conclude that the effect is likely to be due to fer-
romagnetism rather than the radical pair mechanism. This is because the effect is dependent
on the polarity of the field, for subjects in the Northern Hemisphere the alpha wave response
only occurs for horizontal rotations if the static component is directed upwards166. While
this would appear to exclude the radical pair compass which is dependent on inclination,
the corresponding effects would perhaps need to be confirmed for subjects in the Southern
Hemisphere. It has been suggested, in the context of avian migration, that birds employ
both ferromagnetism and the radical pair mechanism for navigation160,167,168 and this could
also be true of humans' magnetic sense. The question being whether there is any evidence
that might link magnetic effects in humans with the radical pair mechanism.
Cryptochrome, the proposed site of the avian compass, is also present in humans169. Foley et
al. use a transgenic approach to show that human cryptochrome can act as magnetosensor
in the magnetoreception of fruit flies170. It is also potentially interesting that is it alpha
waves that are effected by the magnetic field changes. Van Wijk et al. present evidence
that biophoton fluctuation is correlated with the strength of alpha wave production, where
they measure biophoton production in terms of the fluctuations of reactive oxygen species
(ROS)124,171. The production of ROS and how this relates to the radical pair mechanism has
been the motivation for various papers published in the field of quantum biology. Marais
et al. outline a quantum protective mechanism in photosynthesis. They hypothesise that
the high spin iron in the reaction centres of photosystem II of the electron transport chain
25
exerts a magnetic field effect that reduces the triplet yield of radical pairs that are formed
during electron transfer. Triplet states are instrumental in forming ROS which are toxic to
living cells172,173. In another paper, Usselman et al. show how yields of ROS in live cells
are altered by radical pair dynamics, in particular coherent singlet-triplet mixing under the
influence of oscillating magnetic fields at Zeeman resonance174. If ROS are correlated with
alpha wave production124,171 and ROS yields are altered by the Zeeman effect174,175, then
this might point to a mechanism that explains the influence of geomagnetic fields on alpha
waves in the human brain.
Reactive oxygen species participate in cellular signaling176 and have been implicated in aging
and numerous diseases including mental conditions such as depression and schizophrenia177 -- 182.
If humans do have physiological systems that depend on the dynamics of radical pairs
then it is expedient to understand exactly how these function. The avian compass is
disrupted by broadband radiofrequency electromagnetic radiation that is well below the
WHO-recommended level183,184. This radiation might equally be upsetting the balance
of ROS in other species and in turn leading to biological malfunction. There has been
some suggestion that the radical pair mechanism is implicated in the development of can-
cer through circadian rhythm disruption and related ROS levels185. Though the threat of
radical pair mediated carcinogenesis is debatable, being equivalent, as one study finds, to
the risk of travelling some kilometres towards or away from the earth's magnetic poles186.
Nevertheless, the disruption of circadian rhythms has also been linked to mood disorders
and cognitive function187 -- 189. Although the evidence remains contentious, there is also some
documentation of the psychological effects of geomagnetic storms, which alter the earth's
magnetic field for a limited period of time190 -- 192. The radical pair mechanism could offer a
testable hypothesis as to how these effects occur190.
E. Neural entanglement
One of the first and strongest objections to quantum models of consciousness was the
phenomenon of decoherence: that the non-ideal environment of biological systems would
destroy any quantum effects before they could prove useful. As initially calculated by Max
Tegmark in response to Orch OR theory, the timescales on which decoherence occurs in
26
FIG. 4. A scheme for neural entanglement. The Posner molecule structure is interpreted loosely
here as a distorted cube with calcium ions (green) at each vertice and at the centre, and a phosphate
ion (yellow) on each of the six faces198. This diagram is not meant to accurately convey the exact
geometry of the Posner molecule but rather give some sense of Fisher's proposed qubit. The two
Posner molecules share an entangled pair of phosphate ions. When two entangled Posner molecules
are spatially separated and subsequently taken up into different presynaptic neurons in the process
of vesicle endocytosis, their entanglement results in the correlated release of neurotransmitters and
the coordination of firing across neurons84.
the environment of the brain are considerably shorter than neural firing rates47. This con-
clusion has subsequently been challenged not simply by researchers interested in Orch OR,
but also by the development of the field of quantum biology. If quantum effects play a role
in photosynthesis and possibly other biological contexts, then it is not such a stretch to
consider that they may play a role in neural processes. Decoherence, it has been argued,
27
might even enhance energy transfer193. While research suggests that long-lived coherence in
photosynthetic systems lasts for picoseconds194 and the lifetime of the radical pair mecha-
nism is discussed in terms of milliseconds161, it seems unlikely that coherence in biological
systems extends beyond these timescales. That is until a new hypothesis concerning neural
entanglement proposed coherence that could last for hours or even days.
1. Entangled Posner molecules
The hypothesis is this: that phosphorus nuclear spin might function as a neural qubit
to allow for quantum processing to play a role in cognition84 -- 86. Following on from the
proposal by Hu and Wu, that consciousness is linked to quantum spin195, Fisher argues that
the spin-half phosphorus nucleus seems the only likely candidate for this role due to the
fact that for any nucleus with spin greater than half, the electric quadropole moment means
quicker decoherence due to electric field interactions in addition to the slower decoherence
caused by the magnetic fields of nearby nuclei. Spin-half nuclei are thus more favourable
in terms of decoherence times. Out of the various elements and ions that play an essential
role in biological systems, only hydrogen and phosphorus have spin-half nuclei. Fisher goes
on to describe the details of how the quantum information carried by phosphorus nuclei is
created and preserved84 -- 86.
Phosphorus is bound into phosphate or polyphosphate ions such as pyrophosphate. Phos-
phate ions constitute part of the adenosine triphosphate (ATP) molecule, an organic chem-
ical that acts as a source of energy for the many essential actions that sustain living
organisms84,196. The electron transport chain in both photosynthesis and respiration cre-
ates a proton gradient in order to drive ATP production. Fisher postulates that when
adenosine triphosphate is hydrolysed to adenosine monophosphate and pyrophosphate the
two phosphorus nuclei in the pyrophosphate ion will have a specific spin alignment, either
a singlet or one of three triplet states. A quantum mechanical treatment of the enzyme
catalysed reaction that creates two phosphate ions out of pyrophosphate demonstrates that
this reaction depends on the nuclear spin state. More specifically, the enzyme conditional
outcome of the reaction, where the spin dynamics of the triplet states are unfavourable,
mean that the phosphorus nuclear spins in the two distinct phosphate ions will emerge in a
28
singlet entangled state84.
2. Coherence timescales for Posner molecules
Phosphorus in the phosphate ion is surrounded by an oxygen cage, where the oxygen
nuclei all have zero spin. Despite this the coherence time is still very short for solvated
phosphates, approximately one second, due to the fact that hydrogen quickly binds to phos-
phate and the non-zero proton spin contributes to decoherence. A coherence lifetime of a
second is long enough for the quantum effects to be sustained over the process of cellular
diffusion84. However in order for effects such as memory storage and retrieval to utilise
quantum effects, coherence lifetimes need to be significantly longer. Fisher argues that if
binding with the hydrogen can be pre-empted by a spin zero cation such as calcium, then
longer coherence times might be possible. He identifies a possible molecule as the Posner
molecule197, thought to be involved in the formation of hydroxyapatite, an important con-
stituent of bone tissue198 -- 200. A Posner molecule is, simplistically, a distorted cube with a
calcium ion at each vertice, a ninth at the centre and a phosphate ion on each of the six faces
of the cube198. The entangled phosphate spins then result in entangled Posner molecules84.
Fisher estimates that for phosphorus spins in Posner molecules coherence times could
be as long as hours84 or, in a later paper, 21 days85. For solvated Posner molecules the mag-
netic fields from protons in surrounding water molecules cause decoherence but, as Fisher
argues, the rapid tumbling of Posner molecules means that the average magnetic field felt
by a phosphorus nucleus will be zero. Decoherence will then only happen from residual
magnetic field fluctuations and coherence times will thus be much longer84,85. Player et al.
respond to this with their own calculation of coherence times for Posner molecules. They
acknowledge that long lived spin states in nuclei are well accepted, with relaxation times
of up to an hour being recorded198,201. However, they argue that these extended times are
partly due to experimental control of the coherent spin dynamics, a condition that Fisher
does not take into account198. In their analysis of Posner spin dynamics they take a number
of other factors into account such as dipolar and scalar couplings within Posner molecules as
well as the Zeeman interaction of the phosphorus spins with the geomagnetic field198. With
these as the dominant relaxation pathway they arrive at a relaxation time of 37 minutes,
29
as opposed to Fisher's shortest estimation of approximately a day. They also discuss other
ways in which this relaxation could take place even more quickly198.
3. Posner molecules in their biological context
Fisher also outlines the way in which entangled Posner molecules give rise to quantum
effects in neural processes. The entanglement of the spin and rotational states of Pos-
ner molecules leads to correlations in the chemical reactions of spatially separated Posner
molecules. These entangled molecules are taken into presynaptic glutamatergic neurons in
the process of vesicle endocytosis. The acidic environment causes Posner molecules to bind
and release calcium which stimulates exocytosis and further release of glutamate, which
enhances neural firing. Thus, because the chemical reactions of Posner molecules are entan-
gled the subsequent release of glutamate and resultant neural firing might also be considered
entangled84,86. Fisher outlines a number of possible experiments that might verify the var-
ious stages of the theory, most pressingly whether Posner molecules are present in bodily
fluids. Whereas free-floating calcium phosphate clusters resembling Posner molecules have
been found to be stable in simulated body fluids202, there is still some uncertainty as to
whether they are present in vivo84,85. Less fundamental but more interesting in the context
of this review is the suggested investigation of the effects of a chemically viable replacement
of the central calcium in a Posner molecule with lithium ions84,85. There is some evidence
that lithium, used to alleviate the symptoms of bipolar disorder, has isotope dependent
effects on the behaviour of rats84,203. Fisher postulates that the efficacy of lithium as a
treatment for mental disorders could be due to the increased decoherence induced by the
lithium nuclear spins included in the Posner molecule84,86.
4. Neural qubits and quantum computing
While the results of experiments to verify this model of entangled neural processes are
yet to be completed the theory has caught the interest of researchers working in the field of
quantum information theory. Halpern and Crosson apply Fisher's Posner molecule theory
30
in the context of quantum communication, quantum computation and quantum error cor-
rection. They address how a quantum information based model of Posner molecules might
contribute to the understanding of Posner chemistry and vice versa, what insights Posner
molecules have for quantum information processing. They also demonstrate how entan-
glement can change molecular binding rates, which goes some way to supporting Fisher's
neural entanglement hypothesis204. The authors conclude that the quantum information
based formulation of Posner molecules might be a way of framing biological Bell tests204, a
claim that is potentially interesting in light of recent observations of entanglement between
living bacteria and quantised light205.
This is not the first time that phosphorus spin and quantum computing have crossed
paths.
In 1998 Kane proposed a scalable quantum computer where information is en-
coded onto spin half phosphorus nuclei in a substrate of spin zero silicon, ensuring the long
decoherence time necessary for quantum computing. Computations are then performed
through the interaction of nuclear spin with donor electrons206. Kane's idea has recently
been reworked by Tosi et al.
into what they call the flip-flop qubit, where the combined
electron-nuclear spin states of a phosphorus donor are controlled by microwave electric
fields. This research, along with work done by He et al. into engineering a viable exchange
interaction between two phosphorus bound electrons has done much to advance the course
of spin based quantum computing207,208. The fields of quantum computing and quantum
neurobiology might inform each other in other ways too. Quantum dots have been proposed
as an alternative means to implement a quantum computer209. They have also been used
to model the mechanism by which anti-acetylcholine receptor antibodies contribute to the
neuromuscular disorder myasthenia gravis210. Even more recently, graphene quantum dots
have been shown to prevent and even undo the protein clumping of neurons in the brain
that leads to Parkinson's disease211. A separate study study demonstrated similar results
for Alzheimer's disease212.
III. CONCLUSION
The field of neurophysics already attests to the fact that the study of the brain borrows
from ideas across the spectrum of physics: electricity and magnetism, mechanics, thermo-
31
dynamics, optics15.
It is perhaps not completely surprising that quantum physics might
contribute too. While this review outlines working theories as to how quantum effects might
be implicated in neural processes, the research remains largely theoretical. Although some
experimental evidence points to the validity of certain of these theories, conflicting results
mean it is difficult to draw any strong conclusions. However, many of the authors cited in
this review suggest ways in which their theories might be put to the test experimentally.
As already suggested by Craddock et al. advances in experimental techniques such as two-
dimensional electronic spectroscopy, which has been successfully applied to understanding
the quantum nature of photosynthesis, might yield similar success in the context of energy
transfer processes that impact neural signalling64. A number of the hypotheses included in
this review rework, to some extent, ideas already established with respect to other biolog-
ical systems. Coherent energy transfer in photosynthesis is reimagined in the tryptophan
rings of neural microtubules. The vibrational theory of olfaction is reapplied to the binding
action of neurotransmitters. And while Fisher's Posner qubits depend on nuclear spin, un-
like the radical pair model developed with respect to avian migration, their entangled spin
dynamics mediate the outcome of signalling processes in a potentially analogous fashion.
As it stands, the future of quantum neurobiology is the future of quantum biology more
generally; progress made in either will further the other. The questions are then techni-
cal. How exactly to apply experimental techniques to the specific neural context. How to
refine, for example, experiments that measure the differing effects of isotopes without the
confounding influence of contaminants. This mechanistic focus would seem to ignore the
question of consciousness and how it emerges from these underlying neural mechanisms.
Lynn et al. describe the future of brain network research, albeit in a classical capacity, as
a cross-scale approach that links the microscopic to the macroscopic: protein reactions in
neurons to synaptic connectivity to brain region connectivity to social networks15. Jedlicka,
in his review of the future of quantum neurobiology, touches on the evidence that quantum
physics has also been used to describe human behaviour10 -- 12. The use of a quantum frame-
work to accurately describe human behaviour does not necessarily mean that this behaviour
results from quantum effects in neural processes. A clearer understanding of these processes,
however, could also elucidate to what extent quantum physics is involved in the emergence
of cognition from neural activity10. Should it turn out that experimental evidence supports
hypotheses of quantum enhanced neural processing we might turn this knowledge towards
32
thinking about how we think.
FUNDING AND ACKNOWLEDGMENTS
B.A and F.P. were supported by the South African Research Chair Initiative of the
Department of Science and Technology and the National Research Foundation. B.A. was
also supported by the National Institute for Theoretical Physics. Thank you to Angela Illing
for the diagrams.
REFERENCES
1J. Al-Khalili, J. McFadden, Life on the edge:
the coming of age of quantum biology,
London, UK Bantam Press (2014).
2J. McFadden and J. Al-Khalili, Proc. R. Soc. A 474, (2018).
3M. Mohseni,Y. Omar, G.S. Engel, M.B. Plenio (eds). Quantum effects in biology, Cam-
bridge, UK Cambridge University Press (2013).
4N. Lambert, Y.N. Chen, Y. Cheng, C. Li, G. Chen and F. Nori, Nature Physics 9, 10
(2013).
5A. Marais et al., J. R. Soc. Interface 15, 20180640 (2018).
6Zettili N., Quantum Mechanics Concepts and Applications: Second Edition John Wiley
& Sons Ltd, UK (2009).
7Haken H & Wolf HC., Atomic and Quantum Physics Springer-Verlag, Berlin Heidelberg
(1987).
8R. Penrose, The emperor's new mind: Concerning computers, minds, and the laws of
physics, New York, US: Oxford University Press, (1989).
9S.R. Hameroff and R. Penrose, Physics of Life Reviews 11(1), 39 (2014).
10P. Jedlicka, Front Mol Neurosci. 10, 366 (2017).
11P.D. Bruza, Z. Wang, J.R. Busemeyer J, Trends Cogn. Sci. 19, 383 (2015).
12J.R. Busemeyer, P. Fakhari, P. Kvam, Prog. Biophys. Mol. Biol. 130, 53 (2017).
13A. Louveau et al., Nature Neuroscience 21, 1380 (2018).
14A. Louveau et al., Nature 523, 337 (2015).
15C.W. Lynn, D.S. Bassett, Nature Reviews Physics 1, 318 (2019).
33
16K. Batista-Garca-Ram1 and C.I. Fern´andez-Verdecia, Behav Sci 8(4), 39 (2018).
17B. V´azquez-Rodr´ıguez et al., Proceedings of the National Academy of Sciences, 201903403
(2019).
18A. Mess´e, D. Rudrauf, H. Benali and G. Marrelec, PLoS Comput Biol 10(3), e1003530
(2014).
19F.A. Azevedo, L.R. Carvalho, L.T. Grinberg, J.M. Farfel, R.E. Ferretti, R.E. Leite, F.W.
Jacob, R. Lent, S. Herculano-Houzel, J Comp Neurol. 513(5), 532 (2009).
20A. Araque and M. Navarrete, Philos Trans R Soc Lond B Biol Sci. 365(1551), 2375
(2010).
21A. Tameem, H. Krovvidi, Continuing Education in Anaesthesia Critical Care & Pain
13(4), 113 (2013).
22A.L. Hodgkin and A.F. Huxley, J Physiol 117(4), (1952).
23A.E. Pereda, Nat Rev Neurosci. 15(4), 250 (2014).
24H. Lodish, A. Berk, S.L. Zipursky, P. Matsudaira, D. Baltimore and J. Darnell, Molecular
Cell Biology, 4th edition, W.H. Freeman New York (2000).
25F. Cavanna, M.G. Vilas, M. Palmucci, E. Tagliazucchi, NeuroImage 180, 383 (2018).
26A.E. Cavanna, A. Nani, H. Blumenfeld, S. Laureys, Neuroimaging of Consciousness,
Springer Berlin (2013).
27A.M. Owen, Annual Review of Psychology 64, 109 (2013).
28B. Rohaut, A. Eliseyev and J. Claassen, Critical Care 23, (2019).
29M. Boly, M. Massimini, N. Tsuchiya, B.R. Postle, C. Koch and G. Tononi, Journal of
Neuroscience 37(40), 9603 (2017).
30J. Aru, M. Suzuki, R. Rutiku, M.E. Larkum and T. Bachmann, Front. Syst. Neurosci.
13, 43 (2019).
31D. Godwin, R.L. Barry and R. Marois, PNAS 112(12), 3799 (2015).
32G. Tononi, Biol Bull. 215(3), 216 (2008).
33G. Tononi, M. Boly, M. Massimini, C. Koch, Nat. Rev. Neurosci. 17, 450 (2016).
34X.D. Arsiwalla1 and P. Verschure, Front Neurosci. 12, 424 (2018).
35A. de Sousa, Mens Sana Monogr. 11(1), 100 (2013).
36S.R. Hameroff and R. Penrose, Math Comput Simul 40 453 (1996).
37S.R. Hameroff and R. Penrose, J Conscious Stud 3(1), 36 (1996).
38S. Meunier, I. Vernos, Journal of Cell Science 125, 2805 (2012).
34
39M.J. Shelley, Annual Review of Fluid Mechanics 48, 487 (2016).
40P. Dustin, Microtubules (2nd edition), Springer-Verlag, New York (1985).
41G. Guillaud, C. Bosc, A. Fourest-Lieuvin, E. Denarier, F. Priollet, L. Lafanach`ere, D.
Job, Cell Biol 142, 167 (1998).
42T. Craddock, J. Tuszynski, S. Hameroff, PLoS Comput Biol 8(3), e1002421 (2012).
43T.J. Craddock, D. Friesen, J. Mane, S. Hameroff, J. Tuszynski, J R Soc Interface
6;11(100), 20140677 (2014).
44J.T. Toole, P. Kurian, T.J.A. Craddock, Journal of Cognitive Science 19, 115 (2018).
45S. Hameroff, J Biol Phys. 36(1), 71 (2010).
46S. Hameroff, Cogn Sci 31, 1035 (2007).
47M. Tegmark, Physical review. E 61(4), 4194 (2000).
48S. Hagan, S. Hameroff, J. Tuszynski, Phys Rev E 65, 061901 (2001).
49G.S. Engel et al., Nature 446, 782 (2007).
50T. Brixner, J. Stenger, H.M. Vaswani, M. Cho, R.E. Blankenship and G.R. Fleming,
Nature 434, 625 (2005).
51R. van Grondelle and V.I. Novoderezhkin, Proc. Chem. 3, 198 (2011).
52G.S. Schlau-Cohen et al., J. Phys. Chem. B 113, 15352 (2009).
53G. Panitchayangkoon et al. Proc. Natl. Acad. Sci. USA 107, 12766 (2010).
54E. Collini et al. Nature 463, 644 (2010).
55E. Mandelkow, Y.H. Song, E.M. Mandelkow, Trends Cell Biol 5(7), 262 (1992).
56L.K. McKemmish, J.R. Reimers, R.H. McKenzie, A.E. Mark, N.S. Hush, Phys Rev E 80,
021912 (2009).
57J.R. Reimers, L.K. McKemmish, R.H. McKenzie, A.E. Mark, N.S. Hush, Proc Natl Acad
Sci USA 106(11), 4219 (2009).
58J. Pokorn´y, Bioelectrochemistry 63, 321 (2004).
59S. Sahu, S. Ghosh, B. Ghosh, K. Aswani, K. Hirata, D. Fujita, A. Bandyopadhyay, Biosens
Bioelectron 47, 141 (2013).
60S. Sahu, S. Ghosh, K. Hirata, D. Fujita, A. Bandyopadhyay, Phys Lett 102, 123701
(2013).
61D. Emerson, B. Weiser, J. Psonis, Z. Liao, O. Taratula, A. Fiamengo, X. Wang, K.
Sugasawa, A.B. Smith, R.G. Eckenhoff and I.J. Dmochowski, J Am Chem Soc 135(14),
5398 (2013).
35
62M. Bianchi, A.J. Shah, K.C. Fone, A.R. Atkins, L.A. Dawson, C.A. Heidbreder, M.E.
Hows, J.J. Hagan, C.A. Marsden, Synapse 63, 359 (2009).
63A. Rinaldi, EMBO Rep. 15(11), 1113 (2014).
64T.J.A. Craddock, S.R. Hameroff, A.T. Ayoub, M. Klobukowski and J.A. Tuszynski, Cur-
rent Topics in Medicinal Chemistry 15, 523 (2015).
65L. Turin, E.M.C. Skoulakis and A.P. Horsfield, PNAS 111(34), E3524 (2014).
66S.R. Hameroff, R.C. Watt, J.D. Borel and G. Carlson, Physiol. Chem. Phys. 14, 183
(1982).
67S.R. Hameroff and R.C. Watt, Anesth. Analg. 62, 936 (1983).
68S.R. Hameroff, Toxicol. Lett. 100-101, 31 (1998).
69S.R. Hameroff, Anesthesiology 105(2), 400 (2006).
70N. Li et al. Anesthesiology 129(2), 271 (2018).
71L.J. Voss, P.S. Garcia, H. Hentschke and M.I. Banks, Anesthesiology 130(6), 1049 (2019).
72C.D. Richards, BJA: British Journal of Anaesthesia 89(1), 79 (2002).
73M. Perouansky, Anesthesiology 117, 465 (2012).
74E.F. Keller, L.A. Segel, J. Theor. Biol. 26, 399 (1970).
75T. Craddock et al., PloS one 7, e37251 (2012).
76R.G. Eckenhoff, J.S. Johansson, Pharmacol Rev. 49(4), 343 (1997).
77R.G. Eckenhoff, Anesth Analg. 107(3), 859 (2008).
78J.M. Sonner, R.S. Cantor Annu Rev Biophys 42, 143 (2013).
79J.Z. Pan, J. Xi, M.F. Eckenhoff, R.G. Eckenhoff, Proteomics 8, 2983 (2008).
80T. Craddock, A. Priel and J. Tuszynski, Journal of integrative neuroscience 13, 293
(2014).
81G. Celardo, M. Angeli, P. Kurian and T. Craddock, New Journal of Physics 21, (2018).
82T.J.A. Craddock, P. Kurian, J. Preto, K. Sahu, S.R. Hameroff, M. Klobukowski and J.A.
Tuszynski, Scientific Reports 7(9877), (2017).
83A. Albert, Selective Toxicity: The Physico-Chemical Basis of Therapy, Chapman & Hall,
London (1985).
84M.P.A. Fisher, Annals of Physics 362, (2015).
85M.W. Swift, C.G. Van de Walle and M.P.A. Fisher, Physical Chemistry Chemical Physics
20(18), (2017).
36
86C.P. Weingarten, P.M. Doraiswamy and M.P.A. Fisher, Frontiers in Human Neuroscience
10, (2016)
87D.M. Eigler, E.K. Schweizer, Nature 344, 524 (1990).
88C.J. Rourk, Biosystems 171, 48 (2018).
89L. Turin and E.M.C. Skoulakis, Methods Enzymol. 603, 115 (2018).
90J.M. Kater, Science 82(2124), 256 (1935).
91P.J. Cohen, Anesthesiology 39(2), 153 (1973).
92S.A. Pshenichnyuk and A. Modelli, Phys Chem Phys 15(23), 9125 (2013).
93A. Gaitanidis, A. Sotgiu, L. Turin, arXiv:1907.04764 [physics.bio-ph] (2019).
94R.K. Burdick, J.P. Villabona-Monsalve, G.A. Mashour and T. Goodson III, Scientific
Reports 9(11351), (2019).
95J. Pan et al., Translational Psychiatry 8(130), (2018).
96D.J. Nutt, J Clin Psychiatry 69,4 (2008).
97A. Vaziri and M.B. Plenio, New Journal of Physics 12, (2010)
98E. Fischer, Ber. Dtsch. Chem. Ges.27, 2985 (1894).
99A. Tripathi and V.A. Bankaitis, J Mol Med Clin Appl. 2(1), (2017).
100H. Mitsuda et al., Journal of Physical Chemistry Letters 1(7), 1130 (2010).
101A. Reese, N. Holmgaard List, J. Kongsted, I.A. Solovyov, PLOS ONE 11(3), e0152345
(2016).
102M.D.H. Samee, B.G. Bruneau, K.S. Pollard, Cell Systems 8(1), 27 (2019).
103M. Rask-Andersen, M.S. Alm´en, H.B. Schioth, Nat. Rev. Drug Discov. 10, 579 (2011).
104A. Gehrckens, A. Horsfield, E. Skoulakis and L. Turin, preprint on biorxiv 10.1101/650531,
(2019).
105A.P. Horsfield, A. Haase and L. Turin, Advances in Physics: X 2:3, 937 (2017).
106G.M. Dyson, J. Soc. Chem. Industry 57, 647 (1938).
107L. Turin, Chemical Senses 21(6), 773 (1996).
108Franco MI1, Turin L, Mershin A, Skoulakis EM, Proc Natl Acad Sci U S A 108(9), 3797
(2011).
109R.D. Hoehn, D. Nichols, H. Neven and S Kais, Sci Rep. 5, 9990 (2015).
110J. Hara, Experientia 33(5), 618 (1977).
111B.R. Havens and C.D. Melone, Dev. Food. Sci. 37, 497 (1995)
112K.D. Klika, ISRN Org Chem, 515810 (2013).
37
113M. Paoli, D. Mnch, A. Haase, E. Skoulakis, L. Turin and C. G. Galizia, eNeuro 4(3),
0070 (2017).
114H.K Chee and S.J. Oh, Genomics Inform 11(4), 282 (2013).
115H.K. Chee, J. Yang, J. Joung, B. Zhang, S.J. Oh, FEBS Letters 589(4), 548 (2015).
116Ross D. Hoehn, David E. Nichols, John D. McCorvy, Hartmut Neven, and Sabre Kais,
PNAS 114(22), 5595 (2017).
117S. Lin, L. Lee and Y.K. Yang, Clin Psychopharmacol Neurosci. 12(3), 196 (2014).
118M. Krzan, R. Vianello, A. Marsavelski, M. Repic, M. Zaksek, K. Kotnik, E. Fijan and J.
Mavri, PLoS One 11(5), e0154002 (2016).
119S.J. Oh, Genomics Inf. 10(2), 128 (2012).
120K. Palczewski, Annu Rev Biochem. 75, 743 (2007).
121A. Krishnan, M.S. Alm´en, R. Fredriksson, H.B. Schioth, PLoS One, 7(1), e29817 (2012).
122D. Wacker, R.C. Stevens and B.L. Roth, Cell 170(3), 414 (2017).
123D. Fels, PLoS One 4, e5086 (2009).
124M. Rahnama, J.A. Tuszynski, I. B´okkon, M. Cifra, P. Sardar, J. Salari, Integr Neurosci.
10(1), 65 (2011).
125S. Kumar, K. Boone, J. Tuszy´nski, P. Barclay and C. Simon, Scientific Reports 6, 36508
(2016).
126M. Cifra, P.J. Posp´ısil, Photochem Photobiol B 139, 2 (2014).
127M. Kobayashi, M. Takeda, T. Sato, Y. Yamazaki, K. Kaneko, K. Ito, H. Kato, Humio
Inaba, Neurosci. Res. 34, 103 (1999).
128Y. Isojima, T. Isoshima, K. Nagai, K. Kikuchi and H. Nakagawa, NeuroReport 6, 658
(1995).
129Y. Kataoka et al., Biochem. Biophys. Res. Commun. 285, 1007 (2001).
130R. Tang and J. Dai, PLoS One 9, e85643 (2014).
131Z. Wang, N. Wang, Z. Li, F. Xiao and J. Dai, PNAS 113, 8753 (2016).
132V. Salari, I. B´okkon, R. Ghobadi, F. Scholkmann and J.A. Tuszynski, PNAS 113, E5540
(2016).
133Y. Sun, C. Wang, J. Dai, Photochem Photobiol Sci. 9(3), 315 (2010).
134P. Zarkeshian, S. Kumar, J. Tuszynski, P. Barclay, C. Simon, Front Biosci (Landmark
Ed). 23, 1407 (2018).
135A.I. Zhuravlev, O.P. Tsvylev and S.M. Zubkova, Biofizika 18, 1037 (1973).
38
136J.A. Tuszy´nski and J.M. Dixon, Phys. Rev. E 64, 051915 (2001).
137G. Albrecht-Buehler, Cell Motil. Cytoskeleton 27, 262 (1994).
138M. Kato, K. Shinzawa and S. Yoshikawa, Photobiochem. Photobiophys. 2, 263 (1981).
139P. Kurian, T.O. Obisesan, and T. Craddock, Journal of Photochemistry and Photobiology
B: Biology 175, (2017).
140L. Tonello, M. Cocchi, F. Gabrielli, J.A. Tuszynski, J Integr Neurosci. 14(3), 295 (2015).
141L. Tonello, B. Gashi, A. Scuotto, G. Cappello, M. Cocchi, F. Gabrielli and J.A. Tuszynski,
Journal of Integrative Neuroscience 17, 177 (2019).
142S. Baghbanzadeh and I. Kassal, J. Phys. Chem. Lett. 7, 3804 (2016).
143M.W. Gray, G. Burger and B.F. Lang, Science 283, 1476 (1999).
144Z. Wang and M. Wu, Sci. Rep. 5, 7949 (2015).
145A. Harish and C.G. Kurland, Journal of Theoretical Biology 434, 88 (2017).
146J. Martijn, J. Vosseberg, L. Guy, P. Offre and T.J.G. Ettema, Nature 557, 101 (2018).
147P. Husen, I.A. Solov'yov, Journal of Physical Chemistry B 121 3308 (2017).
148A.B. Salo, P. Husen, I.A. Solov'yov, Journal of Physical Chemistry B 121, 1771 (2017).
149A.M. Barragan, K. Schulten, I.A. Solov'yov, Journal of Physical Chemistry B 120, 11369
(2016).
150D. Xia, L. Esser, W. Tang, F. Zhou, Y. Zhou, L.Yu, C. Yu, Biochimica et Biophysica
Acta (BBA) - Bioenergetics 1827, 1278 (20130.
151F. Salehpour et al., Brain Research Bulletin 144, 213 (2019).
152M.R. Hamblin, BBA Clinical 6, 113 (2016).
153M.R. Hamblin, J Neurosci Res. 96(4), 731 (2018).
154J. Chang, Y. Ren, R. Wang, C. Li, Y. Wang, X. Chu, Xiang-Ping. Neuropsychiatry 08,
(2018).
155D.W. Barrett, F. Gonzalez-Lima, Neuroscience 230, 13 (2013).
156N.J. Blanco, W.T. Maddox and F. Gonzalez-Lima, J Neuropsychol. 11(1), 14 (2017).
157N.J.Blanco, C.L.Saucedo, F.Gonzalez-Lima, Neurobiology of Learning and Memory 139,
69 (2017).
158T.I. Karu and S.F. Kolyakov, Photomed Laser Surg. 23(4), 355 (2005).
159K. Schulten, C.E. Swenberg and A. Weller Z. Phys. Chem. 111. 1 (1978).
160W. Wiltschko and R. Wiltschko, J Comp Physiol A 191, 675 (2005).
161C.T. Rodgers and P.J. Hore, PNAS 106(2), 353 (2009).
39
162T. Ritz, T. Yoshii, C. Helfrich-Forster and M. Ahmad, Communicative and Integrative
Biology 3, 24 (2010).
163A. Pinzon-Rodriguez, S. Bensch, R. Muheim, J. R. Soc. Interface 15, 20180058 (2018).
164A. Gunther, A. Einwich, E. Sjulstok, R. Feederle, P. Bolte, K.W. Koch, I.A. Solov'yov,
H. Mouritsen, Curr Biol. 28, 211 (2018).
165M. Liedvogel and H. Mouritsen, J. R. Soc. Interface 7, S147 (2014).
166C.X. Wang, I.A. Hilburn, D. Wu, Y. Mizuhara, C.P. Coust´e, J.N.H. Abrahams, S.E.
Bernstein, A. Matani, S. Shimojo and J.L. Kirschvink, eNeuro 6(2), 0483 (2019).
167R. Wiltschko and W. Wiltschko, Commun Integr Biol. 2(2), 100 (2009).
168D.A. Kishkinev, N.S. Chernetsov, Zh Obshch Biol. 75(2), 104 (2014).
169D.S. Hsu et al., Biochemistry 35, 13871 (1996).
170L.E. Foley, R.J. Gegear and S.M. Reppert, Nat Commun. 2, 356 (2011).
171R. Van Wijk, S. Bosman, J. Ackerman, E.P.A. Van Wijk, NeuroQuantology 6, 452 (2008).
172C. Schweitzer and R. Schmidt, Chem. Rev. 103, 1685 (2003).
173A. Marais, I. Sinayskiy, F. Petruccione and R. van Grondelle, Scientific Reports 5, 8720
(2015).
174R.J. Usselman, C. Chavarriaga, P.R. Castello, M. Procopio, T. Ritz, E.A. Dratz, D.J.
Singel and C.F. Martino, Scientific Reports 6, 38543 (2016).
175R.J. Usselman , I. Hill, D.J. Singel, C.F. Martino, PLoS ONE 9(3), e93065 (2014).
176H. Wang and X. Zhang, Int J Mol Sci. 18(10), 2175 (2017).
177M.K. Shigenaga, T.M. Hagen,and B.N. Ames, Proc. Natl. Acad. Sci. 91, 10771 (1994).
178E.R. Stadtman and B.S. Berlett, Chem. Res. Toxicol. 10, 485 (1997).
179T.M. Michel, D. Pulschen, J. Thome, Curr Pharm Des. 18(36), 5890 (2012).
180S. Salim, J Pharmacol Exp Ther. 360(1), 201 (2017).
181S. Salim, Curr Neuropharmacol. 12(2), 140 (2014).
182P. Kovacic and W. Weston, Chronicles of Pharmaceutical Science 1(6), (2017).
183P. Thalau, T. Ritz, K. Stapput, R. Wiltschko and W. Wiltschko, Naturwissenschaften
92, 86 (2004).
184S. Engels, N.L. Schneider, N. Lefeldt, C.M. Hein, M. Zapka, A. Michalik, D. Elbers, A.
Kittel, P.J. Hore and H. Mouritsen, Nature 509, 353 (2014).
185J. Juutilainen, M. Herrala, J. Luukkonen, J. Naarala, P.J. Hore, Proc Biol Sci.
285(1879), 20180590 (2018).
40
186P.J. Hore, eLife 8, (2019).
187L.M. Lyell et al., Proc Biol Sci. 285(1879), 20180590 (2018).
188P. Boyce, E.A. Barriball, Fam Physician 39(5),307 (2010).
189A. Ferguson et al., EBioMedicine 35, 279 (2018).
190J. Close, Proc Biol Sci. 279(1736), 2081 (2012).
191R.W. Kay, Br. J. Psychiatry 164, 403 (1994).
192V.N. Oraevskii, V.P. Kuleshova, I.F. Gurfinkel, A.V. Guseva, S.I. Rapoport, Biofizika 43,
844 (1998).
193M.B. Plenio and S.F. Huelga, New J. Phys. 10, 113019 (2008).
194G.S. Engel, Procedia Chemistry 3, 222 (2011).
195H. Hu and M. Wu, Medical Hypotheses 63, 633 (2004).
196L.A. Cole, Biology of Life: Biochemistry, Physiology and Philosophy Chapter 10, 65
(2016).
197A.S. Posner and F. Betts, Acc Chem Res. 8(8), 273 (1975).
198T.C Player and P.J. Hore, J. R. Soc. Interface 15(147), (2018).
199L. Wang, S. Li, E. Ruiz-Agudo, C.V. Putnis, A. Putnis, Cryst. Eng Comm. 14(19), 6252
(2012).
200G. Mancardi, C.E.H. Tamargo, T.D. Di, N.H. de Leeuw, J Mater Chem B. 5(35), 7274
(2017).
201G. Stevanato et al., Angew Chem Int Ed. 54(12), 3740 (2015).
202K. Onuma and A. Ito, Chem. Mater. 10, 3346 (1998).
203J.A. Sechzer et al., Biol. Psychiatry 21, 1258 (1986).
204N.Y Halpern and E. Crosson, Annals of Physics 407, 92 (2019).
205C. Marletto et al., J. Phys. Commun. 2, 101001 (2018).
206B.E. Kane, Nature 393, 133 (1998).
207G. Tosi, F.A. Mohiyaddin, V. Schmitt, S. Tenberg, R. Rahman, G. Klimeck and A.
Morello, Nature Communications 8, (2017).
208Y. He, S.K. Gorman, D. Keith, L. Kranz, J. G. Keizer and M.Y. Simmons, Nature 571,
371 (2019).
209C. Kloeffel and D. Loss, Annual Review of Condensed Matter Physics 4, 51 (2013).
210C.W. Lee, H. Zhang, L. Geng and H.B. Peng, PLoS One 9(2), e90187 (2014).
211D. Kim et al., Nature Nanotechnology 13, 812 (2018).
41
212Y. Liu, L. Xu, W. Dai, H. Dong, Y. Wen and X. Zhang, Nanoscale,7, 19060 (2017).
213V.M. Mazhul, and D.G. Shcherbin, Biofizika 44, 676 (1999).
42
|
1209.3829 | 1 | 1209 | 2012-09-18T02:25:50 | Self-organized criticality in a network of interacting neurons | [
"q-bio.NC",
"physics.bio-ph"
] | This paper contains an analysis of a simple neural network that exhibits self-organized criticality. Such criticality follows from the combination of a simple neural network with an excitatory feedback loop that generates bistability, in combination with an anti-Hebbian synapse in its input pathway. Using the methods of statistical field theory, we show how one can formulate the stochastic dynamics of such a network as the action of a path integral, which we then investigate using renormalization group methods. The results indicate that the network exhibits hysteresis in switching back and forward between its two stable states, each of which loses its stability at a saddle-node bifurcation. The renormalization group analysis shows that the fluctuations in the neighborhood of such bifurcations have the signature of directed percolation. Thus the network states undergo the neural analog of a phase transition in the universality class of directed percolation. The network replicates precisely the behavior of the original sand-pile model of Bak, Tang & Wiesenfeld. | q-bio.NC | q-bio | Self-organized criticality in a network of interacting
neurons
J D Cowan1, J Neuman2, W van Drongelen3
1 Dept. of Mathematics, University of Chicago, 5734 S. University Ave., Chicago, IL
60637
2 Dept. of Physics, University of Chicago, 5720 S. Ellis Ave., Chicago, IL, 60637
3 Dept. of Pediatrics, University of Chicago, KCBD 900 E. 57th St., Chicago, IL.,
60637
E-mail: [email protected]
Abstract. This paper contains an analysis of a simple neural network that exhibits
self-organized criticality. Such criticality follows from the combination of a simple
neural network with an excitatory feedback loop that generates bistability,
in
combination with an anti-Hebbian synapse in its input pathway. Using the methods
of statistical field theory, we show how one can formulate the stochastic dynamics
of such a network as the action of a path integral, which we then investigate using
renormalization group methods. The results indicate that the network exhibits
hysteresis in switching back and forward between its two stable states, each of which
loses its stability at a saddle-node bifurcation. The renormalization group analysis
shows that the fluctuations in the neighborhood of such bifurcations have the signature
of directed percolation. Thus the network states undergo the neural analog of a phase
transition in the universality class of directed percolation. The network replicates
precisely the behavior of the original sand-pile model of Bak, Tang & Wiesenfeld.
2
1
0
2
p
e
S
8
1
]
.
C
N
o
i
b
-
q
[
1
v
9
2
8
3
.
9
0
2
1
:
v
i
X
r
a
Self-organized criticality in a neural network
2
1. Introduction
The idea of self-organized criticality (SOC) was introduced by Bak et al. (1988). Their
paper immediately triggered an avalanche of papers on the topic, not the least of which
was a connection with 1/f - or scale-free noise. However it was not until another paper
appeared, by Gil & Sornette (1996), which greatly clarified the dynamical prerequisites
for achieving SOC, that a real understanding developed of the essential requirements
for SOC: (1) an order-parameter equation for a dynamical system with a time-constant
τo, with stable states separated by a threshold, (2) a control-parameter equation with
a time-constant τc, and (3) a steady driving force. In Bak et.al.'s classic example, the
sand-pile model, the order parameter is the rate of flow of sand grains down a sand-pile,
the control parameter is the sand-pile's slope, and the driving force is a steady flow of
grains of sand onto the top of the pile. Gil and Sornette showed that if τo (cid:28) τc then the
resulting avalanches of sand down the pile would have a scale-free distribution, whereas
if τo (cid:29) τc then the distribution would also exhibit one or more large avalanches.
In this paper we will analyze a neural network model which is in one-to-one
correspondence with the Gil-Sornette SOC-model, and therefore also exhibits SOC.
1.1. Neural network dynamics
Consider first the mathematical representation of the dynamics of a neocortical slab
comprising a single spatially homogeneous network of N excitatory neurons. Such
neurons make transitions from a quiescent state q to an activated state a at the rate σ
and back again to the quiescent state q at the rate α, as shown in figure 1.
Figure 1. Neural state transitions
Let Pn(t) be the probability that a fraction n/N is active at time t. Then the
probabilistic dynamics of such a slab can be formulated as a master equation of the
Self-organized criticality in a neural network
form:
dPn(t)
dt
= α[(n + 1)Pn+1 − nPn] + (N − n + 1)f [s(n − 1)]Pn−1
− (N − n)f [s(n)]Pn
3
(1)
where α is the rate at which activated neurons become quiescent, and s(n) is the total
current or excitation driving each neuron in the population to fire at the rate σ = f [s(n)].
We assume in this simplified model that each neuron receives a signal weighted by w0/N
from each of N other neurons in the population, so that
s(n) = w0n + h
where h is an external current.
(2)
Equation 1 can be extended to the spatially inhomogeneous case. Let nr/Nr be
the fraction of active cells at time t in the rth population of Nr cells, and let P [n, t] be
the probability of the configuration n = {n1, n2,··· , nr,··· , nΩ} existing at time t. The
extended master equation then takes the form
(cid:88)
(cid:88)
r
dP [n, t]
dt
= α
[(nr + 1)P [nr+, t] − nrP [n, t]]
[(Nr − nr + 1)f [s(nr − 1)]P [nr−, t]
+
− (Nr − nr)f [s(nr)]P [n, t]]
r
(4)
(5)
(6)
(7)
(3)
where nr± = {n1, n2,··· , nr ± 1,··· , nΩ} and there are a total of Ω locally homogeneous
populations.
Following Van Kampen (1981) equation 3 can be rewritten using the one-step
operators
i.e.,
E±
r f (n) = f (nr ± 1)
(cid:88)
(cid:2)α(E +
dP [n, t]
dt
=
r − 1)nr + (E−
r − 1)(Nr − nr)f [s(nr)](cid:3) P [n, t]
Note that the total number of cells Nr in the rth population comprises nr active
cells and Nr − nr = qr quiescent cells, so that
nr + qr = Nr
Thus equation 5 can be rewritten slightly in the form
(cid:88)
(cid:2)α(E +
dP [n, t]
dt
=
r − 1)nr + (E−
r − 1)qrf [s(nr)](cid:3) P [n, t]
r
r
Self-organized criticality in a neural network
4
1.2. Annihilation and creation operators
We now introduce Fock space annihilation and creation operators satisfying boson
commutation rules
[ar, a†
[ar, as] = [a†
[qr, qs] = [q†
r, a†s] = 0
r, q†
s] = 0
s] = δ(r − s)
s] = [qr, q†
(8)
such that given neural vectors nr(cid:105), mr(cid:48)(cid:105)
rnr(cid:105) = nr + 1(cid:105), arnr(cid:105) = nrnr − 1(cid:105)
a†
rmr(cid:105) = mr + 1(cid:105), qrmr(cid:105) = mrmr − 1(cid:105)
q†
where
r0r(cid:105),mr(cid:105) = (q†
r 0r(cid:105)
r)m
and 0r(cid:105) is a fiducial or vacuum (empty) state such that
nr + mr = 1,nr(cid:105) = (a†)n
r0r(cid:105) = nr = 1r(cid:105), ar0r(cid:105) = 0
a†
r0r(cid:105) = mr = 1r(cid:105), qr0r(cid:105) = 0
q†
and there exists a dual vector (cid:104)0 such that
r = 0,
r = 0,
(cid:104)0rar = (cid:104)nr = 1r,(cid:104)0ra†
(cid:104)0rqr = (cid:104)mr = 1r,(cid:104)0rq†
It follows from equation 9 that
rarnr(cid:105) = nrnr(cid:105)
a†
(9)
(10)
(11)
(12)
i.e. nr is the eigenvalue of the operator a†
or number density operator.
rar, which is therefore referred to as a number
1.3. Bosonizing the master equation
The master equation can now be bosonized by replacing the van Kampen one-step
operators by bosonic equivalents, i.e.
r)ar
r)qr
(E +
r − 1) = (q†
r − a†
r − 1) = (a†
(E−
r − q†
(cid:88)
so eqn 7 becomes:
(cid:88)
r)ar + (a†
Let P (t)(cid:105) be a probability state vector satisfying
(cid:2)α(q†
r − a†
dP [n, t]
dt
=
r
P (t)(cid:105) =
P [n, t]n(cid:105)
n
rar)](cid:3) P [n, t]
r − q†
r)qrf [s(a†
(13)
(14)
(15)
P (t)(cid:105) =
r − a†
r)ar + (a†
r − q†
r)qrf [s(a†
rar)](cid:3)P (t)(cid:105)
Self-organized criticality in a neural network
Then equation 7 can be written in the form
(cid:88)
(cid:2)α(q†
r
or formally as
d
dt
d
dt
where
− H =
P (t)(cid:105) = − HP (t)(cid:105)
(cid:88)
(cid:2)α(q†
r − a†
r)ar + (a†
r − q†
r)qrf [s(a†
rar)](cid:3)
5
(16)
(17)
(18)
(20)
(21)
is the quasi-Hamiltonian operator for the Markov process represented in equation 1.
r
1.4. From bosons to coherent states
Equation 18 is a linear operator equation with formal solution
P (t)(cid:105) = exp[− H(t − t0)P (t0)(cid:105)
We need to re-express this solution in terms of numbers rather than operators. This
can be achieved by introducing coherent states. These were introduced by Schrodinger
(1926) and first used extensively in coherent optics by Glauber (1963). We therefore
introduce such states φr(cid:105) in the form
φr(cid:105) = exp[−1
2
rϕr + ϕra†
ϕ(cid:63)
r]0r(cid:105)
(19)
where ϕr is the right eigenvalue of ar, i.e. arφr(cid:105) = ϕrφr(cid:105). There is also a coherent
state representation of qr in the form θr(cid:105) such that the right eigenvalue of qr is ϑr, i.e.
qrθr(cid:105) = ϑrθr(cid:105). In similar fashion (cid:104)φra†
r = (cid:104)φr ϕr where ϕr, the complex conjugate of
r = (cid:104)θr ϑr, i.e. ϑr is the left eigenvalue
ϕ, is the left eigenvalue of a†
of q†
r, and similarly (cid:104)θrq†
All this suggests that the operator quasi-Hamiltonian has a coherent state representation
in the form
rarφr(cid:105) = (cid:104)φr ϕrϕrφr(cid:105) = ϕrϕr
(cid:88)
(cid:104)
− H =
α( ϑr − ϕr)ϕr + ( ϕr − ϑr)ϑrf [s( ϕrϕr)]
(cid:105)
r. It follows that
(cid:104)φra†
r
Note in passing that operator products involving powers of the number operator
a†
rar must first be normal ordered, i.e. all creation operators a†
r must preceed the
annihilation operators ar, before coherent states can be introduced. For example the
normal order form of exp[a†a] written as : exp[a†a] : is expanded as
: exp[a†a] : = 1 + (a†a) +
= 1 + (a†a) +
(a†a)2 + ···
(a†a + a†2a2) + ···
1
2!
1
2!
sl,ja†jaj
∞(cid:88)
l(cid:88)
1
l!
l=0
j=0
=
(22)
Self-organized criticality in a neural network
6
where the sl,j are Stirling numbers of the second kind. It follows that : exp[a†a] : can
be written as
: exp[a†a] :=
hka†kak
(23)
(cid:88)
k=0
where hk =(cid:80)
l sl,k.
The final preliminary step in this formulation is to take the continuum limit of the
expression for H in equation 21, so that
(cid:90) (cid:90)
(cid:104)
− H =
(24)
where ϕr → ϕ(x, t) ≡ ϕ etc., and the conjugate coherent state ϕ has been shifted to
1 + ϕ.
α( ϑ − ϕ)ϕ + ( ϕ − ϑ)ϑf [s( ϕϕ + ϕ)]
ddxdt
(cid:105)
1.5. Dimensions and the density representation
Before proceeding further we need to assign a dimension to each variable in equation 24.
To do so we use a modified version of the convention used in particle physics so
that [x] = L−1, [t] = L−2 where L is the length scale used, whence [x2/t] = L0.
{This generates a scaling found in Markov random walks and related processes such
as stochastic neural activity.} Then [α] = L2,
[ ϕϕ] = Ld,
[f [s]] = [α] = L2. This last value of [f [s]] implies that the input current function
s( ϕϕ + ϕ) = s(I) = kI where the constant k has the dimensions of inverse current. The
net effect of such a choice leads to the required result that [H] = 0.
[ϕ] = Ld,
[ ϕ] = L0,
To emphasize this choice we further transform the coherent-state quasi-Hamiltonian
by introducing the density representation:
ϕ → exp[n] − 1, ϕ → n[exp[−[n]]
ϑ → exp[p] − 1, ϑ → p[exp[−[p]]
so that equation 24 transforms into
(cid:90) (cid:90)
− H =
ddxdt [α(exp(p − n) − 1)n + (exp(n − p) − 1)pf [s(n)]]
(26)
k(w(cid:63)n+h) where (cid:63) is the spatial convolution operator, i.e. w(cid:63)n =(cid:82) ddx(cid:48)w(x−x(cid:48))n(x(cid:48), t).
Note that in the continuum limit the input current function s(n) = s(I) = kI =
(25)
1.6. From the quasi-Hamiltonian to a neural Path Integral
Using standard methods (Doi 1976, Peliti 1985) Buice & Cowan (2007) incorporated
the quasi-Hamiltonian into the action of a Wiener path integral. This action takes the
form:
(cid:90) (cid:90)
S(n) =
ddxdt [n∂tn + p∂tp+
α(1 − exp(−(n − p))n − (exp(n − p) − 1)pf [s(n)]](27)
Self-organized criticality in a neural network
7
The utility of this action is that it is part of the exponent of the moment generating
functional for the statistical moments of the probability density P [n, t].
At an extremum
(cid:12)(cid:12)(cid:12)(cid:12)n=0
(cid:12)(cid:12)(cid:12)(cid:12)p=0
δS(n)
δn
=
δS(n)
δ p
= 0
which leads to the mean-field equations
∂tn + αn − pf [s(n)] = 0,
∂tp − αn + pf [s(n)] = 0
whence
n + p = ρ
(28)
(29)
(30)
where ρ is the (constant) packing density of excitatory neurons in the population. This
is a mean-field result, so we replace p by ρ − ncl in equation 27, so as to generate
the mean-field Wilson-Cowan equation (Wilson & Cowan 1973) for a single spatially
organized population from the first variation of the action given in equation 27 and (in
principle), all higher moments, and (finally) rewrite equation 27 in the form:
(cid:90) (cid:90)
S(n) =
ddxdt [n∂tn+
α(1 − exp(−n))n − (exp(n) − 1)(ρ − ncl)f [s(n)]] (31)
1.7. Renormalizing the path integral
We expand n about its mean value (cid:104)n(cid:105) = ncl, which satisfies equation 28 in the form:
(32)
∂tncl = −αncl + (ρ − ncl)f [s(ncl)]
where
s(ncl) = k(w (cid:63) ncl + hcl)
(33)
Thus n → n+ncl, n → n, since ncl = 0. So s(n) → s(n+ncl) and f [s(n)] → f [s(n+ncl)].
If follows that s(n + ncl) = k(w (cid:63) (n + ncl) + (h + hcl)) = k(w (cid:63) ncl + hcl) + k(w (cid:63) n + h)) =
s(ncl) + s(n), and therefore f [s(n)] = f [s(ncl) + s(n)]. We next expand f [s(n)] in
a Taylor expansion about the mean-field value ncl, noting that from equation A.1,
s(n) = k(w (cid:63) n + h)) = k(Ln + h).
In what immediately follows we assume that
the external stimulus h(x, t) = 0. It follows that:
f [s(n)] = f [kL(ncl + n)] = f [kLncl] + f (1)[kLncl]kLn
However because of normal ordering, equation 34 leads to the expression:
(cid:88)
f [s(n)] =
gm(kLn)m, where
gm =
m
l=0
f (2)[kLncl](kLn)2 + ···
+
1
2
m(cid:88)
f (l)
l!
sl,m
(34)
(35)
Since the leading terms of gm are proportional to f (m), and given the assumed form for
f [s(n)] to be such that f (1) > 0 and f (2) < 0, then gm > 0 for m odd, and gm < 0 for m
even.
Self-organized criticality in a neural network
We also expand the functions exp(±n). The resulting action S(n) takes the form:
8
(cid:90) (cid:90)
S(n) =
ddxdt [n(∂t + α − (ρ − ncl)g1kL)n
(α + (ρ − ncl)g1kL)n + n((ρ − ncl)g2(kL)2)n2
((ρ − ncl)g2(kL)2)n2 + ···
(cid:21)
(36)
− n2
2
n2
2
+
It follows from the appendix that w2 is small compared to w0, so that in most
expressions the terms proportional to ∇2mnm can be neglected. However this is not
always the case for m = 1. Thus the first term can be written approximately as
2!w2∇2))n = n(∂t + µ− D∇2)n where µ = α− (ρ− ncl)g1kw0
n(∂t + α− (ρ− ncl)g1k(w0 + 1
2(ρ − ncl)g1kw2. So the expression for the action is now reduced to the form:
and D = 1
(cid:90) (cid:90)
S(n) =
ddxdt(cid:2)n(∂t + µ − D∇2)n − n2G1n
(cid:21)
+nG2n2 +
n2G2n2 + ···
1
2
(37)
where G1 = 1/2(α+(ρ−ncl)g1kw0), and G2 = (ρ−ncl)g2k2w2
that the last term in S(n), i.e., 1
of the renormalization group.
0. We need to demonstrate
2 n2G2n2, and all other terms, are irrelevant in the sense
The renormalization group [RG] analysis
is carried out via dimensional
analysis.
It can be shown that all the terms in S(n) are zero-dimensional when
[ddxdt] = L−(d+2) and
integrated over d-dimensional space and over time,
[any term in the integrand] = Ld+2. However, as it stands [n] = Ld, but [n] = L0,
so that [nn] = L0+d = Ld. This is not suitable for the scaling analysis implemented in
the RG process. We therefore introduce a new scaling,
i.e.,
(cid:114) G1
(cid:114)G2
n,
s =
s =
(38)
such that ss = nn where [G2/G1] = L−d. The effect of this scaling is that both s and s
have dimension Ld/2. Let
G2
n
G1
G1G2 = u, G2 = 2τ
(39)
The net effect of this scaling transformation is that
ddxdt(cid:2)s(∂t + µ − D∇2)s + us(s − s)s + τ s2s2 + ···(cid:3)
S(s) =
(40)
But [τ /u] = L−d/2 and therefore scales to zero as L → ∞ under subsequent RG
transforms. So asymptotically the terms τ s2s2 + ··· become irrelevant in the RG sense.
So finally
ddxdt(cid:2)s(∂t + µ − D∇2)s + us(s − s)s(cid:3)
S(s) =
(41)
(cid:90) (cid:90)
(cid:90) (cid:90)
(cid:112)
is the renormalized action of the large-scale neural activity of a single neural population.
Self-organized criticality in a neural network
9
1.8. Directed Percolation
This action is well-known: it is called Reggeon Field Theory, and is found in directed
percolation [DP] in random graphs, in contact processes, in high-energy nuclear physics,
in bacterial colonies, all of which exhibit the characteristic properties of what is called
a universality class, i.e., it is a phase transition with a universal scaling of important
statistical exponents.
It also shows up in branching and annihilating random walks,
catalytic reactions, and interacting particles. Thus we have mapped the mathematics of
large-scale neural activity in a single homogeneous neural population into a percolation
problem in random graphs, or equivalently into a branching and annihilating random
walk. A first version of this work was presented in Buice & Cowan (2007). A more
extensive paper with many applications to neuroscience was presented in Buice & Cowan
(2009).
Here we note that there is an upper critical dimension at which directed percolation
crosses over to mean-field behavior. This upper critical dimension is d = 4. What is
the dimension of the neocortex? To answer this question we look at the number of
synapses per neuron in the neocortex. Using estimates provided by Stevens (1989),
this number is about 4 × 103. Assuming the number of synapses in a terminal axonal
arbor to be about 50, the number of neural neighbors per neuron is about 80, so the
effective dimensionality of a neocortical hyper lattice is about d = 40. Thus the critical
exponents characterizing the neural phase transition are the d = 4 exponents of directed
percolation. These have been calculated by Abarbanel & Broznan (1974), Abarbanel
et al. (1976), and Amati et al. (1976), and appear in the linear response of the neocortical
model to an impulsive stimulus, known to mathematicians as the Green's function and
to physicists as the propagator. This takes the form
G(x − x(cid:48), t − t(cid:48)) ∝
(t − t(cid:48))−2 exp
(cid:105)
(cid:104)− x−x(cid:48))2
(cid:105)
(cid:104)− x−x(cid:48))2
4(t−t(cid:48)) − µ(t − t(cid:48))
(cid:104)(cid:112)µ(t − t(cid:48)) − x − x(cid:48)(cid:105)
4(t−t(cid:48)) )
(t − t(cid:48))−2 exp
µ2Θ
, µ > 0
µ = 0
(42)
µ < 0
,
,
where the cases µ > 0, µ = 0, and µ < 0 correspond, respectively, to the sub-critical,
critical, and super-critical propagators. They correspond, respectively, to solutions of
the cable equation, the diffusion equation, and a nonlinear wave-equation. The critical
propagator is thus the diffusion limit of Brownian motion.
It turns out that there
is a great deal of data supporting the hypothesis that the mean-field propagator of
DP correctly describes the essential features of large-scale neocortical activity on many
spatio-temporal scales. [See Burns (1951), Lampl et al. (1999), Nauhaus et al. (2009).]
In addition data on the statistical structure of large-scale activity recorded in cortical
slices by Beggs & Plenz (2003) supports the hypothesis. In particular, the avalanche-size
distribution of spontaneous activity in cortical slices fits the DP hypothesis.
The analysis can be extended to deal with a neural network comprising both
excitatory and inhibitory neurons. However in this paper we describe how to incorporate
synaptic plasticity into a network comprising only excitatory neurons, as a mechanism
Self-organized criticality in a neural network
10
that tunes the network so that it automatically reaches the critical point of the DP
phase transition, thus exhibiting self-organized criticality.
2. Incorporating synaptic plasticity
Consider first a single excitatory population model with a fixed recurrent excitatory
synapse wE and an input H through an excitatory modifiable synapse wH. The mean-
Figure 2. A recurrent excitatory network
field neural equation for this is a version of equation 32, i.e.
dnE
dt
= −αEnE + (1 − nE)σE [sE(IE)]
where
sE(IE) = IE/IRH,E,
IE = wE (cid:63) nE + wH (cid:63) nH
and the synapse wH is modifiable with
dwH
dt
= −gE(cid:104)(nE − ρE,0 − ρE,SwH) nH(cid:105)t
(43)
(44)
(45)
where ρE,0 is a constant neural activity, ρE,S is a constant (Vogels et al. 2011), and gE
is a state-dependent rate function.
This is also a mean-field equation in which the synaptic weight wH is depressed by
an anti-Hebbian mechanism, and potentiated by the input activity nH. The problem
is to write an action that incorporates these equations. Note that the time scale of
the growth and decay of neural activity is set by the constant αE, whereas that of
the growth and decay of synaptic plasticity is set by gE. Thus the ratio αE/gE is an
important parameter.
3. Deriving an action for synaptic plasticity
The first problem is to develop an action for the modifiable synapse wH. In order to do
so we first note from equation A.2 that wH scales with bH, the synaptic conductance or
weight, and we can write an approximation to eqn. 45 in the form:
dbH
dt
= −gE (nE − ρE,0 − ρE,SkHbH) nH
(46)
Self-organized criticality in a neural network
11
We next reformulate the changes in bH as a Markov process with discrete states
in continuous time. We therefore assume that bH is quantized in units ∆ of synaptic
weight, and similarly for bE. Thus
bE = mE∆,
bH = mH∆
(47)
and we look at the Markov process represented in figure 3:
Figure 3. Synaptic state transitions
and we write a master equation for this process in the form
dP (mH, t)
dt
= t−(mH + 1)P (mH + 1, t) − t−(mH)P (mH, t)
+ t+(mH − 1)P (mH − 1, t) − t+(mH)P (mH, t)
where P (m, t) is the probability that the synaptic weight mH = m at time t. and
t+ = gE(ρE,0 + ρE,SkHmH)nH
t− = gEnEnH
are the transition rates for the excitatory synapse mH.
3.1. The van Kampen ladder operators
We again introduce the van Kampen ladder operators
so that the master equation can be rewritten as:
− 1) + t+(E−
dP (mH, t)
mH
− 1)(cid:3) P (mH, t)
mm = m ± 1
E±
=(cid:2)t−(E+
mH
dt
(48)
(49)
(50)
(51)
An examination of eqns. 48-50 indicates that only the transition rate t+ contains a
term proportional to mH. We will utilize this property in what follows.
Self-organized criticality in a neural network
12
3.2. From the Master Equation to the Action
We now introduce bosonic annihilation and creation operators for mH. Let such
operators be denoted by s† and s respectively, and let m(cid:105) be a column vector
representing the synaptic weight m such that
s†m(cid:105) = m + 1(cid:105),
sm(cid:105) = mm − 1(cid:105)
(52)
Note that s† and s satisfy the same commutation rules and equations that we introduced
earlier, so that
m − 1 → s† − s†s = s†(1 − s)
E+
m − 1 → s − s†s = (1 − s†)s
E−
So we consider the equation:
∂ m(cid:105)
∂t
= (E−
m − 1)m(cid:105) = (1 − s†)sm(cid:105)
We further note that (1 − s†)s is normal ordered. We therefore shift s† and s,
s† → 1 + s, s → s
where s and s are now coherent states, and introduce the density representation
s → exp( m) − 1,
s → m exp(− m)
we find
∂tm(cid:105) = − ssm(cid:105)
= − (exp( m) − 1)m exp(− m)m(cid:105)
= − (1 − exp(− m))mm(cid:105)
Eqn. 21 tells us that the action Sm must contain a term of the form
−gEρE,SkHnH(1 − exp(− mH))mH,
a source term
−gEρE,0nH mH,
and an interaction term of the form
(53)
(54)
(55)
(56)
(57)
(59)
(60)
leading to an action of the form:
+gEnEnH mH,
S(mH) =
ddxdt [ mH∂tmH − gEρE,SkHnH(1 − exp(− mH))mH
− mHgEρE,0nH + mHgEnEnH]
(58)
Using variational techniques, we can derive the mean-field equation [eqn 45] from the
condition:
(cid:90) (cid:90)
(cid:12)(cid:12)(cid:12)(cid:12) mH =0
δS( mH)
δ mH
= 0
For then we obtain the equation:
dmH
dt
= −gE (nE − ρE,0 − ρE,SkHmH) nH
i.e., eqn. 46.
Self-organized criticality in a neural network
13
3.3. Renormalizing the synaptic plasticity action
We now proceed to renormalize the action S(mH) just as we renormalized S(nE). We
therefore expand the exponential term in equation 58 and rewrite S(mH) in the form:
(cid:90) (cid:90)
(cid:20)
S(mH) =
ddxdt
mH∂tmH − H1 mHmHnH +
−H2 mHnH + mHgEnEnH]
1
2
H1 m2
HmHnH
(61)
where H1 = gEρE,SkH, and H2 = gEρE,0.
We now introduce the scaling
(cid:114)H2
H1
(cid:114) H1
H2
mH,
sH =
sH =
(62)
such that sHsH − mHmH, and [H1/H2] = L−d. This scaling is analogous to the scaling
of n and n which we carried out earlier for neural activities. As before the effect of this
scaling is that both sH and sH have dimension Ld/2.
Let
mH
(cid:112)
and recall that equation 38 scales nE to(cid:112)G1/G2sE.
H1H2 = uH, H1 = 2τH
(63)
Following the procedure outlined earlier we can calculate which terms in the
transformed action S(sH) become irrelevant under scaling transformations. The
resulting renormalized synaptic plasticity action takes the form:
S(sH) =
ddxdt [sH∂tsH − uH sHnH]
(64)
(cid:90) (cid:90)
4. Combining the actions
(cid:90) (cid:90)
It follows from this formulation that the full action for the coupled system of equations
for the evolution of nE and mH can be obtained simply by adding the actions S(nE)
and S(mH) together. The combined action therefore takes the form:
ddxdt [nE∂tnE + α(1 − exp(−nE))nE
−(exp(nE) − 1)(ρ − nE,cl)f [s(nE)] + mH∂tmH
−gEρE,SkHnH(1 − exp(− mH))mH
− mHgEρE,0nH + mHgEnEnH]
S(nE, mH) =
(65)
4.1. A simulation of the behavior of the combined mean-field equations
The first variation of equation 65 generates the mean-field equations for nE and mHin
the form:
dnE
dt
dmH
dt
= − αEnE + (1 − nE)σE [sE(IE)]
= − gE (nE − ρE,0 − ρE,SkHmH) nH
(66)
Self-organized criticality in a neural network
where
sE(IE) = kE(mE (cid:63) nE + mH (cid:63) nH)
14
(67)
These equations can be simulated. The results are shown in figure 4. It will be seen that
Figure 4. Neural state transitions between a ground state and an excited state.
Parameter values: mE = 3, nH = 3; α = 0.2. N (cid:63) is the fixed-point value of nE,
and WH is the magnitude of the anti-Hebbian synapse in the input path.
in the "ground-state" of low values of N (cid:63) = n(cid:63)
E the synaptic weight wH (proportional to
mH), increases until it reaches the critical point at a saddle-node bifurcation, at which
point N (cid:63) becomes unstable and the system switches to the "excited-state". But then
the anti-Hebbian term in the synaptic plasticity dynamics kicks in, and WH declines
until the excited-state fixed-point becomes unstable at the upper critical point, also
at a saddle-node, and switches back to the ground-state fixed point, following which
the hysteresis cycle starts over. This is a exact representation of the sand-pile model's
behavior. This is another representation of SOC in a neural network. The reader should
compare this with the synaptic mechanisms for achieving SOC described in Levina et al.
(2007) and in Millman et al. (2010).
4.2. Renormalizing the combined action
To renormalize the combined action we follow the same procedure as before and
expand the exponential functions exp[±nE], exp[± mH], and f [s(nE)]. Note that after
normal ordering and collecting terms, f [s(nE)] can be expanded in the extended form
f [s(nE)] =(cid:80)
renormalized combined action takes the form:
(cid:90) (cid:90)
Self-organized criticality in a neural network
15
m gm(kE(LEnE + LHnH))m. After scaling and dimensional analysis, the
ddxdt(cid:2)sE(∂t + µE − DE∇2)sE + uE sE(sE − sE)sE
S(sE, sH) =
−vE sEnH + sH∂tsH − uH sHnH]
(68)
where uE, uH and vE are renormalized constants.
It would appear that apart from
the source term vE sEnH the coupled action is just the sum of the two uncoupled
renormalized actions. This is indeed the case! All the addition terms which appear
in the current function become irrelevant under renormalization, and do not effect the
renormalized action representing nE. Similarly for mH.
It follows that fluctuations
in the activity nE in the neighborhood of the critical point µE = 0, i.e. those in the
fluctuation-driven regime, should be essentially those of directed percolation.
5. Inhibitory synapses
In case wH is an inhibitory synapse the transitions t+(m) and t−(m) are reversed. so
that:
t+ = gEnEnH
t− = gE(ρE,0 + ρE,SkHmH)nH
(69)
Thus now only the transition rate t− contains a term proportional to mH. The effect of
this is that the action for Sm leads to the mean-field equation:
dmH
dt
= gE (nE − ρE,0 − ρE,SkHmH) nH
at an inhibitory synapse mH.
(70)
Thus inhibitory feedforward synapses are Hebbian with stimulus dependent
depression, whereas excitatory feedforward synapses are anti-Hebbian with stimulus
dependent potentiation, and we have now formulated actions for feedforward excitatory
and inhibitory synaptic plasticity, based on simple microscopic potentiation and
depression processes, involving a unary variable, the synaptic weight mH.
6. Conclusion
In this paper we have indicated how one can formulate and analyze actions for a
simple network of excitatory cells with an input coupled to the network via an activity-
dependent modifiable synapse. This action allows, in principal, the computation of
statistical moments of the fluctuating dynamics of the network. Previous work by Wilson
& Cowan (1972) indicates that the mean-field dynamics of the network is bistable, and
can generate a hysteresis loop. On coupling this dynamics to that of the modifiable
feedforward synapse introduced, all the conditions for the achievement of SOC are
present in the combined system. The simulation of the mean-field dynamics shows
that SOC is indeed achieved. It only remains to simulate the behavior of the combined
Self-organized criticality in a neural network
16
system with intrinsic noise. Our prediction is that the fluctuations in the activity near
the critical points of the system will exhibit the properties of directed percolation. This
will be the subject of another paper.
Acknowledgments
The work reported in this paper was initially developed in large part with Michael. A.
Buice. The current work was supported (in part) by a grant to Wim van Drongelen
from the Dr. Ralph & Marian Falk Medical Trust.
Appendix
Appendix A.1. Expanding the weighting function
We approximate the convolution w (cid:63) n as:
w (cid:63) n =
ddx(cid:48)w(x − x(cid:48))n(x(cid:48), t)
(cid:90)
w2∇2 + ···)n(x(cid:48), t)
(cid:39) (w0 +
≡ Ln
1
2!
where
w(x) → w(r) = b/σde−r/σ, σ = r0.
It follows that
w0 =
w2 =
(cid:90)
(cid:90)
ddxw(x) = b
dΓ(d)
Γ(d/2 + 1)
πd/2,
ddx2w(x) = bσ2 dΓ(d + 2)
Γ(d/2 + 1)
πd/2
(A.1)
(A.2)
(A.3)
In case d = 3, w0 = 8πb, w2 = 12w0σ2, whence w2/w0 = 12σ2. This expansion of the
weighting function w(x) is known as the moment expansion.
References
Abarbanel H & Broznan J 1974 Phys. Lett. B 48, 345–348.
Abarbanel H, Broznan J, Schwimmer A & Sugar R 1976 Phys. Rev. D 14(2), 632–646.
Amati D, Marchesini G, Ciafoloni M & Parisi G 1976 Nuclear Physics B 114, 483–504.
Bak P, Tang C & Wiesenfeld K 1988 Physical Review A 38(1), 364–374.
Beggs J & Plenz D 2003 J. Neurosci. 23(35), 11167–11177.
Buice M A & Cowan J D 2007 Physical Review E 75, 051919.
Buice M & Cowan J 2009 Prog. Biophys. Theor. Biol. 99(2,3), 53–86.
Burns B D 1951 J. Physiol. 112, 156–175.
Doi M 1976 J. Phys. A: Math Gen. 9(9), 1465–1477.
Gil L & Sornette D 1996 Physical Review Letters 76(21), 3991–3994.
Glauber R 1963 Phys. Rev. Lett. 10(3), 84–86.
Lampl I, Reichova I & Ferster D 1999 Neuron 22, 361–374.
Self-organized criticality in a neural network
17
Levina A, Herrmann J & Geisel T 2007 Nature Physics 3(12), 857–860.
Millman D, Mihalas S, Kirkwood A & Niebur E 2010 Nature Physics 6(10), 801–805.
Nauhaus I, Busse L, Carandini M & Ringach D 2009 Nature Neurosci. 12(1), 70–76.
Peliti L 1985 J. Physique 46, 1469–1483.
Schrodinger E 1926 Naturwissenschaften 14, 664–666.
Stevens C 1989 Neural Comp. 1, 473–479.
Van Kampen N 1981 Stochastic Processes in Physics and Chemistry North Holland.
Vogels T, Sprekeler H, Zenke F, Clopath C & Gerstner W 2011 Science 334(6062), 664–666.
Wilson H & Cowan J 1972 Biophys. J. 12, 1–22.
Wilson H R & Cowan J D 1973 Kybernetik 13, 55–80.
|
1801.04623 | 1 | 1801 | 2018-01-15T00:10:29 | Sex differences in network controllability as a predictor of executive function in youth | [
"q-bio.NC",
"eess.SY"
] | Executive function emerges late in development and displays different developmental trends in males and females. Sex differences in executive function in youth have been linked to vulnerability to psychopathology as well as to behaviors that impinge on health. Yet, the neurobiological basis of these differences is not well understood. Here we test the hypothesis that sex differences in executive function in youth stem from sex differences in the controllability of structural brain networks as they rewire over development. Combining methods from network neuroscience and network control theory, we characterize the network control properties of structural brain networks estimated from diffusion imaging data acquired in males and females in a sample of 882 youth aged 8-22 years. We summarize the control properties of these networks by estimating average and modal controllability, two statistics that probe the ease with which brain areas can drive the network towards easy- versus difficult-to-reach states. We find that females have higher modal controllability in frontal, parietal, and subcortical regions while males have higher average controllability in frontal and subcortical regions. Furthermore, average controllability values in the medial frontal cortex and subcortex, both higher in males, are negatively related to executive function. Finally, we find that average controllability predicts sex-dependent individual differences in activation during an n-back working memory task. Taken together, our findings support the notion that sex differences in the controllability of structural brain networks can partially explain sex differences in executive function. Controllability of structural brain networks also predicts features of task-relevant activation, suggesting the potential for controllability to represent context-specific constraints on network state more generally. | q-bio.NC | q-bio | Sex differences in network controllability as a predictor of executive
function in youth
Eli J. Cornblath1,2, Evelyn Tang2, Graham L. Baum1,2,3, Tyler M. Moore3, David R.
Roalf3, Ruben C. Gur3, Raquel E. Gur3, Fabio Pasqualetti4, Theodore D. Satterthwaite3,*,
and Danielle S. Bassett2,5,6,7,*
1Department of Neuroscience, University of Pennsylvania, Philadelphia, PA 19104 USA
2Department of Bioengineering, University of Pennsylvania, Philadelphia, PA 19104 USA
3Department of Psychiatry, University of Pennsylvania, Philadelphia, PA 19104 USA
4Department of Mechanical Engineering, University of California, Riverside, CA 92521 USA
5Department of Electrical and Systems Engineering, University of Pennsylvania,
Philadelphia, PA 19104 USA
6Department of Neurology, University of Pennsylvania, Philadelphia, PA 19104 USA
7To whom correspondence should be addressed: [email protected].
*These authors contributed equally.
January 16, 2018
8
1
0
2
n
a
J
5
1
]
.
C
N
o
i
b
-
q
[
1
v
3
2
6
4
0
.
1
0
8
1
:
v
i
X
r
a
1
Abstract
Executive function is a quintessential human capacity that emerges late in development and dis-
plays different developmental trends in males and females. Sex differences in executive function in
youth have been linked to vulnerability to psychopathology as well as to behaviors that impinge
on health, wellbeing, and longevity. Yet, the neurobiological basis of these differences is not well
understood, in part due to the spatiotemporal complexity inherent in patterns of brain network
maturation supporting executive function. Here we test the hypothesis that sex differences in exec-
utive function in youth stem from sex differences in the controllability of structural brain networks
as they rewire over development. Combining methods from network neuroscience and network con-
trol theory, we characterize the network control properties of structural brain networks estimated
from diffusion imaging data acquired in males and females in a sample of 882 youth aged 8-22 years.
We summarize the control properties of these networks by estimating average and modal control-
lability, two statistics that probe the ease with which brain areas can drive the network towards
easy- versus difficult-to-reach states. We find that females have higher modal controllability in
frontal, parietal, and subcortical regions while males have higher average controllability in frontal
and subcortical regions. Furthermore, average controllability values in the medial frontal cortex
and subcortex, both higher in males, are negatively related to executive function. Finally, we find
that average controllability predicts sex-dependent individual differences in activation during an
n-back working memory task. Taken together, our findings support the notion that sex differences
in the controllability of structural brain networks can partially explain sex differences in executive
function. Controllability of structural brain networks also predicts features of task-relevant activa-
tion, suggesting the potential for controllability to represent context-specific constraints on network
state more generally.
Keywords:
network controllability, neurodevelopment, sex differences, executive function, working memory,
fMRI BOLD, diffusion tensor imaging
Introduction
Executive function is necessary for regulation of goal-directed behavior, and encompasses cogni-
tive processes including working memory, inhibition, task switching, and performance monitor-
ing 1. Deficits in executive function are associated with increased risk taking and associated con-
sequences 2,3, and more generally hamper academic and occupational performance 4. Executive
deficits frequently lead to personal, social, and professional consequences that accumulate through-
out the course of a patient's life 4.
Importantly, the normative capacity for executive function
rapidly increases during adolescence and differs by sex. Sex differences in the developmental tra-
jectory of executive function have been linked to higher rates of impulsivity 5, ADHD diagnosis 6,
criminality 7 and substance use 2 in males. Current interventions for disorders of executive function
are relatively limited, usually do not consider sex, and rely primarily on psychotherapy and global
manipulations via psychopharmacology 8.
A basic understanding of sex-related differences in executive function and their implications for the
diagnosis and treatment of executive deficits requires an understanding of the normative matura-
tion of underlying neural circuitry. Several recent studies highlight the fact that such maturation,
and sex-differences in that maturation, span structure 9, anatomical connectivity 10, functional ac-
tivity 11,12,13, and functional connectivity 14. Using structural MRI, a recent study 15 reported age-
related, non-linear increases in gray matter density with concurrent decreases in cortical thickness.
Interestingly, the maturation of these structural features was markedly different between the sexes,
with females showing higher gray matter density globally and higher cortical thickness in frontal
and insular regions 15. Several older structural studies 16 have also reported linear increases in fron-
toparietal white matter density throughout adolescence. Using diffusion-weighted MRI, another
study found significantly greater within-hemisphere connectivity in males and greater between-
hemisphere connectivity in females 17. Sex differences have also been identified in the clustered (or
modular) structure in patterns of functional connectivity estimated from resting state fMRI data:
males display higher between-module connectivity while females display higher within-module con-
nectivity 18, and these differences were shown to predict individual differences in executive func-
tion 18. Although these descriptive studies have provided important insights, it remains difficult to
specify in a mechanistic sense how executive function might arise from such complex, multimodal
patterns of brain maturation in a sex-dependent manner.
We address this challenge by positing that sex differences in executive function in youth stem from
sex differences in the controllability of structural brain networks as they rewire over development.
This notion intuitively bridges the control of behavior (executive function) with the control of brain
dynamics (network controllability). Specifically, we capitalize on recent advances in network control
theory 19,20, an emerging branch of theoretical physics and systems engineering that builds on early
efforts in control theory 21,22 to offer a mechanistic model of how key nodes, or control points, can
exert disproportionate influence over system function 19. Control points are identified with metrics
that assess the ability of specific nodes to alter a system's state, based on the underlying network
topology 20 (Fig. 1). Specifically, the metric of average controllability reflects the average energy
input required at a node to move the system from some initial state to all possible states.
In
contrast, the metric of modal controllability reflects the ease of transitioning the system from some
initial state to a difficult-to-reach state. Prior work has demonstrated the utility of network control
theory in understanding basic brain architecture and function 23,24,25 across spatial scales 26 and
3
species 27, posited its relation to cognition 28,29, and outlined its developmental course 29.
Here, we tested the hypothesis that developmental sex differences in network control underlie sex
differences in executive functioning. Specifically, we predicted that (i) network controllability differs
by sex, (ii) network controllability changes with age differently in males and females, (iii) sex differ-
ences in network controllability predict executive function, and (iv) network controllability predicts
the activation of brain regions during a working memory task demanding executive function. To
test these hypotheses, we constructed structural brain networks from diffusion tensor imaging data
acquired in 882 healthy youth, ages 8-22 years, in the Philadelphia Neurodevelopmental Cohort
(PNC) 30. Each brain network was comprised of 234 anatomically defined brain regions 31 connected
by white matter tracts estimated from diffusion tractography. We show that regional controllability
is a significant mediator of the relationship between sex and executive function, and that it predicts
the magnitude of fMRI BOLD signal on an n-back working memory task. As described in detail
below, our results suggest that sex differences in the controllability of structural brain networks
predict executive function and the activity profiles that support that function.
Materials and Methods
Participants
Diffusion tensor imaging (DTI) data were obtained from youth who participated in a large community-
based study of brain development, now known as the Philadelphia Neurodevelopmental Cohort
(PNC) 30. Here we study 882 out of a total of 1601 subjects between the ages of 8 and 22 years
(mean age = 15.06, SD = 3.15, 389 males, 493 females). Due to lack of complete diffusion scans
(n = 224) and incidental findings (n = 20), data from 244 participants was deemed unusable. The
remaining 1357 participants underwent a rigorous manual and automated quality assurance proto-
col for DTI datasets 32, eliminating an additional 147 subjects with poor data quality. A subset of
93 of the remaining 1210 participants were excluded for low quality or incomplete FreeSurfer re-
construction of T1-weighted images. Further, 235 of the remaining 1117 participants were excluded
for one or more of the following reasons: gross radiological abnormalities distorting brain anatomy,
medical history that might impact brain function, history of inpatient psychiatric hospitalization,
use of psychotropic medication at the time of imaging, or high levels of in-scanner head motion
during the DTI scan, as defined by a mean relative displacement between non-weighted volumes of
greater than 2 mm. These exclusions left us with a final sample of n = 882 subjects 10,29 between
the ages of 8 and 22 years (mean age = 15.06, SD = 3.15, 389 males, 493 females).
Cognitive Phenotyping
Cognition was measured outside of the scanner using the Penn Computerized Neurocognitive Bat-
tery (CNB) 33,34. Briefly, the 1-hour CNB was administered to all participants, and consisted of
14 tests that evaluated a broad range of cognitive functions. Twelve of the tests measure both
accuracy and speed, while two of the tests (motor and sensorimotor) measure only speed.
Overall cognitive performance was summarized as the average z-transformed accuracy and speed (as
well as the difference between accuracy and median response time: efficiency) scores across all tests
administered (as described in Moore et al. 35 for complete details). Factor scores are described in
4
Moore et al. 35; here, we used the factor score for executive efficiency from a best-fitting four-factor
solution comprising tests from the executive function domain (attention, abstraction and working
memory). The tests contributing to the executive efficiency score include the Penn Continuous
Performance Test, the Letter N-Back task, and (weakly) the Penn Verbal Reasoning Test 35,36,37
(analogical reasoning). In the present study, we use this factor score for executive efficiency as our
primary measure of executive function, hereafter referred to as "executive function".
Imaging Data Acquisition
MRI data were acquired on a 3 Tesla Siemens Tim Trio whole-body scanner and 32-channel head
coil at the Hospital of the University of Pennsylvania. DTI scans were acquired via a twice-refocused
spin-echo (TRSE) single-shot echo-planar imaging (EPI) sequence (TR = 8100ms, TE = 82ms, FOV
= 240mm2/240mm2; Matrix = RL:128/AP:128/Slices:70, in-plane resolution (x and y) 1.875 mm2;
slice thickness = 2mm, gap = 0; flip angle = 90/180/180 degrees, volumes = 71, GRAPPA factor
= 3, bandwidth = 2170 Hz/pixel, PE direction = AP). This sequence utilizes a four-lobed diffusion
encoding gradient scheme combined with a 90-180-180 spin-echo sequence designed to minimize
eddy-current artifacts. The complete sequence consisted of 64 diffusion-weighted directions with
b = 1000s/mm2 and 7 interspersed scans where b = 0s/mm2. Total scan time was approximately
11 min. The imaging volume was prescribed in axial orientation covering the entire cerebrum with
the topmost slice just superior to the apex of the brain.
In addition to the DTI scan, a map of the main magnetic field (i.e., B0) was derived from a
double-echo, gradient-recalled echo (GRE) sequence, allowing us to estimate field distortions in
each dataset. Prior to DTI acquisition, a 5-minute magnetization-prepared, rapid acquisition
gradient-echo T1-weighted (MPRAGE) image (TR 1810 ms, TE 3.51 ms, FOV 180 ×240 mm,
matrix 256 × 192, effective voxel resolution of 1 × 1 × 1 mm) was acquired. This high-resolution
structural image was used for tissue segmentation and parcellating gray matter into anatomically
defined regions in native space. Rigorous manual and automated quality-assurance protocols for
the T1-weighted structural imaging data were performed and cleared for the 882 subjects consid-
ered here 38. Subsequently, all structural images were processed using FreeSurfer (version 5.3) 39.
FreeSurfer reconstructions underwent rigorous quality-assurance protocols 38,40. The T1 image was
parcellated into 234 regions by FreeSurfer according to the Lausanne Atlas 41. We define the sub-
cortex of this 234-region parcellation to be comprised of the left and right hemispheric counterparts
of the thalamus proper, caudate, putamen, pallidum, nucleus accumbens area, hippocampus and
amygdala, while excluding the brainstem (14 regions). We define the cortex of this 234-region
parcellation to be comprised of the remaining regions (219 regions).
fMRI BOLD data were acquired as subjects completed a version of the n-back task using fractal
images 42. See the supplementary methods of ref. 43 for details regarding task presentation and
structure. Functional images were obtained using a whole-brain, single-shot, multislice, gradient-
echo echoplanar sequence (231 volumes; TR = 3000 ms; TE = 32 ms; flip angle = 90 degrees; FOV
= 192× 192 mm; matrix = 64× 64; slices = 46; slice thickness = 3 mm; slice gap = 0 mm; effective
voxel resolution = 3.0 × 3.0 × 3.0 mm).
5
Imaging Data Preprocessing
All DTI datasets were subject to a rigorous manual quality assessment procedure involving visual
inspection of all 71 volumes 32. Each volume was evaluated for the presence of artifact, with the
total number of volumes impacted summed over the series. This scoring was based on previous work
describing the impact of removing image volumes when estimating the diffusion tensor 44,45. Data
was considered "Poor" if more than 14 (20%) volumes contained artifact, "Good" if it contained 1-14
volumes with artifact, and "Excellent" if no visible artifacts were detected in any volumes. All 882
subjects included in the present study had diffusion datasets identified as "Good" or "Excellent,"
and had less than 2mm mean relative displacement between interspersed b = 0 volumes. As
described below, even after this rigorous quality assurance, motion was included as a covariate in
all analyses.
The skull was removed for each subject by registering a binary mask of a standard fractional
anisotropy (FA) map (FMRIB58 FA) to each subject's DTI image using an affine transformation 46.
Eddy currents and subject motion were estimated and corrected using the FSL eddy tool 47. Diffu-
sion gradient vectors were then rotated to adjust for subject motion estimated by eddy. After the
field map was estimated, distortion correction was applied to DTI data using FSL's FUGUE 48.
BOLD time series were processed as described in 49,43. Briefly, FSL 5 48 was used to analyze time se-
ries data from three condition blocks (0-back, 1-back and 2-back), with the primary contrast being
2-back > 0-back. BOLD images were co-registered to the T1 image using boundary-based regis-
tration 50 with integrated distortion correction as implemented in FSL. Generalized linear model
(GLM) beta weights were averaged across all voxels in each parcel of the 234-node Lausanne atlas.
In our assessment of n-back performance-related activation, we use the difference in GLM beta
weights between the 2-back and 0-back condition. For all analyses of fMRI data, we excluded 223
subjects with incomplete data or excessive head motion (mean relative displacement > 0.5 mm or
maximum displacement > 6 mm), leaving n = 659 remaining.
Structural Network Estimation
Structural connectivity was estimated from DTI data in order to generate the adjacency matrix
representing the pattern of white matter tracts between large-scale brain areas. DSI Studio was
used to estimate the diffusion tensor and perform deterministic whole-brain fiber tracking with a
modified FACT algorithm that used exactly 1,000,000 streamlines per subject after removing all
streamline with length < 10 mm 28. To extend regions into white matter, parcels defined using the
Lausanne atlas were dilated by 4 mm 28,29 and registered to the first non-weighted (b = 0) volume
using an affine transform 28,29. The number of streamlines connecting each node of the 234 region
parcel end-to-end was used to define the edge weights aij of the adjacency matrix A.
Network Controllability
We represent the streamline-weighted structural network estimated from diffusion tractography as
the graph G = (V,E), where V and E are the vertex and edge sets, respectively. Let aij be the weight
associated with the edge (i, j) ∈ E, and define the weighted adjacency matrix of G as A = [aij],
where aij = 0 whenever (i, j) /∈ E. We associate a real value with each node to generate a vector
6
describing the network state, and we define the map x : N≥0 ← RN to describe the dynamics of
the network state over time.
It is worth noting that this method assumes that the number of streamlines is proportional to the
strength of structural connectivity with regards to propagation of activity between nodes according
to a specified model of dynamics. Here we employ a simplified noise-free linear discrete-time and
time-invariant model of such dynamics:
x(t + 1) = Ax(t) + BKuK(t),
(1)
where x describes the state (i.e. voltage, firing rate, BOLD signal) of brain regions over time.
Thus, the state vector x has length N , where N is the number of brain regions in the connectome
parcellation, and the value of xi describes the brain activity state of that region. The matrix A
is symmetric, with the diagonal elements satisfying Aii = 0. Prior to calculating controllability
values, we divide A by 1 + ξ0(A), where ξ0(A) is the largest eigenvalue of A. The input matrix
BK identifies the control point K in the brain, where K = k1, ..., km and
(2)
and ei denotes the i-th canonical vector of dimension N . The input uK : R≥0 → RM denotes the
control strategy.
BK = [ek1 ··· ekm],
To study the dynamics by which the activity of one brain region influences structurally connected
regions, we apply the control theoretic notion of controllability to our dynamical model. Classic
results in control theory ensure that controllability of the network, x(t + 1) = Ax(t) + BKuK(t),
from the set of network nodes K is equivalent to the controllability Gramian WK being invertible,
where
∞(cid:88)
τ =0
WK =
Aτ BKBTKAτ .
(3)
We calculate WK, or rather Wi, with BK set equal to one canonical vector ei and repeat this process
for all N nodes 29,28. Although it is well known that the activity of several brain regions and neuronal
ensembles are related via non-linear dynamics, it has been shown that a linear approximation can
explain features of the resting state fMRI BOLD signal 51; this suggests that a linear approximation
can effectively capture the controllability properties of the original non-linear dynamics.
Controllability Metrics
Following ref. 29,28, controllability metrics for structural brain networks were calculated for two
different control strategies which describe the ability to change x(t) in a particular fashion 20.
7
Average controllability describes the ease of transition to energetically similar states, while modal
controllability describes the ease of transition to difficult-to-reach states 20.
Average controllability of a network equals the average input energy applied to a set of control nodes
required to reach all possible target states. It is known that average input energy is proportional
cannot be accurately computed even at more coarse connectome parcellations. In addition to its
to Trace(W−1K ), the trace of the inverse of the controllability Gramian. Instead, as in ref. 29,28, we
use Trace(WK) as a measure of average controllability because (i), Trace(WK) and Trace(W−1K )
are related via inverse proportionality, and (ii), Trace(W−1K ) tends to be very ill-conditioned and
relationship with Trace(W−1K ), Trace(WK) describes the energy of the network impulse response,
or, equivalently, the network H2 norm 21,22. In summary, to compute the average controllability
value for node i in A, we compute the Trace(WK) when node i is the only control node (i.e.
BK = ei).
Modal controllability refers to the ability of a node to control the evolutionary modes of a dynamical
network, and is most interpretable when used to identify states that are poorly controllable given
BK. To calculate modal controllability, one must first obtain the eigenvector matrix V = [vij] of
the adjacency matrix A. If vij is small, then the j-th evolutionary mode of the input-independent
form of Eq. (1), x(t) = Ax(t), is poorly controllable from node i. According to prior work 20, we
ij as a scaled measure of the modal controllability of each of the N
define φi =(cid:80)N
modes ξ0(A), ..., ξN−1(A) from node i.
j=1(1 − ξ2
j (A))v2
Finally, we calculate the mean of the regional controllability values either across the whole brain
(Fig. A.1) or within the cortex and subcortex separately (Figs. 3, 4, 5, 6). For each subject, we
define the mean modal controllability to be the sum of the values of φi for each node within A,
divided by the number of regions. Similarly, we define the mean average controllability to be the
sum of the values of Trace(WK) for each node within A, divided by the number of regions.
Network Synchronizability
While the metrics of average and modal controllability provide complementary views on the diverse
dynamics that a brain network can display, it is also useful to study contrasting metrics that probe
the susceptibility of the network to be constrained within a narrower range of dynamics. Previous
work 29 has demonstrated that a useful contrasting metric to controllability is synchronizability,
which intuitively measures the susceptibility of a network to remain in a single synchronous state
s(t), i.e. x1 = ··· = xn(t + 1) = s(t). The master stability function (MSF) allows one to analyze
(cid:80)
the stability of this synchronous state without fully characterizing the properties of each feature of
the system 52. Within this framework, linear stability depends on the positive eigenvalues λi, i =
1, ..., N − 1 of the Laplacian matrix L defined by Lij = δij
k Aik − Aij, where δij is the Kronecker
delta function.
The condition for stability depends on the shape of the MSF, i.e. the stability of the synchronized
state to linear perturbations holds when the MSF is negative for all non-zero eigenvalues of L. This
condition is more likely for a smaller range of eigenvalues, hence we can use the normalized spread
of these eigenvalues to quantify network synchronizability as
8
(cid:80)N−1
d2(N − 1)
i=1 λi − ¯λ2
1
σ2 =
, where ¯λ :=
1
N − 1
N−1(cid:88)
i=1
λi
(4)
(cid:80)
(cid:80)
and d := 1
j(cid:54)=i Aij, the average coupling strength per node, which normalizes for the overall
N
network strength. Using this definition, we calculated the synchronizability of A for each subject.
i
Linear Regression of Network Control Metrics on Sex
For all analyses of network control metrics, we examined the effect of sex while controlling for
age, total brain volume (segmented brain volume, as defined by FreeSurfer BrainSegVol metric),
handedness, and motion during the diffusion scan. We used multiple ordinary least squares (OLS)
linear regression with the lm() command in R or the MATLAB regress function to fit the following
general equation:
C = 1 + βaa + βvv + βhh + βmdmd + βss ,
(5)
where C is the controllability statistic (either φi for modal controllability or A for average con-
trollability), a is age, v is total intracranial volume, md is the mean framewise displacement as a
summary measure of in-scanner head motion during the diffusion imaging sequence, h is handed-
ness, and s is sex. We then used the R package visreg to calculate 95% confidence intervals around
fitted lines and generate partial residuals. Furthermore, when using the multiple OLS regression
for the node-level analyses of controllability, we applied a false discovery rate (FDR) correction 53
(q < 0.05) to control for Type I error due to multiple testing.
Non-Linear Fits of Executive Function and Network Metrics
Because age and executive function are nonlinearly related, we used a general additive model 54,38
(GAM) to assess the relationship between executive efficiency and age using the mgcv package in
R (Fig. 1b). A GAM is a generalized linear model in which the linear predictor is defined by
unknown smooth functions of predictor variables. In this case, we use penalized regression splines
for age, with conventional parametric components for the remaining predictors.
Following ref. 29, we used non-linear models to examine the relationships between synchronizability,
average controllability, and modal controllability (Fig. 5). We generated estimates of parameters
for models of the form y = a+b exp(cx) via non-linear least squares using the nls() function in R. To
compare these fits between males and females, we performed an analysis of variance (ANOVA). For
example, when considering sex differences in the nonlinear relation between modal controllability
and average controllability, we examined the expression:
φi = (a + αs) + (b + βs) exp((c + γs)A) ,
(6)
where φi is modal controllability, A is average controllability, and s is sex, coded as 0 for males
and 1 for females, so that α, β and γ = 0 for males and the equation is reduced to the base form.
9
The predicted values for males or females obtained from the full model were used to generate the
curves in Fig. 5.
Mediation Testing
After identifying associations between sex, controllability, and executive function, we asked whether
controllability is a statistically significant mediator of the relationship between sex and executive
function. To test for a formal mediation, we performed non-parametric bootstrapping causal me-
diation analysis using the R function mediate() from the package mediation 55. It is worth noting
that we cannot estimate causality in cross-sectional association data. The Average Causal Me-
diation Effect (ACME) quantifies the relationship between mediator and outcome independent of
treatment, while the Average Direct Effect (ADE) quantifies the relationship between treatment
and outcome, independent of the mediator 55.
Linear Regression of BOLD Data on Network Control Metrics
Finally, and in keeping with our linear systems framework, we sought to test whether controllability
would predict brain state (x(t)), as defined by BOLD activation during a task demanding executive
function.
In these analyses (Fig. 8), we used the residuals from Eq. 5 for controllability. For
activation, we used the residuals from the following equation:
A = 1 + βaa + βhh + βmnmn + βss ,
(7)
where A is 2-back minus 0-back activation (hereafter referred to as "activation"), a is age, h is
handedness, mn is the mean framewise displacement during the n-back scan, and s is sex. Because
we had no prior knowledge about where controllability might predict activation, we fit linear models
between controllability and activation at every possible pair of nodes, i.e. we performed 234 × 234
regressions of controllability residuals at each node with activation residuals at each node. The
234× 234 matrices of p-values for the slope of controllability on activation were FDR corrected (q <
0.05) separately for average controllability and for modal controllability. To identify regions where
activation was associated with executive function, we followed previous work 49,43 by examining
d(cid:48), a composite measure of n-back task performance which takes both correct responses and false
positives into account to separate performance from response bias.
Results
Sex differences in executive function
Our analysis of sex differences in the development of executive function and its neurobiological
underpinnings began with a sex-stratified comparison of mean performance on executive efficiency,
a cross-contextual summary measurement of executive function 35. Consistent with prior results
from the PNC data 18,30, the executive efficiency subscore of the PCNB was significantly higher in
females (full model r2 = 0.52, df = 873, p = 0.0015; Fig. 2a) with no interaction between age and
sex. Furthermore, executive efficiency increased with age in a non-linear fashion (r2 = 0.52, df = 2,
p < 10−15; Fig. 2a). This result suggests that development of executive function occurs via a similar
10
course for males and females, but that females display higher scores in this cognitive domain for
all ages studied. After performing this sex-stratified analysis of the developmental course of the
executive efficiency score, we next turned to a consideration of its neurobiological underpinnings
(Fig. 1a).
Sex differences in regional network controllability
Our general hypothesis was that sex differences in executive function in youth stem from sex
differences in the controllability of structural brain networks as they rewire over development, a
notion that bridges the control of behavior (executive function) with the control of brain dynamics
(network controllability). To test this hypothesis, we examined the anatomical distribution of
control points in the structural brain networks of males and females separately. We ranked the
mean average controllability value at each region for males and females separately, which revealed
that the distribution of controller strength is virtually identical between male and females (r = 0.99;
Fig. 2b). This result suggests that there is no sex difference in the spatial distribution of controllers
when classified by their relative magnitudes.
We next investigated whether the actual (rather than ranked) regional controllability estimates dif-
fered by sex, and we addressed this question first by considering the cortex and subcortex separately.
Specifically, we computed mean controllability values for each subject across 219 cortical regions
In the cortex, mean average controllability (r2 = 0.082, df = 876,
and 14 subcortical regions.
p = 0.018; Fig. 3a) and mean modal controllability (r2 = 0.13, df = 876, p = 0.018; Fig. 3a) are
higher in females. Mean modal controllability in the subcortex is also higher in females (r2 = 0.047,
df = 876, p < 10−4; Fig. 3b), whereas mean average controllability in the subcortex is higher in
males (r2 = 0.036, df = 876, p = 0.034; Fig. 3b). Only with average controllability in the subcortex
did males have higher controllability than females; this result suggests that the connectivity profile
of subcortical regions may contribute to sex differences in functional brain dynamics.
In a finer-grained analysis, we investigated whether the controllability of single regions differed by
sex. Modal and average controllability were separately examined at each region, while accounting
for age, total brain volume, handedness, and mean in-scanner head motion as model covariates.
We found that average and modal controllability differed by sex at a subset of network nodes after
FDR correction (q < 0.05) for multiple comparisons. Nodes with average controllability values that
differed by sex were almost all (4/5) higher in males and located in the frontal lobe or subcortex
(Fig. 3c-d). Conversely, nodes with modal controllability values that differed by sex were all
(18/18) higher in females and located in frontoparietal and subcortical systems. These results
suggest that controller strength differs between males and females on a region-specific and control
strategy-specific basis.
Development of network controllability across the sexes
We next turned to assessing whether the developmental arc of network controllability differed by
sex. We found that controllability in the cortex and subcortex tended to increase with age, with
In the cortex (r2 = 0.13, df = 876,
the exception of average controllability in the subcortex.
p = 9.9 × 10−10; Fig. 4b) and subcortex (r2 = 0.047, df = 876, p = 1.0 × 10−5; Fig. 4d), mean
modal controllability increases with age for males and females. Mean cortical average controllability
also increases with age (r2 = 0.082, df = 876, p = 1.9 × 10−15; Fig. 4b). Interestingly, there was a
11
significant age-by-sex interaction only with subcortical average controllability, such that the slope
was positive for males and negative for females (Fig. 4c). However, we also found that subcortical
average controllability remains stable throughout development for both males (βage = 0.47, t = 1.29,
p = 0.20, df = 875) and females ((βage + βage∗sex) = −0.65, p = 0.052, t = −1.94, df = 875) 56.
Taken together, these results suggest that controllability changes with age similarly for males and
females, but that average controllability in the subcortex changes with age differently than modal
controllability or than cortical average controllability.
With the knowledge that controllability has a sex-independent relationship with age, we were
interested in testing the hypothesis that sex influences the relationship between different types of
controllability in the developing brain. Following ref. 29, we fit the relationships between average
and modal controllability with the exponential function y = a + b exp(cx) separately for each sex
(Eq. 6). Our results showed that males and females did not have a statistically different relationship
between average and modal controllability averaged across the whole brain (p = 0.16, df = 3; Fig.
5a,e). In the cortex alone, average and modal controllability followed an increasing exponential
form (Fig. 5b,f), similar to that of the whole brain. In contrast, in the subcortex alone average and
modal controllability followed a decreasing exponential form (Fig. 5d,h). When sex was included
in the model, fits improved significantly (cortex: p = 2.1 × 10−8, df = 3, Fig. 5b,f; subcortex:
p = 6.3 × 10−9, df = 3, Fig. 5c,g), suggesting that these regions may contain sex-dependent
differences in structural connections important for controlling brain network state.
Next, we performed a specificity analysis to determine whether our results could be further con-
firmed by sex differences in a contrasting metric – synchronizability – that probed the susceptibility
of the network to be constrained within a narrow (rather than broad) range of dynamics. We found
that synchronizability did not differ between males and females (A.1b) and decreased with age in a
sex-independent fashion (A.1d). Moreover, the exponential relationship between synchronizability
and average controllability did not differ by sex (p = 0.33, df = 3; Fig. 5d,h), confirming the local
(Fig. 3c-d) rather than global (Fig 2b) differences in metrics of network control and dynamics.
Sex differences in network controllability predict individual differences in exec-
utive function
While sex differences in network controllability are of interest in understanding the structural
drivers of brain dynamics, their impact on behavior requires a link to cognition. Here we examine
the relation between executive function and network controllability across cortex and subcortex
separately, and then at individual brain regions. When considering mean cortical and subcortical
controllability and executive function we found that neither average (r2 = 0.46, df = 873, p = 0.96;
Fig. 6a) nor modal controllability (r2 = 0.46, df = 873, p = 0.25; Fig. 6b) in the cortex were asso-
ciated with performance. A different trend was apparent in the subcortex: average controllability
was negatively correlated with performance (r2 = 0.47, df = 873, p = 5.4 × 10−4; Fig. 6c), while
modal controllability was positively correlated with performance (r2 = 0.47, df = 873, p = 0.024;
Fig. 6d).
Next, we considered the 21 brain areas that we had previously found to display sex differences in
controllability values (5 nodes for average controllability and 18 nodes for modal controllability,
with 2 nodes overlapping). Nodes with significantly higher average controllability in males (right
medial orbitofrontal cortex and bilateral caudate nuclei) had average controllability values that
12
were negatively related to executive function (Fig. 6e; FDR corrected, q < 0.05). Interestingly,
modal controllability at any individual node was not significantly associated with executive function
after FDR correction; however, two nodes with uncorrected p < 0.05 (right caudate nucleus and left
inferior parietal lobe) had modal controllability values that were higher in females and positively
associated with executive function. This result suggests that modal controllability of one particular
node is not as strong of a predictor of cognitive performance as mean modal controllability over
the subcortex as a whole.
One parsimonious explanation for the results reported up to this point is that controllability medi-
ates the relationship between sex and executive function: specifically, subcortical average controlla-
bility is higher in males, and increasing subcortical average controllability is associated with poorer
performance on an assessment of executive function. We explicitly tested for such a mediation and
found that, indeed, average controllability at the right medial orbitofrontal cortex, right superior
frontal cortex, right postcentral gyrus and the bilateral caudate nuclei were statistically significant
mediators of the relationship between sex and executive function (Fig. 7a). Similarly, subcortical
modal controllability was higher in females, and increasing subcortical modal controllability was
associated with better executive function. However, consistent with the fact that no individual
node had modal controllability values associated with executive function, modal controllability was
not found to be a mediator of the relationship between sex and executive function. Taken together,
these results suggest that sex differences in average controllability may be a predictive biomarker
for sex differences in executive function.
Sex-dependent relationships between n-back task activation magnitudes and con-
trollability
Describing the relationship between sex, regional controllability, and cognitive performance provides
us with a better understanding of the importance of structural brain networks in sex differences in
executive function. The final aspect of our hypothesis pertains to whether network control theory
can be used to explain how differences in brain network structure produce divergent patterns of
brain activity that underpin executive function. We hypothesized that regional controllability
values would predict regional n-back task activation magnitudes, and that associations between
controllability and activation would differ by sex.
To obtain a reference point for activation profiles associated with strong executive function, we
separately regressed activation at each node on d(cid:48) 49. This analysis identified 113 nodes for which
activation was associated with successful task performance (Fig. 8a). Note that in contrast to the
use of a general factor score for executive efficiency in the previous analyses, here we considered the
d(cid:48) measure to summarize performance on the n-back task only 57 so that the performance measure
is directly related to the context in which activation values are measured. Interestingly, sex was
not associated with activation at any brain region (FDR corrected, q < 0.05, data not shown).
After identifying regions at which activation was associated with executive function, we sought to
relate these activation values to controllability metrics of structural brain networks. We found a
complex pattern of both positive and negative relationships between average controllability at 8
nodes with activation at 11 nodes (Fig. 8b). Average controllability at the left caudal middle frontal
lobe was associated with activation at 7 different regions, which cluster together in a symmetric
fashion in superior parietal and lateral occipital cortex (Fig. 8c). Modal controllability was not
13
significantly associated with activation.
Next, we tested our hypothesis that the relationship between regional controllability and activation
depends on sex. Specifically, we performed the same analysis on males and females, separately.
Consistent with the pooled analysis, the profile of controllability-activation associations overlapped
significantly between the sexes (Fig. 8d, Fig. A.4a). While there was no overlap in the precise
regions involved in these sex-split analyses, there were gross anatomical similarities between males
and females. In both sexes, controllability at regions near the temporal poles was associated with
activation at medial frontal regions. We observed no significant interaction between controllability
and sex in predicting activation (Fig. A.4b), although the directionality of the interaction terms
was consistent with the observed sex differences (Fig. A.4a-b). These results support the notion
that both similarities and differences in structure-function relationships exist across the sexes.
Weighted degree does not explain controllability trends
Weighted degree has previously been shown to correlate well with controllability metrics 28. We
performed several specificity analyses to determine whether our results could be trivially explained
by sex-differences in the weighted degree of nodes in the network. We identified 13 nodes where
weighted degree was associated with sex (Fig. A.3a), and found that only 5/21 unique nodes with
sex-associated controllability values also had a significant association between weighted degree and
sex. Among those 5 unique nodes, only average controllability at the left caudate nucleus and
the right post central gyrus had the same directionality of association with sex as the association
with weighted degree (i.e. higher in males and females, respectively). When weighted degree was
averaged across the anatomical cortex or subcortex, there were no significant associations between
sex and cortical (r2 = 0.090, df = 876, p = 0.73; Fig. A.2a) or subcortical (r2 = 0.077, df = 876,
p = 0.13; Fig. A.2c) weighted degree. These findings suggest that despite the known associations
between controllability and weighted degree 28, weighted degree does not explain the observed sex
differences in controllability.
Next, we tested whether developmental trends of weighted degree paralleled those of controllability
metrics. We found that when averaged across the cortex or subcortex, weighted degree is positively
associated with average controllability and negatively associated with modal controllability. How-
ever, mean weighted degree increased with age in the cortex (r2 = 0.090, df = 876, p = 0.0034;
Fig. A.2b), but decreased with age in the subcortex (r2 = 0.077, df = 876, p = 0.011; Fig.
A.2d). Neither of these two trends (Fig. A.2b,d) were simple mirrors of the relationships between
age and regional average or modal controllability (Fig. 4), suggesting that developmental trends
of controllability reflect a more complex phenomenon than developmental trends of weighted de-
gree. Furthermore, at nodes with sex-associated weighted degree or sex-associated controllability,
weighted degree did not significantly mediate the relationship between sex and executive function
(Fig. A.6). These findings suggest that weighted degree does not explain developmental trends of
controllability, nor does it mediate the relationship between sex and executive function.
Finally, we assessed whether the associations between controllability and activation could be simply
explained by the weighted degree of nodes in the network. We found no significant associations
between activation and weighted degree in a pooled analysis. Moreover, including each subject's
global mean weighted degree in the model did not change the associations between average con-
trollability and activation (data not shown). While there were significant associations between
14
activation and weighted degree for males alone (Fig. A.5), weighted degree-activation profiles did
not overlap well with average controllability-activation profiles. These results suggest that weighted
degree does not explain the relation between controllability and activation.
Discussion
Our study demonstrates that network control theory can be used to explain how differences in
brain network structure produce divergent patterns of brain activity that underpin executive func-
tion and its differences across males and females. We first showed that the relative locations of
controllers by strength did not differ between males and females, suggesting that the overall struc-
ture of control points is similar across sex. Then, we showed that controllability covaried with age
in a sex-independent fashion in both the cortex and subcortex. While global and developmental sex
differences appear minimal, local sex differences in controllability exist and average controllability
at those regions predicts executive function. Notably, average controllability significantly mediated
the association between sex and executive function, such that subcortical average controllability
was higher in males and is negatively associated with executive function. Crucially, consistent with
predictions from network control theory, we also observed sex-dependent associations between nodal
average controllability and n-back task BOLD activation magnitudes, demonstrating that differ-
ences in brain network structure produce divergent patterns of brain activity supporting executive
function.
Implications for cognitive and clinical neuroscience
Executive function is generally lower in males, which is reflected in higher rates of impulsivity 5,
ADHD diagnoses 6, criminality 7, and substance use 2. While it is known that sex differences exist in
the prevalence of disorders of executive function and the putative neural circuitry involved 58,59, it is
unclear whether the pathophysiology of such disorders is sex-specific or sex-independent. Our study
significantly extends the boundaries of knowledge in demonstrating neurophysiological markers of
sex differences in executive function, and in couching such markers within a general network control
theory of brain function.
The present work also provides important groundwork for clinical therapy. The delivery of psychi-
atric care is shifting towards personalized, targeted interventions. This shift can be supported by
the methodology and computational tools of network neuroscience, where accurate models of brain
structure and dynamics can be constructed for single individuals 60,61. Such subject-, age-, and
sex-specific models of brain structure and function can directly inform neuromodulatory therapies
such as transcranial magnetic stimulation, by offering predictions about how stimulation will affect
both the area being stimulated, and other areas connected to it, thereby producing a complex
spatiotemporal influence on brain state. Specifically, assuming a model of brain dynamics allows us
to determine which nodes could most easily drive transitions in the state of brain activity via some
stimulatory input. The network control theory that we use here to study the internal modulation
of brain state (via undertaking a task requiring executive function) also makes explicit predictions
about the external modulation of brain state (via stimulation or neurofeedback) 62. Indeed, aver-
age controllability values were associated with sex-dependent, symmetric patterns of brain activity
(Fig. 8c-d). Our results suggest that average controllability values contain useful information about
15
brain state, beyond what can be predicted from a simple streamline-weighted adjacency matrix.
These observations are the first steps towards a characterization of brain dynamics that would
allow clinicians to predict the impact of stimulation given a subject-, age-, and sex-specific network
architecture, thereby producing predictable changes in brain state.
An understanding of the relation between network controllability, sex, and executive function could
provide important context for the study and diagnosis of neurological disease and psychiatric dis-
orders whose prevalence may differ by sex and whose presentation includes alterations in executive
function.
In our study, the most robust associations between controllability, sex, and executive
function are found in the subcortex and the prefrontal cortex. It is widely known that the pre-
frontal cortex is important for behavioral planning and working memory 63,64, two key components
of executive function. The subcortex alone is less commonly associated with executive function,
but there is a strong precedent for the requirement of circuitry between dorsolateral prefrontal
cortex (DLPFC) and subcortical structures for executive function 65,66,67,68. Vascular 69 or neu-
rodegenerative 68 (i.e. Huntington's disease, Parkinson's disease) lesions to the subcortex give rise
to executive dysfunction, similar to what is observed with DLPFC lesions. Moreover, ADHD, a
disorder characterized by impulsivity and executive dysfunction, is associated with decreased stri-
atal dopamine transporter expression 58, aberrant striatal fMRI BOLD activation 70, and decreased
striatal volume 71,72. Sex differences in striatal volume 73 and striatal dopamine receptor binding 59
have also been reported. In light of these results, our study provides an important account of sex
differences in subcortical connectivity from the dynamical perspective of network control theory.
Our results suggest that the connectivity profile of subcortical structures is related to sex and is
important for executive function, supporting the notion that disorders of executive function arise
in part from altered structural connectivity.
Methodological Considerations
Several methodological considerations are pertinent to this work. First, we use a time-invariant,
linear model of brain dynamics because its network control properties have been well characterized
mathematically; however, it is known that the brain is highly non-linear 74. Yet, the associations
between controllability and brain activity that we uncover here suggest that a simple linear model
is sufficient to capture some aspect of the underlying brain dynamics. Second, axons transmit
information in a unidirectional fashion, but diffusion imaging and associated tractography tools
cannot elucidate the directionality of large axonal fiber bundles. Thus, we construct a symmetric
adjacency matrix A, assuming bidirectional influence between network nodes, and all interpreta-
tions of regional controllability depend on the realism of that assumption. Third, the eigenvalues of
a symmetric matrix can only ever be real and thus the system will not oscillate, as neural systems
are known to do. Finally, streamline counts and network density vary between scans 75 and with
different tract reconstruction methods 76, which can complicate interpretation. Nevertheless, nu-
merous DTI-based studies of neuropsychiatric illness have described differences in structural brain
networks consistent with clinical and neuroscientific priors 77,78, suggesting that DTI is capable of
capturing subtle but meaningful variation in structural connectivity.
It is also important to acknowledge that our analysis of sex differences focuses on sex assigned at
birth, which we refer to as "sex" throughout this paper. Our data does not include endocrinological
measurements relevant to sex, nor does it include any psychosocial assessment of gender identity.
16
Thus, we were not able to control for or assess gender-based differences in brain structure. Fur-
thermore, while there exist two distinct classes of human genitalia, this fact does not imply that
brains are also sexually dimorphic 79. Both "male" and "female" features exist in both male and
female brains, though some features are more common in one sex than the other 79,80. As a result,
there may be more meaningful variance in neurologic phenotypes within each sex than between
sexes. The divergent controllability-activation profiles in Fig. 8d may be in part due to this fact.
Nevertheless, biological sex is an easily measured variable and identifying correlates of sex makes it
a useful biomarker. In the future, identification of additional covariates might help uncover a more
ubiquitous reason for sex-associated brain features 80.
Conclusion and Future Directions
First and foremost, our analysis of sex differences in structural brain networks showed that males
and females are highly similar from the perspective of network controllability. The organization
of relative controller strength was almost identical between males and females, and sex was not
significantly associated with controllability values at most brain regions. However, the differences in
average controllability between males and females predicted differences in cognitive performance and
effects were most robust in subcortical and frontal regions. Given that BOLD signal is associated
with network controllability, an interesting future study might use time-invariant, linear dynamics
to predict changes in brain activity after stimulation of regions with high versus low controllability.
Our results pave the way for a future study to use sex and age as features that may delineate the
likely responses to brain stimulation a priori, obviating the need for pre-stimulation diffusion image
acquisition and processing. Such an approach would allow the delivery of personalized psychiatric
care at low cost with minimal requirements for imaging equipment and computational power.
Acknowledgments
This work was supported by an administrative supplement to NIH R21-M MH-106799 (Satterth-
waite/Bassett MPI). DSB also acknowledges support from the John D. and Catherine T. MacArthur
Foundation, the Alfred P. Sloan Foundation, the Army Research Office through contract number
W911NF-14-1-0679, the Army Research Laboratory through contract number W911NF-10-2-0022,
the National Institute of Health (2-R01-DC-009209-11, 1R01HD086888-01, R01-MH107235, R01-
MH107703, R01MH109520, 1R01NS099348 and R21-M MH-106799), the Office of Naval Research,
and the National Science Foundation (BCS-1441502, CAREER PHY-1554488, BCS-1631550, and
CNS-1626008). TDS was supported by R01MH107703. DRR was supported by K01MH102609.
FP acknowledges support from NSF-BCS-1631112 and NSF-BCS-1430279. The content is solely
the responsibility of the authors and does not necessarily represent the official views of any of the
funding agencies.
17
References
[1] Vicki A Anderson, Peter Anderson, Elisabeth Northam, Rani Jacobs, and Cathy Catroppa.
Development of executive functions through late childhood and adolescence in an australian
sample. Developmental neuropsychology, 20(1):385–406, 2001.
[2] Daniel Romer, Laura Betancourt, Joan M. Giannetta, Nancy L. Brodsky, Martha Farah,
and Hallam Hurt. Executive cognitive functions and impulsivity as correlates of risk taking
and problem behavior in preadolescents. Neuropsychologia, 47(13):2916 – 2926, 2009. ISSN
0028-3932. doi: http://dx.doi.org/10.1016/j.neuropsychologia.2009.06.019. URL http://www.
sciencedirect.com/science/article/pii/S002839320900270X.
[3] Russell A Barkley, Kevin R Murphy, George J Dupaul, and Tracie Bush. Driving in young
adults with attention deficit hyperactivity disorder: knowledge, performance, adverse out-
comes, and the role of executive functioning. Journal of the International Neuropsychological
Society, 8(5):655–672, 2002.
[4] J. Biederman, C. R. Petty, R. Fried, A. E. Doyle, T. Spencer, L. J. Seidman, L. Gross,
K. Poetzl, and S. V. Faraone. Stability of executive function deficits into young adult years:
a prospective longitudinal follow-up study of grown up males with adhd. Acta Psychiatrica
Scandinavica, 116(2):129–136, 2007. ISSN 1600-0447. doi: 10.1111/j.1600-0447.2007.01008.x.
URL http://dx.doi.org/10.1111/j.1600-0447.2007.01008.x.
[5] Constance L. Chapple and Katherine A. Johnson. Gender differences in impulsivity. Youth
Violence and Juvenile Justice, 5(3):221–234, 2007. doi: 10.1177/1541204007301286. URL
http://dx.doi.org/10.1177/1541204007301286.
[6] Erik G Willcutt. The prevalence of dsm-iv attention-deficit/hyperactivity disorder: a meta-
analytic review. Neurotherapeutics, 9(3):490–499, 2012.
[7] Catharine P Cross, Lee T Copping, and Anne Campbell. Sex differences in impulsivity: a
meta-analysis. Psychological bulletin, 137(1):97, 2011.
[8] Sheik Hosenbocus and Raj Chahal. A review of executive function deficits and pharmacological
management in children and adolescents. Journal of the Canadian Academy of Child and
Adolescent Psychiatry, 21(3):223, 2012.
[9] N Gogtay, J N Giedd, L Lusk, K M Hayashi, D Greenstein, A C Vaituzis, T F 3rd Nugent,
D H Herman, L S Clasen, A W Toga, J L Rapoport, and P M Thompson. Dynamic mapping
of human cortical development during childhood through early adulthood. Proc Natl Acad Sci
U S A, 101(21):8174–8179, 2004.
[10] G L Baum, R Ciric, D R Roalf, R F Betzel, T M Moore, R T Shinohara, A E Kahn, S N
Vandekar, P E Rupert, M Quarmley, P A Cook, M A Elliott, K Ruparel, R E Gur, R C
Gur, D S Bassett, and T D Satterthwaite. Modular segregation of structural brain networks
supports the development of executive function in youth. Curr Biol, 27(11):1561–1572.e8,
2017.
[11] V J Schmithorst, J Vannest, G Lee, L Hernandez-Garcia, E Plante, A Rajagopal, S K Holland,
18
and CMIND Authorship Consortium. Evidence that neurovascular coupling underlying the
BOLD effect increases with age during childhood. Hum Brain Mapp, 36(1):1–15, 2015.
[12] J S Nomi, T S Bolt, C E C Ezie, L Q Uddin, and A S Heller. Moment-to-moment BOLD
signal variability reflects regional changes in neural flexibility across the lifespan. J Neurosci,
37(22):5539–5548, 2017.
[13] E H Keulers, P Stiers, and J Jolles. Developmental changes between ages 13 and 21 years in
the extent and magnitude of the BOLD response during decision making. Neuroimage, 54(2):
1442–1454, 2011.
[14] D A Fair, A L Cohen, J D Power, N U Dosenbach, J A Church, F M Miezin, B L Schlaggar, and
S E Petersen. Functional brain networks develop from a "local to distributed" organization.
PLoS Comput Biol, 5(5):e1000381, 2009.
[15] Efstathios D. Gennatas, Brian B. Avants, Daniel H. Wolf, Theodore D. Satterthwaite, Kosha
Ruparel, Rastko Ciric, Hakon Hakonarson, Raquel E. Gur, and Ruben C. Gur. Age-related
effects and sex differences in gray matter density, volume, mass, and cortical thickness from
childhood to young adulthood. Journal of Neuroscience, 2017.
ISSN 0270-6474. doi: 10.
1523/JNEUROSCI.3550-16.2017. URL http://www.jneurosci.org/content/early/2017/
04/21/JNEUROSCI.3550-16.2017.
[16] Sarah-Jayne Blakemore and Suparna Choudhury. Development of the adolescent brain: im-
plications for executive function and social cognition. Journal of Child Psychology and Psy-
chiatry, 47(3-4):296–312, 2006. ISSN 1469-7610. doi: 10.1111/j.1469-7610.2006.01611.x. URL
http://dx.doi.org/10.1111/j.1469-7610.2006.01611.x.
[17] Madhura Ingalhalikar, Alex Smith, Drew Parker, Theodore D. Satterthwaite, Mark A. Elliott,
Kosha Ruparel, Hakon Hakonarson, Raquel E. Gur, Ruben C. Gur, and Ragini Verma. Sex
differences in the structural connectome of the human brain. Proceedings of the National
Academy of Sciences, 111(2):823–828, 2014. doi: 10.1073/pnas.1316909110. URL http://
www.pnas.org/content/111/2/823.abstract.
[18] Theodore D Satterthwaite, Daniel H Wolf, David R Roalf, Kosha Ruparel, Guray Erus, Simon
Vandekar, Efstathios D Gennatas, Mark A Elliott, Alex Smith, Hakon Hakonarson, et al.
Linked sex differences in cognition and functional connectivity in youth. Cerebral cortex, 25
(9):2383–2394, 2014.
[19] Y Y Liu, J J Slotine, and A L Barabasi. Controllability of complex networks. Nature, 473
(7346):167–173, 2011.
[20] Fabio Pasqualetti, Sandro Zampieri, and Francesco Bullo. Controllability metrics, limitations
and algorithms for complex networks. IEEE Transactions on Control of Network Systems, 1
(1):40–52, 2014.
[21] T Kailath. Linear Systems. Prentice Hall, Englewood Cliffs, 1980.
[22] R. E. Kalman, Y. C. Ho, and K. S. Narendra. Controllability of linear dynamical systems.
Contributions to Differential Equations, 1:189–213, 1963.
[23] S F Muldoon, F Pasqualetti, S Gu, M Cieslak, S T Grafton, J M Vettel, and D S Bassett.
19
Stimulation-based control of dynamic brain networks. PLoS Comput Biol, 12(9):e1005076,
2016.
[24] R F Betzel, S Gu, J D Medaglia, F Pasqualetti, and D S Bassett. Optimally controlling the
human connectome: the role of network topology. Sci Rep, 6:30770, 2016.
[25] S Gu, R F Betzel, M G Mattar, M Cieslak, P R Delio, S T Grafton, F Pasqualetti, and D S
Bassett. Optimal trajectories of brain state transitions. Neuroimage, 148:305–317, 2017.
[26] L Wiles, S Gu, F Pasqualetti, D S Bassett, and D F Meaney. Autaptic connections shift
network excitability and bursting. Scientific Reports, In Press, 2017.
[27] J Kim, J M Soffer, A E Kahn, J M Vettel, F Pasqualetti, and D S Bassett. Role of graph
architecture in controlling dynamical networks with applications to neural systems. Nature
Physics, Epub Ahead of Print, 2017.
[28] Shi Gu, Fabio Pasqualetti, Matthew Cieslak, Qawi K Telesford, B Yu Alfred, Ari E Kahn,
John D Medaglia, Jean M Vettel, Michael B Miller, Scott T Grafton, et al. Controllability of
structural brain networks. Nature communications, 6, 2015.
[29] E Tang, C Giusti, G L Baum, S Gu, E Pollock, A E Kahn, D R Roalf, T M Moore, K Ruparel,
R C Gur, R E Gur, T D Satterthwaite, and D S Bassett. Developmental increases in white
matter network controllability support a growing diversity of brain dynamics. Nat Commun,
8(1):1252, 2017.
[30] T D Satterthwaite, M A Elliott, K Ruparel, J Loughead, K Prabhakaran, M E Calkins,
R Hopson, C Jackson, J Keefe, M Riley, F D Mentch, P Sleiman, R Verma, C Davatzikos,
H Hakonarson, R C Gur, and R E Gur. Neuroimaging of the philadelphia neurodevelopmental
cohort. Neuroimage, 86:544–553, 2014.
[31] L Cammoun, X Gigandet, D Meskaldji, J P Thiran, O Sporns, K Q Do, P Maeder, R Meuli,
and P Hagmann. Mapping the human connectome at multiple scales with diffusion spectrum
MRI. J Neurosci Methods, 203(2):386–397, 2012.
[32] David R Roalf, Megan Quarmley, Mark A Elliott, Theodore D Satterthwaite, Simon N Van-
dekar, Kosha Ruparel, Efstathios D Gennatas, Monica E Calkins, Tyler M Moore, Ryan Hop-
son, et al. The impact of quality assurance assessment on diffusion tensor imaging outcomes
in a large-scale population-based cohort. Neuroimage, 125:903–919, 2016.
[33] R C Gur, J Richard, P Hughett, M E Calkins, L Macy, W B Bilker, C Brensinger, and
R E Gur. A cognitive neuroscience-based computerized battery for efficient measurement of
individual differences: standardization and initial construct validation. J Neurosci Methods,
187(2):254–262, 2010.
[34] Ruben C Gur, Jan Richard, Monica E Calkins, Rosetta Chiavacci, John A Hansen, Warren B
Bilker, James Loughead, John J Connolly, Haijun Qiu, Frank D Mentch, et al. Age group and
sex differences in performance on a computerized neurocognitive battery in children age 8- 21.
Neuropsychology, 26(2):251, 2012.
[35] Tyler M Moore, Steven P Reise, Raquel E Gur, Hakon Hakonarson, and Ruben C Gur. Psy-
20
chometric properties of the penn computerized neurocognitive battery. Neuropsychology, 29
(2):235, 2015.
[36] Tyler M Moore, Steven P Reise, David R Roalf, Theodore D Satterthwaite, Christos Da-
vatzikos, Warren B Bilker, Allison M Port, Chad T Jackson, Kosha Ruparel, Adam P Savitt,
et al. Development of an itemwise efficiency scoring method: Concurrent, convergent, dis-
criminant, and neuroimaging-based predictive validity assessed in a large community sample.
Psychol Assess., 28(12):1529–1542, 2016.
[37] Tyler M Moore, Ruben C Gur, Michael L Thomas, Gregory G Brown, Matthew K Nock,
Adam P Savitt, John G Keilp, Steven Heeringa, Robert J Ursano, Murray B Stein, et al.
Development, administration, and structural validity of a brief, computerized neurocognitive
battery: Results from the army study to assess risk and resilience in service members. Assess-
ment, Epub Ahead of Print, 2017.
[38] Simon N Vandekar, Russell T Shinohara, Armin Raznahan, David R Roalf, Michelle Ross,
Nicholas DeLeo, Kosha Ruparel, Ragini Verma, Daniel H Wolf, Ruben C Gur, et al. Topo-
logically dissociable patterns of development of the human cerebral cortex. Journal of Neuro-
science, 35(2):599–609, 2015.
[39] B Fischl. Freesurfer. Neuroimage, 62:774–781, 2012.
[40] Adon F.G. Rosen, David R. Roalf, Kosha Ruparel, Jason Blake, Kevin Seelaus, Lakshmi P.
Villa, Rastko Ciric, Philip A. Cook, Christos Davatzikos, Mark A. Elliott, Angel Garcia
de La Garza, Efstathios D. Gennatas, Megan Quarmley, J. Eric Schmitt, Russell T. Shi-
nohara, M. Dylan Tisdall, R. Cameron Craddock, Raquel E. Gur, Ruben C. Gur, and
Theodore D. Satterthwaite. Data-driven assessment of structural image quality. NeuroIm-
age, pages –, 2017. ISSN 1053-8119. doi: https://doi.org/10.1016/j.neuroimage.2017.12.059.
URL https://www.sciencedirect.com/science/article/pii/S1053811917310832.
[41] Meritxell Bach Cuadra, Claudio Pollo, Anton Bardera, Olivier Cuisenaire, J-G Villemure, and
J-P Thiran. Atlas-based segmentation of pathological mr brain images using a model of lesion
growth. IEEE transactions on medical imaging, 23(10):1301–1314, 2004.
[42] J Daniel Ragland, Bruce I Turetsky, Ruben C Gur, Faith Gunning-Dixon, Travis Turner, Lee
Schroeder, Robin Chan, and Raquel E Gur. Working memory for complex figures: an fmri
comparison of letter and fractal n-back tasks. Neuropsychology, 16(3):370, 2002.
[43] Sheila Shanmugan, Daniel H Wolf, Monica E Calkins, Tyler M Moore, Kosha Ruparel, Ryan D
Hopson, Simon N Vandekar, David R Roalf, Mark A Elliott, Chad Jackson, et al. Common
and dissociable mechanisms of executive system dysfunction across psychiatric disorders in
youth. American journal of psychiatry, 173(5):517–526, 2016.
[44] D K Jones and P J Basser. "squashing peanuts and smashing pumpkins": how noise distorts
diffusion-weighted MR data. Magn. Reson. Med., 52:979–993, 2004.
[45] Y Chen, O Tymofiyeva, C P Hess, and D Xu. Effects of rejecting diffusion directions on
tensor-derived parameters. Neuroimage, 109:160–170, 2015.
[46] M Jenkinson, P Bannister, M Brady, and S Smith. Improved optimization for the robust and
21
accurate linear registration and motion correction of brain images. Neuroimage, 17:825–841,
2002.
[47] J L R Andersson and S N Sotiropoulos. An integrated approach to correction for off-resonance
effects and subject movement in diffusion MR imaging. Neuroimage, 125:1063–1078, 2016.
[48] M Jenkinson, C F Beckmann, T E Behrens, M W Woolrich, and S M Smith. Fsl. Neuroimage,
62:782–790, 2012.
[49] Theodore D Satterthwaite, Daniel H Wolf, Guray Erus, Kosha Ruparel, Mark A Elliott, Ef-
stathios D Gennatas, Ryan Hopson, Chad Jackson, Karthik Prabhakaran, Warren B Bilker,
et al. Functional maturation of the executive system during adolescence. Journal of Neuro-
science, 33(41):16249–16261, 2013.
[50] Douglas N Greve and Bruce Fischl. Accurate and robust brain image alignment using
boundary-based registration. Neuroimage, 48(1):63–72, 2009.
[51] CJ Honey, O Sporns, Leila Cammoun, Xavier Gigandet, Jean-Philippe Thiran, Reto Meuli,
and Patric Hagmann. Predicting human resting-state functional connectivity from structural
connectivity. Proceedings of the National Academy of Sciences, 106(6):2035–2040, 2009.
[52] Louis M Pecora and Thomas L Carroll. Master stability functions for synchronized coupled
systems. Physical review letters, 80(10):2109, 1998.
[53] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical and pow-
erful approach to multiple testing. Journal of the royal statistical society. Series B (Method-
ological), pages 289–300, 1995.
[54] Simon N Wood. Stable and efficient multiple smoothing parameter estimation for generalized
additive models. Journal of the American Statistical Association, 99(467):673–686, 2004.
[55] Dustin Tingley, Teppei Yamamoto, Kentaro Hirose, Luke Keele, and Kosuke Imai. Mediation:
R package for causal mediation analysis. 2014.
[56] KJ Preacher, PJ Curran, and DJ Bauer. Simple intercepts, simple slopes, and regions of
significance in mlr 2-way interactions. retrieved september 25, 2005, 2004.
[57] Joan G Snodgrass and June Corwin. Pragmatics of measuring recognition memory: appli-
cations to dementia and amnesia. Journal of Experimental Psychology: General, 117(1):34,
1988.
[58] F Xavier Castellanos and Rosemary Tannock. Neuroscience of attention-deficit/hyperactivity
disorder: the search for endophenotypes. Nature Reviews Neuroscience, 3(8):617–628, 2002.
[59] Tiina Pohjalainen, Juha O Rinne, Kjell Nagren, Erkka Syvalahti, and Jarmo Hietala. Sex
differences in the striatal dopamine d2 receptor binding characteristics in vivo. American
Journal of Psychiatry, 155(6):768–773, 1998.
[60] K.E. Stephan, F. Schlagenhauf, Q.J.M. Huys, S. Raman, E.A. Aponte, K.H. Brodersen,
L. Rigoux, R.J. Moran, J. Daunizeau, R.J. Dolan, K.J. Friston, and A. Heinz. Compu-
tational neuroimaging strategies for single patient predictions. NeuroImage, 145(Part B):
ISSN 1053-8119. doi: https://doi.org/10.1016/j.neuroimage.2016.06.038.
180 – 199, 2017.
22
URL http://www.sciencedirect.com/science/article/pii/S1053811916302877.
vidual Subject Prediction.
Indi-
[61] Eugene M Izhikevich and Gerald M Edelman. Large-scale model of mammalian thalamocortical
systems. Proceedings of the national academy of sciences, 105(9):3593–3598, 2008.
[62] A C Murphy and D S Bassett. A network neuroscience of neurofeedback for clinical translation.
Curr Opin Biomed Eng, 1:63–70, 2017.
[63] Jun Tanji and Eiji Hoshi. Role of the lateral prefrontal cortex in executive behavioral control.
Physiological reviews, 88(1):37–57, 2008.
[64] David R Euston, Aaron J Gruber, and Bruce L McNaughton. The role of medial prefrontal
cortex in memory and decision making. Neuron, 76(6):1057–1070, 2012.
[65] Cummings JL. Frontal-subcortical circuits and human behavior. Archives of Neurology, 50(8):
873–880, 1993. doi: 10.1001/archneur.1993.00540080076020. URL +http://dx.doi.org/10.
1001/archneur.1993.00540080076020.
[66] Lisa M Duke and Alfred W Kaszniak. Executive control functions in degenerative dementias:
A comparative review. Neuropsychology review, 10(2):75–99, 2000.
[67] Raphael M Bonelli and Jeffrey L Cummings. Frontal-subcortical circuitry and behavior. Dia-
logues in clinical neuroscience, 9(2):141, 2007.
[68] Sibel Tekin and Jeffrey L Cummings. Frontal–subcortical neuronal circuits and clinical neu-
ropsychiatry: an update. Journal of psychosomatic research, 53(2):647–654, 2002.
[69] St´ephanie Bombois, St´ephanie Debette, Xavier Delbeuck, Am´elie Bruandet, Samuel Lepoit-
tevin, Christine Delmaire, Didier Leys, and Florence Pasquier. Prevalence of subcortical vas-
cular lesions and association with executive function in mild cognitive impairment subtypes.
Stroke, 38(9):2595–2597, 2007. ISSN 0039-2499. doi: 10.1161/STROKEAHA.107.486407. URL
http://stroke.ahajournals.org/content/38/9/2595.
[70] Michael M. Plichta, Nenad Vasic, Robert Christian Wolf, Klaus-Peter Lesch, Dagmar Brum-
mer, Christian Jacob, Andreas J. Fallgatter, and Georg Gron. Neural hyporesponsiveness
and hyperresponsiveness during immediate and delayed reward processing in adult attention-
deficit/hyperactivity disorder. Biological Psychiatry, 65(1):7 – 14, 2009.
ISSN 0006-3223.
doi: https://doi.org/10.1016/j.biopsych.2008.07.008. URL http://www.sciencedirect.com/
science/article/pii/S0006322308008275. Perception, Empathy, and Reward in Attention-
Deficit/Hyperactivity Disorder and Autism.
[71] Ana Cubillo, Rozmin Halari, Anna Smith, Eric Taylor, and Katya Rubia. A review of fronto-
striatal and fronto-cortical brain abnormalities in children and adults with attention deficit
hyperactivity disorder (adhd) and new evidence for dysfunction in adults with adhd during
motivation and attention. Cortex, 48(2):194 – 215, 2012. ISSN 0010-9452. doi: https://doi.
org/10.1016/j.cortex.2011.04.007. URL http://www.sciencedirect.com/science/article/
pii/S001094521100102X. Frontal lobes.
[72] Chandan J Vaidya. Neurodevelopmental abnormalities in adhd. In Behavioral neuroscience of
attention deficit hyperactivity disorder and its treatment, pages 49–66. Springer, 2011.
23
[73] Naftali Raz, Ivan J Torres, and James D Acker. Age, gender, and hemispheric differences in
human striatum: a quantitative review and new data from in vivo mri morphometry. Neuro-
biology of learning and memory, 63(2):133–142, 1995.
[74] Walter J. Freeman and Christine A. Skarda. Spatial eeg patterns, non-linear dynamics
the neo-sherringtonian view. Brain Research Reviews, 10(3):147 – 175,
ISSN 0165-0173. doi: https://doi.org/10.1016/0165-0173(85)90022-0. URL http:
and perception:
1985.
//www.sciencedirect.com/science/article/pii/0165017385900220.
[75] Tong Zhu, Rui Hu, Xing Qiu, Michael Taylor, Yuen Tso, Constantin Yiannoutsos, Bradford
Navia, Susumu Mori, Sven Ekholm, Giovanni Schifitto, et al. Quantification of accuracy
and precision of multi-center dti measurements: a diffusion phantom and human brain study.
Neuroimage, 56(3):1398–1411, 2011.
[76] Klaus H Maier-Hein, Peter F Neher, Jean-Christophe Houde, Marc-Alexandre Cot´e, Eleftherios
Garyfallidis, Jidan Zhong, Maxime Chamberland, Fang-Cheng Yeh, Ying-Chia Lin, Qing Ji,
et al. The challenge of mapping the human connectome based on diffusion tractography. Nature
communications, 8:1349, 2017.
[77] Marek Kubicki, Robert McCarley, Carl-Fredrik Westin, Hae-Jeong Park, Stephan Maier, Ron
Kikinis, Ferenc A Jolesz, and Martha E Shenton. A review of diffusion tensor imaging studies
in schizophrenia. Journal of psychiatric research, 41(1):15–30, 2007.
[78] Michela Pievani, Willem de Haan, Tao Wu, William W Seeley, and Giovanni B Frisoni. Func-
tional network disruption in the degenerative dementias. The Lancet Neurology, 10(9):829–843,
2011.
[79] Daphna Joel, Zohar Berman, Ido Tavor, Nadav Wexler, Olga Gaber, Yaniv Stein, Nisan Shefi,
Jared Pool, Sebastian Urchs, Daniel S Margulies, et al. Sex beyond the genitalia: The human
brain mosaic. Proceedings of the National Academy of Sciences, 112(50):15468–15473, 2015.
[80] Daphna Joel and Margaret M McCarthy. Incorporating sex as a biological variable in neu-
ropsychiatric research: where are we now and where should we be? Neuropsychopharmacology,
42(2):379–385, 2017.
24
Figures
Figure 1: Control theory and schematic of data processing. (a) Schematic depicting ba-
sic methodological approach. Diffusion tensor imaging (DTI) data was acquired from 882 youth
ages 8 to 22 years. Deterministic tractography was used to identify the number of white matter
streamlines, connecting any two regions of interest. These estimates were used to construct a struc-
tural brain network for each subject representing white matter connectivity (edges) between brain
regions (nodes). (b) Diagram illustrating brain state transitions from an activated default mode
system (Initial ). The blue arrow denotes a distant transition to a deactivated default mode system
and activated frontoparietal/dorsal attention system (Distant). The red arrow denotes a nearby
transition to a partially deactivated default mode system (Near ). Regions with high modal con-
trollability can facilitate transitions to energetically distant states while regions with high average
controllability can facilitate transitions to nearby states while requiring very little energy.
25
a.DTIDeterministicTractographyConnectivity MatrixBrain Network0Energy1b.InitialDistantNearModalcontrollabilityAveragecontrollability Figure 2: Cognitive development and network controllability by sex. (a) Executive effi-
ciency, a summary of performance on tasks taxing executive function in the PCNB, fit non-linearly
on age with a penalized regression spline. Red and blue data points and curves represent females
and males, respectively. Each point represents a subject's partial residual with respect to sex; solid
lines represent the predicted mean for males and females separately, and shaded envelopes denote
95% confidence intervals of the prediction. (b) Rank of mean average controllability values at each
node across 389 male subjects and 493 female subjects, separately. The Spearman rank correlation
between regional average controllability averaged across males and regional average controllability
averaged across females was r = 0.9915.
26
a.b.r = 0.99MalesFemales1234−3−2−1012810121416182022Age (y)Executive Efficiency Figure 3: Sex differences in regional controllability. (a-b) Controllability estimates averaged
over 219 cortical nodes (a) and 14 subcortical nodes (b), for average controllability and modal
controllability. Females have higher cortical average controllability, cortical modal controllabillity
and subcortical modal controllability, but males have higher subcortical average controllability. In
panels a-b, red and blue represent females and males, respectively. Each point represents a subject's
partial residual with respect to sex; bar height represents the predicted mean for males and females
separately, and shaded grey envelopes denote 95% confidence intervals of the prediction. (c-d)
Heatmaps where color intensity corresponds to standardized regression coefficient for sex, with
warmer colors indicating higher controllability in females. Average controllability is depicted in
(c) and modal controllability is depicted in (d). Colors reflect values of standardized regression
coefficients for controllability at each node as a predictor of executive function while controlling for
covariates (Eq. 5). *, p < 0.05, **, p < 0.01, ***, p < 10−4.
27
c.a.b.-0.2100.15M> FM > F0F > MF > M101520FemalesMalesCorticalAverage Controllability0.940.960.981.00FemalesMalesCorticalModal Controllability100200FemalesMalesSubcorticalAverage Controllability0.900.951.00FemalesMalesSubcorticalModal Controllabilityd.Average Controllability ~ SexModal Controllability ~ Sex0.18-0.18zz Figure 4: Network controllability as a function of age. (a-d) Relationship between age and
controllability estimates, averaged over 219 cortical nodes (a-b) and 14 subcortical nodes (c-d), for
average controllability (a, c) and modal controllability (b, d). Modal controllability increases with
age in the cortex and in the subcortex; average controllability increases with age in the cortex. In
the subcortex, there was a significant age-by-sex interaction for average controllability. However,
the simple slopes for males (βage) and females (βage + βage×sex) were not significantly different from
0. Age-by-sex interactions in a, b, and d were not significant. Red and blue represent females and
males, respectively. Each point represents a subject's partial residual with respect to age; lines
represent regression slopes for males and females separately, and shaded envelopes denote 95%
confidence intervals of the slopes. The p values for βage are shown in panels (a,b,d) and for βage×sex
are shown in panel (c).
28
a.b.c.d.FemalesMales8121620pβage<10e−41015208121620Age (y)Average ControllabilityCorticalSubcorticalpβint<0.051002008121620Age (y)Average Controllabilitypβage<10e−40.900.951.008121620Age (y)Modal Controllabilitypβage<10e−40.940.960.981.00Age (y)Modal Controllability Figure 5: Sex effects on relationships between metrics of network control and dynamics.
Relationships between whole-brain average and modal controllability (a, e), between synchroniz-
ability and whole-brain average controllability (d, h), and between regional average controllability
and regional modal controllability (b-c, f-g). Data from male subjects is depicted in panels a-d and
data from female subjects is depicted in panels e-h. Non-linear least squares was used on the base
model: y = a+b∗exp(c∗x), and the full model: y = (a+α∗Sex)+(b+β∗Sex)∗exp((c+γ∗Sex)∗x).
Slopes represent coefficients from the full model. The p values were obtained from ANOVA com-
paring base and full models. In these cases, a low p value indicates that allowing for a sex term in
the model significantly increases the explained variance. *, p < 0.05, **, p < 0.01, ***, p < 10−4.
29
912151821Age (y)a.MalesFemalesb.c.d.h.g.e.0.940.950.960.970.98202530Modal ControllabilityGlobal Mean***************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************0.950.960.970.98101520Modal ControllabilityCortical Mean***************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************0.8750.9000.9250.9500.975100200Modal ControllabilitySubcortical Mean1.01.52.02.5202530SynchronizabilitySynchronizability0.940.950.960.970.98202530Average ControllabilityModal Controllability***************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************0.950.960.970.98101520Average ControllabilityModal Controllability***************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************0.8750.9000.9250.9500.975100200Average ControllabilityModal Controllability1.01.52.02.5202530Average ControllabilitySynchronizabilityf. Figure 6: Network controllability and cognitive performance. (a-d) Relationship between
age and controllability estimates, averaged over 219 cortical nodes (a-b) and 14 subcortical nodes
(c-d), for average controllability (a, c) and modal controllability (b, d). Cortical controllability
(a-b) is not significantly related to executive function, but subcortical average controllability (c)
is negatively correlated with executive function while subcortical modal controllability (d) is posi-
tively correlated with executive function. In panels a-d, red and blue represent females and males,
respectively. Each data point represents a subject's partial residual with respect to controllability;
lines represent regression slopes for males and females separately, and shaded envelopes denote 95%
confidence intervals of the slopes. (e) Brain regions with sex-associated average controllability and
their relation to executive function (FDR corrected, q < 0.05). Colors reflect values of standardized
regression coefficients for average controllability at each node as a predictor of executive function
while controlling for covariates. Modal controllability was not significantly associated with execu-
tive function. For panels a-e, there were no significant sex-controllability interactions on executive
function.
30
a.b.c.d.e.-0.080.080zExecutive Efficiency ~ Average ControllabilityNS−3−2−1012101520Average ControllabilityExecutive EfficiencyNS−3−2−10120.960.981.00Modal Controllabilitypβage<0.01−3−2−101250100150200250Average ControllabilityExecutive Efficiencypβage<0.05−3−2−10120.880.920.961.00Modal ControllabilityCorticalSubcortical R Medial orbito frontal, parcel 1
R Superior frontal, parcel 4
R Postcentral, parcel 5
R Caudate
L Caudate
Average Controllability
Control Node ACME
0.03
0.01
0.01
0.03
0.02
p ADE
0.13
0.15
0.15
0.13
0.14
0.00
0.43
0.10
0.01
0.03
p
0.01
0.01
0.00
0.01
0.01
Figure 7: Average controllability mediates the relation between sex and executive func-
tion. (a) Non-parametric causal mediation analysis for sex → average controllability → executive
function, using control nodes with sex-associated average controllability values. ACME = average
causal mediation effect, i.e. average controllability → executive function; ADE = average direct
effect, i.e. sex → executive function. Associations between sex and controllability from Fig. 3c
are shown; cooler colors indicate higher controllability in males. 1000 simulations were performed.
p-values were FDR corrected (q < 0.05). Age, total brain volume, handedness, and head motion
were included in all analyses. * indicates standardized regression coefficient for sex as a predictor
of executive function with the same covariates in the model.
31
SexExecutive Efficiency(Fig. 2c)-0.030.030ControllabilityAverage-0.210.150ADE0.16* Figure 8: Regional controllability predicts n-back task activation and cognitive per-
formance differently for males and females. (a) Brain regions at which the 2-back minus
0-back GLM β weights are associated with 2-back d(cid:48), a summary measure of task performance
in n = 659 subjects with quality fMRI data. (b) Heatmap depicting standardized multiple linear
regression β's for regional average controllability as a predictor of regional 2-back minus 0-back
activation. Only nodes with associations surviving FDR correction are shown (q < 0.05). (c) Vi-
sual representation of the data presented in panel (b); average controllability at highlighted nodes
is associated with activation at nodes of the same colors. (d) Associations between activation and
average controllability for n = 283 males and n = 376 females, separately. Modal controllability
was not significantly associated with activation. For panels (c,d), corresponding colors indicate
association between control node and regional activation.
32
-0.200.2a.b.c.d.d' ~ Activation MalesFemales-0.1600.22zzControllabilityActivation A Supplementary Data
33
Figure A.1: Whole-brain network controllability and synchronizability by sex. (a, c)
Average (a) and modal (c) controllability, averaged across all brain regions (hereafter "global"),
comparing males and females. (b) Whole-brain ("global") synchronizability over age. (d) Global
synchronizability by sex. We observed no significant age-by-sex interactions on synchronizability.
Red and blue represent females and males, respectively. Each point represents a subject's partial
residual with respect to sex; bar height represents the predicted mean for males and females sep-
arately, and shaded envelopes denote 95% confidence intervals of the prediction. *, p < 0.05, **,
p < 0.01, ***, p < 10−4.
34
**1520253035a.b.c.d.FemalesMalesGlobal Mean Average Controllabilitypβage<10e−161.01.52.02.5810121416182022Age (y)Global Synchronizability**0.920.940.960.981.00FemalesMalesGlobal Mean Modal ControllabilityNS1.01.52.02.5FemalesMalesGlobal SynchronizabilityFemalesMales Figure A.2: Trends between regional weighted degree, age, and sex differ from those
with network controllability. (a-d) Weighted degree averaged over 219 cortical nodes (a-b) and
14 subcortical nodes (c-d). Panels (a, c) show that no significant sex differences exist in mean
cortical (a) or subcortical (c) weighted degree. Panels (b, d) show that weighted degree decreases
with age in the cortex (b) and increases with age in the subcortex (d). We observed no significant
age-by-sex interactions in these models. Red and blue represent females and males, respectively.
Each point represents a subject's partial residual with respect to sex; bar height represents the
predicted mean for males and females separately, and shaded envelopes denote 95% confidence
intervals of the prediction. *, p < 0.05, **, p < 0.01, ***, p < 10−4.
35
NS3500400045005000FemalesMalesMean Cortical Weighted Degreepβage<0.013500400045005000810121416182022Age (y)Mean Cortical Weighted DegreeNS5000700090001100013000FemalesMalesMean Subcortical Weighted Degreepβage<0.055000700090001100013000810121416182022Age (y)Mean Subcortical Weighted Degreea.b.c.d.FemalesMales Figure A.3: Relationship between sex, network controllability, and cognitive perfor-
mance is not well-explained by weighted degree. (a) Heatmaps where color intensity corre-
sponds to standardized regression coefficient for sex, with warmer colors indicating higher weighted
degree in females. Five out of twenty-one regions with sex-associated controllability also have sex-
associated weighted degree. Of those five, only average controllability values at the left caudate
nucleus and right postcentral gyrus have the same direction of assocation with sex as weighted de-
gree. (b) Association between executive function and weighted degree at nodes with sex-associated
average controllability, modal controllability, or weighted degree, correcting for multiple compar-
isons with an FDR correction at q < 0.05. No sex × weighted degree interaction terms on executive
function were significant. Age, total brain volume, handedness, and head motion were included as
covariates in all analyses.
36
b.-0.0700.07Executive Efficiency ~ Weighted Degreea.-0.150 Figure A.4: Sex-dependent relationships between network controllability and activation.
(a) Heatmap of standardized multiple linear regression β's for regional average controllability as a
predictor of regional 2-back minus 0-back activation. (Left) Data from n = 376 females. (Right)
Data from n = 283 males. Only nodes with associations surviving FDR correction are shown
(q < 0.05). (b) Heatmap of standardized β's for average controllability × sex interaction. No β
was significant after FDR correction (q < 0.05), but nodes from panel (a) are shown to illustrate
consistent directionality of association. Especially obvious are the left caudal middle frontal (parcel
1) and left banks of STS (parcel 2), where darker colors on the interaction heatmap indicate a larger
effect size in males. The left fusiform gyrus (parcel 4) shows the opposite trend, with lighter colors
on the interaction heatmap indicating a larger effect size in females.
37
a.b.-0.50.5-0.20.200-0.250.250F > MM > Fzz Figure A.5: Relationship between weighted degree and activation. (a) Heatmap of multiple
linear regression β's for regional weighted degree as a predictor of regional 2-back minus 0-back
activation. Only nodes with associations surviving FDR correction are shown (q < 0.05). There
were only significant associations in the sample of n = 283 males, and not in the pooled sample
or in females alone.
(b) Visualization of associations between weighted degree and activation.
Weighted degree at highlighted nodes is associated with activation at nodes of the same colors;
that is, corresponding colors indicate the association between structural connections and regional
activation.
38
Weighted DegreeActivationMalesb.0.150-0.15za. R Medial orbito frontal1
R Rostral middle frontal6
R Superior frontal4
R Postcentral5
R Superior parietal1
R Inferior parietal1
R Inferior parietal3
R Lateral occipital1
R Fusiform4
R Caudate
L Lateral orbitofrontal3
L Pars triangularis1
L Precentral4
L Precentral8
L Postcentral1
L Postcentral3
L Postcentral4
L Superior parietal2
L Superior parietal5
L Inferior parietal3
L Lateral occipital4
L Lateral occipital5
L Fusiform4
L Superior temporal4
L Insula3
L Caudate
L Putamen
L Pallidum
L Amygdala
ACME
0.01
0.01
0.00
0.01
-0.00
-0.00
-0.00
-0.01
0.00
0.01
-0.01
-0.00
0.00
0.00
-0.00
0.00
-0.00
-0.01
-0.00
0.00
0.00
0.00
0.01
-0.01
0.00
0.01
0.00
0.01
0.01
a. Weighted Degree
p ADE
0.15
0.15
0.15
0.15
0.16
0.15
0.16
0.16
0.15
0.15
0.17
0.16
0.16
0.16
0.16
0.15
0.16
0.17
0.16
0.16
0.16
0.15
0.15
0.16
0.15
0.15
0.15
0.15
0.15
0.54
0.25
0.79
0.77
0.91
0.91
0.91
0.54
0.85
0.06
0.25
0.91
0.91
0.91
0.91
0.91
0.91
0.54
0.91
0.86
0.91
0.86
0.78
0.70
0.86
0.54
0.70
0.61
0.54
p Total Effect
0.16
0.16
0.15
0.15
0.15
0.15
0.15
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.15
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.16
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
p Prop. Mediated
0.06
0.06
0.02
0.05
-0.00
-0.00
-0.01
-0.04
0.03
0.08
-0.07
-0.00
0.01
0.01
-0.01
0.00
-0.01
-0.03
-0.01
0.01
0.00
0.02
0.05
-0.04
0.01
0.06
0.02
0.04
0.04
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.01
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
p
0.55
0.27
0.79
0.77
0.91
0.91
0.91
0.55
0.86
0.12
0.27
0.91
0.91
0.91
0.91
0.91
0.91
0.55
0.91
0.86
0.91
0.86
0.78
0.70
0.86
0.55
0.70
0.61
0.55
Figure A.6: Sex-(weighted degree)-executive function mediation relationship. (a-b) Non-
parametric causal mediation analysis for sex → weighted degree → executive function, using control
nodes with sex-associated average controllability, modal controllability, or weighted degree. ACME
= average causal mediation effect, ADE = average direct effect. We performed 1000 simulations.
The p-values were FDR corrected (q < 0.05). Age, total brain volume, handedness, and head
motion were included in all analyses.
39
|
1112.4987 | 1 | 1112 | 2011-12-21T11:07:16 | Biologically-Inspired Electronics with Memory Circuit Elements | [
"q-bio.NC",
"cond-mat.mes-hall"
] | Several abilities of biological systems, such as adaptation to natural environment, or of animals to learn patterns when appropriately trained, are features that are extremely useful, if emulated by electronic circuits, in applications ranging from robotics to solution of complex optimization problems, traffic control, etc. In this chapter, we discuss several examples of biologically-inspired circuits that take advantage of memory circuit elements, namely, electronic elements whose resistive, capacitive or inductive characteristics depend on their past dynamics. We provide several illustrations of what can be accomplished with these elements including learning circuits and related adaptive filters, neuromorphic and cellular computing circuits, analog massively-parallel computation architectures, etc. We also give examples of experimental realizations of memory circuit elements and discuss opportunities and challenges in this new field. | q-bio.NC | q-bio |
Biologically-Inspired Electronics with Memory
Circuit Elements
Massimiliano Di Ventra and Yuriy V. Pershin
Abstract Several unique properties of biological systems, such as adaptation to nat-
ural environment, or of animals to learn patterns when appropriately trained, are
features that are extremely useful, if emulated by electronic circuits, in applications
ranging from robotics to solution of complex optimization problems, traffic con-
trol, etc. In this chapter, we discuss several examples of biologically-inspired cir-
cuits that take advantage of memory circuit elements, namely, electronic elements
whose resistive, capacitive or inductive characteristics depend on their past dynam-
ics. We provide several illustrations of what can be accomplished with these ele-
ments including learning circuits and related adaptive filters, neuromorphic and cel-
lular computing circuits, analog massively-parallel computation architectures, etc.
We also give examples of experimental realizations of memory circuit elements and
discuss opportunities and challenges in this new field.
1 Introduction
Reproducing some of the features that are commonly found in living organisms, in-
cluding - as the ultimate, and most sought after goal - the workings of the human
brain, is what "artificial intelligence" is all about [1]. However, even without aiming
for such an ambitious target, there are several tasks that living organisms perform
seamlessly and, when reproduced in electronic circuits, are of great benefit and help
us solve complicated problems. The main idea of biologically-inspired electronics is
Massimiliano Di Ventra
Department of Physics, University of California San Diego, La Jolla, CA 92093 USA,
e-mail: [email protected]
Yuriy V. Pershin
Department of Physics and Astronomy and USC Nanocenter, University of South Carolina,
Columbia, SC 29208 USA,
e-mail: [email protected]
1
2
Massimiliano Di Ventra and Yuriy V. Pershin
thus in borrowing approaches used by biological systems interacting with their envi-
ronment and in applying them in diverse technological areas requiring, for example,
adaptation to changing inputs, analog solution of optimization problems -- one of
the well-known approaches to this problem is the "ant-search algorithm" [2, 3] -- ,
associative memory and unlearning, binary and fuzzy logic, etc.
If we look closer, it is evident that if we want to reproduce any of these tasks
with electronic circuits, two main characteristics have to be satisfied by some of the
circuit elements or their combinations. These elements need to i) store information
-- they have to have memory of the past -- and ii) be dynamical -- their states have
to vary in time in response to a given input, preferably in a non-linear way. The
latter requirement will help building a wide range of electronic circuits of desired
functionality.
Circuits based on active elements (such as transistors) can clearly perform both
tasks, however, at a high cost of power consumption, low density, and complexity.
It would be much more desirable if we could combine the above features in single,
passive elements, preferably with dimensions at the nanometer scale, and hence
comparable to -- or even smaller than -- biological storing and processing units, such
as synapses and neurons.
Such elements do exist, and go under the name of memristive, memcapacitive
and meminductive systems, or collectively simply named memelements [4]. These
are resistors, capacitors and inductors, respectively, whose state at a given time de-
pends on the history of applied inputs (e.g., charge, voltage, current, or flux) and
states through which the system has evolved. As we have shown, these memele-
ments provide an unifying description of materials and systems with memory [5],
in the sense that all two-terminal electronic devices based on memory materials and
systems, when subject to time-dependent perturbations, behave simply as -- or as a
combination of -- memristive, memcapacitive and meminductive systems, namely,
dynamical non-linear circuit elements with memory [6].
In this Chapter, we will show how the analog memory features of memelements
are ideal to reproduce a host of processes typical of living organisms. We will review
mainly work by the authors in various contexts, ranging from learning (adaptive) cir-
cuits to associative memory, and inherent massive parallelism afforded by networks
of memelements. The Chapter is then organized as follows. Section 2 briefly reviews
both the definition of these memory circuit elements and their main properties (a full
account can be found in the original publications [7, 8, 4] and in our recent review
paper [6]). Several experimental realizations exemplifying memristive, memcapac-
itive and meminductive systems are discussed in Sec. 3. Section 4 is devoted to
biologically-inspired circuits based on memory circuit elements including simple
adaptive circuits (Sec. 4.1), neuromorphic circuits (Sec. 4.2) and massively-parallel
analog processing circuits (Sec. 4.3). Concluding remarks are given in Sec. 5.
Biologically-Inspired Electronics with Memory Circuit Elements
3
Fig. 1 Symbols of memory circuit elements: memristor, memcapacitor and meminductor. Gener-
ally, memelements are asymmetric devices. The thick horizontal lines above the bottom electrodes
are employed to define the device polarity [4]. Reprinted with permission from Ref. [4]. c(cid:13)2009
IEEE.
2 Definitions and properties
Let us consider electronic devices defined by all possible pairs of fundamental cir-
cuit variables u(t) and y(t) (i.e., current, charge, voltage, or flux). For each pair,
we can introduce a response function, g, that, generally, depends on a set of n state
variables, x = {xi}, i = 1, ..,n, describing the internal state of the system [4] 1. For
instance, resistance of certain systems may depend on spin polarization [9, 10] or
temperature [8]; capacitance and inductance of some other elements may exhibit
dependence on system geometry [11, 6] or electric polarization [12]. In general,
all internal microscopic physical properties related to the memory response of these
electronic devices should be included into the vector of internal state variables x. The
resulting memory circuit elements are then described by the following relations [4]
y(t) = g (x,u,t)u(t)
x = f (x,u,t)
(1)
(2)
where f is a continuous n-dimensional vector function. Eqs. (1), (2) have to be sup-
plied by appropriate initial conditions [13]. If u is the current and y is the voltage
then Eqs. (1), (2) define memristive (for memory resistive) systems. In this case g is
the memristance (memory resistance). In memcapacitive (memory capacitive) sys-
tems, the charge q is related to the voltage V so that g is the memcapacitance (mem-
ory capacitance). Finally, in meminductive (memory inductive) systems the flux ϕ
is related to the current I with g the meminductance (memory inductance). There
are still three additional pairs of fundamental circuit variables. However, these do
not give rise to any new devices. For example, the pairs charge-current and voltage-
flux are linked through equations of electrodynamics. Moreover, we could redefine
devices defined by the charge-flux (which is the integral of the voltage) pair in the
current-voltage basis [7]. Circuits symbols of memristive, memcapacitive and me-
minductive systems are presented in Fig. 1.
1 There is no such dependence for traditional basic circuit elements -- resistors, capacitors and
inductors.
MemristorMemcapacitorMeminductorMemristorMemcapacitorMeminductor 4
Massimiliano Di Ventra and Yuriy V. Pershin
Let us consider memristive systems in more details (definitions of all the other
elements can be easily derived by considering the different constitutive variables
[4]). Specifically, a current-controlled memristive system [8, 4] is defined by Eqs.
(1), (2) as
VM(t) = R (x,I,t)I(t),
x = f (x,I,t) ,
(3)
(4)
where VM(t) and I(t) denote the voltage and current across the device, and R is the
memristance. In a simple model of an ideal memristor [7], the memristance depends
only on charge -- the time integral of the current. One particular realization of such
a model has been suggested in Ref. [14] and is formulated as
i j = RONx + ROFF (1− x) ,
RM
(5)
where RON and ROFF are minimal and maximal values of memristance, and x is a
dimensionless internal state variable bound to the region 0 ≤ x ≤ 1. The dynamics
of x can then be simply chosen as [14]
dx
dt = αI(t),
(6)
where α is a constant and I(t) is the current flowing through the memristor.
Another example of memristive systems (which we will make use of later in
this chapter) is a threshold-type memristive system [15]. Its model is specified by a
threshold voltage (and some other parameters) that defines different device response
regions (with regard to the voltage applied across the device). Mathematically, the
threshold-type memristive system is described by the following equations
I = x−1VM,
x = (βVM + 0.5 (α − β ) [VM +Vt−VM −Vt])
×θ (x/R1 − 1)θ (1− x/R2) ,
(7)
(8)
where I and VM are the current through and the voltage drop on the device, respec-
tively, and x is the internal state variable playing the role of memristance, R = x,
θ (·) is the step function, α and β characterize the rate of memristance change at
VM ≤ Vt and VM > Vt, respectively, Vt is a threshold voltage, and R1 and R2 are
limiting values of the memristance R. In Eq. (8), the θ-functions symbolically show
that the memristance can change only between R1 and R2. On a practical level, the
value of x must be monitored at each time step and in the situations when x < R1 or
x > R2, it must be set equal to R1 or R2, respectively. In this way, we avoid situations
when x may overshoot the limiting values by some amount and thus not change any
longer because of the step function in Eq. (8). We have introduced and employed
this model to describe the learning properties of unicellular organisms, associative
memory, etc., as we will describe in some detail in the following sections.
Biologically-Inspired Electronics with Memory Circuit Elements
5
There are several properties that characterize memory circuit elements. We refer
the reader to the original papers [7, 8, 4] and the extensive review [6] where these are
discussed at length. Here, we just mention that they are typically characterized by a
frequency-dependent "pinched hysteresis loop" in their constitutive variables when
subject to a periodic input. Also, normally, the memristance, memcapacitance and
meminductance acquire values between two limits (with exceptions as discussed in
Refs. [6, 12, 16]). Although the hysteresis of these elements under a periodic input
may strongly depend on initial conditions [13], it is generally more pronounced
at frequencies of the external input that are comparable to frequencies of internal
processes that lead to memory. In many cases, at very low frequencies memory
circuit elements behave as non-linear elements while at high frequencies as linear
elements.
Also, the hysteresis loops may or may not show self-crossing - which we have
named type-I and type-II crossing behavior, respectively [6] - and often the internal
state variable remains unchanged for a long time without any input signal applied.
This provides non-volatile memory, which is an important feature for some of the
applications we discuss later.
Finally, we mention that the state variables - whether from a continuum or a
discrete set of states - may follow a stochastic differential equation rather than a de-
terministic one [6]. Interesting effects have been predicted in the presence of noise,
such as noise-induced hysteresis [17]. This may have a large bearing in simulating
biological processes - which necessarily occur under noisy conditions - and could
be used to enhance the performance of certain devices.
3 Experimental realizations
In this section, we briefly discuss experimental realizations of memory circuit el-
ements. There is a large amount of experimental systems showing memristive be-
havior (based, however, on very different physical mechanisms). For example, in
thermistors -- being among the first identified memristive systems [8] -- the memory
effects are related to thermal effects, mainly on how fast the heat dissipation occurs.
In spintronics memristive systems, either based on semiconductor [9] or metal [10]
materials, the memory feature is provided by the electron spin degree of freedom.
Finally, resistance switching memory cells are probably the most important type of
memristive systems. These cells are normally built of two metal electrodes sepa-
rated by a gap of a few tens of nanometers filled by a memristive material. Different
physical processes inside the memristive material can be responsible for the mem-
ory. Fig. 2 shows an example of resistance switching memory cell -- an electrochem-
ical metallization cell -- in which a layer of dielectric material (SiO2) separates two
dissimilar metal electrodes (made of copper and platinum) [18]. Externally applied
voltage induces migration of copper atoms that can bridge the gap between the elec-
trodes thus reducing the cell's resistance. Such a bridge can also be disrupted by the
applied voltage of the opposite polarity (see Fig. 2).
6
Massimiliano Di Ventra and Yuriy V. Pershin
Fig. 2 Current-voltage char-
acteristic of a Cu/SiO2/Pt
electrochemical metallization
cell recorded using a triangu-
lar voltage sweep. The insets
show dynamics of metallic fil-
ament formation. (Reprinted
with permission from [18].
2009 American Institute of
Physics.)
Memcapacitive effects can be related to changes in the capacitor geometry or
specific intrinsic properties of dielectric medium located between the capacitor
plates [6]. For instance, the former mechanism plays the main role in elastic mem-
capacitive systems that could be based on internal elastic states of the capacitor
plate (e.g., direction of bending [11]) or elasticity of a medium between the plates
that could be thought of as a spring [6]. Examples of the latter mechanism include
memcapacitive systems with a delayed polarization response [12, 16] and struc-
tures with permittivity switching [19]. Although meminductive systems are the least
studied memory circuit elements at the moment, there are several known systems
showing such type of functionality. For instance, in bimorph meminductive sys-
tems [20, 21, 22], the inductance depends on the inductor's shape defined by the
inductor temperature. Heat dissipation mechanisms play a significant role in this
type of systems. Many additional examples of memristive, memcapacitive and me-
minductive systems can be found in our recent review paper [6]. The most important
aspect of all these examples is that they relate primarily to structures with nanoscale
dimensions. This is not surprising since (up to a certain limit) the smaller the di-
mensions of the system the easier it is to observe memory effects.
In addition, we would like to mention that several types of memory effects may
be present in a single device. For example, the coexistance of memristive and mem-
capacitive responses has been observed in perovskite oxide thin films [23]. More-
over, several three-terminal transistor-like memristive devices have been investi-
gated in the past [24, 25, 26]. Clearly, different types of memory materials can be
combined to obtain multi-terminal device structures with complex functionalities.
Biologically-Inspired Electronics with Memory Circuit Elements
7
Fig. 3 (a) A possible real-
ization of the learning circuit
consisting of resistor R, in-
ductor L, capacitor C and
memristive system M. (b)
Schematics of the function
f (V ) defining a threshold-
type memristive system (see
also Eqs. (7), (8)). Reprinted
with permission from Ref.
[15]. c(cid:13)2009 American Phys-
ical Society.
4 Biologically-Inspired Circuits
4.1 Modeling adaptive behavior of unicellular organisms
Adaptive behavior is common to all life forms on Earth of all five kingdoms of na-
ture: in Plantae (the plants) [27, 28, 29], Animalia (the animals) [30], Protista (the
single-celled eukaryotes) [31, 32], Fungi (fungus and related organisms) [33, 34],
and Monera (the prokaryotes) [35, 36, 37]. (The literature on this subject is exten-
sive, hence we have given only a few representative references.) To a greater or
lesser extent, representatives of all life forms respond (adapt) to changes of their en-
vironment in a manner that increases the survival of their species. It would thus be
of benefit to mimic and use this important natural feature in artificial structures, in
particular in electronics to allow novel functionalities otherwise nonexistent in stan-
dard circuitry. There are indeed many domains where such learning circuits can be
employed. These range from robot control systems to signal processing. Therefore,
developing circuit models of the adaptive behavior of the natural world is of great
importance in many scientific areas [38]. Note that by "learning" here we simply
mean the ability to adapt to incoming signals with retention of such information for
later use.
In this section we consider a particularly interesting example: the ability of the
slime mold Physarum polycephalum to adapt its locomotion to periodic changes of
the environment [32]. The simplicity of the system - a unicellular organism - and
its well-defined response to specific input signals, make it an ideal test bed for the
application of the notion of memory circuit elements in biology, and a source of in-
spiration for more complex adaptive behavior in living organisms. In addition, this is
a particularly appealing example of the full range of properties of memelements, in
particular, their analog capability, which expands their range far beyond the digital
domain.
V(t)Rf(b)(a)LVVTVCM(t)V-VTC 8
Massimiliano Di Ventra and Yuriy V. Pershin
In particular, it has been shown in a recent experiment [32] that the Physarum
polycephalum subjected to periodic unfavorable environmental conditions (lower
temperature and humidity) not only responds to these conditions by reducing its
locomotion speed, but also anticipates future unfavorable environmental conditions
reducing the speed at the time when the next unfavorable episode would have oc-
curred. While the microscopic mechanism of such behavior has not been identified
by the authors of that work [32], their experimental measurements clearly prove
the ability of Physarum polycephalum to anticipate an impending environmental
change.
More specifically, the locomotion speed of the Physarum polycephalum was
measured when favorable environmental conditions (26◦C and 90% humidity) were
interrupted by three equally spaced 10 min pulses of unfavorable environmental
conditions (22◦C and 60% humidity) [32]. The time separation between the pulses τ
was selected between 30 and 90 minutes. It was observed that the locomotion speed
at favorable conditions (approximately 0.25 mm/10min as shown in Fig. 1 of Ref.
[32]) turns to close to zero each time the unfavorable conditions were presented.
However, spontaneous in-phase slow downs were observed after time intervals τ,
2τ and even 3τ after the last application of unfavorable conditions. In addition, if
-- after a long period of favorable conditions -- a single pulse of unfavorable condi-
tions is applied again then a spontaneous slow down (called a spontaneous in-phase
slow down after one disappearance [32]) after a time interval τ was observed. It
clearly follows form this experiment that the Physarum polycephalum has a mecha-
nism to memorize ("learn") the periodicity of environmental changes and adapts its
behavior in anticipation of next changes to come [32].
We have developed a circuit model [15] of the adaptive behavior of the slime
mold which has been later realized experimentally [39] using vanadium dioxide as
memory element [40, 41]. The learning circuit is shown in Fig. 3(a). Here, the role of
environmental conditions is played by the input voltage V (t) and the speed of loco-
motion is mapped into the voltage across the capacitor C. The learning circuit design
resembles a damped LC contour in which the amount of damping is controlled by
the state of the memristive system M. To understand the circuit operation, we note
that the memristive system employed in the circuit is of a threshold type (see Eqs.
(7) and (8)), namely, its state can be significantly changed only by a voltage (across
the memristive system) with a magnitude exceeding a certain threshold. Fig. 3(b)
presents the switching function f (V ) used to describe a threshold-type memristive
system.
Our simulations of the learning circuit response to irregular and regular se-
quences of pulses are shown in Fig. 4. In these simulations, the scheme described
above has been used with the only restriction that the response signal cannot exceed
a certain value [15] (electronically, a cut-off can be easily obtained by using an addi-
tional resistor and diode). When an irregular sequence of pulses is applied to the cir-
cuit, the voltage oscillations across the capacitor can not exceed the threshold volt-
age of the memristive system M which continues to stay in its initial low-resistance
state, thus damping the circuit. When the pulses are applied periodically with a pe-
riod close to the period of the LC contour oscillations, a sufficiently strong voltage
Biologically-Inspired Electronics with Memory Circuit Elements
9
Fig. 4 Modeling of the spontaneous in-phase slow-down responses [15]. This plot demonstrates
that stronger and longer-lasting responses for both spontaneous in-phase slow down and sponta-
neous in-phase slow down after one disappearance of the stimulus are observed only when the
circuit was previously trained by a periodic sequence of three equally spaced pulses as present in
V2(t). The applied voltage V1(t) is irregular and thus the three first pulses do not "train" the circuit.
Reprinted with permission from Ref. [15]. c(cid:13)2009 American Physical Society.
across the capacitor C is induced. This voltage switches the memristive system into
the high-resistance state. Therefore, in this case, oscillations in the contour are less
damped and last longer as Fig. 4 demonstrates. These oscillations exactly model the
spontaneous in-phase slow down and in-phase slow down after one disappearance
effects observed experimentally [32]. We note that a single learning circuit memo-
rizes past events of a frequency close to the resonance frequency of LC contour. An
array of learning circuits would model the learning of Physarum polycephalum in
the whole frequency range [15].
Recently, an experimental implementation of this learning circuit has been re-
ported [39]. In this work, a learning circuit similar to that in Fig. 3(a) has been built
with the only difference that the memristive system (a vanadium dioxide memristive
device [39]) has been connected in series with a capacitor (Fig. 5(a)). The memris-
tive properties of vanadium dioxide are based on an insulator-to-metal phase tran-
sition occurring in this material in the vicinity of 342K [40, 41]. In order to realize
the memristive functionality, the vanadium dioxide device is heated to a steady-state
temperature of 339.80 K (right below the transition temperature) and subjected to
an externally applied voltage. The Joule heating (due to the applied voltage) incre-
mentally drives the vanadium dioxide material through the phase transition, thus
reducing its resistance. The operation of the learning circuit depicted in Fig. 5(a) is
then clear. While off-resonance signals applied to the circuit can not excite a suffi-
cient current to drive the vanadium dioxide through the phase transition, the current
generated by resonance signals is sufficient for this purpose. Fig. 5(b) demonstrates
modification of the transfer function of the circuit caused by off-resonance and res-
050100950100010501100-3-2-101234-3-2-101234V2(t)Voltage (arb.units) Response (arb. units)Time (s)V1(t) 10
Massimiliano Di Ventra and Yuriy V. Pershin
Fig. 5 Experimental realization of the learning circuit (adaptive filter) based on vanadium dioxide
memristive device. (a) Schematic of the adaptive filter in which the memristive device (with a
small memcapacitive component) is connected in series with a capacitor C and inductor L. We
note that such realization of the learning circuit operates similarly to the learning circuit shown in
Fig. 3(a). (b) Small-signal (10 mV) transfer function (Vout /Vin) for the adaptive filter plotted before
and after off-resonance "A" and on-resonance "B" pulses. Solid lines are RLC bandpass-filter fit
to data, which generates the ω0 and Q values in the legend. Pulse sequence B has a significant
training effect on the circuit, while A has little or no effect. (c) Time series of the off-resonance
"A" sequence of pulses and on-resonance "B" sequence of pulses. Reprinted with permission from
Ref. [39]. c(cid:13)2010 American Institute of Physics.
onance pulse sequences (Fig. 5(c)) applied to the circuit. Fig. 5(b) clearly indicates
a change in the transfer function caused by resonance signals (learning).
Finally, we mention that the formalism of memory circuit elements [4] has also
been useful in modeling biophysical systems whose electric response depends on the
history of applied voltages or currents. An example of such situation is the electro-
osmosis in skin which has been recently described by a memristive model [42].
Physically, the voltage applied to the skin induces a water flow in sweat-duct cap-
illaries changing the skin conductance. The position of the water table (the level
separating dry and wet zones) in capillaries plays the role of the internal state vari-
able whose dynamics is determined by the applied voltage [42]. The memristive
model of electro-osmosis [42] is in a good agreement with experimental data and
further demonstrates the potential of the formalism of memory circuit elements for
modeling biophysical processes.
4.2 Neuromorphic circuits
Our second example of biologically-inspired circuits with memory circuit elements
is from the area of neural networks. Neural networks form a class of circuits whose
Biologically-Inspired Electronics with Memory Circuit Elements
11
Fig. 6 Memristive neural
network with an associative
memory ability. Here, two
input neurons (N1 and N2)
are connected through mem-
ristive synapses (S1 and S2)
to the output neuron N3. The
details of circuit operation are
given in the text. Reprinted
from Ref. [44]. c(cid:13)2010 with
permission from Elsevier.
operation mimics the operation of the human (and animal) brain. Below, we con-
sider electronic implementations of two important processes occurring in biolog-
ical neural networks: associative memory and spike timing-dependent plasticity.
Both features can be implemented in artificial neural networks based on memris-
tive synapses.
4.2.1 Associative memory
The associative memory is one of the most fundamental functionalities of the hu-
man (and animal) brain. By making associations we learn, adapt to a changing en-
vironment and better retain and recall events. One of the most famous experiments
related to associative memory is Pavlov's experiment [43] whereby salivation of a
dog's mouth is first set by the sight of food. Then, if the sight of food is accompanied
by a sound (e.g., the tone of a bell) over a certain period of time, the dog learns to
associate the sound to the food, and salivation can be triggered by the sound alone,
without the intervention of vision.
Recently, we have reproduced [44] the Pavlov's experiment utilizing a neural
network with memristive synapses. As a first example, we have implemented the
well known Hebbian rule introduced by Hebb in 1949: "when an axon of cell A is
near enough to excite a cell B and repeatedly or persistently takes part in firing it,
some growth process or metabolic change takes place in one or both cells such that
A's efficiency, as one of the cells firing B, is increased" [45]. To put it differently,
the neurons that fire together, wire together.
In order to show associative memory, let us consider a simple neural network
consisting of three electronic neurons and two memristive synapses as shown in
Fig. 6. We assume that the first input neuron activates under a specific ("visual")
event, such as "sight of food", and the second input neuron activates under another
("auditory") event, such as a particular "sound".
On the electronic level, an electronic neuron sends forward (to its output) and
backward (to its input) voltage spikes of opposite polarity when the amplitude of
the input signal exceeds a threshold value. Regarding the dynamics of memris-
tive synapses, they have been selected of a threshold-type (Eqs. (7) and (8)) with
a threshold voltage exceeding the output voltage of electronic neurons. In this case,
N1S1Input 1"ihffd"1NS1Output"sight of food"N3SInput 2"salivation"N2S2p"sound" 12
Massimiliano Di Ventra and Yuriy V. Pershin
voltage spikes applied to a single terminal of a memristive synapse is not enough
to induce its change. The latter is possible only if forward and backward propagat-
ing spikes overlap in time across a synapse. We have employed memristor emula-
tors [46, 44] as memristive synapses2. The main components of a memristor emula-
tor are a digital potentiometer, a microcontroller and an analog-to-digital converter.
Using the converter, the microcontroller continuously reads the voltage applied to
the digital potentiometer and updates the potentiometer resistance according to a
pre-programmed model of a voltage- or current-controlled memristive system. The
operation of electronic neurons is realized along similar lines [44]. Operation of the
associative memory is presented in Fig. 7 where a detail of this process is given.
Our work as described above [44] demonstrates the potential of memristive de-
vices for neuromorphic circuits applications. Importantly, it has been recently shown
in numerous experiments that memristive devices can be built at the nanoscale [52,
53, 54, 55, 56, 57, 14, 58, 59, 18, 60, 6, 61, 62]. This opens up the possibility to
Fig. 7 Experimental demonstration of the associative memory with memristive neural networks.
In this experiment, a simple neural network shown in Fig. 6 was realized. The first "probing"
phase demonstrates that, initially, only a signal from N1 neuron activates the output neuron. The
association of the Input 2 signal with the Output develops in the "learning phase" when N1 and
N2 neurons are simultaneously activated. In this case, a signal at the Input 1 excites the third
neuron that sends back-propagating pulses of a negative amplitude. These pulses, when applied
simultaneously with forward propagating pulses from the Input 2 to the second memristive synapse
S2 cause it to learn. The final "probing" phase demonstrates that signals from both N1 and N2
activate the output neuron. A detalied description of the experiment is given in Ref. [44]. Reprinted
from Ref. [44]. c(cid:13)2010 with permission from Elsevier.
2 Several designs of memristor [46, 47, 48, 7] as well as memcapacitor and meminductor [47, 49,
50, 51] emulators are known in the literature. These emulators serve as an important practical tool
to build small-scale circuits with memory circuit elements.
03691215-202468101214ProbingOutput"salivation"Input 2"sound"LearningProbingInput 1"sight of food"Voltage (V)Time (s) Biologically-Inspired Electronics with Memory Circuit Elements
13
fabricate neuromorphic circuits with the same amount of memristive synapses as
the number of biological synapses in the human brain (∼ 1014). In fact, one of the
main challenges for practical realizations of an artificial brain on a chip is related to
the high connectivity of biological neurons. It has been estimated that, on average,
the number of connections per neuron is of the order of 103. Therefore, neural net-
works of memelements built at the nanoscale offer advantages - in terms of density
- unavailable with current active elements (such as transistors).
4.2.2 Spike timing-dependent plasticity
However, the above mentioned simple Hebbian rule does not describe the much
more complicated time-dependent plasticity of biological synapses [63, 64, 65, 66].
The latter has come to be known as spike timing-dependent plasticity (STDP). When
a post-synaptic signal reaches the synapse before the action potential of the pre-
synaptic neuron, the synapse shows long-term depression (LTD), namely its strength
decreases (smaller connection between the neurons) depending on the time differ-
ence between the post-synaptic and the pre-synaptic signals. Conversely, when the
post-synaptic action potential reaches the synapse after the pre-synaptic action po-
tential, the synapse undergoes a long-time potentiation (LTP), namely the signal
transmission between the two neurons increases in proportion to the time difference
between the pre-synaptic and the post-synaptic signals. The learning process and the
storing of information in the brain thus follow non-trivial time-dependent features
which have not been fully understood yet. Implementation of STDP in artificial net-
works can thus help unraveling these mechanisms.
The spike timing-dependent plasticity can be implemented using different types
of memristive systems. Following our previous work [67], neuromorphic circuits
can be based on memristive systems with or without an internal spike-timing track-
ing capability. In the most simple case, memristive systems without spike-timing
tracking capability are of the first order, while those supporting such capability are
of the second order as an additional internal state variable is needed to track timing
of pre-synaptic and post-synaptic pulses [67]. In the first case, an additional exter-
nal hardware is required to implement the spike timing-dependent plasticity. For
example, STDP was recently realized using a combination of memristive systems
with CMOS (complementary metal-oxide-semiconductor) elements [60] (see Fig.
8). Another approach involves utilization of overlapping pulses of opposite polari-
ties [67, 68, 69].
A simple second-order memristive system with time-tracking capability is de-
scribed by the following equations [67]
R = x,
x = γ [θ (VM −Vt )θ (y− yt ) + θ (−VM −Vt )θ (−y− yt )]y,
y =
τ [−VMθ (VM −Vt )θ (yt − y)−VMθ (−VM −Vt )θ (y + yt )− y] ,
1
(9)
(10)
(11)
14
Massimiliano Di Ventra and Yuriy V. Pershin
Fig. 8 (a) Measured change
in the synaptic weight versus
spike separation. Inset: SEM
image of the memristive
crossbar array, scale bar is
300 nm. (b) Measured change
in excitatory postsynaptic
current of rat neurons after
repetitive correlated spiking
versus relative spiking timing
(the plot was reconstructed
from Ref. [65]). Inset: image
of a hippocampal neuron
(the image was adapted with
permission from reference
[70]). Scale bar is 50 µm.
Reprinted with permission
from [60]. Copyright 2010
American Chemical Society.
where x and y are internal state variables, γ is a constant, Vt is a threshold voltage,
yt is the threshold value of y, and τ is a constant defining the time window of STDP.
The second-order memristive system with timing tracking capability defined by the
above equations is very promising for neuromorphic circuits application since neu-
ron's firing can be implemented simply by short single rectangular pulses and no ad-
ditional hardware as in the case of first-order memristors (see, e.g., Ref. [60]). How-
ever, such solid-state second-order memristors need to be developed, even though
their implementation in memristor emulators [46, 47] can be easily realized.
Moreover, several authors have discussed applications of three-terminal transistor-
like electronic devices with memory [25, 26, 71] in the area of neuromorphic com-
puting. Although, formally such devices can not be categorized as memristive sys-
tems, their operation is clearly based on memristive effects. In particular, Lai et
al. [26] have experimentally fabricated a synaptic transistor. For instance, Fig. 9
depicts their experimental scheme and selected measurement results that confirm
realization of spike timing-dependent plasticity in this device.
Biologically-Inspired Electronics with Memory Circuit Elements
15
(a)
(b)
Fig. 9 (a) Structure of synaptic transistor . (b) The relative changes of the postsynaptic currents
measured after application of 120 pairs of temporally correlated pre- and post-synaptic spikes.
From [26], copyright Wiley-VCH Verlag GmbH & Co. KGaA. Reproduced with permission.
4.3 Networks of memory circuit elements
A human brain - and also the brain of many other living organisms - solves many
problems much better than traditional computers. The reason for this is a type of
massive parallelism in which a large number of neurons and synapses participate
simultaneously in the calculation. Here, we consider networks of memory circuit
elements and their ability to i) solve efficiently certain graph theory optimization
problems, and ii) retain such information for later use. In particular, we demonstrate
that a network of memristive devices solves the maze problem much faster than any
existing algorithm [72]. Similar to the brain's operation, such an extraordinary ad-
vance in computational power is due to the massively-parallel network dynamics in
which all network components are simultaneously involved in the calculation. This
type of parallelism could be dubbed as an analog parallelism which is very differ-
ent from that used in conventional supercomputers. In the latter systems, each core
typically runs a separate process that, relatively rarely, exchanges information with
other cores. In calculations done by networks, the information exchange is contin-
uous resulting in a tremendous increase of computational power as we demonstrate
below.
Left panel of Fig. 10 depicts a memristive network (memristive processor) in
which points of a square grid are connected by basic units (memristive system plus
switch (FET)) [72]. Each switch in the network can be in the "connected" or "not-
connected" state. Since the direction of current flow in the network is not known a
priori, the polarity of adjacent memristive devices (indicated by the black thick line
in the memristor symbol in Figure 1) is chosen to be alternating. Experimentally, the
suggested network could be fabricated using, e.g., CMOL (Cmos+MOLecular-scale
devices) architecture [73] combining a single memristor layer with a conventional
CMOS layer. The operation of the massively-parallel processor consists of three
16
Massimiliano Di Ventra and Yuriy V. Pershin
Fig. 10 Maze mapping into a network of memristors. Right panel. The maze is covered by an array
of vertical and horizontal lines having the periodicity of the maze. Left panel. Architecture of the
network of memristors in which each crossing between vertical and horizontal lines in the array (in
the right panel) is represented by a grid point to which several basic units consisting of memristors
and switches (field-effect transistors) are linked in series. The maze topology is encoded into the
state of the switches such that if the short line segment connecting neighboring crossing points
in the array crosses the maze wall then the state of the corresponding switch is "not connected"
(shown with red symbols). All other switches are in the "connected" state. The external voltage (V )
is applied across the connection points corresponding to the entrance (V ) and exit (ground, GND)
points of the maze. Reprinted with permission from Ref. [72]. c(cid:13)2011 American Physical Society.
main stages: initialization, computation and reading out the computation result. All
these stages are realized by externally applied signals (originating, e.g., from the
CMOS layer).
During the first initialization stage, all memristive elements in the network are
switched into the "OFF" state. This can be done, for example, by applying GND
and appropriately selected V1 voltages in a chessboard-like pattern to all grid points
of the memristive network for a sufficiently long period of time [72]. After that,
the maze topology is mapped onto the memristive network by setting appropriate
switches into the "not connected" state. We describe this process in the caption
of Fig. 10. The computation stage consists in the application of a single voltage
pulse of appropriate amplitude and duration across grid points corresponding to the
entrance and exit points of the maze. The solution can be later read or used in further
calculations.
We have modeled the memristive processor operation by numerically solving
Kirchhoff's current law equations complemented by Eqs. (5), (6) which in the
present network case are modified as
Biologically-Inspired Electronics with Memory Circuit Elements
17
Fig. 11 Solution of a multiple-path maze [72]. The maze solution contains two common segments
(red dots connected by a red line), and two alternative segments of different lengths close to the
left bottom corner. The memristance in the shorter segment (blue dots connected by a red line)
is smaller than that in the longer segment (green dots connected by a red line) since the current
through the shorter segment is larger and, consequently, the change of the memristors' state along
this segment is larger. The arrow at the bottom indicates a splitting point of the solution path.
The resistance is in Ohms, the voltage is in Volts and the current is in Amperes. Reprinted with
permission from Ref. [72]. c(cid:13)2011 American Physical Society.
i j = RONxi j + ROFF (1− xi j) ,
RM
(12)
where RON and ROFF are again the minimal and maximal values of memristance, xi j
is the dimensionless internal state variable for each memristor bound to the region
0 ≤ xi j ≤ 1, and (i, j) are grid indexes of a memristor to identify its location in the
network. The dynamics of xi j is then given by
dxi j
dt = αIi j(t),
(13)
with α a constant and Ii j(t) the current flowing through the memristor (i j).
Fig. 11(a) shows a solution of a multiple-path maze. The maze solution is clearly
seen in Fig. 11 as chains of red, blue and green boxes (representing memristive
devices with lower memristance) connected by a red line. Importantly, the memris-
tive processor not only determines all possible solutions of the maze but also stores
them and sorts them out according to their length. This feature is described in more
details in the caption of Fig. 11. Also, the memristive processor requires only one
single step to find the maze solution thus outperforming all known maze solving
approaches and algorithms.
18
Massimiliano Di Ventra and Yuriy V. Pershin
We also note that the wide selection of physical mechanisms of memory we
can "shop" from, offers many opportunities to design novel efficient electronic de-
vices [6]. For example, a memristive processor based on fast switching nanoionic
metal/insulator/metal cells [6] would require just few nanoseconds or even less 3
to solve the maze. More generally, a network of memristors - or other memory cir-
cuit elements - can be considered as an adaptable medium whose state dynamically
changes in response to time-dependent signals or changes in the network config-
uration. Therefore, the use of these processors is not limited to maze solving: We
expect they could help find the solution of many graph theory optimization prob-
lems including, for instance, the traveling salesman problem, shortest path problem,
etc.
5 Conclusions and Outlook
In conclusion, we have shown that the two-terminal electronic devices with memory
-- memristive, memcapacitive and meminductive systems -- are very useful to model
a variety of biological processes and systems. The electronic implementation of
all these mechanisms can clearly lead to a novel generation of "smart" electronic
circuits that can find useful applications in diverse areas of science and technology.
In addition, these memelements and their networks, provide solid ground to test
various hypothesis and ideas regarding the functioning of the human (and animal)
brain both theoretically and experimentally. Theoretically because their flexibility in
terms of what type and how many internal state variables responsible for memory,
or what network topology are required to reproduce certain biological functions can
lead to a better understanding of the microscopic mechanisms that are responsible
for such features in living organisms. Experimentally because with the continuing
miniaturization of electronic devices, memelements can be assembled into networks
with similar densities as the biological systems (e.g., the brain) they are designed to
emulate. In particular, we anticipate potential applications for memcapacitive and
meminductive systems [4] which offer such an important property as low energy
dissipation combined with information storage capabilities. We are thus confident
that the area of biologically-inspired electronics with memory circuit elements will
offer many research opportunities in several fields of science and technology.
Acknowledgements M.D. acknowledges partial support from the NSF Grant No. DMR-0802830.
3 Fast sub-nanosecond switching has been recently reported in tantalum oxide memristive sys-
tems [62].
Biologically-Inspired Electronics with Memory Circuit Elements
References
19
1. S.J. Russell, P. Norvig, Artificial Intelligence: A Modern Approach, 3rd edn. (Prentice Hall,
2009)
2. A. Colorni, M. M. Dorigo, V. Maniezzo, in Actes de la premire confrence europenne sur la vie
artificielle, Paris, France, Elsevier Publishing (1991), pp. 134 -- 142
3. N. Monmarche, F. Guinand, P. Siarry, Artificial Ants (Wiley-ISTE, 2010)
4. M. Di Ventra, Y.V. Pershin, L.O. Chua, Proc. IEEE 97(10), 1717 (2009)
5. M. Di Ventra, Y.V. Pershin, Mat. Today 14, 584 (2011)
6. Y.V. Pershin, M. Di Ventra, Advances in Physics 60, 145 (2011)
7. L.O. Chua, IEEE Trans. Circuit Theory 18, 507 (1971)
8. L.O. Chua, S.M. Kang, Proc. IEEE 64, 209 (1976)
9. Y.V. Pershin, M. Di Ventra, Phys. Rev. B 78, 113309 (2008)
10. X. Wang, Y. Chen, H. Xi, H. Li, D. Dimitrov, El. Dev. Lett. 30, 294 (2009)
11. J. Martinez-Rincon, Y.V. Pershin, IEEE Trans. El. Dev. 58, 1809 (2011)
12. J. Martinez-Rincon, M. Di Ventra, Y.V. Pershin, Phys. Rev. B 81, 195430 (2010)
13. F. Corinto, A. Ascoli, M. Gilli, submitted for publication (2011)
14. D.B. Strukov, G.S. Snider, D.R. Stewart, R.S. Williams, Nature 453, 80 (2008)
15. Y.V. Pershin, S. La Fontaine, M. Di Ventra, Phys. Rev. E 80, 021926 (2009)
16. M. Krems, Y.V. Pershin, M. Di Ventra, Nano Lett. 10, 2674 (2010)
17. A. Stotland, M. Di Ventra, arXiv:1104.4485v2 (2011)
18. C. Schindler, G. Staikov, R. Waser, Appl. Phys. Lett. 94, 072109 (2009)
19. Q. Lai, L. Zhang, Z. Li, W.F. Stickle, R.S. Williams, Y. Chen, Appl. Phys. Lett. 95, 213503
(2009)
Tech. 49(11), 2093 (2001)
20. V. Lubecke, B. Barber, E. Chan, D. Lopez, M. Gross, P. Gammel, IEEE Trans. Microw. Theory
21. I. Zine-El-Abidine, M. Okoniewski, J.G. McRory, in Proceedings of the 2004 International
Conference on MEMS, NANO and Smart Systems (ICMENS'04) (2004), pp. 636 -- 638
22. S. Chang, S. Sivoththaman, IEEE El. Dev. Lett. 27(11), 905 (2006)
23. S. Liu, N. Wu, A. Ignatiev, J. Li, J. Appl. Phys. 100, 056101 (2006)
24. V.V. Erokhin, T.S. Berzina, M.P. Fontana, Crystallogr. Rep. 52, 159 (2007)
25. F. Alibart, S. Pleutin, D. Guerin, C. Novembre, S. Lenfant, K. Lmimouni, C. Gamrat, D. Vuil-
laume, Adv. Funct. Mat. 20, 330 (2010)
26. Q. Lai, L. Zhang, Z. Li, W.F. Stickle, R.S. Williams, Y. Chen, Adv. Mat. 22, 2448 (2010)
27. T. Kozlowski, S. Pallardy, Botanical Review 68, 270 (2002)
28. A. Trewavas, Annals of Botany 92, 1 (1998)
29. P. Calvo Garzon, F. Keijzer, Adaptive Behavior 19, 155 (2011)
30. J.E.R. Staddon, Adaptive behavior and learning (Cambridge University Press, 1983)
31. J. Ojal, Biol. Cybern. 79, 403 (1998)
32. T. Saigusa, A. Tero, T. Nakagaki, Y. Kuramoto, Phys. Rev. Lett. 100, 018101 (2008)
33. Q. Li, B. McNeil, L.M. Harvey, Free Radical Biology and Medicine 44, 394 (2008)
34. J. Hartley, J.W.G. Cairney, A.A. Meharg, Plant and Soil 189, 303 (1997)
35. H. Thieringer, P. Jones, M. Inouye, BioEssays 20, 49 (1998)
36. J. van der Oost, M.M. Jore, E.R. Westra, M. Lundgren, S.J.J. Brouns, Trends in Biochemical
Sciences 34, 401 (2009)
37. J. Larsson, J.A.A. Nylander, B. Bergman, BMC evolutionary biology 11, 187 (2011)
38. J.H. Holland, Adaptation in Natural and Artificial Systems: An Introductory Analysis with
Applications to Biology, Control, and Artificial Intelligence (Bradford Book, 1992)
39. T. Driscoll, J. Quinn, S. Klein, H.T. Kim, B.J. Kim, Y.V. Pershin, M. Di Ventra, D.N. Basov,
Appl. Phys. Lett. 97, 093502 (2010)
40. T. Driscoll, H.T. Kim, B.G. Chae, M. Di Ventra, D.N. Basov, Appl. Phys. Lett. 95, 043503
(2009)
41. T. Driscoll, H.T. Kim, B.G. Chae, B.J. Kim, Y.W. Lee, N.M. Jokerst, S. Palit, D.R. Smith,
M. Di Ventra, D.N. Basov, Science 325, 1518 (2009)
20
Massimiliano Di Ventra and Yuriy V. Pershin
42. G.K. Johnsen, C.A. Lutken, O.G. Martinsen, S. Grimnes, Phys. Rev. E 83, 031916 (2011)
43. I. Pavlov, Conditioned Reflexes: An Investigation of the Physiological Activity of the Cerebral
Cortex (London: Oxford University Press, 1927). (translated by G. V. Anrep)
44. Y.V. Pershin, M. Di Ventra, Neural Networks 23, 881 (2010)
45. D.O. Hebb, The Organization of Behavior; A Neuropsychological Theory (New York: Wiley,
1949)
46. Y.V. Pershin, M. Di Ventra, IEEE Trans. Circ. Syst. I 57, 1857 (2010)
47. D. Biolek, V. Biolkova, Int. J. Numer. Mod. (in press) (2011)
48. D. Biolek, V. Biolkova, Z. Kolka, in 5th International Conference on Circuits, Systems and
Signals (CSS'11). (2011), p. 171
49. D. Biolek, V. Biolkova, El. Lett. 46, 1428 (2010)
50. Y.V. Pershin, M. Di Ventra, El. Lett. 47, 243 (2011)
51. Y.V. Pershin, M. Di Ventra, El. Lett. 46, 517 (2010)
52. T. Tamura, T. Hasegawa, K. Terabe, T. Nakayama, T. Sakamoto, H. Sunamura, H. Kawaura,
S. Hosaka, M. Aono, Jpn. J. Appl. Phys. 45, L364 (2006)
53. R.R. Waser, M. Aono, Nat. Mat. 6, 833 (2007)
54. D. Lee, D. jun Seong, I. Jo, F. Xiang, R. Dong, S. Oh, , H. Hwang, Appl. Phys. Lett. 90,
122104 (2007)
55. S. Dietrich, M. Angerbauer, M. Ivanov, D. Gogl, H. Hoenigschmid, M. Kund, C. Liaw,
M. Markert, IEEE J. Sol.-State Circ. 42, 839 (2007)
56. S. Raoux, G.W. Burr, M.J. Breitwisch, C.T. Rettner, Y.C. Chen, R.M. Shelby, M. Salinga,
D. Krebs, S.H. Chen, H.L. Lung, C.H. Lam, IBM J. Res. Dev. 52, 465 (2008)
57. I.H. Inoue, S. Yasuda, H. Akinaga, H. Takagi, Phys. Rev. B 77, 035105 (2008)
58. A. Sawa, Mat. Today 11, 28 (2008)
59. S.H. Jo, K.H. Kim, W. Lu, Nano Lett. 9, 870 (2009)
60. S.H. Jo, T. Chang, I. Ebong, B.B. Bhadviya, P. Mazumder, W. Lu, Nano Lett. 10, 1297 (2010)
61. M.J. Lee, C.B. Lee, D. Lee, S.R. Lee, M. Chang, J.H. Hur, Y.B. Kim, C.J. Kim, D.H. Seo,
S. Seo, U.I. Chung, I.K. Yoo, K. Kim, Nature Materials 10(8), 625 (2011)
62. A.C. Torrezan, J.P. Strachan, G. Medeiros-Ribeiro, R.S. Williams, Nanotechnology 22(48),
485203 (2011)
63. W.B. Levy, O. Steward, Neuroscience 8, 791 (1983)
64. H. Markram, J. Lubke, M. Frotscher, B. Sakmann, Science 275, 213 (1997)
65. G.Q. Bi, M.M. Poo, J. Neurosci. 18, 10464 (1998)
66. R.C. Froemke, Y. Dan, Nature 416, 433 (2002)
67. Y.V. Pershin, M. Di Ventra, Proc. IEEE (in press); arXive:1009.6025 (2010)
68. G.S. Snider, SciDAC Review 10, 58 (2008)
69. S. Parkin, Innately three dimensional spintronic memory and logic devices: Racetrack memory
and spin synapses (2010). Talk at 2010 MRS SprinG Meeting
70. S. Kaech, G. Banker, Nat Protoc. 1, 2406 (2006)
71. W.S. Zhao, G. Agnus, V. Derycke, A. Filoramo, J.P. Bourgoin, C. Gamrat, Nanotechnology
21(17), 175202 (2010)
72. Y.V. Pershin, M. Di Ventra, Phys. Rev. E 84, 046703 (2011)
73. K.K. Likharev, D.B. Strukov, in Introducing Molecular Electronics, vol. 657, ed. by G.F.
G. Cuniberti, K. Richter (Springer, 2005), pp. 447 -- 477
|
1305.4160 | 1 | 1305 | 2013-05-17T19:17:38 | A generative spike train model with time-structured higher order correlations | [
"q-bio.NC",
"math.DS"
] | Emerging technologies are revealing the spiking activity in ever larger neural ensembles. Frequently, this spiking is far from independent, with correlations in the spike times of different cells. Understanding how such correlations impact the dynamics and function of neural ensembles remains an important open problem. Here we describe a new, generative model for correlated spike trains that can exhibit many of the features observed in data. Extending prior work in mathematical finance, this generalized thinning and shift (GTaS) model creates marginally Poisson spike trains with diverse temporal correlation structures. We give several examples which highlight the model's flexibility and utility. For instance, we use it to examine how a neural network responds to highly structured patterns of inputs. We then show that the GTaS model is analytically tractable, and derive cumulant densities of all orders in terms of model parameters. The GTaS framework can therefore be an important tool in the experimental and theoretical exploration of neural dynamics. | q-bio.NC | q-bio |
A generative spike train model with time-structured higher order
correlations
James Trousdale, Yu Hu, Eric Shea-Brown and Kresimir Josi´c
Abstract
Emerging technologies are revealing the spiking activity in ever larger neural ensembles. Frequently,
this spiking is far from independent, with correlations in the spike times of different cells. Under-
standing how such correlations impact the dynamics and function of neural ensembles remains an
important open problem. Here we describe a new, generative model for correlated spike trains
that can exhibit many of the features observed in data. Extending prior work in mathematical
finance, this generalized thinning and shift (GTaS) model creates marginally Poisson spike trains
with diverse temporal correlation structures. We give several examples which highlight the model's
flexibility and utility. For instance, we use it to examine how a neural network responds to highly
structured patterns of inputs. We then show that the GTaS model is analytically tractable, and
derive cumulant densities of all orders in terms of model parameters. The GTaS framework can
therefore be an important tool in the experimental and theoretical exploration of neural dynamics.
1
Introduction
Recordings across the brain suggest that neural populations spike collectively -- the statistics of
their activity as a group are distinct from that expected in assembling the spikes from one cell
at a time [6, 48, 64, 62, 67, 8, 31, 32, 10, 24, 57]. Advances in electrode and imaging technology
allow us to explore the dynamics of neural populations by simultaneously recording the activity of
hundreds of cells. This is revealing patterns of collective spiking that extend across multiple cells.
The underlying structure is intriguing: For example, higher-order interactions among cell groups
have been observed widely [67, 68, 4, 56, 48, 64, 76, 24]. A number of recent studies point to
mechanisms that generate such higher-order correlations from common input processes, including
unobserved neurons. This suggests that, in a given recording or given set of neurons projecting
downstream, higher-order correlations may be quite ubiquitous [51, 79, 44, 9]. Moreover, these
higher-order correlations may impact encoded information [55, 17, 24] as well as the firing rate of
downstream neurons [46].
What exactly is the impact of such collective spiking on the encoding and transmission of
information in the brain? This question has been studied extensively, but much remains unknown.
Results to date show that the answers will be varied and rich. Patterned spiking can impact
responses at the level of single cells [78, 62, 46] and neural populations [59, 58, 74, 5]. Neurons with
even the simplest of nonlinearities can be highly sensitive to correlations in their inputs. Moreover,
such nonlinearities are sufficient to accurately decode signals from the input to correlated neural
populations [66].
1
An essential tool in understanding the impact of collective spiking is the ability to generate
artificial spike trains with a predetermined structure across cells and across time [29, 45, 50, 14].
Such synthetic spike trains are the grist for testing hypotheses about spatiotemporal patterns in
coding and dynamics. In experimental studies, such spike trains can be used to provide structured
stimulation of single cells across their dendritic trees via glutamate uncaging [23, 26, 12, 13], or
entire populations of neurons via optical stimulation of microbial opsins [19, 30]. Computationally,
they are used to examine the response of nonlinear models of downstream cells [62, 46, 18].
Therefore, much effort has been devoted to developing statistical models of population activity.
A number of flexible, yet tractable probabilistic models of joint neuronal activity have been pro-
posed. Pairwise correlations are the most common type of interactions obtained from multi-unit
recordings. Therefore many earlier models were designed to generate samples of neural activity
patterns with predetermined first and second order statistics [29, 45, 50, 14].
In these models,
higher-order correlations are not explicitly and separately controlled.
A number of different models have been used to analyze higher-order interactions. However,
most of these models assume that interactions between different cells are instantaneous (or near-
instantaneous) [70, 46, 40]. A notable exception is the work of [11], which developed such methods
for use in financial applications. In these previous efforts, correlations at all orders were character-
ized by the increase, or decrease, in the probability that groups of cells spike together at the same
time, or have a common temporal correlation structure regardless of the group.
The aim of the present work is to provide a statistical method for generating spike trains with
more general correlation structures across cells and time. Specifically, we allow distinct temporal
structure for correlations at pairwise, triplet, and all higher orders, and do so separately for different
groups of cells in the neural population. Our aim to describe a model that can be applied in
neuroscience, and can potentially be fit to emerging datasets.
A sample processes from our model is shown in Fig. 1. The multivariate spike train consists of
six marginally Poisson processes. Each event was either uncorrelated with all other events across the
population, or correlated in time with an event in all other spike trains. This model was configured
to exhibit activity that cascades through a sequence of neurons. Specifically, neurons with larger
index tend to fire later in a population wide event (this is similar to a synfire chain [2], but with
variable timing of spikes within the cascade). In Fig. 1B, we plot the "population cross-cumulant
density" for three chosen neurons -- the summed activity of the population triggered by a spike in
a chosen cell. The center of mass of this function measures the average latency by which spikes of
the neuron in question precede those of the rest of the population [48]. Finally, Fig. 1C shows the
third-order cross-cumulant density for the three neurons. The triangular support of this function
is a reflection of a synfire-like cascade structure of the spiking shown in the raster plot of panel A:
when firing events are correlated between trains, they tend to proceed in order of increasing index.
We demonstrate the impact of such structured activity on a downstream network in Section 2.2.
2 Results
Our aim is to describe a flexible multivariate point process capable of generating a range of high
order correlation structures. To do so we extend the TaS (thinning and shift) model of temporally-
and spatially-correlated, marginally Poisson counting processes [11]. The TaS model itself gener-
alizes the SIP and MIP models [46] which have been used in theoretical neuroscience [17, 59, 73].
However the TaS model has not been used as widely. The original TaS model is too rigid to
generate a number of interesting activity patterns observed in multi-unit recordings [37, 49, 48].
2
Figure 1: (A) Raster plot of event times for an example multivariate Poisson process X =
(X1, . . . , X6) generated using the methods presented below. This model exhibits independent
marginal events (blue) and population-level, chain-like events (red). (B) Some second order pop-
ulation cumulant densities (i.e., second order correlation between individual unit activities and
population activity) for this model [48]. Greater mass to the right (resp.
left) of τ = 0 indi-
follow) in pairwise-correlated events. (C) Third-order
cates that the cell tends to lead (resp.
cross-cumulant density for processes X1, X2, X3. The quantity κX
123(τ1, τ2) yields the probability of
observing spikes in cells 2 and 3 at an offset τ1, τ2 from a spike in cell 1, respectively, in excess of
what would be predicted from the first and second order cumulant structure. All quantities are
precisely defined in the Methods. Note: system parameters necessary to reproduce results are given
in the Appendix for all figures.
We therefore developed the generalized thinning and shift model (GTaS) which allows for a more
diverse temporal correlation structure.
We begin by describing the algorithm for sampling from the GTaS model. This constructive
approach provides an intuitive understanding of the model's properties. We then present a pair
of examples, the first of which highlights the utility of the GTaS framework. The second example
demonstrates how sample point processes from the TaS models can be used to study population
dynamics. Next, we present the analysis which yields the explicit forms for the cross-cumulant
densities derived in the context of the examples. We do so by first establishing a useful distribu-
tional representation for the GTaS process, paralleling [11]. Using this representation, we derive
cross-cumulants of a GTaS counting process, as well as explicit expressions for the cross-cumulant
3
densities. After explaining the derivation at lower orders, we present a theorem which describes
cross-cumulant densities at all orders.
2.1 GTaS model simulation
D
The GTaS model is parameterized first by a rate λ which determines the intensity of a "mother
process" - a Poisson process on R. The events of the mother process are marked, and the markings
determine how each event is distributed among a collection of N daughter processes. The daughter
processes are indexed by the set D = {1, . . . , N}, and the set of possible markings is the power
, the set of all subsets D. We define a probability distribution p = (pD)D⊂D, assigning a
set 2
probability to each possible marking, D. As we will see, pD determines the probability of a joint
event in all daughter processes with indices in the set D. Finally, to each marking, D, we assign a
probability distribution QD, giving a family of shift (jitter) distributions (QD)D⊂D. Each (QD) is
a distribution over RN .
The rate λ, the distribution p over the markings, and the family of jitter distributions (QD)D⊂D,
define a vector X = (X1, . . . , XN ) of dependent daughter Poisson processes described by the fol-
lowing algorithm, which yields a single realization (see Fig. 2):
1. Simulate the mother Poisson process of rate λ on R, generating a sequence of event times
{tj}. (Fig. 2A)
2. With probability pDj assign the subset Dj ⊂ D to the event of the mother process at time tj.
This event will be assigned only to processes with indices in Dj. (Fig. 2B)
3. Generate a vector (Y j
1 , . . . , Y j
N ) = Yj from the distribution QDj . For each i ∈ D, the time
tj + Y j
i
is set as an event time for the marginal counting process Xi. (Fig. 2C)
Hence copies of each point of the mother process are placed into daughter processes after a shift
in time. A primary difference between the GTaS model and the TaS model presented in [11] is the
dependence of the shift distributions QD on the chosen marking. This allows for greater flexibility
in setting the temporal cumulant structure.
2.2 Examples
Relation to SIP/MIP processes Two simple models of correlated, jointly Poisson processes
were defined in [46]. The resulting spike trains exhibit spatial correlations, but only instanta-
neous temporal dependencies. Each model was constructed by starting with independent Poisson
processes, and applying one of two elementary point process operations: superposition and thin-
ning [21]. We show that both models are special cases of the GTaS model.
In the single interaction process (SIP), each marginal process Xi is obtained by merging an
independent Poisson process with a common, global Poisson process. That is,
tively. An SIP model is equivalent to a GTaS model with mother process rate λ = λc +(cid:80)N
where Zc and each Zi are independent Poisson counting processes on R with rates λc, λi, respec-
i=1 λi,
i = 1, . . . , N,
Xi(·) = Zi(·) + Zc(·),
and marking probabilities
λi
λ D = {i}
λ D = D
0
otherwise
λc
.
pD =
4
Figure 2: An illustration of a GTaS simulation. (A) Step 1: Simulate the mother process - a
time-homogeneous Poisson process with event times {tj}. (B) Step 2: For each tj in step 1, select
a set Dj ⊂ D according to the distribution pD, and project the event at time tj to the subsets with
indices in Dj. The legend indicates the colors assigned to three possible markings in this example.
(C) Step 3: For each pair (tj, Dj) generated in the previous two steps, draw Yj from QDj , and
shift the event times in the daughter processes by the corresponding values Y j
i .
coordinate (i.e., qD(y1, . . . , yN ) ≡(cid:81)N
Note that if λc = 0, each spike will be assigned to a different process Xi, resulting in N independent
Poisson processes. Lastly, each shift distribution is equal to a delta distribution at zero in every
i=1 δ(yi) for every D ⊂ D). Thus, all joint cumulants (among
distinct marginal processes) of orders 2 through d are delta functions of equal magnitude, λpD.
The multiple interaction process (MIP) consists of N Poisson processes obtained from a common
mother process with rate λm by thinning [21]. The ith daughter process is formed by independent
(across coordinates and events) deletion of events from the mother process with probability p =
(1 − ). Hence, an event is common to k daughter processes with probability k. Therefore, if we
take the perspective of retaining, rather than deleting events, the MIP model is equivalent to a
GTaS process with λ = λm, and pD = D(1− )d−D. As in the SIP case, the shift distributions are
singular in every coordinate. Below, we present a general result (Theorem 1) which immediately
yields as a corollary that the MIP model has cross-cumulant functions which are δ functions in all
dimensions, scaled by k, where k is the order of the cross-cumulant.
Generation of synfire-like cascade activity The GTaS framework provides a simple, tractable
way of generating cascading activity where cells fire in a preferred order across the population -- as
in a synfire chain, but (in general) with variable timing of spikes [2, 37, 3, 1, 7]. More generally, it
can be used to simulate the activity of cell assemblies [32, 34, 16, 10], in which the firing of groups
of neurons is likely to follow a particular order.
In the Introduction, we briefly presented one example in which the GTaS framework was used to
generate synfire-like cascade activity (see Fig. 1), and we present another in Fig. 3. In what follows,
we will present the explicit definition of this second model, and then derive explicit expressions for
its cumulant structure. Our aim is to illustrate the diverse range of possible correlation structures
that can be generated using the GTaS model.
Consider an N -dimensional counting process X = (X1, . . . , XN ) of GTaS type, where N ≥ 4.
We restrict the marking distribution so that pD ≡ 0 unless D ≤ 2 or D = D. That is, events
are either assigned to a single, a pair, or all daughter processes. For sets D with D = 2, we set
5
QD ∼ N (0, Σ) - a Gaussian distributions of zero mean and some specified covariance. The choice
of the precise pairwise shift distributions is not important. Shifts of events attributed to a single
process have no effect on the statistics of the multivariate process. (To see this, note that the
integrals with respect to t in Eq. (2) below, for example, may be viewed as a marginalization over
shifts applied to events in the first process.)
It remains to define the jitter distribution for events common to the entire population of daughter
processes, i.e. events marked by D. We will show that we can generate cascading activity, and
analytically describe the resulting correlation structure. We generate random vectors Y ∼ QD
according to the following rule, for each i = 1, . . . , N :
1. Generate independent random variables ϕi ∼ Exp(αi) where αi > 0.
2. Set Yi =(cid:80)i
j=1 ϕj.
In particular, note that these shift times satisfy YN ≥ . . . ≥ Y2 ≥ Y1 ≥ 0, indicating the chain-like
structure of these joint events.
From the definition of the model and our general result (Theorem 1) below, we immediately
have that κX
ij (τ ), the second order cross-cumulant density for the process (i, j), is given by
where
c2
ij(τ ) = λp{i,j}
(cid:90)
κX
ij (τ ) = c2
ij(τ ) + cN
ij (τ ),
{i,j}
{i,j}(t, t + τ )dt,
q
cN
ij (τ ) = λpD
(cid:90)
{i,j}
D
q
(t, t + τ )dt
(1)
(2)
define the contributions to the second order cross-cumulant density by the second-order, Gaussian-
jittered events and the population-level events, respectively. The functions qD(cid:48)
D indicate the densities
associated with the distribution QD, projected to the dimensions of D(cid:48). All statistical quantities
are precisely defined in the methods.
By exploiting the hierarchical construction of the shift times, we can find an expression for the
joint density qD, necessary to explicitly evaluate Eq. (1). For a general N -dimensional distribution,
f (y1, . . . , yN ) = f (yNy1, . . . , yN−1)f (yN−1y1, . . . , yN−2)··· f (y2y1)f (y1).
(3)
Since Y1 ∼ Exp(α1), we have f (y1) = exp [−α1y1] Θ(y1), where Θ(y) is the Heaviside step function.
Further, as Yi(Y1, . . . , Yi−1) ∼ Yi−1 + Exp(αi) for i ≥ 2, the conditional densities of the yi's take
the form
f (yiy1, . . . , yi−1) = f (yiyi−1) = αi exp [−αi(yi − yi−1)] Θ(yi − yi−1),
i ≥ 2.
Substituting this in to the identity Eq. (3), we have
(cid:40)
α1 exp [−α1y1](cid:81)N
qD(y1, . . . , yN ) =
0
i=2 αi exp [−αi(yi − yi−1)]
yN ≥ . . . ≥ y2 ≥ y1 ≥ 0
otherwise
.
(4)
Using Theorem 1 (Eq. (33)) we obtain the N th order cross-cumulant density (see the Methods),
κX
1···N (τ1, . . . , τN−1) = λpD
qD(t, t + τ1, . . . , t + τN−1)
= λpD ·
i=1 αi+1 exp [−αi+1(τi − τi−1)]
τi ≥ τi−1 i = 1, . . . , N − 1
otherwise
,
(5)
(cid:90)
(cid:40)(cid:81)N−1
0
6
Figure 3: An example of a six dimensional GTaS model exhibiting synfire-like cascading firing
patterns. (A) A raster-plot of spiking activity over a 100ms window. Blue spikes indicate either
marginal or pairwise events (i.e., events corresponding to markings for sets D ⊂ D with D ≤ 2.
Red spikes indicate population-wide events which have shift-times given by cumulative sums of
independent exponentials, as described in the text. Arrows indicate the location of the first spike
in the cascade. (B) A second-order cross-cumulant κX
13 (black line) of this model is composed of
contributions from two sources: correlations due to second-order markings, which have Gaussian
shifts (c2
13 -- dashed red line), and correlations due to the the occurrence of population wide events
(cN
13 -- dashed blue line). (C) Density plots of the third-order cross-cumulant density for triplets i)
(1, 2, 3) and ii) (1, 2, 4) -- the latter is given explicitly in Eq. (6). System parameters are given in
the Appendix.
where, for notational convenience, we define τ0 = 0. A raster plot of a realization of this model
is shown in Fig. 3A. We note that the cross-cumulant densities of arbitrary subcollections of the
counting processes X can be obtained by finding the appropriate marginalization of qD via inte-
gration of Eq. (4). In the case that common distributions are used to define the shifts, symbolic
calculation environments (i.e., Mathematica) can quickly yield explicit formulas for cross-cumulant
densities. Mathematica notebooks for Figure 1 available upon request.
As a particular example, we consider the cross-cumulant density of the marginal processes
X1, X3. Using Eqs. (2, 4), we find
13(τ ) = λpDΘ(τ ) ·
cN
(cid:40) α2α3
α3−α2
α2α3τ exp [−α2τ ]
{exp [−α2τ ] − exp [−α3τ ]} α2 (cid:54)= α3
α2 = α3
.
An expression for c2
for all i, j. In Fig. 3B, we plot these contributions, as well as the full covariance density.
13(τ ) may be obtained similarly using Eq. (2) and recalling that Q{i,j} ≡ N (0, Σ)
Similar calculations at third order yield, as an example,
124(τ1, τ2) = λpD ·
κX
α4−α3
α2α3α4(τ2 − τ1) exp [−α2τ1 − α3(τ2 − τ1)]
exp [−α2τ1]{exp [−α3(τ2 − τ1)] − exp [−α4(τ2 − τ1)]} α3 (cid:54)= α4
α3 = α4
, (6)
(cid:40) α2α3α4
7
124(τ1, τ2) is supported only on τ2 ≥ τ1 ≥ 0. Plots of the third-
where the cross-cumulant density κX
order cross-cumulants for triplets (1, 2, 3) and (1, 2, 4) in this model are shown in Fig. 3C. Note that,
for the specified parameters, the conditional distribution of Y4 -- the shift applied to the events
of X4 in a joint population event -- given Y2 follows a gamma distribution, whereas Y3Y2 follows
an exponential distribution, explaining the differences in the shapes of these two cross-cumulant
densities.
General cross-cumulant densities of at least third order for the cascading model will have a form
similar to that given in Eq. (6), and will contain no signature of the correlation of strictly second
order events. This highlights a key benefit of cumulants as a measure of dependence: although
they agree with central moments up to third order, we know from Eq. (23) below (or Eq. (22) in
the general case) that central moments necessarily exhibit a dependence on lower order statistics.
On the other hand, cumulants are "pure" and quantify only dependencies which cannot be inferred
from lower order statistics [27].
One useful statistic for analyzing population activity through correlations is the population
cumulant density [48]. The second order population cumulant density for cell i is defined by (see
the Methods)
κX
i,pop(τ ) =
κX
ij (τ ).
(cid:88)
j(cid:54)=i
(cid:88)
k(cid:54)=i,j
This function is linearly related to the spike-triggered average of the population activity conditioned
on that of cell i. In Fig. 4 we show three different second-order population-cumulant functions for
the cascading GTaS model of Fig. 3A. When the second order population cumulant for a neuron
is skewed to the right of τ = 0 (as is κX
1,pop -- blue line), a neuron tends to precede its partners in
pairwise spiking events. Similarly, skewness to the left of τ = 0 (κX
6,pop -- orange line) indicates a
neuron which tends to trail its partners in such events. A symmetric population indicates a neuron
is a follower and a leader. Taken together, these three second order population cumulants hint at
the chain structure of the process.
Greater understanding of the joint temporal statistics in a multivariate counting process can
be obtained by considering higher-order population cumulant densities. We define the third-order
population cumulant density for the pair (i, j) to be
κX
ij,pop(τ1, τ2) =
κX
ijk(τ1, τ2).
The third-order population cumulant density is linearly related to the spike-triggered population
activity, conditioned on spikes in cells i and j separated by a delay τ1. In Fig. 4B,C,D, we present
three distinct third-order population cumulant densities. Examining κX
12,pop(τ1, τ2) (panel B), we
see only contributions in the region τ2 > τ1 > 0, indicating that the pairwise event 1 → 2 often
precedes a third spike elsewhere in the population. The population cumulant κX
34,pop(τ1, τ2) has
contributions in two sections of the plane (panel C). Contributions in the region τ2 > τ1 > 0 can be
understood following the preceding example, while contributions in the region τ2 < 0 < τ1 imply
that the firing of other neurons tends to precede the joint firing event 1 → 2. Lastly, contributions
to κX
16,pop(τ1, τ2) (panel D) are limited to 0 < τ2 < τ1, indicating an above chance probability of
joint firing events of the form 1 → i → 6, where i indicates a distinct neuron within the population.
A distinct advantage of the study of population cumulant densities as opposed to individual
cross-cumulant functions in practical applications is related to data (i.e., sample size) limitations.
In many practical applications, where the temporal structure of a collection of observed point
processes is of interest, we often deal with a small, noisy samples. It may therefore be difficult to
8
estimate third- or higher-order cumulants. Population cumulants partially circumvent this issue by
pooling [59, 58, 74] (or summing) responses, to amplify existing correlations and average out the
noise in measurements.
Figure 4: Population cumulants for the synfire-like cascading GTaS process of Fig. 3. See Eq. (25)
for the definition of population cumulants. (A) Second order population cumulant densities for
processes 1,3 and 6. Greater mass to the right (resp. left) of τ = 0 indicates that a cell tends to lead
(resp.
follow) in pairwise-correlated events. (B) Third order population cumulant for processes
X1, X2 in the cascading GTaS process. Concentration of the mass in different regions of the plane
indicates temporal structure of events correlated between X1, X2 relative to the remainder of the
population (see the text). (C) Same as (B), but for processes X3, X4. (D) Same as (B), but for
processes X1, X6. System parameters are given in the Appendix.
We conclude this section by noting that even cascading GTaS examples can be much more
general. For instance, we can include more complex shift patterns, overlapping subassemblies
within the population, different temporal processions of the cascade, and more.
Timing-selective network The responses of single neurons and neuronal networks in experi-
mental [10, 69, 53] and theoretical studies [28, 36, 75, 39, 41] can reflect the temporal structure of
their inputs. Here, we present a simple example that shows how a network can be selective to fine
temporal features of its input, and how the GTaS model can be used to explore such examples.
9
As a general network model, we consider N leaky integrate-and-fire (LIF) neurons with mem-
brane potentials Vi obeying
= −Vi +
dVi
dt
N(cid:88)
j=1
wij(F ∗ zj)(t) + winxi(t),
i = 1, . . . , N.
(7)
When the membrane potential of cell i reaches a threshold V th, an output spike is recorded and
the membrane potential is reset to zero, after which evolution of Vi resumes the dynamics in
Eq. (7). Here wij is the synaptic weight of the connection from cell j to i, win is the input
weight, and we assume time to be measured in units of membrane time constants. The function
−1e−(t−τd)/τsynΘ(t − τd) is a delayed, unit-area exponential synaptic kernel with time-
F = τsyn
constant τsyn and delay τd. When the membrane potential of a cell reaches threshold, V th, a spike
is generated and the membrane potential is reset to zero. The output of the ith neuron is
where tj
i is the time of the jth spike of neuron i. In addition, the input {xi}N
i=1 is
(cid:88)
j
(cid:88)
j
zi(t) =
δ(t − tj
i ),
xi(t) =
δ(t − sj
i ),
i} correspond to those of a GTaS counting process X. Thus, each input
where the event times {sj
spike results in a jump in the membrane potential of the corresponding LIF neuron of amplitude win.
The particular network we consider will have a ring topology (nearest neighbor-only connectivity)
-- specifically, for i, j = 1, . . . , N, we let
(cid:40)
wij =
wsyn
0
i − j mod N ≡ 1 or N − 1
otherwise
.
We further assume that all neurons are excitatory, so that wsyn > 0.
A network of LIF neurons with synaptic delay is a minimal model which can exhibit fine-scale
discrimination of temporal patterns of inputs without precise tuning [38]. To exhibit this depen-
dence we generate inputs from two GTaS processes. The first (the cascading model ) was described
in the preceding example. To independently control the mean and variance of relative shifts we
replace the sum of exponential shifts with sums of gamma variates. We also consider a model
featuring population-level events without shifts (the synchronous model ), where the distribution
QD is a δ distribution at zero in all coordinates.
The only difference between the two input models is in the temporal structure of joint events.
In particular, the rates, and all long timescale spike count cross-cumulants (equivalent to the total
"area" under the cross-cumulant density, see the Methods) of order two and higher are identical for
the two processes. We focus on the sensitivity of the network to the temporal cumulant structure
of its inputs.
In Fig. 5A,B, we present two example rasters of the nearest-neighbor LIF network receiving
synchronous (left) and cascading (right) input. In the second case, there is an obvious pattern in
the outputs, but the firing rate is also increased. This is quantified in Fig. 5C, where we compare
the number of output spikes fired by a network receiving synchronous input (horizontal axis) with
the same for a network receiving cascading input (vertical axis), over a large number of trials. On
10
average, the cascading input increases the output rate by a factor of 1.5 over the synchronous inputs
-- we refer to this quantity as the cascade amplification factor (CAF).
Finally, in Fig. 5D, we study how the the cascade amplification factor depends on the parameters
that define the timing of spikes for the cascading inputs. First, we study the dependence on
the standard deviation σshift of the gamma variates determining the shift distribution. We note
that amplification factors above 1.5 hold robustly (i.e., for a range of shift σshift values). The
amplification factors decrease with shift variance. In the inset to panel D, we show how the gain
depends on the mean of the shift distribution µshift. On an individual trial, the response intensity
will depend strongly on the total number of input spikes. Thus, in order to enforce a fair comparison,
the mother process and markings used were identical in each trial of every panel of Fig. 5.
Figure 5:
(A) Example input (left) and output (right) for the nearest neighbor LIF network
receiving input with synchronous input. (B) Same as (A), but for cascading input. (C) Scatter
plot of the output spike count of the network receiving synchronous (horizontal axis) and cascading
input (vertical axis) with µshift = 2, σshift = 0.3. The red line is the diagonal. (D) Average gain
(rate in response to cascading input divided by rate in response to synchronous input) as a function
of the standard deviation of the gamma variates which compose the shift vectors for population-
level events (µshift was fixed at 2). The red dot indicates the value of σshift used in panel C. Inset
shows the same gain as panel D, but for varying the mean of the shift distribution (σshift = 0.3).
Spike counts in panels C and D were obtained for trials of length T = 100. Other system parameters
are given in the Appendix.
These observations have simple explanations in terms of the network dynamics and input statis-
tics. Neglecting, for a moment, population-level events, the network is configured so that correla-
tions in activity decrease with topographic distance. Accordingly, the probability of finding neurons
that are simultaneously close to threshold also decreases with distance. Under the synchronous in-
put model, a population-level event results in a simultaneous increase of the membrane potentials
of all neurons by an amount win, but unless the input is very strong (in which case every, or almost
every, neuron will fire regardless of fine-scale input structure), the set of neurons sufficiently close
to threshold to "capitalize" on the input and fire will typically be restricted to a topographically
11
adjacent subset. Neurons which do not fire almost immediately will soon have forgotten about this
population-level input. As a result, the output does not significantly reflect the chain-like structure
of the inputs (Fig. 5A, right).
On the other hand, in the case of the cascading input, the temporal structure of the input and
the timescale of synapses can operate synergistically. Consider a pair of adjacent neurons in the ring
network, called cells 1 and 2, arranged so that cell 2 is downstream from cell 1 in the direction of the
population-level chain events. When cell 1 spikes, it is likely that cell 2 will also have an elevated
membrane potential. The potential is further elevated by the delayed synaptic input from cell 1. If
cell 1 spikes in response to a population-level chain event, then cell 2 imminently receives an input
spike as well. If the synaptic filter and time-shift of the input spikes to each cell align, then the
firing probability of cell 2 will be large relative to chance. This reasoning can be carried on across
the network. Hence synergy between the temporal structure of inputs and network architecture
allows the network to selectively respond to the temporal structure of the inputs (Fig. 5B, right).
In [46], the effect of higher order correlations on the firing rate gain of an integrate -- and -- fire
neuron was studied by driving single cells using sums of SIP or MIP processes with equivalent firing
rates (first order cumulants) and pairwise correlations (second order cumulants). In contrast, in
the preceding example, the two inputs have equal long time spike count cumulants, and differ only
in temporal correlation structure. An increase in firing rate was due to network interactions, and
is therefore a population level effect. We return to this comparison in the Discussion.
These examples demonstrate how the GTaS model can be used to explore the impact of spatio-
temporal structure in population activity on network dynamics. We next proceed with a formal
derivation of the cumulant structure for a general GTaS process.
2.3 Cumulant structure of a GTaS process
The GTaS model defines an N -dimensional counting process. Following the standard description
for a counting process, X = (X1, . . . , XN ) on RN , given a collection of Borel subsets Ai ∈ B(R), i =
1, . . . , N , then X(A1 × ··· × AN ) = (X1(A1), . . . , XN (AN )) ∈ NN is a random vector where the
value of each coordinate i indicates the (random) number of points which fall inside the set Ai.
Note that the GTaS model defines processes that are marginally Poisson.
For each D ⊂ D = {1, . . . , N}, define the tail probability ¯pD by
(cid:88)
D⊂D(cid:48)⊂D
¯pD =
pD(cid:48).
(8)
Since pD is the probability that exactly the processes in D are marked, ¯pD is the probability that all
processes in D, as well as possibly other processes, are marked. An event from the mother process
is assigned to daughter process Xi with probability ¯p{i}. As noted above, an event attributed to
process i following a marking D (cid:51) i will be marginally shifted by a random amount determined
{i}
by the distribution Q
D which represents the projection of QD onto dimension i. Thus, the events
in the marginal process Xi are shifted in an independent and identically distributed (IID) manner
according to the mixture distribution Qi given by
(cid:80)
(cid:80)
{i}
D(cid:51)i pDQ
D
D(cid:51)i pD
.
Qi =
Note that IID shifting of the event times of a Poisson process generates another Poisson process of
identical rate. Thus, the process Xi is marginally Poisson with rate λ¯p{i} [60].
12
.
In deriving the statistics of the GTaS counting process X, it will be useful to express the
distribution of X as
X1(A1)
...
=distr
XN (AN )
(cid:80)
(cid:80)
D(cid:51)1 ξ(D; A1, . . . , AN )
...
D(cid:51)N ξ(D; A1, . . . , AN )
Here, each ξ(D; A1, . . . , AN ) is an independent Poisson process. This process counts the number of
points which are marked by a set D(cid:48) ⊃ D, but (after shifting) only the points with indices i ∈ D
lie in the corresponding set Ai. Precise definitions of the processes ξ and a proof of Eq. (9) may
be found in the Appendix. We emphasize that the Poisson processes ξ(D) do not directly count
points marked for the set D, but instead points which are marked for a set containing D that, after
shifting, only have their D-components lying in the "relevant" sets Ai.
{Xij}ij∈ ¯D for some set ¯D ⊂ D, consisting of ¯D = k distinct members of the collection of counting
Suppose we are interested in calculating dependencies among a subset of daughter processes,
processes X. Then the following alternative representation will be useful:
Xi1(Ai1)
...
=distr
(cid:80)
(cid:80)
(cid:88)
D(cid:48)⊃D
( ¯D\D)∩D(cid:48)=∅
i1∈D⊂ ¯D ζD(A1, . . . , AN )
...
ik∈D⊂ ¯D ζD(A1, . . . , AN )
ξ(D(cid:48); A1, . . . , AN ).
where
Xik (Aik )
ζD(A1, . . . , AN ) =
We illustrate this decomposition in the cases k = 2, 3 in Fig. 6. The sums in Eq. (10) run over
all sets D ⊂ D containing the indicated indices ij and contained within ¯D. The processes ζD are
comprised of a sum of all of the processes ξ(D(cid:48)) (defined below Eq. (9)) such that D(cid:48) contains all
of the indices D, but no other indices which are part of the subset ¯D under consideration. These
sums are non-overlapping, implying that the ζD are also independent and Poisson.
The following examples elucidate the meaning and significance of Eq. (10). We emphasize that
the GTaS process is a completely characterized, joint Poisson process, and we use Eq. (10) to
calculate cumulants of a GTaS process. In principle, any other statistics can be obtained similarly.
(9)
(10)
Second order cumulants (covariance) We first generalize a well-known result about the de-
pendence structure of temporally jittered pairs of Poisson processes, X1, X2. Assume that events
from a mother process with rate λ, are assigned to two daughter processes with probability p. Each
event time is subsequently shifted independently according to a univariate distribution f . The
cross-cumulant density (or cross-covariance function; see the Methods for cumulant definitions)
then has the form [14]
We generalize this result within the GTaS framework. At second order, Eq. (10) has a particularly
nice form. Following [11] we write for i (cid:54)= j (see Fig. 6A)
.
(11)
(cid:90)
κX
12(τ ) = λp
f (t)f (t + τ )dt = λp(f (cid:63) f )(τ ).
(cid:18)Xi(Ai)
(cid:19)
Xj(Aj)
=distr
(cid:18)ζ{i,j}(Ai, Aj) + ζ{i}(Ai)
(cid:19)
ζ{i,j}(Ai, Aj) + ζ{j}(Aj)
13
Figure 6: (A) Illustrating the representation given by Eq. (10) in the case of two distinct processes
(see Eq. (11)) with N = 4 and ¯D = {1, 2}. (B) Same as (A), for three processes with ¯D = {1, 2, 3}
(see Eq. (16)).
The process ζ{i,j} sums all ξ(D(cid:48)) for which {1, 2} ⊂ D(cid:48), while the process ζ{i} sums all ξ(D(cid:48)) such
that i ∈ D(cid:48), j /∈ D(cid:48), and ζ{j} is defined likewise.
Using the representation in Eq. (11), we can derive the second order cumulant (covariance)
structure of a GTaS process. First, we have
cov[Xi(Ai), Xj(Aj)] = κ[Xi(Ai), Xj(Aj)]
= κ[ζ{i,j}(Ai, Aj), ζ{i,j}(Ai, Aj)] + κ[ζ{i}(Ai), ζ{i,j}(Ai, Aj)]
+ κ[ζ{i,j}(Ai, Aj), ζ{j}(Aj)] + κ[ζ{i}(Ai), ζ{j}(Aj)]
= κ2[ζ{i,j}(Ai, Aj)] + 0
= E(cid:2)ζ{i,j}(Ai, Aj)(cid:3) .
The third equality follows from the construction of the processes ζD: if D (cid:54)= D(cid:48), then the processes
ζD, ζD(cid:48) are independent. The final equality follows from the observation that every cumulant of a
Poisson random variable equals its mean.
The covariance may be further expressed in terms of model parameters (see Theorem 1 for a
(cid:90)
(cid:88)
D(cid:48)⊃{i,j}
generalization of this result to arbitrary cumulant orders):
cov[Xi(Ai), Xj(Aj)] = λ
pD(cid:48)
P (t + Yi ∈ Ai, t + Yj ∈ Aj Y ∼ QD(cid:48)) dt.
(12)
In other words, the covariance of the counting processes is given by the weighted sum of the
probabilities that the (i, j) marginal of the shift distributions yield values in the appropriate sets.
14
The weights are the intensities of each corresponding component processes ξ(D) which contribute
events to both of the processes i and j.
In the case that QD ≡ Q, Eq. (12) reduces to the solution given in [11]. Using the tail
probabilities defined in Eq. (8), if QD ≡ Q for all D, the integral in Eq. (12) no longer depends on
the subset D(cid:48), and the equation may be written as
cov[Xi(Ai), Xj(Aj)] = λ¯p{i,j}
P (t + Yi ∈ Ai, t + Yj ∈ Aj Y ∼ Q) dt.
Using Eq. (12), we may also compute the second cross-cumulant density (also called the covari-
ance density) of the processes. From the definition of the cross-cumulant density (Eq. (24) in the
Methods), this is given by
cov[Xi([0, ∆t)), Xj([τ, τ + ∆t))]
κX
ij (τ ) = lim
∆t→0
(cid:88)
(cid:90)
= λ
D(cid:48)⊃{i,j}
pD(cid:48)
lim
∆t→0
∆t2
P (t + Yi ∈ [0, ∆t), t + Yj ∈ [τ, τ + ∆t) Y ∼ QD(cid:48))
(13)
dt.
∆t2
Before continuing, we note that given a random vector Y = (Y1, . . . , YN ) ∼ Q, where Q has density
q(y1, . . . , yN ), the vector Z = (Y2 − Y1, . . . , YN − Y1) has density qZ given by
(cid:90)
(cid:90)
qZ(τ1, . . . , τN−1) =
q(t, t + τ1, . . . , t + τN−1)dt.
(14)
Assuming that the distributions QD(cid:48) have densities qD(cid:48), and denoting by q
density of the variables Yi, Yj under QD(cid:48), we have that
{i,j}
D(cid:48)
the bivariate marginal
κX
ij (τ ) = λ
pD(cid:48)
{i,j}
D(cid:48)
q
(t, t + τ )dt.
(15)
(cid:90)
(cid:88)
D(cid:48)⊃{i,j}
According to Eq. (14), the integrals present in Eq. (15) are simply the densities of the variables
Yj − Yi, where Y ∼ QD(cid:48).
Thus κX
ij (τ ), which captures the additional probability for events in the marginal processes Xi
and Xj separated by τ units of time beyond what can be predicted from lower order statistics is
given by a weighted sum (in this case, the lower order statistics are marginal intensities -- see the
discussion around Eq. (24) of the Methods). The weights are the "marking rates" λpD(cid:48) for markings
contributing events to both component processes, while the summands are the probabilities that
the corresponding shift distributions yield a pair of shifts in the proper arrangement - specifically,
the shift applied to the event as attributed to Xi precedes that applied to the event mapped to Xj
by τ units of time.
Third order cumulants To determine the higher order cumulants for a GTaS process, one can
again use the representation given in Eq. (10). The distribution of a subset of three processes may
be expressed in the form (see Fig. 6B)
Xi(Ai)
Xj(Aj)
Xk(Ak)
=distr
ζ{i,j,k} + ζ{i,j} + ζ{i,k} + ζ{i}
ζ{i,j,k} + ζ{i,j} + ζ{j,k} + ζ{j}
ζ{i,j,k} + ζ{i,k} + ζ{j,k} + ζ{k},
,
(16)
15
where, for simplicity, we suppressed the arguments of the different ζD on the right hand side. Again,
the processes in the representation are independent and Poisson distributed. The variable ζ{i,j,k}
is the sum of all random variables ξ(D) (see Eq. (9)) with D ⊃ {i, j, k}, while the variable ζ{i,j} is
now the sum of all ξ(D) with D ⊃ {i, j}, but k /∈ D. The rest of the variables are defined likewise.
Using properties (C1) and (C2) of cumulants given in the Methods, and assuming that i, j, k are
distinct indices, we have
κ(Xi(Ai), Xj(Aj), Xk(Ak)) = κ3(ζ{i,j,k}) = E(cid:2)ζ{i,j,k}(cid:3) .
The second equality follows from the fact that all cumulants of a Poisson distributed random
variable equal its mean. Similar to Eq. (12), we may write
κ(Xi(Ai), Xj(Aj), Xk(Ak)) = λ
pD(cid:48)
P (t + Yi ∈ Ai, t + Yj ∈ Aj, t + Yk ∈ Ak Y ∼ QD(cid:48)) dt.
The third cross-cumulant density is then given similarly to the second order function by
κX
ijk(τ1, τ2) = λ
pD(cid:48)
{i,j,k}
D(cid:48)
q
(t, t + τ1, t + τ2)dt.
(cid:90)
(cid:88)
D(cid:48)⊃{i,j,k}
(cid:90)
(cid:88)
D(cid:48)⊃{i,j,k}
Here, we have again assumed the existence of densities qD(cid:48), and denote by q
the joint marginal
density of the variables Yi, Yj, Yk under qD(cid:48). The integrals appearing in the expression for the third
order cross-cumulant density are the probability densities of the vectors (Yj − Yi, Yk − Yi), where
Y ∼ QD(cid:48).
{i,j,k}
D(cid:48)
General cumulants Finally, consider a general subset of k distinct members of the vector count-
ing process X as in Eq. (10). The following theorem provides expressions for the cross-cumulants
of the counting processes, as well as the cross-cumulant densities, in terms of model parameters in
this general case. The proof of Theorem 1 is given in the Appendix.
Theorem 1. Let X be a joint counting process of GTaS type with total intensity λ, marking
distribution (pD)D⊂D, and family of shift distributions (QD)D⊂D. Let A1, . . . , Ak be arbitrary sets
in B(R), and ¯D = {i1, . . . , ik} ⊂ D with ¯D = k. The cross-cumulant of the counting processes
may be written
κ(Xi1(A1), . . . , Xik (Ak)) = λ
pD(cid:48)
P (t1 + Y
¯D ∈ A1 × ··· × AkY ∼ QD(cid:48))dt
(17)
where Y ¯D represents the projection of the random vector Y on to the dimensions indicated by the
members of the set ¯D. Furthermore, assuming that the shift distributions possess densities (qD)D⊂D,
the cross-cumulant density is given by
κX
i1···ik
(τ1, . . . , τk−1) = λ
pD(cid:48)
¯D
D(cid:48)(t, t + τ1,··· , t + τk−1)dt,
q
(18)
(cid:90)
(cid:88)
D(cid:48)⊃ ¯D
(cid:90)
(cid:88)
D(cid:48)⊃ ¯D
where q ¯D
D(cid:48) indicates the kth order joint marginal density of qD(cid:48) in the dimensions of ¯D.
16
An immediate corollary of Theorem 1 is a simple expression for the infinite-time-window cu-
mulants, obtained by integrating the cumulant density across all time lags τi. From Eq. (33), we
have
(cid:90)
(cid:90)
(cid:88)
D(cid:48)⊃ ¯D
γX
i1···ik
(∞) =
···
κX
i1···ik
(τ1, . . . , τk−1)dτk−1 ··· dτ1 = λ
pD(cid:48) · 1 = λ¯p ¯D.
(19)
This shows that the infinite time window cumulants for a GTaS process are non-increasing with
respect to the ordering of sets, i.e.,
γX
i1···ik
(∞) ≥ γX
i1···ikik+1
(∞).
We conclude this section with a short technical remark: Until this point, we have considered
only the cumulant structure of sets of unique processes. However occasionally, one may wish
to calculate a cumulant for a set of processes including repeats. Take, for example, a cumulant
κ(X1(A1), X1(A2), X3(A3)). Owing to the marginally Poisson nature of the GTaS process, we
would have (referring to the Methods for cumulant definitions)
κ(X1(A1), X1(A2), X3(A3)) = κ(2,1)(X1(A1 ∩ A2), X3(A3))
if X ∼ GTaS.
For a general counting process X, it may be shown that
κX
113(τ1, τ2) = δ(τ1)κX
13(τ2) + "non-singular contributions".
In addition, the second order auto-cumulant density may be written [21]
κX
ii (τ ) = riδ(τ ) + "non-singular contributions",
(20)
(21)
where ri is the stationary rate. The singular contribution shown in Eq. (21) at third order is
in analogy to the delta contribution proportional to the firing rate which appears in the second-
order auto-cumulant density. For a GTaS process, the non-singular contributions in Eq. (21) are
identically zero, following directly from Eq. (20). Expressions similar to Eqs. (20, 21) hold for
general cases.
3 Discussion
We have introduced a general method of generating spike trains with flexible spatiotemporal struc-
ture. The GTaS model is completely analytically tractable: all statistics of interest can be obtained
directly from the distributions used to define it. It is based on an intuitive method of selecting and
shifting point processes from a "mother" train. Moreover, the GTaS model can be used to easily
generate partially synchronous states, cluster firing, cascading chains, and other spatiotemporal
patterns of neural activity.
Processes generated by the GTaS model are naturally described by cumulant densities of pair-
wise and higher orders. This raises the question of whether such statistics are readily computable
from data, so that realistic classes of GTaS models can be defined in the first place. One approach
is to fit mechanistic models to data, and to use the higher order structure that follows from the
underlying mechanisms [79]. A synergistic blend of other methods with the GTaS framework may
also be fruitful -- for example, the CuBIC framework of [70] could be used to determine relevant
marking orders, and the parametrically-described GTaS process could then be fit to allow gen-
eration of surrogate data after selection of appropriate classes of shift distributions. When it is
17
necessary to infer higher order structure in the face of data limitations, population cumulants are
an option to increase statistical power (albeit at the cost of spatial resolution; see Figure 4).
While the GTaS model has flexible higher order structure, it is always marginally Poisson.
While throughout the cortex, spiking is significantly irregular [65, 35], the level of variability differs
across cells, with Fano factors ranging from below 0.5 to above 1.5 -- in comparison with the Poisson
value of 1 [20]. Changes in variability may reflect cortical states and computation [77, 47]. A model
that would allow flexible marginal variability would therefore be very useful. Unfortunately, the
tractability of the GTaS model is closely related to the fact that the marginal processes are Poisson.
Therefore an immediate generalization does not seem possible.
A number of other models have been used to describe population activity. Maximum entropy
(ME) approaches also result in models with varied spatial activity; these are defined based on
moments or other averaged features multivariate spiking activity [63, 61]. Such models are often
used to fit purely spatial patterns of activity, though [72, 52] have extended the techniques to
treat temporal correlations as well. Generalized linear models (GLMs) have been used successfully
to describe spatiotemporal patterns at second [57], and third order [56].
In comparison to the
present GTaS method, both GLMs and ME models are more flexible. They are feature well-defined
approaches for fitting to data, including likelihood-based methods with well-behaved convexity
properties. What the GTaS method contributes is an explicit way to generate population activity
with explicitly specified high order spatio-temporal structure. Moreover, the lower order cumulant
structure of a GTaS process can be modified independently of the higher order structure, though
the reverse is not true.
There are a number of possible implications of such spatio-temporal structure for communication
within neural networks. In Section 2.2, we showed that these temporal correlations can play a role
similar to that of spatial correlations established in [46] for determining network input-output
transfer. Our model allowed us to examine that impact of such temporal correlations on the
network-level gain of a downstream population (cascade amplification factor). Even in a very
simple network it was clear that the strength of the response is determined jointly by the temporal
structure of the input to the network, and the connectivity within the network. Kuhn et al.
examined the effect of higher order structure on the firing rate gain of an integrate -- and -- fire neuron
by driving it with a mixture of SIP or MIP processes [46]. However, in these studies, only the spatial
structure of higher order activity was varied. The GTaS model allows us to concurrently change
the temporal structure of correlations. In addition, the precise control of the cumulants allows us
to derive models which are equivalent up to a certain cross-cumulant order, when the configuration
of marking probabilities and shift distributions allow it (as for the SIP and MIP processes of [46],
which are equivalent at second order).
Such patterns of activity may be useful when experimentally probing dendritic information
processing [26], or investigating the response of neuronal networks to complex patterns of input [42].
Spatiotemporal patterns may also be generated by cell assemblies [10]. The firing in such assemblies
can be spatially structured, and this structure may not be reflected in the activity of participating
cells. Assemblies can exhibit persistent patterns of firing, sometimes with millisecond precision [33].
The GTaS framework is well suited to describe exactly such activity patterns. The examples we
presented can be easily extended to generate more complex patterns of activity with overlapping
cell assemblies, different cells leading the activity, and other variations.
Understanding impact of spatiotemporal patterns on neural computations remains an open and
exciting problem. Progress will require cooperation among simulation, theory, and experimental
work -- the latter taking advantage of novel stimulation techniques. We hope that the GTaS model,
18
as a practical and flexible method for generating high-dimensional, correlated spike trains, will play
a significant role along the way.
4 Methods
Cumulants as a measure of dependence We first define cross-cumulants (also called joint
cumulants) [71, 25, 43] and review some important properties of these quantities. Define the
cumulant generating function g of a random vector X = (X1, . . . , XN ) by
g(t1, . . . , tN ) = log
N(cid:88)
E
exp
.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)t1=···=tN =0
where r = (r1, . . . , rN ) is a N -vector of positive integers, and r =(cid:80)N
The r-cross-cumulant of the vector X is given by
∂r
1 ··· ∂trN
∂tr1
N
g(t1, . . . , tN )
κr(X) =
.
tjXj
j=1
i=1 ri. We will generally deal
with cumulants where all variables are considered at first order, without excluding the possibility
that some variables are duplicated. In this case, we define the cross-cumulant κ(X), of the variables
in the random vector X = (X1, . . . , XN ) as
κ(X) := κ1(X) =
∂N
∂t1 ··· ∂tN
This relationship may be expressed in combinatorial form:
g(t1, . . . , tN )
(cid:12)(cid:12)(cid:12)(cid:12)t1=···=tN =0
(cid:88)
(π − 1)!(−1)π−1 (cid:89)
where 1 = (1, . . . , 1).
(cid:35)
Xi
(22)
(cid:34)(cid:89)
i∈B
E
B∈π
κ(X1, . . . , XN ) =
π
(cid:88)
where π runs through all partitions of D = {1, . . . , N}, and B runs over all blocks in a partition
π. More generally, the r-cross-cumulant may be expressed in terms of moments by expanding the
cumulant generating function as a Taylor series, noting that
g(t1, . . . , tN ) =
κr(X1, . . . , XN )
1 ··· xrN
xr1
d
with r! =
ri!,
r
r!
N(cid:89)
i=1
similarly expanding the moment generating function M (t) = eg(t), and matching the polynomial
coefficients. Note that the nth cumulant κn of a random variable X may be expressed as a joint
cumulant via
κn(X) = κ(X, . . . , X)
.
(cid:124)
(cid:123)(cid:122)
(cid:125)
n copies of X
We will utilize the following two principal properties of cumulants [71, 15, 70, 54]:
(C1) Multilinearity - for any random variables X, Y,{Zi}N
i=2, we have
κ(aX + bY, Z2, . . . , ZN ) = aκ(X, Z2, . . . , ZN ) + bκ(Y, Z2, . . . , ZN ).
This holds regardless of dependencies amongst the random variables.
19
(C2) If any subset of the random variables in the cumulant argument is independent from the
remaining, the cross-cumulant is zero - i.e., if {X1, . . . , XN1} and {Y1, . . . , YN2} are sets of
random variables such that each Xi is independent from each Yj, then
for all rX ∈ NN1
κ(rX ,rY )(X1, . . . , XN1, Y1, . . . , YN2) = 0
+ , rY ∈ NN2
+ .
To exhibit another key property of cumulants, consider a 4-vector X = (X1, X2, X3, X4) with
non-zero fourth cumulant and a random variable Z independent of each Xi. Define Y = (X1 +
Z, X2 + Z, X3 + Z, X4). Using properties (C1), (C2) above, it follows that
κ(Y1, Y2, Y3) = κ(X1, X2, X3) + κ3(Z).
On the other hand, it is also true that
κ(Y) = κ(X),
that is, adding the variable Z to only a subset of the variables in X results in changes to cumulants
involving only that subset, but not to the joint cumulant of the entire vector. In this sense, an rth
order cross-cumulant of a collection of random variables captures exclusively dependencies amongst
the collection which cannot be described by cumulants of lower order. In the example above, only
the joint statistical properties of a subset of X were changed. As a result, the total cumulant κ(X)
remained fixed.
From Eq. (22), it is apparent that κ(Xi) = E[Xi], and κ(Xi, Xj) = cov[Xi, Xj]. In addition,
the third cumulant, like the second, is equal to the corresponding central moment:
κ(Xi, Xj, Xk) = E[(Xi − E[Xi])(Xj − E[Xj])(Xk − E[Xk])] .
As cumulants and central moments agree up to third order, central moments up to third order
inherit the properties discussed above at these orders. On the other hand, the fourth cumulant is
not equal to the fourth central moment. Rather:
κ(Xi,Xj, Xk, Xl) = E[(Xi − E[Xi])(Xj − E[Xj])(Xk − E[Xk])(Xl − E[Xl])]
− cov[Xi, Xj] cov[Xk, Xl] − cov[Xi, Xk] cov[Xj, Xl] − cov[Xi, Xl] cov[Xj, Xk] .
(23)
Higher cumulants have similar (but more complicated) expansions in terms of central moments.
Accordingly, central moments of fourth and higher order do not inherit properties (C1), (C2).
Temporal statistics of point processes
In the Results, we present an extension of previous
work [11] in which we construct and analyze multivariate counting processes X = (X1, . . . , XN )
where each Xi is marginally Poisson.
Formally, a counting process X is an integer-valued random measure on B(RN ). Evaluated
on subset A1 × ··· × AN of B(RN ), the random vector (X1(A1), . . . , XN (AN )) counts events in d
distinct categories whose times of occurrence fall in to the sets Ai. A good general reference on the
properties of counting processes (marginally Poisson and otherwise) is [22].
The assumption of Poisson marginals implies that for a set Ai ∈ B(R), the random variable
Xi(Ai) follows a Poisson distribution with mean λi(cid:96)(Ai), where (cid:96) is the Lebesgue measure on R, and
λi is the (constant) rate for the ith process. The processes under consideration will further satisfy a
joint stationarity condition, namely that the distribution of the vector (X1(A1 +t), . . . , XN (AN +t))
does not depend on t, where Ai + t denotes the translated set {a + t : a ∈ Ai}.
20
We now consider some common measures of temporal dependence for jointly stationary vector
counting processes. We will refer to the quantity Xi[0, T ] as the spike count of process i over [0, T ].
The quantity γX
(T ) (which we will refer to as a spike count cumulant) is given by
i1···ik
γX
i1···ik
(T ) =
1
T
κ[Xi1[0, T ], . . . , Xik [0, T ]]
measures kth order correlations amongst spike counts for the listed processes which occur over
windows of length T . At second order, γX
ij (T ) measures the covariance of the spike counts of
processes i, j over a common window of length T . The infinite window spike count cumulant
quantifies dependencies in the spike counts of point processes over arbitrarily long windows, and is
given by
γX
i1···ik
(∞) = lim
T→∞ γX
i1···ik
(T ).
A related measure is the kth order cross-cumulant density κX
i1,...,ik
(τ1, . . . , τk−1), defined by
κX
i1···ik
(τ1, . . . , τk−1) = lim
∆t→0
1
∆tk κ[Xi1[0, ∆t], Xi2[τ1, τ1 + ∆t], . . . , Xik [τk−1, τk−1 + ∆t]].
(24)
The cross-cumulant density should be interpreted as a measure of the likelihood -- above what may
be expected from knowledge of the lower order cumulant structure -- of seeing events in processes
i2, . . . , ik at times τ1 + t, . . . , τk−1 + t, conditioned on event in process i1 at time t. The infinite
window spike count cumulant is equal to the total integral under the cross-cumulant density,
(cid:90)
(cid:90)
γX
i1···ik
(∞) =
···
κX
i1···ik
(τ1, . . . , τk−1)dτk−1 ··· dτ1.
As an example, we again consider the familiar second-order cross-cumulant density κX
ij (τ ) - often
referred to as the cross-covariance density or cross-correlation function. Defining the conditional
intensity hij(τ ) of process j, conditioned on process i to be
hX
ij (τ ) = lim
∆t→0
1
∆t
P (Xj[τ, τ + ∆t] > 0Xi[0, ∆t] > 0),
that is, the intensity of j conditioned on an event in process i which occurred τ units of time in
the past, then it is not difficult to show that
ij (τ ) = λihij(τ ) − λiλj.
κX
That is, the second order cross-cumulant density supplies the probability of chance of observing
an event attributed to process i, followed by one attributed to process j, τ units of time later,
above what would be expected from knowledge of first order statistics (given by the product of
the marginal intensities, λiλj). More generally, at higher orders, the cross-cumulant density should
be interpreted as a measure of the likelihood (above what may be expected from knowledge of
the lower order correlation structure) of seeing events attribute to processes i2, . . . , ik at times
τ1 + t, . . . , τk−1 + t, conditioned on an event in process i1 at time t.
Another statistic useful in the study of a correlated vector counting process X is the population
cumulant density. At second-order, the population cumulant density for Xi takes the form [48]
κX
i,pop(τ ) =
κX
ij (τ ).
More generally, the kth order population cumulant density corresponding to the processes Xi1, . . . , Xik−1
is given by
κX
i1···ik−1,pop(τ1, . . . , τk−1) =
κX
i1···ik−1j(τ1, . . . , τk−1).
(25)
(cid:88)
(cid:88)
j(cid:54)=i
j(cid:54)=i1,...,ik
21
References
[1] M Abeles and Y Prut. Spatio-temporal firing patterns in the frontal cortex of behaving
monkeys. J Physiology-Paris, 90(3):249 -- 250, 1996.
[2] Moshe Abeles. Corticonics: Neural circuits of the cerebral cortex. Cambridge University Press,
1991.
[3] AMHJ Aertsen, M Diesmann, and MO Gewaltig. Propagation of synchronous spiking activity
in feedforward neural networks. J Physiology-Paris, 90(3):243 -- 247, 1996.
[4] Shun-ichi Amari, Hiroyuki Nakahara, Si Wu, and Yutaka Sakai. Synchronous firing and higher-
order interactions in neuron pool. Neural Comput, 15(1):127 -- 142, 2003.
[5] AM Amjad, DM Halliday, JR Rosenberg, and BA Conway. An extended difference of coherence
test for comparing and combining several independent coherence estimates: theory and appli-
cation to the study of motor units and physiological tremor. J Neurosci Meth, 73(1):69 -- 79,
1997.
[6] Bruno B Averbeck, Peter E Latham, and Alexandre Pouget. Neural correlations, population
coding and computation. Nat Rev Neurosci, 7(5):358 -- 366, May 2006.
[7] Yuval Aviel, Evgeny Pavlov, Moshe Abeles, and David Horn. Synfire chain in a balanced
network. Neurocomputing, 44:285 -- 292, 2002.
[8] W. Bair, E. Zohary, and W.T. Newsome. Correlated firing in macaque visual area mt: time
scales and relationship to behavior. J Neurosci, 21(5):1676 -- 1697, 2001.
[9] Andrea K Barreiro, Julijana Gjorgjieva, Fred Rieke, and Eric Shea-Brown. When are feedfor-
ward microcircuits well-modeled by maximum entropy methods? Arxiv preprint, 2010.
[10] Brice Bathellier, Lyubov Ushakova, and Simon Rumpel. Discrete neocortical dynamics predict
behavioral categorization of sounds. Neuron, 76(2):435 -- 449, 2012.
[11] N. Bauerle and R. Grubel. Multivariate counting processes: copulas and beyond. Astin
Bulletin, 35(2):379, 2005.
[12] Tiago Branco, Beverley A Clark, and Michael Hausser. Dendritic discrimination of temporal
input sequences in cortical neurons. Sci Signal, 329(5999):1671, 2010.
[13] Tiago Branco and Michael Hausser. Synaptic integration gradients in single cortical pyramidal
cell dendrites. Neuron, 69(5):885 -- 892, 2011.
[14] Romain Brette. Generation of correlated spike trains. Neural Comput, 21(1):188 -- 215, 2009.
[15] David R Brillinger. An introduction to polyspectra, 1964.
[16] Gyorgy Buzs´aki. Neural syntax:
cell assemblies, synapsembles, and readers. Neuron,
68(3):362 -- 385, 2010.
[17] Nicholas Cain and Eric Shea-Brown. Impact of correlated neural activity on decision-making
performance. Neural Comput, 25(2):289 -- 327, 2013.
22
[18] Catherine E Carr, Hagai Agmon-Snir, and John Rinzel. The role of dendrites in auditory
coincidence detection. Nature, 393(6682):268 -- 272, May 1998.
[19] Brian Y Chow, Xue Han, Allison S Dobry, Xiaofeng Qian, Amy S Chuong, Mingjie Li,
Michael A Henninger, Gabriel M Belfort, Yingxi Lin, Patrick E Monahan, et al. High-
performance genetically targetable optical neural silencing by light-driven proton pumps. Na-
ture, 463(7277):98 -- 102, 2010.
[20] Mark M Churchland, Byron M Yu, John P Cunningham, Leo P Sugrue, Marlene R Cohen,
Greg S Corrado, William T Newsome, Andrew M Clark, Paymon Hosseini, Benjamin B Scott,
David C Bradley, Matthew A Smith, Adam Kohn, J Anthony Movshon, Katherine M Arm-
strong, Tirin Moore, Steve W Chang, Lawrence H Snyder, Stephen G Lisberger, Nicholas J
Priebe, Ian M Finn, David Ferster, Stephen I Ryu, Gopal Santhanam, Maneesh Sahani, and
Krishna V Shenoy. Stimulus onset quenches neural variability: a widespread cortical phe-
nomenon. Nat Neurosci, 13(3):369 -- 378, March 2010.
[21] D David Roxbee Cox and Valerie Isham. Point processes, volume 12. Chapman & Hall/CRC,
1980.
[22] DJ Daley and D Vere-Jones. An Introduction to the Theory of Point Processes: Volume I:
Elementary Theory and Methods, volume 1. Springer, 2002.
[23] Gaddum Duemani Reddy, Keith Kelleher, Rudy Fink, and Peter Saggau. Three-dimensional
random access multiphoton microscopy for functional imaging of neuronal activity. Nat Neu-
rosci, 11(6):713 -- 720, June 2008.
[24] Elad Ganmor, Ronen Segev, and Elad Schneidman. Sparse low-order interaction network
underlies a highly correlated and learnable neural population code. Proc Natl Acad Sci,
108(23):9679 -- 9684, 2011.
[25] CW Gardiner. Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sci-
ences. Springer-Verlag, Berlin, 2009.
[26] Sonia Gasparini and Jeffrey C Magee. State-dependent dendritic computation in hippocampal
CA1 pyramidal neurons. J Neurosci, 26(7):2088 -- 2100, February 2006.
[27] Sonja Grun and Stefan Rotter. Analysis of parallel spike trains. Springer, 2010.
[28] Robert Gutig and Haim Sompolinsky. The tempotron: a neuron that learns spike timing -- based
decisions. Nat Neurosci, 9(3):420 -- 428, 2006.
[29] Diego A Gutnisky and Kresimir Josic. Generation of spatio-temporally correlated spike-trains
and local-field potentials using a multivariate autoregressive process. J Neurophysiol, December
2009.
[30] Xue Han and Edward S Boyden. Multiple-color optical activation, silencing, and desynchro-
nization of neural activity, with single-spike temporal resolution. PloS one, 2(3):e299, 2007.
[31] Bryan J Hansen, Mircea I Chelaru, and Valentin Dragoi. Correlated variability in laminar
cortical circuits. Neuron, 76(3):590 -- 602, 2012.
[32] Kenneth D Harris. Neural signatures of cell assembly organization. Nat Rev Neurosci, 6(5):399 --
407, 2005.
23
[33] Kenneth D Harris, Darrell A Henze, Hajime Hirase, Xavier Leinekugel, George Dragoi, Andras
Czurk´o, Gyorgy Buzs´aki, et al. Spike train dynamics predicts theta-related phase precession
in hippocampal pyramidal cells. Nature, 417(6890):738 -- 741, 2002.
[34] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology
Press, 1949.
[35] G. R. Holt, W. R. Softky, C. Koch, and R. J. Douglas. Comparison of discharge variability in
vitro and in vivo in cat visual cortex neurons. J Neurophysiol, 75(5):1806 -- 1814, 1996.
[36] JJ Hopfield. Pattern recognition computation using action potential timing for stimulus rep-
resentation. Nature, 376(6535):33 -- 36, 1995.
[37] Yuji Ikegaya, Gloster Aaron, Rosa Cossart, Dmitriy Aronov, Ilan Lampl, David Ferster, and
Rafael Yuste. Synfire chains and cortical songs: temporal modules of cortical activity. Science
Sig, 304(5670):559, 2004.
[38] Eugene M Izhikevich. Polychronization: Computation with spikes. Neural Comput, 18(2):245 --
282, 2006.
[39] Lloyd A Jeffress. A place theory of sound localization. J Comp Physiol Psychol, 41(1):35 -- 39,
1948.
[40] Don H Johnson and Ilan N Goodman. Jointly poisson processes. arXiv preprint, 2009.
[41] P X Joris, P H Smith, and TCT Yin. Coincidence detection in the auditory system: 50 years
after Jeffress. Neuron, 21(6):1235 -- 1238, December 1998.
[42] I Kahn, U Knoblich, M Desai, J Bernstein, AM Graybiel, ES Boyden, RL Buckner, and
CI Moore. Optogenetic drive of neocortical pyramidal neurons generates fMRI signals that are
correlated with spiking activity. Brain Res, 2013.
[43] Maurice G Kendall, Alan Stuart, and JK Ord. The advanced theory of statistics (volume 1),
1969.
[44] U. Koster, J. Sohl-Dickstein, C. Gray, and B. Olshausen. Higher Order Correlations within
arXiv preprint q-
Cortical Layers Dominate Functional Connectivity in Microcolumns.
bio/1301.0050, 2013.
[45] Michael Krumin and Shy Shoham. Generation of spike trains with controlled auto- and cross-
correlation functions. Neural Comput, 21(6):1642 -- 1664, June 2009.
[46] A. Kuhn, A. Aertsen, and S. Rotter. Higher-order statistics of input ensembles and the response
of simple model neurons. Neural Comput, 15(1):67 -- 101, 2003.
[47] Ashok Litwin-Kumar and Brent Doiron. Slow dynamics and high variability in balanced
cortical networks with clustered connections. Nat Neurosci, 15(11):1498 -- 1505, September
2012.
[48] Artur Luczak, Peter Bartho, and Kenneth D Harris. Gating of sensory input by spontaneous
cortical activity. J Neurosci, 33(4):1684 -- 1695, January 2013.
24
[49] Artur Luczak, Peter Barth´o, Stephan L Marguet, Gyorgy Buzs´aki, and Kenneth D Harris. Se-
quential structure of neocortical spontaneous activity in vivo. Proc Natl Acad Sci, 104(1):347 --
352, 2007.
[50] Jakob H Macke, Philipp Berens, Alexander S Ecker, Andreas S Tolias, and Matthias Bethge.
Generating spike trains with specified correlation coefficients. Neural Comput, 21(2):397 -- 423,
February 2009.
[51] Jakob H Macke, Manfred Opper, and Matthias Bethge. Common input explains higher-order
correlations and entropy in a simple model of neural population activity. Phys Rev Lett,
106(20):208102, 2011.
[52] O. Marre, S. El Boustani, Y. Fre?gnac, and A. Destexhe. Prediction of spatiotemporal patterns
of neural activity from pairwise correlations. Physical Review Letters, 102:138101, 2009.
[53] M Meister and MJ Berry II. The neural code of the retina. Neuron, 22(3):435, 1999.
[54] J.M. Mendel. Tutorial on higher-order statistics (spectra) in signal processing and system
theory: Theoretical results and some applications. Proceedings of the IEEE, 79(3):278 -- 305,
1991.
[55] Fernando Montani, Elena Phoka, Mariela Portesi, and Simon R. Schultz.
tical modelling of higher-order correlations in pools of neural activity.
392(14):10.1016/j.physa.2013.03.012, 2013.
Statis-
Physica A,
[56] Ifije E Ohiorhenuan, Ferenc Mechler, Keith P Purpura, Anita M Schmid, Qin Hu, and
Jonathan D Victor. Sparse coding and high-order correlations in fine-scale cortical networks.
Nature, 466(7306):617 -- 621, July 2010.
[57] Jonathan W Pillow, Jonathon Shlens, Liam Paninski, Alexander Sher, Alan M Litke, E J
Chichilnisky, and Eero P Simoncelli. Spatio-temporal correlations and visual signalling in a
complete neuronal population. Nature, 454(7207):995 -- 999, July 2008.
[58] Robert Rosenbaum, James Trousdale, and Kresimir Josi´c. The effects of pooling on spike train
correlations. Front Neurosci, 5, 2011.
[59] Robert J Rosenbaum, James Trousdale, and Kresimir Josi´c. Pooling and correlated neural
activity. Front Comput Neurosci, 4, 2010.
[60] Sheldon M. Ross. Stochastic Processes. Wiley, 2 edition, 1 1995.
[61] Yasser Roudi, Sheila Nirenberg, and Peter E Latham. Pairwise maximum entropy models for
studying large biological systems: when they can work and when they can't. PLoS Comp Biol,
5(5):e1000380, May 2009.
[62] E Salinas and T J Sejnowski. Correlated neuronal activity and the flow of neural information.
Nat Rev Neurosci, 2(8):539 -- 50, Aug 2001.
[63] Elad Schneidman, Michael J Berry, Ronen Segev, and William Bialek. Weak pairwise correla-
tions imply strongly correlated network states in a neural population. Nature, 440(7087):1007 --
1012, April 2006.
25
[64] Elad Schneidman, Micheal J. Berry, Ronen Segev, and William Bialek. Weak pairwise corre-
lations imply strongly correlated network states in a neural population. Nature, 440(20):1007 --
1012, 2006.
[65] M. N. Shadlen and W. T. Newsome. The variable discharge of cortical neurons: implications
for connectivity, computation, and information coding. J Neurosci, 18:3870 -- 3896, 1998.
[66] Maoz Shamir and Haim Sompolinsky. Nonlinear population codes. Neural Comput, 16(6):1105 --
1136, June 2004.
[67] J. Shlens, G.D. Field GD, J.L. Gauthier, M.I. Grivich MI, D. Petrusca D, A. Sher A, A.M.
Litke, and E.J. Chichilnisky EJ. The structure of multi-neuron firing patterns in primate
retina. J Neurosci, 26:8254 -- 8266, 2006.
[68] Jonathon Shlens, Greg D Field, Jeffrey L Gauthier, Martin Greschner, Alexander Sher, Alan M
Litke, and EJ Chichilnisky. The structure of large-scale synchronized firing in primate retina.
J Neurosci, 29(15):5022 -- 5031, 2009.
[69] Wolf Singer. Neuronal synchrony: A versatile code review for the definition of relations?
Neuron, 24:49 -- 65, 1999.
[70] B. Staude, S. Rotter, and S. Grun. Cubic: cumulant based inference of higher-order correlations
in massively parallel spike trains. J Comput Neurosci, 29(1):327 -- 350, 2010.
[71] R.L. Stratonovich. Topics in the Theory of Random Noise, volume 2. Gordon and Breach,
New York, 1967.
[72] A. Tang et al. A maximum entropy model applied to spatial and temporal correlations from
cortical networks in vitro. Journal of Neuroscience, 28:505 -- 518, 2008.
[73] T. Tetzlaff, S. Rotter, E. Stark, M. Abeles, A. Aertsen, and M. Diesmann. Dependence
of neuronal correlations on filter characteristics and marginal spike train statistics. Neural
Comp, 20(9):2133 -- 2184, 2008.
[74] Tom Tetzlaff, Michael Buschermohle, Theo Geisel, and Markus Diesmann. The spread of rate
and correlation in stationary cortical networks. Neurocomputing, 52:949 -- 954, 2003.
[75] Simon Thorpe, Arnaud Delorme, and Rufin Van Rullen. Spike-based strategies for rapid
processing. Neural Netw, 14(6-7):715 -- 725, 2001.
[76] Juan Carlos Vasquez, Olivier Marre, Adrian G Palacios, MJ Berry II, and Bruno Cessac. Gibbs
distribution analysis of temporal correlations structure in retina ganglion cells. J Physiology-
Paris, 106(3):120 -- 127, 2012.
[77] Benjamin White, L F Abbott, and J´ozsef Fiser. Suppression of cortical neural variability is
stimulus- and state-dependent. J Neurophysiol, 108(9):2383 -- 2392, November 2012.
[78] Ning-long Xu, Mark T Harnett, Stephen R Williams, Daniel Huber, Daniel H O'Connor, Karel
Svoboda, and Jeffrey C Magee. Nonlinear dendritic integration of sensory and motor input
during an active sensing task. Nature, 492(7428):247 -- 251, December 2012.
[79] Shan Yu, Hongdian Yang, Hiroyuki Nakahara, Gustavo S Santos, Danko Nikoli´c, and Dietmar
Plenz. Higher-order interactions characterized in cortical activity. J Neurosci, 31(48):17514 --
17526, 2011.
26
5 Appendix
5.1 Proof of the distributional representation of the GTaS model in Eq. (9)
The construction of the GTaS model allows us to provide a useful distributional representation of
the process. We describe this representation in a theorem that generalizes Theorem 1 in [11]. This
theorem also immediately implies that the GTaS process is marginally Poisson.
Some definitions are required: first, for subsets A1, . . . , AN ∈ B(R) and D, D(cid:48) ⊂ D with D ⊂ D(cid:48),
let
M (D, D(cid:48); A1, . . . , AN ) := B1 × ··· × BN with Bi :=
Ai,
Ac
i ,
R,
for i ∈ D,
for i ∈ D(cid:48)\D,
otherwise
In addition, setting 1 = (1, . . . , 1) to be the N -dimensional vector with all components equal to
unity, and if QD is a measure on RN , then we define the measure ν(QD) by
ν(QD)(A) :=
for A ∈ B(RN )
QD(A − t1)dt
P (Y + t1 ∈ AY ∼ QD)dt.
(26)
(cid:90)
(cid:90)
=
The measure ν(QD) may be interpreted as giving the expected Lebesgue measure of the subset
L of R for which uniform shifts by the elements of L translate a random vector Y ∼ QD in to
A. Heuristically, one may imagine sliding the vector Y over the whole real line, and counting the
number of times every coordinate ends up in the "right" set -- the projection of A on to that
dimension. In equation form, this means
ν(QD)(A) = EY[(cid:96)({t ∈ R : Y + t1 ∈ A})Y ∼ QD] .
(27)
where the subscript Y indicates that we take the average over the distribution of Y ∼ QD. A
short proof of this representation is presented below. We now present the theorem, with a proof
indicating adjustments necessary to that of [11].
Theorem 0 Let X be an N -dimensional counting process of GTaS type with base rate λ, thinning
mechanism p = (pD)D⊂D, and family of shift distributions (QD)D⊂D. Then, for any Borel subsets
A1, . . . , AN of the real line, we have the following distributional representation:
(cid:88)
D(cid:48)
27
X1(A1)
...
XN (AN )
(cid:80)
(cid:80)
=distr
(cid:88)
,
D(cid:51)1 ξ(D; A1, . . . , AN )
...
D(cid:51)d ξ(D; A1, . . . , AN )
where the random variables ξ(D; A1, . . . , AN ),∅ (cid:54)= D ⊂ D, are independent and Poisson distributed
with
E[ξ(D; A1, . . . , AN )] = λ
pD(cid:48)ν(QD(cid:48))(M (D, D(cid:48); A1, . . . , AN )).
D(cid:48)⊃D
Proof. For each marking D(cid:48) ⊂ D, define XD(cid:48)
mother process rate λpD(cid:48), shift distribution QD(cid:48), and markings (pD(cid:48)
and is zero otherwise (i.e., the only possible marking for XD(cid:48)
to be an independent TaS [11] counting process with
D = 1 if D = D(cid:48)
D )D⊂D where pD(cid:48)
is D(cid:48)). We first claim that
X =distr
XD(cid:48)
.
(29)
(28)
To see this, note that spikes in the mother process of the GTaS process of X marked for a set D(cid:48)
occur at a rate λpD(cid:48), which is the rate of the process XD(cid:48)
. In addition, these event times are then
shifted by QD(cid:48), exactly as they are for XD(cid:48)
. Thus, the distribution of event times (and hence the
counting process distributions) are equivalent.
Let A1, . . . , AN be any Borel subsets of the real line. Applying Theorem 1 of [11] to each XD(cid:48)
gives the following distributional representation:
,
(D; A1, . . . , AN )
...
(D; A1, . . . , AN )
X D(cid:48)
...
1 (A1)
X D(cid:48)
N (AN )
=distr
(cid:80)
D(cid:51)1 ξD(cid:48)
(cid:80)
D(cid:51)N ξD(cid:48)
(cid:105)
= λpD(cid:48) (cid:88)
D(cid:48)(cid:48)⊃D
(cid:104)
ξD(cid:48)
E
(D; A1, . . . , AN )
D(cid:48)(cid:48)ν(QD(cid:48))(M (D, D(cid:48)(cid:48); A1, . . . , AN ))
pD(cid:48)
= λpD(cid:48)ν(QD(cid:48))(M (D, D(cid:48); A1, . . . , AN )).
where the random variables ξD(cid:48)
In the latter case, they are independent and Poisson distributed with
(D; , A1, . . . , AN ) are taken to be identically zero unless D ⊂ D(cid:48).
(30)
The second equality above follows from the fact that pD(cid:48)
D(cid:48)(cid:48) = 1 if D(cid:48)(cid:48) = D(cid:48) and is zero otherwise.
Next, define
ξ(D; A1, . . . , AN ) =
(cid:88)
D(cid:48)
(cid:88)
D(cid:48)⊃D
ξD(cid:48)
(D; A1, . . . , AN ) =
ξD(cid:48)
(D; A1, . . . , AN ).
As the sum of independent Poisson variables is again Poisson with rate equal to the sum of the
rates, we have that ξ(D; A1, . . . , AN ) is Poisson with mean
E[ξ(D; A1, . . . , AN )] = λ
pD(cid:48)ν(QD(cid:48))(M (D, D(cid:48); A1, . . . , AN )).
(31)
Finally, combining Eqs. (29, 30), we may write
X1(A1)
...
XN (AN )
D(cid:48)⊃D
D(cid:51)1 ξD(cid:48)
(cid:88)
=distr
(cid:80)
D(cid:48)(cid:80)
D(cid:48)(cid:80)
(cid:80)
(cid:80)
(cid:80)
D(cid:51)N ξD(cid:48)
D(cid:48) ξD(cid:48)
(cid:80)
(cid:80)
(cid:80)
D(cid:48) ξD(cid:48)
(cid:80)
D(cid:51)1
=
=
...
(D; A1, . . . , AN )
...
(D; A1, . . . , AN )
,
D(cid:51)N
D(cid:51)1 ξ(D; A1, . . . , AN )
D(cid:51)N ξ(D; A1, . . . , AN )
(D; A1, . . . , AN )
...
(D; A1, . . . , AN )
,
,
which, along with Eq. (31), establishes the theorem.
A short note: The variable ξ(D; A1, . . . , AN ) counts the number of points which are marked by
a set D(cid:48) ⊃ D, but after shifting, only the points attributed to the processes with indices i ∈ D
28
remain in the corresponding subsets Ai. Thus, to determine the number of points attributed to the
ith process which lie in Ai (Xi(Ai)), one simply sums the variables ξ for all D containing i, as in
Eq. (28). Thus, the intensity of ξ(D; A1, . . . , AN ),
λpD(cid:48)ν(QD(cid:48))(M (D, D(cid:48); A1, . . . , AN )),
is simply the expected number of such points. Keeping in mind these natural interpretations of
terms, Theorem 1 is easier to digest, and the result is not surprising.
5.2 Proof of Eq. (27)
In Eq. (27), we gave a more intuitive representation of the measure ν(QD) than the one first defined
in [11], which we prove here. Suppose that Q is a measure on B(Rd), and A ∈ B(Rd). Then we
have
(cid:90)
(cid:90)(cid:90)
(cid:90)(cid:90)
(cid:90)
ν(Q)(A) =
=
=
Q(A − t1)dt
1A−t1(y)Q(dy)dt
1{t∈R:y+t1∈A}(t)dtQ(dy)
(cid:96)({t ∈ R : y + t1 ∈ A})Q(dy)
=
= EY[(cid:96)({t ∈ R : Y + t ∈ A})Y ∼ Q] ,
thus proving Eq. (27)
5.3 Proof of Theorem 1
Theorem 1 Let X be a joint counting process of GTaS type with total intensity λ, marking
distribution (pD)D⊂D, and family of shift distributions (QD)D⊂D. Let A1, . . . , Ak be arbitrary sets
in B(R), and ¯D = {i1, . . . , ik} ⊂ D with ¯D = k. The cross-cumulant of the counting processes
may be written
κ(Xi1(A1), . . . , Xik (Ak)) = λ
pD(cid:48)
P (t1 + Y
¯D ∈ A1 × ··· × AkY ∼ QD(cid:48))dt
(32)
where Y ¯D represents the projection of the random vector Y on to the dimensions indicated by
the members of the set ¯D. Furthermore, assuming that the shift distributions possess densities
(qD)D⊂2D, the cross-cumulant density is given by
κX
i1···ik
(τ1, . . . , τk−1) = λ
pD(cid:48)
¯D
D(cid:48)(t, t + τ1,··· , t + τk−1)dt,
q
(33)
(cid:90)
(cid:88)
D(cid:48)⊃ ¯D
(cid:90)
(cid:88)
D(cid:48)⊃ ¯D
where q ¯D
D(cid:48) indicates the kth order joint marginal density of qD(cid:48) in the dimensions of ¯D.
Proof. First, as noted in the text, we may rewrite the distributional representation of Theorem 0
(Eq. (28)) as
Xi1(Ai1)
...
=distr
Xik (Aik )
(cid:80)
(cid:80)
i1∈D⊂ ¯D ζD(A1, . . . , AN )
...
ik∈D⊂ ¯D ζD(A1, . . . , AN )
29
(34)
where
ζD(A1, . . . , AN ) =
(cid:88)
D(cid:48)⊃D
( ¯D\D)∩D(cid:48)=∅
ξ(D(cid:48); A1, . . . , AN ).
(35)
Repeating the description from the main text, the processes ζD are comprised of a sum of all of the
processes ξ(D(cid:48)) (defined above, in Theorem 0) such that D(cid:48) contains all of the indices D, but no
other indices which are part of the subset ¯D under consideration. These sums are non-overlapping,
implying that the ζD are also independent and Poisson.
Using the representation of Eq. (34), we first find that
κ(Xi1(A1), . . . , Xik (Ak)) = κ
=
(cid:88)
(cid:88)
i1∈D1⊂ ¯D
ζD1 , . . . ,
··· (cid:88)
(cid:88)
ik∈Dk⊂ ¯D
ζDk
κ[ζD1, . . . , ζDk ].
i1∈D1⊂ ¯D
ik∈Dk⊂ ¯D
where we suppressed the dependence of the variables ζD on the subsets Ai. The first equality
in the previous equation is simply the representation defined in Eq. (35), and the second is from
the multilinear property of cumulants (property (C1) in the Methods). Note that the sums are
over the sets D1, . . . , Dk satisfying the given conditions. Recall that, by construction, the Poisson
processes ζD (see Eq. (35)) are independent for distinct marking sets. Accordingly, the cumulant
κ[ζD1, . . . , ζDk ] is zero unless D1 = . . . = Dk, by property (C2) of cumulants -- that is,
(cid:40)
κk(ζ ¯D(A1, . . . , AN )) Dj = ¯D for each j
0
otherwise
.
κ[ζD1(A1, . . . , AN ), . . . , ζDk (A1, . . . , AN )] =
Hence,
κ(Xi1(A1), . . . , Xik (Ak)) = κk(ζ ¯D(A1, . . . , AN )) = E[ζ ¯D(A1, . . . , AN )] ,
(36)
where we have again used that all cumulants of a Poisson-distributed random variable are equal to
its mean.
For what follows, taking D0, D(cid:48) ⊂ D fixed with D0 ⊂ D(cid:48), the sets M (D, D(cid:48); A1, . . . , AN ) with
D0 ⊂ D ⊂ D(cid:48) are disjoint, and
(cid:40)
Ai,
R,
i ∈ D0
i /∈ D0
.
(37)
∪D0⊂D⊂D(cid:48) M (D, D(cid:48); A1, . . . , AN ) = B1 × ··· × BN with Bi =
In particular, note the independence of the above union from D(cid:48).
Substituting Eq. (35) in to Eq. (36), we have
κ(Xi1(A1), . . . , Xik (Ak)) =
D⊃ ¯D
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
D⊃ ¯D
D(cid:48)⊃ ¯D
D(cid:48)⊃ ¯D
D(cid:48)⊃ ¯D
= λ
= λ
= λ
E[ξ(D; A1, . . . , Ak]
D(cid:48)⊃D
pD(cid:48)ν(QD(cid:48))(M (D, D(cid:48); A1, . . . , AN ))
(cid:88)
pD(cid:48) (cid:88)
pD(cid:48)ν(QD(cid:48))(∪ ¯D⊂D⊂D(cid:48)M (D, D(cid:48); A1, . . . , AN ))
(cid:90)
ν(QD(cid:48))(M (D, D(cid:48); A1, . . . , AN ))
¯D⊂D⊂D(cid:48)
= λ
pD(cid:48)
P (t + Y
¯D ∈ A1 × ··· × AkY ∼ QD(cid:48))dt,
30
where the third equality above is a simple exchange of the order of summation, the fourth equality
uses the independence of the inner union from the set D(cid:48) as indicated by Eq. (37), and the final
equality follows from the definition of the measure ν(QD(cid:48)) in Eq. (26) and the value of the set union
given in Eq. (37).
This completes the proof of Eq. (32), and Eq. (33) follows from the definition of the cross-
cumulant density in Eq. (24) of the Methods.
5.4 Other details
Parameters for figures in the text
Figure 1 For figure 1, the GTaS process of size N = 6 consisted of only first order and population-
level events which were assigned marking probabilities
0.05 D = D
0.95
6
0
pD =
D = {i} for some i ∈ D
otherwise
.
The rate of the mother process was λ = 0.5 kHz, and the shift times for population level events
were generated as in Section 2.2 with
ϕi ∼ Γ(2, 1) − 1,
i = 1, . . . , 6.
Figures 3, 4 For figures 3, 4, the GTaS process of size N = 6 consisted of first and second order
as well as population-level events. These events had marking probabilities
0.05 D = D
0.95
0
otherwise
pD =
21 D = {i},{i, j} for some i, j ∈ D
.
The rate of the mother process was λ = 0.5 kHz, and the shift times for population level events
were generated as in Section 2.2 with
ϕi ∼ Exp(0.5),
i = 1, . . . , 6.
The shift times of the second order events were drawn from an independent Gaussian distribution
with each coordinate having standard deviation 5ms.
Figure 5 For figure 5, the network parameters were win = 0.4, wsyn = 6, τsyn = 0.1, τd = 1.75.
The GTaS input had the same size as the network (N = 10). As in the example of figures 3, 4, the
GTaS input included first and second order as well as population level events. Here, we set
0.2 D = D
0.95
0
otherwise
pD =
5 D = {i},{i, j} for some i, j ∈ D
.
The rate of the mother process was λ = 1.5 kHz, and the shift times for population level events
were generated as in Section 2.2 with
ϕi ∼ Γ(α, β),
i = 1, . . . , 6.
31
The shift parameters k, θ (representing shape and scale) were determined by the given shift mean
µshift and standard deviation σshift as
√
µshift = kθ,
σshift =
kθ2.
The shift times of the second order events were drawn from an independent Gaussian distribution
with each coordinate having standard deviation 0.3ms.
5.4.1 Notation table
D
(pD)D⊂D
(QD)D⊂D
B(R)
D = {1, 2, . . . , N} where N is the system size of the GTaS process
under consideration
Marking probabilities of a GTaS process.
Family of shift distributions on RN for a GTaS process.
Borel subsets of the real line R.
ξ(D; A1, . . . , AN )
Independent Poisson variables which count points which, after shift-
ing, lie in the sets Ai only along the dimensions corresponding to
the indices of D. These counts consist of contributions from subsets
marked for D(cid:48) ⊃ D, but indices in D(cid:48)\D end up outside the corre-
sponding Ai. Defined in the statement of Theorem 0.
ζD(A1, . . . , AN )
Independent Poisson variables which are context-dependent resum-
mations of the variables ξ(D; A1, . . . , AN ). Defined below Eq. (10).
κ(X1, . . . , XN )
Cross-cumulant of the random variables X1, . . . , XN defined in the
Methods.
κX
i1···ik
(τ1, . . . , τk−1)
Cross-cumulant density defined in Eq. (24).
i1···ik−1,pop(τ1, . . . , τk−1) Population cumulant density defined in Eq. (25).
κX
Table 1: Common notation utilized in the text.
32
|
1411.7916 | 4 | 1411 | 2016-06-09T14:23:52 | The effect of heterogeneity on decorrelation mechanisms in spiking neural networks: a neuromorphic-hardware study | [
"q-bio.NC"
] | High-level brain function such as memory, classification or reasoning can be realized by means of recurrent networks of simplified model neurons. Analog neuromorphic hardware constitutes a fast and energy efficient substrate for the implementation of such neural computing architectures in technical applications and neuroscientific research. The functional performance of neural networks is often critically dependent on the level of correlations in the neural activity. In finite networks, correlations are typically inevitable due to shared presynaptic input. Recent theoretical studies have shown that inhibitory feedback, abundant in biological neural networks, can actively suppress these shared-input correlations and thereby enable neurons to fire nearly independently. For networks of spiking neurons, the decorrelating effect of inhibitory feedback has so far been explicitly demonstrated only for homogeneous networks of neurons with linear sub-threshold dynamics. Theory, however, suggests that the effect is a general phenomenon, present in any system with sufficient inhibitory feedback, irrespective of the details of the network structure or the neuronal and synaptic properties. Here, we investigate the effect of network heterogeneity on correlations in sparse, random networks of inhibitory neurons with non-linear, conductance-based synapses. Emulations of these networks on the analog neuromorphic hardware system Spikey allow us to test the efficiency of decorrelation by inhibitory feedback in the presence of hardware-specific heterogeneities. The configurability of the hardware substrate enables us to modulate the extent of heterogeneity in a systematic manner. We selectively study the effects of shared input and recurrent connections on correlations in membrane potentials and spike trains. Our results confirm ... | q-bio.NC | q-bio | The effect of heterogeneity on decorrelation mechanisms in spiking neural networks:
a neuromorphic-hardware study
Thomas Pfeil†,1, ∗ Jakob Jordan,2, ∗ Tom Tetzlaff,2 Andreas Grübl,1
Johannes Schemmel,1 Markus Diesmann,2, 3, 4 and Karlheinz Meier1
1Kirchhoff-Institute for Physics, Heidelberg University, Heidelberg, Germany
2Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation
(IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany
3Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, Aachen, Germany
4Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany
(Dated: June 10, 2016)
6
1
0
2
n
u
J
9
]
.
C
N
o
i
b
-
q
[
4
v
6
1
9
7
.
1
1
4
1
:
v
i
X
r
a
High-level brain function such as memory, classification or reasoning can be realized by means
of recurrent networks of simplified model neurons. Analog neuromorphic hardware constitutes a
fast and energy efficient substrate for the implementation of such neural computing architectures in
technical applications and neuroscientific research. The functional performance of neural networks
is often critically dependent on the level of correlations in the neural activity. In finite networks,
correlations are typically inevitable due to shared presynaptic input. Recent theoretical studies have
shown that inhibitory feedback, abundant in biological neural networks, can actively suppress these
shared-input correlations and thereby enable neurons to fire nearly independently. For networks of
spiking neurons, the decorrelating effect of inhibitory feedback has so far been explicitly demon-
strated only for homogeneous networks of neurons with linear sub-threshold dynamics. Theory,
however, suggests that the effect is a general phenomenon, present in any system with sufficient
inhibitory feedback, irrespective of the details of the network structure or the neuronal and synaptic
properties. Here, we investigate the effect of network heterogeneity on correlations in sparse, ran-
dom networks of inhibitory neurons with non-linear, conductance-based synapses. Emulations of
these networks on the analog neuromorphic hardware system Spikey allow us to test the efficiency
of decorrelation by inhibitory feedback in the presence of hardware-specific heterogeneities. The
configurability of the hardware substrate enables us to modulate the extent of heterogeneity in a
systematic manner. We selectively study the effects of shared input and recurrent connections on
correlations in membrane potentials and spike trains. Our results confirm that shared-input correla-
tions are actively suppressed by inhibitory feedback also in highly heterogeneous networks exhibiting
broad, heavy-tailed firing-rate distributions. In line with former studies, cell heterogeneities reduce
shared-input correlations. Overall, however, correlations in the recurrent system can increase with
the level of heterogeneity as a consequence of diminished effective negative feedback.
I.
INTRODUCTION
Dynamical systems in nature often exhibit a remark-
able degree of diversity, specialization or anticorrelation
across their components, despite equalizing factors such
as common input or homogeneity in component and in-
teraction parameters. In many cases, these observations
can be explained by the effect of negative feedback. Cell
differentiation caused by lateral inhibition [1], formation
of new species driven by competition [2] or antiferro-
magnetism [3] constitute just a few examples. In recur-
rent neuronal networks, inhibitory feedback constitutes a
powerful decorrelation mechanism which allows different
neurons to respond nearly independently, even if they are
driven by largely overlapping local or external inputs [4–
6]. Decorrelation by negative feedback hence implements
∗ These authors contributed equally to this study.
† Correspondence: Thomas Pfeil
Kirchhoff-Institute for Physics
Im Neuenheimer Feld 227
69120 Heidelberg, Germany
tel: +49-6221-549813
[email protected]
an efficient form of redundancy reduction. In biological
systems, it may serve similar purposes as decorrelation in
technical applications, where it is used in data compres-
sion (e.g., principal-component analysis [7]), crosstalk re-
duction (e.g., in digital signal processing [8]), echo sup-
pression (e.g., in acoustics [9]) or random-number genera-
tion in hardware [10]. Moreover, inhibitory feedback sup-
presses "quantization noise" at low frequencies and can
thereby increase the dynamical range and signal-to-noise
ratio for the encoding of analog signals in the spiking ac-
tivity of recurrent neural networks [11]. It is tempting to
exploit these mechanisms in synthetic, neurally inspired
architectures such as analog neuromorphic hardware.
Analog neuromorphic hardware mimics properties of
biological neural systems using physical models of neu-
rons and synapses (capacitors, for example, emulate in-
sulating cell membranes) [12, 13]. The temporal evolu-
tion of the analog circuits represents a solution to the
corresponding model equations. In consequence, neural-
network emulations on analog neuromorphic hardware
are massively parallel, extremely fast and energy efficient.
Analog neuromorphic devices are therefore highly attrac-
tive as tools for neuroscientific research, e.g., for the in-
vestigation of learning on long time scales, and technical
applications [14–17]. A biologically inspired neural net-
work (olfactory system of insects) performing rapid on-
line data (odor) classification, for example, has recently
been successfully implemented on the analog neuromor-
phic hardware system Spikey [18, 19].
In this applica-
tion, decorrelation by inhibition is an essential ingredi-
ent to guarantee high classification performance. The
suppression of quantization noise by inhibitory feedback
[11] has been used as a means of noise shaping in several
neuromorphic-hardware applications, aiming at the con-
struction of biologically inspired ultra-low power analog-
to-digital converters [? ? ].
For the functional performance of neuronal architec-
tures, the level of correlations between the activities of
individual neurons is often pivotal. Whether such cor-
relations are beneficial or not is context dependent. A
number of previous studies emphasize a functional bene-
fit of certain types of correlation for encoding/decoding of
information in/from populations of neurons [20–22], in-
formation transmission [23–25], robustness against noise
[26], or gain control of postsynaptic neurons [27]. Other
studies argue that positive cross-correlations are detri-
mental as they decrease the precision or sparseness of
population codes [19, 28–31]. Cohen & Maunsell [32], for
example, have shown that decreased spike-train correla-
tions in macaque visual area V4 are accompanied by in-
creased behavioral performance in an orientation change-
detection task. Depending on the similarity between the
trial-averaged responses of different neurons to external
stimuli (signal correlation), noise correlations (correla-
tions not explained by signal correlations) can either in-
crease or decrease the amount of information that can be
encoded in or decoded from a population of neurons. In
populations of neurons with high signal correlation, van-
ishing or even negative noise correlations are desirable to
improve the population code [21].
In finite neural networks, an inevitable source of corre-
lated neural activity is common presynaptic input, shared
by multiple postsynaptic neurons.
In network models
and in-vivo recordings, however, pairwise correlations in
the activity of neighboring neurons have been found to
be substantially smaller than expected given the amount
of shared input [4, 5, 33–36].
In several studies, this
observation has been explained by inhibitory coupling.
While Ly et al. [37] and Middleton et al. [38] primarily fo-
cused on the effect of feedforward inhibition, Renart et al.
[4], Wiechert et al. [39], and Tetzlaff et al. [5] attributed
the smallness of correlations to an active decorrelation of
neural activity by inhibitory feedback. The mechanism
underlying this active decorrelation has already been de-
scribed by Mar et al. [11].
In this study, the authors
focused on the suppression of low-frequency fluctuations
of the population firing rate by recurrent dynamics. As
the amplitude of population-rate fluctuations is directly
linked to pair-wise correlations (see, e.g., [40]), the effect
described in [11] corresponds to a suppression of pairwise
correlations in the spiking activity. The theory underly-
ing decorrelation by inhibitory feedback suggests the ef-
2
fect to be general: Decorrelation should be observable in
any system with sufficiently strong inhibitory feedback,
irrespective of the details of the network structure and
the cell and synapse properties. For networks of spik-
ing neurons, however, the effect has so far been explic-
itly demonstrated only for the homogeneous case, where
all neurons have identical properties, receive (approxi-
mately) the same number of inputs, and, hence, fire at
about the same rate [4, 5]. Moreover, the sub-threshold
dynamics of individual neurons was assumed to be linear.
Biological neuronal networks typically exhibit broad,
heavy-tailed firing-rate distributions [41–47], indicating a
high degree of heterogeneity, e.g., in synaptic weights [48–
52], in-degrees [53] or time constants [46, 54]. The same
holds for neural networks implemented on analog neuro-
morphic hardware. All analog circuits suffer from device
variations caused by unavoidable variability in the man-
ufacturing process. Neurons and synapses implemented
in analog neuromorphic hardware therefore exhibit het-
erogeneous response properties, similar to their biological
counterparts [55, 56]. To understand the dynamics and
function of recurrent neural networks in both biological
and synthetic substrates, it is therefore essential to ac-
count for such heterogeneities.
Previous work on recurrent neural networks has shown
that heterogeneity in single-neuron properties or connec-
tivity broadens the distribution of firing rates [46, 57]
and affects the stability of asynchronous or oscillatory
states [53, 58–62]. A number of studies pointed at a
potential benefit of heterogeneity for the information-
processing capabilities of neural networks [62–73]. The
effect of heterogeneity on correlations in the activity of
recurrent networks of spiking neurons, however, remains
unclear. Padmanabhan & Urban [67] have shown that
the responses of a population of unconnected neurons are
decorrelated by heterogeneity in the neuronal response
properties. These results are supported by the subse-
quent theoretical analysis in [70]. In the following, we re-
fer to this type of decorrelation by heterogeneity as feed-
forward decorrelation. It does not account for the effect
of the recurrent network dynamics. Active decorrelation
due to inhibitory feedback [see above; 4, 5], in contrast,
constitutes a very different mechanism. The effect of het-
erogeneity on this feedback decorrelation has lately been
studied by Bernacchia & Wang [74] in the framework of
a recurrent network of linear firing-rate neurons. In this
setup, correlations are suppressed by heterogeneity in the
network connectivity (distributions of coupling strengths
or random dilution of connectivity). It remains unclear,
however, whether this holds true for networks of (nonlin-
ear) spiking neurons.
In this study, we investigate the impact of hetero-
geneity on input and output correlations in the asyn-
chronous regime of sparse networks of leaky integrate-
and-fire (LIF) neurons with conductance-based synapses.
Emulation of the networks on the analog neuromorphic
hardware system Spikey (Figure 1) [18, 75] enable us to
investigate the impact of substrate specific properties on
3
difficult to achieve with hardware studies alone. In anal-
ogy to the necessity of performing experiments on biolog-
ical neural systems to verify assumptions made in Com-
putational Neuroscience, actual emulations on neuromor-
phic hardware are essential to understand its properties
and develop efficient neural algorithms for these devices.
The fact that our main findings hold true for both emu-
lations on hardware and simulations with software, and
that they can be distilled to simple linear models, sup-
port their broad relevance and robustness.
II. METHODS
A. Network model
Details on the network, neuron and synapse model are
provided in Table I. Parameter values are given in Ta-
ble II. Briefly: We consider a purely inhibitory, sparse
network of N (N = 192, unless stated otherwise) LIF
neurons with conductance-based synapses. Each neuron
receives input from a fixed number K = 15 of randomly
chosen presynaptic sources, independently of the network
size N. Self-connections and multiple connections be-
tween neurons are excluded. Resting potentials El are
set above the firing thresholds Θ (equivalent to applying
a constant supra-threshold input current). We thereby
ensure autonomous firing in the absence of any further
external input. Due to temporal noise, the initial condi-
tions are essentially random.
B. Network emulations on the
neuromorphic-hardware system Spikey
The Spikey chip (Figure 1) consists of physical models
of LIF neurons and conductance-based synapses with ex-
ponentially decaying dynamics (for details, see Table I).
The emergent dynamics of these physical models repre-
sents a solution for the model equations of neurons and
synapses in continuous time, in parallel for all units. In
contrast, in classical simulations on von-Neumann archi-
tectures, model equations are solved by step-wise numer-
ical integration, where parallelization is limited by the
available number of processor cores. To emphasize the
difference between simulations using software and simu-
lations using physical models, the term emulation is used
for the latter [18].
The response properties of physical neurons and
synapses vary across the chip due to unavoidable vari-
ations in the production process that manifest in a spa-
tially disordered pattern (fixed-pattern noise).
In con-
trast to the approximately static fixed-pattern noise,
temporal noise, including electronic noise and transient
experiment conditions (e.g., chip temperature), impairs
the reproducibility of emulations.
In general, two net-
work emulations with identical configuration and stimu-
lation do not result in identical network activity. Both
The neuromorphic hardware system Spikey. (a)
FIG. 1.
Photograph of the Spikey chip (size 5 × 5 mm2). It comprises
analog circuits of 384 neurons and 98304 synapses, is highly
configurable and emulates neural-network dynamics with a
speed-up of 104 with respect to biological real-time. (b) Pho-
tograph of the partly cased Spikey system, carrying the Spikey
chip (covered by a black round seal) and conventional mem-
ory. The system is connected to the host computer via USB
2.0, consumes 6 W of power in total and less than 1 nJ per
synaptic transmission (see Supplements 1).
the network dynamics. Insights about the interplay be-
tween features of the computing substrate and network
dynamics are a necessary prerequisite for the develop-
ment of algorithms that exploit the benefits of analog
neuromorphic systems at best. The configurability of this
system [18] enables us to systematically vary the level of
heterogeneity, and to disentangle the effects of hetero-
geneity on feedforward and feedback decorrelation (see
above). For simplicity, we focus on purely inhibitory net-
works, thereby emphasizing that active decorrelation by
inhibitory feedback does not rely on a dynamical balance
between excitation and inhibition [5, 6]. We show that
decorrelation by inhibitory feedback is effective even in
highly heterogeneous networks with broad distributions
of firing rates (Section III A). Increasing the level of het-
erogeneity has two effects: Feedforward decorrelation is
enhanced, feedback decorrelation is impaired. Due to the
latter, the overall input and output correlations do not
necessarily become smaller with increasing heterogeneity.
They can even increase (Section III B).
Note that results from specific network emulations on
hardware do not directly translate to those obtained by
simulations on conventional computers, because the dy-
namics, parametrization and interplay of analog circuits
is very complex and difficult to reproduce with classical
simulations.
If simplified models for spatial and tem-
poral variability are considered in software simulations,
however, emulation results can be reproduced qualita-
tively, thereby verifying the design of the hardware sys-
tem. While our hardware system is designed to physi-
cally implement biologically realistic neural algorithms in
a fast and energy-efficient way, software simulations are
used as a complementary tool to isolate, verify and in-
vestigate different hardware features, such as spatial and
temporal parameter variations. Due to the limited access
and configurability of network parameters, this would be
ab fixed-pattern and temporal noise need to be taken into
account when developing models for analog neuromor-
phic hardware.
The key features of the Spikey chip are the high ac-
celeration and configurability of the analog network im-
plementation. Some network parameters, e.g., synaptic
weights and leak conductances, are configurable for each
unit, while other parameters are shared for several units
(for details see [18]). The hardware system is optimized
for spike in- and output and allows to record the mem-
brane potential of one (arbitrarily chosen) neuron with a
sampling frequency of 96 MHz in hardware time. On the
Spikey chip, capacitances are smaller and conductances
are much higher than in biological nervous systems. In
consequence, networks on the Spikey chip are emulated
with a speed-up of approximately 104 with respect to
biological real-time. Due to this high acceleration of the
neuromorphic chip, the data bandwidth of the connection
between the neuromorphic system and the host computer
is not sufficient to communicate with the chip in real
time. Consequently, input and output spikes (for stimu-
lation and from recordings, respectively) are buffered in
a local memory next to the chip. The high acceleration of
the Spikey chip allows most of the transistors to operate
outside of weak inversion, thereby reducing the effect of
transistor variations and minimizing fixed-pattern noise.
In contrast to such accelerated systems, most other
configurable, analog neuromorphic substrates are de-
signed for real-time emulations at very low power con-
sumption [76–82] and implement fewer, but more com-
plex, neurons [83, 84].
Access to the Spikey system is encapsulated by the
simulator-independent language PyNN [85, 86], providing
a stable and user-friendly interface. PyNN integrates the
hardware into the computational neuroscience tool chain
and has facilitated the implementation of several network
models on the Spikey chip [18, 19, 87–89].
On the Spikey system, a spiking neural network is emu-
lated as follows (Figure 2a): First, the network described
in PyNN is mapped to the Spikey chip, i.e., neurons and
synapses are allocated and parametrized. Second, input
spikes, if available, are prepared on the host computer
and transferred to the local memory on the hardware
system. Third, the emulation is triggered and available
input spikes are generated. Output spikes and mem-
brane data are recorded to local memory. Last, spike and
membrane data are transferred to the host computer and
scaled back into the biological domain of the PyNN model
description.
For consistency with the model description and sim-
plified comparison to the existing literature, all hardware
times and all hardware voltages are expressed in terms
of the quantities they represent in the neurobiological
model, throughout this study.
C. Experimental setup
4
To differentiate and compare the effects of shared in-
puts and feedback connections on correlations, we inves-
tigate two different emulation scenarios: First, we emu-
late networks with intact feedback (FB, Figure 2b), and
second, the contribution of shared input is isolated by
randomizing the temporal order of this feedback (RAND,
Figure 2d).
In the RAND scenario, the inputs of neurons are de-
coupled from their outputs. Spatio-temporal correlations
in presynaptic spike trains are removed by randomizing
the presynaptic spike times.
Input correlations between neurons are measured via
their free membrane potential, i.e., the membrane poten-
tial with disabled spiking mechanism (technically, the
threshold is set very high). Because membrane poten-
tial traces can be recorded in the hardware only one at
a time, traces are obtained consecutively, while repeat-
edly replaying the previously recorded activity of the FB
network to a population of unconnected neurons of equal
size. We keep the connectivity the same, and hence each
neuron receives the same number of spikes as in the recur-
rent network during the whole emulation, either without
(FBreplay, Figure 2c) or with randomization of presynap-
tic spike times (RAND, Figure 2d), respectively. To pre-
serve the fixed pattern of variability of synaptic weights
in hardware, the same hardware synapses are used for
each connection in both scenarios.
If network dynam-
ics were reproduced perfectly, membrane potential traces
and spike times would be identical in the FB and FBreplay
cases (see also Section II D).
Drawing two different network realizations (i.e., the
connectivity matrix) results in the allocation of different
hardware synapses, and, due to fixed-pattern noise, in
different values of synaptic weights. To average over this
variability, throughout this study, emulation results are
averaged over M = 100 network realizations, if not stated
otherwise.
D. Reproducibility of hardware emulations
Since the initial conditions of the recurrent network on
hardware are undefined, consecutive emulations of the
FB network result in different network activities. In the
RAND and FBreplay case, however, the input of neurons
is decoupled from their output. Although unavoidable
temporal noise is present, the system's state-space trajec-
tory returns to the trajectory of the previously recorded
FB case. A certain degree of reproducibility is required
for two reasons: First, the investigated effect of decor-
relation by inhibitory feedback requires a precise rela-
tion between spike input and output. Thus our method
of replacing the feedback loop by replay is only valid if
temporal noise does not substantially corrupt this rela-
tionship. Second, to record the membrane potentials of
all neurons, as if recorded at once, neuron dynamics have
5
FIG. 2. Experimental setup. (a) Data flow of the Spikey system. For details see Section II B. (b) Network with on-chip
feedback connections (FB). Spikes from all neurons are recorded to the local memory. (c) Spikes of the FB network in (b)
replayed from memory via off-chip spike sources ξi to neurons i (FBreplay). Spike times of ξi correspond to those recorded from
neuron i in (b). Spikes from all neurons or the free membrane potential of one selected neuron are recorded. (d) Like (c), but
spike times from (b) are randomized for each source (RAND).
to be reasonably similar in consecutive emulations.
We measure the reproducibility of neuron dynamics by
comparing consecutive emulations with identical config-
uration, i.e., connectivity and stimulation. For this pur-
pose the spiking activity of a FB network is first recorded
(Figure 2b) and then repeatedly replayed (Figure 2c).
Reproducibility is quantified by the correlations (κX in
Table III) of free membrane potential traces and output
spike trains obtained for individual neurons in L = 25
different trials.
Free membrane potentials are reproduced quite well,
while spike trains show larger deviations across trials
(Figure 3). Small deviations in the membrane poten-
tial (Figure 3b) are amplified by the thresholding proce-
dure [39, 90, 91] and can lead to large differences between
spike trains (Figure 3c). Consequently, measures based
on data of several consecutive replays are more precise for
membrane potentials than for spike trains. Nevertheless,
results have to be interpreted with care in both cases.
E. Calibration
The heterogeneity of the Spikey hardware is adjusted
by calibrating the leak conductance1 for each individ-
ual neuron, compensating for fixed-pattern noise of neu-
ron parameters. To this end, a population of uncon-
nected neurons is driven by a constant supra-threshold
current and the time-averaged population activity ¯r is
measured. Then, we applied the bisection method [92] to
adjust the leak conductance gl of each neuron, such that
the neuron's firing rate matches the target rate ¯r. This
results in calibration values b for the leak conductance
gl = gl,0(1 + b), where gl,0 is the leak conductance before
1 since capacitances and potentials can not be configured individ-
ually for each hardware neuron [18]
calibration. Because emulations on hardware are not per-
fectly reproducible, more precise calibration was achieved
by evaluating the median over 25 identically configured
trials instead of single trials. Furthermore, the bisection
method was modified for noisy systems (for details, see
Supplements 2).
Intermediate calibration states are obtained by linearly
scaling the full calibration:
gl = gl,0(1 + (1 − a)b)
.
(1)
The heterogeneity a is chosen in [0, 1] for calibrations
between the uncalibrated (a = 1) and calibrated state
(a = 0). In the following, the fully calibrated chip (a = 0)
is used, if not stated otherwise.
This calibration substantially narrows the distribution
of firing rates compared to the uncalibrated state (Fig-
ure 4). With respect to the stationary firing rate, vari-
ability on the neuron level is reduced from 35.1 to 0.9 s−1.
Even in the fully calibrated state, leak conductances
can still be widely distributed. Due to the chosen cal-
ibration procedure, they are likely to be correlated to
other parameters that influence the neurons' response to
a constant supra-threshold current after calibration. This
mutual compensation can lead to similar phenomenology
(here: firing rates) despite disparate parameter values,
similar to what is observed in biology [93]. In addition
to remaining variations in neuron parameters, synaptic
parameters are significantly distributed [19, 94].
F. Correlation measures
In the following, we introduce definitions used to ana-
lyze the recorded data. For clarity, all relevant equations
and their parametrization are listed in Table III and IV,
respectively.
We quantify correlations of membrane potentials vi(t)
and spike trains si(t) by the population-averaged low-
6
FIG. 4. Calibration of the Spikey chip. (a) Histogram of
firing rates r for a population of unconnected neurons with
supra-threshold input currents, before (gray) and after (black)
calibration, each neuron averaged over L = 100 trials. The
arrow denotes the target rate ¯r. (b) Difference ∆r = rP 75 −
rP 25 of 75th and 25th percentile of the histograms in (a), as a
function of network heterogeneity a (Equation 1). The mean
firing rate over all values of a is (73.4 ± 0.3) s−1.
tials, and output correlations for correlations between
spike trains. Shared-input correlations are membrane-
potential correlations that are exclusively caused by over-
ignoring possible correla-
lapping presynaptic sources,
tions in the presynaptic activity.
In homogeneous net-
works, the average pairwise shared-input correlation
κV =
K
N
(2)
is given by the connectivity K/N [5]. In heterogeneous
networks, shared-input correlations can be reduced. In
the presence of heterogeneous synaptic weights, for ex-
ample, the shared-input correlation
κV =
1
1 + CV 2
J
K
N
(3)
J ), where CVJ de-
is decreased by a factor 1/(1 + CV 2
notes the coefficient of variation of the (non-zero) synap-
tic weights. Note, however, that heterogeneities which
affect only the spike generation but not the integration
of synaptic inputs, e.g. distributions of firing thresholds,
have no effect on the shared-input correlation.
We assess the significance of correlations by compar-
ing the results from emulations to correlations in sur-
rogate data, in which we removed spatial correlations.
For every neuron, we randomly shuffled bins of the mem-
brane potential trace, and assigned a new timestamp uni-
formly drawn from the emulation interval to every spike,
respectively. We thereby remove all spatio-temporal cor-
relations between neurons recorded in parallel. By this
procedure we create 100 surrogate trials, across which we
calculate the average correlations and the standard error.
To quantify fluctuations in the population activity ¯s
(Figure 5a–c, horizontal histograms) we compute the
power spectrum ¯A(f ) of the population activity (Fig-
i (t) and sl
FIG. 3.
Reproducibility of free membrane potentials and
spiking activity in the FBreplay case. (a) Low-frequency co-
i (t) and
herence κV and κS of free membrane potentials vk
i(t), respectively, for
i(t) and binned spike trains sk
vl
each neuron i averaged over L = 25 trials k, l with k (cid:54)= l, for
M = 50 different network realizations. The diamond marks
the average across all neurons i and M network realizations
(b) Free single-trial membrane
(κV = 0.96, κS = 0.72).
potentials vk
i (t)
(black), and (c) spike density ξi(t) of a single neuron i for
L = 25 identical trials. The selected neuron i has membrane
potential coherence and spike train coherence closest to the
diamond in (a).
i (t) (gray) and average over trials 1
(cid:80)L
k=1 vk
L
frequency coherence κV and κS, respectively. At fre-
quency zero, the coherence corresponds to the normal-
ized integral of the cross-covariance function, i.e., it mea-
sures correlations on all time scales. We define the low-
frequency coherence κX, with X ∈ {S, V }, to be the aver-
age coherence over a frequency interval from 0.1 to 20 Hz.
In this interval, the suppression of population-rate fluc-
tuations in recurrent networks due to inhibitory feedback
is most pronounced, and the coherence is approximately
constant. Before calculating the coherence, we convolve
the power- and cross-spectra with a rectangular window
to average out random fluctuations. This measure, or
a variant of it, is commonly used in the neuroscientific
literature [4, 5, 70, 91, 95–97]. We use the terms low-
frequency coherence and correlation interchangeably.
Throughout this study, the term input correlations
is used for correlations between free membrane poten-
0.60.81.0Free membrane potential corr. κV0.20.40.60.8Spike-train corr. κS-74.0-70.0-66.0-62.0-58.0vki(t) (mV)single trialtrial average450046004700480049005000Time t (ms)04812162024Trial index k10-410-310-2Rel. frequencybca050100150Rate r(s−1)050100150200Neuron counta=1a=00.00.51.0Heterogeneity a010203040∆r (s−1)ab 7
FIG. 5. Typical spiking and membrane-potential activity of a random inhibitory network of LIF neurons with intact and
cut feedback loop emulated on the fully calibrated system. (a–c) Spiking activity (raster plots), population activity ¯s(t)
(horizontal histograms; bin size 50 ms) and time-averaged single-neuron firing rates ri (vertical histograms) in the network
with intact feedback (a) and for cases where the feedback loop is cut (b and c). (a) Intact recurrent network (FB scenario).
(b) Population of mutually unconnected neurons receiving input spike trains identical to those in (a) (FBreplay scenario). (c)
As in (b), but after randomization of presynaptic spike times (RAND scenario). (d,e) Population-averaged cross-correlation
functions c(τ ) (after offset subtraction) of pairs of spike trains (d) and power spectra ¯A(f ) (e; log-log representation) of the
population activity ¯s(t) (cf. horizontal histograms in (a–c)) for the FB (dark gray), FBreplay (black) and RAND scenario (light
gray). Inset in (e): Population-averaged power spectra A(f ) of individual single-cell spike trains (same scales as in main panel).
Correlation functions and spectra are averaged across M = 100 network realizations. (f) Membrane potential of a neuron in
the RAND scenario (with firing rate of 23.20 s−1 close to population average of 23.24 s−1; see black arrow in (c)) with intact
(black curve) and removed threshold (gray curve; free membrane potential). The threshold potential is marked by the horizontal
dashed line. The time frame corresponds to the gray-shaded region in (c).
ure 5e), which we scale with the duration T of the em-
ulation. Consequently, the population power spectrum
¯A(f ), scaled by the population size, coincides with the
time-averaged population activity ¯r for high frequencies:
limf→∞ 1
N
¯A(f ) = ¯r [35].
As a measure of pairwise correlations in the time do-
main (Figure 5d), we compute the population-averaged
cross-correlation function c(τ ) by Fourier transforming
the population-averaged cross-spectrum C(f ) to time do-
main.
III. RESULTS
In this study, we investigate the roles of shared in-
put, feedback and heterogeneity on input and output
correlations in random, sparse networks of inhibitory LIF
neurons with conductance-based synapses (Table I), im-
182430050100150420043004400450046004700Time t (ms)85756555Mem. potential vi (mV)without thresholdwith threshold200030004000500060007000Time t (ms)1824300501001506040200204060Time lag τ (ms)0.020.010.000.010.02Correlation c(s−1)FBFBreplayRAND100101102Frequency f (Hz)10-1100101102Power spectrum ¯A(s−1)A(s−1)182430¯s(s−1)03060ri(s−1)050100150Neuron index iabcdef plemented on the analog neuromorphic hardware chip
Spikey (Figure 1). Similarly to [5], we separate the con-
tributions of shared input and feedback by studying dif-
ferent network scenarios (Figure 2): In the FB case, we
emulate the recurrent network with intact feedback loop
(Figure 2b) and record its spiking activity (Figure 5a).
In the FBreplay case (Figure 2c), the feedback loop is
cut and replaced by the activity recorded in the FB net-
work. Ideally, the input to each neuron in the FBreplay
case should be identical to the input of the correspond-
ing neuron in the FB network. As the replay of spikes
and the resulting postsynaptic currents and membrane
potentials are not perfectly reproducible on the Spikey
chip, the neural responses in the FB and in the FBreplay
scenario are slightly different (compare Figures 5a and
5b).
In the RAND case (Figures 2d and 5c), we use
the same setup as in the FBreplay case. However, the
spike times in each presynaptic spike train are random-
ized. While the average presynaptic firing rates and the
shared-input structure are exactly preserved in this sce-
nario, the spatio-temporal correlations in the presynaptic
spiking activity are destroyed.
Using this setup, we first demonstrate in Section III A
that active decorrelation by inhibitory feedback [4, 5]
is effective in heterogeneous networks with conductance-
base synapses over a range of different network sizes. In
Section III B, we show that decreasing the level of het-
erogeneity by calibration of hardware neurons leads to an
enhancement of this active decorrelation.
A. Decorrelation by inhibitory feedback
The time-averaged population activities in the FB,
FBreplay and RAND scenarios are roughly identical (ver-
tical histograms in Figures 5a–c; see also high-frequency
power in Figure 5e).
In the FB and FBreplay scenario,
fluctuations in the population-averaged activity are small
(horizontal histograms in Figure 5a and b). The removal
of spatial and temporal correlations in the presynaptic
spike trains in the RAND case leads to a significant in-
crease in the fluctuations of the population-averaged re-
sponse activity (horizontal histogram in Figure 5c). At
low frequencies (≤ 20 Hz), the population-rate power in
the FB and in the RAND case differs by about two orders
of magnitudes (dark and light gray curves in Figure 5e).
This increase in low-frequency fluctuations in the RAND
case is mainly caused by an increase in pairwise correla-
tions in the spiking activity (Figure 5d; the power spec-
tra of individual spike trains [inset in Figure 5e] are only
marginally affected by a randomization of presynaptic
spike times) [5].
In other words, shared-input correla-
tions, i.e., those leading to large spike-train correlations
in the RAND scenario, are efficiently suppressed by the
feedback loop in the FB case.
On the neuromorphic hardware, the replay of net-
work activity is not perfectly reproducible (Section II D).
While the across-trial variability in membrane potentials
8
is small, postsynaptic spikes are dithered by few millisec-
onds (Figure 3). In the FBreplay case, the suppression of
shared-input correlations by correlations in presynaptic
spike trains is slightly less efficient as compared to the in-
tact network (FB). The differences in the population-rate
power spectra and in the spike-train correlations between
the FBreplay and RAND case, respectively, are neverthe-
less substantial (solid black and light gray curves in Fig-
ure 5d and e; note the logarithmic scale; for a detailed
investigation of spike dither see Supplements 4 and Sup-
plements Figure 7).
Note that the suppression of correlations and, hence,
population-rate fluctuations by inhibitory feedback is re-
stricted to low frequencies (here, to frequencies < 50 Hz;
see Figure 5e). In the remainder of this study, we will
quantify pairwise correlations by the low-frequency co-
herence in the range 0.1–20 Hz (see Section II F). At
higher frequencies, the population-rate power spectra in
the FB, FBreplay and RAND case are similar.
In Fig-
ure 5e, the peaks at ∼ 50 Hz and higher harmonics re-
sult from the single-cell spike-train statistics (they are
also visible in the single-cell spectra; see inset): A large
fraction of cells, in particular those firing at higher rates,
generate regular spike trains with low inter-spike-interval
(ISI) variability (cf. Figure 6d–f). These (fast spiking)
cells contribute maxima to the spike-train spectra at fre-
quencies close to their firing rates (and higher harmon-
ics). The structure of the population-rate spectra at
higher frequencies (≥ 50 Hz) is reproduced using surro-
gate data where the ISI distributions of the individual
neurons (and, hence, their firing rates and ISI variabil-
ity) are preserved, but serial ISI correlations and cross-
correlations between spike trains are destroyed (data not
shown).
In the RAND case, presynaptic spike-train correlations
free-membrane-
were removed, and hence input (i.e.,
potential) correlations are exclusively determined by the
number of shared presynaptic sources (Equation 2).
If
the in-degree K is fixed, input correlations will decrease
with network size N (Equation 2, light gray curve and
symbols in Figure 7a). In purely inhibitory networks with
intact feedback loop (FB scenario), correlations in presy-
naptic spike trains are on average significantly smaller
than zero (dark gray diamonds in Figure 7b, [5]), and
largely cancel the positive contribution from shared-input
correlations. Average input correlations are therefore sig-
nificantly reduced (black symbols in Figure 7a). As both
shared-input and spike-train correlations scale with the
inverse of the network size (N−1; light gray curve in Fig-
ure 7a and inset in Figure 7b, respectively) [91], this sup-
pression of correlations in the FB (and FBreplay) case is
observed for all investigated network sizes N. Note that
output correlations are negative even though input cor-
relations are positive. This effect is predicted by theory
and also observed in linear network models as well as
LIF-network simulations on conventional computers (see
Figure 9, Supplements 4, Supplements 5 and Section IV).
9
FIG. 6. Modulation of network heterogeneity by leak-conductance calibration (see Section II E). Input (top row) and firing
statistics (bottom row) in the intact recurrent networks (FB scenarios) for fully calibrated (a and d; a = 0), partially calibrated
(b and e; a = 0.625) and uncalibrated neurons (c and f; a = 1). (a–c) Effect of calibration on input statistics. Distributions of
relative mean input D = (¯v − Θ)/σ(v) (distance of time averaged free membrane potential ¯v from firing threshold Θ in units
of the standard deviation σ(v)) across the population of neurons. Gray areas in (a), (b) and (c) highlight [−3, 3] intervals,
containing 74%, 53% and 42% of the total mass of the distribution, respectively. Inset in (a): Distributions of free membrane
potentials v for three neurons α, β and γ with D = −3, D = 0 and D = 3 (arrows in (a)), respectively. Dotted lines mark
threshold potentials that may vary due to fixed-pattern noise. (d–f) Effect of calibration on spike-train statistics. Joint
(scatter plots) and marginal distributions of single-neuron firing rates r (horizontal histograms; log-linear scale) and coefficients
of variation CVISI of inter-spike intervals (vertical histograms; log-linear scale). Dashed lines mark mean of firing rate (22.6 s−1,
28.7 s−1, 34.8 s−1) and CVISI distributions (0.35, 0.28, 0.25), respectively. Gray bars (bottom panels) represent fractions of
silent neurons. Data obtained from M = 50 different network realizations. Percentage of dead neurons: 8%, 26%, 37%.
FIG. 7. Dependence of population-averaged input correlations (a) and spike-train correlations (b) on the network size N, for
the intact network (FB, dark gray diamonds), the FBreplay (black circles) and the RAND (light gray circles) case (fixed in-degree
K = 15). Symbols and error bars denote mean and one standard deviation, respectively, across M = 100 network realizations
(error bars are partly covered by markers). The gray curve in (a) depicts shared-input correlations in a homogeneous network
(Equation 2). The inset in (b) shows a magnified view of the spike-train correlations in the FB case (dark gray diamonds)
with a power-law fit ∼ N−1 (dark gray curve). The light gray horizontal band represents mean ± three standard deviations
of (spurious) correlations in surrogate data where correlations were removed. Note that free membrane potentials cannot be
recorded in the FB case (see Section II). Hence, there are no gray diamonds in (a).
04080120Rate r (s−1)0.00.51.01.52.02.5CVISI0408012004080120201001020Rel. mean input D0.000.050.100.15Rel. neuron countαβγ201001020201001020αβ756555vγdefca=1ba=0.625aa=096112128144160176192Network size N0.000.040.080.12Correlation κSSpike trains96112128144160176192Network size N0.000.040.080.120.16Correlation κVFree membrane potentialsRANDFBreplayFB0.0080.0060.004ab B. Effect of heterogeneity on decorrelation
In neural networks implemented in analog neuromor-
phic hardware, neuron and synapse parameters vary sig-
nificantly across the population of cells (fixed-pattern
noise; see Section II B). For a population of mutually
unconnected neurons with distributed parameters, injec-
tion of a constant (supra-threshold) input current leads
to a distribution of response firing rates (Figure 4). In
this study, we consider the width of this firing-rate distri-
bution as a representation of neuron heterogeneity. It is
systematically varied by calibration of leak conductances.
The extent of heterogeneity is quantified by the calibra-
tion parameter a (a = 1 and a = 0 correspond to the un-
calibrated and the fully calibrated system, respectively;
for details, see Section II E). For an unconnected popu-
lation of neurons subject to constant input, the width of
the firing-rate distribution increases monotonically with
a.
As shown in Figure 6, the level of heterogeneity (i.e.,
the calibration state a) is clearly reflected in the activ-
ity of the recurrent network (FB case). Both the width
of the distribution of mean free membrane potentials
(Figure 6a–c) as well as the width of the firing-rate dis-
tribution increases with a (Figure 6d–f; horizontal his-
tograms). In the uncalibrated system (a = 1), a substan-
tial fraction of neurons is predominantly driven by con-
stant supra-threshold input currents and therefore gen-
erates highly regular spike trains (CVISI ≈ 0) with high
firing rates (r > 60 s−1). Simultaneously, about 37% of
the neurons are silent (r = 0 s−1). Neurons with inter-
mediate firing rates (0 s−1 < r < 15 s−1), however, show
quite irregular activity (CVISI > 0.5). After calibration,
the firing-rate distribution is narrowed. For a = 0, the
fraction of silent neurons is reduced to about 8%. Max-
imum rates are limited to < 66 s−1. Note that our cal-
ibration routine compensates only for the distribution
of neuron parameters, but not for the heterogeneity in
synapse properties (synaptic weights, synaptic time con-
stants; see Section IV). Even for the fully calibrated net-
work (a = 0), the firing-rate distribution is therefore still
broad. In the RAND case, we obtain similar firing-rate
and ISI statistics as in the FB case (Supplements 6).
For all levels of heterogeneity attainable by our calibra-
tion procedure (a ∈ [0, 1]), input and output correlations
are significantly suppressed by the recurrent-network dy-
namics (cf. black and dark gray vs. light gray symbols
in Figure 8). In a homogeneous, random (Erdős-Rényi)
network with fixed in-degree K and linear sub-threshold
dynamics, the contribution of shared input to the input
(free-membrane-potential) correlation is given by the net-
work connectivity K/N [Equation 2; reference 5] (thin
light gray curves in Figure 7a and Figure 8a). Nonlinear-
ities in synaptic and/or spike-generation dynamics [91] as
well as heterogeneity in neuron (and synapse) parameters
lead to a suppression of this contribution (Equation 3,
[67]). Here, we refer to this type of decorrelation as feed-
forward decorrelation. In fact, in our setup the input and
10
spike-train correlations in the RAND case decrease with
increasing heterogeneity (light gray symbols in Figure 8).
Even in the fully calibrated case input correlations are
slightly smaller than K/N (gray symbols vs. thin gray
curve in Figure 8a) due to remaining heterogeneities.
Spike-train correlations decrease slightly faster with in-
creasing heterogeneity than input correlations. These ob-
servations indicate that both the synaptic integration and
spike generation are affected by heterogeneities on hard-
ware. To illustrate the different effects of heterogeneity in
synaptic integration and spike generation, we performed
network simulations on conventional computers where we
distributed either firing thresholds (Figure 9) or synaptic
weights (Supplements Figure 4). In the RAND case, the
decrease of input correlations on the hardware can be at-
tributed to an increase in heterogeneity in parameters af-
fecting synaptic integration (compare light gray symbols
in Figure 8a and Supplements Figure 4a). Heterogene-
ity in spike thresholds, in contrast, does not affect input
correlations in the RAND scenario, but strongly reduces
spike-train correlations (light gray symbols in Figure 9).
Overall, feedforward decorrelation, i.e. the suppression of
correlations in the RAND case, becomes more effective
in networks with heterogeneous cell parameters.
Despite this enhancement of feedforward decorrelation,
input and output correlations increase with the level of
heterogeneity in the presence of an intact feedback loop
(FB and FBreplay scenarios; black and dark gray symbols
in Figure 8). We attribute this effect to a weakening of
the effective feedback loop in the recurrent circuit: In het-
erogeneous networks with broad firing-rate distributions,
neurons firing with low or high rates, corresponding to
mean inputs far below or far above firing threshold (see
Figure 6a–c), are less sensitive to input fluctuations than
moderately active neurons (see Supplements Figure 2).
Hence, they contribute less to the overall feedback. In
consequence, feedback decorrelation is impaired by het-
erogeneity (see also Section IV).
IV. DISCUSSION
We have shown that inhibitory feedback effectively
suppresses correlations in heterogeneous recurrent neu-
ral networks of leaky integrate-and-fire (LIF) neurons
with nonlinear subthreshold dynamics, emulated on ana-
log neuromorphic hardware (Spikey; [18, 75]). Both in-
put and output correlations are substantially smaller in
networks with intact feedback loop (FB), as compared
to the case where the feedback is replaced by random-
ized input while preserving the connectivity structure
and presynaptic firing rates (RAND). Our results hence
show that active decorrelation of network activity by in-
hibitory feedback [4, 5] is a general phenomenon which
can be observed in realistic, highly heterogeneous net-
works with nonlinear interaction and sufficiently strong
negative feedback. Moreover, the study serves as a proof-
of-principle that network activity can be efficiently decor-
11
FIG. 8. Dependence of population-averaged input correlations (a) and spike-train correlations (b) on the heterogeneity of
the neuromorphic substrate for the intact network (FB, dark gray diamonds), the FBreplay (black circles) and the RAND (light
gray circles) case. Symbols and error bars denote mean and one standard deviation, respectively, across M = 100 network
realizations (error bars are partly covered by markers). The gray line in (a) depicts shared-input correlations in a homogeneous
network (Equation 2). Insets in (a) and (b) depict magnified views of input correlations in the FBreplay case and spike-train
correlations in the FB case, respectively. The light gray horizontal band represents mean ± three standard deviations of
(spurious) correlations in surrogate data where correlations were removed. Note that free membrane potentials cannot be
recorded in the FB case (see Section II). Hence, there are no gray diamonds in (a).
related even on heterogeneous hardware, which can be
exploited in functional applications, e.g., in the neuro-
morphic algorithms developed by Pfeil et al. [18] and
Schmuker et al. [19].
Functional neural architectures often rely on stochas-
tic dynamics of its constituents or on some form of back-
ground noise [see, e.g., 18, 19, 98]. Deterministic recur-
rent neural networks with inhibitory feedback could pro-
vide decorrelated noise to such functional networks, both
in artificial as well as in biological substrates. In neuro-
morphic hardware applications, these "noise networks"
could thereby replace conventional random-number gen-
erators and avoid a costly transmission of background
noise from a host computer to the hardware substrate
(which may be particularly relevant for mobile appli-
cations with low power consumption; see Supplements
1). It needs to be investigated, however, how well func-
tional stochastic circuits perform in the presence of such
network-generated noise.
Partial calibration of hardware neurons allowed us to
modulate the level of network heterogeneity and, there-
fore, to systematically study its effect on correlations in
the network activity. The analysis revealed two counter-
acting contributions: As shown in previous studies [e.g.,
67], neuron heterogeneity decorrelates (shared) feedfor-
ward input (feedforward decorrelation). On the other
hand, however, heterogeneity impairs feedback decorre-
lation (see next paragraph). In our network model, this
weakening of feedback decorrelation is the dominating
factor. Overall, we observed a slight increase in correla-
tions with increasing level of heterogeneity. We cannot
exclude that feedforward decorrelation may play a more
significant role for different network configurations (e.g.,
different connection strengths or network topologies, dif-
ferent structure of external inputs, different types of het-
erogeneity). Our study demonstrates, however, that het-
erogeneity is not necessarily suppressing correlations in
recurrent systems. In this context, it would be interest-
ing to investigate the interplay of signal and noise corre-
lations in the presence of network heterogeneities in re-
current systems. We leave this intriguing topic to future
studies.
As shown in [5], feedback decorrelation in recurrent
networks becomes more (less) efficient with increasing
(decreasing) strength of the effective negative feedback.
For networks of spiking neurons, the effective connection
strength wij between two neurons j and i corresponds to
the total number of extra spikes emitted by neuron i in re-
sponse to an additional input spike generated by neuron j
(see, e.g., [99]). Assuming that the effect of a single addi-
tional input spike is small, the effective connectivity can
be obtained by linear-response theory [91]. Note that the
effective weights wij depend on the working point, i.e.,
the average firing rates of all pre- and postsynaptic neu-
rons (mathematically, wij is given by the derivative of the
stationary response firing rate ri = φi(r1, . . . , rj, . . . , rN )
of neuron i with respect to the input firing rate rj, eval-
uated at the working point; for details, see [5]). Neurons
firing at very low or very high rates are typically less sen-
sitive to input fluctuations than neurons firing at inter-
mediate rates (due to the shape of the response function
φi(r1, . . . , rN )). Their dynamical range is reduced.
In
consequence, they hardly mediate feedback in a recurrent
network. In heterogeneous networks with broad distribu-
tions of firing rates, the number of these insensitive neu-
rons is increased. Hence, the effective feedback is weak-
0.00.20.40.60.81.0Heterogeneity a0.010.000.010.020.030.040.05Correlation κSSpike trains0.00.20.40.60.81.0Heterogeneity a0.000.020.040.060.08Correlation κVFree membrane potentialsRANDFBreplayFB0.00380.00320.0050.0100.015ab ened (see Supplements 3). We can qualitatively repro-
duce this effect of heterogeneity on correlations in recur-
rent networks (FB case) by means of a simplified linear
rate model where increasing heterogeneity is described as
a decrease in the effective-weight amplitudes (see Supple-
ments 5). A more quantitative analysis requires an ex-
plicit mapping of the synaptic weights in the LIF-neuron
network to the effective weights of the linear model (as
in, e.g., [100]) in the presence of distributed firing rates.
We commit this task to future studies. Note that the rate
dependence of the effective weights and the resulting ef-
fects on correlations are consistent with our observation
that neuron pairs with very low firing rates exhibit spike-
train correlations close to zero, whereas pairs with high
firing rates are positively correlated (see Supplements 7).
Pairs with one neuron firing at an intermediate rate of-
ten exhibit negative spike-train correlations. As shown
in [4, 5], these negative spike-train correlations are essen-
tial for compensating the positive contribution of shared
inputs to the total input correlation (at least in purely in-
hibitory networks). Narrowing the firing rate distribution
(e.g., by calibration of hardware neurons) increases the
number of neurons contributing to the negative feedback,
which, in turn, leads to more neuron pairs with negative
spike-train correlations and, therefore, to smaller overall
correlations.
Seemingly contrary to our findings, Bernacchia &
Wang [74] report a decrease in correlations with increas-
ing level of heterogeneity. The results of their study are
obtained for a linear network model, which can be consid-
ered the outcome of the linearization procedure described
above. Hence, the connectivity of their model corre-
sponds to an effective connectivity (see above). Their
study neglects the rate (working-point) dependence of the
effective weights and can therefore not account for the ef-
fect of firing-rate heterogeneity. In [74], heterogeneity is
quantified by the variance of the (effective) weight matrix
(Equations 2.2 and 2.4 in [74]). For sparse connectivity
matrices (with a large number of zero elements), the vari-
ance of the weight matrix reflects not only the width
of the non-zero-weight distribution, but also its mean
(Equation 2.4 in [74]). For networks of nonlinear spik-
ing neurons, heterogeneities in neuron and/or synapse
parameters broadens the distribution of non-zero effec-
tive weights, but may simultaneously reduce its mean
(see above, Supplements 3, and [46, 100]). Hence, the
variance of the full weight matrix may decrease (for il-
lustration, see Supplements Figure 9). In other words,
increasing heterogeneity in the nonlinear system may cor-
respond to decreasing heterogeneity in the linearized sys-
tem. A direct test of this hypothesis requires an explicit
linearization of the nonlinear heterogeneous system.
The results of this study were obtained by network
emulations on analog neuromorphic hardware. We re-
produced the main findings by means of conventional
computer simulations of LIF-neuron networks with dis-
tributed firing thresholds (see Figure 9). The focus on
threshold heterogeneity allows us to isolate the effect
12
of firing-rate distributions on correlations.
It does not
affect shared-input correlations (see Section II F). Al-
though networks simulated on conventional computers
and those emulated on the neuromorphic hardware dif-
fer in several respects (e.g., in the exact implementation
of heterogeneity or the synapse model; compare Supple-
ments Table I and II to Table I and II, respectively),
the qualitative results are very similar: In networks with
intact feedback loop, input and output correlations are
substantially reduced (as compared to the case where the
feedback is replaced by randomized input), but increase
with the extent of heterogeneity. As predicted by the
theory for homogeneous inhibitory networks, we observe
positive input correlations and negative output correla-
tions (see Equation 21 in [5] and in the paragraph which
follows; see also [103] and Supplements 5). Further, note
that heterogeneity in neuron parameters does not "av-
erage out" in larger networks. Upscaling the network
size by a factor of 25 (N = 4800, in-degree K = 375)
yields smaller spike-train correlations, but the qualita-
tive results are similar to those obtained for the smaller
network (N = 192, K = 15) emulated on the Spikey chip
(compare Figure 8 to Supplements Figure 3).
In networks with intact feedback loop (FB and FBreplay
scenarios), the precise spatio-temporal structure of spike
trains arranges such that the self-consistent input and
output correlations are suppressed. Perturbations of this
structure in the local input typically lead to an increase
in correlations [5]. In this study, we demonstrate this by
replaying spiking activity after randomization of spike
times, i.e., by replacing the time of each input spike by a
random number uniformly drawn from the full emulation
time interval [0, T ) (RAND case). However, even subtle
modifications of input spike trains, such as random dither
of spike times by few milliseconds, lead to an increase of
correlations. On the neuromorphic hardware, replay of
spike trains is not entirely reproducible (see Section II D
and Supplements 9). Hence, spike-train correlations mea-
sured in the FBreplay mode are slightly larger than in the
FB case. We would expect the same effect on the in-
put side (free membrane potentials). Due to hardware
limitations, however, we can measure input correlations
only in replay mode (FBreplay or RAND), but not in the
fully connected network (FB). Therefore, all reported in-
put correlations are likely to be slightly overestimated.
In conventional network simulations, we mimicked the
effect of unreliable replay by input-spike dithering and,
indeed, find a gradual increase in input and output cor-
relations (see Supplements Figure 6). These results seem
to be contrary to the study by Rosenbaum & Josic [104],
in which synaptic noise leads to a decrease of output
correlations in a feedforward scenario. In our case, spike-
train correlations, which suppress shared-input correla-
tions, are removed by dithering spikes, thereby increas-
ing correlations on the output side. In [104], in contrast,
spike-train correlations are always zero, and shared-input
correlations are decreased by synaptic failure, explaining
the decreased output correlations. We attribute this con-
13
FIG. 9. Dependence of population-averaged input correlations (a), and spike-train correlations (b) on the width of threshold
distributions in networks simulated with NEST [101] and PyNN [102], for the intact network (FB, dark gray circles) and the
RAND (light gray circles) case. Symbols and error bars denote mean and standard deviation, respectively, across M = 30
network realizations (error bars are partly covered by markers). The gray line in (a) depicts shared-input correlations in a
homogeneous network (Equation 2). The inset in (b) shows a magnified view of the spike-train correlations in the FB case.
Note that in simulations the FBreplay is identical to the FB case, and is hence not shown. For details see Supplements 4.
tradiction to the missing feedback loop in their system,
and expect correlations to increase in recurrent networks
subject to similar perturbations.
Despite the imperfect replay of input spikes, the decor-
relation effect is clearly visible in hardware emulations,
both on the input and on the output side. The repro-
ducibility of emulations on neuromorphic hardware could
be improved by stabilizing the environment of the sys-
tem, e.g., the chip temperature or the support electron-
ics (under development). Analog hardware, however, will
never reach the level of reproducibility of digital com-
puters. But note that, similar to analog hardware, bi-
ological neurons exhibit a considerable amount of trial-
to-trial variability, even under controlled in-vitro condi-
tions [90]. So far, the details of how neuronal noise, for
example, stochastic synapses (spontaneous postsynap-
tic events, stochastic spike transmission, synaptic failure
[105]), affects correlations in recurrent neural circuits re-
main unclear.
Although different Spikey chips exhibit different real-
izations of fixed-pattern noise, they show a comparable
extent of heterogeneity and yield results which are qual-
itatively similar to those presented in this article (Sup-
plements 8).
We have shown that negative feedback in recurrent cir-
cuits can efficiently suppress correlations, even in highly
heterogeneous systems such as the analog neuromorphic
architecture Spikey. Correlations can be further reduced
by minimizing the level of network heterogeneity. In this
study, we reduced the level of heterogeneity through cal-
ibration of neuron parameters in the unconnected case
(see Section II E). The calibration could, in principle, be
improved by calibrating neuron (and possibly synapse)
parameters in the full recurrent network. Such calibra-
tion procedures are however time consuming and cumber-
some. In biological substrates, homeostasis mechanisms
[56, 106] keep neurons in a responsive regime and reduce
the level of firing-rate heterogeneity in a self-regulating
manner. Future neuromorphic devices could mimic this
behavior, thereby reducing the necessity of time consum-
ing calibration procedures. Alternatively, the analog cir-
cuits could be optimized to reduce fixed-pattern noise.
This would likely require the allocation of more chip re-
sources, hence reducing the network size per chip area.
For simplicity, this work focuses on purely inhibitory
networks (as in [11]). This demonstrates that decorrela-
tion by inhibitory feedback does not rely on a dynami-
cal balance between excitation and inhibition (note that
the external "excitatory" drive is constant in our model)
[5, 6]. Previous studies have shown that, for the ho-
mogeneous case, decorrelation by inhibitory feedback is
a general phenomenon, which also occurs in excitatory-
inhibitory networks, provided the overall inhibition is
sufficiently strong (which is typically the case to ensure
stability) [4–6, 74]. For the heterogeneous case, com-
puter simulations of excitatory-inhibitory networks show
qualitatively the same results as purely inhibitory net-
works (compare Figure 8 to Supplements Figure 5), con-
firming that our results generalize to the case of mixed
excitatory-inhibitory coupling.
Similar to our study, Giulioni et al. [107] use a theory-
guided approach to implement, verify and investigate
network dynamics on analog neuromorphic hardware. In
their study, an attractor network is implemented that is
inspired by a mean-field model. Due to heterogeneities
in synaptic efficacies on the hardware, stability analy-
sis of attractor states requires the authors to measure
effective response functions of populations of hardware
neurons. To this end, they replace recurrent connections
in one population of neurons by external input. This al-
0.000.020.040.060.080.100.120.14Rel. std. deviation of threshold potential σΘµΘ0.010.000.010.020.030.040.05Correlation κS0.000.020.040.060.080.100.120.14Rel. std. deviation of threshold potential σΘµΘ0.000.020.040.060.080.10Correlation κVFBRAND0.0060.0030.000ab 14
mized by processing the data directly on the hardware, or
by choosing a data representation which is closer to the
format used on the Spikey chip, but this would impair
user-friendliness, and hence, the effectiveness of proto-
typing. While the Spikey system permits the monitoring
of the spiking activity of all neurons simultaneously, ac-
cess to the membrane potentials is limited to a single (al-
beit arbitrary) neuron in each emulation run. Monitor-
ing of membrane potentials of a population of n neurons
therefore requires n repetitions of the same emulation.
Extending the hardware system to enable access to the
membrane potentials of at least two neurons simultane-
ously would allow for a direct observation of input cor-
relations in the intact network (and thereby avoid prob-
lems with replay reproducibility; see above) and reduce
execution time (the Spikey chip itself permits recording
of up to eight neurons in parallel, the support electron-
ics, however, does not). While the Spikey system does
not significantly outperform conventional computers in
terms of computational power, emulations on this sys-
tem are much more energy efficient (Supplements 1). A
substantial increase of computational power is expected
for large systems exploiting the scalability of this tech-
nology without slow-down [108].
ACKNOWLEDGMENTS
We would like to thank Matthias Hock, Andreas Har-
tel, Eric Müller and Christoph Koke for their support
in developing and building the Spikey system, as well as
Mihai Petrovici and Sebastian Schmitt for fruitful dis-
cussions. This research is partially supported by the
Helmholtz Association portfolio theme SMHB, the Jülich
Aachen Research Alliance (JARA), EU Grant 269921
(BrainScaleS), and EU Grant 604102 (Human Brain
Project, HBP). We acknowledge financial support by
Deutsche Forschungsgemeinschaft and Ruprecht-Karls-
Universität Heidelberg within the funding programme
Open Access Publishing. All network simulations car-
ried out with NEST (http://www.nest-simulator.org).
Appendix A: Network description
Details on the network model as well as parameter val-
ues are provided in Table I and II, respectively.
Appendix B: Description of data analysis
A summary of the quantities and the data analysis used
in this study is provided in Table III. Parameter values
for the data analysis are given in Table IV.
FIG. 10. Acceleration factor as a function of emulated net-
work time T for the record (black) and the replay case (gray).
The acceleration factor is defined as the ratio between the
emulated network time T (in biological time) and the exe-
cution time (wall clock time). In the record case, a network
realization is generated on the host computer and uploaded
to the chip. During the subsequent emulation, spike trains
are recorded. In the replay case, spikes are replayed and the
membrane potential of one neuron is recorded with full sam-
pling frequency (9.6 kHz). The execution time covers the full
data flow from a network description in PyNN to the emula-
tion on the Spikey system and back to the network repre-
sentation in PyNN. The time-averaged population firing rate
is ¯r = (23.7 ± 0.2) s−1. The vertical dashed line depicts the
runtime used in this study. The hardware system has to be
initialized once before usage (< 1 s), which is not considered
here.
lows them to measure the firing rate of the population
as a function of the external input, while the activity of
the population is in equilibrium with that of other re-
currently connected populations. This study represents
another example illustrating that investigations of actual
hardware emulations are a prerequisite for successful ap-
plication of analog neuromorphic hardware.
This study demonstrates that the Spikey system has
matured to a level that permits its use as a tool for neu-
roscientific research. For the results presented in this
study, we recorded in total 1011 membrane-potential and
spike-train samples, representing more than 100 days of
biological time. Due to the 104-fold acceleration of the
Spikey chip, this corresponds to less than 15 minutes
in the hardware-time domain. Interfacing the hardware
system, however, reduces the acceleration to an approx-
imately 50-fold speed-up (Figure 10). The translation
between the network description and its hardware rep-
resentation claims the majority of execution time, more
than the network emulation and the transfer of data to
and from the hardware system together. Encoding and
decoding spike times on the host computer is particu-
larly expensive. Obviously, the system could be opti-
100101102103Emulated network time T(s)100101102103Acceleration factorrecordreplay 15
A
Populations
Topology
Connectivity
Neuron model
Channel models
Synapse model
Plasticity
External input
Measurements
Other
B
Name
I
Model summary
One (inhibitory)
None
Random convergent connections (fixed in-degree)
Leaky integrate-and-fire (LIF), fixed firing threshold, fixed absolute refractory time
None
Exponentially decaying conductances, fixed delays
None
Resting potential higher than threshold (= constant current) (El > Θ)
Spikes and membrane potentials from all neurons
No autapses, no multapses
Populations
Elements
LIF neuron
Size
N
C
Source Target Pattern
I
I
Random convergent connections, in-degree K
Connectivity
D
Type
Subthreshold
dynamics
Conductance
dynamics
Spiking
Neuron and synapse model
d
Cm
dt v(t) = −gl(v(t) − El) − gsyn(t)(v(t) − Einh)
Leaky integrate-and-fire, exponentially decaying conductances
Subthreshold dynamics (t (cid:54)∈ (t∗, t∗ + τref)):
Reset and refractoriness (t ∈ (t∗, t∗ + τref)):
This model is emulated by analog circuitry on the Spikey chip [13].
For each presynaptic spike at time t∗ (t > t∗ + d):
v(t) = vreset
gsyn(t) ≈ J exp(− t−t∗−d
τsyn
)Θ(t), with J = whwgmax and Heaviside function Θ(t).
This model is emulated by analog circuitry on the Spikey chip [18].
If v(t∗−) < Θ ∧ v(t∗+) ≥ Θ:
emit spike with time stamp t∗
TABLE I. Description of the network model (according to [109]).
[1] Gierer, A., & Meinhardt, H. (1974). Biological pat-
tern formation involving lateral inhibition. Lectures on
mathematics in the life sciences 7, 163.
[2] Rosenzweig, M. L. (1978). Competitive speciation. Bi-
ological Journal of the Linnean Society 10 (3), 275–289.
[3] Neel, L. (1952). Antiferromagnetism and ferrimag-
netism. Proceedings of the Physical Society. Section
A 65 (11), 869.
[4] Renart, A., De La Rocha, J., Bartho, P., Hollender, L.,
Parga, N., Reyes, A., & Harris, K. D. (2010). The asyn-
chronous state in cortical circuits. Science 327, 587–590.
[5] Tetzlaff, T., Helias, M., Einevoll, G., & Diesmann, M.
(2012). Decorrelation of neural-network activity by in-
hibitory feedback. PLoS Comput. Biol. 8 (8), e1002596.
[6] Helias, M., Tetzlaff, T., & Diesmann, M. (2014). The
correlation structure of local cortical networks intrinsi-
cally results from recurrent dynamics. PLoS Comput.
Biol. 10 (1), e1003428.
[7] Pearson, K. (1901). Liii. on lines and planes of closest
fit to systems of points in space. Philosophical Magazine
Series 6 2 (11), 559–572.
B
Name
N
C
Name
K
D
Name
vreset
El
Θ
Einh
τref
Cm
gl
τ eff
m
D
Name
gmax
whw
V eff
max
τ eff
syn
deff
Other
Name
SpikeyHAL
PyNN
vmodule
logger
Other
Name
16
Values
{96, 112, 128, 144, 160, 176, 192}
Populations
Description
network size
Connectivity
Values
15
Values
−80 mV
−52 mV
−62 mV
−80 mV
1 ms
0.2 nF
tuned during calibration process
uncalibrated: 10.50(5.94, 17.90)ms
calibrated: 10.77(6.22, 18.30)ms
Description
number of presynaptic partners
Neuron
Description
reset potential
resting potential
firing threshold
inhibitory reversal potential
refractory period
membrane capacitance
leak conductance
effective membrane time constant
Synapse
Description
conductance amplitude
synaptic weight (in hardware values ∈ [0, 15])
Values
in the order of 1 nS
3
uncalibrated: -5.57(−12.50,−2.62)mV effective post-synaptic potential prefactor
calibrated: -6.05(−14.37,−2.92)mV
uncalibrated: 3.78(2.52, 5.58)ms
calibrated: 4.17(2.76, 6.18)ms
uncalibrated: 2.16(1.91, 2.48)ms
calibrated: 2.26(1.93, 2.55)
effective synaptic-current time constant
effective synaptic delay
Values
9e86d11c
2fe40b43
76ef3b44
826c5ed6
Software
Description
git revision
git revision
git revision
git revision
Hardware
Values
chip 508 (version 5)
chip 503, 504 (version 5) and 603, 605,
666 (version 4)
Description
chip used for main manuscript
chips used for supplements
TABLE II. Parameter values for the network model described in Table I. Bold numbers indicate default values. Gray numbers
indicate target values not considering fixed-pattern noise. Leak conductances gl are adjusted in the calibration process (see
Section II E). τ eff
max describe effective values measured from spike-triggered averages in RAND emulations
as described in Supplements 10. Effective values denote the median across synapses and trials followed by the 25th and 75th
percentiles in brackets.
syn, deff and V eff
m , τ eff
A
Measure
spike density
spike train
population activity
time-averaged population activity
membrane potential
(finite time) Fourier transform
(single unit) power spectrum
population-averaged power spectrum
population power spectrum
pairwise cross spectrum
population-averaged cross spectrum
sliding window filter
coherence
low-frequency coherence
pop.-averaged cross-correlation function
time average
17
Analysis measures
Details
ξi(t) =(cid:80)
k δ(t − tk
i )
(cid:80)
si(tk) = number of spikes of neuron i per bin [k∆t, (k + 1)∆t)
¯s(t) = 1
N
¯r = (cid:104)¯s(t)(cid:105)t
vi(tk) = membrane potential of neuron i in bin [k∆t, (k + 1)∆t)
i si(t)
0 dt xi(t)e−2πif t (with inverse F−1)
j Sj(f ))
(cid:80)
T X∗
T ((cid:80)
Xi(f ) = F[xi(t)](f ) =(cid:82) T
i (f )Xi(f )
i (f ))((cid:80)
i Ai(f )
i S∗
(cid:80)
i(cid:54)=j Cij(f ) ≡
Ai(f ) = 1
A(f ) = 1
N
¯A(f ) = 1
i (f )Xj(f ), i (cid:54)= j
T X∗
Cij = 1
C(f ) =
N (N−1)
(note: C(f ) ∈ R)
X(f ) → X(f ) ∗ H(f )
with H(f ) = 1
κ(f ) = C(f )
A(f )
1
fmax−fmin
N (N−1)
f1−f0
κX =
1
1
c(τ ) =
(cid:104). . .(cid:105)t
N (N−1) ( ¯A(f ) − N A(f ))
1
Θ(f − f0)Θ(f1 − f )
(cid:82) fmax
(cid:80)
i(cid:54)=j (cid:104)si(t)sj(t + τ )(cid:105)t ≡ F−1[C(f )](τ )
df κ(f )
fmin
TABLE III. Summary of the data analysis. Here i ∈ [1, N ], X ∈ S, V .
A
Parameter
∆t
∆tm
Twarmup
T
M
∆F
fmin, fmax
a
Analysis parameters
Description
bin size for spike trains
bin size for membrane potential traces
initial warm-up time (not considered in analysis)
emulated network time (biol. time domain)
number of network realizations
width of sliding window
interval boundaries for low-frequency coherence
calibration state
Values
1 ms
0.52 ms
1 s
10 s
{50, 100}
1 Hz
0.1 Hz, 20 Hz
{0, 1
8 , 4
8 , 5
8 , 3
8 , 2
8 , 6
8 , 7
8 , 1}
TABLE IV. Summary of analysis parameters (default values in bold).
[8] Ginis, G., & Peng, C.-N. (2006). Alien crosstalk
cancellation for multipair digital subscriber line sys-
tems. EURASIP Journal on Advances in Signal Pro-
cessing 2006 (1), 1–12.
[9] Benesty, J., Gänsler, T., Morgan, D. R., Sondhi, M. M.,
& Gay, S. L. (2001). Advances in Network and Acoustic
Echo Cancellation (1 ed.). Berlin, Germany: Springer.
[10] Trichina, E., Bucci, M., De Seta, D., & Luzzi, R. (2001).
Supplemental cryptographic hardware for smart cards.
Micro, IEEE 21 (6), 26–35.
[11] Mar, D., Chow, C., Gerstner, W., Adams, R., & Collins,
J. (1999). Noise shaping in populations of coupled model
neurons. Proc. Nat. Acad. Sci. USA 96 (18), 10450–
10455.
[12] Mead, C. (1990). Neuromorphic electronic systems.
Proc. IEEE 78 (10), 1629–1636.
[13] Indiveri, G., Linares-Barranco, B., Hamilton, T. J., van
Schaik, A., Etienne-Cummings, R., Delbruck, T., Liu,
S.-C., Dudek, P., Häfliger, P., Renaud, S., Schemmel, J.,
Cauwenberghs, G., Arthur, J., Hynna, K., Folowosele,
F., Saighi, S., Serrano-Gotarredona, T., Wijekoon, J.,
Wang, Y., & Boahen, K. (2011). Neuromorphic silicon
neuron circuits. Front. Neurosci. 5 (73).
[14] SenseMaker (2009). SenseMaker: A multi-sensory, task-
specific, adaptable perception system; project website.
Available at: http://cordis.europa.eu/project/rcn/
62746_en.html.
[15] FACETS (2010). Fast Analog Computing with Emer-
gent Transient States, project website. Available at:
http://www.facets-project.org.
[16] BrainScaleS (2014). Project website. Available at:
http://www.brainscales.eu.
[17] Human Brain Project (2014). Project website. Available
at: http://www.humanbrainproject.eu.
[18] Pfeil, T., Grübl, A., Jeltsch, S., Müller, E., Müller, P.,
Petrovici, M. A., Schmuker, M., Brüderle, D., Schem-
mel, J., & Meier, K. (2013). Six networks on a uni-
versal neuromorphic computing substrate. Front. Neu-
rosci. 7 (11).
[19] Schmuker, M., Pfeil, T., & Nawrot, M. P. (2014). A neu-
romorphic network for generic multivariate data classi-
fication. Proc. Natl. Acad. Sci. USA 111 (6), 2081–2086.
[20] Ecker, A. S., Berens, P., Tolias, A. S., & Bethge, M.
(2011). The effect of noise correlations in populations
of diversely tuned neurons. J. Neurosci. 31 (40), 14272–
14283.
[21] Averbeck, B. B., Latham, P. E., & Pouget, A. (2006).
Neural correlations, population coding and computa-
tion. Nat. Rev. Neurosci. 7, 358–366.
[22] Moreno-Bote, R., Beck, J., Kanitscheider, I., Pitkow,
Information-
X., Latham, P., & Pouget, A. (2014).
limiting correlations. Nat. Neurosci. 17, 1410–1417.
[23] Fries, P. (2005). A mechanism for cognitive dynamics:
neuronal communication through neuronal coherence.
Theor. Comput. Sci. 9 (10), 474–480.
[24] Abeles, M. (1991). Corticonics: Neural Circuits of the
Cerebral Cortex (1st ed.). Cambridge: Cambridge Uni-
versity Press.
[25] Diesmann, M., Gewaltig, M.-O., & Aertsen, A. (1999).
Stable propagation of synchronous spiking in cortical
neural networks. Nature 402 (6761), 529–533.
[26] Tabareau, N., Slotine, J.-J., & Pham, Q.-C. (2010).
How synchronization protects from noise. PLoS Com-
put. Biol. 6 (1). doi:10.1371/journal.pcbi.1000637.
18
[27] Salinas, E., & Sejnowski, T. J. (2001). Correlated neu-
ronal activity and the flow of neural information. Nat.
Rev. Neurosci. 2 (8), 539–550.
[28] Shadlen, M. N., & Newsome, W. T. (1998). The variable
discharge of cortical neurons: Implications for connec-
tivity, computation, and information coding. J. Neu-
rosci. 18 (10), 3870–3896.
[29] Abbott, L. F., & Dayan, P. (1999). The effect of corre-
lated variability on the accuracy of a population code.
Neural Comput. 11, 91–101.
[30] Tripp, B., & Eliasmith, C. (2007). Neural popula-
tions can induce reliable postsynaptic currents without
observable spike rate changes or precise spike timing.
Cereb. Cortex 17 (8), 1830–1840.
[31] Schmuker, M., & Schneider, G. (2007). Processing and
classification of chemical data inspired by insect olfac-
tion. Proc. Natl. Acad. Sci. USA 104 (51), 20285–20289.
[32] Cohen, M. R., & Maunsell, J. H. R. (2009). Attention
improves performance primarily by reducing interneu-
ronal correlations. Nat. Neurosci. 12, 1594–1600.
[33] Tetzlaff, T., Morrison, A., Geisel, T., & Diesmann,
M. (2004). Consequences of realistic network size on
the stability of embedded synfire chains. Neurocomput-
ing 58–60, 117–121.
[34] Kriener, B., Tetzlaff, T., Aertsen, A., Diesmann, M., &
Rotter, S. (2008). Correlations and population dynam-
ics in cortical networks. Neural Comput. 20, 2185–2226.
[35] Tetzlaff, T., Rotter, S., Stark, E., Abeles, M., Aertsen,
A., & Diesmann, M. (2008). Dependence of neuronal
correlations on filter characteristics and marginal spike-
train statistics. Neural Comput. 20 (9), 2133–2184.
[36] Ecker, A. S., Berens, P., Keliris, G. A., Bethge, M., &
Logothetis, N. K. (2010). Decorrelated neuronal firing
in cortical microcircuits. Science 327 (5965), 584–587.
[37] Ly, C., Middleton, J., & Doiron, B. (2012). Cellular and
circuit mechanisms maintain low spike co-variability
and enhance population coding in somatosensory cor-
tex. Front. Comput. Neurosci. 6 (7), 1–25.
[38] Middleton, J., Omar, C., Doiron, B., & Simons, D.
(2012). Neural correlation is stimulus modulated by
feedforward inhibitory circuitry. J. Neurosci. 32 (2),
506–518.
[39] Wiechert, M. T., Judkewitz, B., Riecke, H., & Friedrich,
R. W. (2010). Mechanisms of pattern decorrelation by
recurrent neuronal circuits. Nat. Neurosci. 13, 1003–
1010.
[40] Harris, K. D., & Thiele, A. (2011). Cortical state and
attention. Nat. Rev. Neurosci. 12, 509–523.
[41] Griffith, J. S., & Horn, G. (1966). An analysis of spon-
taneous impulse activity of units in the striate cortex of
unrestrained cats. J. Physiol. (Lond.) 186 (3), 516–534.
[42] Koch, K. W., & Fuster, J. M. (1989). Unit activity in
monkey parietal cortex related to haptic perception and
temporary memory. Exp. Brain Res. 76 (2), 292–306.
[43] Shafi, M., Zhou, Y., Quintana, J., Chow, C., Fuster,
J., & Bodner, M. (2007). Variability in neuronal ac-
tivity in primate cortex during working memory tasks.
Neuroscience 146 (3), 1082–1108.
[44] Hromadka, T., DeWeese, M. R., & Zador, A. M. (2008).
Sparse representation of sounds in the unanesthetized
auditory cortex. PLoS Biol. 6 (1).
[45] OĆonnor, D. H., Peron, S. P., Huber, D., & Svoboda,
K. (2010). Neural activity in barrel cortex underly-
ing vibrissa-based object localization in mice. Neu-
ron 67 (6), 1048–1061.
[46] Roxin, A., Brunel, N., Hansel, D., Mongillo, G., & van
Vreeswijk, C. (2011). On the distribution of firing rates
in networks of cortical neurons. J. Neurosci. 31 (45),
16217–16226.
[47] Buzsaki, G., & Mizuseki, K. (2014). The log-dynamic
brain: how skewed distributions affect network opera-
tions. Nat. Rev. Neurosci. 15, 264–278.
[48] Song, S., Sjöström, P., Reigl, M., Nelson, S., &
Chklovskii, D. (2005). Highly nonrandom features of
synaptic connectivity in local cortical circuits. PLoS
Biol. 3 (3), e68.
[49] Lefort, S., Tomm, C., Sarria, J.-C. F., & Petersen, C.
C. H. (2009). The excitatory neuronal network of the C2
barrel column in mouse primary somatosensory cortex.
Neuron 61, 301–316.
[50] Koulakov, A. A., Hromadka, T., & Zador, A. M. (2009).
Correlated connectivity and the distribution of firing
rates in the neocortex. J. Neurosci. 29 (12), 3685–3694.
[51] Avermann, M., Tomm, C., Mateo, C., Gerstner, W.,
& Petersen, C. (2012). Microcircuits of excitatory and
inhibitory neurons in layer 2/3 of mouse barrel cortex.
J. Neurophysiol. 107 (11), 3116–3134.
[52] Ikegaya, Y., Sasaki, T., Ishikawa, D., Honma, N., Tao,
K., Takahashi, N., Minamisawa, G., Ujita, S., & Mat-
suki, N. (2013). Interpyramid spike transmission stabi-
lizes the sparseness of recurrent network activity. Cereb.
Cortex 23 (2), 293–304.
[53] Roxin, A. (2011). The role of degree distribution in
shaping the dynamics in networks of sparsely connected
spiking neurons. Front. Comput. Neurosci. 5 (8).
[54] Kuhn, A., Aertsen, A., & Rotter, S. (2004). Neuronal
integration of synaptic input in the fluctuation-driven
regime. J. Neurosci. 24 (10), 2345–2356.
[55] Stein, R. B., Gossen, E. R., & Jones, K. E. (2005).
Neuronal variability: Noise or part of the signal? na-
trevnsci 6, 389–397.
[56] Marder, E., & Goaillard, J.-M. (2006). Variability, com-
pensation and homeostasis in neuron and network func-
tion. Nat. Rev. Neurosci. 7 (7), 563–574.
[57] Van Vreeswijk, C., & Sompolinsky, H. (1998). Chaotic
balanced state in a model of cortical circuits. Neural
Comput. 10 (6), 1321–1371.
[58] Tsodyks, M., Mitkov, I., & Sompolinsky, H. (1993). Pat-
tern of synchrony in inhomogeneous networks of oscil-
lators with pulse interactions. Phys. Rev. Lett. 71 (8).
[59] Golomb, D., & Rinzel, J. (1993). Dynamics of globally
coupled inhibitory neurons with heterogeneity. Phys.
Rev. E 48 (6), 4810–4814.
[60] Neltner, L., Hansel, D., Mato, G., & Meunier, C. (2000).
Synchrony in heterogeneous networks of spiking neu-
rons. Neural Comput. 12 (7), 1607–1641.
[61] Denker, M., Timme, M., Diesmann, M., Wolf, F., &
Geisel, T. (2004). Breaking synchrony by heterogeneity
in complex networks. Phys. Rev. Lett. 92 (7), 074103–
1–074103–4.
[62] Mejias, J. F., & Longtin, A. (2012). Optimal hetero-
geneity for coding in spiking neural networks. Phys.
Rev. Lett. 108.
[63] Stocks, N. G. (2000). Suprathreshold stochastic reso-
nance in multilevel threshold systems. Physical Review
Letters 84 (11), 2310.
[64] Shamir, M., & Sompolinsky, H. (2006). Implications of
neuronal diversity on population coding. Neural Com-
19
put. 18 (8), 1951–1986.
[65] Chelaru, M. I., & Dragoi, V. (2008). Efficient coding in
heterogeneous neuronal populations. Proc. Natl. Acad.
Sci. USA 105 (42).
[66] Osborne, L. C., Palmer, S. E., Lisberger, S. G., &
Bialek, W. (2008). The neural basis for combinato-
rial coding in a cortical population response. J. Neu-
rosci. 28 (50), 13522–13531.
[67] Padmanabhan, K., & Urban, N. N. (2010).
Intrinsic
biophysical diversity decorrelates neuronal firing while
increasing information content. Nat. Neurosci. 13 (10),
1276–1282.
[68] Marsat, G., & Maler, L. (2010). Neural heterogeneity
and efficient population codes for communication sig-
nals. J. Neurophysiol. 104 (5), 2543–2555.
[69] Holmstrom, L. A., Eeuwes, L. B., Roberts, P. D., &
Portfors, C. V. (2010). Efficient encoding of vocaliza-
tions in the auditory midbrain. J. Neurosci. 30 (3), 802–
819.
[70] Yim, M. Y., Aertsen, A., & Rotter, S. (2013). Impact of
intrinsic biophysical diversity on the activity of spiking
neurons. Phys. Rev. E 87.
[71] Lengler, J., Jug, F., & Steger, A. (2013). Reliable neu-
ronal systems: The importance of heterogeneity. PLoS
One.
[72] Mejias, J. F., & Longtin, A. (2014). Differential ef-
fects of excitatory and inhibitory heterogeneity on the
gain and asynchronous state of sparse cortical networks.
Frontiers in computational neuroscience 8.
[73] Tripathy, S. J., Padmanabhan, K., Gerkin, R. C., &
Urban, N. N. (2013).
Intermediate intrinsic diversity
enhances neural population coding. Proc. Natl. Acad.
Sci. USA 110 (20), 8248–8253.
[74] Bernacchia, A., & Wang, X.-J. (2013). Decorrelation
by recurrent inhibition in heterogeneous neural circuits.
Neural Comput. 25 (7), 1732–1767.
[75] Schemmel, J., Grübl, A., Meier, K., & Müller, E. (2006).
Implementing synaptic plasticity in a VLSI spiking neu-
ral network model. In Proceedings of the 2006 Interna-
tional Joint Conference on Neural Networks (IJCNN),
Vancouver, pp. 1–6. IEEE Press.
[76] Badoni, D., Giulioni, M., Dante, V., & Del Giudice, P.
(2006). An aVLSI recurrent network of spiking neurons
with reconfigurable and plastic synapses. In Proceedings
of the 2006 International Symposium on Circuits and
Systems (ISCAS), Island of Kos, pp. 4. IEEE Press.
[77] Hafliger, P. (2007). Adaptive WTA with an analog
VLSI neuromorphic learning chip. IEEE Trans. Neu-
ral Netw. 18 (2), 551–572.
[78] Vogelstein, R., Mallik, U., Vogelstein, J., & Cauwen-
berghs, G. (2007). Dynamically reconfigurable sili-
con array of spiking neurons with conductance-based
synapses. IEEE Trans. Neural Netw. 18 (1), 253–265.
[79] Indiveri, G., Chicca, E., & Douglas, R. (2009). Artifi-
cial cognitive systems: From VLSI networks of spiking
neurons to neuromorphic cognition. Cognitive Compu-
tation 1 (2), 119–127.
[80] Serrano-Gotarredona, R., Oster, M., Lichtsteiner,
P., Linares-Barranco, A., Paz-Vicente, R., Gomez-
Rodriguez, F., Camunas-Mesa, L., Berner, R., Rivas-
Perez, M., Delbruck, T., Liu, S.-C., Douglas, R.,
Hafliger, P., Jimenez-Moreno, G., Ballcels, A., Serrano-
Gotarredona, T., Acosta-Jimenez, A., & Linares-
Barranco, B. (2009). CAVIAR: A 45k neuron, 5M
synapse, 12G connects/s AER hardware sensory - pro-
cessing - learning - actuating system for high-speed vi-
sual object recognition and tracking. IEEE Trans. Neu-
ral Netw. 20 (9), 1417–1438.
[81] Brink, S., Nease, S., Hasler, P., Ramakrishnan, S., Wun-
derlich, R., Basu, A., & Degnan, B. (2013). A learning-
enabled neuron array IC based upon transistor channel
models of biological phenomena. IEEE Trans. Biomed.
Circuits Syst. 7 (1), 71–81.
[82] Benjamin, B., Gao, P., McQuinn, E., Choudhary, S.,
Chandrasekaran, A., Bussat, J., Alvarez-Icaza, R.,
Arthur, J., Merolla, P., & Boahen, K. (2014). Neurogrid:
A mixed-analog-digital multichip system for large-scale
neural simulations. Proc. IEEE 102 (5), 699–716.
[83] Renaud, S., Tomas, J., Lewis, N., Bornat, Y., Daouzli,
A., Rudolph, M., Destexhe, A., & Saïghi, S. (2010).
PAX: A mixed hardware/software simulation platform
for spiking neural networks. Neural Networks 23 (7),
905–916.
[84] Yu, T., & Cauwenberghs, G. (2010). Analog VLSI
biophysical neurons and synapses with programmable
membrane channel kinetics. IEEE Trans. Biomed. Cir-
cuits Syst. 4 (3), 139–148.
[85] Davison, A., Brüderle, D., Eppler, J. M., Kremkow, J.,
Muller, E., Pecevski, D., Perrinet, L., & Yger, P. (2009).
PyNN: a common interface for neuronal network simu-
lators. Front. Neuroinformatics 2 (11).
[86] Brüderle, D., Müller, E., Davison, A., Muller, E.,
Schemmel, J., & Meier, K. (2009). Establishing a
novel modeling tool: A Python-based interface for a
neuromorphic hardware system. Front. Neuroinformat-
ics 3 (17).
[87] Kaplan, B., Brüderle, D., Schemmel, J., & Meier, K.
(2009). High-conductance states on a neuromorphic
hardware system. In Proceedings of the 2009 Interna-
tional Joint Conference on Neural Networks (IJCNN),
Atlanta, pp. 1524–1530. IEEE Press.
[88] Bill, J., Schuch, K., Brüderle, D., Schemmel, J., Maass,
W., & Meier, K. (2010). Compensating inhomogeneities
of neuromorphic VLSI devices via short-term synaptic
plasticity. Front. Comput. Neurosci. 4 (129).
[89] Brüderle, D., Bill, J., Kaplan, B., Kremkow, J., Meier,
K., Müller, E., & Schemmel, J. (2010). Simulator-
like exploration of cortical network architectures with
a mixed-signal VLSI system. In Proceedings of the 2010
International Symposium on Circuits and Systems (IS-
CAS), Paris, France. IEEE Press.
[90] Mainen, Z. F., & Sejnowski, T. J. (1995). Reliability of
spike timing in neocortical neurons. Science 268, 1503–
1506.
[91] De la Rocha, J., Doiron, B., Shea-Brown, E., Kresimir,
J., & Reyes, A. (2007). Correlation between neural spike
trains increases with firing rate. Nature 448 (16), 802–
807.
[92] Press, W. H., Teukolsky, S. A., Vetterling, W. T., &
Flannery, B. P. (2007). Numerical Recipes: The Art of
Scientific Computing (3rd ed.). Cambridge University
Press.
[93] Prinz, A. A., Bucher, D., & Marder, E. (2004). Similar
network activity from disparate circuit parameters. Nat.
Neurosci. 7, 1345–1352.
20
[94] Pfeil, T., Scherzer, A.-C., Schemmel, J., & Meier, K.
(2013). Neuromorphic learning towards nano second
precision. In Neural Networks (IJCNN), The 2013 In-
ternational Joint Conference on, pp. 1–5. IEEE Press.
[95] Bair, W., Zohary, E., & Newsome, W. (2001). Corre-
lated firing in Macaque visual area MT: time scales and
relationship to behavior. J. Neurosci. 21 (5), 1676–1697.
[96] Kohn, A., & Smith, M. A. (2005). Stimulus dependence
of neuronal correlations in primary visual cortex of the
Macaque. J. Neurosci. 25 (14), 3661–3673.
[97] Moreno-Bote, R., & Parga, N. (2006). Auto- and cross-
correlograms for the spike response of leaky integrate-
and-fire neurons with slow synapses. Phys. Rev. Lett. 96,
028101.
[98] Maass, W. (2014). Noise as a resource for computa-
tion and learning in networks of spiking neurons. Proc.
IEEE 102 (5), 860–880.
[99] London, M., Roth, A., Beeren, L., Häusser, M., &
Latham, P. E. (2010). Sensitivity to perturbations in
vivo implies high noise and suggests rate coding in cor-
tex. Nature 466 (1), 123–128.
[100] Grytskyy, D., Tetzlaff, T., Diesmann, M., & Helias, M.
(2013). A unified view on weakly correlated recurrent
networks. Front. Comput. Neurosci. 7, 131.
[101] Gewaltig, M.-O., & Diesmann, M. (2007). NEST (NEu-
ral Simulation Tool). Scholarpedia 2 (4), 1430.
[102] Davison, A., Brüderle, D., Eppler, J., Kremkow, J.,
Muller, E., Pecevski, D., Perrinet, L., & Yger, P.
(2008).
PyNN: a common interface for neuronal
network simulators. Front. Neuroinformatics 2 (11).
doi:10.3389/neuro.11.011.2008.
[103] Helias, M., Tetzlaff, T., & Diesmann, M. (2013). Echoes
in correlated neural systems. New J. Phys. 15, 023002.
[104] Rosenbaum, R., & Josic, K. (2011). Mechanisms that
modulate the transfer of spiking correlations. Neural
Comput. 23 (5), 1261–1305.
[105] Ribrault, C., Sekimoto, K., & Triller, A. (2011). From
the stochasticity of molecular processes to the variabil-
ity of synaptic transmission. natrevnsci 12, 375–387.
[106] O'Leary, T., Williams, A. H., Franci, A., & Marder,
E. (2014). Cell types, network homeostasis, and patho-
logical compensation from a biologically plausible ion
channel expression model. Neuron 82 (4), 809–821.
[107] Giulioni, M., Camilleri, P., Mattia, M., Dante, V.,
Braun, J., & Del Giudice, P. (2012). Robust working
memory in an asynchronously spiking neural network re-
alized in neuromorphic VLSI. Front. Neurosci. 5 (149).
[108] Schemmel, J., Brüderle, D., Grübl, A., Hock, M., Meier,
K., & Millner, S. (2010). A wafer-scale neuromorphic
hardware system for large-scale neural modeling.
In
Proceedings of the 2010 International Symposium on
Circuits and Systems (ISCAS), Paris, pp. 1947–1950.
IEEE Press.
[109] Nordlie, E., Gewaltig, M.-O., & Plesser, H. E. (2009).
Towards reproducible descriptions of neuronal network
models. PLoS Comput. Biol. 5 (8), e1000456.
[110] Sharp, T., Galluppi, F., Rast, A., & Furber, S. (2012).
Power-efficient simulation of detailed cortical microcir-
cuits on SpiNNaker. J. Neurosci. Methods 210 (1), 110–
118.
Supplements
Supplements 1: Power consumption
i
The Spikey system consumes approximately 6 W of power, and the chip itself less than 0.6 W. On the chip most
In
power is consumed by digital communication infrastructure, which is not part of the neuromorphic network.
the following, we estimate the power consumption for a single synaptic event using the data set partly shown in
Figure 5a. This emulation lasts T = 10 s in biological time and generates approximately 45 · 103 spikes. Considering
the acceleration of the hardware network (104) and the synapse count per neuron (K = 15), the system generates
7 · 108 synaptic events per second in hardware time.
If we consider the total power consumption of the Spikey
chip, the upper bound of energy consumed by each synaptic transmission will be approximately 1 nJ. Because these
measurements include the communication infrastructure and other support electronics to observe spike times and
membrane traces, the real energy consumption for synaptic transmissions is estimated to be approximately ten times
smaller. Network simulations on conventional supercomputers a far less energy efficient and consume tens of µJ for
each synaptic transmission [110].
Supplements 2: Modification of the bisection method
In each iteration of the bisection method that is used to calibrate the leak conductances of hardware neurons
(Section II E), we evaluated the firing rate for each neuron by the median over L = 25 identical trials. However, if this
measure is compared between consecutive identical iterations, temporal noise on time scales longer than the duration
of one iteration may still lead to variability. In the original bisection method, the interval of possible solutions is
halved after each iteration step [92]. To improve the convergence of this method in the context of our calibration we
expanded the halved interval by 20% at both ends after each iteration. This prevents the algorithm to get stuck in
an interval wrongly chosen by random fluctuations of the firing rates.
Supplements 3: Effective weights
We quantify the effect of a single spike of neuron j on the firing rate of a postsynaptic neuron i by the effective
weight wij of the connection i ← j. Assuming that the activity of neuron i does not affect the activity of neuron j
(i.e., the RAND case), we define wij as the cross-correlation between the spike trains sj(t) and si(t + τ ), where, due
to causality, τ is positive. Then, we average the effective weight over the emulated network time T , and subtract the
baseline determined by the average correlation for negative τ:
(cid:104)sj(t)si(t + τ )(cid:105)tdτ − 1
τmin
−τmin
(cid:104)sj(t)si(t + τ )(cid:105)tdτ
.
(S1)
(cid:90) τmax
1
wij =
τmax
0
(cid:90) 0
SUP. FIG. 1. Distribution of effective weights w (a; log scale) and population-averaged effective weights ¯w (b) for different
levels of heterogeneity, averaged over M = 100 network realizations. Symbols and error bars denote mean and one standard
deviation, respectively.
0.00.20.40.60.81.0Heterogeneity a0.200.150.100.050.00Effective weight wi (s−1)0.00.20.40.60.81.0Heterogeneity a0.0500.0450.0400.0350.030Avg. effective weight ¯w (s−1)10-410-310-2Rel. frequencyab ii
SUP. FIG. 2. Distribution of average incoming effective weights wi over the rates of the postsynaptic neurons i, for three
different levels of heterogeneity (a = 0, a = 0.625 and a = 1), averaged over M = 100 network realizations. The average
incoming effective weights are obtained by summing over all presynaptic partners ( wi =(cid:80)
j wij).
Here, we chose τmax = 50 ms and τmin = 50 ms. τmax was determined by measuring the average duration in which a
spike from neuron j has an influence on neuron i (data not shown). τmin was then chosen symmetrically. The mass of
the effective weights density shifts towards less negative effective weights for increasing heterogeneity (Supplements
Figure 1a).
We obtain ¯w by averaging over all possible connections:
(cid:88)
i,j
¯w =
1
N K
wij
.
(S2)
For increasing heterogeneity the absolute value of the average effective weight ¯w decreases (Supplements Figure 1b).
This can be explained by the dependence of the effective weight on the firing rate of the postsynaptic neuron (Sup-
plements Figure 2). In the regime of small rates (< 10 s−1), incoming spikes hardly affect the neuron's firing and the
effective weight is small. Similarly, in the case for large rates (> 40 s−1). Neurons with intermediate firing rates are
sensitive to input and hence have a more negative effective weight. Heterogeneity increases the number of neurons
with small and large firing rates, and hence the absolute value of the average effective weight decreases, which in turn
weakens the effective negative feedback of the network.
Supplements 4: Simulations with software
We validate our results by comparing them to simulations with software (NEST [101], PyNN [102]).
In these
simulations we modulated the degree of heterogeneity by distributing the firing thresholds of all neurons according to
a normal distribution with mean Θ and standard deviation σΘ. Details about the network, neuron and synapse models
and their parameters can be found in Supplements Table I and II, respectively. The results are qualitatively similar
to network emulations on the Spikey chip (compare Figure 8 to 9) and also hold for larger network sizes (Supplements
Figure 3). In the FB case, input correlations and output correlations increase with network heterogeneity. In contrast
to hardware emulations, input correlations stay approximately constant in the RAND case. Output correlations
strongly decrease with the standard deviation σΘ. Here, heterogeneity only affects the output spike times, and not
the integrative properties of the neurons (see also Section III B and [70]).
In addition, we compare the effect of heterogeneity in firing thresholds on correlations to that of heterogeneity
in synaptic weights. Details about the network, neuron and synapse models and their parameters can be found
in Supplements Table I and III, respectively. If we distribute synaptic weights, shared-input correlations decrease
with the width of the weight distribution (Supplements Figure 4). Output correlations are also reduced, however,
only proportionally to the reduction of input correlations (compare insets in Supplements Figure 4). While output
correlations are overall smaller than input correlations due to the non-linearity in spike generation, we do not observe
a boost of this decrease for large heterogeneities. Overall, the dynamics of the recurrent system is more sensitive to
heterogeneities in firing thresholds than in synaptic weights (compare scale of abscissas of Figure 9 to Supplements
Figure 4).
To investigate the generalization of our results to mixed excitatory-inhibitory networks, correlations were measured
04080120160Rate of neuron i (s−1)0.20000.12670.05330.0200Effective weight wi (s−1)a=1.004080120160a=0.62504080120160a=0.010-410-310-2Rel. frequency iii
SUP. FIG. 3.
Like Figure 9, but for N = 4800 neurons and K = 375 inputs for each neuron. Note the different scales of
the ordinate in Figure 9 and (b). Symbols and error bars denote mean and standard deviation, respectively, across M = 20
network realizations (error bars are partly covered by markers).
SUP. FIG. 4. Dependence of population-averaged input correlations (a), and spike-train correlations (b) on the width of
the weight distribution, for the intact network (FB, dark gray circles) and the RAND (light gray circles) case. Symbols and
error bars denote mean and standard deviation, respectively, across M = 30 network realizations (error bars are partly covered
by markers). The gray line in (a) depicts shared-input correlations in a homogeneous network (Equation 2). The insets show
correlations normalized to unity at σJ = 0, with the same abscissa as in the main plot. Note that in simulations the FBreplay
is identical to the FB case, and is hence not shown. Networks simulated with NEST [101] and PyNN [102].
in a network of N = 192 neurons consisting of half excitatory and half inhibitory neurons. Details about the network,
neuron and synapse models and their parameters can be found in Supplements Table IV and V, respectively. As
described above, we distribute the firing thresholds of all neurons according to a normal distribution. The results
are consistent with those obtained from purely inhibitory networks on hardware demonstrating the generality of our
findings (compare Figure 8 to Supplements Figure 5). By modulating the level of heterogeneity separately for the
excitatory or the inhibitory population, we observe that network dynamics are more sensitive to heterogeneities in
the inhibitory than in the excitatory population (data not shown).
We investigate the effect of temporal noise on correlations by dithering spikes in the FBreplay case before replaying
them to the network (for network, neuron and synapse models see Supplements Table I and II). Each spike time t is
replaced by a spike time t(cid:48) randomly drawn from a normal distribution with width ϑ: t(cid:48) ∼ N (t, ϑ). Even for small
ϑ, correlations increase on the input as well as on the output side, demonstrating the sensitivity of correlations to
perturbations in the feedback loop (see Supplements Figure 6 and [5]). This effect is also reflected in an increase
of the power of the population activity at small frequencies (see Supplements Figure 7). These results suggest that
on hardware temporal noise is responsible for the increase of correlations and population power in the FBreplay case
(compare FBreplay to FB in Figure 5, 7 and 8).
0.000.020.040.060.080.100.120.14Rel. std. deviation of threshold potential σΘµΘ0.0000.0020.0040.0060.008Correlation κS0.000.020.040.060.080.100.120.14Rel. std. deviation of threshold potential σΘµΘ0.000.020.040.060.080.10Correlation κVFBRAND0.00020.00010.0000ab0.00.51.01.52.02.53.0Rel. std. deviation of weights σJµJ0.010.000.010.020.030.040.05Correlation κS0.00.51.01.52.02.53.0Rel. std. deviation of weights σJµJ0.000.020.040.060.080.10Correlation κVFBRAND0.51.00.51.0ab iv
SUP. FIG. 5. Dependence of population-averaged input correlations (a), and spike-train correlations (b) on the width of the
threshold distribution, for the intact network (FB, dark gray circles) and the RAND (light gray circles) case. The network
consists of two populations, one excitatory and one inhibitory, of equal size. Symbols and error bars denote mean and standard
deviation, respectively, across M = 30 network realizations (error bars are partly covered by markers). The gray line in (a)
depicts shared-input correlations in a homogeneous network (Equation 2). The inset in (b) shows a magnified view of the
spike-train correlations in the FB case. Note that in simulations the FBreplay is identical to the FB case, and is hence not
shown. Networks are simulated with NEST [101] and PyNN [102].
SUP. FIG. 6. Dependence of population-averaged input correlations (a), and spike-train correlations (b) on the strength of
temporal noise, for the FBreplay (black circles) and the RAND (light gray circles) case. Symbols and error bars denote mean
and standard deviation, respectively, across M = 30 network realizations (error bars are partly covered by markers). Gray
curve in (a) depicts shared-input correlations in a homogeneous network (Equation 2). Correlations in the FB case correspond
to correlations in the FBreplay case for ϑ = 0 ms. Networks simulated with NEST [101] and PyNN [102].
Supplements 5: Linear model
We investigate the consistency of our results with a linear rate model that allows us to numerically calculate the
average correlations from a given connectivity matrix W. The model is defined as (according to, e.g., [100])
r(t) = (W(r + x) ∗ h)(t)
.
(S3)
Here, r(t) denotes the rate of the individual neurons and x(t) a Gaussian white noise input that is independent for
each neuron. The linear filter kernel h(t) depends on the details of the model, is not relevant in our calculation, and
hence is not further specified here. Equation S3 can be transformed to Fourier domain, where the input and output
0.000.020.040.060.080.100.120.14Rel. std. deviation of threshold potential σΘµΘ0.010.000.010.020.030.040.05Correlation κS0.000.020.040.060.080.100.120.14Rel. std. deviation of threshold potential σΘµΘ0.000.020.040.060.080.10Correlation κVFBRANDEEEIII0.0100.0050.000ab0102030405060Dither of input spikes ϑ (ms)0.010.000.010.020.030.040.05Correlation κS0102030405060Dither of input spikes ϑ (ms)0.000.020.040.060.080.10Correlation κVRANDFBreplayab v
SUP. FIG. 7. Power spectrum of the population activity of the networks in Supplements Figure 6 for the intact network (solid
dark gray), different degrees of dither on input spikes (dashed, dashed dotted and dotted black) and the RAND case (solid
light gray), averaged across M = 30 network realizations.
spectral matrices can be expressed by
(S4)
(S5)
with T (ω) = (1 − H(ω)W)−1 [100]. In the RAND case, the linear equation for the rate of the (unconnected) neurons
reads
CRR(ω) = T (ω)T (ω)†
,
RR(ω) = WCRR(ω)WT
Cin
,
q(t) = (W(r + x) ∗ h)(t)
,
(S6)
where r(t) has the same auto-correlations as r(t) but zero cross correlations, i.e., C R R = diag(CRR), since the
randomization of spike times removes all spatio-temporal correlations. According to Tetzlaff et al. [5] spectral matrices
in the RAND case are given by
We calculate the population-averaged power- and cross-spectra from the full matrices:
¯AX(ω) =
1
N
CXX,ii
(cid:88)
i
1
QQ(ω) =WC R RWT
Cin
CQQ(ω) =H(ω)2(Cin
,
QQ + ρ)
.
¯CXX(ω) =
N (N − 1)
,
CXX,ij
.
(cid:88)
i(cid:54)=j
(S7)
(S8)
(S9)
(S10)
Here, X ∈ {R, Q} denotes the FB and RAND case, respectively. The low frequency coherence is the cross-spectra
normalized by the power spectra:
κX (0) =
¯CXX(0)
¯AX(0)
.
(S11)
Note that in the linear model we are actually taking the zero frequency coherence.
As in the spiking model, we consider a sparse network, i.e., we randomly choose for each neuron i ∈ [1, N ] an
identical number of presynaptic partners (K = 15). In the linear model we do not consider a distribution of non-zero
effective weights. Instead, each realized connection is assigned the same weight value −w. To mimic the effect of
calibration we vary the absolute value of the effective weight by scaling the weights of the non-zero connections with
100101102Frequency f (Hz)10-1100101102Power spectrum ¯A (s−1)FBDither (2ms)Dither (10ms)Dither (40ms)RAND vi
SUP. FIG. 8. Dependence of population-averaged input correlations (a), and output correlations (b) on the absolute magnitude
of the effective weights ("heterogeneity" a, see text), for the intact network (FB, dark gray diamonds) and the RAND (light
gray circles) case. Gray curve in (a) depicts shared-input correlations in a homogeneous network (Equation 2). The inset in
(b) shows a magnified view of the spike-train correlations in the FB case (dark gray diamonds). Note that in the linear model
the FBreplay is identical to the FB case, and is hence not shown.
SUP. FIG. 9. Distributions of effective weights in a sparse network. Different values for the mean µw and standard deviation
σw of the distribution of non-zero weights in (a) and (b), respectively, result in different values for the mean µW and standard
deviation σW of the effective weight matrix.
a sigmoidal function of a ∈ [0, 1]:
w =
1
1 + e10×(a−0.5)
w .
(S12)
This procedure changes the variance of the weight matrix [74] and hence a is denoted the heterogeneity of the
network. More homogeneous (heterogeneous) networks have larger (smaller) effective weights and hence stronger
(weaker) feedback. We obtain qualitatively similar results as we observe on the Spikey chip (compare Supplements
Figure 8 to Figure 8). Output correlations in the RAND case decrease, while both input and output correlations in
the FB case increase with network heterogeneity, i.e., with the variance of the effective weight matrix. Note that, in
this simplified model, input correlations in the RAND case are not affected by the level of "heterogeneity" because
the non-zero weights are homogeneous, i.e. have zero variance, irrespectively of a (cf. Equation 3).
In Supplements Figure 9 we illustrate, how in a sparse network the variance of the weight matrix can increase,
although the distribution of non-zero weights becomes narrower. The standard deviation σw of a distribution of
non-zero weights with mean µw is (in this example) smaller than the standard deviation σW of the full effective weight
matrix, due to the sparseness of the matrix (Supplements Figure 9a; here, we chose = 0.8). If we, at the same
time, increase the mean µw and decrease the standard deviation σw of non-zero weights, the standard deviation of
the weight matrix σW can increase significantly (Supplements Figure 9). While the distribution of effective weights
0.00.20.40.60.81.0Heterogeneity a0.010.000.010.020.030.040.05Output correlations0.00.20.40.60.81.0Heterogeneity a0.000.020.040.060.08Input correlationsFBRAND-0.004-0.0020.000ab0.00.20.40.60.8Effective weight (a.u.)10-310-210-1100Relative countµw=0.37, σw=0.04,µW=0.21, σW=0.190.00.20.40.60.8Effective weight (a.u.)µw=0.67, σw=0.01,µW=0.37, σW=0.33ab vii
SUP. FIG. 10.
CVISI distributions (0.54, 0.53, 0.5) are marked with dashed lines. Percentage of dead neurons: 0.00%, 0.02%, 0.08%.
(a–c) Like Figure 6d–f, but for the RAND case. The mean of firing rate (24.0 s−1, 25.2 s−1, 27.8 s−1) and
Firing-rate dependence of pairwise low-frequency spike-train coherence in the uncalibrated (a) and fully
SUP. FIG. 11.
calibrated system (b). Each bin represents the low-frequency spike-train coherence κij =
, averaged
across all neuron pairs {i, j} with firing rates in the interval [ri, ri + ∆r) and [rj, rj + ∆r). Here, Re(Cij(f )) denotes the real
part of the complex cross-spectrum at frequency f. fmin = 0.1 Hz, fmax = 20 Hz, ∆r = 5 s−1. Data averaged across M = 100
network realizations.
Re(Cij (f ))
Ai(f )Aj (f )
fmax−fmin
1
(cid:82) fmax
fmin
√
is broadened, the mean is decreased, which has a greater impact on the size of correlations. This observation could
explain the decrease of correlations with increased calibration of the neuromorphic chip.
Supplements 6: Firing statistics in the RAND case
The firing rate distributions in the RAND case are similar to those in the FB scenario (compare Supplements
Figure 10 to Figure 6d–f). As in the FB case they become broader for increasing heterogeneity. Due to a larger
variance of the membrane potentials, inactive neurons in the FB case have a higher probability to fire in the RAND
case (compare gray bars in horizontal histograms of Supplements Figure 10a–c and Figure 6d–f). Neurons with firing
rates larger than the average firing rate are strongly driven by constant current influx, and hence show similar firing
rates than in the FB scenario. The irregularity of firing increases for the RAND compared to the FB case. This can
also be traced back to an increased variance of the membrane potential for the RAND case (data not shown).
Supplements 7: Correlation matrices
In addition to the population-averaged measures from the main manuscript, we calculated the pairwise correlations
for each pair i, j of neurons with i (cid:54)= j ∈ [1, N ], and ordered these by the time-averaged rate of the corresponding
04080120Rate r (s−1)0.00.51.01.52.02.5CVISI0408012004080120abc0255075100125150Rate (s−1)a=10255075100125150Rate (s−1)0255075100125150Rate (s−1)a=00.0100.0080.0060.0040.0020.0000.0020.0040.0060.0080.010Correlationab viii
SUP. FIG. 12. Like Figure 4 and 8 in (A) and (B), respectively, but for two additional chips of version 5.
neurons (Supplements Figure 11). This reveals a dependence of the pairwise correlation on the firing rate of the
respective neurons. If both neurons fire at low rate (here < 5 s−1), correlations will be close to zero similar to the
results in [91]. For high rates (here > 25 s−1) we find mostly positive correlations. However, for neurons firing
at intermediate rates, the activity of the pair is often anti-correlated, and can hence effectively suppress (positive)
shared-input correlations. After calibration, the amount of neurons firing at intermediate rates and showing negative
correlations increases (Supplements Figure 11b, Figure 6). Shared-input correlations are hence suppressed by more
neurons leading to smaller input correlations than in the uncalibrated case (Figure 8).
Supplements 8: Results for different Spikey chips
We have performed the same experiments as described in the main text using two additional Spikey chips of
the current version (5). Different chips show different realizations of fixed-pattern noise, and hence calibration was
repeated for each chip separately (Figure 4, Supplements Figure 12A). For all chips we find qualitatively the same
results as described above: in the FB case, free membrane potentials (and spike trains) are decorrelated by inhibitory
feedback and correlations increase with an increasing level of heterogeneity (Figure 8, Supplements Figure 12B).
This also applies to the results of identical experiments on three previous-generation chips (version 4). These old
chips show more pronounced heterogeneity in their parameters. In addition, we observe larger quantitative differences
between old chips for the uncalibrated case, which is likely to be caused by different extents of intrinsic fixed-pattern
noise. Unfortunately, the old systems show an artifact affecting the population power at around 100 Hz (Supplements
Figure 13), consistently across all chips of identical revision. We can exclude that this effect can be traced back
to network effects (it occurs across the FB, FBreplay and RAND cases) or to single-cell power spectra (data not
shown). We hence have to consider it an artificial spurious synchronization. As the artifact lies outside the range of
frequencies that are relevant for our measures of correlation it does not affect the main results and experiments on
the old chips qualitatively confirm our results in the main manuscript (compare Supplements Figure 14 to Figure 4,
8 and Supplements Figure 12).
Supplements 9: Reproducibility of networks with intact feedback
We measured the standard deviation of free membrane potential and spike train correlations over several trials for
a single network realization (Supplements Figure 15). This standard deviation is smaller than the standard deviation
we observe over different network realizations (compare to Figure 8 and Supplements Figure 12), which indicates that
(A)Calibration050100150Rate r(s−1)050100150200Neuron counta=1a=00.00.51.0Heterogeneity a010203040∆r (s−1)ab(B)Correlation0.00.20.40.60.81.0Heterogeneity a0.010.000.010.020.030.040.05Correlation κSSpike trains0.00.20.40.60.81.0Heterogeneity a0.000.020.040.060.08Correlation κVFree membrane potentialsRANDFBreplayFB0.00380.00320.0050.0100.0150.020ab(A)Calibration050100150Rate r(s−1)050100150200Neuron counta=1a=00.00.51.0Heterogeneity a010203040∆r (s−1)ab(B)Correlation0.00.20.40.60.81.0Heterogeneity a0.010.000.010.020.030.040.05Correlation κSSpike trains0.00.20.40.60.81.0Heterogeneity a0.000.020.040.060.08Correlation κVFree membrane potentialsRANDFBreplayFB0.00380.00320.0050.0100.0150.020ab ix
SUP. FIG. 13. Spectra of trial-averaged population power ¯A(f ) and trial-averaged single-cell power A(f ) for chip 666 (version
4, left) and chip 508 (version 5, right) illustrating the synchronization artifact in the 100 Hz range for the intact network (FB,
dark gray), the FBreplay (black) and the RAND (light gray) case.
SUP. FIG. 14. Like Figure 4 and 8 in (A) and (B), respectively, but for three chips of version 4.
100101102Frequency f (Hz)10-1100101102Power spectrum ¯A(s−1)10010110210-1100101102A(s−1)100101102Frequency f (Hz)10-1100101102Power spectrum ¯A(s−1)10010110210-1100101102A(s−1)(A)Calibration050100150Rate r(s−1)050100150Neuron counta=1a=00.00.51.0Heterogeneity a010203040∆r (s−1)ab(B)Correlation0.00.20.40.60.81.0Heterogeneity a0.010.000.010.020.030.040.050.06Correlation κSSpike trains0.00.20.40.60.81.0Heterogeneity a0.000.020.040.060.08Correlation κVFree membrane potentialsRANDFBreplayFB0.0040.003ab(A)Calibration050100150Rate r(s−1)050100150Neuron counta=1a=00.00.51.0Heterogeneity a010203040∆r (s−1)ab(B)Correlation0.00.20.40.60.81.0Heterogeneity a0.010.000.010.020.030.040.050.06Correlation κSSpike trains0.00.20.40.60.81.0Heterogeneity a0.000.020.040.060.08Correlation κVFree membrane potentialsRANDFBreplayFB0.0040.003ab(A)Calibration050100150Rate r(s−1)050100150Neuron counta=1a=00.00.51.0Heterogeneity a010203040∆r (s−1)ab(B)Correlation0.00.20.40.60.81.0Heterogeneity a0.010.000.010.020.030.040.050.06Correlation κSSpike trains0.00.20.40.60.81.0Heterogeneity a0.000.020.040.060.08Correlation κVFree membrane potentialsRANDFBreplayFB0.0040.003ab x
SUP. FIG. 15. Like Figure 8 and for the same chip, but for L = 20 trials of one network realization.
the latter is mostly caused by the different realizations of the connectivity, not by trial-to-trial variability. Note that
the variability between trials of networks with intact feedback is likely to be larger than between replays of network
activity as shown in Figure 3, because network dynamics may be chaotic. The data shown in Supplements Figure 15
has to be interpreted with care, because the reproducibility of only a single network realization is considered. For
different realizations, the mean and standard deviation will in general be different.
Supplements 10: Extraction of effective PSP amplitudes, time constants and delays
In order to provide an estimation for the order of magnitude of various neuron parameters, we fitted the postsynaptic
potentials (PSPs) of current-based LIF neurons with exponential postsynaptic currents (PSCs) to the spike-triggered
average (STA) of free membrane potentials obtained from hardware emulations. The differential equations describing
the dynamics of membrane potentials Vj(t) and synaptic currents Ij(t) for this simplified model are given by
τ eff
m
τ eff
syn
Vj(t) = − Vj(t) + RIj(t)
Ij(t) = − Ij(t) + τ eff
syn
Jjksk(t − deff) .
(S13)
(S14)
(cid:88)
k
syn the synaptic time constant and Jjk
where R is the resistance of the membrane, τ eff
the synaptic weight of the connection from neuron k to j. Here the superscript eff indicates that these are not the
parameters realized on hardware, but of an abstract model that we fit to hardware measurements. Data were obtained
from 100 network realizations of the RAND case (same dataset as for Figure 8). For each neuron we computed the
STA of its free membrane potential individually for each spike train of its presynaptic partners. The PSP for the
current based model described above is given by
m the membrane time constant, τ eff
(cid:16)
V (t) =V eff
max
τ eff
syn
syn − τ eff
τ eff
m
(cid:17)
e−(t−deff)/τ eff
syn − e−(t−deff)/τ eff
m
Θ(t − deff) ,
(S15)
max, τ eff
where we introduced V eff
syn, τ eff
parameters: V eff
max = RJ to simplify the fitting procedure. For each pair of neurons i, j we obtain a set of
The fitted curves (using the function scipy.optimize.curve_fit with default parameters from Python's scipy module
v0.17.0) match fairly well to the empirical data to allow an estimation of the effective values of the neuron parameters
governing the shape of the postsynaptic potentials (Supplements Figure 16a). However, note that the resulting
parameter distributions do not precisely reflect the actual parameter distributions on the chip.
m and deff.
The distributions of neuron and synapse parameters change slightly with calibration. For the calibrated network the
m and deff are larger than for the uncalibrated network (see Table II and Supplements
max, τ eff
syn, τ eff
absolute values of V eff
Figure 16b).
0.01.0Heterogeneity a-0.010.00.010.020.030.040.050.06Correlation κSSpike trainsFBFBreplayRAND0.01.0Heterogeneity a0.00.020.040.060.08Correlation κVFree mem. pot.ab xi
SUP. FIG. 16. (a) Fits of theoretical PSPs (dashed; for model, see Equation S15) to STAs of free membrane potentials in the
FB case (solid). (b) Distributions of effective parameters obtained from fitting. Black and gray lines correspond to the fully
calibrated (a = 0) and uncalibrated (a = 1) network, respectively. Median, 25th and 75th percentile of all fitted parameters
are reported in Table II.
50510152025303540Time (ms)4.03.53.02.52.01.51.00.50.00.5STA (mV)a6050403020100Synaptic weight RJ10-310-210-1Rel. count01020304050Membrane time constant τm0.000.020.040.060.080.100.120.140510152025Synaptic time constant τs0.000.050.100.150.200.25Rel. count01234567Delay d0.00.20.40.60.81.01.21.41.6b xii
A
Populations
Topology
Connectivity
Neuron model
Channel models
Synapse model
Plasticity
External input
Measurements
Other
B
Name
I
Model summary
One (inhibitory)
-
Random convergent connections (fixed in-degree)
Leaky integrate-and-fire (LIF), fixed firing threshold, fixed absolute refractory time
-
Exponentially decaying currents, fixed delays
-
Resting potential higher than threshold (= constant current) (El > Θ)
Spikes and membrane potentials
No autapses, no multapses
Populations
Elements
LIF neuron
Size
N
C
Source Target Pattern
I
I
Connectivity
Random convergent connect, in-degree K
Neuron and synapse model
Leaky integrate-and-fire, exponential currents
Subthreshold dynamics (t (cid:54)∈ (t∗, t∗ + τref)):
Reset and refractoriness (t ∈ (t∗, t∗ + τref)):
dt v(t) = −gl(v(t) − El) + Isyn(t)
Cm
d
v(t) = vreset
dt Isyn(t) = −Isyn(t) +(cid:80)
d
τsyn
i,k Jδ(t − tk
i )
D
Type
Subthreshold
dynamics
Current dynamics
Spiking
E
Spike trains
Membrane potentials
Here the sum over i runs over all presynaptic neurons and the sum over k over all spike
times of the respective neuron i
If v(t∗−) < Θ ∧ v(t∗+) ≥ Θ:
emit spike with time stamp t∗
Measurements
recorded from Ps neurons
recorded from Pv neurons
SUP. TABLE I. Description of the network model (according to [109]).
xiii
B
Name
N
C
Name
K
D
Name
Cm
τref
vreset
El
Θ
gl
D
Name
τsyn
J
d
E
Name
Ps
Pv
Values
{192, 4800}
Values
{15, 375}
Values
0.2 nF
0.1 ms
−80 mV
−52 mV
∼ N (−62, [0, 8.8]) mV
10 nS
Values
5 ms
−0.254 nA
1.0 ms
Values
{192, 4800}
{150, 150}
Populations
Description
network size
Connectivity
Description
number of presynaptic partners
Neuron
Description
membrane capacitance
refractory period
reset potential
resting potential
firing threshold
leak conductance
Synapse
Description
synaptic time constant
synaptic weight
synaptic delay
Measurements
Description
number of neurons spike trains are recorded from
number of neurons membrane potentials are recorded from
SUP. TABLE II. Parameter values for the network model described in Supplements Table I with distributed thresholds.
xiv
B
Name
N
C
Name
K
D
Name
Cm
τref
vreset
El
Θ
gl
D
Name
τsyn
J
d
E
Name
Ps
Pv
Values
192
Values
15
Values
0.2 nF
0.1 ms
−80 mV
−52 mV
−62 mV
10 nS
Populations
Description
network size
Connectivity
Description
number of presynaptic partners
Neuron
Description
membrane capacitance
refractory period
reset potential
resting potential
firing threshold
leak conductance
Values
5 ms
∼ [N (0.254, [0.00254, 0.75])]+
1.0 ms
Synapse
Description
synaptic time constant
synaptic weight, clipped to positive values
synaptic delay
Values
192
150
Measurements
Description
number of neurons spike trains are recorded from
number of neurons membrane potentials are recorded from
SUP. TABLE III. Parameter values for the network model described in Supplements Table I with distributed weights.
xv
A
Populations
Topology
Connectivity
Neuron model
Channel models
Synapse model
Plasticity
External input
Measurements
Other
B
Name
E
I
Model summary
Two (excitatory, inhibitory)
-
Random convergent connections (fixed in-degree)
Leaky integrate-and-fire (LIF), fixed firing threshold, fixed absolute refractory time
-
Exponentially decaying currents, fixed delays
-
Resting potential higher than threshold (= constant current) (El > Θ)
Spikes and membrane potentials
No autapses, no multapses
Populations
Elements
LIF neuron
LIF neuron
Size
NE
NI
Connectivity
Random convergent connect, in-degree KE, weight JE
Random convergent connect, in-degree KE, weight JE
Random convergent connect, in-degree KI, weight JI
Random convergent connect, in-degree KI, weight JI
Neuron and synapse model
Leaky integrate-and-fire, exponential currents
Subthreshold dynamics (t (cid:54)∈ (t∗, t∗ + τref)):
Reset and refractoriness (t ∈ (t∗, t∗ + τref)):
dt v(t) = −gl(v(t) − El) + Isyn(t)
Cm
d
v(t) = vreset
dt Isyn(t) = −Isyn(t) +(cid:80)
d
τsyn
i,k Jδ(t − tk
i )
C
Source Target Pattern
E
E
I
I
E
I
E
I
D
Type
Subthreshold
dynamics
Current dynamics
Spiking
E
Spike trains
Membrane potentials
Here the sum over i runs over all presynaptic neurons and the sum over k over all spike
times of the respective neuron i
If v(t∗−) < Θ ∧ v(t∗+) ≥ Θ:
emit spike with time stamp t∗
recorded from P E
recorded from P E
s
v
Measurements
excitatory and P I
excitatory and P I
s
v
inhibitory neurons
inhibitory neurons
SUP. TABLE IV. Description of the network model consisting of an excitatory and an inhibitory population (according to
[109]).
B
Name
NE
NI
C
Name
KE
KI
D
Name
Cm
τref
vreset
El
Θ
gl
D
Name
τsyn
JE
JI
d
E
Name
P E
s
P I
s
P E
v
P I
v
Values
96
96
Values
7
8
Values
0.2 nF
0.1 ms
−80 mV
−52 mV
−62 mV
10 nS
Values
5 ms
0.0635 nA
−0.254 nA
1.0 ms
Values
96
96
150
150
xvi
Populations
Description
size of the excitatory population
size of the inhibitory population
Connectivity
Description
number of excitatory presynaptic partners
number of inhibitory presynaptic partners
Neuron
Description
membrane capacitance
refractory period
reset potential
resting potential
firing threshold
leak conductance
Synapse
Description
synaptic time constant
excitatory synaptic weight
inhibitory synaptic weight
synaptic delay
Measurements
Description
number of excitatory neurons spike trains are recorded from
number of inhibitory neurons spike trains are recorded from
number of excitatory neurons membrane potentials are recorded
from
number of inhibitory neurons membrane potentials are recorded
from
SUP. TABLE V. Parameter values for the network model described in Supplements Table IV.
|
1705.08261 | 1 | 1705 | 2017-05-17T15:51:37 | Warnings and Caveats in Brain Controllability | [
"q-bio.NC",
"physics.bio-ph"
] | In this work we challenge the main conclusions of Gu et al work (Controllability of structural brain networks. Nature communications 6, 8414, doi:10.1038/ncomms9414, 2015) on brain controllability. Using the same methods and analyses on four datasets we find that the minimum set of nodes to control brain networks is always larger than one. We also find that the relationships between the average/modal controllability and weighted degrees also hold for randomized data and the there are not specific roles played by Resting State Networks in controlling the brain. In conclusion, we show that there is no evidence that topology plays specific and unique roles in the controllability of brain networks. Accordingly, Gu et al. interpretation of their results, in particular in terms of translational applications (e.g. using single node controllability properties to define target region(s) for neurostimulation) should be revisited. Though theoretically intriguing, our understanding of the relationship between controllability and structural brain network remains elusive. | q-bio.NC | q-bio | Warnings and Caveats in Brain Controllability
Chengyi Tu1,7, Rodrigo P. Rocha1,2,7, Maurizio Corbetta3,4,7, Sandro Zampieri5,7, Marzo Zorzi6,7,8 & S.
Suweis1,7∗
1Dipartimento di Fisica e Astronomia, 'G. Galilei' & INFN, Università di Padova, Padova, IT.
2Departamento de Física, Universidade Federal de Santa Catarina, 88040-900, Florianópolis-SC, Brazil.
3Dipartimento di Neuroscienze, Università di Padova, Padova, IT. 4Departments of Neurology, Radiology,
Neuroscience, and Bioengineering, Washington University, School of Medicine, St. Louis, USA.
5Dipartimento di Ingegneria dell'informazione, Università di Padova, Padova, IT. 6Dipartimento di
Psicologia Generale, Università di Padova, Padova, IT. 7Padua Neuroscience Center, Università di
Padova, Padova, IT. 8IRCCS San Camillo Hospital Foundation, Venice, IT,
There is large consensus that the complex, self-organizing structure of the human brain can be well described
by the mathematical framework based on network theory. A recent article by Gu et al.1 proposed to
characterize brain networks in terms of their "controllability", drawing on concepts and methods of control
theory. The analysis of controllability has the potential to unveil how specific nodes and/or sets of nodes
control the dynamics of the entire network1-4 and thus might provide insights on whether manipulating the
local activity of specific nodes would fully or partially restore network functions after brain damage. Gu et
al. applied the tools of control theory to quantify how the networks structure, defined by human connectome
data (i.e., white matter pathways derived from diffusion tensor or diffusion weighted imaging), constrain or
facilitate changes in brain state trajectories. They proposed that single brain regions that are densely
connected facilitate the brain to easily reach many of its cognitive states (i.e. high average controllability),
while weakly connected nodes promote the movement of the brain to difficult-to-reach states (i.e. high
modal controllability). In other words, they claim that brain activity is controllable by a single node, and
1
topology of brain networks provides an explanation for the types of control roles that different regions play in
the brain. Here we present new results based on the analysis of four dataset and numerical simulations that
challenge their main conclusion and undermine their use of the controllability framework.
In order to study single node controllability (in Kalman sensu), the authors first normalize the matrix
describing the brain structural connectivity, and then calculate the controllability Gramian from each node
(see Methods). Their first finding is that the brain can be theoretically controllable by a single region/node,
i.e. the smallest (in absolute value) of the eigenvalues of the controllability Gramian 𝜆!"# (see Methods) from
each brain region as a control node is greater than zero. Then they measure the average controllability and
the modal controllability from each node (see Methods), and find that average controllability is strongly
correlated with the weighted degree, and the modal controllability is strongly anti-correlated with the
weighted degree. Finally they evaluate the control contribution of different brain regions associated with
known cognitive process (i.e. resting state networks (RSNs)) finding that different RSNs have different
control roles.
To investigate further their results, we apply the controllability framework on four empirical open access
datasets of large-scale structural brain connectivity from different studies5–8. In contrast to Gu et al.1, we
observe that for all the analyzed brain networks 𝜆!"# is not greater than zero (we note that already in their
region as a control node is 𝜆!"# = 2.5×10!!"± 4.8×10!!" and thus it is not significantly greater than
work, also Gu et al.1 have found that the smallest eigenvalue of their controllability Gramian from each brain
zero). Indeed, according to our data, the controllability Gramian from each node is not invertible. Therefore, it
is not possible to assess the global controllability of the brain performed by a single region/node. We highlight
that the structural properties of the DTI/DSI network data (symmetric with no self-loops) crucially affect the
controllability of the system. Moreover, it is important also to quantify the energy needed to control the system
from a specific region. In fact, although Gu et al.1 have found that on average 𝜆!"# ~10!!", in order to control
such a system from a single region (node) would need an energy ε≈𝜆!"#!!~10!", in practice the system is not
2
controllable. Our results are in accordance with prior theoretical results9 showing that for symmetric networks
(like DTI/DSI brain connectivity data) the energy ε needed to control from a single node the system grows
exponentially with the network size (i.e. to practically control the system (ε not too large) a fixed fraction of
the networks nodes is necessary9).
Next, we study the role of nodes topological properties in controlling the brain network. Inspired by Gu and
collaborators claim on the importance of nodes centrality, we first rank all networks nodes according to five
distinct centrality measures. Then we build two ranking lists, following respectively the ascending and
descending order of centralities magnitude. Then we generate a list of nodes in random order, irrespective of
the nodes properties. For each of these three nodes sequences, we look for the minimum subsets of nodes
needed to control the brain network. We have repeated this experiment for all brain connectivity datasets as
well as for synthetic brain networks generated with a desired network topology (see Methods and Table 1). If
topology is important we expect that the minimum number of control nodes will depend on the specific nodes
properties (e.g. the most or the least central nodes). Our results are presented in Table 1. It shows the size of
the minimum subset of control nodes with respect to five different centralities measures for the APOE-4 data5
(𝑁=110 nodes), random Barabasi-Albert (BA) scale free networks10 and uniform networks10, both with the
same number of nodes, edges, and edge weight distribution. No matter which centrality measure is adopted and
how we generated the node sequences, the size of the minimum subset is always greater than one and similar in
all three cases (see Table 1). This result strongly suggests that the topological structure of the DTI/DSI
connectivity networks do not play an important role in brain controllability. We thus want to better understand
the relationships between average/modal controllability and weighted degree and test if it is due to the specific
brain network topology, or rather if it holds also for random networks obtained by randomizing the structure of
real brain data, indeed, even if the controllability Gramian is not invertible, the average/modal controllability
can still be calculated. Following Gu et al. (Fig. 2 b and d in1), we plot the average of the ranks for weighted
degrees versus average of the ranks of the average controllability in Fig. 1a, and the average of ranks for
weighted degrees versus the average of ranks of the modal controllability in Fig. 1c, (for details on average
3
rank plots see Gu et al.'s methods section) of our four dataset. As Fig. 1a and 1c show, we find the same
relation between controllability and network centrality reported by Gu et al. We then randomize the structural
connectivity data and calculate average/modal controllability as a function of nodes properties in the random
networks. Surprisingly, as shown in Fig. 1b and 1d, we found exactly the same relationship as before. We
stress that the relationships in average controllability and modal controllability between real and randomized
networks are not an artifact of the rank-rank plots. In fact, controllability measures between real and
randomized networks are not distinguishable also by examining unranked magnitudes.
Finally we want to test if different RSNs are involved in different modes of controllability. To do that, we have
built RSNs for the Hagmann dataset (based on RSNs template in the literature11). Following Gu et al.
procedure, we consider the thirty nodes with highest average or modal controllability, and then calculate a
vector cdata, where its j-th component gives the percentage of the nodes that belongs to the j-th RSN. We then
repeat the same analysis for a proper random null model10,12, obtaining crand from random networks with the
same degree distribution of the original dataset, but with no spatial correlations among regions (as observed in
real RSNs). We find that both real and randomized data display very similar results on the modes of
controllability of the different RSNs; indeed the Pearson correlation between cdata and crand is ρ =0.7 (P-value ≈
0.05). In other words we find that there are not specific roles played by RSNs in controlling the brain.
In summary, our results challenge the main conclusions of Gu et al work on brain controllability. Using the
same methods and analyses on four datasets we find that the minimum set of nodes to control brain networks is
always larger than one. We also find that the relationships between the average/modal controllability and
weighted degrees also hold for randomized data and the there are not specific roles played by RSNs in
controlling the brain. In conclusion, we show that there is no evidence that topology plays specific and unique
roles in the controllability of brain networks. Accordingly, Gu et al. interpretation of their results1, 13–15, in
particular in terms of translational applications (e.g. using single node controllability properties to define target
region(s) for neurostimulation) should be revisited. Though theoretically intriguing, our understanding of the
relationship between controllability and structural brain network remains elusive.
4
y
t
i
l
i
b
a
l
l
o
r
t
n
o
c
e
g
a
r
e
v
a
f
o
k
n
a
r
f
o
e
g
a
r
e
v
a
y
t
i
l
i
b
a
l
l
o
r
t
n
o
c
l
a
d
o
m
f
o
k
n
a
r
f
o
e
g
a
r
e
v
a
250
a
200
150
100
50
0
250
200
150
100
50
0
c
0
0
y
t
i
l
i
b
a
l
l
o
r
t
n
o
c
e
g
a
r
e
v
a
f
o
k
n
a
r
f
o
e
g
a
r
e
v
a
y
t
i
l
i
b
a
l
l
o
r
t
n
o
c
l
a
d
o
m
f
o
k
n
a
r
f
o
e
g
a
r
e
v
a
250
b
200
150
100
50
0
0
250
d
200
150
100
50
50
0
0
50
Rockland
APOE-4
Autism
Hagmann
50
100
150
200
250
average of rank of weighted degree
Rockland
APOE-4
Autism
Hagmann
50
100
150
200
250
average of rank of weighted degree
randomized Rockland
randomized APOE-4
randomized Autism
randomized Hagmann
250
200
average of rank of weighted degree
100
150
randomized Rockland
randomized APOE-4
randomized Autism
randomized Hagmann
100
200
average of rank of weighted degree
150
250
Figure 1: Comparing controllability measures between empirical data and randomized data.
Scatter plot of the average of the ranks for weighted degrees versus average of the ranks of the
average controllability for (a) the empirical data and (b) their randomized counterpart.
Scatter plot of the average of ranks for weighted degrees versus the average of ranks of the
modal controllability for (c) empirical data and (d) their randomized counterpart.
APOE-4 data
Low
31/0.28/86.16
29/0.26/80.60
30/0.27/83.34
29/0.26/80.60
30/0.27/83.38
High
30/0.27/83.37
31/0.28/86.15
31/0.28/86.15
30/0.27/83.37
30/0.27/83.37
BA network
Low
28.04/0.25/77.95
28/0.25/77.84
28.04/0.25/77.95
28/0.25/77.84
27.04/0.25/77.94
High
28.04/0.25/77.91
28.16/0.26/78.25
27.8/0.25/77.25
27.88/0.25/77.47
28.04/0.25/77.93
Uniform network
Low
28.16/0.26/78.27
27.68/0.25/76.94
28/0.25/77.83
28.16/0.26/78.27
27.46/0.25/76.33
High
27.8/0.25/77.27
27.96/0.25/77.71
27.84/0.25/77.38
27.8/0.25/77.27
27.66/0.25/76.88
Centrality measure
Degree centrality
Betweenness centrality
Eigenvector centrality
Pagerank centrality
Random sequence
5
Table 1: The minimum number / fraction (with respect the size of the network) of nodes and its energy
Trace{WK} to guarantee the controllability of the APOE-4 data and the corresponding BA and uniform random
networks (of same size, connectivity and edge weight distribution) following the procedure described in the
main text. For a given sequence, "High" is the rank from the largest to the smallest, while "Low" is the rank
from the smallest to the largest. Hagmann, Autism and Rockland dataset have been also analyzed and we
found qualitatively the same results, i.e. in order to control the network we need a fraction of nodes ranging
from 16% to 25% of the whole brain, regardless of the centrality measures used to rank the nodes.
Methods
Neural dynamics model. Similarly to Gu et. al., we employed a simplified linear discrete-time neural
dynamics model. This model is a noise-free variation of the Galán model2 and can be derived from the
linearization of a general Wilson-Cowan system (for strengths and weakness of this model see Refs.
𝑥𝑡+∆𝑡 =𝐴𝑥𝑡 +𝐵!𝑢!𝑡
(2),(3),(12)). We then discretized the linearized dynamics and obtained the following equation:
where 𝑥𝑡 is the state variable representing the neural activity of the brain regions, 𝐴= 1−𝛼∆𝑡𝐼+
𝑐𝑀∆𝑡 is the 𝑁×𝑁 Jacobian matrix, 𝛼 is the inverse of the relaxation time, ∆𝑡 is the discretized time step, 𝐼 is
the 𝑁×𝑁 identity matrix, 𝑀 is the symmetric and weighted DTI structural matrix and c is a normalization
constant. Following the original work of Galan2, we fixed 𝛼=1.0,∆𝑡=0.2. The input matrix 𝐵! identifies
the set of control points K in the brain, where 𝐾={𝑘!,···,𝑘!} and
𝐵!= 𝑒!!,···,𝑒!! ,
where 𝑒! denotes the i-th canonical vector of dimension 𝑁. The input 𝑢!:𝑅!!→𝑅! denotes the control
Controllability Gramian. Given the neural dynamics, Eq. (1), we can set up the Lyapunov equation, 𝑊!−
𝐴𝑊!𝐴!=𝐵!𝐵!! . If A is stable, then this equation has a unique solution called Gramian:
strategy.
(1)
(2)
6
𝑊!=
𝐴!𝐵!𝐵!!
!!!!
𝐴! !
(3)
adopted by Gu et. al. The results presented in Table 1 were employed using the averaged DTI matrix, i.e.
The results presented in the main text were obtained as follows. In Figure 1, we applied the controllability
total number of healthy subjects. We calculate the average and the modal controllability for each of the
For the discretized dynamics (1) and 𝐴 symmetric, stability is ensured if all eigenvalues of 𝐴 lies in the interval
𝜆!∈(−1,1) con i=1,…,N. Therefore, in order to guarantee stability of our dynamical system we set the
normalization constant to 𝑐=1/(1+𝜆!"# ), where 𝜆!"#=𝑀𝑎𝑥{𝜆!,𝑖=1,…,𝑁}. Following this procedure
we constrain the eigenvalue spectrum of 𝐴 to be in the desired interval.
framework in each of the individual DTI matrices within a given dataset, 𝑀!, where 𝑖 ranges from 1 to 𝑛, the
corresponding Jacobian matrix 𝐴! and then we performed the average of the rank plots following the procedure
𝑀!"= 𝑀! 𝑛.
Trace{WK}. As shown by Gu and collaborators1, the strong correlation between node degree and average
evolutionary mode of a dynamical network. It is given by 𝜙!=
where 𝑉=𝜈!"! is
!!!!
the eigenvector matrix and 𝜆1,···,𝜆𝑁 are the corresponding eigenvalues of 𝐴.
and edge weight distribution of the averaged DTI matrix, 𝑀!"= 𝑀! 𝑛 for each dataset. We implemented
these constraints by fitting the edge weights of 𝑀!" to a Pareto distribution10. The synthetic networks were
Barabasi-Albert scale-free networks10 and the uniform networks10, using the same number of nodes, edges
follows the expressions reported by Gu et. al.1. The average controllability can be approximate by
Controllability measures. The average and the modal controllability measures employed in this study
Synthetic random networks. We generated two type of synthetic DTI random networks, namely, the
controllability can be understood analytically and it does not depend on the specific topological properties of
the structural connectivity M. The modal controllability represents the ability of a node to control each
1−𝜆!𝐴 !𝜈!"!
generated using standard routines available in the Mathematica software. Controllability measures displayed
in Table 1 were averaged over 50 realizations of the synthetic random networks.
7
Betweenness centrality is a measure of centrality based on shortest paths and is given by
Centralities measures. Degree centrality refers to the number of links a node has to other nodes10.
𝑛!" 𝑣 𝑛!"
where 𝑛!" is the total number of shortest paths from node 𝑠 to node 𝑡 and 𝑛!"𝑣 is the number of those paths
that pass through the node 𝑣10. Eigenvector centrality is based on the centralities of its neighbor nodes and can
be expressed as 𝑐=!!𝐴!𝑐, where 𝐴 is the adjacency matrix and 𝜆 is the corresponding largest eigenvalue10.
!!!!!
PageRank centrality is a generalization of the eingevector centrality, weighting in a non-linear way the number
and quality of links connected to the node10.
DTI brain connectivity datasets. We applied the controllability framework on four empirical datasets of large
scale structural (DTI or DSI) brain connectivity from different studies 5–8. All datasets are open access and were
obtained from the Human Connectome Project4, namely, the APOE-4 dataset5 (𝑛=30 APOE-4 non-carrier
and n=25 APOE-4 carrier individuals; gray matter parcellation into 𝑁=110 large scale regions); The
Rockland dataset6 (𝑛=195 healthy subjects, 𝑁=188 large scale regions); The Hagmann dataset8 (average
matrix corresponding to 𝑛=5 healthy subjects, 𝑁=66 cortical regions) and the Autism dataset7 (𝑛=94
healthy subjects, 𝑁=264 large scale regions). For more specific details on the data acquisition and
Randomized data set. In Figure 1 we randomized the topological structure of the DTI matrix 𝑀 by randomly
random with zero diagonal entries. In building random RSNs, we randomized the DTI matrix 𝑀 by randomly
rewiring all its matrix elements, but keeping symmetry. The randomized DTI matrices are thus symmetric
preprocessing we refer to the original studies.
rewiring all its matrix elements, but fixed the degree sequence of the original dataset12.
References.
1. Gu, S., et al. Controllability of structural brain networks. Nature communications 6, 8414,
doi:10.1038/ncomms9414 (2015).
2. Gala´n, R. F. On how network architecture determines the dominant patterns of spontaneous neural
activity. PloS one 3, e2148 (2008).
8
3. Honey, C. J., et al. Predicting human resting-state functional connectivity from structural connectivity. Proc.
Natl Acad. Sci. USA 106, 20352040 (2009).
4. USC Multimodal Connectivity Database: http://umcd.humanconnectomeproject.org/umcd/default/browse
5. Brown, J. A., et al. Brain network local interconnectivity loss in aging APOE-4 allele carriers. Proceedings
of the National Academy of Sciences 108, 20760-20765 (2011).
6. Brown, J. A., Rudie, J. D., Bandrowski, A., Van Horn, J. D. and Bookheimer, S. Y. The UCLA multimodal
connectivity database: a web-based platform for brain connectivity matrix sharing and analysis. Front
Neuroinform 6, 28, doi:10.3389/fninf.2012.00028 (2012).
7. Brown, Jesse A., et al. The UCLA multimodal connectivity database: a web-based platform for brain
connectivity matrix sharing and analysis. Frontiers in neuroinformatics 6: 28 (2012).
8. Hagmann, P., et. al. Mapping the Structural Core of Human Cerebral Cortex PLoS Biol 6(7):e159 (2008).
9. Pasqualetti, F., Zampieri, S., Bullo, F. Controllability metrics, limitations and algorithms for complex
networks. IEEE Transactions on Control of Network Systems, 1(1), 40-52 (2014)
10. Newman, M. E. J. Networks: An Introduction. (Oxford University Press Inc., New York 2010).
11. Sizemore, A., Giusti, C., Betzel, R. F., & Bassett, D. S. Closures and cavities in the human
connectome. arXiv:1608.03520 (2016).
12. Molloy, Michael, and Bruce Reed. "A critical point for random graphs with a given degree
sequence." Random structures & algorithms 6.2‐3 (1995): 161-180.
13. Muldoon, Sarah Feldt, et al. Stimulation-based control of dynamic brain networks. PLoS Comput Biol
12.9: e1005076 (2016).
14. Gu, Shi, et al. Optimal Trajectories of Brain State Transitions. NeuroImage (2017).
15. Betzel, Richard F., et al. "Optimally controlling the human connectome: the role of network topology."
Scientific Reports 6 (2016).
Authors Contributions: SS., S.Z, M.Z., M.C. designed the research, C.T., R.R., S.S. performed the research,
R.R., C.T. analyzed the data. All authors contributed in writing the manuscript.
9
|
1912.06207 | 1 | 1912 | 2019-12-12T20:54:08 | From deep learning to mechanistic understanding in neuroscience: the structure of retinal prediction | [
"q-bio.NC",
"cs.LG",
"physics.bio-ph"
] | Recently, deep feedforward neural networks have achieved considerable success in modeling biological sensory processing, in terms of reproducing the input-output map of sensory neurons. However, such models raise profound questions about the very nature of explanation in neuroscience. Are we simply replacing one complex system (a biological circuit) with another (a deep network), without understanding either? Moreover, beyond neural representations, are the deep network's computational mechanisms for generating neural responses the same as those in the brain? Without a systematic approach to extracting and understanding computational mechanisms from deep neural network models, it can be difficult both to assess the degree of utility of deep learning approaches in neuroscience, and to extract experimentally testable hypotheses from deep networks. We develop such a systematic approach by combining dimensionality reduction and modern attribution methods for determining the relative importance of interneurons for specific visual computations. We apply this approach to deep network models of the retina, revealing a conceptual understanding of how the retina acts as a predictive feature extractor that signals deviations from expectations for diverse spatiotemporal stimuli. For each stimulus, our extracted computational mechanisms are consistent with prior scientific literature, and in one case yields a new mechanistic hypothesis. Thus overall, this work not only yields insights into the computational mechanisms underlying the striking predictive capabilities of the retina, but also places the framework of deep networks as neuroscientific models on firmer theoretical foundations, by providing a new roadmap to go beyond comparing neural representations to extracting and understand computational mechanisms. | q-bio.NC | q-bio |
From deep learning to mechanistic understanding in
neuroscience: the structure of retinal prediction
Hidenori Tanaka1,2, Aran Nayebi3, Niru Maheswaranathan3,5, Lane McIntosh3, Stephen A.
Baccus4, and Surya Ganguli2,5
1Physics & Informatics Laboratories, NTT Research, Inc., East Palo Alto, CA, USA
2Department of Applied Physics, Stanford University, Stanford, CA, USA
3Neurosciences PhD Program, Stanford University, Stanford, CA, USA
4Department of Neurobiology, Stanford University, Stanford, CA, USA
5Google Brain, Google, Inc., Mountain View, CA, USA
Abstract
Recently, deep feedforward neural networks have achieved considerable success in
modeling biological sensory processing, in terms of reproducing the input-output
map of sensory neurons. However, such models raise profound questions about the
very nature of explanation in neuroscience. Are we simply replacing one complex
system (a biological circuit) with another (a deep network), without understanding
either? Moreover, beyond neural representations, are the deep network's computa-
tional mechanisms for generating neural responses the same as those in the brain?
Without a systematic approach to extracting and understanding computational
mechanisms from deep neural network models, it can be difficult both to assess the
degree of utility of deep learning approaches in neuroscience, and to extract experi-
mentally testable hypotheses from deep networks. We develop such a systematic
approach by combining dimensionality reduction and modern attribution methods
for determining the relative importance of interneurons for specific visual computa-
tions. We apply this approach to deep network models of the retina, revealing a
conceptual understanding of how the retina acts as a predictive feature extractor
that signals deviations from expectations for diverse spatiotemporal stimuli. For
each stimulus, our extracted computational mechanisms are consistent with prior
scientific literature, and in one case yields a new mechanistic hypothesis. Thus
overall, this work not only yields insights into the computational mechanisms
underlying the striking predictive capabilities of the retina, but also places the
framework of deep networks as neuroscientific models on firmer theoretical founda-
tions, by providing a new roadmap to go beyond comparing neural representations
to extracting and understand computational mechanisms.
1
Introduction
Deep convolutional neural networks (CNNs) have emerged as state of the art models of a variety
of visual brain regions in sensory neuroscience, including the retina [1, 2], primary visual cortex
(V1), [3, 4, 5, 6], area V4 [3], and inferotemporal cortex (IT) [3, 4]. Their success has so far been
primarily evaluated by their ability to explain reasonably large fractions of variance in biological
neural responses across diverse visual stimuli. However, fraction of variance explained is not of
course the same thing as scientific explanation, as we may simply be replacing one inscrutable black
box (the brain), with another (a potentially overparameterized deep network).
Indeed, any successful scientific model of a biological circuit should succeed along three fundamental
axes, each of which goes above and beyond the simple metric of mimicking the circuit's input-
33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.
Figure 1: Deep learning models
of the retina trained only on nat-
ural scenes reproduce an array
of retinal phenomena with artifi-
cial stimuli (reproduced from ref.
[2]).
(A) Training procedure: We ana-
lyzed a three-layer convolutional
neural network (CNN) model of the
retina which takes as input a spa-
tiotemporal natural scene movie and
outputs a nonnegative firing rate,
corresponding to a retinal ganglion
cell response. The first layer con-
sists of eight spatiotemporal convo-
lutional filters (i.e., cell types) with
the size of (15×15×40), the second
layer of eight convolutional filters
(8×11×11), and the fully connected
layer predicting the ganglion cells'
response. As previously reported
in [2], the deep learning model re-
produces (B) an omitted stimulus re-
sponse, (C) latency coding, (D) the
motion reversal response, and (E)
motion anticipation.
output map. First, the intermediate computational mechanisms used by the hidden layers to generate
responses should ideally match the intermediate computations in the brain. Second, we should be able
to extract conceptual insight into how the neural circuit generates nontrivial responses to interesting
stimuli (for example responses to stimuli that cannot be generated by a linear receptive field). And
third, such insights should suggest new experimentally testable hypotheses that can drive the next
generation of neuroscience experiments.
However, it has been traditionally difficult to systematically extract computational mechanisms, and
consequently conceptual insights, from deep CNN models due to their considerable complexity
[7, 8]. Here we provide a method to do so based on the idea of model reduction, whose goal is to
systematically extract a simple, reduced, minimal subnetwork that is most important in generating
a complex CNN's response to any given stimulus. Such subnetworks then both summarize compu-
tational mechanisms and yield conceptual insights. We build on ideas from interpretable machine
learning, notably methods of input attribution that can decompose a neural response into a sum of
contributions either from individual pixels [9] or hidden neurons [10]. To achieve considerable model
reduction for responses to spatiotemporal stimuli, we augment and combine such input attribution
methods with dimensionality reduction, which, for carefully designed artificial stimuli employed in
neurophysiology experiments, often involves simple spatiotemporal averages over stimulus space.
We demonstrate the power of our systematic model reduction procedure to attain mechanistic insights
into deep CNNs by applying them to state of the art deep CNN models of the retina [1, 2]. The retina
constitutes an ideal first application of our methods because the considerable knowledge (see e.g. [11])
about retinal mechanisms for transducing spatiotemporal light patterns into neural responses enables
us to assess whether deep CNNs successfully learn the same computational structure. In particular,
we obtain deep CNN models from [2] which were trained specifically to mimic the input-output
transformation from natural movies to retinal ganglion cell outputs measured in the salamander retina.
The model architecture involved a three-layer CNN model of the retina with ReLU nonlinearities
(Fig. 1A). This network was previously shown [1, 2] to: (i) yield state of the art models of the retina's
response to natural scenes that are almost as accurate as possible given intrinsic retinal stochasticity;
(ii) possess internal subunits with similar response properties to those of retinal interneurons, such as
bipolar and amacrine cell types; (iii) generalize from natural movies, to a wide range of eight different
classes of artificially structured stimuli used over decades of neurophysiology experiments to probe
2
ΔtxyΔtxMotion reversalt [s]Schwartz et al. (2007)Rate [Hz]xLatency codingGollisch & Meister (2008)Figure 1Training input Natural scene movieTraining output Ganglion cells' responseExperimental dataOmitted stimulus responseModel outputSchwartz et al. (2007)sRate [Hz]t [s]Training procedure: natural scenesTesting procedure: structured stimulusMotion anticipationBerry et al. (1999)Position(A)(B)(C)(D)(E)N. Maheswaranathan et al. (2018) retinal response properties. This latter generalization capacity from natural movies to artificially
structured stimuli (that were never present in the training data) is intriguing given the vastly different
spatiotemporal statistics of the artificial stimuli versus natural stimuli, suggesting the artificial stimuli
were indeed well chosen to engage the same retinal mechanisms engaged under natural vision [2].
Here, we focus on understanding the computational mechanisms underlying the deep CNN's ability
to reproduce the neural responses to four classes of artificial stimuli (Fig. 1B-E), each of which,
through painstaking experiments and theory, have revealed striking nonlinear retinal computations
that advanced our scientific understanding of the retina. The first is the omitted stimulus response
(OSR) [12, 13] (Fig. 1B), in which a periodic sequence of full field flashes entrains a retinal ganglion
cell to respond periodically, but when a single flash is omitted, the ganglion cell produces an even
larger response at the expected time of the response to the omitted flash. Moreover, the timing of this
omitted stimulus response occurs at the expected time over a range of frequencies of the periodic
flash train, suggesting the retina is somehow retaining a memory trace of the period of the train of
flashes. The second is latency encoding [14], in which stronger stimuli yield earlier responses (Fig.
1C). The third is motion reversal [15], in which a bar suddenly reversing its motion near a ganglion
cell receptive field generates a much larger response after the motion reversal (Fig. 1D). The fourth
is motion anticipation [16], where the neural population responding to a moving bar is advanced in
the direction of motion to compensate for propagation delays through the retina (Fig. 1E). These
responses are striking because they imply the retina has implicitly built into it a predictive world
model codifying simple principles like temporal periodicity, and the velocity based extrapolation of
future position. The retina can then use these predictions to improve visual processing (e.g. in motion
anticipation), or when these predictions are violated, the retina can generate a large response to signal
that deviation (e.g. in the OSR and motion reversal).
While experimentally motivated prior theoretical models have been employed to explain the OSR
[17, 18], latency encoding [14, 19], motion reversal [20, 21], and motion anticipation [16], to date,
no single model other than the deep CNN found in [2] has been able to simultaneously account for
retinal ganglion cell responses to both natural scenes and all four of these classes of stimuli, as well
as several other classes of artificial stimuli. However, it is difficult to explain the computational
mechanisms underlying the deep CNN's ability to generate these responses simply by examining the
complex network in Fig. 1A. For example, why does the deep CNN fire more when a stimulus is
omitted, or when a bar reverses? How can it anticipate motion to compensate for propagation delays?
And why do stronger responses cause earlier firing?
These are foundational scientific questions about the retina whose answers require conceptual insights
that are not afforded by the existence of a complex but highly predictive CNN alone. And even more
importantly, if we could extract conceptual insights into the computational mechanisms underlying
CNN responses, would these mechanisms match those used in the biological retina? Or is the
deep CNN only accurate at the level of modelling the input-output map of the retina, while being
fundamentally inaccurate at the level of underlying mechanisms? Adjudicating between these two
possibilities is essential for validating whether the deep learning approach to modelling in sensory
neuroscience can indeed succeed in elucidating biological neural mechanisms, which has traditionally
been the gold-standard of circuit based understanding in systems neuroscience [11, 22, 23, 24].
In the following we will show how a combination of dimensionality reduction and hidden neuron or
stimulus attribution can yield simplified subnetwork models of the deep CNNs response to stimuli,
finding models that are consistent with prior mechanistic models with experimental support in the
case of latency encoding, motion reversal, and motion anticipation. In addition, our analysis yields a
new model that cures the inadequacies of previous models of the OSR. Thus our overall approach
provides a new roadmap to extract mechanistic insights into deep CNN function, confirms in the
case of the retina that deep CNNs do indeed learn computational mechanisms that are similar to
those used in biological circuits, and yields a new experimentally testable hypothesis about retinal
computation. Moreover, our results in the retina yield hope (to be tested in future combined theory
and experiments) that more complex deep CNN models of higher visual cortical regions, may not only
yield accurate black box models of input-output transformations, but may also yield veridical and
testable hypotheses about intermediate computational mechanisms underlying these transformations,
thereby potentially placing deep CNN models of sensory brain regions on firmer epistemological
foundations.
3
2 From deep CNNs to neural mechanisms through model reduction
To extract understandable reduced models from the millions of parameters comprising the deep CNN
in Fig. 1A and [2], we first reduce dimensionality by exploiting spatial invariance present in the
artificial stimuli carefully designed to specifically probe retinal physiology (Fig.1B-E), and then carve
out important sub-circuits using modern attribution methods [9, 10]. We proceed in 3 steps:
Step (1): Quantify the importance of a model unit with integrated gradients. The nonlinear
input-output map of our deep CNN can be expressed as r(t) = F[s(t)], where r(t) ∈ R+ denotes
the nonnegative firing rate of a ganglion cell at time bin t and s(t) ∈ R50×50×40 denotes the recent
spatiotemporal history of the visual stimulus spanning two dimensions of space (x, y) (with 50
spatial bins in each dimension) as well as 40 preceding time bins parameterized by ∆t. Thus
a single component of the vector s(t) is given by sxy∆t(t), which denotes the stimulus contrast
at position (x, y) at time t − ∆t. We assume a zero contrast stimulus yields no response (i.e.
F[0] = 0). We can decompose, or attribute the response r(t) to each preceding spacetime point by
considering a straight path in spatiotemporal stimulus space from the zero stimulus to s(t) given
by s(t; α) = αs(t) where the path parameter α ranges from 0 to 1 [9]. Using the line integral
F[s(t; 1)] =(cid:82) 1
∂α , we obtain
∂s(cid:12)(cid:12)s(t,α) · ∂s(t,α)
0 dα ∂F
50(cid:88)x=1
40(cid:88)∆t=1
50(cid:88)y=1
r(t) = F[s(t)] =
sxy∆t(t)(cid:90) 1
0
dα
∂F
∂sxy∆t(t)(cid:12)(cid:12)(cid:12)(cid:12)αs(t) ≡
50(cid:88)x=1
50(cid:88)y=1
40(cid:88)∆t=1
Axy∆t.
(1)
∂s(t)(cid:12)(cid:12)s=0 · s(t) is only approximate. The coefficient vector ∂F
This equation represents an exact decomposition of the response r(t) into attributions Axy∆t from
each preceding spacetime stimulus pixel (x, y, ∆t). Intuitively, the magnitude of Axy∆t tells us
how important each pixel is in generating the response, and the sign tells us whether or not turning
on each pixel from 0 to sxy∆t(t) yields a net positive or negative contribution to r(t). When F
is linear, this decomposition reduces to a Taylor expansion of F about s(t) = 0. However, in
the nonlinear case, this decomposition has the advantage that it is exact, while the linear Taylor
expansion r(t) ≈ ∂F
of this Taylor
expansion is often thought of as the linear spacetime receptive field (RF) of the model ganglion cell,
a concept that dominates sensory neuroscience. Thus choosing to employ this attribution method
enables us to go beyond the dominant but imperfect notion of an RF, in order to understand nonlinear
neural responses to arbitrary spatiotemporal stimuli. In supplementary material, we discuss how this
theoretical framework of attribution to input space can be temporally extended to answer different
questions about how deep networks process spatiotemporal inputs.
However, since our main focus here is model reduction, we consider instead attributing the ganglion
cell response back to the first layer of hidden units, to quantify their importance. We denote by
cxy(t) = W [1]
z[1]
cxy (cid:126) s(t) + bcxy the pre-nonlinearity activation of the layer 1 hidden units, where
Wcxy and bcxy are the convolutional filters and biases of a unit in channel c (c = 1, . . . , 8) at
convolutional position (x, y) (with x, y = 1, . . . , 36). Now computing the line integral F[s(t; 1)] =
(cid:82) 1
0 dα ∂F
∂α over the same stimulus path employed in (1) yields
∂s(t)(cid:12)(cid:12)s=0
∂z[1](cid:12)(cid:12)s(t,α) · ∂z[1]
r(t) = (cid:88)x,y,c(cid:34)(cid:90) 1
dα
0
∂F
∂z[1]
cxy(cid:12)(cid:12)(cid:12)(cid:12)s(t,α)(cid:35) (W [1]
cxy (cid:126) s) = (cid:88)x,y,c
[Gcxy(s)] (W [1]
cxy (cid:126) s) = (cid:88)x,y,c
Acxy.
(2)
This represents an exact decomposition of the response r(t) into attributions Acxy from each subunit
at the same time t (since all CNN filters beyond the first layer are purely spatial). This attribution
further splits into a product of W [1]
cxy (cid:126) s, reflecting the activity of that subunit originating from
spatiotemporal filtering of the preceding stimulus history, and an effective stimulus dependent weight
Gcxy(s) from each subunit to the ganglion cell, reflecting how variations in subunit activity z[1]
cxy
as the stimulus is turned on from 0 to s(t) yield a net impact on the response r(t). A positive
(negative) effective weight indicates that increasing subunit activity along the stimulus path yields a
net excitatory (inhibitory) effect on r(t).
Step (2): Exploiting stimulus invariances to reduce dimensionality. The attribution of the re-
sponse r(t) to first layer subunits in (2) still involves 8 × 36 × 36 = 10, 368 attributions. We can,
however, leverage the spatial uniformity of artificial stimuli used in neurophysiology experiments to
4
reduce this dimensionality. For example, in the OSR and latency coding, stimuli are spatially uniform,
implying W [1]
cxy (cid:126) s is independent of spatial indices (x, y). Thus, we can reduce the
number of attributions to the number of channels via
c (cid:126) s ≡ W [1]
r(t) =
8(cid:88)c=1(cid:18) 36(cid:88)x=1
36(cid:88)y=1
Gcxy(s)(cid:19) · (W [1]
c (cid:126) s) ≡
8(cid:88)c=1
Gc(s) · (W [1]
c (cid:126) s) ≡
8(cid:88)c=1
Ac.
(3)
For the moving bar in both motion reversal and motion anticipation, W [1]
independent of the y index and we can reduce the dimensionality from 10,368 down to 288 by
cxy (cid:126) s is
r(t) =
8(cid:88)c=1
36(cid:88)x=1(cid:18) 36(cid:88)y=1
Gcxy(s)(cid:19) · (W [1]
cx (cid:126) s) ≡
8(cid:88)c=1
36(cid:88)x=1
cx (cid:126) s ≡ W [1]
36(cid:88)x=1
8(cid:88)c=1
Gcx(s) · (W [1]
cx (cid:126) s) ≡
(4)
Acx.
More generally for other stimuli with no obvious spatial invariances, one could still attempt to reduce
dimensionality by performing PCA or other dimensionality reduction methods on the space of hidden
unit pre-activations or attributions over time. We leave this intriguing direction for future work.
Step (3): Building reduced models from important subunits. Finally, we can construct minimal
circuit models by first identifying "important" units defined as those with large magnitude attributions
A. We then construct our reduced model as a one hidden layer neural network composed of only
the important hidden units, with effective connectivity from each hidden unit to the ganglion cell
determined by the effective weights G in (2), (3), or (4).
3 Results: the computational structure of retinal prediction
We now apply the systematic model reduction steps described in the previous section to each of
the retinal stimuli in Fig. 1B-E. We show that in each case the reduced model yields scientific
hypotheses to explain the response, often consistent with prior experimental and theoretical work,
thereby validating deep CNNs as a method for veridical scientific hypothesis generation in this setting.
Moreover, our approach yields integrative conceptual insights into how these diverse computations
can all be simultaneously produced by the same set of hidden units.
3.1 Omitted stimulus response
As shown in Fig. 1B, periodic stimulus flashes trigger delayed periodic retinal responses. However,
when this periodicity is violated by omitting a flash, the ganglion cell signals the violation with a
large burst of firing [25, 26]. This OSR phenomenon is observed across several species including
salamander [12, 13]. Interestingly, for periodic flashes in the range of 6-12Hz, the latency between
the last flash before the omitted one, and the burst peak in the response, is proportional to the period
of the train of flashes [12, 13], indicating the retina retains a short memory trace of this period.
Moreover, pharmacological experiments suggest ON bipolar cells are required to produce the OSR
[13, 17], which have been shown to correspond to the first layer hidden units in the deep CNN [1, 2].
These phenomena raise two fundamental questions: what computational mechanism causes the large
amplitude burst, and how is the timing of the peak sensitive to the period of the flashes? There are
two theoretical models in the literature that aim to answer these questions. One proposes that the
bipolar cell activity responds to each individual flash with an oscillatory response whose period
adapts to the period of the flash train [18]. However, recent direct recordings of bipolar cells suggest
that such period adaptation is not present [27]. The other model claims that having dual pathways of
ON and OFF bipolar cells are enough to reproduce most of the aspects of the phenomena observed in
experiments [17]. However, the model only reproduces the shift of the onset of the burst, and not a
shift in the peak of the burst, which has the critical predictive latency [18].
Direct model reduction (Fig. 2) of the deep CNN in Fig. 1A using the methods of section 2 yields a
more sophisticated model than any prior model, comprised of three important pathways that combine
one OFF temporal filter with two ON temporal filters. Unlike prior models, the reduced model
exhibits a shift in the peak of the OSR burst as a function of the frequency of input flashes.
5
Figure 2: Omitted stimulus response. (A-1,2) Schematics of the model reduction procedure by
only leaving three (1 OFF, 2 ON) highly contributing units. (B) Attribution for each of the cell types
Ac over time. (C) Effective stimulus dependent weight for each of the cell types Gc over time. (D)
The combination of the two pathways of filter 2 and 6 reproduces the period dependent latency. (E)
Two ON bipolar cells are necessary to capture the predictive latency. Cell 2 with earlier peak is
only active in a high-frequency regime, while the cell 6 with later peak is active independent of the
frequency.
Fig. 2A presents a schematics of the model reduction steps described in (3). We first attribute a
ganglion cell's response to 8 individual channels and then average across both spatial dimensions
(Fig. 2A-1) as in steps (1) and (2). Then we build a reduced model from the identified important
subunits that capture essential features of the omitted stimulus response phenomenon (Fig. 2A-2).
In Fig. 2B, we present the time dependence of the attribution Ac(s(t)) in (3) for the eight channels,
or cell-types. Red (blue) traces reflect positive (negative) attributions. Channel temporal filters are
to the left of each attribution row and the total output response r(t) is on the top row. The stimulus
consists of three flashes, yielding three small responses, and then one large response after the end of
the three flashes (grey line). Quantitatively, we identify that cell-type 3 dominantly explains the small
responses to preceding flashes, while cell types 2 and 6 are necessary to explain the large burst after
the flash train ends. The final set of units included in the reduced model should be the minimal set
required to capture the defining features of the phenomena of interest. In the case of omitted stimulus
response, the defining feature is the existence of the large amplitude burst whose peak location is
sensitive to the period of the applied flashes. Once we identify the set of essential temporal filters,
we then proceed to determine the sign and magnitude of contribution (excitatory or inhibitory) of
the cell types. In Fig. 2C, we present the time-dependent effective weights from Gc(s(t)) in (3) for
the eight cell types, or channels. Red (blue) reflects positive (negative) weights. Given the product
of the temporal filters and the weights, cell-types 2 and 6 are effectively ON cells, which cause
positive ganglion cell responses to contrast increments, while cell-type 3 is an OFF cell, which is a
cell type that causes positive responses to contrast decrements. Following the prescribed procedures,
carving out the 3 important cell-types and effective weights yields a novel, mechanistic three pathway
model of the OSR, with 1 OFF and 2 ON pathways. Unlike prior models, the reduced model exhibits
a shift in the peak of the OSR burst as a function of the frequency of input flashes (with dark to
light blue indicating high to low frequency variation in the flash train) as in Fig. 2D. Furthermore,
the reduced model is consistent across the frequency range that produces the phenomena. Finally,
model reduction yields conceptual insights into how cell-types 2 and 6 enable the timing of the burst
peak to remember the period of the flash train (Fig. 2E). The top row depicts the decomposition
of the overall burst response r(t) (grey) into time dependent attributions A2 (red) and A6 (blue),
obeying the relation r(t) ≈ A2 + A6. Cell-type 2, which has an earlier peak in its temporal filter,
preferentially causes ganglion cell responses in high-frequency flash trains (left) compared to low
frequency trains (right), while cell-type 6 is equally important in both. The middle row shows the
temporal filter Wc=2(∆t), which has an earlier peak with a long tail, enabling it to integrate across
not only the last flash, but also preceding flashes (yellow bars). Time increases into the past from
left to right. Thus, the activation of this cell type 2 decreases as the flash train frequency decreases,
explaining the decrease in attribution in the top row. The bottom row shows that the temporal filter
Wc=6(∆t) of cell type 6, in contrast, has a later peak with a rapidly decaying tail. Thus the temporal
convolution Wc=6(∆t) (cid:126) s(∆t) of this filter with the flash train is sensitive only to the last flash, and
6
(A-2)Ac=6(t)timeintensityFigure 2: Omitted Stimulus Response(A-1)(B)(C)~(D)s(t)(E)12345678Ac=Gc·(W[1]c~s)r(t)=8Xc=1AcGcGc(s)W[1]c32612345678high freqlow freq26Ac=2(t)r(t)⇡+t t tW[1]c=2~sW[1]c=6~shigh
freqlow
freqPositive Negative is therefore independent of flash train frequency. The late peak and rapid tail explain why it supports
the response at late times independent of frequency in the top row.
Thus, our systematic model reduction approach yields a new model of the OSR that cures important
inadequacies of prior models. Moreover, it yields a new, experimentally testable scientific hypothesis
that the OSR is an emergent property of three bipolar cell pathways with specific and diverse temporal
filtering properties.
3.2 Latency coding
Rapid changes in contrast (which often occur
for example right after saccades) elicit a burst
of firing in retinal ganglion cells with a latency
that is shorter for larger contrast changes [14]
(Fig. 1C). Moreover, pharmacological studies
demonstrate that both ON and OFF bipolar cells
(corresponding to first layer hidden neurons in
the deep CNN [1, 2]) are necessary to produce
this phenomenon [19].
Model reduction via (3) in section 2 reveals that
a single pair of slow ON and fast OFF pathways
can explain the shift in the latency Fig. 3. First,
under a contrast decrement, there is a strong, fast
excitatory contribution from the OFF pathway.
Second, as the magnitude of the contrast decre-
ment increases, delayed inhibition from the slow
ON pathway becomes stronger. This negative
delayed contribution truncates excitation from
the OFF pathway at late times, thereby causing a
shift in the location of the total peak response to
earlier times (Fig. 3). The dual pathway mech-
anism formed by slow ON and fast OFF bipolar
cells is consistent with all existing experimental
facts. Moreover, it has been previously proposed
as a theory of latency coding [14, 19]. Thus this
example illustrates the power of a general nat-
ural scene based deep CNN training approach,
followed by model reduction, to automatically
generate veridical scientific hypotheses that were previously discovered only through specialized
experiments and analyses requiring significant effort [14, 19].
Figure 3: Latency coding (A) The decomposition
of the overall response r(t) (grey) into dominant
attributions A3(t) (blue) from an OFF pathway,
and A2(t) (red) from an ON pathway, obeying the
relation r(t) ≈ A3 + A2. Under a contrast decre-
ment, the OFF pathway activated first, followed by
delayed inhibitory input from the ON pathway. (B)
As the amount of contrast decrement increases (yel-
low bars), delayed inhibition from the ON pathway
(red) strengthens, which cuts off the total response
in r(t) at late times more strongly, thereby shifting
the location of the peak of r(t) to earlier times.
3.3 Motion reversal
As shown in Fig. 1D and [15], when a moving bar suddenly reverses its direction of motion, ganglion
cells near the reversal location exhibit a sharp burst of firing. While a ganglion cell classically
responds as the bar moves through its receptive field (RF) center from left to right before the motion
reversal, the sharp burst after the motion reversal does not necessarily coincide with the spatial
re-entry of the bar into the center of the RF as it moves back from right to left. Instead, the motion
reversal burst response occurs at a fixed temporal latency relative to the time of motion reversal,
for a variety of reversal locations within 110 µm of the RF center. These observations raise two
fundamental questions: why does the burst even occur and why does it occur at a fixed latency?
The classical linear-nonlinear model cannot reproduce the reversal response; it only correctly re-
produces the initial peak associated with the initial entry of a bar into the RF center [15]. Thus a
nonlinear mechanism is required. Model reduction of the deep CNN obtained via (4) reveals that
two input channels arrayed across 1D x space can explain this response through a specific nonlinear
mechanism (Fig. 4). Moreover, the second important channel revealed by model reduction yields a
cross cell-type inhibition that explains the fixed latency (Fig. 4D). Intriguingly, this reduced model is
7
Figure 3: Latency Coding23inputoutputLow intensity(A)(B)High intensityA2A3r(t)⇡+ Figure 4: Motion reversal of a moving bar.
(A) Schematics of (x, t) spatiotemporal model
reduction obtained via (4). By averaging over y we obtain 8 cell types at 36 different x positions
yielding 288 units. Attribution values reveal that only cell types 2 and 3 play a dominant role in
motion reversal. (B) The properties of cell-type 3 explains the existence of the burst. On the left are
time-series of pre-nonlinearity activations W [1]
c=3,x (cid:126) s of hidden units whose RF center is at spatial
position x. Time t = 0 indicates the time of motion reversal. The boxed region indicates the spatial
and temporal extent of the retinal burst in response to motion reversal. The offset of the box from
time t = 0 indicates the fixed latency. A fixed linear combination with constant coefficents of this
activation cannot explain the existence of the burst due to cancellations along the vertical x-axis in the
boxed region. However, due to downstream nonlinearities, the effective weight coefficients Gc=3,x
from subunits to ganglion cell responses rapidly flip in sign (middle), and generating a burst of motion
reversal response (right). (C) Schematics of the reduced model keeping only important subunits. (D)
Attribution contributions from the two dominant cell types A2 (in pink) and A3 (in blue), where
Ac =(cid:80)36
x=1 Acx. With only cell-type 3, the further the reversal location is from a ganglion cell's RF
center, the longer we would expect it to take to generate a reversal response. However, the inhibition
coming from cell type 2 increases the further away the reversal occurs, truncating the late response
and thus fixing the latency of the motion reversal response.
qualitatively consistent with a recently proposed and experimentally motivated model [20] that points
out the crucial role of dual pathways of ON and OFF bipolar cells.
3.4 Motion anticipation
As shown in Fig. 1E and [16] the retina already
starts to compensate for propagation delays by
advancing the retinal image of a moving bar
along the direction of motion, so that the retinal
image does not lag behind the instantaneous
location as one might naively expect.
Model reduction of our deep CNN reveals a
mechanism for this predictive tracking. First,
since ganglion cell RFs have some spatial extent,
a moving bar naturally triggers some ganglion
cells before entering their RF center, yielding
a leading edge of a retinal wave. What is then
required for motion anticipation is some addi-
tional motion direction sensitive inhibition that
cuts off the lagging edge of the wave so its peak
activity shifts towards the leading edge. Indeed,
model reduction reveals a computational mech-
anism in which one cell type feeds an excitatory
signal to a ganglion cell while the other provides
direction sensitive inhibition that truncates the lagging edge. This model is qualitatively consistent
with prior theoretical models that employ such direction selective inhibition to anticipate motion [16].
Figure 5: Motion anticipation of a moving bar.
Contributions from the two dominant cell types.
A2 in pink, A3 in blue, r(t) ≈ A2 + A3 in grey,
where Ac =(cid:80)36
x=1 Acx. Depending on the direc-
tion of motion of a bar, activity that lags behind
the leading edge gets asymmetrically truncated by
the inhibition from the cell type 2 (pink). (A) The
bar is moving to the right and the inhibition (pink)
is slightly stronger on the left side. (B) the bar
is moving to the left and the inhibition (pink) is
stronger on the right side.
8
Figure 4: Motion Reversal(A)(B)(D)(C)txW[1]cx~sGcxr(t)=8Xc=136Xx=1AcxW[1]c=3,x~sGc=3,xAc=3,x=Gc=3,x·(W[1]c=3,x~s)Ac=3,x=Gc=3,x·(W[1]c=3,x~s)xA2A3r(t)⇡+32Figure 5: Motion Anticipation(A)(B)A2A3r(t)⇡+ 4 Discussion
Figure 6: A unified framework
to reveal computational structure
in the brain. We outlined an auto-
mated procedure to go from large-
scale neural recordings to mechanis-
tic insights and scientific hypotheses
through deep learning and model re-
duction. We validate our approach
on the retina, demonstrating how
only three cell-types with different
ON/OFF and fast/slow spatiotempo-
ral filtering properties can nonlin-
early interact to simultaneously gen-
erate diverse retinal responses.
In summary, in the case of the retina, we have shown that complex CNN models obtained via machine
learning can not only mimic sensory responses to rich natural scene stimuli, but also can serve as a
powerful and automatic mechanism for generating valid scientific hypotheses about computational
mechanisms in the brain, when combined with our proposed model reduction methods (Fig. 6).
Applying this approach to the retina yields conceptual insights into how a single model consisting
of multiple nonlinear pathways with diverse spatiotemporal filtering properties can explain decades
of painstaking physiological studies of the retina. This suggests in some sense an inverse roadmap
for experimental design in sensory neuroscience. Rather than carefully designing special artificial
stimuli to probe specific sensory neural responses, and generating individual models tailored to each
stimulus, one could instead fit a complex neural network model to neural responses to a rich set of
ethologically relevant natural stimuli, and then apply model reduction methods to understand how
different parts of a single model can simultaneously account for responses to artificial stimuli across
many experiments. The interpretable mechanisms extracted from model reduction then constitute
specific hypotheses that can be tested in future experiments. Moreover, the complex model itself can
be used to design new stimuli, for example by searching for stimuli that yield divergent responses
in the complex model, versus a simpler model of the same sensory region. Such stimulus searches
could potentially elucidate functional reasons for the existence of model complexity.
In future studies, it will be interesting to conduct a systematic exploration of universality and
individuality [28] in the outcome of model reduction procedures applied to deep learning models
which recapitulate desired phenomena, but are obtained from different initializations, architectures,
and experimental recordings. An intriguing hypothesis is that the reduced models required to explain
specific neurobiological phenomena arise as universal computational invariants across the ensemble
of deep learning models parameterized by these various design choices, while many other aspects of
such deep learning models may individually vary across these choices, reflecting mere accidents of
history in initialization, architecture and training.
It would also be extremely interesting to stack this model reduction procedure to obtain multilayer
reduced models that extract computational mechanisms and conceptual insights into deeper CNN
models of higher cortical regions. The validation of such extracted computational mechanisms would
require further experimental probes of higher responses with carefully chosen stimuli, perhaps even
stimuli chosen to maximize responses in the deep CNN model itself [29, 30]. Overall the success of
this combined deep learning and model reduction approach to scientific inquiry in the retina, which
was itself not at all a priori obvious before this work, sets a foundation for future studies to explore
this combined approach deeper in the brain.
Acknowledgments
We thank Daniel Fisher for insightful discussions and support. We thank the Masason foundation
(HT), grants from the NEI (R01EY022933, R01EY025087, P30-EY026877) (SAB), and the Simons,
James S. McDonnell foundations, and NSF Career 1845166 (SG) for funding.
9
1. High-throughput neural recordings2. Train deep-learning model and perform neurophysiology experiments in silico3. Identify important sub-circuits
and derive an array of interpretable models236Figure 6 Final References
[1] Lane McIntosh, Niru Maheswaranathan, Aran Nayebi, Surya Ganguli, and Stephen Baccus. Deep learning
models of the retinal response to natural scenes. Advances in neural information processing systems, pages
1369 -- 1377, 2016.
[2] Niru Maheswaranathan, Lane T McIntosh, David B Kastner, Josh Melander, Luke Brezovec, Aran Nayebi,
Julia Wang, Surya Ganguli, and Stephen A Baccus. Deep learning models reveal internal structure and
diverse computations in the retina under natural scenes. bioRxiv, page 340943, 2018.
[3] Daniel L. K. Yamins, Ha Hong, Charles F. Cadieu, Ethan A. Solomon, Darren Seibert, and James J. DiCarlo.
Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings
of the National Academy of Sciences, 111(23):8619 -- 8624, 2014. doi: 10.1073/pnas.1403112111.
[4] Seyed-Mahdi Khaligh-Razavi and Nikolaus Kriegeskorte. Deep supervised, but not unsupervised, models
may explain it cortical representation. PLoS computational biology, 10(11):e1003915, 2014.
[5] Umut Güçlü and Marcel AJ van Gerven. Deep neural networks reveal a gradient in the complexity of
neural representations across the ventral stream. The Journal of Neuroscience, 35(27):10005 -- 10014, 2015.
[6] Santiago A Cadena, George H Denfield, Edgar Y Walker, Leon A Gatys, Andreas S Tolias, Matthias
Bethge, and Alexander S Ecker. Deep convolutional models improve predictions of macaque v1 responses
to natural images. bioRxiv, page 201764, 2017.
[7] David GT Barrett, Ari S Morcos, and Jakob H Macke. Analyzing biological and artificial neural networks:
challenges with opportunities for synergy? Current opinion in neurobiology, 55:55 -- 64, 2019.
[8] Joshua I Glaser, Ari S Benjamin, Roozbeh Farhoodi, and Konrad P Kording. The roles of supervised
machine learning in systems neuroscience. Progress in neurobiology, 2019.
[9] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Doina
Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning,
volume 70 of Proceedings of Machine Learning Research, pages 3319 -- 3328, International Convention
Centre, Sydney, Australia, 06 -- 11 Aug 2017. PMLR. URL http://proceedings.mlr.press/v70/
sundararajan17a.html.
[10] Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron. In International Con-
ference on Learning Representations, 2019. URL https://openreview.net/forum?id=SylKoo0cKm.
[11] Tim Gollisch and Markus Meister. Eye smarter than scientists believed: neural computations in circuits of
the retina. Neuron, 65(2):150 -- 164, 2010.
[12] Greg Schwartz, Rob Harris, David Shrom, and Michael J Berry II. Detection and prediction of periodic
patterns by the retina. Nature neuroscience, 10(5):552, 2007.
[13] Greg Schwartz and Michael J Berry 2nd. Sophisticated temporal pattern recognition in retinal ganglion
cells. Journal of neurophysiology, 99(4):1787 -- 1798, 2008.
[14] Tim Gollisch and Markus Meister. Rapid neural coding in the retina with relative spike latencies. science,
319(5866):1108 -- 1111, 2008.
[15] Greg Schwartz, Sam Taylor, Clark Fisher, Rob Harris, and Michael J Berry II. Synchronized firing among
retinal ganglion cells signals motion reversal. Neuron, 55(6):958 -- 969, 2007.
[16] Michael J Berry II, Iman H Brivanlou, Thomas A Jordan, and Markus Meister. Anticipation of moving
stimuli by the retina. Nature, 398(6725):334, 1999.
[17] Birgit Werner, Paul B Cook, and Christopher L Passaglia. Complex temporal response patterns with a
simple retinal circuit. Journal of neurophysiology, 100(2):1087 -- 1097, 2008.
[18] Juan Gao, Greg Schwartz, Michael J Berry, and Philip Holmes. An oscillatory circuit underlying the
detection of disruptions in temporally-periodic patterns. Network: Computation in Neural Systems, 20(2):
106 -- 135, 2009.
[19] Tim Gollisch and Markus Meister. Modeling convergent on and off pathways in the early visual system.
Biological cybernetics, 99(4-5):263 -- 278, 2008.
[20] Eric Y Chen, Janice Chou, Jeongsook Park, Greg Schwartz, and Michael J Berry. The neural circuit
mechanisms underlying the retinal response to motion reversal. Journal of Neuroscience, 34(47):15557 --
15575, 2014.
10
[21] Eric Y Chen, Olivier Marre, Clark Fisher, Greg Schwartz, Joshua Levy, Rava Azeredo da Silveira, and
Michael J Berry. Alert response to motion onset in the retina. Journal of Neuroscience, 33(1):120 -- 132,
2013.
[22] David I Vaney, Benjamin Sivyer, and W Rowland Taylor. Direction selectivity in the retina: symmetry and
asymmetry in structure and function. Nature Reviews Neuroscience, 13(3):194, 2012.
[23] Masakazu Konishi. Coding of auditory space. Annual review of neuroscience, 26(1):31 -- 55, 2003.
[24] Eve Marder and Dirk Bucher. Understanding circuit dynamics using the stomatogastric nervous system of
lobsters and crabs. Annu. Rev. Physiol., 69:291 -- 316, 2007.
[25] Theodore H Bullock, Michael H Hofmann, Frederick K Nahm, John G New, and James C Prechtl. Event-
related potentials in the retina and optic tectum of fish. Journal of Neurophysiology, 64(3):903 -- 914,
1990.
[26] Theodore H Bullock, Sacit Karamürsel, Jerzy Z Achimowicz, Michael C McClune, and Canan Ba¸sar-
Eroglu. Dynamic properties of human visual evoked and omitted stimulus potentials. Electroencephalog-
raphy and clinical neurophysiology, 91(1):42 -- 53, 1994.
[27] Nikhil Rajiv Deshmukh. Complex computation in the retina. PhD thesis, Princeton University, 2015.
[28] Niru Maheswaranathan, Alex H Williams, Matthew D Golub, Surya Ganguli, and David Sussillo. Univer-
sality and individuality in neural dynamics across large populations of recurrent networks. Advances in
neural information processing systems, 2019.
[29] Pouya Bashivan, Kohitij Kar, and James J DiCarlo. Neural population control via deep image synthesis.
Science, 364(6439):eaav9436, 2019.
[30] Edgar Y Walker, Fabian H Sinz, Emmanouil Froudarakis, Paul G Fahey, Taliah Muhammad, Alexander S
Ecker, Erick Cobos, Jacob Reimer, Xaq Pitkow, and Andreas S Tolias. Inception in visual cortex: in
vivo-silico loops reveal most exciting images. bioRxiv, page 506956, 2018.
[31] Robert J Aumann and Lloyd S Shapley. Values of non-atomic games. Princeton University Press, 1974.
11
Supplemental Materials
5 Path-integrated gradients for nonlinear dynamical systems
A specific firing burst of a ganglion cell r(t) (e.g. omitted stimulus response, motion reversal etc.) is a
result of spatiotemporal visual stimuli s(t, x, y) processed through a non-linearly entangled web of neurons,
r(t) = F[s(t, x, y)]. Here we aim to identify and isolate the minimal neural circuit at play, in order to obtain
theoretical and biophysical insights responsible for an array of retinal phenomena. Towards this goal, we need to
attribute the contributions to each of the input pixels or interneurons. For a linear system, the product of the
response function and an input stimulus χ(∆t, x, y)s(t − ∆t, x, y) quantifies how much an input pixel changes
the output. However, it is not obvious how to introduce such measure local in (t, x, y) space for non-linear
systems, thus raises questions: how can we, in non-linear systems, quantify the contributions of each of the (1)
spatiotemporal input pixels, (2) internal neurons, and (3) different cell types (filters)? Here we harness attribution
methods based on path-integrated gradients, first introduced in game theory [31], and then recently in machine
learning [9]. We then extend the theoretical framework to a movie input with an additional dimension of time.
Retina as a non-linear functional
Mathematically, a mapping from a space of spatiotemporal stimulus s(t, x, y) to a scalar response of a ganglion
cell r(t) at time t can be represented by a functional,
(5)
Here we assumed (i) causality and (ii) finite temporal memory of a ganglion cell with a cutoff at ∆tm. Thus, the
input s(∆t, x, y; t) is a subset of the entire spatiotemporal stimulus s(t(cid:48), x, y) in a finite region t ∈ [t − ∆tm, t]
and parametrized by t.
Expansion of a general functional v.s. linear response theory
The Taylor expansion of the above functional around s(∆t, x, y; t) ∼ 0 is,
r(t) = F(cid:2)s(∆t, x, y; t)(cid:3).
r(t) = F(cid:2)s(∆t, x, y; t)(cid:3)
(cid:90)
= F [0] +
d(∆t1)dx1dy1
δs(∆t1, x1, y1; t)
In comparison, the linear response theory takes the form of,
r(t) = F(cid:2)s(∆t, x, y; t)(cid:3) = F[0] +
(cid:12)(cid:12)(cid:12)(cid:12)s=0
δF(cid:2)s(t)(cid:3)
(cid:90)
s(∆t1, x1, y1; t) + O(s2)
d(∆t)dxdyχ(∆t, x, y)s(∆t, x, y; t),
(6)
(7)
where χ is the linear response function (e.g. receptive field around zero stimuli).
Two main approximations are made in the linear response theory; (i) no stimulus s dependence of δF
δs , (ii) trun-
cation of the Taylor expansion at the first order. These two assumptions are valid only when s(∆t, x, y; t) (cid:28) 1
holds everywhere. However, this condition hardly holds in visual systems in action, and the above conventional
formulation is due to theoretical and experimental limitations rather than a well-justified approximation as in
Figure 7A for schematics.
Figure 7: How important is an input pixel or a cell type (filter)? (A) [Linear theory: red]
Conventional linear theory, such as classical receptive field, extrapolates linear trend to all over the
input stimulus space. Deep learning model trained on natural scenes is able to compute local gradient,
"instantaneous receptive field", given an input of structured stimuli with high efficiency. However,
the instantaneous receptive field only explains the local property of the non-linear function that does
not necessalily reflect the global structure. [IntegratedGradients in 1D: blue] By integrating the
local gradients captured by a deep learning model over input axis, we can obtain a more acculate
description of the non-linear function. (B) The path-integrated gradients becomes dependents on
the path in higher dimensions (D > 1), but straight path is shown to satisfy desired axioms [9].
Schematic illustrates the case of two-dimensions.
12
@F(↵x0,↵y0)@x@F(↵x0,↵y0)@y(x0,y0)F(x,y)(A)xx0rr0=dF(0)dx·x0(B)Instantaneous receptive field1D2Dr0=F(x0)=x0Z10d↵dF(↵x0)dxFigure 2: IntegratedGradientss0s1G While both of the biological retina and the deep learning model gives r(t) = F(cid:2)s(∆t, x, y; t)(cid:3) upon a
presentation of a stimulus, the stimulus-dependent instantaneous receptive field
(8)
is easily computable in the deep learning model. Below, we harness the quantity δF/δss to investigate where a
ganglion cell is looking at during a burst of neural firings.
Review: Path integrated gradients for an image input
Here we first review a case of an image input r = F[s(x, y)], where s(x, y) represents a two dimensional image.
For a general path γ(α) parameterized by α,
r(α = 1) − r(α = 0) = F[s(x, y, α = 1)] − F [s(x, y, α = 0)]
(cid:90)
(cid:20) δF(x, y; α)
(cid:90) 1
0
=
dxdy
dα
δs(x, y; α)
∂α
(cid:21)
(cid:90)
∂s(x, y; α)
≡
dxdyAγ(α)(x, y).
(9)
Sundararajan et al. [9] proved that a straight path s(x, y; α) = αs(x, y) satisfies an array of desirable axioms:
Sensitivity, Insensitivity, Linearity, Preservation, Implementation invariance, Completeness, and Symmetry. In
that case, the above further simplifies to,
AIG(x, y) = s(x, y)
F(x, y; α)
δs(x, y; α)
dα
.
(10)
See Figure 7 for schematics.
Space-time attribution by path integrated gradients for a movie input
Here we try to extend the theoretical framework of the above IntegratedGradients to the case of a movie input.
In addition to the spatial components (x, y), a movie input has an additional dimension of time t.
In this case, the attribution score A(∆t, x, y; t) can be computed by taking a path integral parametrized by α,
(cid:12)(cid:12)(cid:12)(cid:12)s
δF
δs
(cid:90)
(cid:90)
(cid:90)
d(∆t)dxdy
dα
(cid:20)(cid:90)
r(t) = F(cid:2)s(∆t, x, y; t)(cid:3) − F(cid:2)s(∆t, x, y; α = 0)(cid:3)
δF(cid:2)s(∆t, x, y; α)(cid:3)
(cid:12)(cid:12)(cid:12)(cid:12)s(α)
(cid:12)(cid:12)(cid:12)(cid:12)s(α)
δF(cid:2)s(∆t, x, y; α)(cid:3)
d(∆t)dxdyA(∆t, x, y; t).
δs(∆t, x, y; α)
δs(∆t, x, y; α)
(cid:20)(cid:90)
=
≡
dα
(cid:21)
(cid:21)
.
s(∆t, x, y; α)
∂α
s(∆t, x, y; α)
∂α
Thus, attribution score can be defined as
A(∆t, xx, y; t) ≡
(11)
(12)
For a linear system, where δF
δs is independent of s, we can simply replace it by a linear response function
δs s=0 = χ, and the attribution score reduces to a product of the linear response function and the difference
δF
between the current stimulus and the baseline,
A(∆t, x, y; t) = χ(∆t, x, y; t)(cid:0)s(∆t, x, y; t) − s(∆t, x, y; α = 0)(cid:1).
(13)
δs s depends on s and the attribution score A(∆t, x, y) depends on the
However, for a non-linear system, δF
space of function that we integrate over. Namely, each "path" corresponds to a different attribution method and
provides different explanations to different questions. In particular, for a movie input, the history of stimuli
forms a path naturally parametrized by time "t".
IntegratedGradients: Effective linear response function for non-linear systems
Here we focus on the straight path,
s(∆t, x, y; α) = αs(∆t, x, y; t).
(14)
In this case, the integral of the gradients act as an effective linear response function for a nonlinear system as
below,
AIG(∆t, x, y; t) ≡
dα
(cid:20)(cid:90)
δF(cid:2)αs(∆t, x, y; t)(cid:3)
δ(cid:0)αs(∆t, x, y; t)(cid:1) (cid:21)
= χ(∆t, x, y; t)s(∆t, x, y; t) = χ[s](t, t
(cid:48)
(cid:48)
, x, y)s(t
, x, y).
s(∆t, x, y; t)
(15)
By computing the attribution A(∆t, x, y; t) over the full temporal region, we can obtain the full attribution
A(t, t(cid:48), x, y). Note that the function now depends both on t and t(cid:48) due to nonlinearity. Furthermore, we can
reformulate the above equation into two forms to answer specific questions raised below.
13
input stimuli(cid:82) ∞
Figure 8: Straight path integration from the null stimulus We computed IntegratedGradients
for a periodic flashes with omitted stimulus by taking a straight path integration of gradients from
the null stimulus s0 = 0 to an actual stimulus s1. (A) (Left) Attribution map on response time t v.s
input time t(cid:48) space. Causality limits the attribution in the region of t(cid:48) < t. (Right) Attribution on
t(cid:48) dtA(t, t(cid:48)) quantifies the total amount of output created by each of the pixels of an
input movie. (B) (Left) Attribution map on response time t v.s. integration time ∆t space. (Right)
Attribution on integration time ∆t integrated over a single periodic burst triggered by flashes (blue)
and omitted stimulus response (red). The peak of attribution for the flash response (blue) is placed
earlier than the peak of attribution for the omitted stimulus response (orange). This is consistent with
our finding that the responses for periodic flashes are caused by a fast OFF bipolar cell, while the
omitted stimulus response is triggered by slow ON bipolar cells. (See main text.)
the attribution in the region of t(cid:48) < t. (Right) Attribution on input stimuli(cid:82) ∞
Figure 9: Straight path integration between two stimuli We computed IntegratedGradients to
discriminate between two stimuli continuing periodic flashes s0, and periodic flashes with omitted
stimulus s1. (A) (Left) Attribution map on response time t v.s input time t(cid:48) space. Causality limits
t(cid:48) dtA(t, t(cid:48)) quantifies
the total amount of output created by each of the pixels of ano input movie. Strikingly, the attribution
concentrates at the location of the first omitted stimulus among the four omitted stimulus. (B) (Left)
Attribution map on response time t v.s. integration time ∆t space. (Right) Attribution on integration
time ∆t integrated over a single periodic burst triggered by flashes (blue) and omitted stimulus
response (orange).
Q1. Which pixel in the spatiotemporal input stimulus "(t, x, y)" is responsible for a ganglion cell's re-
sponse?
(cid:90) t1
(cid:90)
(cid:18)(cid:90) t1
dtr(t) =
(cid:48)
dt
dxdy
dtA(t, t
(cid:48)
, x, y)
t0
t0
(cid:19)
See Figure 8 for an example with OSR.
Q2. Which part of the spatiotemporal kernel is "(∆t, x, y)" responsible for a ganglion cell's response?
(16)
(cid:19)
(cid:90) t1
(cid:90)
(cid:90) t1
(cid:90) t1
dtr(t) =
dxdy
dt
t0
t0
t0−∆tm
See Figure 9 for an example with OSR.
, x, y)δ(cid:0)(t−t
(cid:48)
)−∆t(cid:1) =
(cid:90)
(cid:48)A(t, t
dt
(cid:48)
(cid:18)(cid:90) t1
d∆tdxdy
dtA(t, t−∆t, x, y)
t0
(17)
6 Extension to deeper CNNs without spatial invariances in stimuli
In the main text, we demonstrated how we can reduce CNNs with three layers to multi-pathways linear non-linear
models, specifically to validate deep CNNs in the retina where we had neurobiological ground truth. Here, we
discuss one possible way to extend our method to deeper CNNs of depth D processing natural movies through a
dynamic programming (DP) approach that works backwards from layer D to layer 1. First, note a natural movie
of limited duration without spatial invariances is still well approximated by a low dimensional trajectory in both
pixel space and every hidden layer. Let K be the max dimensionality for spatial input patterns for any channel
14
SI Figure: Path-integrated gradients for spatiotemporal inputOSRFlashesOSRFlashess0s1Straight path integration from the null stimuliStraight path integration between stimulus, and s0s1(A)(B)(A)(B)s0s1SI Figure: Path-integrated gradients for spatiotemporal inputOSRFlashesOSRFlashess0s1Straight path integration from the null stimuliStraight path integration between stimulus, and s0s1(A)(B)(A)(B)s0s1 in any layer. Then the basic idea is to attribute the response in layer D to the K dimensional space of inputs
to each channel in layer D − 1 using integrated gradients. We first find the important channels in layer D − 1
using methods presented in the main text. Then we recursively iterate via the same method to layer D − 2 and
so on down to the pixel layer. Because of the DP-like nature of our algorithm, the computational complexity
(after dimensionality reduction to K) is O(DKC) where C is the max number of channels in a layer, and not
exponential in D. The end result is a set of important channels in each layer, along with, for each important
channel ≤ K linear combinations of neurons that matter for generating the response in layer D.
15
|
1801.09900 | 1 | 1801 | 2018-01-30T09:01:00 | Human Echolocation in Static Situations: Auditory Models of Detection Thresholds for Distance, Pitch, Loudness and Timbre | [
"q-bio.NC"
] | We investigated, by using auditory models, how three perceptual parameters, loudness, pitch and sharpness, determine human echolocation. We used acoustic recordings from two previous studies, both from stationary situations, and their resulting perceptual data as input to our analysis. An initial analysis was on the room acoustics of the recordings. The parameters of interest were sound pressure level, autocorrelation and spectral centroid. The auditory models were used to analyze echolocation resulting from the perceptual variables, i.e. loudness, pitch and sharpness. Relevant auditory models were chosen to simulate each variable. Based on these results, we calculated psychophysical thresholds for detecting a reflecting object with constant physical size. A non-parametric method was used to determine thresholds for distance, loudness, pitch and sharpness. Difference thresholds were calculated for the psychophysical variables, since a 2-Alternative-Forced-Choice Paradigm had originally been used. We found that (1) blind persons could detect objects at lower loudness values, lower pitch strength, different sharpness values and at further distances than sighted persons, (2) detection thresholds based on repetition pitch, loudness and sharpness varied and depended on room acoustics and type of sound stimuli, (3) repetition pitch was useful for detection at shorter distances and was determined from the peaks in the temporal profile of the autocorrelation function, (4) loudness at shorter distances provides echolocation information, (5) at longer distances, timbre aspects, such as sharpness, might be used to detect objects. We also discuss binaural information, movements and the auditory model approach. Autocorrelation was assumed as a proper measure for pitch, but the question is raised whether a mechanism based on strobe integration is a viable possibility. | q-bio.NC | q-bio | 1
Human Echolocation in Static Situations: Auditory Models
of Detection Thresholds for Distance, Pitch, Loudness and
Timbre
Bo N. Schenkman1*¶, Vijay Kiran Gidla2¶,
1CTT - Centre for Speech Technology, Department of Speech, Music and Hearing
Royal Institute of Technology, Stockholm, Sweden.
*Corresponding author, email: [email protected]
2Blekinge Institute of Technology, Karlskrona, Sweden.
Email: [email protected]
¶The authors contributed equally to this work.
2
Human echolocation describes how people use reflected sounds to obtain information about
their ambient world. We investigated, by using auditory models, how three perceptual
parameters, loudness, pitch and sharpness, determine echolocation. We used acoustic
recordings from two previous studies, both from stationary situations, and their resulting
perceptual data as input to our analysis. An initial analysis was on the room acoustics of the
recordings. The parameters of interest were sound pressure level, autocorrelation and spectral
centroid. The auditory models were used to analyze echolocation resulting from the
perceptual variables, i.e. loudness, pitch and sharpness. Relevant auditory models were
chosen to simulate each variable. Based on these results, we calculated psychophysical
thresholds for detecting a reflecting object with constant physical size. A non-parametric
method was used to determine thresholds for distance, loudness, pitch and sharpness.
Difference thresholds were calculated for the psychophysical variables, since a 2-Alternative-
Forced-Choice Paradigm had originally been used. We found that (1) blind persons could
detect objects at lower loudness values, lower pitch strength, different sharpness values and at
further distances than sighted persons, (2) detection thresholds based on repetition pitch,
loudness and sharpness varied and depended on room acoustics and type of sound stimuli, (3)
repetition pitch was useful for detection at shorter distances and was determined from the
peaks in the temporal profile of the autocorrelation function, (4) loudness at shorter distances
provides echolocation information, (5) at longer distances, timbre aspects, such as sharpness,
might be used to detect objects. We also discuss binaural information, movements and the
auditory model approach. Autocorrelation was assumed as a proper measure for pitch, but the
question is raised whether a mechanism based on strobe integration is a viable possibility.
Keywords: blind; echolocation; psycho-acoustics; auditory models; detection thresholds
Author summary
Blind people use other senses than vision to orient and move around. Hearing, especially
echolocation, i.e. reflection of echoes from objects, might be used to detect objects. Sounds
can be produced by the person, by mouth or cane tapping, but can also arise from the
environment, e.g. from traffic. We have been studying perceptual processes when blind
people use echolocation by reanalyzing results from earlier studies. We used models of
hearing to understand how these sounds are processed. Blind people could detect objects in
more difficult conditions and at longer distances than sighted people. Detection was based on
pitch, loudness and a timbre aspect, sharpness. The perception of pitch was useful at shorter
distances and depended on the time properties of the sound and on loudness. At longer
distances timbre aspects, such as sharpness, might be used to detect objects. Understanding
echolocation is important for how people manage to move around in their surroundings and
how our senses process spatial information of the environment. The issues involved are of
interest for physiologists, psychologists and biologists. Most importantly, the issues are of
practical relevance for blind people learning how to safely move around and for the design of
mobility aids.
3
Introduction
Persons with blindness use echolocation to obtain information about their surroundings. A
person or a source in the environment emits a sound and the reflection is perceived. Both
static and dynamic means are used for this sensory information. In both cases the person has
to perceive if an object or obstacle is in front of him/her. This perceptual decision is
determined by a threshold of detection. The threshold may vary as a function of a number of
variables, like the character of the sound emitted, the rates of this sound, its position relative
to the person, if motion is involved and the experience and expertise in echolocation. For a
review of human echolocation, see Stoffregen and Pittenger [1], Kolarik et al [2] and Thaler
and Goodale [3]. Physical properties may have different effects on psychoacoustic parameters
that are used to determine if an object is in front or not. Three psychoacoustic parameters are
particularly important as sources for human echolocation, viz. pitch in the form of repetition
pitch, loudness and spectral information that is perceived as timbre. We describe how the
information provided by pitch, loudness or timbre may result in their respective detection
thresholds for echolocation. We limit ourselves to stationary situations, i.e. when neither
object nor person is moving. When movement is involved, more potential information may be
provided (Wilson, [4]; Wallmeier and Wiegrebe, [5]). We also determine at what distance
there is a threshold when a person may detect a reflecting object. A number of auditory
models were applied to the physical stimuli and we related the results of these models to the
perceptual responses of participants from two previous empirical studies, Schenkman and
Nilsson [6] and Schenkman, Nilsson and Grbic [7] which also will be referred to as SN2010
and SNG2016, respectively.
Psychoacoustic and neuroimaging methods are very useful for describing the high
echolocating ability of the blind and their underlying processes. However, they do not fully
reveal the information in the acoustic stimulus that determines echolocation (at least when the
source for the information is not known) and how this information is encoded in the auditory
system. We wanted to know how this information is represented and processed in the human
auditory system. One fruitful way to study the information necessary for human echolocation
is by signal analysis on the acoustic stimulus. However, such an analysis is only directed at
the physical properties of the sound, and does not fully show how the information is
represented in the human auditory system. To investigate this, we used auditory models which
mimic human hearing. Analyzing the acoustic stimulus using these models provide insight
into the processes for human echolocation. They may also allow testing of hypotheses by
comparing models (Dau [8]).
Loudness, pitch (Schenkman and Nilsson [9]) and timbre are three perceptual
attributes of an acoustic sound that are relevant for human echolocation. The backgrounds of
these attributes for modeling human echolocation are discussed next.
Loudness
Loudness is the perceptual attribute of sound intensity and is defined as that attribute of
auditory sensation in terms of which sounds can be ordered on a scale from quiet to loud
(ASA 1973 [10]). The dynamic range of the auditory system is wide and different
mechanisms play a role in intensity discrimination. Psychophysical experiments suggest that
neuron firing rates, spread of excitation and phase locking play a role in intensity perception,
but the latter two may not always be essential. A disadvantage with the neuron firing rates is
that, although the single neurons in the auditory nerve can be used to explain the intensity
discrimination, this does not explain why the intensity discrimination is not better than
observed, suggesting that the discrimination is limited by the capacity of the higher levels in
the auditory system, which may also play a role in intensity discrimination (Moore 2013
4
[11]). Several models (Moore [11] pp 139 - 140) have been proposed to calculate the average
loudness that would be perceived by listeners. The basic structure of these models is that,
initially the outer and middle ear transformations are performed and then the excitation
pattern is calculated. The excitation pattern is transformed into specific loudness, which
involves a compressive non-linearity. The total area calculated for the specific loudness
pattern is assumed to be proportional to the overall loudness. Therefore, independent of the
mechanism underlying the perception of loudness, the excitation pattern is the essential
information that is needed for an auditory model of loudness, and thus also for understanding
human echolocation in its utilizing of loudness.
Pitch
Pitch is "that attribute of auditory sensation in terms of which sounds may be ordered on a
musical scale" (ASA 1960 [12]). One view of the underlying mechanisms of pitch is that, as
the cochlea is assumed to perform a spectrum analysis, the acoustic vibrations are transformed
into the spectrum, coded as a profile of discharge rate across the auditory nerve. An
alternative view proposes that the cochlea transduces the acoustic vibrations into temporal
patterns of neural firing. These two views are known as the place and time hypotheses.
According to the place hypothesis, pitch is determined from the position of maximum
excitation along the basilar membrane, within the cochlea. This explains how the pitch is
perceived by pure tones at low levels, but it fails to explain the perception of pure tones at
higher levels. At such levels, due to the non-linearity of the basilar membrane, the peaks
become broader and tend to shift towards a lower frequency place. This should lead to a
decrease in pitch, but psychophysical experiments show that the pitch is stable. Another case
where the place hypothesis fails is its inability to explain the pitch of stimuli whose
fundamental is absent. According to the paradox of the missing fundamental, the pitch evoked
by a pure tone remains the same if we add additional tones with frequencies that are integer
multiples of that of the original pure tone, i.e. harmonics. It also does not change if we then
remove the original pure tone, the fundamental (De Cheveigné 2010 [13]).
Since the time hypothesis states that pitch is derived from the periodic pattern of the
acoustic waveform it overcomes the problem of the missing fundamental. However, the main
difficulty with the time hypothesis is that it is not easy to extract one pulse per period, in a
way that is reliable and fully general. Psychoacoustic studies also show that pitch exists for
sounds which are not periodic. Of interest to our subject matter, human echolocation, is an
instance of such sounds, namely iterated ripple noise. It is a sound that models some of the
human echolocation signals (e.g. Bilsen [14]).
In order to overcome the limitations of the place and time hypothesis two new
theories have been proposed, pattern matching (De Boer [15-16]; De Cheveigné [13]), and a
theory based on autocorrelation (Licklider [17]; De Cheveigné [13]). De Boer [15-16]
described pattern matching such that the fundamental partial is the necessary correlate of
pitch, but it may be absent if other parts of the pattern are present. In this way pattern
matching supports the place hypothesis. Later Goldstein [18], Wightman [19] and Terhardt
[20] described other models for pattern matching. One problem with the pattern matching
theory is that it fails to account for pitch whose stimuli have no resolved harmonics.
The autocorrelation hypothesis assumes temporal processing in the auditory system.
It states that, instead of detecting the peaks at regular intervals, the periodic neural pattern is
processed by coincidence detector neurons that calculate the equivalent of an autocorrelation
function (Licklider [17]; De Cheveigné [13]). The spike trains are delayed within the brain by
various time lags (using neural delay lines) and are combined or correlated with the original.
When the lag is equal to the time delay between spikes, then the correlation is high and
outputs of the coincidence detectors tuned to that lag are strong. Spike trains in each
5
frequency channel are processed independently and the results are combined into an aggregate
pattern. However, De Cheveigné [13] argued that the autocorrelation hypothesis works too
well: It predicts that, pitch should be equally salient for stimuli with resolved and unresolved
partials, but this is not the case. An alternative to the theory based on an autocorrelation like
function is the strobe temporal integration (STI) of Patterson Allerhand, and Giguere [21]. In
accordance with STI the auditory image underlying the perception of pitch is obtained by
using triggered, quantized, temporal integration, instead of an autocorrelation function. The
STI works by finding the strobes from the neural activity pattern and integrating it over a
certain period.
Thus, there is no full understanding of how pitch is perceived. Irrespective if
temporal, spectral or multi mechanisms determine pitch perception, the underlying
information that the auditory system uses to detect pitch has to be the excitation pattern on the
basilar membrane. Hence, the excitation pattern is the crucial information that should be
simulated by an auditory model for pitch perception, and thus also for human echolocation.
Repetition pitch
Human echolocation signals consist of an original sound along with a reflected or delayed
signal. Several studies have been presented to explain the pitch perception of such sounds.
Bassett and Eastmond [22] examined the physical variations in the sound field close to a
reflecting wall. They reported a perceived pitch caused by the interference of direct and
reflected sound at different distances from the wall; the pitch value being equal to the inverse
of the delay. In a similar way, Small and McClellan [23] and Bilsen [24], delayed identical
pulses and found that the pitch perceived was equal to the inverse of the delay, naming it time
separation pitch, and repetition pitch, respectively. When a sound and the repetition of that
sound are listened to, a subjective tone is perceived with a pitch corresponding to the
reciprocal value of the delay time (Bilsen and Ritsma [25]). They explained the repetition
pitch phenomenon by the autocorrelation peaks or the spectral peaks. Yost [26] performed
experiments using iterated ripple noise stimuli and concluded that autocorrelation was the
underlying mechanism, used by the listeners to detect repetition pitch.
Timbre
When the loudness and pitch of an acoustic sound are similar, the subjective attribute of
sound which may distinguish or identify the sound is its timbre. Timbre has been defined as
that attribute of an auditory sensation which enables a listener to judge that two non-identical
sounds, similarly presented and having the same loudness and pitch, are dissimilar (ANSI
1994 [27]). One example is the difference between two musical instruments playing the same
tone, e.g. a guitar and a piano. Timbre is a multidimensional percept and there is no single
scale on which we can order timbre. To quantify timbre one approach is to consider the
overall distribution of the spectral energy. Plomp and co-workers [28-29] showed that the
perceptual differences between different sounds were closely related to the levels in 18 1/3
octave bands, thus relating the timbre to the relative level produced by the sound in each
critical band. Hence, generally, for both speech and non-speech sounds, the timbre of steady
tones is determined by their magnitude spectra, although the relative phases may play a small
role [30]. When we consider time varying patterns, there are several factors that may
influence the perception of timbre: (i) periodicity; (ii) variation of the envelope of the
waveform; (iii) spectrum changes over time; and (iv) what the preceding and following
sounds were like. Using auditory models, timbre information can be assessed by the levels in
the spectral envelope and by the variations of the temporal envelope. Another way to preserve
the fine grain time interval information that is necessary for timbre perception is by the strobe
temporal integration (STI) method of Patterson, Allerhand, and Giguere [21].
6
and
PsySound3
were
Signal analysis
Signal analysis and auditory models may enable us to understand the processing of sounds for
persons using echolocation, since one then can consider the transmission of the acoustic
sound from the source via the internal representation to the final percept of the person. The
acoustic sound travels and undergoes transformation because of the room acoustics. One
should therefore first understand the information that is received at the human ear. Signal
analysis is useful for this purpose, as we then can analyze the characteristics of the sound,
which have been transformed due to room conditions. The second step is to analyze how the
desired characteristics of the acoustic sound, that contains the information, are represented in
the auditory system. Here auditory models are useful. The desired information is transformed
in an analogous way to how the auditory system is known to process it. Keeping track of the
information from the outer ear to the central nervous system will be an important part for
describing how listeners perceive sounds and explaining the differences between groups of
listeners with different characteristics, e.g. visually handicapped vs sighted persons. This is
the methodology which was used for this report.
downloaded
To model the auditory analysis performed by the human auditory system we used the
auditory image model of Patterson, Allerhand, and Giguere [21], the loudness models of
Glasberg and Moore [31-32] and the sharpness model of Fastl and Zwicker [33]. Matlab was
used as the implementation environment. The auditory image model has been implemented in
matlab by Bleeck, Ives, and Patterson [34] and the current version is known as AIM-MAT.
The loudness and the sharpness models were implemented in PsySound3 (Cabrera, Ferguson,
and Schubert [35]), a GUIdriven Matlab environment for analysis of audio recordings. AIM-
MAT
from
https://code.soundsoftware.ac.uk/projects/aimmat and http:// www.psysound.org, respectively.
Aims and hypothesis
The aims of the present study were:
(1) To study different components of the information in the acoustic stimulus, that determines
echolocation.
(2) To determine the thresholds for different components of the information in the acoustic
stimulus, that are important factors for the detection distance to reflecting objects.
(3) To find out how the acoustic information that determines the high echolocation ability of
the blind is represented in the human auditory system.
More specifically, our hypotheses were:
(1) Detection thresholds based on repetition pitch, loudness and sharpness will vary and will
depend on the room acoustics and type of the sound stimuli that is used.
(2) Repetition pitch is useful for detection at shorter distances and is determined from the
peaks in the temporal profile of the autocorrelation function, computed on the neural activity
pattern.
(3) Detection at shorter distances, based on loudness provides information for listeners.
(4) At longer distances timbre aspects, such as sharpness information might be used by
listeners to detect objects.
Structure of the report
This report is hereafter structured as follows: In the next part "Method", subtitled "Room
acoustics: Basic acoustic analyses of physical parameters" we describe the recordings used in
the studies by Schenkman and Nilsson, SN2010 [6] and Schenkman, Nilsson, and Grbic
SNG2016 [7], since these form the data for the present report. In the same part we present a
7
signal analysis conducted on these recordings. We describe basic room acoustic parameters of
the signals that form the basis for the physical information on the room to the persons,
whether objects are present or not. No consideration is here done to auditory models. In the
next part, "Models", subtitled "Auditory analysis of acoustic information", we describe the
auditory models, their designs and implementation. The loudness, pitch and sharpness for the
recordings of SN2010 [6] and SNG2016 [7] when analyzed using the auditory models are
presented in the Results, "Loudness analysis: Excitation patterns, binaural loudness, short and
long term loudness", "Pitch analysis: autocorrelation with dual profiles" and "Sharpness
analysis" sections, respectively. In the section thereafter, "Threshold values, absolute and
difference, for echolocation in static situations based on auditory model analysis," the
thresholds for object detection are presented. This is followed by Discussion and
Conclusions.
Method
Room acoustics: Basic acoustic analyses of physical parameters
Sound recordings used
Here we describe briefly how the sound recordings of SN2010 [6] and SNG2016 [7] were
made. For more detailed descriptions, see the original articles. In SN2010 [6], the binaural
sound recordings were conducted in an ordinary conference room and in an anechoic chamber
using an artificial manikin. The object was a reflecting 1.5 mm thick aluminum disk with a
diameter of 0.5 m. Recordings were conducted at distances of 0.5, 1, 2, 3, 4, and 5 m between
microphones and the reflecting object. In addition, recordings were made with no obstacle in
front of the artificial manikin. Durations of the noise signal were 500, 50, and 5 ms; the
shortest corresponds perceptually to a click. The electrical signal was a white noise. However,
the emitted sound was not perfectly white, because of the non-linear frequency response of
the loudspeaker and the system. A loudspeaker generated the sounds, resting on the chest of
the artificial manikin.
In SNG2016 [7] recordings were conducted in an ordinary lecture room. Recordings
were conducted at 100 and 150 cm distances between microphones and the reflecting object.
The emitted sounds were either bursts of 5 ms each, varying in rates from 1 to 64 bursts per
500 ms or a 500 ms white noise. In contrast to SN2010 [6], the sounds in SNG2016 [7] were
generated by a loudspeaker placed 1 m straight behind the center of the head of the artificial
manikin. The sound recording set ups can be seen in Fig 1.
8
Fig 1. Sound recordings used. (A) anechoic room (B) conference room, with loudspeaker on
the chest of the artificial manikin in Schenkman and Nilsson [6] (C) lecture room with
loudspeaker behind the artificial manikin in Schenkman, Nilsson, and Grbic [7].
Calibration of signals
Analyses were performed to determine basic room acoustical parameters relevant for human
echolocation: sound pressure level, autocorrelation, and spectral centroid. Before analyzing
the recordings, the recordings were calibrated by the calibrating constants (CC) using
Equation (1). Based on the Sound Pressure Level (SPL) of 77, 79 and 79 dBA for the 500 ms
recording without the object at the ear of the artificial manikin in the anechoic and conference
rooms of SN2010 and in the lecture room of SNG2016, the CC's were calculated to be
2.4663, 2.6283 and 3.5021, respectively. A-weighting was not included in Equation (1), since
the difference of using it was less than 0.5 dB and could thus be neglected. See S.1
Supporting Information for more details.
𝐶𝐶𝐶𝐶= 10�𝑆𝑆𝑆𝑆𝑆𝑆−20∗𝑙𝑙𝑙𝑙𝑙𝑙10�𝑟𝑟𝑟𝑟𝑟𝑟(𝑟𝑟𝑠𝑠𝑙𝑙𝑠𝑠𝑠𝑠𝑙𝑙)
20∗10−6 �
20
�
(1)
As the recordings were binaural, both left and right ear recordings were analyzed.
The recordings in SN2010 [6] and SNG2016 [7] had 10 versions of each duration and
distance. For the condition with no reflecting object, there were two sets of recordings in
SN2010 [6], 10 versions in each, while for the same condition in SNG2016 [7] there was one
set of recordings with 10 versions. The two versions with no object in SNG2010 [6], resulted
9
in very similar values, but are for fullness presented separately in this report. It should be
noted that the recordings vary over the versions causing the term "rms(signal)" in Equation
(1) to vary, thereby varying the calibrated constants for the 10 different versions. However, as
the variation was very small between the versions, we decided to use only the 9th version of
the 500 ms first recording without the object in SN2010 [6] and the 9th version of the 500 ms
recording without the object in SNG2016 [7] to establish the calibrated constants. Another
reason to choose only one version, the 9th, is that although the other versions may not have
exactly the identical CC's they will be relatively calibrated with respect to the recording of
version 9. For example, suppose the recording in the anechoic chamber version 1 had 67 dB
SPL and version 9 had 66 dB SPL before calibration, then the levels obtained by calibrating
the recordings to 77 dB SPL using the CC of the 9th version would be 78 dB SPL for version
1 and 77 dB SPL for version 9. In other words, they will give the same level difference, also
after calibration.
Sound Pressure Level (SPL)
Detection of objects by echolocation is to a certain extent based on intensity information.
Hence, the SPL in dBA were calculated using Equation (2), where "RMS" is the root mean
square amplitude of the signal analyzed. As has been pointed out by various authors (e.g.
Rowan et al. [36]), binaural information may be utilized for echolocation purposes. We
therefore calculated the SPL values for both ears. The mean SPL values of the 500 ms
recordings in study SN2010 [6] and in study SNG2016 [7] are shown in Tables 1 and 2,
respectively. The values for the 5 ms and 50 ms recordings are for reasons of space not shown
here. As mentioned above, for the recordings with no object, two series were conducted in
SN2010, each with 10 recordings. As can be seen in Table 1, the values are very close to each
other.
𝑆𝑆𝑆𝑆𝑆𝑆=20∗𝑙𝑙𝑙𝑙𝑙𝑙10�𝐶𝐶𝐶𝐶∗𝑟𝑟𝑟𝑟𝑟𝑟(𝑟𝑟𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠)
20∗10−6
� (2)
Table 1. Mean of the sound pressure levels (dBA) for the left and right ears over the 10
versions of the 500 ms duration signals in the anechoic and conference room used by
Schenkman and Nilsson [6]. For the recording with no object, two series were conducted.
Object Distance (cm)
No Object, recording 1
No Object, recording 2
50
100
200
300
400
500
Anechoic chamber
Conference room
Left ear
Right ear
Left ear
Right ear
77.2
77.6
85.2
81.9
77.1
77
77.1
77
77.9
77.4
88.2
82.6
78
78.2
78
78
79
79
87.5
82.8
79.6
78.9
79
79
78.8
78.8
87.5
82.4
79.5
78.9
78.9
78.8
10
Table 2. Mean of the sound pressure level (dBA) for the left and right ears over the 10
versions of the 500 ms duration signals in the lecture room used by Schenkman, Nilsson
and Grbic [7].
Lecture room
Object Distance (cm) Left ear Right ear
No Object
100
150
79.2
79.6
79.4
79.6
81.5
79.7
The tabulated SPL values in Tables 1 and 2 show the effect of room acoustics on
level differences, between the ears and between the rooms. The level differences between the
recording without object and the recordings with object were in SNG2016 less than those in
SN2010. This may be due to the differences in experimental setup (Fig 1) or to the acoustics
of the room. The extent to which this information affected the listeners in these studies is not
obvious, as loudness perceived by the human auditory system cannot be related directly to the
SPL (e.g. Moore [11]). This issue is analyzed further in the section "Loudness analysis:
Excitation patterns, binaural loudness, short and long term loudness".
Autocorrelation Function (ACF)
Repetition pitch is an important aspect of how we perceive complex sounds (Bilsen [24];
Bilsen and Ritsma [25]). Schenkman and Nilsson [9] showed that this pitch, rather than
loudness, is used by listeners to detect an object by echolocation. As noted above, pitch
perception can often be explained by the peaks in the autocorrelation function and therefore
an autocorrelation analysis was performed, which we present here.
The theoretical values for repetition pitch for the recordings of SN2010 [6] and
SNG2016 [7] were calculated using Equation (3). The corresponding values for recordings
with objects at distances of 50, 100, 150, 200, 300, 400 and 500 cm would be approximately
344, 172, 114, 86, 57, 43 and 34.4 Hz, assuming sound velocity to be 344 m/s. As the theory
based on autocorrelation uses temporal information, repetition pitch perceived at the above
frequencies can be explained by the peaks in the ACF at the inverse of the frequencies, i.e.
approximately at 2.9, 5.8, 8.7, 11.6, 17.4, 23.2 and 29 ms, respectively. The autocorrelation
analysis was performed using a 32 ms frame, which would cover the required pitch period. A
32 ms hop size was used to analyze the ACF for the next time instants of 64 ms, 96 ms etc. In
order to compare the peaks among all the recordings, the ACF was not normalized to the
limits -1 to 1.
𝑅𝑅𝑆𝑆=
𝑟𝑟𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑜𝑜𝑜𝑜 𝑟𝑟𝑜𝑜𝑠𝑠𝑠𝑠𝑠𝑠
2∗𝑠𝑠𝑠𝑠𝑟𝑟𝑑𝑑𝑠𝑠𝑠𝑠𝑑𝑑𝑠𝑠 𝑜𝑜𝑜𝑜 𝑑𝑑ℎ𝑠𝑠 𝑜𝑜𝑜𝑜𝑜𝑜𝑠𝑠𝑑𝑑𝑑𝑑
where RP is Repetition Pitch.
(3)
In the study SN2010 [6] the participants performed well with the longer duration
signals. For a single short burst the person had only one chance to perceive the signal and its
echo. This can be visualized from the ACFs in Figs 2 and 3, where for the 5ms recording the
peak was present only for the initial 32 ms frame. For the 500 ms recording the peak was also
present for frames with time instants greater than 32 ms. (Note that for each duration of the
signals an additional 450 ms silence was padded and presented to the test persons. The ACF
were analyzed in the same manner, and hence the 5ms duration signal had a total duration of
455 ms and the 500 ms signal had a total duration of 950 ms.)
11
Fig 2. The autocorrelation function of a 5 ms signal recorded in the anechoic chamber
(SN2010) [6] with reflecting object at 100 cm. The sub figures A-E show the autocorrelation
function (ACF index) at 32, 64, 96, 128, 160 and 192 ms time instants of the signal (Lag),
respectively. As the recording is only 5ms in duration the autocorrelation function is only
present in the first 32ms frame.
Fig 3. The autocorrelation function of a 500 ms signal recorded in the anechoic chamber
(SN2010) [6] with reflecting object at 100 cm. The sub figures A-E show the autocorrelation
function (ACF index) at 32, 64, 96, 128, 160 and 192 ms time instants of the signal (Lag),
respectively.
Schenkman and Nilsson [6] argued that the higher detection ability of their
participants for the longer duration signals may have been a result of attention, a confounding
12
factor. Even if a test person at first was not attentive, he or she had a longer time interval
available to perceive the signal. However, in Schenkman, Nilsson and Grbic [7] the
performance decreased at one distance for the longer duration noise, although the repetitions
were present for the frames with a time instant greater than 32 ms. These ACF functions are
for reasons of space not shown here. Therefore, that longer duration signals are always
beneficial for human echolocation cannot be concluded on the available results.
The peak heights at the pitch period for the recordings with object at 100 cm for the 5
ms duration signal in the conference room in SG2010 [6] were greater than those in the
lecture room in SNG2016 [7]. The 500 ms duration signal with object at 100 cm in the lecture
room in SN2016 [7] had a greater peak height than the 5 ms signal in the conference room in
SN2010 [6], but the peak is not distinct enough when compared to the 500 ms duration signal
in the conference room in SN2010 [6].
The cause for these differences in the peak heights between the two rooms,
conference room in SN2010 [6] and the lecture room in SNG2016 [7], are probably the
different room acoustics. The ACF depends on the spectrum of the signal, and the acoustics of
the room certainly influences the peaks in the ACF. The reverberation time, T60, for the
conference and the lecture room were 0.4 and 0.6 seconds, respectively. A fuller discussion of
how the information carried by the peaks is represented in the auditory system is further
discussed in the section "Pitch analysis: autocorrelation with dual profiles".
Timbre: Spectral Centroid (SC)
Detection of an object could be provided by the timbre information available in a sound.
Timbre is constituted by a number of different characteristics, e.g. roughness. Another
characteristic of timbre perception is the spectral centroid (Peeters et al. [37]), which gives a
time varying value characterizing the subjective center of the timbre for a sound. As
mentioned in the Introduction, in the timbre section previously, the timbre of steady tones are
mainly determined by their magnitude spectra. We believe that the spectral centroid is an
important feature of human echolocation in static situations. As the recordings in SN2010 [6]
and SN2016 [7] were static, the spectral centroid using the magnitude spectra of the
recordings was computed to depict the timbre information.
To compute the spectral centroid, the recordings were analyzed using a 32 ms frame
with a 2 ms overlap. The spectral centroid for each frame was computed by Equation (4). As
the spectral centroid for each frame is a time varying function, it is plotted as a function of
time. The means of the spectral centroid for the 10 versions at each condition for the 500 ms
of the left ear recordings are shown in Figs 4 to 6.
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑙𝑙𝐶𝐶𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆= ∑(𝐹𝐹𝑟𝑟𝑠𝑠𝐹𝐹𝑠𝑠𝑠𝑠𝑠𝑠𝑑𝑑𝐹𝐹∗𝐹𝐹𝐹𝐹𝐹𝐹(𝑜𝑜𝑟𝑟𝑠𝑠𝑟𝑟𝑠𝑠))
∑(𝐹𝐹𝐹𝐹𝐹𝐹(𝑜𝑜𝑟𝑟𝑠𝑠𝑟𝑟𝑠𝑠))
In SN2010 [6] for the recordings without the object, the spectral centroid was
approximately below 5000 Hz. For the recordings with the object at 50 and 100 cm, the
spectral centroids were approximately above 5000 Hz (Fig 4). This difference might provide
information to listeners to distinguish conditions with an object from those without an object.
The recordings with the object at 200 to 500 cm did not vary much when compared with the
recording without the object. In SNG2016 [7] the spectral centroid was approximately 6000
Hz for all recordings (Fig 6), showing very small changes. The timbre information in
SNG2016 [7] may thus at first sight not seem to be useful for echolocation. This physical
analysis indicates that there was variation in the spectral centroid in the recordings of SN2010
[6] with object at shorter distances (distances shorter than 200 cm) but for longer distances the
difference in the spectral centroid was almost negligible.
(4)
13
Fig 4. The means of the spectral centroid of the 10 versions as a function of time of the
left ear for a 500ms recording in the anechoic chamber (SN2010) [6]. The sub figures A,
B are for the two no object recordings and C-H are for the recordings with object at 50, 100,
200, 300, 400 and 500 cm, respectively.
Fig 5. The means of the spectral centroid of the 10 versions as a function of time of the
left ear for a 500 ms recording in the conference room (SN2010) [6]. The sub figures A, B
are for the two no object recordings and C-H are for the recordings with object at 50, 100,
200, 300, 400 and 500 cm, respectively.
The conclusions above are based on a purely physical analysis, a Fast Fourier
Transform (FFT) analysis of the sounds. However, the spectral analysis performed by the
auditory system is more complex than a FFT that we have used to compute the spectral
14
centroid. In the next section we will show that the conclusions will be modified, when we
consider how human hearing works. This we do by using auditory models to analyze the
sounds.
Fig 6. The mean of the spectral centroid for the 10 versions as a function of time of the
left ear for a 500 ms recording in the lecture room (SNG2016) [7]. The sub figure A is for
the no object recording, while sub figures B and C are for the recordings with object at 100
and 150 cm, respectively.
Models
Auditory analysis of acoustic information
Premises
In the previous sections in the Method part, we presented physical parameters from the two
studies SN2010 [6] and SNG2016 [7], that determine human echolocation. In this Models part
we take the physical parameters from above and study how they may provide the basis for
relevant auditory information to a person. In the Results part thereafter, we will connect this
auditory analysis with the behavioural results.
For the auditory analysis we used the auditory image model (AIM), originally
developed by Patterson et al. [38, 21] with extensions added by other authors. It is a time-
domain, functional model of the signal processing performed in the auditory pathway as the
system converts a sound wave into the perception that we experience when presented with a
sound. This representation is referred to as an auditory image by analogy with the visual
image of a scene that we experience in response to optical stimulation. The AIM simplifies
the peripheral and the central auditory systems into modules. A summarily description of the
AIM and how the modules were implemented in the present analysis is given below. A more
detailed description of each module of AIM can be found at http://www.acousticscale.
org/wiki/index.php/AIM2006_Documentation.
We used the modules described below to analyse the recordings. All the processing
modules of AIM are written in matlab. The current version, as mentioned earlier, is referred to
as AIM-MAT and can be downloaded from https://code.soundsoftware.ac.uk/projects/aimmat.
The autocorr module was only present in the 2003 version of AIM and can be downloaded
from http://w3.pdn.cam.ac. uk/groups/cnbh/aimmanual/download/downloadframeset.htm.
15
First step: Pre Cochlear Processing (PCP)
The outer middle ear transformation of the acoustic sound is simulated in AIM by the Pre
Cochlear Processing (PCP) module. The PCP module consists of four different Finite Impulse
Response (FIR) filters, designed for different applications. These are: (i) Minimum audible
field (MAF), which is suitable for signals presented in free field. (ii) Minimum audible
pressure (MAP), which is suitable for systems which produce a flat frequency response. (iii)
Equal loudness contour (ELC) and (iv) filter gm2002 (Glasberg and Moore [31]) are almost
identical and include the factors associated with the extra internal noise at low and high
frequencies. However, gm2002 uses more recent data of Glasberg and Moore [31]).
The MAF, MAP, ELC were designed using Parks-McClellan optimal equi-ripple FIR
filter design algorithm, while the gm2002 was designed using a frequency sampling method.
The transmission of the acoustic sound through the PCP filter can be modelled using Equation
(5), where Signalinput is the input to the AIM and Signalpcp is the filtered output of the
corresponding PCP filter.
An example of the frequency response used to generate a PCP filter is shown in Fig 7.
𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑠𝑠𝑑𝑑𝑠𝑠=𝑓𝑓𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑆𝑆�𝑆𝑆𝐶𝐶𝑆𝑆𝑜𝑜𝑠𝑠𝑠𝑠𝑑𝑑𝑠𝑠𝑟𝑟,𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑑𝑑�
(5)
Fig 7. The frequency response used to design the gm2002 filter of the Pre Cochlear
Processing (PCP) module in the AIM. The frequency response was obtained from the
frontal field to cochlea correction data of Glasberg and Moore [31].
Second step: Basilar Membrane Motion (BMM)
The non-linear spectral response of the basilar membrane is an important feature of the
peripheral auditory system. This response is implemented in the AIM by a dynamic
compressive gammachirp filter bank, dcGC (Irino and Patterson [39]). Two important
properties of the Basilar Membrane Motion (BMM) are the asymmetry and the compression
of the auditory filters that are made in proportion to the intensity level. These properties were
designed using a compressive gammachirp filter (cGC). It is a generalized form of the
gammatone filter, which was derived with operator techniques (Irino and Patterson [39]). The
developments of both the gammatone and gammachirp filters are described in Patterson,
Unoki, and Irino [40]. The cGC is simulated by cascading a passive gammachirp filter (pGC)
with a high pass asymmetric function (HP-AF). The asymmetrical property is simulated by
16
the pGC filter and its output is used to adjust the level dependency of the active part, i.e. the
HP-AF.
Other options available for generating the BMM in AIM are the gammatone function
and the pole zero filter cascade. Since the gammatone function does not depict the non-
linearity of the basilar membrane, we used the default filterbank dcGC to simulate the BMM.
The transformation of the BMM can be modelled using Equations (6) and (7). SignalpGC(fc) is
the filtered output of the pGC filterbank, fc is the centre frequency of the filter, ACF(fc) is the
high pass asymmetric compensation filters and SignalcGC(fc) is the final compressed output of
the BMM. For a detailed description of the pGC and cGC filterbanks, see Irino and Patterson
[39].
𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑠𝑠𝑝𝑝𝐶𝐶(𝑓𝑓𝑑𝑑)=𝑓𝑓𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑆𝑆�𝑆𝑆𝑃𝑃𝐶𝐶(𝑓𝑓𝑑𝑑),𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑠𝑠𝑑𝑑𝑠𝑠� (6)
𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑑𝑑𝑝𝑝𝐶𝐶(𝑓𝑓𝑑𝑑)=𝑓𝑓𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑆𝑆�𝐴𝐴𝐶𝐶𝐴𝐴(𝑓𝑓𝑑𝑑),𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑠𝑠𝑝𝑝𝐶𝐶(𝑓𝑓𝑑𝑑)� (7)
Third step: Neural Activity Pattern (NAP)
The basilar membrane motion is transduced into an electrical potential by the inner hair cells.
The Neural Activity Pattern (NAP) is implemented in AIM by half wave rectification
followed by low pass filtering. Low pass filtering is executed as phase locking is not feasible
for high frequencies in the human ear.
There are three modules in the AIM to generate the NAP: (i) half wave rectification
followed by compression and low pass filtering (H-C-L) (ii) half wave rectification followed
by low pass filtering (H-L) (iii) two dimensional adaptive threshold (similar to H-C-L but it
has adaptation which is more realistic). The choice of NAP module depends on the choice of
BMM module. As noted above, we used a dcGC filter bank in our analyses, and the
compression of the basilar membrane was simulated by it. The H-L module was therefore
chosen to generate the NAP. This transformation can be modelled using Equation (8), where
abs(Signalbmm(fc)) is the half wave rectified signal of the basilar membrane, fc is the centre
frequency of the filter, LPF is the low pass filter and Signalnap(fc) is the modelled NAP.
𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑠𝑠𝑠𝑠𝑠𝑠(𝑓𝑓𝑑𝑑)=𝑓𝑓𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑆𝑆(𝑆𝑆𝑆𝑆𝐴𝐴,𝑆𝑆𝑎𝑎𝑎𝑎(𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑙𝑙𝑜𝑜𝑟𝑟𝑟𝑟(𝑓𝑓𝑑𝑑)))
(8)
Fourth step: Strobe Temporal Integration (STI)
The fourth stage in the AIM represents processing in the central nervous system. Perceptual
research suggests that at least some of the fine grain time interval information is needed to
preserve timbre information (Krumbholz et al [41], Patterson [42 - 43]). Auditory models
often time average the NAP information, which unfortunately then loses the fine grain
information. To prevent this, AIM uses a procedure called Strobe Temporal Integration (STI),
which is subdivided into two modules, (i) strobe finding, and (ii) temporal integration.
Strobe Finding (SF): The sub module sf2003 is used to find the strobes from the
NAP. It uses an adaptive strobe threshold to issue a strobe and the time of the strobe is that
associated with the peak of the NAP pulse. After the strobe is initiated the threshold initially
rises along a parabolic path and then returns to the linear decay to avoid spurious strobes. The
duration of the parabola is proportional to the centre frequency of the channel and the height
to the height of the strobe. After the parabolic section of the adaptive threshold, its level
decreases linearly to zero in 30 ms. An additional feature of sf2003 is the inter channel
interaction, i.e. a strobe in one channel reduces the threshold in the neighbouring channels. An
example of how the threshold varies and how the strobes are calculated is shown in Fig 8.
17
Fig 8. The Neural Activity Pattern (NAP) of a 200 Hz pure tone in the 253 Hz frequency
channel. The dashed line shows the threshold variation and the dots indicate the calculated
strobes.
Temporal Integration (TI): The temporal integration is implemented in AIM by a
module called stabilized auditory image (SAI). The SAI in its turn uses a sub module, ti2003,
to accomplish this. The ti2003 module changes the time dimension of the NAP into a time
interval dimension. This works as follows: Initially, a temporal integration is initiated when a
strobe is detected. If no further strobes are detected, the process continues for 35 ms and
stops. If strobes are detected within the 35 ms interval, each strobe initiates a temporal
integration process. To preserve the shape of the SAI to that of the NAP, ti2003 uses
weighting. The new strobes are initially weighted high (the weights are also normalized so
that the sum of the weights is equal to 1) making the older strobes contribute relatively less to
the SAI.
Autocorrelation Function (ACF)
As was mentioned above, the AIM offers a module, autocorr, to analyze autocorrelation
processes. Corresponding physiological processes are presumed to take place in the central
nervous system (e.g. Licklider [17]; Bilsen [24; Yost [26]). By using the autocorr module one
can implement models of hearing based on autocorrelation processes. The autocorr module
takes the NAP as input and computes the ACF on each center frequency channel of the NAP
by using a duration of 70 ms, hop time of 10ms and a maximum delay of 35ms.
Results
Loudness analysis: Excitation patterns, binaural loudness, short
and long term loudness
In the room acoustics part earlier, for the sound pressure level analysis we described the
physical intensity of a sound that may affect human echolocation. This description is
necessary for understanding echolocation, but it is not sufficient. Intensity is related to the
perception of loudness which is a psychological attribute, but loudness also depends on a
number of other parameters, primarily but not only, frequency selectivity, bandwidth and
duration of the sound. In this section we consider perceptual aspects of loudness for
echolocating sounds, using the loudness model of Glasberg and Moore [31]. We chose this
model instead of the AIM for the following reason.
18
The loudness model of Glasberg and Moore (2002) computes the frequency
selectivity and compression of the basilar membrane in two stages, (1) by computing the
excitation pattern and (2) by the specific loudness of the input signal. Physiologically they are
interlinked and a time domain filter bank which simulates both the selectivity and the
compression might be appropriate. Although there are different time domain models of the
level dependent auditory filters available in AIM (e.g. dcGC), they do not give a sufficiently
good fit to the equal loudness contours in ISO 2006 (Moore [44]). Since we consider this fit to
be an important aspect, so instead of choosing the AIM model to model loudness, we used the
model of Glasberg and Moore [31].
A loudness model should consider the outer middle ear filtering, the non-linearity of
the basilar membrane and the temporal integration of the auditory system. The loudness
model of Glasberg and Moore [31] estimates the loudness of steady sounds and of time
varying sounds, by accounting for these features of the human auditory system. Each stage of
this model is described briefly below.
Outer middle ear transformation
The outer middle ear transformation was modelled using a finite impulse response (FIR) filter
with 4097 coefficients. The response at the inner ear can be represented using Equation (9),
where x and yomt are the signals before and after transformation, and h is the impulse response
of the filter.
𝑦𝑦𝑜𝑜𝑟𝑟𝑑𝑑=𝑓𝑓𝑆𝑆𝑙𝑙𝑆𝑆𝑆𝑆𝑆𝑆(ℎ,𝑥𝑥) (9)
Excitation pattern
The excitation pattern is defined as the magnitude of the output of each auditory filter as a
function of the filter center frequency. To compute the excitation pattern from the time
domain signal, Glasberg and Moore [31] used six FFTs in parallel based on Hanning-
windowed segments with durations of 2, 4, 8, 16, 32 and 64 ms, all aligned at their temporal
centers. The windowed segments are zero padded, and all FFTs are based on 2048 sample
points. All FFTs are updated at 1 ms intervals and each FFT was used to calculate the spectral
magnitudes at specific frequency ranges. Values outside the range were discarded.
The running spectrum was the input to the auditory filters, and their output was
calculated at the center frequency of 0.25 Equivalent Rectangular Bandwidth (ERB) intervals
taking into account the known variation of the auditory filter shape regarding center frequency
and level. The excitation pattern is defined as the output of the auditory filter as a function of
center frequency (Glasberg and Moore [31]). This can be represented with Equation (10),
where E(fc) is the magnitude of the output of each auditory filter with center frequency fc,
Yomt is the power spectrum of yomt calculated using six parallel FFT's, as mentioned above,
over a 1ms interval and W(fc) is the frequency response of the auditory filter at center
frequency fc .
𝐸𝐸(𝑓𝑓𝑑𝑑)= 𝑌𝑌𝑜𝑜𝑟𝑟𝑑𝑑∗𝑊𝑊(𝑓𝑓𝑑𝑑) (10)
Specific loudness (SL)
To model the non-linearity of the basilar membrane, the excitation pattern has to be
converted to specific loudness. Specific loudness is the loudness in a critical band. This
conversion is done in the model of Glasberg and Moore [31] using three conditions (Equation
11).
19
𝑆𝑆𝑆𝑆(𝑓𝑓𝑑𝑑)=⎩⎪⎨⎪⎧𝐶𝐶∗�
𝐸𝐸(𝑜𝑜𝑐𝑐)+ 𝐹𝐹𝑄𝑄(𝑜𝑜𝑐𝑐)�1.5+((𝑃𝑃∗𝐸𝐸(𝑓𝑓𝑑𝑑)+𝐴𝐴 )𝛼𝛼− 𝐴𝐴𝛼𝛼)
2𝐸𝐸(𝑜𝑜𝑐𝑐)
((𝑃𝑃∗𝐸𝐸(𝑓𝑓𝑑𝑑)+𝐴𝐴 )𝛼𝛼− 𝐴𝐴𝛼𝛼)
𝐶𝐶∗� 𝐸𝐸(𝑜𝑜𝑐𝑐)
1.04∗106�0.5
𝑠𝑠𝑜𝑜 𝐸𝐸(𝑜𝑜𝑐𝑐)≤ 𝐹𝐹𝑄𝑄(𝑜𝑜𝑐𝑐)
𝑠𝑠𝑜𝑜 1010 ≥𝐸𝐸(𝑜𝑜𝑐𝑐)≥ 𝐹𝐹𝑄𝑄(𝑜𝑜𝑐𝑐)
𝑠𝑠𝑜𝑜 𝐸𝐸(𝑜𝑜𝑐𝑐) ≥ 1010
(11)
TQ(fc) is the threshold of excitation, which is frequency dependent. G represents the low level
gain in the cochlear amplifier, relative to the gain at 500 Hz and above, and is frequency
dependent. The parameter A is used to bring the input-output function close to linear around
the absolute threshold. α is a compressive exponent which varies between 0.20 and 0.27. C is
a constant which scales the loudness to conform to the sone scale, where the loudness of 1
kHz tone at 40 dB SPL corresponds to 1 sone and C is equal to 0.047.
Loudness depends on the intensity and bandwidth of the sound, but among other
factors, also on its duration. Duration of signals is of relevance for human echolocation, and
we will therefore briefly describe models of duration of loudness. The effect of the duration
on the loudness was modeled by Glasberg and Moore [31] using three concepts for duration
of sounds, viz. Instantaneous loudness, Short Term Loudness and Long Term Loudness. They
depict the temporal integration of loudness in the auditory system and are described next.
Instantaneous loudness (IL). The specific loudness in each critical band has a
pattern and the specific loudness over all critical bands is called specific loudness pattern.
Usually the area under the specific loudness pattern is summed to give the instantaneous
loudness. If the sound is binaural, then the area under the specific loudness patterns at the two
ears are summed together to give the instantaneous loudness. The instantaneous loudness is an
intervening variable which is used for calculations and is not a perceptual variable.
Short Term Loudness (STL). The Short Term loudness is determined by averaging
the instantaneous loudness using an attack constant, αa = 0.045, and a decay constant, αr =
0.02 (Equation 12). The values of αa and αr were chosen so that the model will give
reasonable predictions for variations of loudness with duration and amplitude modulated
sounds (see Moore [44]).
𝑆𝑆𝑆𝑆𝑆𝑆 (𝑆𝑆)= �∝𝑠𝑠∗ 𝐼𝐼𝑆𝑆𝑠𝑠+(1− ∝𝑠𝑠)∗ 𝑆𝑆𝑆𝑆𝑆𝑆𝑠𝑠−1
∝𝑟𝑟∗ 𝐼𝐼𝑆𝑆𝑠𝑠+(1− ∝𝑟𝑟)∗ 𝑆𝑆𝑆𝑆𝑆𝑆𝑠𝑠−1 𝑠𝑠𝑜𝑜 𝐼𝐼𝐼𝐼(𝑠𝑠) ≥ 𝑆𝑆𝐹𝐹𝐼𝐼(𝑠𝑠−1)
𝑠𝑠𝑜𝑜 𝐼𝐼𝐼𝐼(𝑠𝑠) ≤ 𝑆𝑆𝐹𝐹𝐼𝐼(𝑠𝑠−1)
Long Term Loudness (LTL). The Long Term Loudness parameter was calculated
by averaging the instantaneous loudness using an attack constant, αa1 = 0.01 and a decay
constant, αr1 = 0.0005 (Equation 13). The values of αa1 and αr1 were chosen so that the model
may give reasonable predictions for the overall loudness of sounds that are amplitude
modulated at low rates (Moore [44]).
𝑆𝑆𝑆𝑆𝑆𝑆 (𝑆𝑆)= �∝𝑠𝑠1∗ 𝑆𝑆𝑆𝑆𝑆𝑆𝑠𝑠+(1− ∝𝑠𝑠1)∗ 𝑆𝑆𝑆𝑆𝑆𝑆𝑠𝑠−1
∝𝑟𝑟1∗ 𝑆𝑆𝑆𝑆𝑆𝑆𝑠𝑠+(1− ∝𝑟𝑟1)∗ 𝑆𝑆𝑆𝑆𝑆𝑆𝑠𝑠−1 𝑠𝑠𝑜𝑜 𝑆𝑆𝐹𝐹𝐼𝐼(𝑠𝑠) ≥ 𝐼𝐼𝐹𝐹𝐼𝐼(𝑠𝑠−1)
𝑠𝑠𝑜𝑜 𝑆𝑆𝐹𝐹𝐼𝐼(𝑠𝑠) ≤ 𝐼𝐼𝐹𝐹𝐼𝐼(𝑠𝑠−1)
As noted above, loudness is also affected by binaural hearing. To model binaural
loudness, a number of psychoacoustic facts have been considered (for details see Moore,
2014). Early results suggested that the level difference required for equal loudness of
monaurally and diotically presented sounds was 10 dB. The subjective loudness of a sound
doubles with about every 10 dB increase in physical intensity, and therefore it was assumed in
(13)
(12)
20
the early loudness model of Glasberg and Moore [31] that loudness sums across ears.
However, later results suggested that the level difference required for equal loudness is rather
between 5 to 6 dB. Glasberg and Moore therefore presented a new model to account for the
lower dB values based on the concept of inhibition. Inhibition occurs when a strong input in
one ear lowers or even stops, i.e. inhibits, the internal response evoked by a weaker input at
the other ear (Moore [44]).
Glasberg and Moore [31] implemented inhibition for binaural hearing by a gain
function. Initially, the specific loudness pattern was smoothed with a Gaussian weighting
function and the relative values of the smoothed function at the two ears were used to
compute the gain functions of the ears. The gains were then applied to the specific loudness
patterns at the two ears. The loudness for each ear was calculated by summing the specific
loudness over the center frequencies and the binaural loudness was obtained by summing the
loudness values across the two ears (Moore [44]). We used this procedure to calculate the
binaural loudness values in this report. The binaural loudness model of Glasberg and Moore
[31] has been implemented in PsySound3, a GUI-driven Matlab environment for analysis of
audio recordings (http://www.psysound.org).
Glasberg and Moore [31] assumed that the loudness of a brief sound is determined
by the maximum of the short term loudness, while the long term loudness may correspond to
the memory for the loudness of an event that can last for several seconds. For a time varying
sound (e.g. an amplitude modulated tone) it is appropriate to consider the long time loudness
as a function of time to calculate the time varying loudness. However, in this report, as the
stimuli presented to the participants were noise bursts and can be considered steady and brief,
we follow the assumption of Glasberg and Moore [31] of using the maximum of short time
loudness as a measure of the loudness of the recordings.
The means of the maxima values of Short Term Loudness in sones for the 10
versions for the 5, 50 and 500 ms recordings in the rooms of SN2010 [6] and SNG2016 [7]
are presented in Tables 3, 4 and 5, respectively. From these tables, one sees that the loudness
difference between the recordings without the object and with the object at 100 cm was less in
the case of the lecture room of SNG2016 [7] than for the anechoic or conference room in
SN2010 [6]. This may explain the low performance of the participants in the lecture room of
SNG2016 [7]. The loudness values follow the same pattern as the sound pressure level
analysis of the room acoustics chapter (Tables 2 and 5). However, the values in Tables 3 to 5
are psychophysical and depict not only the acoustics of the rooms but do also account for
aspects of human hearing that are important for human echolocation. A comparison of the
loudness results with the echolocation of persons will be made in the section "Threshold
values, absolute and difference, for echolocation in static situations based on auditory model
analysis", where we relate these psychophysical values to the echolocation performances.
Table 3. Means of the maxima of Short Term Loudness, in sones, of the 10 versions for
the recordings in the anechoic and conference room in SN2010 [6] and in the lecture
room in SNG2016 [7] with a 5 ms duration signal. The blank cells indicate that no
recordings were made at those distances. For the SN2010 there were two series of recordings
with no reflecting object.
Object distance (cm)
No Object, recording 1
No Object, recording 2
50
Schenkman and Nilsson [6]
Conference
Anechoic
room
13.4
13.3
20.7
room
19.3
19.4
26.7
Schenkman, Nilsson and Grbic [7]
Lecture room
15.5
21
100
150
200
300
400
500
20.2
24.4
17.2
16.2
14.4
13.3
13.4
13.4
21.5
19.7
20
19.5
45
45.1
69.6
55.7
47.6
45.1
45.2
45.0
Table 4. Means of the maxima of Short Term Loudness, in sones, of the 10 versions for
the recordings in the anechoic and conference room in SN2010 [6] with a 50 ms duration
signal. There were two series of recordings with no reflecting object.
Schenkman and Nilsson [6]
Object distance (cm)
Anechoic room
Conference room
No Object, recording 1
No Object, recording 2
40.1
40
63.7
52.3
50
100
150
200
40.3
300
40.3
400
40.2
500 40.1
Table 5. Means of the maxima of Short Term Loudness, in sones, of the 10 versions for
the recordings in the anechoic and conference room in SN2010 [6] and in the lecture
room in SNG2016 [7] with a 500 ms duration signal. The blank cells indicate that no
recordings were made at those distances. For the SN2010 there were two series of recordings
with no reflecting object.
Schenkman and Nilsson [6]
Anechoic
Conference
room
48.1
48.1
76.1
62.2
room
52.4
52.5
78.7
63.6
48.4
48.4
48.2
48.1
54.6
52.4
52.6
52.5
Schenkman, Nilsson and Grbic [7]
Lecture room
52
54.7
52.5
Object distance (cm)
No Object, recording 1
No Object, recording 2
50
100
150
200
300
400
500
22
Pitch analysis: autocorrelation with dual profiles
Repetition pitch is a percept that underlies human echolocation for detecting objects. It is
usually experienced as a coloration of the sound, perceived at a frequency equal to the inverse
of the delay time between the sound and its reflection (Bilsen [14]; Bilsen and Ritsma [25];
see also Bassett and Eastmond [22]). As mentioned in the ACF section of Method, Room
acoustics, in SN2010 and SNG2016 the reflecting objects were at distances of 50, 100, 150,
200, 300, 400 and 500 cm (although not all distances were used in both studies) resulting in
delays of 2.9, 5.8, 8.7, 11.6, 17.4, 23.2 and 29 ms, where the repetition pitches would
correspond to 344, 172, 114, 86, 57, 43 and 34 Hz, respectively. However, the actual delays
might vary because of factors like the recording set up, speed of sound etc. and therefore the
actual repetition pitch would be different. To test the presence of repetition pitch at these
frequencies together with how this information would be represented in the auditory system,
we used the PCP, BMM and NAP modules of the AIM, summarily presented above, to
analyze the recordings from SN2010 [6] and SNG2016 [7].
The perception of repetition pitch can be created by presenting iterated rippled noise
stimuli. The peaks in the autocorrelation function of these sounds are seen as the basis for
repetition pith (Yost [26]; Patterson et al [45]). Hence, instead of the strobe finding and the
temporal integration modules in AIM, we used the autocorr module as the final stage in our
analysis to quantify repetition pitch information. Analysis by autocorrelation provides a
feasible way to quantify repetition pitch, which we need to explain echolocation.
We chose not to use the strobe temporal integration as the final stage, but it does not
exclude that this might be how pitch information for echolocation is represented in the
auditory system. To determine whether it is autocorrelation or strobe temporal integration that
better explains repetition pitch perception and possibly also physiological processes involved
in the auditory system, further experiments and analysis are needed. For the interested reader,
we present as an example, some results obtained using the strobe temporal integration module
for a 500 ms signal, see S.2 Supporting Information.
After generating the ACF with the autocorr module in AIM, it has a dual profile
development module, which sums up the ACF along both the temporal and the spectral
domain. These features are relevant for human hearing in depicting how temporal and spectral
information might be represented and are useful for analyzing repetition pitch. We therefore
used this module to analyze the temporal and spectral results. The dual profile module plots in
a single plot both the temporal and the spectral sum on frequency axis. The temporal profile
and the spectral profile were scaled for this, and the inverse relation of time versus frequency
(f = 1/t) was used to plot both time and frequency on a frequency scale. As an example, the
dual plot for a 200 Hz pure tone is shown in Fig 9.
All the recordings were not analyzed in this way. The recordings with the object at
300 to 500 cm in study SN2010 [6] and with 2, 4, 8, 16, 32 and 64, 5 ms clicks in study
SNG2016 [7] do not provide any additional information for the module and were therefore
not included in this autocorrelation analysis.
23
Fig 9. The autocorrelation function of the autcorr module in the AIM for a 200Hz pure
tone, where the sum of the temporal axis (spectral profile, dashed line) and the
frequency axis (temporal profile, unbroken line) can be seen. The peak in the temporal
profile is at 200Hz. The peak in the spectral profile is approximately above 200Hz, this is due
the high pass asymmetric function used in the dynamic compressive gammachirp filter, whose
centre frequency shifts up as stimulus level increases.
The temporal profile (unbroken line in the figures below) was calculated by
summing the ACF output along 100 critical bands (50Hz to 8000Hz) at each time delay. The
spectral profile (dashed line in the figures below) was calculated by summing the ACF output
in each critical band along a 35 ms time delay. Therefore, the temporal profile consists of the
sum of 100 critical bands at every sample of 35 ms time delay, and the spectral profile
consists of the sum of 35 ms delay samples at every critical band.
In the two studies that we analyze, the recordings had been presented to the
participants with durations of 5, 50 or 500 ms plus an additional 450 ms of silence. Therefore,
we had the same presentation duration, i.e. the whole signal was analyzed. For example, a 5
ms recording had 5 ms duration plus 450 ms of silence. However, when presenting the figures
graphically, we use the first 70 ms time interval of the recordings.
The dual profiles are presented for the recordings for the three signal durations and
the two rooms in study SN2010 [6] and for the 5 ms and 500 ms signal in study SNG2016 [7].
It is important to note that the amplitude scale of the y-axis is different in each sub figure of
each figure. The investigated attribute here is pitch and each sub figure with reflecting object
should be compared with the sub figure with No object for each condition. A distinct peak in
a sub figure which is absent in the sub figure with No object indicates the potential occurrence
of the perception of a pitch. One should remember that the visual impression of a peak in a
sub figure with No object may misleadingly indicate an auditory peak, unless one observes
the different scales on the different y-axis for the different sub figures. The next section will
deal with how to select peaks based on their peak strength.
As mentioned above, the theoretical frequency of the repetition pitch for recordings
with object at 100, 150 and 200 cm is 172, 114 and 86 Hz. The analysis of the 5 ms
recordings show peaks approximately at these frequencies. For example, Fig 10c, 11c, 12b
had peaks (marked by arrows in the subfigures) approximately at 172 Hz; Fig10d, 11d had
peaks (marked by arrows in the subfigures) approximately at 82 Hz; Fig12c had a peak
(marked by an arrow in the subfigure) approximately at 114 Hz. One reason that the peaks
were not exactly at the theoretical values is probably due to the experimental setups of
SN2016 and SNG2017 and of the room acoustics. Fig10b, 11b had no peaks at their
24
corresponding theoretical frequencies. However, this is due to a wider range in the y axis
scale (cf Fig 10, Fig11 y axis labels). The spectral profiles on the other hand did not have
peaks any closer to the theoretical frequencies (cf Fig10-12 dashed lines). There were small
spectral differences but these may provide timbre information but not pitch information.
In the 50 ms and 500 ms signal recordings, distinct peaks that could account for pitch
perception were absent in the spectral profiles (Figs 13 to 17). We conclude that the spectral
profiles (dashed line) did not provide information for pitch perception. The temporal profiles
in Figs 13 to 17 might have some peaks approximately at the theoretical frequencies of the
repetition pitch, but they were not clearly visible in these figures due to the scaling of the
figures. Therefore, it is too premature to conclude that the temporal profile (unbroken line) is
necessary for detecting the objects, if based on repetition pitch. A further analysis was
therefore needed which quantified the peaks in the temporal profile.
To determine the role of temporal information for detecting objects based on
repetition pitch, the pitch strength development module of AIM was used. It measures the
pitch perceived based on the peak strength. We elaborate this in the next section. The
temporal profiles will be shown to have peaks at the theoretical frequencies of repetition pitch
which, we believe, explains the perception of repetition pitch and thus also a major cause for
detection by echolocation of the reflecting objects in the two studies.
Fig 10. The dual profile of a 5 ms signal recorded in the anechoic room (in SN2010 [6]).
The solid line is the sum of the ACF along the spectral axis and the dashed line is the sum of
the ACF along the time delay axis for 70 ms time interval. The temporal and spectral profiles
are scaled to be compared to each other. The 'x' axis is changed in the temporal profile by
using the inverse relationship of time to frequency, f=1/t. The sub figure A is for the no object
recording and sub figures B, C and D are for the recordings with object at 50, 100 and 200 cm
respectively.
25
Fig 11. The dual profile of a 5 ms signal recorded in the conference room (in SN2010
[6]). The solid line is the sum of the ACF along the spectral axis and the dashed line is the
sum of the ACF along the time delay axis for 70 ms time interval. The temporal and spectral
profiles are scaled to be compared to each other. The 'x' axis is changed in the temporal
profile by using the inverse relationship of time to frequency, f=1/t. The sub figure A is for
the no object recording and sub figures B, C and D are for the recordings with object at 50,
100 and 200 cm respectively.
Fig 12. The dual profile of a 5ms signal recorded in the lecture room (SNG2016 [7]). The
solid line is the sum of the ACF along the spectral axis and the dashed line is the sum of the
ACF along the time delay axis for 70 ms time interval. The temporal and spectral profiles are
scaled to be compared to each other. The 'x' axis is changed in the temporal profile by using
26
the inverse relationship of time to frequency, f=1/t. The sub figure A is for the no object
recording and sub figures B and C are for the recordings with object at 100 and150 cm
respectively.
Fig 13. The dual profile of a 50 ms signal recorded in the anechoic room (SN2010 [6]).
The solid line is the sum of the ACF along the spectral axis and the dashed line is the sum of
the ACF along the time delay axis for 70 ms time interval. The temporal and spectral profiles
are scaled to be compared to each other. The 'x' axis is changed in the temporal profile by
using the inverse relationship of time to frequency, f=1/t. The sub figure A is for the no object
recording and sub figures B, C and D are for the recordings with object at 50, 100 and 200 cm
respectively.
27
Fig 14. The dual profile of a 50 ms signal recorded in the conference room (SN2010 [6]).
The solid line is the sum of the ACF along the spectral axis and the dashed line is the sum of
the ACF along the time delay axis for 70 ms time interval. The temporal and spectral profiles
are scaled to be compared to each other. The 'x' axis is changed in the temporal profile by
using the inverse relationship of time to frequency, f=1/t. The sub figure A is for the no object
recording and sub figures B, C and D are for the recordings with object at 50, 100 and 200 cm
respectively.
Fig 15. The dual profile of a 500 ms signal recorded in the anechoic room (SN2010 [6]).
The solid line is the sum of the ACF along the spectral axis and the dashed line is the sum of
the ACF along the time delay axis for 70 ms time interval. The temporal and spectral profiles
are scaled to be compared to each other. The 'x' axis is changed in the temporal profile by
using the inverse relationship of time to frequency, f=1/t. The sub figure A is for the no object
recording and sub figures B, C and D are for the recordings with object at 50, 100 and 200 cm
respectively.
28
Fig 16. The dual profile of a 500 ms signal recorded in the conference room (SN2010
[6]). The solid line is the sum of the ACF along the spectral axis and the dashed line is the
sum of the ACF along the time delay axis for 70 ms time interval. The temporal and spectral
profiles are scaled to be compared to each other. The 'x' axis is changed in the temporal
profile by using the inverse relationship of time to frequency, f=1/t. The sub figure A is for
the no object recording and sub figures B, C and D are for the recordings with object at 50,
100 and 200 cm respectively.
Fig 17. The dual profile of a 500 ms signal recorded in the lecture room (SNG2016 [6]).
The solid line is the sum of the ACF along the spectral axis and the dashed line is the sum of
the ACF along the time delay axis for 70 ms time interval. The temporal and spectral profiles
29
are scaled to be compared to each other. The 'x' axis is changed in the temporal profile by
using the inverse relationship of time to frequency, f=1/t. The sub figure A is for the no object
recording and sub figures B and C are for the recordings with object at 100 and 150 cm
respectively.
Pitch strength
The peaks in the temporal profile of the autocorrelation function that we computed with the
dual profile module of AIM were distributed without apparent order or meaning. It is not
obvious which peak that corresponds to a pitch. There is a solution: The AIM model has a
pitch strength module which calculates the pitch strength to determine if a particular peak is
random or not. This module first calculates the local maxima and their corresponding local
minima. The ratio of peak height to the peak width of the peak (local maxima) is subtracted
from the mean of the peak height between two adjacent local minima to obtain the pitch
strength (PS) of a particular peak.
Two modifications were made by us in the pitch strength algorithm of AIM to
improve its performance for the analysis. (1) The low pass filtering was removed as it
smooths out the peaks and, (2) the pitch strength was measured with Equation (14). Figure 18
illustrates the pitch strength algorithm that was used. The peak with the greatest peak height
has the greatest pitch strength and would be the perceived frequency of repetition pitch.
PS = PH− PHLM�������� (14)
where PS is the calculated pitch strength, PH is the height of the peak and PHLM is the
mean of the peak height between two adjacent local minima.
Fig 18. An illustration of the pitch strength measure computed using the pitch strength
module of the AIM. The dot indicates the local maxima and the two stars are the
corresponding local minima. The vertical dashed line is the pitch strength calculated using
Equation (14). The frequency in Hz was computed by inverting the time delay, f = 1/t.
The results of the calculated pitch strength for recordings of studies SN2010 [6] and
SNG2016 [7] are presented in Tables 6 to 8. As can be seen, peaks were misleadingly
identified for recordings without an object, which would not have caused a pitch perception.
This happens because the pitch strength algorithm identifies all local maxima and minima in a
30
sound and thus also calculates the pitch strength for all random peaks (that have local
maxima).
Table 6. Mean of the pitch strength (autocorrelation index) of the 10 versions for
recordings in SN2010 [6] and in SNG2016 [7] with the 5 ms duration signal. The blank
cells indicate that no recordings were made at those distances.
Schenkman and Nilsson [6]
Object distance (cm)
No Object, recording 1
No Object, recording 2
50
100
150
200
Anechoic room
Conference room
0.37
0.40
2.54
1.06
-
0.78
0.80
9.65
2.43
-
0.35
0.85
Schenkman, Nilsson
and Grbic [7]
Lecture room
0.19
-
-
0.29
0.28
-
Table 7. Mean of the pitch strength (autocorrelation index) of the 10 versions for
recordings in SN2010 [6] with the 50 ms duration signal.
Schenkman and Nilsson [6]
Object distance (cm)
Anechoic room
Conference room
No Object, recording 1
No Object, recording 2
50
100
150
0.47
0.44
4.17
1.67
0.42
0.55
0.52
6.59
2.22
0.54
Table 8. Mean of the pitch strength (autocorrelation index) of the 10 versions for
recordings in SN2010 [6] and in SNG2016 [7] with the 500 ms duration signal. The blank
cells indicate that no recordings were made at those distances.
Schenkman and Nilsson [6]
Schenkman,
Nilsson and Grbic
[7]
Object distance (cm)
No Object, recording 1
No Object, recording 2
50
100
150
200
Anechoic room
Conference room
Lecture room
0.71
0.78
4.75
2.44
-
0.70
0.84
0.90
7.74
2.91
-
1.35
1.30
-
-
1.36
1.42
-
The unit for pitch strength in our analysis is the autocorrelation index, as it is
computed on the autocorrelation function. The tabulated data show that for the 5 ms and 50
31
ms duration signals the pitch strength was greater than 1 for the object distances of 50 and 100
cm in the anechoic and conference rooms of SN2010 [6] (Tables 6 and 7). For the 500 ms
duration signal, the strength was greater than 1 at distances of 50 and 100 cm in the anechoic
room and at the distances of 50, 100 and 200 cm in the conference room. The lecture room in
SNG2016 [7] also had a pitch strength greater than 1 at this condition, but the computed pitch
strength was not consistent over a single frequency and it lasted for only 4 to 8 time frames.
(The time frames had 35ms time delay computed from a 70 ms interval NAP signal. Each
frame had a hop time of 10ms). This was not the case for the anechoic and conference rooms
in SN2010 [6] that both had high pitch strengths at a particular frequency and lasted for 14 to
18 time frames of 35 ms intervals, each with a hop time of 10 ms. Additionally, in the lecture
room in SNG2016 [7] with a reflecting object present, the pitch strength was not much
different from when there was no object. This further illustrates the echolocating difficulties
of the test persons in that study.
The perceptual results of SN2010 [6] showed that the participants were able to detect
the objects with a high percentage correct at object distances 50 and 100 cm in the anechoic
room and at 50, 100 and 200 cm in the conference room (Schenkman and Nilsson [6]). As
presented in the previous paragraph, the pitch strength was greater than 1 at these conditions.
Pitch seems to be the most important information that listeners use to detect objects at these
distances (see e.g. Schenkman and Nilsson [9]). Therefore, these results imply that there
might be a perceptual threshold approximately equal to 1 (autocorrelation index) for pitch
strength in echolocating situations. The peak with that pitch strength must exist for certain
time frames for a person to perceive a repetition pitch. This is also dependent on the acoustics
of the room. Relating the pitch strength results with the performance of the participants in the
two studies SN2010 [6] and SN2016 [7] will be made in a following section on threshold
values. Before this we shall analyze the results in terms of a timbre property, namely
sharpness.
Sharpness analysis
In the Method part for analysis of room acoustics we described how the spectral centroid was
used as a measure for timbre perception. The spectral centroid was computed on the time
varying Fourier Transform. To address more specifically how human hearing perceives
timbre, Fastl and Zwicker [33] computed the weighted centroid of the specific loudness rather
than of the Fourier Transform. This psychophysical measure is called sharpness and indicates
how sound extends from being perceived to vary from dull to sharp. We conducted the
sharpness analysis for our recordings using code available from Psysound (Cabrera, Ferguson,
and Schubert [35]). Sharpness varies over time and therefore the median was used as a
representative measure for the perceived sharpness. The results of the mean of the medians of
the perceived sharpness over the 10 versions in the anechoic, conference and lecture room in
SN2010 [6] and SNG2016 [7] for 5, 50 and 500 ms duration signals are presented in Tables
11 to 14. The unit for sharpness is called acum.
There are, to our knowledge, only a few studies on thresholds of sharpness.
Pedrielli, Carletti, and Casazza [46] found that their participants had a just noticeable
difference for sharpness of 0.04 acum. You and Jeon [47] found in a study on refrigerator
noise that their participants had a just noticeable difference for sharpness of 0.08 acum.
Assuming that 0.04 acum is a threshold value for sharpness, then the results in Tables 9 to 11
show that the difference in median sharpness in SN2010 [6] was greater than threshold for the
object at 50 and 100 cm compared to the recordings without the object. In SNG2016 [7], the
differences between the recordings with and without object were smaller than in SN2010 [6],
although greater than 0.04 acum. It is possible that at shorter distances (say less than 200 cm)
Object distance (cm)
No Object, recording 1
No Object, recording 2
50
100
150
200
300
400
500
1.89
1.9
2.05
2.14
1.92
1.91
1.89
1.89
1.97
1.98
2.03
2.03
2
2.01
1.98
1.99
1.85
1.78
1.83
32
repetition pitch and loudness information might be more relevant for providing echolocation
information than sharpness information.
In study SN2010 [6] with reflecting object at distances 200, 300, 400 and 500 cm for
5 ms (anechoic and conference rooms), 50ms (anechoic and conference rooms) and 500ms
signal (conference room) durations, the recordings had differences in mean of median
sharpness of less than 0.04 acum when compared to the recordings without an object.
However, in the anechoic room of SN2010 for the 500 ms signal duration [6], the recordings
with object at 400 cm and 500 cm had differences in sharpness approximately greater than
0.04 acum when compared to the recordings without the object (Tables 9 to 11). This is
information that blind persons might use to detect and identify a reflecting object at longer
distances than, say at 200 cm. We discuss this issue further in the next section.
Table 9. Mean of the 10 versions of mean of median of the sharpness (acum) for the
recordings in anechoic, conference and the lecture room with the 5 ms duration signal.
Schenkman and Nilsson [6]
Schenkman,
Nilsson and
Grbic [7]
Anechoic room
Conference room
Lecture room
Table 10. Mean of the 10 versions of mean of median of the sharpness (acum) for the
recordings in anechoic, conference and the lecture room with the 50 ms duration signal.
Object distance (cm)
No Object, recording 1
No Object, recording 2
50
100
150
200
300
400
500
Schenkman and Nilsson [6]
Anechoic room
Conference room
1.89
1.9
2.07
2.14
1.91
1.9
1.87
1.88
1.89
1.89
1.96
1.95
1.94
1.91
1.92
1.89
33
Table 11. Mean of the 10 versions of mean of median of the sharpness (acum) for the
recordings in anechoic, conference and the lecture room with the 500 ms duration
signal.
Schenkman and Nilsson [6]
Conference
Anechoic
room
1.86
1.88
2.12
2.12
room
1.94
1.94
2.1
2.04
Schenkman, Nilsson
and Grbic [7]
Lecture room
2.07
2.2
2.11
Object distance (cm)
No Object, recording 1
No Object, recording 2
50
100
150
200
300
400
500
1.89
1.86
1.83
1.84
1.97
1.95
1.95
1.94
Threshold values, absolute and difference, for echolocation in
static situations, based on auditory model analysis
Background
The stimuli in the studies [6, 7] were presented to the participants in a two-alternative-forced-
choice (2AFC) manner. The participants had to compare two sounds and detect the sound
with the echo. This perceptual decision is based on information in attributes of the stimuli.
If a person can detect an attribute of an object with a 75 percentage of correct
response, this limit is usually termed the absolute threshold. The judgments are commonly
based on only one source of information. In the experiments in [6, 7] two sources of
information were used. The sounds were presented in a 2AFC manner, where the participants
compared the information of two sounds, with and without object, and made the decision.
This is a difference threshold (Gescheider [48]), which is the threshold at which the
participant can make a discrimination with about 75 percentage of correct response.
As regards echolocation, we define the absolute threshold as the value of a
perceptual attribute (e.g. loudness, pitch or sharpness) of a recording with a reflecting object
at which a person has a 75 percentage correct response. The difference threshold is thus the
difference of responses between the recordings with and without object, respectively, at which
the person has 75 percentage of correct response. The experimental procedure used in the
experiments was 2AFC, and therefore the difference threshold is the relevant measure of the
echolocation ability of the persons. However, as the difference threshold is calculated from
absolute thresholds both are for clarity shown here. The procedure for finding the difference
thresholds and the corresponding results are presented below.
Non-parametric versus parametric modeling of psychometric function for
threshold analysis
Psychometric functions relate perceptual results to physical parameters of a stimulus.
Commonly the psychometric function is estimated by parametric fitting, i.e. it is assumed that
the underlying relationship can be described by a specific parametric model. The parameters
of such a model are then estimated by maximizing the likelihood. However, the most correct
parametric model underlying the description of the psychometric function is unknown.
34
Therefore, estimating the psychometric function based on the assumptions of a parametric
model may lead to incorrect interpretations (Zychaluk and Foster [49]). To address this
problem, Zychaluk and Foster [49] implemented a non-parametric model to estimate the
psychometric function that does not require assumptions about the empirical state of the
underlying phenomenon. The psychometric function is thus modeled locally without
assuming a "true" underlying function. Since the true relationship for the variables that
determine human echolocation is unknown, we chose the method proposed by Zychaluk and
Foster [49] in our analysis of the perceptual data. Next is a brief description of the non-
parametric model for estimating the underlying psychometric function. Thereafter our
analysis of the perceptual results is presented.
A generalized linear model (GLM) is usually used when fitting a psychometric
function. It consists of three components: a random component from the exponential family, a
systematic component η and a monotonic differentiable link function g, that relates the two.
Hence, a psychometric function, P(x), can be modeled by using Equation (15). The
parameters of the GLM are estimated by maximizing the appropriate likelihood function
(Zychaluk and Foster [49]). The efficiency of the GLM relies on how much the chosen link
function, g , approximates the "true" underlying function.
(15)
𝜂𝜂(𝑥𝑥)=𝑙𝑙[𝑆𝑆(𝑥𝑥)]
However, as mentioned, the true function one can never know for certain. In non-
parametric modelling, instead of fitting the link function g, the function η is fitted using a
local linear method. For a given point, x, the value η(u) at any point u in a neighborhood of x
is approximated by Equation (16) (Zychaluk and Foster [49]).
𝜂𝜂(𝑢𝑢) ≈ 𝜂𝜂(𝑢𝑢)−(𝑢𝑢−𝑥𝑥)𝜂𝜂′(𝑥𝑥) (16)
where η'(x) is the first derivative of η. The actual estimate of the value of η(x) is obtained by
fitting this approximation to the data over the prescribed neighborhood of x. Two features are
important for this purpose, kernel K and the bandwidth h. A Gaussian kernel is preferred, as it
has unbounded support and is best for widely spaced levels. An optimal bandwidth can be
chosen using plugin, bootstrap or cross validation methods (Zychaluk and Foster [49]). As no
method is guaranteed to always work, to find the optimal bandwidth for the analysis we chose
a bootstrap method with 30 replications. When the bootstrap method failed to find an optimal
bandwidth, cross validation was used to establish the optimal bandwidth.
Distance thresholds for object detection
The psychometric function was initially fitted to the mean proportion of correct responses
with respect to the distance to the reflecting object. Figures 19, 20 and 21 show both the non-
parametric modeling (local linear fit) and the parametric modeling of the perceptual results for
the blind participants in study SN2010 [6]. The mean percentage correct is plotted as a
function of distance for recordings with 5, 50 and 500 ms signals in the anechoic and
conference rooms. The link function used for the parametric modeling was the Weibull
function. Visual inspection shows that this link function was not appropriate, since the fit does
not correspond well with the perceptual results. As mentioned, if one knows the underlying
link function for the psychophysical data, then the parametric fit is a better fit than the local
non-linear fit, but for the data we are analyzing one cannot assume a particular link function.
However, the local linear fit correlates well with the perceptual results. This demonstrates
35
some of the advantages of using non-parametric modeling for the purposes of the present
investigation.
We used the means of the proportion of correct responses of the participants for the
psychometric fitting. If we had used the individual responses, the individual thresholds would
vary but the local linear fit would probably still, we believe, be well correlated with the
perceptual results. Therefore, the results in the remaining part of the present section will be
based on the psychometric function using local linear fit for the mean proportion correct
answers. We used the implementation of the non-parametric model fitting in Matlab by
Zychaluk and Foster [49].
The local linear fit needs at least three stimulus values to make a mathematically
valid fit. As the recordings in the lecture room (in SNG2016 [7]) had only two stimulus
values, at 100 and 150 cm, it was not possible to make a psychometric fit for these recordings.
Fig 19. The parametric (Weibull fit) and non-parametric (Local linear fit) modeling of
the mean proportion of correct responses for the 5 ms recordings of the blind
participants in SN2010 [6] as a function of distance: (A) anechoic chamber. (B)
conference room.
Fig 20. The parametric (Weibull fit) and non-parametric (Local linear fit) modeling of
the mean proportion of correct responses for the 50 ms recordings of the blind
participants in SN2010 [6] as a function of distance: (A) anechoic chamber. (B)
conference room.
36
Fig 21. The parametric (Weibull fit) and non-parametric (Local linear fit) modeling of
the mean proportion of correct responses for the 500 ms recordings of the blind
participants in SN2010 [6] as a function of distance: (A) anechoic chamber. (B)
conference room.
Distance perception of an object is not a perceptual attribute that was presented
directly to the participants in studies SN2010 [6] and SNG2016 [7]. Therefore, the distance
threshold obtained from the psychometric fit is a derived quantitative threshold. The distance
threshold is the distance at which a person may detect an object with a probability of 75%. As
the fitted psychometric function is discrete, it was not always possible for the fit to have an
exact value of 0.75. Therefore, the threshold values at the proportion of correct responses
within the range of 0.73 to 0.75 were initially calculated and the mean of the threshold values
was determined as the actual threshold.
The distance threshold in study SN2010 [6] for which the blind and the sighted could
detect the object using echolocation, with a proportion of correct responses between 0.73 and
0.75, are shown in Table 12. The distances at which the blind participants could detect the
reflecting object were farther away than for the sighted in both rooms and for all three sound
signals. The threshold is positively related to the signal duration for both groups, i.e. the
longer durations give a longer range of detection. We can also see that the blind persons could
detect objects farther away in the conference room for all signals, but for the sighted this was
only the case for the 500 ms signal. In the original study SN2010 [6, table 3], the calculations
of at what distance a blind or a sighted person might detect a reflecting object were based on a
parametric approach, yielding in general lower distance values for thresholds than our non-
parametric approach. In other words, our reanalysis of the data show a higher sensitivity of
both blind and sighted, than the values in [6].
Table 12. Distance thresholds (cm) for duration, room, and listener groups in study
SN2010 [6]. The threshold values were calculated from the psychometric function of the blind
and sighted participants' responses at the mean percentage of correct responses values of 0.73
to 0.75.
Sound duration (ms)
5
Sighted
130
121
Blind
150
158
50
500
Blind
166
176
Sighted
160
147
Blind
172
247
Sighted
166
207
Room
Anechoic
Conference
37
Loudness thresholds, absolute and difference, for object detection
Loudness is also a source of information for detecting reflecting objects by echolocation
(Schenkman and Nilsson [9]; Kolarik et al, 2014 [2]). A common psycho-acoustical measure
to express loudness is in the unit of sones (Yost [50]; Gescheider [48]). Therefore, we used
values in sones for the local linear fitting to determine the absolute and difference threshold of
loudness, where blind and sighted could detect a reflecting object. As elsewhere, the criterion
was detection with a percentage correct between 73 and 75%. The mean loudness values, and
the mean percentage of correct responses calculated from study SN2010 [6] were used as
inputs for the psychometric fit. The resulting absolute threshold values of loudness for
detecting the object are presented in Table 13. The data show that the thresholds for loudness
for the blind participants were lower compared to those of the sighted, roughly 1 sone less in
the anechoic chamber and 2 sones less in the conference room.
Table 13. Absolute threshold values of loudness (sones) for duration, room, and listener
groups in study SN2010 [6]. The threshold values were calculated from the psychometric
function of the blind and sighted participants' responses at the mean percentage of correct
responses value of 0.73 to 0.75.
Room
Anechoic
Conference
Sound duration (ms)
5
Sighted
17.5
24.1
Blind
16.8
22.6
50
500
Blind
43.7
49.4
Sighted
45.1
53.1
Blind
52.9
53.6
Sighted
53.2
55.3
The values in Table 13 may misleadingly lead a reader to infer that the shortest
sounds had the lowest threshold. This is not the case. If we look at tables 3 and 5 the
recording without object for the 5 ms signal had a loudness of approximately 13 sones and the
500 ms signal had approximately 45 sones. Considering these values, it is more appropriate to
use the difference rather than the absolute threshold, since the detection based on loudness in
SN2010 [6], as a consequence of the 2AFC method that had been used, was based on a
relative judgment, a comparison, and not on an absolute judgment. The difference threshold
values were calculated by subtracting the absolute threshold values in Table 13 with their
corresponding loudness values for the recording without object, see Table 14.
Table 14. Difference threshold values of loudness (sones) for duration, room, and
listener groups in study SN2010 [6]. The threshold values were calculated from the
psychometric function of the blind and sighted participants' responses at the mean percentage
of correct responses value of 0.73 to 0.75.
Room
Anechoic
Conference
Sound duration (ms)
5
Sighted
4.1
4.7
Blind
3.4
3.2
50
Sighted
5
8
Blind
3.6
4.3
500
Blind
4.8
1.1
Sighted
5.1
2.8
The data in Table 14 give a different and more realistic picture than the data in Table
13. The blind detect objects at lower loudness values for all sounds, and both groups could
detect with lower relative loudness levels (loudness difference between the recording with and
without object) for 500ms duration signals in the conference room. The loudness model used
to compute the mean loudness was the same for both test groups, and therefore we conclude
38
that the low thresholds of the blind persons are an effect of their perceptual ability. This
conclusion is further discussed in the Discussion part of this report.
Pitch thresholds, absolute and difference, for object detection
The absolute and difference threshold values of pitch strength, as calculated by the
autocorrelation index for which the blind and the sighted test persons in study SN2010 could
detect the reflecting object, are presented in Tables 15 and 16, respectively. We will first
discuss the results in terms of the absolute thresholds, but the more appropriate conclusions
will be based on the difference thresholds.
The absolute threshold varies for blind and sighted persons depending on signal
durations and room conditions. The blind, for all conditions, had lower thresholds, the pitch
strength increased with signal duration and the thresholds were lower in the anechoic room. It
is possible that for shorter duration signals, the person may be inattentive and miss the signal
and thus also the pitch information. The performance (percentage of correct response) of the
participants with 5 and 50 ms signals may thus not only be based on pitch strength but also on
cognitive factors such as attention.
Schenkman and Nilsson [9] showed that when pitch and loudness information were
presented together, at distances up to 200 cm to the object, the participants' performance was
almost 100 percentage correct. The 500 ms recordings with the object at 50 and 100 cm in
study SN2010 [6] had almost 100 percent correct response for both the blind and the sighted.
Therefore, for the 500 ms signal condition it is likely that the likelihood to miss a signal and
its pitch information because of non-attention is lower, and the perceptual results of the
participants are likely to be based mostly on pitch information.
There are two possible theoretical ways to regard how the hearing system treats the
ACF values. We will here focus the analysis on the 500 ms signal, since as noted above, the 5
ms and 50 ms signals may have cognitive aspects that could bias the auditory model analysis.
(1) Based on the above reasoning, and if we assume that the auditory system analyses the
pitch information absolutely i.e. it does not compare the peak heights in the ACF between the
recordings (when presented in a 2AFC manner), then the results indicate that the absolute
threshold for detecting a pitch based on an autocorrelation process should be greater than 1.10
and 1.23 (as indicated by the autocorrelation index for the 500 ms signal) for the blind and the
sighted, respectively, as shown in Table 15. (2) On the other hand, if we assume that the
auditory system analyses the pitch information relatively i.e. it compares the peak heights in
the ACF between the recordings (when presented in a 2AFC manner) then the results indicate
that the difference threshold for detecting the pitch based on autocorrelation should be greater
than 0.27 and 0.49 (autocorrelation index) for the blind and the sighted, respectively, as
shown in Table 16. For all cases, the values show that the blind persons could detect echo
reflections of objects having lower peak heights in the ACF, than the sighted could.
Table 15. Absolute threshold values of pitch strength (autocorrelation index) for
duration, room, and listener groups in study SN2010 [6]. The threshold values were
calculated from the psychometric function of the blind and sighted participants' responses at
the mean percentage of correct response values of 0.73 to 0.75.
Sound duration (ms)
5
Sighted
0.88
2.21
Blind
0.77
1.54
50
500
Blind
0.8
1.07
Sighted
0.96
1.69
Blind
1.1
1.14
Sighted
1.23
1.41
Room
Anechoic
Conference
39
Table 16. Difference threshold values of pitch strength (autocorrelation index) for
duration, room, and listener groups in study SN2010. The threshold values were
calculated from the psychometric function of the blind and sighted participants' responses at
the mean percentage of correct response values of 0.73 to 0.75.
Sound duration (ms)
Room
Anechoic
Conference
Blind
0.39
0.75
5
Sighted
0.50
1.42
50
500
Blind
0.35
0.54
Sighted
0.51
1.16
Blind
0.36
0.27
Sighted
0.49
0.54
Sharpness thresholds, absolute and difference, for object detection
As discussed previously, timbre indicates how we experience various qualities of sounds. We
chose to study one aspect of timbre, sharpness, as a potential information source for object
detection by echolocation. In analog to the previous psycho-acoustical parameters, we
calculated the absolute and difference threshold values of sharpness for which the blind and
the sighted test persons in study SN2010 [6] could detect a reflecting object using
echolocation with a correct response value of 0.73 to 0.75.
For quantitative values for sharpness, we used the psychophysical unit acum (e.g.
Fastl and Zwicker [33]). Tables 17 and 18 show that for the blind and sighted participants
both their absolute and difference thresholds, the sharpness values were about the same.
However, unlike loudness and pitch strength the sharpness information need not be greater in
value for the participants to detect the object. For sharpness, a listener must distinguish
between timbres. This may include cognition, involving e.g. memory processes.
When a participant in SN2010 [6] and SNG2016 [7] were presented with two stimuli
in a 2AFC method, they distinguished the recording with the object from the recording
without the object by identifying the one with the higher loudness level, stronger pitch
strength or both. However, when a person uses sharpness for echolocation, e.g. in a 2AFC
method, it is not necessary that the recording with the reflecting object has the higher
sharpness value. The recording with the reflecting object might sound duller, i.e. having a
lower value of sharpness, than the recording without the object. A person might use this
information to detect or identify an object.
Table 17. Absolute threshold values of the mean of the median sharpness (in the unit
acum) for duration, room, and listener groups in study SN2010 [6]. The threshold values
were calculated from the psychometric function of the blind and sighted participants'
responses at the mean percentage of correct response values of 0.73 to 0.75.
Room
Anechoic
Conference
Blind
1.97
2.01
Sound duration(ms)
5
Sighted
1.98
2.03
50
500
Blind
1.96
1.94
Sighted
1.98
1.94
Blind
1.94
1.97
Sighted
1.96
1.97
Table 18. Difference threshold values of the mean of the median sharpness (in the unit
acum) for duration, room, and listener groups in study SN2010 [6]. The threshold values
were calculated from the psychometric function of the blind and sighted participants'
responses at the mean percentage of correct response values of 0.73 to 0.75.
Sound duration(ms)
5
50
500
40
Room
Anechoic
Conference
Blind
0.07
0.03
Sighted
0.08
0.05
Blind
0.06
0.05
Sighted
0.08
0.05
Blind
0.07
0.03
Sighted
0.09
0.03
Table 11 showed that the sharpness values for the 500 ms duration recordings, with
reflecting object at 400 and 500 cm in the anechoic room, had smaller sharpness values than
the recording without an object. Interestingly, two blind participants (no. 2 and 6) performed
better at these conditions than all the remaining participants, i.e. their proportion correct were
approximately 0.7, even at 400 and 500 cm. We looked deeper into the performance of these
two high-performing echolocators by making a local linear fit for the proportion correct of
these two persons and the sharpness values of the 500ms recordings in the anechoic chamber
that were calculated and are shown in Table 11. Cross validation was used to find the
bandwidth of the local linear fit kernel. Figs 22 and 23 show the corresponding local linear
fits.
One can see clearly that when the proportion correct was approximately equal to 0.7,
there were two absolute threshold values for sharpness, one higher and one lower. If we
consider the mean of the sharpness of the two no object recordings at this condition (i.e.
anechoic room and 500ms signal duration) it was about 1.87 (Table 11). Hence the difference
threshold for the blind participant 2 would be 1.94 -1.87 = 0.07 acum and 1.83-1.87 = -0.04
acum. Similarly, the difference threshold for the blind participant 6 would be 1.97-1.87 = 0.10
acum and 1.83-1.87 = -0.04 acum. Perceptually, this means that the two high-performing
blind participants could detect the object even when the recording with the object was duller
than the recording without an object. A more detailed discussion on the importance of
sharpness information for human echolocation is presented in the Discussion part below.
Fig 22. The non-parametric (local linear fit) modeling of the proportion of correct
response for the 500 ms recordings of the 2nd blind participant in SN2010 [6] as a
function of the mean of median sharpness. As the function is discrete, the values of
proportion correct responses approximately equal to 0.7 are considered.
41
Fig 23. The non-parametric (local linear fit) modeling of the proportion of correct
response for the 500 ms recordings of the 6th blind participant in SN2010 [6] as a
function of the mean of median sharpness. As the function is discrete, the values of
proportion correct responses approximately equal to 0.7 are considered.
Discussion
We wanted to elucidate factors that determine echolocation and the differences of
echolocation ability between blind and sighted persons. Neuroimaging and psychoacoustic
methods give us insight into the high echolocating ability of blind persons, but these methods
do not necessarily reveal the information in the acoustic stimulus that determines echolocation
(at least when the information is not known) and also not fully how this information is
represented and processed in the human auditory system. The implementation of auditory
models for human echolocation was intended to establish the information that determines
echolocation ability and its variability, and how this information is represented and processed
in the human auditory system.
Signal analysis was conducted on the physical signals and is presented in the Method
part on Room acoustics. The aim was to find the physical information that is used for
echolocation and to analyze the effects of room acoustics on human echolocation. We studied
in particular the sound pressure levels, auto correlations and spectral centroids. The analyses
were performed on the recordings of studies SN2010 [6] and SNG2016 [7]. The results
demonstrate, as expected, that the acoustics of the room affect the sounds and thereby the
physical attributes that are associated with it. However, the information represented in the
auditory system is complex and it uses this physical information for further processing
through the auditory neural system. To understand better what takes place in the auditory
system of a person using echolocation, we used what we consider are the most relevant
auditory models for human echolocation, that today are available in the literature. We thus
studied how the corresponding perceptual attributes of sound pressure level, auto correlation
and spectral centroid are encoded in the human auditory system.
The results of the auditory models suggest that repetition pitch, loudness and
sharpness all provide potential information for people to echolocate at distances below 200
cm. The results also indicate that at longer distances sharpness information may be used for
human echolocation. A detailed discussion of how loudness, pitch and sharpness are essential
for human echolocation and how they might be represented in the auditory system is
42
presented below in sections "Echolocation and loudness", "Echolocation and pitch" and
"Echolocation and timbre". A discussion of how room acoustics and binaural information as
well as movement affect human echolocation is presented in the section "Echolocation and
room acoustics" and "Echolocation, binaural information and movement". We finally
conclude by some comments on using auditory models for understanding human echolocation
together with some theoretical implications.
Echolocation and loudness
Of the existing loudness models, we chose the one by Glasberg and Moore [31], since it has a
good fit to the equal loudness contours in ISO 2006. The results of the model were related to
the proportion of correct responses of the listeners in studies SN2010 [6] and SNG2016 [7]
for calculating estimates of threshold values based on loudness. The loudness values are
tabulated in Tables 3 to 5 and the resulting threshold values for detecting a reflecting object,
when based on the percentage correct, are shown in Tables 13 and 14.
The difference in loudness level between the loudness threshold and loudness level
of the recordings without the object for 5, 50 and 500 ms duration signals in the anechoic
room were approximately 4.2, 5 and 5 sones for the sighted persons. As an example, the 5 ms
signal in the anechoic room Table 13 had a threshold of 17.5 sones for the sighted persons,
while the mean for loudness with no object in this room was 13.3 sones, as shown in Table 3.
The difference of these two values, 17.5 and 13.3, is 4.2 sones. For the conference room those
differences were 5, 8 and 3 sones, respectively (Tables 3 to 5 and Table 13). These differences
in loudness level make it possible for persons to echolocate, making loudness a potential
information source for echolocation (see also Schenkman and Nilsson [9]; Kolarik et al, [2]).
Comparing the loudness thresholds of sighted and blind persons, the thresholds of the blind
persons were lower than those of the sighted test persons (Table 13). If loudness information
is encoded in the same manner for both groups of test persons, which we believe is a
reasonable assumption, then this analysis shows that blind persons may echolocate at lower
loudness levels than sighted persons.
Echolocation and pitch
Repetition pitch is one of the important information sources that blind people use to detect a
reflecting object at shorter distances (e.g. Bilsen [14]; [6]). We studied how this information is
represented in the auditory system. For this purpose, a dual profile analysis was performed
and is presented above in the section "Pitch analysis: autocorrelation with dual profiles". The
results suggest that repetition pitch can be explained by the peaks in the temporal profile
rather than by peaks in the spectral profile of the autocorrelation function. This is in
agreement with a study by Yost [26], where the peaks in the temporal domain of the
autocorrelation function form the basis for explaining the perception of repetition pitch.
However, the dual profile analysis was not sufficient to determine the strength of the
perceived pitch as the peaks were more random in the temporal profile of the autocorrelation
function. A measure of pitch strength was therefore used that showed if the peaks were
random or not and thereafter computed the pitch strength (Equation 14). The means of the
resulting pitch strengths are shown in Tables 6 to 8 and the thresholds for detecting objects in
Tables 15 and 16. Only the pitch strength values obtained from the 500 ms duration
recordings in Table 8 were considered, as these recordings are not likely to be influenced by
cognitive factors. The pitch strength threshold was lower for blind persons than for sighted. A
reasonable assumption is that pitch information is encoded in the same manner for both blind
and sighted persons. Then it appears that blind persons may echolocate with a lower pitch
strength than sighted persons. The auditory models were used without changing its parameters
43
for the analysis of the two groups and it is thus not possible to infer what factors that
determine the perceptual differences.
Echolocation and timbre
To determine to what extent that sharpness is useful for echolocation, we computed the
weighted centroid of the sounds from studies SN2010 [6] and SNG2016 [7] for specific
loudness using the code of Psysound3. Pedrielli, Carletti, and Casazza [46] showed in their
analysis that the just noticeable difference for sharpness in their study was 0.04 acum. We
used this value as a criterion of sharpness for detecting reflecting objects. The results from our
analysis (Tables 9 to 11) show that the difference in sharpness was greater than 0.04 acum for
recordings with the object at distances of 50, 100, 150 and 200 cm. However, at these
distances both loudness and pitch information are more prominent. Hence, at distances shorter
than 200 cm, sharpness might not be a major information source for echolocation.
One may note that, in study SN2010 [6], for the 500 ms recording in the anechoic
chamber, with the reflecting object at 400 cm and 500 cm, the sharpness difference was
approximately equal to 0.04 acum when compared to the mean of the two recordings without
the object (Table 11). A few of the blind test persons in SN2010 were able to detect objects at
400 cm. If the just noticeable difference for sharpness of 0.04 acum as found by Pedrielli,
Carletti, and Casazza [46] also is a difference threshold for sharpness when echolocating
sounds at longer distances than about 2 m, then sharpness can be used as vital information for
blind people to detect objects at 400 cm. Here we point out that Tables 17 and 18 present a
linear relationship between sharpness and percentage correct, i.e. if there is a higher value of
sharpness then a higher probability of detection will result. However, as discussed previously,
in distinction to loudness or pitch, sharpness does not need to be larger to indicate detection,
since this may be indicated by either being perceived as dull or sharp. A psychometric
function cannot depict this. An experiment controlling for sharpness information of the sound
could clarify its role for echolocation.
Echolocation and room acoustics
Loudness, pitch and sharpness provide people with information useful for echolocation, but
the efficiency of these also depend on the acoustics of the room and the character of the
sounds. The results of many studies of human echolocation are evidence of this (Kolarik et al
[2]; Tonelli, Brayda and Gori [52]; Vercillo et al [53]).
The conference room in SN2010 [6] increased pitch strength and hence enabled the
participants to echolocate at farther distances, while the lecture room in SNG2016 [7]
decreased pitch strength and the listeners presumably had to rely more on other kinds of
information like loudness, resulting in deterioration of object detection. The physical cause
for the deterioration was probably because in SN2010 [6] the loudspeaker was on the chest of
the artificial head, while in SNG2016 [7] it was 1 m behind the artificial head. Another
physical cause for the deterioration might be that the reverberation time in the conference
room in SN2010 [6] was 0.4s, but in the lecture room of SNG2016 [7] it was 0.6s.Too little
reverberation does not seem to be beneficial for human echolocation (see also Tonelli, Brayda
and Gori [52]), but too much cannot neither be useful. We hypothesize that there might be an
optimal amount of reverberation for successful echolocation. Careful design of room
acoustics should improve the possibilities for echolocation by blind persons.
The effects of room acoustics for echolocation are also provided by the recordings in
the anechoic room in study SN2010 [6]. The recordings with object at all distances in this
room, including those at 400cm and 500cm, had no other reflections than from the reflecting
object. The slight sharpness differences that resulted might be used by very skilled listeners to
detect an object.
44
Echolocation, binaural information and movement
Binaural information may provide additional information for echolocation. Both inter-aural
level differences and time differences provide information for echolocation. Papadopoulos et
al. [54] argued that information for obstacle discrimination were found for the frequency
dependent inter aural level differences especially in the range from 5.5 to 6.5 kHz. Nilsson
and Schenkman [55] found that the blind people in their study used interaural level
differences more efficiently than the sighted.
There is evidence that self-generated sounds (e.g. Kellogg [56]; Rice, [57]), as well
as binaural information (e.g. Dunai, Peris-Fajarnés and Brusola [58]) is beneficial for
echolocation. The recordings of the analyzed studies in this report, SN2010 [6] and SNG2016
[7], had the reflecting object directly in front of the recording microphones of the dummy
head, and very little binaural information was therefore provided to the test persons. It was
thus not considered as a source of information in this report, although we did calculate the
SPL values at each ear. As can be seen in Tables 1 and 2, the SPL values at both ears were
very similar.
We may here add that the static nature of the recordings might have resulted in a
poorer echolocation of the test persons. In a real situation movement gives additional
information. Blind persons may move their heads and body, or the object might be moving.
Additionally, they would also use their own sounds. It is reasonable to conclude that such
sounds, together with motion, offer more information to blind persons than static sounds that
are provided by an external source or by an experimenter [2-3].
The present article is based on two studies, both using stationary objects and
stationary listeners. Already Wilson [4] showed theoretical reasons for the benefits for
echolocation that could be achieved through motion. The same conclusions are obtained by
the findings of Bassett and Eastmond [22] of how pitch changes depending on the distance to
the reflecting object. More recent findings by Rosenblum et al [59] showed advantages for
walking when echolocating. Furthermore, self-motion has been found to be beneficial for
echolocation (Wallmeier and Wiegrebe [5]). In a study on blind children walking along a
path, Ashmead, Hill and Talor [60] found that the children could avoid a box by utilizing non-
visual information. These authors stated that congenitally blind children could utilize some of
the auditory information and that they could coordinate this information with functionally
important behavior, such as goal-directed locomotion.
Arias et al [61] saw echolocation as a combination of action and perception, while
Thaler and Goodale [3] in their review stressed that echolocation is an active process. Our
present analysis should be extended to also include studies of motion of objects or persons, as
well as of self-generated sounds. The same kinds of auditory models, possibly with some
modification, could then be used to test processes or hypothesis pertaining to the additional
benefits of motion and of self-generated sounds for echolocation.
Comments on
echolocation
In psychoacoustic experiments a sound is usually presented to participants in a controlled
manner and the perceptual or behavioral responses are measured. This makes it possible to
identify cause-effect relationships between stimuli and response. However, although the
stimuli are presented in a controlled manner in e.g. echolocation studies, the underlying cause
for the echolocation is not obvious, e.g. whether it is biological, perceptual or psychological.
For example, in the study SN2010 [6], the blind participants were able to perform better than
the sighted, but the main cause for the high performance is not evident, except that the cause
is related to blindness.
the auditory model approach
to human
45
De Volder et al [62] were maybe the first to describe different brain activities in
blind and sighted persons in distance tasks by using positron emission tomography. Other
advanced scanning techniques like functional magnetic resonance imaging of the brain may
locate areas in the brain that are activated when persons use echolocation. These techniques
can describe physiological processes relating to or underlying echolocation abilities (Thaler et
al [63]; Thaler et al [64]), but such analyses do not fully reveal how the information used by
blind persons for echolocation is processed and represented in the auditory system, how it
cognitively is perceived and analyzed. Recently there has also been some serious criticism
against MRI studies, that their results may be inflated by false-positive rates (Eklund, Nichols
and Knutsson [65]).
To address the issues of representation and processing, we implemented a number of
auditory models; the binaural loudness model of Moore and Glasberg [32], the auditory image
model, AIM, of Patterson, Allerhand, and Giguere [21] and the sharpness model of Fastl and
Zwicker [33]. We chose the loudness model of Moore and Glasberg [32] since it agrees well
with the equal loudness contours of ISO 2006 and also gives an accurate representation of
binaural loudness (Moore [44]). We chose the auditory image model, AIM, since instead of
using two different modules to depict frequency selectivity and compression; it uses a
dynamic compressive gammachirp filterbank (dcGC) module to depict both the frequency
selectivity and the compression performed by the basilar membrane. The AIM model of
Bleeck, Ives, and Patterson [34], was thus used to analyze the information provided by
repetition pitch. This model is also physiologically inspired. Finally, we used the loudness
model of Glasberg and Moore [31] to analyze sharpness. The sharpness information was
obtained from the weighted centroid of the specific loudness (Fastl and Zwicker [33]). A
general review of auditory processing models is given by Dau [8].
The signal analysis performed on the physical stimuli showed how sound pressure
level, autocorrelation and spectral centroid varied with the recordings. The results with AIM
showed that the peaks in the temporal information was the likely source for echolocation at
shorter distances, and this explanation is thus in line with the analysis by Bilsen [14] and Yost
[26] in how the perception of repetition pitch is represented in people, i.e. that the information
necessary for pitch perception is represented temporally in the auditory system. The analysis
performed with the sharpness model showed that blind participants in our analyzed studies
could have used sharpness to detect objects at longer distances and that both temporal and
spectral information are required to encode this attribute. Our analysis has some similarities to
that of Rowan et al [66] in utilizing models to analyze the perception of level information.
Similar to their analysis we saw cross channel cues or spectral spread information as relevant
for object detection, which we used in our quantification of sharpness.
The analyses with the auditory models we have used do not fully explain how
information necessary for the high echolocation ability of blind persons is represented in the
auditory system. In order to investigate by auditory models, the neural and physiological
causes to the different echolocation abilities of blind and sighted people, the parameters of the
models could be varied to fit participants´ perceptual results. However, we have assumed that
the high echolocation ability was due mainly to a perceptual ability common to both groups.
Therefore, the thresholds for the blind and the sighted persons were obtained by comparing
the results of the auditory models with the perceptual results of both test groups in SN2010
[6] and SNG2016 [7]. As mentioned, the blind participants were consistently better than
sighted when echolocating, and they had lower thresholds of detection than the sighted for all
parameters that were studied.
The analysis with the auditory models confirmed that repetition pitch and loudness
are important information sources for people when echolocating at shorter distances, which is
in agreement with earlier results (e.g. Schenkman and Nilsson [6, 9]; Kolarik, Cirstea,
Pardhan, and Moore [2]). Sharpness is a candidate for being an important source for
46
echolocation both at short and long distance. Psychoacoustic experiments could determine the
usefulness of sharpness and other timbre qualities for echolocation. Highest ecological
validity is probably reached by experiments with real objects in actual environments, but
laboratory studies are a viable alternative. Today simulations of rooms and objects provide
another option for how to study human echolocation, see e.g. Pelegrin-Garcia, Rychtáriková
and Glorieux [66].
Conclusions
We used auditory models to analyze how information for human echolocation in static
situations is represented and processed in the auditory system, i.e. when no movement is
involved. We focused on three perceptual attributes. Two of these, loudness and pitch, are
known to be important for human echolocation. The third attribute, sharpness, an aspect of
timbre, we considered important for echolocation and was therefore also studied. We used a
number of auditory models: The binaural loudness model of Moore and Glasberg [32], the
AIM model of Bleeck, Ives, and Patterson [34], the loudness model of Glasberg and Moore
[31], and for sharpness information we used the weighted centroid of the specific loudness, as
formulated by Fastl and Zwicker [33].
The main results of our analysis are the following. At shorter distances between
person and reflecting object, repetition pitch, loudness and also sharpness provide information
to detect objects by echolocation. At longer distances, sharpness information might be used
for the same purpose. This tentative conclusion has to be justified experimentally by varying
in particular the sharpness characteristics of echolocation sounds. Our analysis confirmed that
repetition pitch can best be represented in the auditory system by the peaks in the temporal
profile rather than by the spectral profile (see also Yost [26]). As the median sharpness
information is computed by using the centroid of the specific loudness varying over time, it is
represented by both the spectral and temporal information.
For our analysis we have assumed that the auditory information for both blind and
sighted persons is represented and processed in the same way. However, this assumption may
not be true. The high echolocation ability of the blind may be the outcome of physiological
changes in the neural system, as some studies have indicated (Thaler et al [63]; Thaler et al
[64]). To investigate this in further detail, one should change the parameters of the auditory
models and analyze the results together with data from neuroimaging and psychoacoustic
experiments. If it is established that the underlying ability of blind persons is because of
certain physiological conditions, then the parameters of the auditory models can be varied
until the results from the auditory models agree with the psychoacoustic results.
We used different auditory models to analyze loudness, pitch and sharpness
attributes. It was assumed that the high echolocation ability of many blind persons is a result
of a general perceptual ability, and we therefore computed perceptual thresholds for the blind
and the sighted persons by using the same models for both groups. The analysis showed that
the blind had lower thresholds than the sighted and could echolocate at both lower loudness
and lower pitch strength levels. As noted above, the recordings in SN2010 [6] and SNG2016
[7], that form the basis for our analysis were recorded in static positions, i.e. there was no
movement between the person and the object. In real life, a blind person would likely be
moving, or the object would be so. In addition, the person is often using his/her own sounds,
which is advantageous for echolocation (e.g. Kellogg, [56]; Rice, [57]). When movement is
provided, and self-produced sounds are used, we believe that the thresholds for the blind
persons would be even lower. These ideas are in alignment with the concept of surplus
information in Schenkman and Nilsson [6] that more information makes perceptual tasks
easier to perform, while lack of information makes perception ambiguous and difficult. This
47
concept follows from Gibsons´s [67] theory of ecological perception. In summary, we have
shown the importance of pitch, loudness and timbre for human echolocation. These three
characteristics have to be further studied, but especially the role of timbre attributes like
sharpness, needs a deeper understanding. Neuroimaging, psychoacoustic experiments, and
signal analysis including auditory models, may help us to understand how information
necessary for the high ability of many blind persons or with visual impairments is represented
and perceived.
Acknowledgment
The original data were collected together with Mats E. Nilsson. This work was partially
supported by the Swedish Council for Working Life and Social Research, (grant number
2008-0600)
12006)
https://www.promobilia.se/, both to BS. The funders had no role in study design, data
collection and analysis, decision to publish, or preparation of the manuscript.
Author contributions
Conceived and controlled the study: BS VG. Performed the simulations: BS VG. Analysed
the data: BS VG. Wrote the paper: BS VG. Funding acquisition: BS. Contributed analysis
tools: VG. Developed the theory: BS VG. Wrote the paper: BS VG.
References
https://forte.se/,
Promobilia
number
(grant
1. Stoffregen TA, Pittenger JB (1995) Human echolocation as a basic form of
and
by
perception and action. Ecological Psychology 7: 181- 216.
2. Kolarik AJ, Cirstea S, Pardhan S, Moore BC (2014) A summary of research
investigating echolocation abilities of blind and sighted humans. Hearing Research
310: 60-68.
3. Thaler L, Goodale MA (2016) Echolocation in humans: an overview. WIREs
Cogn Sci 2016, doi:10.1002/wcs.1408.
4. Wilson JP (1967) Psychoacoustics of obstacle detection using ambient or self-
generated noise. In: Busnel RG, editor, Animal Sonar Systems. Biology and
Bionics, vol 1, (pp 89 – 114) Jouy.en-Josas, France: Laboratoire de Physiologie
Acoustique, INRA-CNRZ.
5. Wallmeier L, Wiegrebe L (2014) Self-motion facilitates echo-acoustic orientation
in humans. Royal Society Open Science, 1:140185.
6. Schenkman BN, Nilsson ME (2010) Human echolocation: Blind and sighted
persons' ability to detect sounds recorded in the presence of a reflecting object.
Perception 39: 483-501.
7. Schenkman BN, Nilsson ME, Grbic N (2016) Human echolocation: Acoustic gaze
trains and acoustic noise. Applied Acoustics 106: 77-86.
for burst
http://dx.doi.org/10.1016/j.apacoust.2015.12.008.
8. Dau T (2008) Auditory processing models. In: Havelock D, Kuwano S, Vorländer
M, editors, Handbook of Signal Processing in Acoustics. New York: Springer.
9. Schenkman BN, Nilsson ME (2011) Human echolocation: Pitch versus loudness
information. Perception 49: 840-852.
48
10. ASA (1973) American national psychoacoustical terminology, s3.20–1973. New
York: American Standards Association.
11. Moore BC (2013) An Introduction to the Psychology of Hearing. San Diego:
12. ASA (1960) Acoustical terminology si, 1–1960. New York: American Standards
Academic Press.
Association.
13. De Cheveigné A (2010) "Pitch perception". In: Plack GJ, editor, Oxford Handbook
of Auditory Science – Auditory Perception, Oxford University Press, Oxford, pp.
71–104.
14. Bilsen FA (1968) On the interaction of a sound with its repetitions. Doctoral
dissertation, University of Technology, Delft, The Netherlands.
15. de Boer E (1956) On the "residue" in hearing. Academic thesis, Amsterdam.
16. de Boer E (1976) On the "residue" and auditory pitch perception. In Keidel WD,
Neff WD, editors, Handbook of Sensory Physiology, vol V-3. Berlin: Springer, pp.
479-583.
17. Licklider JC (1951) A duplex theory of pitch perception. Cellular and Molecular
Life Sciences 7: 128–134.
18. Goldstein JL (1973) An optimum processor theory for the central formation of the
pitch of complex tones. The Journal of the Acoustical Society of America,
54:1496–1516.
19. Wightman FL (1973) The pattern-transformation model of pitch. The Journal of
doi:
54: 407-416.
America.
of
the
http://dx.doi.org/10.1121/1.1913592.
Acoustical
Society
20. Terhardt E (1974) Pitch, consonance, and harmony. The Journal of the Acoustical
Society of America 55: 1061-1069.
21. Patterson RD, Allerhand MH, Giguere C (1995) Time-domain modeling of
peripheral auditory processing: A modular architecture and a software platform.
The Journal of the Acoustical Society of America 98: 1890-1894
22. Bassett IG, Eastmond EJ (1964) Echolocation: Measurement of pitch versus
distance for sounds reflected from a flat surface. The Journal of the Acoustical
Society of America 36: 911-916.
23. Small AM, jr, McClellan ME (1963) Pitch associated with time delay between two
pulse trains. The Journal of the Acoustical Society of America 35: 1246-1255.
24. Bilsen F (1966) Repetition pitch: monaural interaction of a sound with the
repetition of the same, but phase shifted sound. Acustica 17: 295–300.
25. Bilsen F (1969) Repetition pitch and its implications for hearing theory. Acustica
26. Yost WA (1996) Pitch strength of iterated rippled noise. The Journal of the
Acoustical Society of America 100: 3329-3335.
27. ANSI (1994) American national standard acoustical terminology, ansi s1.1-1994.
New York: American National Standard Institute.
28. Plomp R (1970) Timbre as a multidimensional attribute of complex tones. In
Plomp R, Smoorenburg GF, editors, Frequency Analysis and Periodicity Detection
in Hearing, Leiden: W.J.Sitjthoff, pp. 397-414.
29. Plomp R (1976) Aspects of Tone Sensation: A psychophysical study. London, UK:
Academic Press.
30. Plomp R, Steeneken, HJM (1969) Effect of phase on the timbre of complex
tones. The Journal of the Acoustical Society of America, 46(2B), 409-421.
31. Glasberg BR, Moore BCJ (2002) A model of loudness applicable to time-varying
sounds. Journal of Audio Engineering Society. 50.5: 331-342.
22: 63-73.
49
32. Moore BC, Glasberg BR (2007) Modeling binaural loudness. The Journal of the
Acoustical Society of America. 121: 1604–1612
33. Fastl H, Zwicker E (2007) Psychoacoustics. Facts and Models. Berlin: Springer.
34. Bleeck S, Ives T, Patterson RD (2004) Aim-mat: the auditory image model in
matlab. Acta Acustica United with Acustica 90: 781–787.
35. Cabrera D, Ferguson S, Schubert E (2007). 'Psysound3': Software for Acoustical
and Psychoacoustical Analysis of Sound Recordings. Proceedings of the 13th
International Conference on Auditory Display, Montréal, Canada, June 26-29.
36. Rowan D, Papadopoulos T, Edwards D, Holmes H, Hollingdale A, Evans L, Allen
R (2013) Identification of the lateral position of a virtual object based on echoes by
humans. Hearing Research 300: 56-65.
37. Peeters G, Giordano BL, Susini P, Misdariis N, McAdams S (2011) The timbre
toolbox: Extracting audio descriptors from musical signals. The Journal of the
Acoustical Society of America 130: 2902– 2916.
38. Patterson R D, Robinson K, Holdsworth J, McKeown D, Zhang C, Allerhand M
(1992) Complex sounds and auditory images. Auditory Physiology and Perception
83: 429–446.
39. Irino T, Patterson RD (2006).A dynamic compressive gammachirp auditory
Filterbank. IEEE Trans Audio Speech Lang Processing 14:2222-2232.
40. Patterson RD, Unoki M, Irino T (2003) Extending the domain of center
frequencies for the compressive gammachirp auditory filter. The Journal of the
Acoustical Society of America 114: 1529–1542.
41. Krumbholz K, Patterson RD, Nobbe A, Fastl H (2003) Microsecond temporal
resolution in monaural hearing without spectral cues? The Journal of the
Acoustical Society of America 113(5): 2790-2800.
42. Patterson, R. D. (1994). The sound of a sinusoid: Spectral models. The Journal of
the Acoustical Society of America 96(3): 1409-1418.
43. Patterson, R. D. (1994). The sound of a sinusoid: Time-interval models. The
Journal of the Acoustical Society of America 96(3): 1419-1428.
44. Moore BC (2014) Development and current status of the "Cambridge" loudness
models. Trends in Hearing 18: 1–29. doi: 10.1177/2331216514550620.
45. Patterson RD, Handel S, Yost WA, Datta AJ (1996) The relative strength of the
tone and noise components in iterated rippled noise. The Journal of the Acoustical
Society of America 100: 3286-3294.
46. Pedrielli P, Carletti E, Casazza C (2008) Just noticeable differences of
loudness and sharpness for earth moving machines. Proceedings of
Acoustics08, June 29-July 4, 2008, Paris, France, pp. 1231-1236.
47. You J, Jeon JY (2008) Just noticeable differences in sound quality metrics for
refrigerator noise. Noise Control Engineering Journal 56(6): 414-424.
48. Gescheider GA (1985) Psychophysics, 2nd ed. New Jersey: Lawrence Erlbaum.
49. Zychaluk K, Foster DH (2009) Model-free estimation of the psychometric
function. Attention, Perception, and Psychophysics 71: 1414–1425.
50. Yost WA (2007) Fundamentals of hearing: An introduction. San Diego: Elsevier
Academic Press.
51. Yost WA (1997) Pitch strength of iterated rippled noise when the pitch is
ambiguous. Journal of the Acoustical Society of America 101: 644-1648.
52. Tonelli A, Brayda L, Gori M (2016) Depth echolocation learnt by novice sighted
people. PLosOne 11(6):e0156654. Doi10.371/journal.pone.0156654.
50
53. Vercillo T, Jennifer L, Milne JL, Gori M, Goodale MA (2015) Enhanced auditory
spatial localization in blind echolocators. Neuropsychologia 67: 35-40.
54. Papadopoulos T, Edwards DS, Rowan D, Allen R (2011) Identification of auditory
cues utilized in human echolocation - objective measurement results. Biomedical
Signal Processing and Control 6: 280-290 doi:10.1016/j.bspc.2011.03.005.
55. Nilsson ME, Schenkman BN (2016). Blind people are more sensitive than sighted
people to binaural sound-location cues, particularly inter-aural level differences.
Hearing Research 332: 223-232.
56. Kellogg WN (1962) Sonar system of the blind. Science 137: 399-404.
57. Rice CE (1969) Perceptual enhancement in the early blind. Psychological
Research 19: 1-14.
58. Dunai L, Peris-Fajarnés G, Brusola F (2015) Virtual sound localization by blind
people. Archives of Acoustics 40: 561-567.
59. Rosenblum LD., Gordon MS, Jarquin L. (2000) Echolocating distance by moving
and stationary listeners. Ecological Psychology 12: 181-206.
60. Ashmead DH, Hill EW, Talor CR (1989) Obstacle perception by congenitally
blind children. Perception & Psychophysics 46: 425-433.
61. Arias C, Bermejo F, Hüg MX, Venturelli N, Rabinovich D, Skarp AO (2012)
Echolocation: an action-perception phenomenon. New Zealand Acoustics 25: 20–
27.
62. De Volder AG., Catalan-Ahumada M., Robert A, Bol A, Labar D, Coppens A,
Michel C, Veraart C (1999) Changes in occipital cortex activity in early blind
humans using a sensory substitution device. Brain Research 826: 128-134.
63. Thaler L, Arnott SR, Goodale, MA (2011) Neural correlates of natural human
echolocation in early and late blind echolocation experts. PLoS One, 6, e20162.
64. Thaler L, Milne JL, Arnott SR, Kish D, Goodale MA (2014) Neural correlates of
motion processing through echolocation, source hearing, and vision in blind
echolocation
and
of
Neurophysiology 111: 112-27
echolocation
novices.
Journal
experts
sighted
65. Eklund A, Nichols TE, Knutsson H (2016) Cluster failure: Why fMRI inferences
for spatial extent have inflated false-positive rates. Proc Natl Acad Sci USA 113:
7900–7905.
66. Rowan D, Papadopoulos T, Archer L, Goodhew A, Cozens H, Guzman Lopez R,
Edwards D, Holmes H, Allen R (2017) The detection of 'virtual' objects using
echoes by humans: spectral cues. Hearing Research 350: 205 – 216.
67. Pelegrin-Garcia D, Rychtáriková M, Glorieux C (2017) Single simulated reflection
audibility thresholds for oral sounds in untrained sighted people. Acta Acustica
United with Acustica 103: 492 – 505.
68. Gibson JJ (1986). The Ecological Approach to Visual Perception. Hillsdale, NJ:
Lawrence Erlbaum.
51
S1 Supporting information: Calculation of calibration constant
The reference sound pressure level (SPL) for the calibration constants in the anechoic,
conference and lecture rooms in SN2010 [6] and SNG2016 [7], were documented in dB(A),
77, 79 and 79 dB(A), respectively. The calibration constant of the recordings should, in
principle, be A weighted. The results in Table S.1 show the calibrated levels calculated using
the calibration constants in Equations A.1 and A.2, respectively. The A weighted signal gives
an increase in calibrated levels by approximately less than 0.5 dB, which is negligible for
human hearing. We therefore used Equation A.1, and not Equation A.2, for calculating the
calibration constants for all recordings in this report.
The results of Equations A.1 and A.2 for the calibrated levels with and without A
weighting were calculated for the recordings of the 9th version of the left ear, 500 ms signal
with no object of the first recording in the anechoic and conference rooms and of the 9th
version of the left ear, 500 ms with no object recording in the lecture room, see Table S.1.
Schenkman and Nilsson [6]
Schenkman, Nilsson and Grbic [7]
Anechoic room
Conference room
Lecture room
Without A weighting:
Equation A1
With A weighting:
Equation A2
77.0
77.5
79.0
79.5
79.0
79.3
Table S.1: Calibrated levels with and without A weighting for the 9th version at the left ear
for a 500ms sound with no reflecting object, first recording, in anechoic and conference
rooms, and for the 9th version of left ear for 500ms sound with no reflecting object in lecture
room.
20
𝐶𝐶𝐶𝐶= 10�𝑆𝑆𝑆𝑆𝑆𝑆−20∗𝑙𝑙𝑙𝑙𝑙𝑙10�𝑟𝑟𝑟𝑟𝑟𝑟(𝑟𝑟𝑠𝑠𝑙𝑙𝑠𝑠𝑠𝑠𝑙𝑙)
20∗10−6 �
�
𝐶𝐶𝐶𝐶= 10�𝑆𝑆𝑆𝑆𝑆𝑆−20∗𝑙𝑙𝑙𝑙𝑙𝑙10�𝑟𝑟𝑟𝑟𝑟𝑟(𝐴𝐴𝐴𝐴𝐴𝐴𝑠𝑠𝑙𝑙ℎ𝑡𝑡𝑠𝑠𝑠𝑠𝑙𝑙(𝑟𝑟𝑠𝑠𝑙𝑙𝑠𝑠𝑠𝑠𝑙𝑙))
20∗10−6
20
(A.1)
�
�
(A.2)
52
S2 Supporting information: Pitch strength using strobe temporal integration
The temporal profile of the stabilized auditory image for a recording of a 500 ms in the
conference room in SN2010 [6] is shown in Fig S.2. As stated in the section "Auditory
analysis of acoustic information", the stabilized auditory image was implemented with two
modules, sf2003 and ti2003. A brief description of this implementation is given below.
Fig S.2. The temporal profiles of the stabilized auditory image for a 500 ms signal
recorded in the conference room (SN2010 [6]) at 495 ms time frame. The dot indicates
the highest peak and the corresponding values indicate the pitch strength (calculated by
Equation 14) and frequency in Hz (using the inverse relationship of time and frequency, f =
1/t).The sub figure A is for the no object recording and sub figures B, C and D are for the
recordings with object at 50, 100 and 200 cm respectively.
Initially, the sf2003 module uses an adaptive strobe threshold to issue a strobe on the
NAP. After the strobe is initiated, the threshold initially rises along a parabolic path and then
returns to the linear decay to avoid spurious strobes (cf. Fig 8). When the strobes have been
computed for each frequency channel of the NAP, the ti2003 module uses the strobes to
initiate a temporal integration.
The time interval between the strobe and the NAP value determines the position
where the NAP value is entered into the SAI. For example, if a strobe is identified in the
200Hz channel of the NAP at 5 ms time instant, then the level of the NAP sample at 5 ms
time instant is added to the 1st position of the 200 Hz channel in the SAI. The next sample of
the NAP is added to the 2nd position of the SAI. This process of adding the levels of the NAP
samples continues for 35 ms and terminates if no further strobes are identified.
In the case of strobes detected within the 35ms interval, each strobe initiates a
temporal integration process. To preserve the shape of the SAI to that of the NAP, ti2003 uses
weighting, viz. new strobes are initially weighted high (also the weights are normalized such
that the sum of the weights is equal to 1) so that the older strobes contribute relatively less to
the SAI. In this way the time axis of the NAP is converted into a time interval axis of the SAI.
The temporal profile in the sub figures of Fig S.2 was generated by summing the SAI
along the center frequencies. Fig S.2 shows that the recording with no reflecting object had a
53
pitch strength of 0.07, while the recording with the object at 200 cm (the fourth subfigure in
Fig S.2) had a pitch strength of 0.1 at the corresponding frequencies of the repetition pitch. If
this is the case for all the recordings has to be verified.
Previous researchers (Yost [26]; Patterson et al [45]) analyzed the perception of
repetition pitch by the autocorrelation function. We followed the same approach, since
autocorrelation appears to be a good description of how repetition pitch is processed in the
auditory system. To determine whether it is autocorrelation or strobe temporal integration that
best accounts for human echolocation, repetition pitch and the relevant physiological
processes of the auditory system, further analysis is needed, where these two concepts are
studied and compared in a number of different conditions.
|
1908.00077 | 2 | 1908 | 2019-08-02T13:42:25 | Learning in brain-computer interface control evidenced by joint decomposition of brain and behavior | [
"q-bio.NC"
] | Motor imagery-based brain-computer interfaces (BCIs) use an individuals ability to volitionally modulate localized brain activity as a therapy for motor dysfunction or to probe causal relations between brain activity and behavior. However, many individuals cannot learn to successfully modulate their brain activity, greatly limiting the efficacy of BCI for therapy and for basic scientific inquiry. Previous research suggests that coherent activity across diverse cognitive systems is a hallmark of individuals who can successfully learn to control the BCI. However, little is known about how these distributed networks interact through time to support learning. Here, we address this gap in knowledge by constructing and applying a multimodal network approach to decipher brain-behavior relations in motor imagery-based brain-computer interface learning using MEG. Specifically, we employ a minimally constrained matrix decomposition method (non-negative matrix factorization) to simultaneously identify regularized, covarying subgraphs of functional connectivity, to assess their similarity to task performance, and to detect their time-varying expression. Individuals also displayed marked variation in the spatial properties of subgraphs such as the connectivity between the frontal lobe and the rest of the brain, and in the temporal properties of subgraphs such as the stage of learning at which they reached maximum expression. From these observations, we posit a conceptual model in which certain subgraphs support learning by modulating brain activity in regions important for sustaining attention. To test this model, we use tools that stipulate regional dynamics on a networked system (network control theory), and find that good learners display a single subgraph whose temporal expression tracked performance and whose architecture supports easy modulation of brain regions important for attention. | q-bio.NC | q-bio | Learning in brain-computer interface control evidenced by joint decomposition of
brain and behavior
Jennifer Stiso1,2, Marie-Constance Corsi3,4, Jean M. Vettel5,2,6, Javier Garcia5,2, Fabio
Pasqualetti7, Fabrizio De Vico Fallani3,4, Timothy H. Lucas9, and Danielle S. Bassett2,8,9,10,11,12,13
1Neuroscience Graduate Group, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA 19104, USA
2Department of Bioengineering, School of Engineering & Applied Science,
University of Pennsylvania, Philadelphia, PA 19104, USA
3Inria Paris, Aramis project-team, F-75013, Paris, France
4Institut du Cerveau et de la Moelle Epinire, ICM, Inserm, U 1127,
CNRS, UMR 7225, Sorbonne Universit, F-75013, Paris, France
5Human Research & Engineering Directorate, US CCDC Army Research Laboratory, Aberdeen, MD, USA
6Department of Psychological & Brain Sciences, University of California, Santa Barbara, CA, USA
7Department of Mechanical Engineering, University of California, Riverside, CA 92521
8Department of Electrical & Systems Engineering, School of Engineering & Applied Science,
University of Pennsylvania, Philadelphia, PA 19104, USA
9Department of Neurology, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA 19104, USA
10Department of Psychiatry, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA 19104, USA
11Department of Physics & Astronomy, College of Arts & Sciences,
University of Pennsylvania, Philadelphia, PA 19104, USA
12The Santa Fe Institute, Santa Fe, NM 87501, USA and
13To whom correspondence should be addressed: [email protected]
Motor imagery-based brain-computer interfaces (BCIs) use an individuals ability to volitionally
modulate localized brain activity, often as a therapy for motor dysfunction or to probe causal rela-
tions between brain activity and behavior. However, many individuals cannot learn to successfully
modulate their brain activity, greatly limiting the efficacy of BCI for therapy and for basic scien-
tific inquiry. Formal experiments designed to probe the nature of BCI learning have offered initial
evidence that coherent activity across spatially distributed and functionally diverse cognitive sys-
tems is a hallmark of individuals who can successfully learn to control the BCI. However, little
is known about how these distributed networks interact through time to support learning. Here,
we address this gap in knowledge by constructing and applying a multimodal network approach to
decipher brain-behavior relations in motor imagery-based brain-computer interface learning using
magnetoencephalography. Specifically, we employ a minimally constrained matrix decomposition
method -- non-negative matrix factorization -- to simultaneously identify regularized, covarying sub-
graphs of functional connectivity, to assess their similarity to task performance, and to detect their
time-varying expression. We find that learning is marked by diffuse brain-behavior relations: good
learners displayed many subgraphs whose temporal expression tracked performance.
Individuals
also displayed marked variation in the spatial properties of subgraphs such as the connectivity be-
tween the frontal lobe and the rest of the brain, and in the temporal properties of subgraphs such
as the stage of learning at which they reached maximum expression. From these observations, we
posit a conceptual model in which certain subgraphs support learning by modulating brain activ-
ity in regions important for sustaining attention. To test this model, we use tools that stipulate
regional dynamics on a networked system (network control theory), and find that good learners
display a single subgraph whose temporal expression tracked performance and whose architecture
supports easy modulation of brain regions important for attention. The nature of our contribution
to the neuroscience of BCI learning is therefore both computational and theoretical; we first use
a minimally-constrained, individual specific method of identifying mesoscale structure in dynamic
brain activity to show how global connectivity and interactions between distributed networks sup-
ports BCI learning, and then we use a formal network model of control to lend theoretical support
to the hypothesis that these identified subgraphs are well suited to modulate attention.
9
1
0
2
g
u
A
2
]
.
C
N
o
i
b
-
q
[
2
v
7
7
0
0
0
.
8
0
9
1
:
v
i
X
r
a
INTRODUCTION
Both human and non-human animals can learn to voli-
tionally modulate diverse aspects of their neural activity
from the spiking of single neurons to the coherent activity
of brain regions [1 -- 3]. Such neural modulation is made
possible by routing empirical measurements of the user's
neural activity to a screen or other external display de-
vice that they can directly observe [1, 4, 5]. Referred to
as a brain-computer interface (BCI), this technology can
be used to causally probe the nature of specific cognitive
processes [6 -- 9], and offers great promise in the treatment
of neural dysfunction [10 -- 13]. However, translating that
promise into a reality has proven difficult [14 -- 16] due to
the extensive training that is required and due to the fact
that some individuals who undergo extensive training will
only achieve moderate control [5, 17, 18]. A better under-
standing of the neural processes supporting BCI learning
is an important first step towards the development of
BCI therapies and the identification of specific individu-
als who are good candidates for treatment [17, 18].
While BCIs vary widely in their nature, we focus on
the common motor imagery based BCIs where subjects
are instructed to imagine a particular movement to mod-
ulate activity in motor cortex. Performance on motor
imagery based BCIs has been associated with a diverse
array of neural features, demographic factors, and be-
havioral measures [17 -- 21]. Neural features predicting
performance are frequently identified in areas associated
with either performing or imagining action; for example,
better performance is associated with higher pre-task ac-
tivity in supplementary motor areas [22] and larger grey
matter volume in somatomotor regions [22].
Interest-
ingly, performance has also been predicted by activity in
a diverse range of other cognitive systems relevant for
sustained attention, perhaps due to the high cognitive
demands associated with BCI learning [18]. Specifically,
better performance is associated with greater parietal
power suppression in the α band, midline power suppres-
sion in the β band, and frontal and occipital activation
with motor power suppression in the γ band [20, 23, 24].
The role of sustained attention in BCI control is corrob-
orated by the fact that personality and self-report mea-
sures of attention predict successful learning [19]. The
heterogeneity of predictors suggests the possibility that
individual differences in the interactions between cogni-
tive systems necessary for action, action planning, and
attention might explain the idiosyncratic nature of BCI
control, although these interactions are challenging to
quantify [6, 25].
Assessing the interactions between cognitive systems
has historically been rather daunting,
in part due to
the lack of a common mathematical language in which
to frame relevant hypotheses and formalize appropriate
computational approaches. With the recent emergence
and development of network science [26], and its appli-
cation to neural systems [27], many efforts have begun
to link features of brain networks to BCI learning specif-
ically and to other types of learning more generally. In
this formal modeling approach [28], network nodes rep-
resent brain regions or sensors and network edges repre-
sent statistical relations or so-called functional connec-
tions between regional time series [29]. Recent studies
have demonstrated that patterns of functional connec-
tions can provide clearer explanations of the learning
process than activation alone [30], and changes in those
functional connections can track changes in behavior [31].
During BCI tasks, functional connectivity reportedly in-
creases within supplementary and primary motor areas
[15] and decreases between motor and higher-order as-
2
sociation areas as performance becomes more automatic
[32]. Data-driven methods to detect putative cognitive
systems as modules in functional brain networks [33, 34]
have been used to demonstrate that a particularly clear
neural marker of learning is reconfiguration of the net-
work's functional modules [31]. Better performance is
accompanied by flexible switching of brain regions be-
tween distinct modules as task demands change [35 -- 39].
While powerful, such methods for cognitive system de-
tection are built upon an assumption that limits their
conceptual relevance for the study of BCI learning.
Specifically, they enforce the constraint that a brain re-
gion may only affiliate with one module at a time [40],
in spite of the fact that many regions, comprised of het-
erogeneous neural populations, might participate in mul-
tiple neural processes. To address this limitation, recent
efforts have begun to employ so-called soft-partitioning
methods that detect coherent patterns in mesoscale neu-
ral activity and connectivity [40 -- 44]. Common examples
of such methods are independent component analysis and
principal component analysis, which impose pragmatic
but not biological constraints on the orthogonality or in-
dependence of partitions. An appealing alternative is
non-negative matrix factorization (NMF), which achieves
a soft partition by decomposing the data into the small
set of sparse, overlapping, time-varying subgraphs that
can best reconstruct the original data with no require-
ment of orthogonality or independence [45]. Previous
applications of this method to neuroimaging data have
demonstrated that the detected subgraphs can provide a
description of time varying mesoscale activity that com-
plements descriptions provided by more traditional ap-
proaches [40]. For example, some subgraphs identified
with NMF during the resting state have similar spatial
distributions to those found with typical module detec-
tion methods, while others span between modules [40].
As a minimally constrained method for obtaining a soft
partition of neural activity, NMF is a promising candi-
date for revealing the time-varying neural networks that
support BCI learning.
Here, we investigate the properties of dynamic func-
tional connectivity supporting BCI learning.
In indi-
viduals trained to control a BCI, we calculate single
trial phase-based connectivity in magnetoencephalogra-
phy (MEG) data in three frequency bands with stereo-
typed behavior during motor imagery: α (7-14 Hz), β
(15-25 Hz), and γ (30-45 Hz). We construct multimodal
brain-behavior time series of dynamic functional connec-
tivity and performance, and apply NMF to those time
series to obtain a soft partition into additive subgraphs
[45] (Fig. 1). We determine the degree to which a sub-
graph tracks performance by defining the performance
loading as the similarity between each subgraph's tem-
poral expression and the time course of task accuracy.
We first identify subgraphs whose performance loading
predicted the rate of learning and then we explore the
spatial and temporal properties of subgraphs to iden-
tify common feature across participants. We hypothesize
that subgraphs predicting learning do so by being struc-
tured and situated in such a way as to easily modulate
patterns of activity that support sustained attention, an
important component of successful BCI control [18]. Af-
ter demonstrating the suitability of this approach for our
data (Fig. S1A-B), we test this hypothesis by capitaliz-
ing on recently developed tools in network control theory,
which allowed us to operationalize the networks ability
to activate regions involved in sustained attention as the
energy required for network control [46]. Collectively,
our efforts provide a careful network-level description of
neural correlates of BCI performance and learning rate,
and a formal network control model that explains those
descriptions.
METHODS
Participants
Written informed consent was obtained from twenty
healthy, right-handed subjects (aged 27.45 ± 4.01 years;
12 male), who participated in the study conducted in
Paris, France. Subjects were enrolled in a longitudinal
electroencephalography (EEG) based BCI training with
simultaneous MEG recording over four sessions, spanning
2 weeks. All subjects were BCI-naive and none presented
with medical or psychological disorders. The study was
approved by the ethical committee CPP-IDF-VI of Paris.
BCI task
The BCI task consisted of a standard 1 dimensional,
two-target box task (14) in which the subjects modulated
their α [8-12 Hz] and/or β [14-29 Hz] activity to control
the vertical position of a cursor moving with constant ve-
locity from the left side of the screen to the right side of
the screen. Both cursor and target were presented using
the software BCI 2000 [47]. To hit the target-up, the
subjects performed a sustained motor imagery of their
right-hand grasping and to hit the target-down they re-
mained at rest. Some subjects reported that they imag-
ined grasping objects while others reported that they
simply imagined clenching their hand to make a fist.
Each trial lasted 7 s and consisted of a 1 s inter-stimulus
interval, followed by 2 s of target presentation, 3 s of
feedback, and 1 s of result presentation. In the present
study, we restricted our analysis to the feedback portion
of the motor imagery task because we were interested in
the neural dynamics associated with learning to volition-
ally regulate brain activity rather than in the neural dy-
namics occurring at rest. BCI control features including
EEG electrode and frequency were selected in a calibra-
tion phase at the beginning of each session, by instructing
the subjects to perform the BCI tasks without any visual
feedback. The subjects were seated in front of a screen
at a distance of 90 cm. To ensure a stable position of the
3
hands, each subject rested their arms on a comfortable
support, with palms facing upward. We also recorded
electromyogram (EMG) signals from the left and right
arm of subjects, electrooculograms, and electrocardio-
grams. EMG activity was manually inspected to ensure
that subjects were not moving their forearms during the
recording sessions.
MEG Data
Preprocessing
As a preliminary step, temporal Signal Space Sepa-
ration (tSSS) was performed using MaxFilter (Elekta
Neuromag) to remove environmental noise from MEG
activity. All signals were downsampled to 250 Hz and
segmented into trials.
ICA was used to remove blink
and heartbeat artifacts. An FFT of the data from each
subject was inspected for line noise, although none was
found in the frequency bands studied here. We note
that the frequency of the line noise (50 Hz) was out-
side of our frequency bands of interest. In the present
study, we restricted our analyses to gradiometer sen-
sors. Gradiometers sample from a smaller area than
magnetometers, which is important for ensuring a sep-
arability of nodes as expected by network models [48].
Furthermore, gradiometers are typically less susceptible
to noise than magnetometers [49]. We combined data
from planar gradiometers in the voltage domain using
the 'sum' method from Fieldtrip's ft combine planar()
function (http://www.fieldtriptoolbox.org/).
Connectivity Analysis
To estimate phase-based connectivity, we calculated
the weighted phase-locking index (wPLI) [50]. The wPLI
is an estimate of the extent to which one signal consis-
tently leads or lags another, weighted by the imaginary
component of the cross-spectrum of the two signals. Us-
ing phase leads or lags allows us to take zero phase lag
signals induced by volume conduction and to reduce their
contribution to the connectivity estimate, thereby ensur-
ing that estimates of coupling are not artificially inflated
[50]. By weighting the metric by the imaginary com-
ponent of the cross spectrum, we enhance robustness to
noise [50]. Formally, the wPLI between two time series x
and y is given by
E{imag(Γxy)}
E{imag(Γxy)} ,
φ(x, y) =
(1)
where E{} denotes the expected value across estimates
(here, centered at different samples), Γxy denotes the
cross spectrum between signals x and y, and imag() se-
lects the imaginary component.
We first segment MEG data from gradiometers into 3
second trials, sampled at 250 Hz. The cross spectrum
4
FIG. 1. Schematic of non-negative matrix factorization. (1) MEG data recorded from 102 gradiometers is segmented
into windows (w1, w2, w3, w4, ... wn) that each correspond to a single BCI trial. (2) In each window, functional connectivity
is estimated as the weighted phase-locking index between sensor time series. The subject's performance on each trial is also
recorded. (3) The lower diagonal of each trial (highlighted in grey in panel (2)) is reshaped into a vector, and vectors from all
trials are concatenated to form a single data matrix. The subject's time-varying performance forms an additional row in this
configuration matrix. (4) The NMF algorithm decomposes the configuration matrix (composed of neural and behavioral data)
into m subgraphs (where m is a free parameter), with two types of information: (i) the weight of each edge in each subgraph,
also referred to as the connection loading (viridis color scale), and (ii) the time varying expression of each subgraph (black
line graphs). From these data, we calculate the performance loading, or how similar the time-varying performance is to each
subgraph's expression (shades of purple).
is then estimated using wavelet coherence [51] in each
of three frequency bands of interest (α 7-14 Hz, β 15-
20 Hz, and γ 31-45 Hz), with wavelets centered on each
timepoint. We chose to compute the wavelet coherence
due to the fact that -- unlike Welch's method -- it does not
assume stationarity of the signal [51]. We implemented
the procedure in the Fieldtrip package in MATLAB, with
a packet width of 6 cycles and zero-padding up to the
next power of two ('nextpow2'). We then calculate the
wPLI as the mean of the imaginary component of the
cross spectrum, divided by the imaginary component of
the mean of the cross spectrum.
We then construct a network model of these statistical
relationships where sensors (N = 102) are nodes, and the
weight of the edge between node i and node j is given by
the weighted phase-locking value. The graph, G, com-
posed of these nodes and edges is a weighted, undirected
graph that is encoded in an adjacency matrix A. By con-
structing this network model, we can use statistics from
graph theory and computational approaches from control
theory to quantify the structure of inter-sensor functional
relations [6, 28].
Uniformly Phase Randomized Null Model
In order to ensure that our results are not due
to choices in preprocessing, the time invariant cross-
correlation of neural signals, or the autocorrelation of
neural signals, we repeated all of the preprocessing and
analysis steps with a uniformly phase randomized null
model [52]. To enhance the simplicity and brevity of the
exposition, we will also sometimes refer to this construct
simply as the null model. Surrogate data time series from
the null model were calculated using a custom function in
MATLAB. Essentially, the FFT of the raw data is taken,
the same random phase offset is added to every channel,
and then the inverse FFT is taken to return the signal
to the time domain [53]. Mathematically, this process
is achieved by taking the discrete Fourier transform of a
time series yv:
V −1(cid:88)
v=0
Y (u) =
yvei2πuv/V ,
(2)
where V is the length of the time series, v indexes time,
and u indexes frequencies. We then multiply the Fourier
transform by phases chosen uniformly at random before
transforming back to the time domain:
V −1(cid:88)
v=0
yv =
1√
V
eiau Y (u) e−i2πkv/V ,
(3)
where the phase at ∈ [0, 2π).
Construction of a Multimodal Configuration Matrix
In this work, we wished to use a data-driven matrix
decomposition technique to identify time-varying sub-
graphs of functional connectivity that support learning.
Specifically, we created a multimodal configuration ma-
trix of edge weights and BCI performance over time, prior
to submitting this matrix to a decomposition algorithm
that we describe in more detail below. To construct the
matrix, we first vectorize the upper triangle (not includ-
ing the diagonal) of each trial's connectivity matrix, and
then we concatenate all of the vectors and our one per-
formance measure into an E × τ matrix, where τ is the
number of trials (384, if no trials were removed), and E
is the number of edges (5151) plus the number of behav-
ioral measures (1). This concatenation process results in
a 5152 × 384 multimodal (brain-behavior) matrix.
In
this task, each subject's performance is recorded as their
5
percentage of successful trials (out of 16) on each run,
and we therefore interpolate the performance time series
to obtain a graded estimate of percentage correct in each
trial that is τ time points long. The performance vector
is then normalized to have the same mean as the other
rows of the configuration matrix.
Non-negative Matrix Factorization
We used a data-driven matrix decomposition method
-- non-negative matrix factorization (NMF) -- to iden-
tify time-varying groups of neural interactions and be-
havior during BCI learning [45].
Intuitively, NMF de-
composes a matrix into a set of additive subgraphs with
time-varying expression such that a linear combination
of these subgraphs weighted by temporal expression will
recreate the original matrix with minimal reconstruction
error [40, 45]. The NMF algorithm can also be thought
of as a basis decomposition of the original matrix, where
the subgraphs are basis sets and the temporal coefficients
are basis weights. Unlike other graph clustering methods
[54, 55], NMF creates a soft partition of the original net-
work, allowing single edges to be a part of multiple sub-
graphs. Additionally, unlike other basis decomposition
methods [54, 55], NMF does not impose harsh constraints
of orthogonality, or independence of the subgraphs; it
simply finds the most accurate partition, given that the
original matrix is non-negative.
In many systems (in-
cluding phase-locking) the non-negativity constraint is
not difficult to satisfy, and is beneficial in physical sys-
tems where the presence of a negative weight would be
difficult to interpret.
Formally, the NMF algorithm will approximate an
E × T configuration matrix A by the multiplication of
two matrices: W, the subgraph matrix with dimensions
E × m, and H, with dimensions m × T . Here, E is the
number of time varying processes (behavior and func-
tional connections derived from MEG data), T is the
number of time points, and m is the number of subgraphs.
We solve for W and H such that:
min
W,H
1
2
A − WH2
F + αW2
F + β
H(:, t)2
1
(4)
T(cid:88)
t=1
where β is the penalty to impose sparse basis weights, and
α is the regularization for the basis set. Regularization is
frequently used in machine learning algorithms to avoid
overfitting data, which is especially important when em-
ploying these techniques to examine highly variable single
trial estimates of functional connectivity [56]. Addition-
ally, selecting for sparsity will encourage the characteri-
zation of local neural processes where many edges do not
contribute [40]. From many such local processes arises
the diversity of cognitive functions involved in complex
tasks such as BCI control [57].
To solve the NMF equation, we use an alternating non-
negative least squares with block-pivoting method with
100 iterations for fast and efficient factorization of large
matrices, where W and H with non-negative weights are
drawn from a uniform random distribution on the inter-
val [0, 1][58]. The parameter m is drawn from the range
(2,20), and α and β are drawn from the range (0.001,2).
We select for parameters that will both minimize the
residual error, and maximize the temporal and subgraph
sparsity [40]. Specifically, we select the optimal parame-
ters ¯m, ¯α, and ¯β that are in the lowest 25th percentile for
residual error, and the highest 25th percentile for tempo-
ral and subgraph sparsity. This procedure resulted in an
average ¯m of 7.4, an average ¯α of 0.46, and an average ¯β
of 0.45. Distributions of parameters and reliability across
runs are shown in Fig. S2 and S3.
Given the non-deterministic nature of this approach,
we also test for the stability of our identified clusters us-
ing a consensus clustering algorithm [59]. Our procedure
was comprised of the following ordered steps: (1) run the
NMF algorithm r = 100 times per multimodal configu-
ration matrix, (2) concatenate the subgraph matrix W
across r runs into an aggregate matrix with dimensions
E × (r ¯m), and (3) apply NMF to the aggregate matrix
to determine a final set of subgraphs and expression co-
efficients [40]. While the implementation is heuristic in
nature, we found that across two runs of the algorithm,
we obtain highly consistent selections for parameters (see
Supplement), bolstering confidence in the robustness of
the subsequent analyses.
Subgraph Inclusion
Most subgraphs are sparse, with distributions of tem-
poral coefficients skewed towards zero (see Fig. S4).
However, for every subject and every frequency band,
one subgraph showed very little regularization (no edges
were equal to 0) and had a uniform, rather than skewed
distribution of temporal coefficients. These subgraphs
are clear outliers from the others, and appear to be cap-
turing global phase-locking across the entire brain, rather
than any unique subsystem. To answers questions about
the time varying interactions between neural systems, we
were interested in differences between the subgraphs that
were spatially localized, having edges regularized to zero.
Because including these outlier subgraphs would obscure
those differences, we removed these subgraphs from all
further analyses.
Group Average Subgraphs
After applying NMF to the multimodal brain-behavior
matrix, we next turned to a study of the nature of the
detected subgraphs after ranking them by performance
loading. Specifically, we were initially interested in de-
termining which edges contributed to each ranked sub-
graph most consistently across the population. For this
purpose, we used a consistency based approach to cre-
6
FIG. 2. BCI performance. Each subject's average perfor-
mance across four days within two weeks. BCI Score is the
percentage of correct trials during that session.
ate a group representative subgraph for each ranked sub-
graph [60].
In this procedure, each subject's subgraph
was first thresholded to retain only the 25% strongest
connections (see Fig. S5 for evidence that results are ro-
bust to variations in this choice). We then constructed
an average N × N subgraph G, where N is the number
of channels and where each element Gij quantifies how
many subjects (out of 20) displayed an edge between re-
gion i and region j in their thresholded subgraph.
In
addition to visually depicting these group representative
subgraphs, we also wished to summarize their content in
large anatomical regions. We therefore binned edges into
10 anatomically defined areas in frontal, motor, parietal,
occipital, and temporal lobes in both hemispheres. Lobes
were obtained from BrainStorm [61] software (neuroim-
age.usc.edu/brainstorm/Tutorials/MontageEditor) (Fig.
S8).
Optimal Control
Our final broad goal was to provide a theoretical ex-
planation for why certain networks support BCI learning.
We hypothesized that these regularized networks might
have structures that make it easier for the brain to mod-
ulate the patterns of activity that are necessary for BCI
control. This hypothesis motivated us to formulate and
validate a model to explain how the sparse statistical re-
lationships characteristic of each subgraph could support
the production of brain activity patterns implicated in
BCI learning [62, 63]. Additionally, this model should
account for the brain's ability to reach these patterns of
activity in the context of the BCI task, where there is
increased volitional modulation of the left motor cortex.
Here, we use tools from network control theory to sat-
isfy these conditions [64]. Specifically, we characterize
the theoretical brain activity at each sensor as a vector
x(t), and we use the adjacency matrix A of a subgraph
40507090Session 1Session 2Session 3Session 4TimeBCI Score(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)(cid:31)405060708090Sess1Sess2Sess3Sess4TimeBCI Scoreas.factor(sess_cont)(cid:31)(cid:31)(cid:31)(cid:31)6080 to quantify the ease with which that activity can affect
other regions. We then incorporate volitional input con-
trol as input into the brain (u(t)) at a specific region
(given by B). Then, by stipulating
x(t) = Ax(t) + Bu(t),
(5)
we model the linear spread of activity along the connec-
tions in A in the context of input to regions given in
B.
With this model of network dynamics, optimal control
trajectories can be formalized and identified by develop-
ing a cost function that seeks to minimize two terms: (i)
the distance of the current state from the target state
and (ii) the energy required for control. Specifically, we
solve the following minimization problem:
(xT − x(t))T (xT − x(t)) + ρuκ(t)T uκdt,
(cid:90) T
min
u
0
7
node "b", an assumption that is violated by measures
of statistical similarity such as the Pearson correlation
coefficient which is the measure of functional connectiv-
ity most commonly employed in neuroimaging studies.
Because we are using neither structural connectivity nor
common measures of functional connectivity, it was nec-
essary for us to first prove that the networks we are study-
ing are consistent with our model. To address the first
point regarding the propagation of activity along edges,
we demonstrate that the structure of the subgraphs used
have utility in predicting empirical brain state transi-
tions, and that the relative contribution of each subgraph
is related to its temporal expression (Fig. S1C-D). It is
only in light of these validations that we are able to in-
terpret our results as a potential model for driving brain
activity. To address the second point regarding isolation
of pairwise relations not due to third party effects, we
note that the matrix A that we study reflects statistical
similarity in phase after strict regularization that removes
redundant statistical relationships (Fig. S1A-B).
s.t.
x = Ax(t) + Bu(t), x(0) = x0,
and x(T ) = xT ,
(6)
Target state definition
where ρ is a free parameter that weights the input con-
straint, xT is the target state, and T is the control hori-
zon, which is a free parameter that defines the finite
amount of time given to reach the target state. Dur-
ing BCI control, there is specific, targeted control to a
specific area of the brain (here, the left motor cortex) in
addition to other ongoing control and sensory processes.
We wished for our selection of the input matrix B to
reflect this richness and also allow for computationally
tractable calculations of optimal control, which is diffi-
cult for sparse control sets. Therefore, we constructed
the input matrix B so as to allow input that was domi-
nated by the BCI control site, while maintaining minor
contributions from other areas. More specifically, rather
than being characterized by binary state values, chan-
nels other than the one located over left motor cortex
were given a value of approximately 5 × 10−5 at their
corresponding diagonal entry in B. See Supplement for
the full derivation from [62].
It is important to note that in general the tools from
linear controllability theory are not applicable to the
functional networks commonly derived from neuroimag-
ing data for two reasons. The first reason is that the
model which the tools are built upon stipulates a time-
dependent propagation of activity along edges; such a
propagation is physically true for structural connections
derived from white matter, but is not generally true for
other types of connections used in network models, such
as morphometric similarity or most common functional
connectivity measures. While we do not expect that
these simple models will fully capture neural dynamics, it
is important to explore how much variance these models
do explain, even if we expect that amount to be small.
The second reason is that the model assumes that in-
teractions between nodes "a" and "c" are not due to
A central hypothesis in this work is that certain reg-
ularized subgraphs are better suited to drive the brain
to patterns of activity that are beneficial for BCI con-
trol than others. To test this hypothesis, we create tar-
get states that reflect these beneficial patterns, based
on previous literature. Target states for motor im-
agery and attention are obtained for each band indi-
vidually from references [20, 23, 24], and can be briefly
described as follows: α contralateral motor suppression
for motor imagery and parietal suppression for atten-
tion, β contralateral motor suppression and ipsilateral
motor activation for motor imagery and vertex suppres-
sion for attention, and γ contralateral motor activation
for motor imagery and motor cortex suppression with
frontal and occipital activation for attention (Fig. S9).
Channels were divided into lobes using standard mon-
tages provided by Brainstorm [61] software (neuroim-
age.usc.edu/brainstorm/Tutorials/MontageEditor). The
target state of channels in brain regions where we did
not have specific hypotheses for their activity were set
to zero; the target state of channels with activation were
set to 1 and that of channels with deactivation were set
to -1. Initial states were set to 0 for all channels. We
then calculate the optimal energy (using the optimal con-
trol equation described above) required to reach each of
these target states to test the hypothesis that subgraphs
that support learning will have lower energy requirements
than those that do not.
Statistical Analyses
Much of our analyses involve testing differences in
distributions across subjects for different subgraphs
the
lm.beta
package
calculated
coefficients
or sessions, both for phase-randomized and empirical
data. We also compare these distributions to subject
learning rate defined as the slope of performance over
time. For the results displayed in Fig. 2 here in the
main manuscript, we used a repeated measures ANOVA
to test for the presence of a main effect across conditions
given that the distributions of performances were normal
(see Fig. S10). In Fig. 3 here in the main manuscript, we
sought to predict learning rate with ranked performance
loading.
After plotting quantile-quantile plots (see
Fig. S11-S13) for the learning rate, and each of the
performance loadings, it became clear that the lowest
loadings were not normally distributed. Therefore, we
used a linear model combined with non-parametric test-
ing utilizing 5000 permutations (lmPerm package in R
https://cran.r-project.org/web/packages/lmPerm).
were
us-
Standardized
ing
in R (https://cran.r-
project.org/web/packages/lm.beta/lm.beta.pdf). We
use a Bonferroni correction to control
false positive
errors due to multiple comparisons across all 6 pre-
dictors (α = 0.008). To obtain an estimate of how
sensitive our results are to our specific sample, we also
plot summary statistics from 500 models obtained from
bootstrapping a sample of equal size (N = 60, 3 band
and 20 subjects). To examine differences in consistency
(Fig. 4 here in the main manuscript), we use a linear
model (consistency ∼ band + dataT ype + rank) to
test for a main effect of data type (null or empirical),
band, and subgraph on consistency (see Fig. S14). We
next sought to determine if different subgraphs had
consistently different temporal expression for null and
empirical data (Fig. 5 here in the main manuscript).
We also used a repeated measures ANOVA to test for a
main effect of subgraph across bands, and paired t-tests
to test for differences amongst individual subgraphs
(Fig. S15). Lastly, for the results shown in Fig. 6 here
in the main manuscript, we test the relationship between
learning rate and optimal control energy differences for
several different models. Pearson's correlations were
used, given that the data appears normally distributed
and has few outliers (see Fig. S16-S19).
Data and Code
Code for analyses unique to this manuscript are avail-
able at github.com/jastiso/netBCI. Code for the NMF al-
gorithm and the NMF parameter selection is available at
github.com/akhambhati/Echobase/tree/master/Echobase
for
/Network/Partitioning/Subgraph.
at
optimal
is
github.com/jastiso/NetworkControl.
neces-
sary to reproduce each figure will be made available
upon request.
Code
available
analyses
control
Data
8
RESULTS
BCI Learning Performance
Broadly, our goal was to examine the properties of
dynamic functional connectivity during BCI learning,
and to offer a theoretical explanation for why a certain
pattern of connectivity would support individual differ-
ences in learning performance. We hypothesized that de-
composing dynamic functional connectivity into additive
N × N subgraphs would reveal unique networks that are
well suited to drive the brain to patterns of activity as-
sociated with successful BCI control. We use MEG data
from 20 healthy adult individuals who learned to control
a motor-imagery based BCI over four separate sessions
spanning a two week period. Consistent with prior re-
ports of this experiment [32], we find a significant im-
provement in performance across the four sessions (one-
way ANOVA F (3, 57) = 13.8, p = 6.8−7) (Fig. 2). At
the conclusion of training, subjects reached a mean per-
formance of 68%, which is above chance (approximately
55 - 60%) level for this task [65].
Dynamic patterns of functional connectivity
supporting performance
To better understand the neural basis of learning per-
formance, we detected and studied the accompanying
patterns of dynamic functional connectivity. First, we
calculated single trial phase-based connectivity in MEG
data in three frequency bands: α (7-14 Hz), β (15-25 Hz),
and γ (30-45 Hz). We then used non-negative matrix fac-
torization (NMF) -- a matrix decomposition method -- to
separate the time-varying functional connectivity into a
soft partition of additive subgraphs. We found that the
selected parameters led to an average of 7.4 subgraphs,
with a range of 6 to 9, and that all frequency bands had
a decomposition error lower than 0.47 (mean α error =
3.52, mean β error = 0.379, mean γ error = 0.465) (Fig.
S2). The error is the Frobenius norm of the squared dif-
ference between our observed and estimated connectivity
matrices (with dimensions 5152 × 384) and takes values
between 0 and 1. For each band, the error value is low,
giving us confidence that we have fairly accurately recon-
structed relevant neural dynamics. To determine whether
any properties of the identified subgraphs were trivially
due to preprocessing choices, NMF parameters, or time-
invariant autocorrelation in neural activity, we repeated
the full decomposition process after permuting the phases
of all time series uniformly at random. We found that
the statistics of subgraph number and decomposition er-
ror were similar for the uniformly phase randomized data,
indicating that any differences in subgraph and tempo-
ral expression between null and empirical data is not due
to the NMF algorithm's inability to find a good decom-
position, but rather due to the structure of the chosen
decomposition (Fig. S2).
We quantified the similarity between each subgraph's
temporal expression and the time course of performance,
and we refer to this quantity as the subgraph's per-
formance loading (Fig. 1). We hypothesized that the
ranked performance loading would predict task learn-
ing, as operationalized by the slope of performance over
time. It is important to note the distinction between per-
formance and learning: performance is defined as task
accuracy and therefore varies over time, while learning
is defined as the linear rate of change in that perfor-
mance over the course of the experiment (384 trials over
4 days). We tested whether learning was correlated with
the performance loading of subgraphs. Because the min-
imum number of subgraphs in a given subject was 6,
we decided to investigate the top four highest perfor-
mance loading subgraphs, and the smallest and second
smallest nonzero loading subgraphs. We found a general
trend that the performance loading from high loading
subgraphs was negatively associated with learning rate,
and the performance loading from low loading subgraphs
was positively associated with learning rate (Fig. 3AB).
We assessed the statistical significance of these trends
and found that only the third highest loading subgraph
displayed a performance loading that was significantly
correlated with learning rate after Bonferroni correction
for multiple comparisons (linear model with permuta-
tion tests slope ∼ loading3 + band : p = 0.005). Per-
formance loading from uniformly phase randomized sur-
rogate data for this subgraph did not predict learning
rate (p = 0.292). The direction of the observed effect in
the empirical data is notable; subjects with lower load-
ing onto high loading subgraphs learned the task better,
suggesting that learning is facilitated by a dynamic in-
terplay between several subnetworks. It is also notable
that the highest loading subgraphs are not the strongest
predictors of learning, indicating that the subgraphs that
most closely track performance are not the same as the
subgraphs that track changes in performance.
Spatial properties of dynamic patterns of functional
connectivity
Next we sought to better understand why the third
highest loading subgraph most robustly predicted learn-
ing. We hypothesized that because of this subgraph's
predictive power across subjects, it might recruit consis-
tent brain regions and reflect the involvement of specific
cognitive systems across subjects. To evaluate this hy-
pothesis, we began by investigating the shared spatial
properties of this subgraph in comparison to the others.
To identify shared spatial features we grouped subgraphs
together by their ranked performance loading, and then
quantified how consistent edges were across participants
[60] (see Methods). We found that the average consis-
tency varied by frequency band, and differed between the
empirical and surrogate data, but not across ranked sub-
graphs (linear model consistency ∼ band + rank + data :
9
FIG. 3. Performance loading predicts learning.
(A)
Here we show the p-values for empirical (green) and uniformly
phase randomized (grey) data for linear models predicting
the slope of performance with ranked performance loading
from each frequency band. The black line corresponds to p =
0.05, while the red dashed line corresponds to the Bonferroni
corrected α = 0.008. Error bars show the standard error
and median of p values from 500 models with bootstrapped
samples. (B) The standardized regression coefficients for the
same models. Error bars show the standard error and mean
of coefficients from 500 models with bootstrapped samples.
Fband(2, 17) = 90.36, pband = 9.00×10−10, Fdata(1, 17) =
41.8, pdata = 5.78 × 10−6). The α band had the most
consistent edges, followed by the γ band, and then the β
band (tαβ = −12.68, pαβ = 4.3 × 10−10, tαγ = −10.41,
pαγ = 1.2 × 10−8). In the uniformly phase randomized
surrogate data, we observed less consistent subgraphs
than those observed in the empirical data (t = −6.47,
p = 5.78 × 10−6). These observations support the con-
clusion that across the population, despite heterogeneous
performance, similar regions interact to support perfor-
mance and learning to varying degrees.
Anatomically, subgraphs were dominated by connec-
tivity in the frontal lobe sensors, with subtle differences
in the pattern of connections from the frontal lobe sen-
sors to sensors located in other areas of the brain (Fig. 4).
To determine which functional edges were most consis-
tent in each subgraph and frequency band, we calculated
the average consistency over each lobe and motor cortex
in both hemispheres (for the same analysis in surrogate
data, see Fig. S6).
In the α band, the most consis-
tent edges on average were located in the left frontal lobe
in the highest performance loading subgraph, in the left
occipital lobe in the second highest performance load-
ing subgraph, between right frontal and right motor in
-2.0-1.0-0.50.0Highest2nd H 3rd 4thLowest2nd Llog10(p−value)EmpiricalNull00.2Standardized Coefficientp = 0.05ABHighest2nd H 3rd 4thLowest2nd L-0.2 p = 0.006-1.5-0.4 the third highest performance loading subgraph, and be-
tween left frontal lobe and right parietal lobe in the low-
est performance loading subgraph.
In the β band, the
most consistent edges were located between right and
left frontal lobe for the highest and second highest per-
formance loading subgraph, between left frontal lobe and
right motor for the third highest performance loading
subgraph, and between left and right frontal lobe for the
lowest performance loading subgraph.
In the γ band,
the most consistent edges were located in the left frontal
and right frontal lobes for the highest performance load-
ing subgraph, in the left frontal lobe and right motor for
the second highest performance loading subgraph, and in
left frontal and right frontal lobe for the third highest and
lowest performance loading subgraphs. We also note that
the most consistent individual edges for each subgraph
are still only present in 10-12 individuals, indicating a
high amount of individual variability. Collectively, these
observations suggest widespread individual variability in
the spatial composition of ranked subgraphs, with the
most consistent connectivity being located in the frontal
lobe during BCI learning.
10
later than all others (paired t-test N = 20, t2H = 10.9,
p2H = 1.39 × 10−9; t3H = 7.56, p3H = 3.57 × 10−7;
tlow = 8.07, plow = 1.49−7). In the γ band, the highest
performance loading subgraph peaked significantly later
than the second highest, and lowest loading subgraphs
(paired t-test N = 20, t2H = 4.50, p2H = 2.46 × 10−4;
tlow = 8.06, plow = 1.49 × 10−7).
(Fig. 5). Finally,
we asked whether the time of the peak in the third
highest performance loading subgraph predicted learn-
ing. We did not find a relationship between peak time
and learning in any frequency band (Pearson's correla-
tion: α : r = 0.005, p = 0.98, β : r = 0.047, p = 0.84,
γ : r = −0.21, p = 0.037). To summarize these findings,
we note that across participants and especially in the
β band, subgraphs that support performance are highly
expressed late in learning, when performance tends to
be highest. However, subgraphs that support learning
do not have consistent peaks across subjects, and each
individual's peak does not predict their learning rate, in-
dicating that some other feature of these subgraphs must
explain their role in learning.
Temporal properties of dynamic patterns of
connectivity supporting BCI learning via network
functional connectivity
control theory
Explaining dynamic patterns of functional
Importantly, subgraphs can be characterized not only
by their spatial properties, but also by their temporal
expression. We therefore next examined the temporal
properties of each subgraph to better understand why
the third highest performance loading subgraph most ro-
bustly predicted learning. As a summary marker of tem-
poral expression, we calculated the total energy of the
time series operationalized as the sum of squared values,
as well as the time of the peak value of the time series.
Across frequency bands, we found no significant depen-
dence between energy and subgraph ranking. We did find
a significant effect of rank for the peak time of temporal
expression obtained from the empirical data (repeated
measures ANOVA peak ∼ rank + band : Frank(3, 215) =
6.67, prank = 2.53 × 10−4 but not from the uniformly
phase randomized surrogate data (Frank(3, 215) = 1.28,
p = 0.282). Overall, peak times are widely distributed
across individuals. However we find that across bands,
the highest performance loading subgraph has a later
peak, which is intuitive since performance is generally
increasing over time and these subgraphs most strongly
track performance.
We then performed post-hoc paired t-tests corrected for
multiple comparisons (Bonferroni correction α = 0.006)
between the highest performance loading subgraph and
all other ranked subgraphs in each band. In the α band,
the highest performance loading subgraph only peaked
significantly later than the lowest (paired t-test N =
20, tlow = 8.06, plow = 1.49 × 10−7) after Bonfer-
roni correction (α = 0.006).
In the β band, the high-
est performance loading subgraph peaked significantly
Lastly we asked how the third highest loading sub-
graph could facilitate successful BCI performance, as
shown in Fig. 3. Here, we considered an edge -- extracted
under penalties of spatial and temporal sparsity -- as a po-
tential path for a brain region to affect a change in the ac-
tivity of another brain region [66 -- 68]. Assuming the true
connectivity structure is sparse, the regularization ap-
plied in the NMF algorithm can remove large statistical
relationships between regions that are not directly con-
nected, but might receive common input from a third re-
gion [69, 70] (see Methods for addition discussion, and see
Fig. S1A-B for the effect of regularization on the preva-
lence of triangles). We hypothesized that the pattern of
edges in this subgraph would facilitate brain states, or
patterns of activity, that were predictive of BCI literacy.
Specifically, we expected that when the brain mirrored
the connectivity of the third subgraph, the brain could
more easily reach states of sustained motor imagery or
sustained attention than when the brain mirrored the
connectivity of the lowest performance loading subgraph.
We also hypothesized that the magnitude of this differ-
ence would be predictive of each subject's learning rate.
To test these hypotheses, we used mathematical mod-
els from network control theory to quantitatively esti-
mate the ease with which the brain can reach a desired
pattern of activity given a pattern of connectivity (see
Methods and Fig. S1C-D for analyses demonstrating the
efficacy of the regularized subgraphs in linearly predict-
ing changes in activity). Specifically we calculated the
optimal control energy required to reach a target state
(either sustained motor imagery or sustained attention)
11
FIG. 4. Spatial distribution of subgraph edges that are consistent across participants. Consistent edges for each
frequency band and for each ranked subgraph. Left images show individual edges plotted on a topographical map of the brain.
Right images show the mean edge weight over sensors for a given region. We studied 10 regions, including the frontal lobe,
temporal lobe, parietal lobe, occipital lobe, and motor cortex in both hemispheres. The weight of the edge corresponds to the
number of individual participants for whom the edge was among the 25% strongest for that subgraph.
from an initial state when input is applied primarily to
the left motor cortex, which was the site of BCI control
(Fig. 6A-B).
We tested whether the third highest performance load-
ing subgraph supported the transition to states of sus-
tained motor imagery or sustained attention with smaller
energy requirements than other subgraphs that did not
support learning in the same way. We chose the lowest
performance loading subgraph for comparison because it
was the only subgraph with a large positive standard-
ized regression coefficient for predicting learning, which
contrasts sharply with the large negative coefficient for
the third subgraph. For both states (motor imagery and
attention), we found no population level differences in
energy requirements by the two subgraphs (paired t-test
tα = −0.005, pα = 0.565,
N = 20, motor imagery:
tβ = 1.38, pβ = 0.184, tγ = −1.00, pγ = 0.329. atten-
tion: tα = −1.35, pα = 0.193, tβ = −0.344, pβ = 0.735,
tγ = −0.937, pγ = 0.360). We next tested whether the
magnitude of the difference in energy required by the
two subgraphs to reach a given state tracked with learn-
ing rate. In the β band, we observed a significant cor-
relation between the magnitude of the energy difference
to reach attentional states and learning rate over sub-
jects (Pearson's correlation coefficient r = 0.560, p =
0.0103, Bonferroni corrected for multiple comparisons
12
(Pearson's correlation r = 0.40, p = 0.077), the second
highest with the third highest performance loading sub-
graph (Pearson's correlation r = 0.266, p = 0.257), or
the second highest with the lowest performance loading
subgraph (Pearson's correlation r = −0.072, p = 0.764).
This pattern of null results underscores the specificity of
our finding.
Reliability and specificity of inferences from network
control theory
Collectively, our findings are consistent with the hy-
pothesis that during BCI learning, one subnetwork of
neural activity arises, separates from other ongoing pro-
cesses, and facilitates sustained attention. An alternative
hypothesis is that our results are due to trivial factors re-
lated to the magnitude of the attentional state, or could
have just as easily been found if we had placed input to
a randomly chosen region of the brain, rather than to
the left motor cortex which was the actual site of the
BCI control. To determine whether these less interest-
ing factors could explain our results, we performed the
same network control calculation but with a spatially
non-overlapping target state, and then -- in a separate
simulation -- with a mirrored input region (right motor
cortex rather than left motor cortex). We performed the
spatial shifting by ordering the nodes anatomically (to
preserve spatial contiguity), and then circular shifting
the attention target state by random number between 1
and N−1. For 500 circularly shifted states, only 3 (0.6%)
had a correlation value equal to or stronger than the one
observed (Fig. S7). Furthermore, we found no significant
relationship between learning rate and the difference in
energy required by the two subgraphs to reach the true
attention state when input was applied to the right motor
cortex instead of the left motor cortex (Pearson's corre-
lation t = 0.711, p = 0.313). Together, these two findings
suggest that the relationship identified is specific to BCI
control.
Finally, we assessed the robustness of our results to
choices in modeling parameters. First we performed the
computational modeling with two different sets of control
parameter values (see Supplement). In both cases, the
significant relationship remained between learning rate
and the difference in energy required by the two sub-
graphs to reach the attentional state (set one Pearson's
correlation coefficient r = 0.476, p = 0.0338; set two
Pearson's correlation coefficient r = 0.514, p = 0.0204).
Second, since our target states were defined from prior
literature, there was some flexibility in stipulating fea-
tures of those states. To ensure that our results were not
unduly influenced by these choices, we tested whether
ideologically similar states would provide similar results.
Namely, we assessed (i) the impact of varying the mag-
nitude of (de)activation by changing (-)1 to (-)2, (ii) the
impact of the neutral state by changing 0 to 1, and (iii)
the impact of negative states by changing -1, 0 and 1 to
FIG. 5. Temporal expression of ranked subgraphs. The
peak temporal expression for every subject (black data point),
for each frequency band (indicated by color) and for each
subgraph (ordered vertically). Violin plots show the density
distribution of all subjects' peaks. The median is marked with
a solid line through the violin plot.
across frequency bands; Fig. 6). Notably, the relation-
ship remained significant when controlling for subgraph
density (linear model slope ∼ energy dif f erence +
density dif f erence: tenergy = 2.68, penergy = 0.0158,
tdensity = −0.266, pdensity = 0.794). When using sub-
graphs derived from the uniformly phase randomized sur-
rogate data, the relationship was not observed (Pearson's
correlation r = −0.0568, p = 0.819). We next asked
which subgraph contributed most to this effect. We
found no significant relationship between learning rate
and the energy required to reach the attentional state by
the third highest performance loading subgraph (Pear-
son's correlation r = −0.389, p = 0.702) or by the low-
est performance loading subgraph (Pearson's correlation
r = 0.227, p = 0.335). This finding suggests that learn-
ing rate depends on the relative differences between sub-
graphs, rather than the energy conserving architecture
of one alone. As a final test of specificity, we assessed
whether this difference was selective to the third highest
and lowest performance loading subgraph. We found no
significant relationship when testing the difference of the
highest with the third highest performance loading sub-
graph (Pearson's correlation r = −0.554, p = 0.586), the
highest with the lowest performance loading subgraph
0.250.50.751Lowest2HHighestNormalized Peak03HNormalized TimeBand:AlphaBetaGammaRank(last trial)((cid:31)rst trial)0 13
FIG. 6. Separation of the ability to modulate attention predicts learning. Different patterns of connections will
facilitate transitions to different patterns of brain activity. We hypothesize that the ease with which connections in certain
regularized subgraphs facilitate transitions to patterns of activity that support either motor imagery (A) or attention (B) will
predict learning rate. We use network control theory to test this hypothesis. We model how much energy (u(t)) is required
to navigate through state space from some initial pattern of activity x(0) to a final pattern of activity x(T ). Some networks
(e.g., the brown network in panel A) will require very little energy (schematized here with a smaller, solid colored arrow) to
reach patterns that support motor imagery, while other networks (e.g., the pink network in panel B ) will have small energy
requirement to reach patterns of activity that support attention. (C) The relationship between learning rate and the difference
in energy required to reach the attention state when the underlying network takes the form of the lowest versus third highest
performance loading subgraphs for empirical data (green) and uniformly phase randomized surrogate data (grey). (D) The
relationship between the learning rate and the energy required to reach the attention state when the underlying network takes
the form of the lowest performance loading subgraph, or when the underlying network takes the form of the third highest
performance loading subgraph.
1, 2, and 3. We found a consistent relationship between
learning rate and the difference in energy required by the
two subgraphs to reach the attentional state when we
changed the magnitude of activation/deactivation (Pear-
son's correlation coefficient r = 0.560, p = 0.0103), as
well as when we changed the neutral state (Pearson's
correlation coefficient r = 0.520, p = 0.0188). However,
we found no significant relationship when removing neg-
ative states (Pearson's correlation coefficient r = 0.350,
p = 0.130), indicating that this result is dependent on our
choice to operationalize deactivation as a negative state
value. After performing these robustness checks, we con-
clude that a selective separation of the third highest and
lowest performance loading subgraphs impacts their abil-
ity to drive the brain to patterns of sustained attention
in the β band in the context of BCI control. This result
is robust to most of our parameter choices, is selective
for biologically observed states, and is not observed in
surrogate data.
DISCUSSION
In this work, we use a minimally constrained decom-
position of dynamic functional connectivity during BCI
learning to investigate which groups of phase locked
brain regions (subgraphs) support BCI control. The per-
formance loading onto these subgraphs favors the the-
ory that dynamic involvement of several subgraphs dur-
ing learning supports successful control, rather than ex-
tremely strong expression of a single subgraph. Addi-
tionally, we find a unique role for the third highest load-
ing subgraph in predicting learning at the population
level. This result shows that learning is not simply pre-
dicted by the subset of edges that has the most simi-
lar temporal expression, but rather that a subnetwork
with a middling range of similarity has the strongest re-
lationship with performance improvement. While the
spatiotemporal distribution of this subgraph was vari-
able across individuals, we did observe some consistencies
at the group level. Spatially, the third highest loading
subgraph showed strong edges between left frontal and
right motor cortices for low frequencies, and left frontal
and left motor cortices for the γ band. Lower frequen-
cies showed stronger connectivity to the ipsilateral (to
imagined movement) motor cortex, suggesting a possible
role in suppression for selective control. This subgraph
also showed the highest expression earlier than the other
ranked subgraphs we investigated, perhaps linking it to
the transition from volitional to automatic control.
We next wished to posit a theory of how these sub-
graphs fit with previously identified neural processes im-
portant for learning, despite their heterogeneity across
subjects. After quantifying the extent to which NMF reg-
ularization removed potentially redundant relationships
between regions (Fig. S1A-B), we suggested that the
regularized pattern of statistical relationships identified
in this subgraph could comprise an avenue through which
brain activity could be modulated via cognitive control
or external input. We then hypothesized that these net-
works would be better suited to modulate activity in
either regions implicated in attention or in motor im-
agery than other subgraphs, and further that individuals
whose networks better modulated activity in these re-
gions would display greater task learning [18]. We chose
to operationalize the "ease of modulation" with a met-
ric from network control theory called optimal control
energy. Optimal control energy quantifies the minimum
input needed to drive the brain from an initial pattern of
activity to a final pattern of activity, while also assuring
that the pattern of activity stays close to the target state
at every point in time. This last constraint assures that
we do not pass through biologically unfeasible patterns
of activity to reach our desired pattern. The notion of
optimal control energy that we use here assumes a par-
ticular linear model of how neural dynamics change given
potential avenues of communication between regions. Im-
portantly, in the supplement (Fig. S1C-D) we show that
our subgraphs predict empirical brain state changes ac-
cording to this model, and that the contribution of each
subgraph to empirical changes in brain state is related
to its temporal expression. Using this model, we did not
find any population differences in optimal control energy
when the simulation was enacted on the third highest
performance loading subgraph compared to the lowest
performance loading subgraph. However, we did find
that the magnitude of this difference predicted learning
in individual subjects. This result was specific to the β
band and to brain regions implicated in attention. Crit-
ically, the relation to learning could not be explained by
the energy of either subgraph alone, was not present in
surrogate data derived from a uniformly phase randomize
null model, and was robust to parameter choices. Over-
all, the observations support our hypothesis that in the
β band the subgraphs we identified that support learn-
ing are well suited to modulate activity in brain regions
associated with attention.
14
A delicate balance of interactions is required for BCI learning
Our initial analysis explored the relationship between
performance loading and learning.
It is important to
note the behavioral difference between performance and
learning: we use the term performance to refer to task
accuracy over time, whereas we use the term learning to
refer to how well a subject is able to increase that accu-
racy. With that distinction in mind, we aimed to better
understand how subgraphs that vary similarly to perfor-
mance (those with high performance loading) relate to
learning. We found that the subgraph with the third
highest performance loading best predicted learning and
that a narrow distribution of performance loading across
all subgraphs was associated with better learning. To-
gether, these two observations are in line with previous
research in motor and spatial learning, which shows that
some brain structures display differential activity during
learning that is independent of performance [71, 72]. Our
work adds to this literature by demonstrating that in ad-
dition to targeted differences in individual brain regions
or networks, a minimally constrained decomposition of
dynamic functional connectivity across the whole brain
reveals that separable processes are most associated with
performance and with learning.
Additionally, we find that BCI learning is not pre-
dicted simply by the processes most strongly associated
with performance and learning individually, but by a dis-
tributed loading across many different subgraphs. This
notion is supported by the modestly predictive role of
standard deviation of loading predicting learning, and
also by the sign of beta value predictors for ranked sub-
graphs. Generally, subgraphs with higher ranked load-
ing were negative predictors, while subgraphs with lower
ranked loading were positive predictors. A wealth of
whole brain connectivity analyses have similarly shown
that the interaction between systems is an important
component of skill learning specifically, and other do-
mains of learning more generally [30, 31, 73, 74]. While
we observed marked interactions between many regions,
the majority were located in the frontal lobe for all fre-
quency bands. Previous work has also demonstrated
changes in frontal-motor [75, 76] and fronto-parietal [77]
connectivity during motor skill learning. In BCI learn-
ing specifically, the strength of white matter connectivity
between frontal and occipital regions predicts control of
motor imagery based BCIs [78]. Additionally, analyses of
this same experiment have shown task related changes in
functional connectivity were spatially diffuse, and found
in frontal, temporal, and occipital regions in the α band
[32], and were strongest in frontal, motor, central, and
parietal regions in the β band. Our results add to these
findings by demonstrating that the most consistent re-
gions that covary in their functional connectivity are in-
teractions between the frontal lobe and other regions.
Our work shows that broad motifs like the dynamic inte-
gration of multiple systems (including cognitive systems
involving the frontal lobe) found in other types of learn-
ing are also important for BCI learning. Additionally,
we add to previous work on BCI learning specifically by
quantifying the structure of covarying subgraphs of con-
nectivity.
BCI learning is heterogenous across individuals
We find population level consistencies in spatial and
temporal properties of ranked subgraphs despite hav-
ing no constraint to assure consistency across individ-
uals. However, we also note that there is a high degree
of variability in both of these measures. The variabil-
ity is mirrored in the subjects' performance, with final
performances varying from 38.1 % to 89.3 %. Our ob-
servations are in line with previous literature demon-
strating variability in subjects' performance and learn-
ing for psychological, cognitive, and neurological predic-
tors [18, 79]. Such pervasive and marked individual dif-
ferences presents a challenge for the use of BCIs clini-
cally [80]. To address this challenge, researchers have
explored ways to optimize BCI features and algorithms
for neurofeedback itself [81, 82] and to identify selection
criteria for BCI based therapies [57, 79]. The results of
our study support the idea that different individuals will
have slightly different neural correlates of both perfor-
mance and learning based on a variety of features such
as demographics [83], spatial manipulation skills [84], re-
lationship with the technology [85], and attention span
[23, 86]. Our findings also highlight the importance of
studying models fit to each individual when searching for
selection criteria for BCI therapies. Here, despite tem-
poral and edge level heterogeneity, our minimally con-
strained, individual specific method of brain connectiv-
ity decomposition revealed a robust predictor of learning
with a theoretical role that aligns well with previous liter-
ature. Further development and expansion of this model
to incorporate resting state neuroimaging data and other
physiological predictors could be a promising direction
for selection of candidates for BCI therapies before train-
ing.
Role of beta oscillations in BCI learning
Prominent theories describing the neural processes
that give rise to cognition and shape our behavior often
involve integration of complex multimodal information
using a combination of top-down predictions (built from
prior experience) and bottomup, sensory-driven represen-
tations of the dynamic world around us [87 -- 89]. These
generalized frameworks, in turn, require the precise co-
ordination of ensemble neural activity both within and
between brain regions. Several theoretical approaches
have examined how these two scales of functional activ-
ity may harmonize to produce the desired behavior [90],
and empirical research has shown that there is consistent
cross-talk between these scales [91]. Within human neu-
15
roimaging work, synchronous oscillations have been crit-
ical to understanding this complex coordination, where
cortico-cortical propagation delays and membrane poten-
tials give rise to observed oscillatory activity in the brain
[92, 93]. Here, we study the time varying connectivity
within α, β, and γ bands. Much like how specialized
functions arise from different brain regions, different nar-
rowband oscillations have been implicated in diverse but
specialized processes, where some generalizable theories
suggest a role for α in disengagement of task irrelevant
areas or a lack of sensory processing [94], β in sustain-
ing the current cognitive state [95] and γ in task active
local cortical computation [96]. Specifically in the con-
text of motor imagery based BCIs, α and β bands have
prominent signatures in motor imagery [97]. Our results
show that only the β bands functional connectivity is well
suited to modulate patterns of activity that support sus-
tained attention (not motor imagery), which is a critical
process for BCI control. While our results are in line with
generalized theories on the role of oscillations in cogni-
tion, the specificity of the β band in our results extends
classic studies that discuss the role of this oscillation in
attention [98] and in maintaining the current cognitive
state [95]. Our results suggest that this maintenance, a
consistent control (or attention to) internally generated
activity, may play a crucial role in longterm BCI use.
Methodological Considerations
NMF Non-negative matrix factorization is a machine
learning technique for separating, in our case, a multi-
modal configuration matrix into a soft-partition of sub-
graphs with time-varying expression. This process has
several advantages, such as being able to link behav-
ioral and neural data, and creating a quantification of
mesoscale structure where brain regions can participate
in multiple functional groups. Nevertheless, the method
also faces several limitations that are common to other
large-scale machine learning techniques. NMF yields a
low rank approximation of a large configuration matrix,
and can sometimes be rank deficient for large number of
subgraphs, for very large datasets, or for datasets with
high covariance. Because of this sensitivity, we were not
able to test our data against independently phase ran-
domized null models.
Spatial Resolution We chose to complete our analyses
in sensor, rather than source space. This choice limits
the anatomical resolution of our data, and therefore the
specificity of the claims that we can make about the spa-
tial distribution of the regions involved. However, source
reconstruction requires many parameter choices and has
potentially confounding effects on estimates of functional
connectivity [99 -- 101]. Additionally, we were not inter-
ested in finer resolution distribution of the identified sub-
graphs, but more in the process of identifying them, in
validating the hypothesis that features of these subgraphs
predicted learning, and in their theoretical functions.
Optimal Control We chose to use tools from network
control theory to quantify the ease with which each net-
work can modulate brain activity. Network control the-
ory relies on several assumptions that should be consid-
ered when interpreting these results. First, the model of
dynamics that we employ is linear and noise free, unlike
the brain [102], but has proven useful in gaining intuitions
about the behavior of nonlinear systems [103, 104]. How-
ever, we still sought to quantify the ability of this linear
model to explain empirical changes in brain state. Specif-
ically, we asked two questions: (1) do the regularized
subgraphs used in our analyses have the ability to predict
state transitions, and do they do so better than randomly
rewired networks, and (2) is the contribution of each sub-
graph to explaining a given state transition proportional
to its temporal expression, and is it more proportional
than a different subgraph's temporal expression? To eval-
uate these questions, we generated brain states for every
trial (band specific power at each channel) and simulated
Eq. 5 (see Supplement). Regarding the similarity of pre-
dicted and empirical state transitions, we find modest
correlation values (mean Pearson's r = 0.25) that are
significantly greater than the correlations observed from
randomized networks. Similarly for our second question,
we found small but positive correlations between the con-
tribution of each subgraph to a given transition and its
temporal expression (mean Pearson's r = 0.03), which
was also significantly greater than correlations to tempo-
ral expression from mismatched subgraphs. While it is
unsurprising that our linear model did not fully capture
neural dynamics across a three second trial, it is worth
considering extensions that can maximize this similar-
ity for future analyses investigating how connections be-
tween regions facilitate changes to activity. One option
is to use effective connectivity -- such as autoregressive
models [69, 105] -- that solve for a network of connec-
tions that best predicts the evolution of brain states in
time. However, effective connectivity matrices are of-
ten sparse, and therefore not well suited to NMF matrix
decomposition used in the present work. Alternatively,
one could use non-linear models of dynamics [106] and
non-linear control theory [107] to capture a wider range
of dynamic behaviors, although non-linear control does
not currently support the same scope of tools available
for linear control theory. Lastly, future work could use
functional approximation[108] in order to identify a set
of simple basis functions that well approximate the data.
If a sparse approximation can be found, it supports the
idea that the underlying non-linear dynamics can be cap-
tured with linear combinations of these basis functions,
and therefore are suitable to be modeled with simplified
linear models.
Additionally, network control is typically applied to
time invariant, structural connections that have a clear
role as an avenue along which brain activity can prop-
agate. Here we used functional connectivity (weighted
phase locking) which is a statistical relationship that (1)
does not imply the presence of a physical connection and
16
(2) is not time invariant. Due to (1), our original func-
tional connectivity matrix can have large values between
two regions that are not directly connected, but might
both connect to the same region. This situation would
lead to a triangle composed of three connections in a
functional connectivity matrix where in reality there are
only two connections. However, the regularization ap-
plied by the NMF algorithm mitigates this concern in a
manner that is similar to the regularization applied in
effective connectivity metrics [69, 70]. We also explicitly
quantify the effect of regularization on triangles in our
subgraphs and find a dramatic reduction from the origi-
nal functional connectivity (Fig. S1A-B). This quantifi-
cation, along with the two validations discussed above,
show that our model is a suitable way to evaluate the role
of regularized subgraphs in modulating different patterns
of activity.
In relation to (2), we note that functional
connectivity in not time-invariant, unlike the state ma-
trix more commonly employed in linear control models.
However, it is important to note that NMF identifies sub-
graphs that are separable from their temporal expression,
and that we expect that the hypothesized role in control
would only be prominent when the subgraph was highly
expressed.
Conclusion
In conclusion, we use a minimally constrained method
of matrix decomposition that is specific to each human
participant to investigate the dynamic neural networks
that support BCI learning. We find that the subgraphs
that most tightly mirror performance are not the same
subgraphs that most strongly support learning. Addi-
tionally, we find that the interaction between many differ-
ent neural processes is important for BCI learning. While
the subgraphs identified are heterogeneous (as is subject
performance), we find consistent involvement of frontal
and motor cortices in subgraphs that support learning.
We also observe differential temporal expression amongst
subgraphs, and perhaps most notably that the subgraphs
that vary more similarly with performance reach their
highest expression later in learning. Lastly, we test the
hypothesis that subgraphs that support learning are bet-
ter suited to modulate activity in brain regions important
for attention than other subgraphs. We find evidence to
support this hypothesis in the β band specifically, ulti-
mately suggesting that the separation of processes for
maintaining attention is important for successful BCI
learning. Our results align with prior work from dynamic
functional connectivity in other types of skill learning,
and also highlight a method for identifying individual
predictors of successful BCI control with theoretical sup-
port.
17
ACKNOWLEDGEMENTS
We would like to thank Ankit N. Khambhati for help-
ful discussions regarding the application of NMF to
BCI learning, and Marcelo G. Mattar for helpful dis-
cussions regarding joint decomposition of brain and be-
havior. We would also thank Pragya Srivastava, Ja-
son Kim, Lia Papadopoulos, and Eli Cornblath for their
helpful comments on the manuscript. D.S.B. and J.S.
acknowledge support from the John D. and Catherine
T. MacArthur Foundation, the Alfred P. Sloan Foun-
dation, the ISI Foundation, the Paul Allen Founda-
tion, the Army Research Laboratory (W911NF-10-2-
0022), the Army Research Office (Bassett-W911NF-14-
1-0679, Grafton-W911NF-16-1-0474, DCIST- W911NF-
17-2-0181), the Office of Naval Research, the National
Institute of Mental Health (2-R01-DC-009209-11, R01
MH112847, R01-MH107235, R21-M MH-106799), the
National Institute of Child Health and Human Devel-
opment (1R01HD086888-01), National Institute of Neu-
rological Disorders and Stroke (R01 NS099348), the Na-
tional Science Foundation (BCS-1441502, BCS-1430087,
NSF PHY-1554488 and BCS-1631550), and French pro-
gram Investissements davenir ANR-10-IAIHU-06; AN-
RNIH CRCNS ANR-15-NEUC-0006-02. The content is
solely the responsibility of the authors and does not nec-
essarily represent the official views of any of the funding
agencies. 4
[1] R. Sitaram, T. Ros, L. Stoeckel,
S. Haller,
F. Scharnowski,
J. Lewis-Peacock, N. Weiskopf,
M. L. Blefari, M. Rana, E. Oblak, N. Birbaumer, and
J. Sulzer, Nature Reviews Neuroscience 18, 86 (2017).
(1969),
[2] E.
Science
163,
E.
Fetz,
955
http://science.sciencemag.org/content/163/3870/955.full.pdf.
Biological Psychology 89, 80 (2012).
[20] A. Bamdadian, C. Guan, K. K. Ang, and J. Xu, Journal
of Neuroscience Methods 235, 138 (2014).
[21] A. Guillot, C. Collet, V. A. Nguyen, F. Malouin,
and J. Doyon, NeuroImage 41, 1471
C. Richards,
(2008).
[3] M. D. Sacchet, J. Mellinger, R. Sitaram, C. Braun,
N. Birbaumer, and E. Fetz, Frontiers in Neuroscience
6, 1 (2012), arXiv:15334406.
[4] B. Graimann, B. Allison,
and G. Pfurtscheller,
, 1
(2010), arXiv:arXiv:1011.1669v3.
[5] K. A. Moxon and G. Foffani, Neuron 86, 55 (2015).
[6] D. S. Bassett and O. Sporns, Nature Neuroscience 20,
353 (2017), arXiv:0106096v1 [arXiv:cond-mat].
[7] M. Reiner, R. Rozengurt, and A. Barnea, Biological
Psychology 95, 45 (2014).
[22] S. Halder, D. Agorastos, R. Veit, E. M. Hammer,
S. Lee, B. Varkuti, M. Bogdan, W. Rosenstiel, N. Bir-
baumer, and A. Kubler, NeuroImage 55, 1779 (2011),
arXiv:j.neuroimage.2011.01.021. [10.1016].
[23] M. Grosse-Wentrup, B. Scholkopf, and J. Hill, (2011),
10.1016/j.neuroimage.2010.04.265.
[24] J. Frey, C. Muhl, F. Lotte, and M. Hachet, Proceedings
of the International Conference on Physiological Com-
puting Systems , 2019 (2013), arXiv:1311.2222.
[25] F. De Vico Fallani and D. S. Bassett, Physics of Life
[8] T. Ros, M. A. M. Munneke, L. A. Parkinson, and J. H.
Reviews 1, 6 (2018), arXiv:1807.05616.
Gruzelier, Biological Psychology 95, 54 (2014).
[9] V. P. Buch, A. G. Richardson, C. Brandon, J. Stiso,
M. N. Khattak, D. S. Bassett, and T. H. Lucas, Fron-
tiers in Neuroscience 12, 790 (2018).
[10] S. Shahid, R. Kumar Sinha, and G. Prasad,
(2010),
10.1186/1471-2202-11-S1-P127.
[11] T. Ros, B. J. Baars, R. A. Lanius, and P. Vuilleumier,
Frontiers in Human Neuroscience 8 (2014), 10.3389/fn-
hum.2014.01008.
[12] N. Lofthouse, L. E. Arnold, S. Hersch, E. Hurt, and
(2012),
R. DeBeus, Journal of Attention Disorders
10.1177/1087054711427530.
[13] H. Gevensleben, B. Holl, B. Albrecht, D. Schlamp,
O. Kratz, P. Studer, A. Rothenberger, G. H. Moll, and
H. Heinrich, 10.1007/s00787-010-0109-5.
[14] R. T. Thibault, M. Lifshitz, and A. Raz, Cortex 74,
247 (2016).
[15] M. Hamedi, S. H. Salleh, and A. M. Noor, Neural Com-
putation 28, 999 (2016), arXiv:1309.2848v1.
[16] M. Ahn and S. C. Jun, Journal of Neuroscience Methods
243, 103 (2015).
[17] E. A. Curran and M. J. Stokes, Brain and Cognition 51,
326 (2003).
[18] C. Jeunet, B. N'kaoua, and F. Lotte, Progress in Brain
Research 228 (2016), 10.1016/bs.pbr.2016.04.002.
[19] E. M. Hammer, S. Halder, B. Blankertz, C. Sannelli,
T. Dickhaus, S. Kleih, K. R. Muller, and A. Kubler,
[26] M. E. J. Newman, Networks: An Introduction (2010).
[27] E. Bullmore and O. Sporns, Nature Neuroscience
(2009), 10.1038/nrn2575.
[28] D. S. Bassett, P. Zurn, and J. I. Gold, Nature Reviews
Neuroscience , 1 (2018).
[29] F. De Vico Fallani,
J. Richiardi, M. Chavez,
and S. Achard, Philosophical Transactions of
the
Royal Society B: Biological Sciences 369 (2014),
10.1098/rstb.2013.0521.
[30] D. S. Bassett, M. Yang, N. F. Wymbs,
and
S. T. Grafton, Nature Neuroscience 18, 744 (2015),
arXiv:1403.6034.
[31] D. S. Bassett and M. G. Mattar, Trends in Cognitive
Sciences 21, 250 (2017).
[32] M.-C. Corsi, M. Chavez, D. Schwartz, N. George,
L. Hugueville, A. E. Khan, S. Dupont, D. S. Bassett,
and F. De Vico Fallani, (2018), 10.1101/487074.
[33] A. N. Khambhati, A. E. Sizemore, R. F. Betzel, and
D. S. Bassett, "Modeling and interpreting mesoscale
network dynamics," (2018).
[34] M. Fukushima, R. F. Betzel, Y. He, M. A. de Reus,
M. P. van den Heuvel, X. N. Zuo,
and O. Sporns,
"Fluctuations between high- and low-modularity topol-
ogy in time-resolved functional connectivity," (2018),
arXiv:1511.06427.
[35] D. S. Bassett, N. F. Wymbs, M. A. Porter, P. J. Mucha,
and S. T. Grafton, Proceedings of
J. M. Carlson,
the National Academy of Sciences 108, 7641 (2010),
arXiv:1010.3775.
[36] M. Pedersen, A. Zalesky, A. Omidvarnia, and G. D.
[57] C. Jeunet, B. Nkaoua, S. Subramanian, M. Ha-
and F. Lotte, PLoS ONE 10, 1 (2015),
chet,
arXiv:journal.pone.0143962 [10.1371].
Jackson, bioRxiv (2018), 10.1073/pnas.1814785115.
[58] J. Kim, Y. He, and H. Park, Journal of Global Opti-
18
[37] J. M. Shine, M. Breakspear, P. T. Bell, K. E. Martens,
R. Shine, O. Koyejo, O. Sporns, and R. A. Poldrack,
NATuRE NEuROSciENcE -- 22, 10.1038/s41593-018-
0312-0.
[38] U. Braun, A. Schafer, D. S. Bassett, F. Rausch,
J. I. Schweiger, E. Bilek, S. Erk, N. Romanczuk-
Seiferth, O. Grimm, L. S. Geiger, L. Haddad, K. Otto,
S. Mohnke, A. Heinz, M. Zink, H. Walter, E. Schwarz,
A. Meyer-Lindenberg, and H. Tost, Proceedings of the
National Academy of Sciences 113, 12568 (2016).
[39] R. T. Gerraty, J. Y. Davidow, K. Foerde, A. Galvan,
D. S. Bassett, and D. Shohamy, J Neurosci 2084, 17
(2018).
[40] A. N. Khambhati, M. G. Mattar, N. F. Wymbs, S. T.
and D. S. Bassett, NeuroImage 166, 385
Grafton,
(2018).
[41] L. R. Chai, A. N. Khambhati, R. Ciric, T. M. Moore,
R. C. Gur, R. E. Gur, T. D. Satterthwaite, and D. S.
Bassett, 10.1162/netn a 00001.
[42] A. N. Khambhati, D. S. Bassett, B. S. Oommen, S. H.
Chen, T. H. Lucas, K. A. Davis, and B. Litt, Eneuro
4, ENEURO.0091 (2017).
[43] J.-L. Chen, T. Ros, and J. H. Gruzelier, Human Brain
Mapping 34, 852 (2013).
[44] R. Monti, R. Lorenz, P. Hellyer, R. Leech, C. Anag-
nostopoulos,
and G. Montana, Graph embeddings of
dynamic functional connectivity reveal discriminative
patterns of task engagement in HCP data, Tech. Rep.
(2015) arXiv:1506.05219v1.
[45] D. D. Lee and H. S. Seung, Nature 401, 788 (1999),
arXiv:arXiv:1408.1149.
[46] S. Gu, R. F. Betzel, M. G. Mattar, M. Cieslak, P. R.
Delio, S. T. Grafton, F. Pasqualetti, and D. S. Bassett,
NeuroImage 148, 305 (2017), arXiv:1607.01706.
[47] G. Schalk, D. J. McFarland, T. Hinterberger, N. Bir-
baumer,
and J. R. Wolpaw, IEEE Transactions on
Biomedical Engineering 51, 1034 (2004), arXiv:1741-
2560/2/4/008 [10.1088].
[48] C.
T.
Butts,
Science
325,
414
(2009),
arXiv:arXiv:1010.0725v1.
[49] P. Garc´es, D. L´opez-Sanz, F. Maest´u, and E. Pereda,
10.3390/s17122926.
[50] M. Vinck, R. Oostenveld, M. Van Wingerden,
F. Battaglia, and C. M. Pennartz, NeuroImage 55, 1548
(2011), arXiv:0006269v1 [hep-ph].
[51] J.-P. Lachaux, A. Lutz, D. Rudrauf, D. Cosmelli, M. Le
Van Quyen, J. Martinerie, and F. Varela, Estimating
the time-course of coherence between single-trial brain
signals: an introduction to wavelet coherence, Tech.
Rep. (2002).
[52] S. Heitmann and M. Breakspear, Network Neuroscience
, 1 (2018), arXiv:arXiv:1404.2263v1.
[53] J. Theiler, Testing for nonlinearity in timeseries:
the
method of surrogate data, Tech. Rep.
[54] D. J. Bartholomew, International Encyclopedia of Edu-
cation 2, 374 (2010).
[55] P. Comon, P. Comon, I. Component, and P. Comon,
Signal Processing 1684, 287 (2015).
[56] J. Kim and H. Park, SIAM J. Sci. Comput, Tech. Rep.
1 (2011).
mization 58, 285 (2014).
[59] D. Greene, G. Cagney, N. Krogan, and P. Cunningham,
Bioinformatics 24, 1722 (2008).
[60] J. A. Roberts, A. Perry, G. Roberts, P. B. Mitchell, and
M. Breakspear, NeuroImage 145, 118 (2017).
[61] F. Tadel, S. Baillet, J. C. Mosher, D. Pantazis, and
R. M. Leahy, Intell. Neuroscience 2011, 8:1 (2011).
[62] S. Gu, R. F. Betzel, M. G. Mattar, M. Cieslak, P. R.
Delio, S. T. Grafton, F. Pasqualetti, and D. S. Bassett,
, 1 (2016), arXiv:1607.01706.
[63] R. F. Betzel, S. Gu, J. D. Medaglia, F. Pasqualetti,
and D. S. Bassett, Scientific Reports 6, 1 (2016),
arXiv:1603.05261.
[64] F. Pasqualetti, S. Zampieri, and F. Bullo, IEEE Trans-
actions on Control of Network Systems 1, 40 (2014).
[65] G. R. Muller-Putz, R. Scherer, C. Brunner, R. Leeb,
and G. Pfurtscheller, Ijbem.Org 10, 52 (2008).
[66] M. D. Fox, R. L. Buckner, M. P. White, M. D. Gre-
icius, and A. Pascual-Leone, Biological Psychiatry 72,
595 (2012).
[67] M. D. Fox, H. Liu, and A. Pascual-Leone, NeuroImage
66, 151 (2013).
[69] Y. Shen, B. Baingana,
[68] F. Ferreri, F. Vecchio, D. Ponzo, P. Pasqualetti, and
P. M. Rossini, Human Brain Mapping 35, 1969 (2014).
and G. B. Giannakis, IEEE
TRANSACTIONS ON MEDICAL IMAGING, Tech.
Rep. (2016) arXiv:1610.06551v1.
[70] A. Das, A. L. Sampson, C. Lainscsek, L. Muller, W. Lin,
J. C. Doyle, S. S. Cash, E. Halgren,
and T. J.
Sejnowski, Neural Computation, Tech. Rep. 3 (2017)
arXiv:1309.2848v1.
[71] A. L. Shelton and J. D. E. Gabrieli, American Psycho-
logical Association 18, 442 (2004).
[72] A. Purushotham, R. D. Seidler, K. Uğurbil,
J. Ashe, S.-G. Kim, and D. Willingham, Science, Tech.
Rep. 5575 (2002).
[73] N. Altman and M. Krzywinski, Nature Methods 14, 545
(2017), arXiv:NIHMS150003.
[74] A. V. Mantzaris, D. S. Bassett, N. F. Wymbs,
E. Estrada, M. A. Porter, P. J. Mucha, S. T. Grafton,
and D. J. Higham, Journal of Complex Networks 1, 83
(2013).
[75] L. Ma, B. Wang, S. Narayana, E. Hazeltine, X. Chen,
D. A. Robin, P. T. Fox, and J. Xiong, Brain Research
(2009), 10.1016/j.brainres.2009.12.073.
[76] D. R. Leff, A. Darzi, K. Shetty, G.-Z. Yang,
and
J. Andreu-Perez, Brain Connectivity 6, 375 (2016).
[77] C. H. Janice Lin, M. C. Chiang, B. J. Knowlton, M. Ia-
coboni, P. Udompholkul, and A. D. Wu, Human Brain
Mapping 34, 1542 (2013).
[78] R. Sitaram, S. Halder, N. Birbaumer, M. Bog-
and A. Kubler,
(2013),
dan, W. Rosenstiel, B. Varkuti,
Frontiers
arXiv:fnhum.2013.00105 [10.3389].
in Human Neuroscience 7,
1
[79] S. Halder, E. M. Hammer, A. Kubler, B. Blankertz,
C. Sannelli, T. Dickhaus, G. Curio, and K.-R. Muller,
NeuroImage 51, 1303 (2010).
[80] C. Brunner, V. Kaiser, B. Z. Allison, G. Pfurtscheller,
G. R. Muller-Putz, and C. Neuper, Journal of Neural
Engineering 7, 026007 (2010).
19
[81] C. Vidaurre, C. Sannelli, K. R. Muller,
and
B. Blankertz, Journal of Neural Engineering 8 (2011),
10.1088/1741-2560/8/2/025009.
[82] A. Kubler, M. Tangermann, K.-R. Muller, J. Hohne,
F. Cincotti, R. Rupp, J. d. R. Millan, D. Mattia,
R. Leeb, and G. Muller-Putz, Proceedings of the IEEE
103, 926 (2015).
[83] J. Schumacher, C. Jeunet, and F. Lotte, in 2015 IEEE
International Conference on Systems, Man, and Cyber-
netics (2015) pp. 3169 -- 3174.
[84] A. Vuckovic and B. A. Osuagwu, Clinical Neurophysi-
ology 124, 1586 (2013).
[85] M. J. Brosnan, Journal of Computer Assisted Learning
14, 223 (1998).
[86] M. Grosse-Wentrup and B. Scholkopf, Journal of Neural
Engineering 9 (2012), 10.1088/1741-2560/9/4/046001.
[87] O. Doehrmann and M. J. Naumer, Brain Research
1242, 136 (2008).
[88] P. Kok, J. F. M. Jehee, and F. P. de Lange, Neuron
75, 265 (2012).
[89] D. Talsma, Frontiers in Integrative Neuroscience 09, 1
(2015).
[90] S. L. Bressler and J. A. Kelso, Trends in cognitive sci-
ences 5, 26 (2001).
[91] G. Buzs´aki, Nature Neuroscience 7, 446 (2004),
arXiv:arXiv:1608.08828v1.
[92] A. M. Bastos, W. M. Usrey, R. A. Adams, G. R. Man-
gun, P. Fries, and K. J. Friston, Neuron 76, 695 (2012).
[93] W. Singer, A. K. Engel, and P. Fries, Nature Reviews.
[96] P. Fries, Annual Review of Neuroscience 32, 209 (2009).
[97] D. J. McFarland, L. A. Miner, T. M. Vaughan, and
J. R. Wolpaw, Brain Topography, Tech. Rep. 3 (2000).
[98] G. Pfurtscheller A'b', C. Neuper, D. Flotzinger, and
M. Pregenzer, Electroencephalography and clinical Neu-
rophysiology, Tech. Rep. (1997).
[99] M. J. Brookes, J. R. Hale, J. M. Zumer, C. M. Steven-
son, S. T. Francis, G. R. Barnes, J. P. Owen, P. G. Mor-
ris, and S. S. Nagarajan, NeuroImage 56, 1082 (2011).
[100] A. Hillebrand, G. R. Barnes, J. L. Bosboom, H. W.
Berendse, and C. J. Stam, NeuroImage 59, 3909 (2012).
[101] G. L. Colclough, M. W. Woolrich, P. K. Tewarie, M. J.
Brookes, A. J. Quinn, and S. M. Smith, NeuroImage
138, 284 (2016).
[102] S. Gu, F. Pasqualetti, M. Cieslak, Q. K. Telesford, A. B.
Yu, A. E. Kahn, J. D. Medaglia, J. M. Vettel, M. B.
Miller, S. T. Grafton, and D. S. Bassett, Nature Com-
munications 6, 1 (2015), arXiv:1406.5197.
[103] S. F. Muldoon, F. Pasqualetti, S. Gu, M. Cieslak,
S. T. Grafton, J. M. Vettel,
and D. S. Bassett,
PLoS Computational Biology 12 (2016), 10.1371/jour-
nal.pcbi.1005076, arXiv:1601.00987.
[104] C. J. Honey, R. Kotter, M. Breakspear, and O. Sporns,
Proceedings of the National Academy of Sciences 104,
10240 (2007).
[105] A. Neumaier and T. Schneider, ACM Trans. Math.
Softw. 27, 27 (2001).
[106] V. Jirsa and H. Haken, Physical Review Letters, Tech.
Rep. 5 (1996).
Neuroscience, Tech. Rep. 10 (2001).
[107] J. G. T. Zanudo, G. Yang, and R. Albert,
(2016),
[94] S. Palva and J. M. Palva, Trends in Neurosciences 30,
10.1073/pnas.1617387114, arXiv:1605.08415.
150 (2007).
[95] A. K. Engel and P. Fries, Current Opinion in Neurobi-
ology 20, 156 (2010).
[108] S. L. Brunton, J. L. Proctor,
and J. N. Kutz, Pro-
ceedings of the National Academy of Sciences 113, 3932
(2016).
|
1605.01270 | 1 | 1605 | 2016-05-04T13:26:30 | Towards tailoring non-invasive brain stimulation using real-time fMRI and Bayesian optimization | [
"q-bio.NC"
] | Non-invasive brain stimulation, such as transcranial alternating current stimulation (tACS) provides a powerful tool to directly modulate brain oscillations that mediate complex cognitive processes. While the body of evidence about the effect of tACS on behavioral and cognitive performance is constantly growing, those studies fail to address the importance of subject- specific stimulation protocols. With this study here, we set the foundation to combine tACS with a recently presented framework that utilizes real-time fRMI and Bayesian optimization in order to identify the most optimal tACS protocol for a given individual. While Bayesian optimization is particularly relevant to such a scenario, its success depends on two fundamental choices: the choice of covariance kernel for the Gaussian process prior as well as the choice of acquisition function that guides the search. Using empirical (functional neuroimaging) as well as simulation data, we identified the squared exponential kernel and the upper confidence bound acquisition function to work best for our problem. These results will be used to inform our upcoming real- time experiments. | q-bio.NC | q-bio |
Towards tailoring non-invasive brain stimulation
using real-time fMRI and Bayesian optimization
Romy Lorenz∗†, Ricardo P Monti∗†, Adam Hampshire†, Yury Koush‡, Christoforos Anagnostopoulos†,
Aldo A Faisal†, David Sharp†, Giovanni Montana†§, Robert Leech† and Ines R Violante†
†Imperial College London, UK, ‡EPFL, Switzerland, §King's College London, UK
∗ These authors contributed equally to this work.
Abstract-Non-invasive brain stimulation, such as transcranial
alternating current stimulation (tACS) provides a powerful tool
to directly modulate brain oscillations that mediate complex
cognitive processes. While the body of evidence about the effect
of tACS on behavioral and cognitive performance is constantly
growing, those studies fail to address the importance of subject-
specific stimulation protocols. With this study here, we set the
foundation to combine tACS with a recently presented framework
that utilizes real-time fRMI and Bayesian optimization in order to
identify the most optimal tACS protocol for a given individual.
While Bayesian optimization is particularly relevant to such a
scenario, its success depends on two fundamental choices: the
choice of covariance kernel for the Gaussian process prior as
well as the choice of acquisition function that guides the search.
Using empirical (functional neuroimaging) as well as simulation
data, we identified the squared exponential kernel and the upper
confidence bound acquisition function to work best for our
problem. These results will be used to inform our upcoming real-
time experiments.
Keywords-Bayesian optimization, fMRI, non-invasive brain
stimulation, transcranial alternating current stimulation
I.
INTRODUCTION
Studies involving non-invasive brain stimulation have re-
ported remarkable changes in cognitive and behavioral perfor-
mance [1]. Among those techniques, transcranial alternating
current stimulation (tACS) is particularly promising as it
directly allows for the modulation of physiologically relevant
brain oscillations that subserve cognitive operations [2]. The
effects of tACS are highly dependent on the frequency and
phase of stimulation [3]. The conventional tACS approach
involves defining the frequency and phase of stimulation ad
hoc and testing them on a cohort of subjects. However, this
approach exhibits two limitations: (1) the brain networks
targeted by the stimulation cannot be verified without si-
multaneous fMRI; (2) those stimulation parameters may vary
across subjects due to difference in anatomy or heterogeneity
in disease. Yet,
there is a combinatorial explosion in the
biologically plausible range of stimulation frequencies (1-100
Hz) and phases (0-359◦), resulting in up to thousands of
possibilities. Identifying the optimal stimulation protocol for a
given individual is like "finding a needle in a haystack". There-
fore, using conventional methodology makes tailoring tACS to
an individual highly unfeasible. To address this fundamental
challenge in the clinical use of tACS, we aim to combine tACS
with a recently presented framework that behavior real-time
functional magnetic resonance imaging (fMRI) and Bayesian
optimization: The Automatic Neuroscientist [4]. Employing
Fig. 1. High-level overview of the Automatic Neuroscientist combined with
tACS. (1) The experiment starts with applying a random combination of
tACS parameters to the subject in the scanner. (2) Whole-brain functional
images are acquired and analyzed in real-time in response to the block of
stimulation. (3) Information about the current brain state is extracted and (4)
compared to the pre-defined target brain state. This result is then fed into the
Bayesian optimization approach. (5) Based on this, the algorithm chooses a
new combination of tACS parameters that optimizes for the target brain state.
This closed-loop cycle continues until the optimal tACS parameters are found.
such a framework allows us to start with a target brain state
and find a set of tACS parameters (frequency and phase of
stimulation) for each individual that maximally activates it
(Fig. 1).
Bayesian optimization is particularly relevant to such a
problem as it provides a powerful strategy for finding maxima
of objective functions that are expensive to evaluate and might
contain noisy measurements (both criteria are met with func-
tional neuroimaging data). Moreover, Bayesian optimization is
suited in situations where we do not have an analytical expres-
sion of the objective function nor can make formal statements
regarding its properties [5], [6]. However, the implementation
of Bayesian optimization requires several fundamental choices
that are paramount to the success of the technique; such as
(1) the choice of covariance function for the Gaussian process
(GP) prior and (2) the choice of acquisition function which
determines the manner in which the space of parameters is
explored.
The aim of this study is to objectively compare the per-
formance of a variety of distinct kernel functions (resulting in
distinct GP priors) and acquisitions functions. Offline fMRI
data from eight healthy volunteers receiving blocks of non-
invasive tACS stimulation over left frontoparietal brain regions
as well as simulation data were employed. The results obtained
will be used to propose a combination of kernel and acquisition
functions for upcoming real-time experiments.
II. METHODS
The objective of this work is to obtain a better under-
standing of the relationship between various combinations of
tACS parameters (i.e., frequency and phase) and each subject's
neural response. In the framework of Bayesian optimization,
such a relationship is summarized by a latent objective function
which we wish to infer. With this in mind, a small neuroimag-
ing study was conducted where healthy participants received
tACS consisting of various combinations of frequency and
phase. This data was subsequently employed to perform model
selection on the GP covariance function. In addition to this,
simulation analyses were carried out to compare two popular
acquisition functions.
A. Empirical Data
1) Subjects and Experimental Design: Eight healthy volun-
teers (5 female, 24.75 ± 3.49 years) took part in our study. The
study was approved by the Hammersmith Hospital (London,
UK) research ethics committee. Each participant performed
three runs of a switch task [7] in the MR scanner. Each run
consisted of task blocks (36 s) interleaved with rest blocks
(24 s). During task blocks, tACS was applied via a pair of
MR-compatible conductive rubber electrodes over left frontal
(F3) and left parietal (P3) regions (as determined by the
International 10-20 EEG system). Both return electrodes were
placed on the ipsilateral shoulder. The current was fixed to
1 mA (peak-to peak). The experiment parameter space was
limited to ten different frequencies (0,1,5,8,12,16,20,26,40 and
80 Hz) and five different phases (0, 60, 90, 180 and 270◦),
resulting in 50 different frequency-and-phase combinations
(see Fig. 2). For each run and each subject, we tested a subset
of 14 identical frequency-and-phase combinations (grey shaded
squares in Fig. 2). The order of those frequency-and-phase
combinations was random for each run in order to eliminate
any order effect of the stimulation on the subject's neural
response. The condition without any stimulation (0 Hz with
0◦) was the only condition that was repeated two (n=4) to
four (n=4) times within each run.
2) Target brain state extraction: Whole-brain coverage
images were acquired by a Siemens Verio 3T scanner using
an EPI sequence (TR: 1.5s). Data was minimally preprocessed
by applying motion-correction, high-pass filtering (100s) and
spatial smoothing (5mm FWHM Gaussian kernel). We than ex-
tracted 20 regions of interest (ROIs) that have been previously
shown to predict trial-by-trial performance in task switching
(for ROI definition we created 8mm spheres around the peak
coordinates reported in [8]). The extracted timecourses were
further cleaned by removing high-frequency noise and large
signal spikes using a modified Kalman filter [9]. We then
assessed changes in functional connectivity between those
regions for blocks of tACS using psychophysiological inter-
action (PPI) analyses [10] (whilst also including six motion
parameters as confounds). For the analyses presented here,
Exhaustive two-dimensional
Fig. 2.
tACS parameter space with each
dimension corresponding to the frequency or phase of non-invasive brain
stimulation, respectively. Grey shaded squares are combinations sampled in
our (offline) fMRI study.
we selected the two regions that exhibited the highest inter-
run robustness across all subjects (assessed using Spearman's
rank correlation coefficient): the left inferior parietal lobule
and the posterior cingulate gyrus. In order to assess how
likely the pattern of results between those regions could have
occurred by chance, we performed non-parametric permutation
testing (10,000 permutations) and found the result to be highly
significant (z=3.71, p<.001).
B. Bayesian optimization
The underlying intuition behind the method of Lorenz et al.
[4] is that the target brain pattern is a function of experimental
conditions. As such, the authors propose to learn the relation-
ship by modelling the observed brain state as a sample from a
Gaussian process (GP). This facilitates the use of a Bayesian
optimization framework [5], [6] in a closed-loop form where
subjects are presented with an experimental condition and
real-time fMRI provides instantaneous information about the
subject's brain state. Based on this, we can iteratively update
our beliefs about the unknown objective function, captured in
the posterior distribution of the GP. While Lorenz et al. [4]
used simple perceptual stimuli in their original experiment, our
study here involves blocks of non-invasive brain stimulation
with different combinations of frequency and phase. This can
be considered as a far more challenging problem as only a
handful of studies exist investigating the effects of tACS on
the blood oxygenation level dependent (BOLD). Beyond that,
high inter-subject variability can be expected as outlined in the
Introduction.
In contrast to the fMRI setting employed in [4], there is
far less prior knowledge regarding the properties of the latent
objective function which can be leveraged in this context.
Therefore, data was collected over a cohort of eight subjects
(see previous section). Using this empirical data we begin by
studying the choice of kernel function in Section II-B1. Given
a choice of kernel and by carrying out simulation analyses,
the effects of various acquisition functions are subsequently
assessed in Section II-B2.
1) Covariance function: The choice of kernel function in
GP regression is fundamental [11]. The kernel directly spec-
ifies a measure of similarity across various inputs and there-
fore determines the generalization properties of the estimated
model. Model selection in the context of GPs is often posed
as a parameter estimation task whereby the hyper-parameters
for a specific choice of kernel can be learnt [11]. However,
comparing the performance of potentially many distinct kernels
is a challenging problem. Recently, [12], [13] have proposed
a method for automatically searching over the space of kernel
structures in a principled manner.
Briefly, the proposed approach proceeds by considering
compositional structures (either addition or multiplication)
over four base kernels; squared exponential (as employed in
[4]), periodic, linear or rational quadratic. When comparing
various distinct kernel structures the Bayesian Information Cri-
terion (BIC) is employed to score the various models (where
type-2 maximum likelihood estimates of parameters are used).
Initially, each of the base kernels is proposed. Thereafter,
the selected kernel can be extended by either performing
addition or multiplication with any base kernel. While the
original approach described in [12] proposes kernels for each
dimension independently, in this work both dimensions were
studied simultaneously (thus two dimensional kernels were
proposed at each step). This decision is based on the hypothesis
that each dimension (either frequency or phase) would share a
similar relationship with the response, therefore leading to the
selection of similar (if not the same) kernel.
2) Acquisition function: Of equal importance for the suc-
cess of the Bayesian optimization is the choice of acquisition
function. The role of the acquisition function is to iteratively
propose the experimental condition to be presented to the
subject. It therefore serves to guide the search over all com-
binations of parameters (in our case frequency and phase)
and must implicitly balance exploration with exploitation. In
this work we considered the performance of two popular
acquisition functions:
the expected improvement (EI) and
upper confidence bound (GP-UCB) acquisition function [5]1.
The two acquisition functions differ in their trade-off between
exploration and exploitation of the experimental parameters
space, giving rise to distinct sampling behaviors over time.
Informally, the EI acquisition function can be seen as trying
to maximize the expected improvement over the current best.
While the GP-UCB also favors the selection of points with
high mean value (similar to the EI acquisition function), it also
favors points with high variance (i.e., regions worth to explore).
In certain settings,
in a more explorative
behavior of the GP-UCB compared to the more "greedy" EI
acquisition function. For algorithmic details, please refer to
[5].
this will result
While the choice of kernel function can be guided by model
selection and information theoretic measures, the comparison
of multiple acquisition functions requires explicit knowledge
of the underlying objective function. As a result, a simulation
study was employed to the benchmark the performance of the
two aforementioned acquisition functions.
A complex and multimodal objective function was pro-
posed for the simulation study (see Fig. 3a). This was moti-
vated by our empirical results, showing multiple optima and
minima for each subject (not shown). As the presence of
non-neural noise is well documented for fMRI experiments,
we also studied how different levels of contrast-to-noise-ratio
1 The probability of improvement (PI) acquisition function was not included
as it is typically seen to be over exploitative.
(CNR) affected the Bayesian optimization. We repeated our
simulations analyses with CNR values ranging from 0.1 and
1.8 (typically reported CNR values in fMRI literature vary
between 0.5 and 1.8 [14]). For each CNR value tested, we
ran 100 simulations. The maximum number of iterations was
set to 100 with the first five (randomly selected) observations
serving as burn-in phase. Thereafter, at each iteration, a new
frequency-and-phase combination was proposed by maximiz-
ing the respective acquisition function and sampled by the
Bayesian optimization algorithm in the next iteration.
As a measure of accuracy we computed spatial correlation
between the algorithm's predictions for the whole parameter
space and the "ground truth" objective function. Considering
our underlying motivation to gain a holistic understanding of
the whole parameter space (learning about maxima and min-
ima), spatial correlations seemed most appropriate to capture
similarity across the whole space. This procedure was identical
to the one reported in [4].
III. RESULTS
A. Covariance kernel selection
In this work we follow [12], [13], and look to compare
several distinct kernels via the use of BIC. Furthermore, we
also employ a greedy search algorithm whereby we begin with
a base kernel and iteratively extend this kernel.
linear and Matern ( 3
There are two significant differences in the approaches
described in this work to those of [12], [13]. The first is our
choice of base kernels; in this work the squared exponential,
periodic,
2 ) where employed as base
kernels. The second difference is that two-dimensional kernels
were proposed at each iteration. This choice was motivated by
an assumption that the relationship between parameter (fre-
quency or phase) and the response shared a similar functional
form. Finally, a minor difference between this work and that of
[13] is that change-point functions were not considered here.
Following [12], we begin by proposing each of the base
kernels and calculating the associated BIC scores. In this first
step the squared exponential kernel was selected. We then
proceeded to consider compositional structures. Each of these
structures was composed by either adding or multiplying the
squared exponential kernel with each of the base kernels.
For each compositional kernel, the parameters were estimated
using the previous parameters as warm starts. Again, BIC
scores were employed to score each of the compositional
models. At this stage, there was no reduction in BIC scores
resulting in the choice of the squared exponential kernel.
B. Acquisition function selection
Simulation analyses were carried out to assess the perfor-
mance of the EI and GP-UCB acquisition functions. A range of
CRN values were employed in order to recreate the properties
of fMRI data as well as study the performance of each of
the acquisition functions in scenarios with low signal-to-noise
ratio. Throughout this simulation the underlying GP model was
maintained constant, allowing us to attribute any differences in
performance to the acquisition functions.
For each CNR value, 100 simulations where performed.
Each simulation allowed the proposed Bayesian optimization
Fig. 3. Results of simulation analyses. (a) Modeled objective function used for simulations. Mean ± SEM (shaded areas) spatial correlation between predicted
and modeled parameter space for (b) EI and (c) GP-UCB acquisition function. As the first five iterations were used as a burn-in, they are not depicted here
(gray dashed line). Simulations were performed for 10 different CNR values, ranging from 0.1 (bright blue) to 1.8 (dark blue).
algorithm to explore the parameter space over 100 iterations.
The spatial correlation between the predictive posterior under
the GP model and the true objective function was calculated
at each iteration.
The mean ± SEM spatial correlation results are depicted in
Fig. 3b for the EI and in Fig. 3c for the GP-UCB acquisition
function. Not surprisingly, we note that both acquisition func-
tions fail in the presence of too much noise (for CNR values
< 0.2). The results also indicate that the GP-UCB acquisition
function clearly outperforms the EI acquisition function for
typical CNR values reported in the literature (0.5-1.8).
IV. DISCUSSION AND FUTURE WORK
The results presented here serve as an initial exploration
into the choices of kernel and acquisition functions which can
be used in the context of real-time optimization for tACS.
Our results propose the use of a squared exponential kernel
function in combination with a GP-UCB acquisition function.
In a series of simulations, the latter was shown to capture
the whole parameter space (as identified by using spatial
correlations) when compared to the EI acquisition function.
This result was consistent for a range of CNR values tested.
These findings serve as an important basis for upcoming real-
time experiments. An avenue of future work is the definition
of appropriate stopping criteria to determine convergence [15].
This is particularly important
in our context due to high
scanning costs, limited attentional capacities of subjects and
neural habituation to the stimulation.
To our knowledge, this is the first study to propose a
framework that allows tailoring non-invasive brain stimulation
to an individual. The next step consists of applying this
technology to patients suffering from traumatic brain injury; a
condition that disrupts frontoparietal brain networks and hence
results in cognitive impairments [16]. While closed-loop deep
brain stimulation has already been shown to outperform the
conventional approach in Parkinson's patients [17] we envision
that our framework will advance personalized treatment by
means of non-invasive brain stimulation. This will be of
particular importance for neurological and psychiatric deficits
that are diffuse and widely heterogeneous in their origin.
ACKNOWLEDGMENT
This research was supported by the NIHR Imperial BRC
and the Wellcome Trust (103045/Z/13/Z).
REFERENCES
[1] MF Kuo, MA Nitsche: Effects of transcranial electrical stimulation on cognition,
Clin. EEG Neurosci.,43(3), 92–199 (2012)
[2] CS Herrmann, S Rach, T Neuling, D Strueber: Transcranial alternating current
stimulation: a review of the underlying mechanisms and modulation of cognitive
processes, Frontiers Human Neuroscience, 7:279 (2013)
[3] A Antal, W Paulus: Transcranial alternating current stimulation (tACS), Frontiers
Human Neuroscience, 7:317 (2013)
[4] R Lorenz, RP Monti, IR Violante et al.: The Automatic Neuroscientist: A
framework for optimizing experimental design with closed-loop real-time fMRI.
NeuroImage,129, 320–334 (2016)
[5] E Brochu, VM Cora, N de Freitas: A tutorial on Bayesian optimization of
expensive cost functions, with application to active user modeling and hierarchical
reinforcement learning. arXiv:1012.2599 (2010)
[6] B Shahriari, K Swersky, Z Wang, RP Adams, N de Freitas: Taking the Human
Out of the Loop: A Review of Bayesian Optimization, Proceedings of the IEEE,
104(1), 148–175 (2016)
[7] A Dove, S Pollmann, T Schubert, CJ Wiggins, DY von Cramon: Prefrontal cortex
activation in task switching: an event-related fMRI study. Cognitive Brain Research,
9(1): 103–109 (2000)
[8] AB Leber, NB Turk-Browne, MM Chun: Neural predictors of moment-to-moment
fluctuations in cognitive flexibility, PNAS, 105(36), 13592–13597 (2008)
[9] Y Koush, M Zvyagintsev, M Dyck, KA Mathiak, K Mathiak: Signal quality and
Bayesian signal processing in neurofeedback based on real-time fMRI, NeuroImage,
59(1) (2012)
[10] KJ Friston: Functional and Effective Connectivity: A Review, Brain Connectivity,
1(1),13–36 (2011)
[11] CE Rasmussen, CK Williams: Gaussian processes for machine learning. MIT Press
(2006)
[12] D Duvenaud, JR Lloyd, R Grosse, JB Tenenbaum, Z Ghahramani.: Structure
discovery in nonparametric regression through compositional kernel search, ICML
(2013)
JT Lloyd, D Duvenaud, R Grosse, JB Tenenbaum, Z Ghahramani.: Automatic
construction and natural-language description of nonparametric regression models,
Association for the Advancement of Artificial Intelligence (2014)
[13]
[14] M Welvaert, Y Rosseel: On the definition of signal-to-noise ratio and contrast-to-
noise ratio for fMRI data, PLoS One, 8, doi: 10.1371/journal.pone.0077089 (2013)
[15] R Lorenz, RP Monti, IR Violante et al.: Stopping criteria for boosting au-
tomatic experimental design using real-time fMRI with Bayesian optimization,
arXiv:1511.07827 (2015)
SR Jilka, G Scott, T Ham, A Pickering, V Bonnelle, RM Braga, R Leech, DJ
Sharp: Damage to the Salience Network and Interactions with the Default Mode
Network, J. Neurosci., 34(33), 0798-10807 (2014)
S Little, A Pogosyan, S Neal et al.:Adaptive deep brain stimulation in advanced
Parkinson disease, Ann. Neurol., 74(3), 449–457 (2013)
[16]
[17]
|
1503.07610 | 1 | 1503 | 2015-03-26T03:44:43 | Eye-Movement Control During the Reading of Chinese: An Analysis Using the Landolt-C Paradigm | [
"q-bio.NC"
] | Participants in an eye-movement experiment performed a modified version of the Landolt-C paradigm (Williams & Pollatsek, 2007) in which they searched for target squares embedded in linear arrays of spatially contiguous "words" (i.e., short sequences of squares having missing segments of variable size and orientation). Although the distributions of single- and first-of-multiple fixation locations replicated previous patterns suggesting saccade targeting (e.g., Yan, Kliegl, Richter, Nuthmann, & Shu, 2010), the distribution of all forward fixation locations was uniform, suggesting the absence of specific saccade targets. Furthermore, properties of the "words" (e.g., gap size) also influenced fixation durations and forward saccade length, suggesting that on-going processing affects decisions about when and where (i.e., how far) to move the eyes. The theoretical implications of these results for existing and future accounts of eye-movement control are discussed. | q-bio.NC | q-bio | Eye-Movement Control During the Reading of Chinese: An Analysis
Using the Landolt-C Paradigm
Yanping Liu1, Erik D. Reichle2, & Ren Huang3
1Key Laboratory of Behavioral Science
Institute of Psychology, Chinese Academy of Sciences
2School of Psychology, University of Southampton, UK
3Department of Psychology, Sun Yat-Sen University, China
Author note
This research was supported by two grants from the China Postdoctoral Science
Foundation (2013M541073 & 2014T70132) awarded to the first author, and by a
Grant HD075800 from the National Institutes of Health awarded to the second author.
Correspondence should be addressed to Yanping Liu, 16 Lincui Road, Key Laboratory
of Behavioral Science, Institute of Psychology, Chinese Academy of Sciences,
Beijing, China. Email: [email protected].
1
Abstract
Participants in an eye-movement experiment performed a modified version of the
Landolt-C paradigm (Williams & Pollatsek, 2007) in which they searched for target
squares embedded in linear arrays of spatially contiguous “words” (i.e., short
sequences of squares having missing segments of variable size and orientation).
Although the distributions of single- and first-of-multiple fixation locations replicated
previous patterns suggesting saccade targeting (e.g., Yan, Kliegl, Richter, Nuthmann,
& Shu, 2010), the distribution of all forward fixation locations was uniform,
suggesting the absence of specific saccade targets. Furthermore, properties of the
“words” (e.g., gap size) also influenced fixation durations and forward saccade length,
suggesting that on-going processing affects decisions about when and where (i.e., how
far) to move the eyes. The theoretical implications of these results for existing and
future accounts of eye-movement control are discussed.
Keywords: Chinese reading; eye-movement control; Landolt-C paradigm; lexical
processing; reading; saccade targeting
2
Neural systems are extremely adept at exploiting the regularities that exist in the
environment for the purposes of making (near) optimal inferences to guide behavior
(Anderson, 1990). For example, the systems that mediate vision exploit regularities of
objects (e.g., the Gestalt principle of similarity) to represent those objects in a manner
that affords their invariant perception despite variability of both the objects and
viewing angle, partial occlusion of the objects, etc. (Smith, 1988). It should therefore
come as little surprise that complex visual tasks like reading also exploit regularity,
allowing the task to be performed in a (near) optimal manner (Liu & Reichle, 2010;
Liu, Reichle, & Gao, 2013; Reichle & Laurent, 2006; see also Bicknell & Levy,
2012). For example, because low-frequency words require more time to identify than
high-frequency words (Forster & Chambers, 1973; Schilling, Rayner, & Chumbley,
1998), readers of alphabetic languages like English tend to spend more time fixating
the low- than high-frequency words (Inhoff & Rayner, 1986; Kliegl, Nuthmann, &
Engbert, 2006; Rayner, Ashby, Pollatsek, & Reichle, 2004). And in a similar manner,
because words can be identified most rapidly when they are fixated near their center
(O’Regan, 1981; Rayner, 1979; Rayner & Morrison, 1981), readers of alphabetic
languages use the blank spaces between words to direct their eyes to locations near
the center of words (McConkie, Kerr, Reddix, & Zola, 1988; Rayner, Sereno, &
Raney, 1996). Such eye-movement behavior takes advantage of the inherent
regularities of the text to support reading that is rapid, but that also maintains some
overall level of comprehension. What is less clear, however, is whether such
behaviors are universal, generalizing across languages and writing systems that
exhibit less—or perhaps different—patterns of regularity. One prime example that has
been the focus of much recent research is Chinese (for a review, see Zang, Liversedge,
Bai, & Yan, 2011).
3
The Chinese language and writing system has many properties that make it
different from the alphabetical languages that have most often been used in
experiments to understand reading. For example, in contrast to alphabetic writing
systems, Chinese words are not comprised of letters and are not spatially segregated
by blank spaces. Instead, Chinese words consist of 1-4 equally sized box-shaped
characters that are comprised of 1-36 “strokes” (i.e., line segments that originally
corresponding to the brush strokes used to write the characters), with the characters
being arranged in a spatially adjacent linear array (see Figure 1 for an example). The
fact that Chinese words are comprised of a variable number of characters and not
demarcated by blank spaces means that Chinese readers must use their knowledge of
their language to somehow segment character sequences into their corresponding
words (e.g., see Li, Rayner, & Cave, 2009). This task of segmenting characters into
words is intrinsically difficult because there is often ambiguity in how a character
sequences might be segmented, with a given sequence of four characters, for example,
possibly corresponding to two 2-character words or one 4-character word (Hoosain,
1991, 1992; Liu, Li, Lin, & Li, 2013).
----- Figure 1 -----
Because of these properties of written Chinese words, there is some debate about
both the extent to which and how the eye movements of Chinese readers are
influenced by the lexical properties of words (as they clearly are in alphabetic
languages; see Rayner, 1998, 2009). For example, one theoretical “camp” has argued
that characters are more important than words during the reading of Chinese, and that
characters are the basic units that determine both when the eyes move (Chen, 1996;
Chen, Song, Lau, Wong, & Tang, 2003; Chen & Zhou, 1999; Feng, 2008; Hoosain,
1991, 1992) and where the eyes move (Tsai & McConkie, 2003; Yang & McConkie,
4
1999). In contrast, the other theoretical “camp” maintains that, because words have a
psychological reality in the minds of Chinese readers (Cheng, 1981; Li et al., 2009;
Li, Bicknell, Liu, Wei, & Rayner, 2014; Li & Logan, 2008; Li, Gu, Liu, & Rayner,
2013), they also play an important functional role in determining when the eyes move
during the reading of Chinese (Liversedge, Hyönä, & Rayner, 2013; Yan, Tian, Bai, &
Rayner, 2006). However, this second perspective further diverges in their accounts of
how readers of Chinese decide where to move their eyes. One contingent has argued
that the locations of initial fixations on words are a function of whether the words are
successfully segmented in the parafovea, with saccades towards segmented words
being directed towards their centers but saccades towards un-segmented words being
directed towards their beginnings (Yan et al., 2010). In contract, the other contingent
has proposed that there is no “default” saccade target within a word, but that the
length of the saccade moving the eyes into a word is instead influenced by the relative
difficulty of foveal and/or parafoveal processing, with easier processing resulting in
longer forward saccades (Liu, Reichle, & Li, 2014; Wei, Li, & Pollatsek, 2013). Given
these different theoretical perspectives on eye-movement control during the reading of
Chinese, it is clear that—in contrast to alphabetic languages—we still know very little
about the basic mechanisms that determine when and where the eyes move during the
reading of non-alphabetic languages, or more specifically, about if and, if so, how
lexical processing mediates eye-movement control during the reading of Chinese.
One strategy that has been used to examine the influence of lexical processing in
guiding eye movements during the reading of alphabetic languages has involved
comparisons of such eye movements to those that are observed during tasks that do
not require lexical processing but that do entail many of the same visual and
oculomotor constraints, such as the Landolt-C “reading” task (Corbic, Glover, &
5
Radach, 2007; Williams & Pollatsek, 2007; Vanyukov, Warren, Wheeler, & Reichle,
2012; see also Williams, Pollatsek, & Reichle, 2014). In experiments that have used
this task, participants were instructed to scan through linear arrays of Landolt Cs (i.e.,
ring-shaped stimuli having missing segments (i.e., gaps) of variable size and/or
orientations) to search for target stimuli—rings containing no missing gaps. Because
these linear arrays are comprised of Landolt-C stimuli arranged into sequences of
“words” (i.e., clusters of the stimuli separated by blank spaces between clusters), the
task has been used to study saccadic targeting under conditions that resemble those
experienced during the reading of alphabetic languages, minus all of the language-
related behaviors that are normally engaged during word, sentence, and discourse
processing. The key findings from these experiments is that participants seem to treat
the Landolt-C clusters as “words”, directing their eyes towards their centers and
fixating the easier-to-process clusters (i.e., those with larger gaps sizes and/or that
occur more often during the course of the experiment) for shorter durations than the
more difficult-to-process clusters (i.e., those with smaller gap sizes and/or that occur
less often during the experiment).
In the experiment that is reported in this article, the Landolt-C task was modified
so that it more closely resembled the reading of Chinese (see Figure 2), thereby
providing a novel way to examine if the segmentation of “words” (i.e., spatially
adjacent stimuli having similarly sized and orientated gaps) affects fixation durations
and/or saccadic targeting1. That is, our participants were instructed to scan linear
arrays of Landolt-squares, or boxes with a missing segment of varying size and
orientation, and to indicate the number of targets (i.e., squares that do not contain
missing segments; e.g., □). The advantages of this method are that it provides a way
of precisely manipulating both word processing difficulty (i.e., gap size) and size (i.e.,
6
the number of identical spatially adjacent non-target squares) in the absence of
language processing.
The main goal of this experiment was therefore to determine if/how word
segmentation and identification affects eye movements in this task so as to better
inform our understanding of the word-based effects that have been observed during
the actual reading of Chinese text. More specifically, we focus on: (1) if/how fixation
durations on individual words are affected by their identification difficulty; and (2)
if/how saccadic targeting is affected by their segmentation difficulty. If easier-to-
identify words are fixated for shorter durations but saccade targeting is unaffected by
the ease of word segmentation, then this pattern of results would lend further support
to the hypothesis that the processing associated with words influences both when and
where the eyes move during the reading of Chinese (e.g., Liu et al., 2014; Wei et al.,
2013), but that parafoveal word segmentation is not the primary determinant of where
the eyes are moved (cf., Yan et al., 2010).
Method
Participants
Twenty participants (12 males) recruited from Sun Yat-Sen University were paid
for their participation. All participants had normal or corrected-to-normal vision and
were naive to the purpose of the experiment.
Apparatus
Stimuli were displayed on a 23-in. LCD monitor (Samsung SyncMaster2233)
using SR-Research Experiment Builder software. The monitor had a resolution of
1,680 × 1,050 pixels and refresh rate of 120 Hz. Eye movements were recorded using
7
a SR-Research Eyelink 1000 eye tracker (Kanata, ON, Canada) with a spatial
resolution of 0.01° and sampling rate of 1,000 Hz.
Stimuli and Design
The materials were designed to be as similar to Chinese sentences as possible
(e.g., cf., Figures 1 vs. 2). Each “character” was a Landolt-Square (40 pixels × 40
pixels) with a gap that was 2, 4, 6, or 8 pixels in size and that occurred in the left,
right, top, or bottom of the character. Each cluster or “word” was composed of 1, 2, 3,
or 4 characters. Both the gap size and location were held constant within a given
word, but varied between words. The spaces between any two successive characters
were 6 pixels. Each array of words or “sentence” contained 10 randomly selected
words and ranged from 16 to 33 characters in length (M = 25 characters). The
combination of the number of characters per word and the gap size and orientation in
a given word gave 64 unique words. Each word was repeated 40 times across the
experiment. Each sentence had 0, 1, or 2 targets (i.e., characters without gaps) that
could appear with equal probability within any word except the first and last word
within a sentence. Participants were instructed to scan through each sentence and then
indicate via button presses how many targets occurred in the sentence, with one
sentence being displayed per trial. There were four blocks containing 64 trials each,
with the order of blocks being completely random.
----- Figure 2 -----
Procedure
Upon arrival, participants were given task instructions and then seated 63 cm
from the video monitor. A chinrest was used to minimize head movements during the
experiment. Although viewing was binocular, eye-movement data were only collected
8
from right eye. The eye-tracker was calibrated and validated using a 9-dot procedure
at the beginning of each block of trials, with additional calibrations and validations
being conducted as necessary. After that, participants completed 8 practice trials (not
included in our analyses) and then completed the 4 blocks of experimental trials. A
drift-check procedure was performed before each trial; a sentence was displayed after
the participants successfully fixated the white box (1° × 1°) at the location of the first
character in the sentence. Participants were instructed to view a right-bottom region of
the screen to terminate a trial and to then press the number button on the keyboard
corresponding to the number of targets that appeared in the trial. Accuracy feedback
was randomly provided on half of the trials.
Results
Accuracy
The mean overall accuracy of the experiment was 91% (SD = 0.05).
Eye-Movement Results
Fixations on the first and last words in each sentence were removed from our
analyses because the former coincided with the abrupt appearance of the sentence and
the latter coincided with the termination of a trial. Fixations on words containing
targets and immediately preceding or following targets were also removed from our
analyses. Thus, fixations on 64.3% of the total number of words were included in our
analyses. From these data, saccades longer than 3 SDs above the mean for a given
participant were excluded (3.05% of the total saccades).
To determine if/how the processing of words affected eye movements during our
task, our analyses focused on the relationship between various eye-movement
dependent measures (related to when and where to move eyes) and properties of both
9
the fixated word (i.e., word N) and the two spatially adjacent words (i.e., words N-1
and N+1). To do this, we examined the first-pass fixation-location distributions
corresponding to: (1) first fixations; (2) single fixations; (3) first-of-multiple fixations;
and (4) all forward fixations. We also examined the following measures: (5) forward-
saccade length, or the distance between two consecutive forward fixations during
first-pass scanning; (6) first-fixation duration, or the duration of the first fixation on a
word during first-pass scanning; (7) gaze duration, or the sum of all first-pass fixation
durations on a word; and (8) total-viewing time, or the sum of all fixation durations on
a word. We then built linear mixed-effects models (LMMs) for each measure,
specifying participants, items, and random slopes of each predictor (e.g., word
properties) as the random effects so that our reported significance values reflect the
variability of participants, items, and the slopes of the predictors (see also Barr, Levy,
Scheepers, & Tily, 2013). Based on prior results and our priori hypotheses, we also
included practice (i.e., ordinal trial number) and the properties of words N-1, N, and
N+1 (i.e., gap size and number of characters) as fixed-effect factors in our LMMs.
The models were then fitted using the lme4 package (ver. 1.1-7; Bates, Maechler,
Bolker, & Walker, 2014; Pinheiro & Bates, 2000) and p-values were estimated using
lmerTest package (ver. 2.0-20; Kuznetsova, Brockhoff, & Christensen, 2013) in R
(ver. 3.1.2; R Development Core Team, 2015).
Fixation Locations
To determine if parafoveal word segmentation influences saccadic targeting, we
analyzed our various measures of fixation location during first-pass scanning. Figure
3 respectively shows the distributions of initial-, single-, first-of-multiple-, and all
forward fixation locations as a function of the number of characters with the words,
their gap sizes, and practice (i.e., ordinal block number). As shown, the locations of
10
initial fixations (panels a-b) and the first-of-multiple fixations (panels c-d) were more
likely to be near the beginnings of words, whereas single fixations (panels e-f) were
more likely to be located near the centers of words. However, the locations of all
forward fixations (panels g-h) were uniform (all values of χ2 ≤ 1.74, ps ≥ 0.755; also
see Table 1). All of these results are consistent with previous findings that have been
observed with Chinese reading (Li, Liu, & Rayner, 2011; Yan et al., 2010).
----- Figure 3 & Table 1 -----
LMMs were further used to examine if practice and the properties of words N-1,
N, and N+1 affected all forward fixation locations on word N (measured from the left
edge of that word). Unsurprisingly, Table 2 shows that the mean fixation location
across all forward saccades moved further to the right of word N as its length
increased (b = 0.99, SE = 0.01, t = 97.27, p < 0.001). The fact that the mean fixation
location was not influenced by the gap size of word N, the properties of the spatially
adjacent words, or the amount of practice (all values of t ≤ 1.24, all ps ≥ 0.214)
indicates that the uniformity of distributions of forward fixation locations was robust,
which in turn suggests the absence of any (strongly) preferred viewing location on the
words in our task. Importantly, the “preferred viewing locations” which were
seemingly evident in the distributions of first and first-of-multiple fixations (i.e.,
located near the word beginnings) and of the single fixations (i.e., located near the
word centers) were specious because of the arbitrary way in which these fixations
were extracted from the uniform distribution of all forward fixations. We will say
more about this in the Discussion.
----- Table 2 -----
Saccade Length
11
Although the results from fixation-location analyses indicated that properties of
neither the fixated word nor its spatially adjacent words influenced the distribution of
all of the forward fixation locations, it is possible that foveal and/or parafoveal
processing influenced the length of forward saccade. As Table 3 indicates, forward
saccade length increased with the lengths of word N-1 (b = 0.01, SE = 0.01, t = 2.43, p
= 0.015), word N (b = 0.10, SE = 0.01, t = 9.88, p < 0.001), and word N+1 (b = 0.04,
SE = 0.01, t = 5.05, p < 0.001). Forward saccade length also increased with the gap
size of word N-1 (b = 0.01, SE = 0.003, t = 1.72, p = 0.085), word N (b = 0.01, SE =
0.01, t = 2.23, p = 0.037), and word N+1 (b = 0.02, SE = 0.003, t = 5.58, p < 0.001).
These results were entirely consistent with those of previously reported experiments
involving Chinese reading which demonstrated that forward saccade length increases
with both the processing ease (frequency) and length of the fixated word (Li et al.,
2014; Liu et al., 2014; Liu, Li, &Pollatsek, submitted for review; Wei et al., 2013),
and with the processing ease (frequency) and length of upcoming (parafoveal) word
(Li et al., 2014; Liu et al., 2014; Liu et al., submitted for review).
Fixation Durations
----- Table 3 -----
Our analyses have shown that properties of the fixated words and its spatially
adjacent words did not influence the uniformity of all forward fixation location
distributions, but that word properties did influence the length of forward saccades.
This suggests that ongoing processing influences saccade length rather than fixation
locations on words. It is therefore also necessary to examine how word properties and
practice affected ongoing processing.
As Table 4 shows, first-fixation durations decreased with the length of word N (b
= -1.35, SE = 0.67, t = -2.01, p = 0.051) and the gap size of word N (b = -1.16, SE =
12
0.33, t = -3.50, p = 0.002). Conversely, as Table 5 shows, gaze durations increased
with the length of word N (b = 135.70, SE = 10.91, t = 12.44, p < 0.001), but
decreased with the lengths of word N-1 (b = -4.03, SE = 1.21, t = -3.34, p < 0.001)
and word N+1 (b = -6.84, SE = 1.53, t = -4.46, p < 0.001). Similarly, gaze durations
also decreased with increasing gap size of word N (b = -9.75, SE = 0.72, t = -13.47, p
< 0.001).
----- Table 4 & 5 -----
Finally, as Table 6 shows, total-viewing times (like gaze durations) increased with
the length of word N (b = 158.70, SE = 3.66, t = 43.37, p < 0.001), but decreased with
the lengths of word N-1 (b = -5.06, SE = 1.41, t = -3.59, p < 0.001) and word N+1 (b =
-6.53, SE = 1.41, t = -4.62, p < 0.001). And like first-fixation and gaze durations, the
total-viewing times deceased with increasing gap size of word N (b = -13.16, SE =
0.85, t = -15.44, p < 0.001).
----- Table 6 -----
The results of the preceding analyses are consistent with other experimental
results showing that the properties of the fixated word and its spatially adjacent
neighbors can influence fixation durations during Chinese reading. For example,
several studies have demonstrated that fixation durations decrease with increasing
processing ease (e.g., frequency) of both the fixated word (Yan et al., 2006; Yang &
McConkie, 1999; Rayner, Li, Juhasz, & Yan, 2005) and the previously fixated word
(i.e., spillover effects; Li et al., 2014, see Table 1). Our results are also consistent with
evidence that gaze durations tend to increase as the length of the fixated word
increases and as the length of the previously fixated word decreases (Li et al., 2014,
13
see Table 1). The theoretical implications of these results for our understanding of
eye-movement control during the reading of Chinese will be discussed next.
General Discussion
The present experiment aimed to determine if and how “word” processing affects
eye movements in a Chinese-like visual search task so as to better inform our
understanding of the word-related effects that have been observed in the actual
reading of Chinese text. We replicated the well-established pattern of fixation
locations that has been reported in Chinese reading: Whereas the distribution of first-
fixation locations peaked at the word beginnings when the words were fixated more
than once, the distribution of single-fixation locations peaked near the word centers
(Li et al., 2011; Yan et al., 2010). However, an examination of the distributions of all
of the forward fixation locations indicated that they were clearly uniform and not
influenced by word length, word-processing difficulty (i.e., gap size), or practice.
Finally, there were effects of word length and word-processing difficulty on the
various fixation-duration measures and forward saccade length, suggesting that
variables that influence that rate of on-going word processing also dynamically
influences both when and where the eyes move during our task and—as we will
argue, by extension—the reading of Chinese text (Liu et al., 2014). In the remainder
of this discussion, we indicate precisely why we maintain this position, and discuss in
more detail what we believe each of the preceding results says about how readers of
Chinese make “decisions” about when and where to move their eyes.
First, the finding that the distributions of all forward fixation locations were
uniform and invariant to both “word” properties and practice strongly suggests that
there is no preferred viewing location in our task; otherwise, the distributions would
have tended to be normal, peaked on whatever viewing location actually afforded
14
efficient performance of our task. We believe that the patterns of fixation-location
distributions observed for first-of-multiple and single fixations in our study and others
(e.g., Li et al., 2011) and that have been suggested as indicative of a preferred viewing
location that is conditional upon the successful segmentation of parafoveal words
(e.g., see Yan et al., 2010) is instead an artifact of the simple fact that words that are
fixated near their centers are less likely to be refixated than words that are fixated
nears their beginnings (Liu et al., submitted for review). Li et al. (2011) provide
convincing support for this conjecture by running simulations of eye movements
during Chinese reading that included the simple assumption that saccades are of
constant mean length but with some variability. These simulations reproduced the
findings that the distributions of single- and first-of-multiple fixation locations tend to
be near the centers and beginnings of words, respectively, thereby demonstrated
precisely how a pattern of results that might otherwise suggest a type of saccade
targeting that is conditional upon parafoveal word segmentation can emerge from
simple assumptions that involve no specific saccade targets (see also Ma, Li, &
Pollatsek, in press). These simulations, in conjunction with our findings, are therefore
consistent with the hypothesis that there are no specific saccade targets during the
reading of Chinese. In other words, a saccade target is not a true target per se, but is
instead the consequence of a saccade that is intended to move the eyes to an
informative viewing location.
Second, the finding that properties of words influenced the various fixation-
duration measures in our task is completely consistent with findings that lexical
properties of Chinese words also influence such measures during the reading of
Chinese (Li et al., 2014; Liu et al., 2014; Liu et al., submitted for review; Yan et al.,
2006; Yang & McConkie, 1999; Rayner et al., 2005). These findings indicate the
15
validity of our method of using our version of the Landolt-C paradigm to examine
eye-movement control during Chinese reading by replicating key results showing that
the relative difficulty of “word” processing (Williams & Pollatsek, 2007; Williams et
al., 2014) can influence the time spent fixating (processing) a given “word”,
analogous to what is known to happen in actual reading (e.g., Schilling et al., 1998).
These findings are therefore important because they permits one to infer that whatever
perceptual (e.g., segmenting individual “words” based on character gap size and
orientation) and cognitive (e.g., discriminating non-targets from targets) are operative
in our task are likely to have analogs in Chinese reading (e.g., segmenting individual
words, identifying characters and words, etc.), and that, in both tasks, these processes
are likely to play causal roles in the patterns of eye movements that are observed. The
most transparent of these causal roles is undoubtedly related to the moment-by-
moment “decisions” about when to move the eyes from one location to the next, but
as our experiment demonstrates, the causal influence is also likely to extend to
decisions about where to move the eyes.
This last conjecture is consistent with our third main finding that the properties of
“words” influences forward saccade length in our task. On some level, this finding
should not be too surprising if one considers the range on possible ways that readers
might in theory guide their eyes while reading Chinese. On one end of the continuum,
it is possible—thought not likely—that readers simply move their eyes randomly to
new locations, perhaps adopting some simple heuristic (e.g., the “fixed saccade
length” assumption used in the simulations reported by Li et al., 2011; also see Yan et
al., 2010). At the other end of the continuum, readers might use word boundary
information to move their eyes to the center of the next unidentified word, as posited
to occur in models that simulate the eye movements of people reading alphabetic
16
languages like English and German (e.g., E-Z Reader: Reichle, Pollatsek, Fisher, &
Rayner, 1998; Reichle, Warren, & McConnell, 2009; Reichle, 2011; SWIFT: Engbert,
Nuthmann, Richter, & Kliegl, 2005; Schad & Engbert, 2012). This second possibility
also seems unlikely given the empirical evidence against preferred viewing locations
in both this experiment and in experiments involving Chinese reading (e.g., Li et al.,
2011), and given the inherent difficulties associated with word segmentation in
Chinese reading (e.g., see Li et al., 2009). By the process of elimination, then, the fact
that these two extreme possible accounts of saccade targeting in Chinese are not very
feasible suggests a third possibility—that readers decide where to move their eyes
using some type of information other than word boundaries. What might this
information be?
One possibility that we believe is a viable candidate is whatever processing
difficulty is associated with the currently fixated word and/or the word(s) immediately
to the right of fixation. By this account, readers might use the foveal load, or the
difficulty associated with identifying the fixated word, as a metric to gauge how far to
move their eyes with the next saccade. The intuition behind this simple heuristic is
that, because difficult-to-process fixated words afford less parafoveal preview of the
upcoming word than do less difficult-to-process fixated words (Henderson & Ferriera,
1990; Kennison & Clifton, 1995), foveal load might provide the reader with a metric
of how much processing still remains to be done on the parafoveal word when it is
fixated, and thus a metric of how far the eyes should be moved into the word. (A
parafoveal word that has received little processing during preview is likely to require
a substantial amount of processing after it is fixated, making a more conservative
fixation near the beginning of the word, thereby allowing for the possibility of one or
more additional fixations on the word during the forward pass through the text.) And
17
for the same reason, information that becomes available about the parafoveal word
from the foveal word might also be expected to play an important role in modulating
the length of the saccade leaving the foveated word if the goal is to move the eyes just
past the location where the parafoveal characters and/or words have been sufficiently
processed for their identification to be imminent. Such a proposal has already been
suggested to account for experimental results showing that the saccade length leaving
a word is modulated by the difficulty (e.g., frequency) of that word during Chinese
reading (Liu et al., 2014; Wei et al., 2013), and as a general account of how people
decide where to move their eyes during the reading of Chinese (Liu et al., submitted
for review).
Thus, our variant Landolt-C paradigm has provided new information that is
consistent with prior research using this paradigm (e.g., Williams & Pollatsek, 2007;
Vanyukov et al., 2012) and with prior research on Chinese reading (e.g., Li et al.,
2014; Liversedge et al., 2013; Yan et al., 2006), lending further support to a
theoretical account of eye-movement control in Chinese reading—an account based
on local character- and word-based control of when the eyes move that in turn affects
where (or more precisely, how far) the eyes move (Liu et al., 2014, submitted for
review; Wei et al., 2013). This represents significant empirical and theoretical
progress in understanding eye-movement control during reading because it reveals the
limitations of existing accounts of eye-movement control (e.g., the computational
models cited earlier) that are based of years of research of alphabetic languages. It
goes without saying that this progress will be extremely useful for developing formal
accounts (i.e., computational models) of eye-movement control in non-alphabetic
languages like Chinese, and more generally for informing our understanding of how
18
differences among languages and writing systems influence the perceptual, cognitive,
and motoric processes that move the eyes during reading.
19
References
Anderson, J. R. (1990). The adaptive character of thought. Hillsdale, NJ: Erlbaum.
Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for
confirmatory hypothesis testing: Keep it maximal. Journal of Memory and
Language, 68, 255-278.
Bates, D., Maechler, M., Bolker, B., & Walker, S. (2014). lme4: Linear mixed-effects
models using Eigen and S4. URL http://lme4.r-forge.r-project.org/.
Bicknell, K., & Levy, R. (2012).The utility of modelling word identification from
visual input within models of eye movements in reading. Visual Cognition, 20,
422-456.
Chen, H. (1996). Chinese reading and comprehension: A cognitive psychology
perspective. In M. H. Bond (Ed.), Handbook of Chinese psychology (pp. 43-
62). Hong Kong: Oxford University Press.
Chen, H., Song, H., Lau, W. Y., Wong, K. F. E., & Tang, S. L. (2003). Chinese reading
and comprehension: A cognitive psychology perspective. In C. McBride-
Chang & H. Chen (Eds.), Reading development in Chinese children (pp. 157-
169). Westport, CT: Praeger.
Chen, H., & Zhou, X. (1999). Processing East Asian languages: An introduction.
Language and cognitive processes, 14, 425-428.
Cheng, C. (1981). Perception of Chinese character. Act Psychological Taiwanica, 23,
137-153.
Corbic, D., Glover, L., & Radach, R. (2007). The Landolt-C string scanning task as a
proxy for visuomotor processing in reading. A pilot study. In Poster session
presented at the 14th European Conference on Eye Movements.
20
Engbert, R., Nuthmann, A., Richter, E., & Kliegl, R. (2005). SWIFT: A dynamical
model of saccade generation during reading. Psychological Review, 112, 777-
813.
Feng, G. (2008). Orthography and eye movements: The paraorthographic linkage
hypothesis. In K. Rayner, D. Shen, X. Bai, & G. Yan (Eds.), Cognitive and
cultural influences on eye movements (pp. 395-420). Tianjin, China: Tianjin
People’s Publishing House.
Forster, K. I. & Chambers, S. M. (1973). Lexical access and naming time. Journal of
Verbal Learning and Verbal Behavior, 12, 627-635.
Henderson, J. M., & Ferreira, F. (1990). Effects of foveal processing difficulty on the
perceptual span in reading: Implications for attention and eye movement
control. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 16, 417-429.
Hoosain, R. (1991). Aspects of the Chinese language. In R. Hoosain (Ed.),
Psycholinguistic implications for linguistic relativity: A case study of Chinese
(pp. 5–21). Hillsdale, NJ: Erlbaum.
Hoosain, R. (1992). Psychological reality of the word in Chinese. In H. C. Chen & O.
J. L. Tzeng (Eds.), Language processing in Chinese (pp. 111-130).
Amsterdam, the Netherlands: North-Holland.
Inhoff, A. W., & Rayner, K. (1986). Parafoveal word processing during eye fixations
in reading: Effects of word frequency. Perception & Psychophysics, 40, 431-
439.
Kennison, S. M., & Clifton, C., Jr. (1995). Determinants of parafoveal preview benefit
in high and low working memory capacity readers: implications for eye
21
movement control. Journal of Experimental Psychology: Learning, Memory
and Cognition, 21, 68-81.
Kliegl, R., Nuthmann, A., & Engbert, R. (2006). Tracking the mind during reading:
The influence of past, present, and future words on fixation durations. Journal
of Experimental Psychology: General, 135, 12-35.
Kuznetsova, A., Brockhoff, P. B., & Christensen, R. H. B. (2013). lmerTest: Tests for
random and fixed effects for linear mixed effect models (lmer objects of lme4
package). URL http://lmertest.r-forge.r-project.org/.
Li, X., Bicknell, K., Liu, P., Wei, W., & Rayner, K. (2014). Reading Is Fundamentally
Similar Across Disparate Writing Systems: A Systematic Characterization of
How Words and Characters Influence Eye Movements in Chinese Reading.
Journal of Experimental Psychology: General, 143, 895-913.
Li, X., Gu, J., Liu, P., & Rayner, K. (2013). The advantage of word-based processing
in Chinese reading: Evidence from eye movements. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 39, 879-889.
Li, X., Liu, P., & Rayner, K. (2011). Eye movement guidance in Chinese reading: Is
there a preferred viewing location? Vision Research, 51, 1146-1156.
Li, X., & Logan, G. (2008). Object-based attention in Chinese readers of Chinese
words: Beyond Gestalt principles. Psychonomic Bulletin & Review, 15, 945-
949.
Li, X., Rayner, K., & Cave, K. R. (2009). On the segmentation of Chinese words
during reading. Cognitive Psychology, 58, 525-552.
Liu, P., Li, W., Lin, N., & Li, X. (2013). Do Chinese readers follow the National
Standard Rules for word segmentation during reading? PLoS One, 8, e55440.
22
Liu, Y., Li, X., & Pollatsek, A. (2014). Cognitive Control of Saccade Amplitude
during the Reading of Chinese: A Theoretical Analysis and New Evidence.
Manuscript submitted to review.
Liu, Y., & Reichle, E. D. (2010). The emergence of adaptive eye movements in
reading. In S. Ohlsson & R. Catrabone (Eds.), Proceedings of the 32nd Annual
Conference of the Cognitive Science Society (pp. 1136-1141). Austin, TX:
Cognitive Science Society.
Liu, Y., Reichle, E. D., & Gao, D.-G. (2013). Using reinforcement learning to
examine dynamic attention allocation during reading. Cognitive Science, 37,
1507-1540.
Liu, Y., & Reichle, E. D., & Li, X. (2014). Parafoveal Processing Affects Outgoing
Saccade Length During the Reading of Chinese. Journal of Experimental
Psychology: Learning, Memory, and Cognition. Manuscript in press.
Liversedge, S.P., Hyönä, J., & Rayner, K. (2013). Eye movements during Chinese
reading. Journal of Research in Reading, 36, S1-S3.
Ma, G., Li, X., & Pollatsek, A. (in press). There is no relationship between the
preferred viewing location and word segmentation in Chinese reading. Visual
Cognition. doi: 10.1080/13506285.2014.1002554.
McConkie, G. W., Kerr, P. W., Reddix, M. D., & Zola, D. (1988). Eye movement
control during reading: I. The location of the initial eye fixations on words.
Vision Research, 28, 1107-1118.
O’Regan, J. K. (1981). The convenient viewing position hypothesis. In D. F. Fisher,
R. A. Monty, & J. W. Senders (Eds.), Eye movements: Cognition and visual
perception (pp. 363–383). Hillsdale, NJ: Lawrence Erlbaum Associates.
23
Pinheiro, J.C., & Bates D.M. (2000). Mixed-effects Models in S and S-PLUS.
Springer, New York, USA.
Rayner, K. (1979). Eye guidance in reading: Fixation locations within words.
Perception, 8, 21-30.
Rayner, K. (1998). Eye movements in reading and information processing: 20 years of
research. Psychological Bulletin, 124, 372-422.
Rayner, K. (2009). Eye movements and attention in reading, scene perception, and
visual search. Quarterly Journal of Experimental Psychology, 62, 1457-1506.
Rayner, K., Ashby, J., Pollatsek, A., & Reichle, E. D. (2004). The effects of frequency
and predictability on eye fixations in reading: Implications for the E-Z
Reader model. Journal of Experimental Psychology: Human Perception and
Performance, 30, 720-730.
Rayner, K., Li, X., Juhasz, B. J., & Yan, G. (2005). The effect of word predictability
on the eye movements of Chinese readers. Psychonomic Bulletin & Review,
12, 1089-1093.
Rayner, K., & Morrison, R. E. (1981). Eye movements and identifying words in
parafoveal vision. Bulletin of the Psychonomic Society, 17, 135-138.
Rayner, K., Sereno, S. C., & Raney, G. E. (1996). Eye movement control in reading: A
comparison of two types of models. Journal of Experimental Psychology:
Human Perception and Performance, 22, 1188-1200.
Reichle, E. D. (2011). Serial attention models of reading. In S. P. Liversedge, I. D.
Gilchrist, & S. Everling (Eds.), Oxford handbook on eye movements (pp. 767-
786). Oxford, U.K.: Oxford University Press.
24
Reichle, E. D. & Laurent, P. (2006). Using reinforcement learning to understand the
emergence of “intelligent” eye-movement behavior during reading.
Psychological Review, 113, 390-408.
Reichle, E. D., Pollatsek, A., Fisher, D. L., & Rayner, K. (1998). Toward a model of
eye movement control in reading. Psychological Review, 105, 125-157.
Reichle, E. D., Warren, T., & McConnell, K. (2009). Using E-Z Reader to model the
effects of higher-level language processing on eye movements during reading.
Psychonomic Bulletin & Review, 16, 1-21.
Schad, D.J., & Engbert, R. (2012). The zoom lens of attention: Simulating shuffled
versus normal text reading using the SWIFT model. Visual Cognition, 20, 391-
421.
Schilling, H. E. H., Rayner, K., & Chumbley, J. I. (1998). Comparing naming, lexical
decision, and eye fixation times: Word frequency effects and individual
differences. Memory and Cognition, 26, 1270-1281.
Smith, B. (1988). Gestalt Theory: An Essay in Philosophy. In B. Smith (Ed.),
Foundations of Gestalt Theory (pp. 11-81), Munich and Vienna: Philosophia
Verlag.
Tsai, J. L., & McConkie, G. W. (2003). Where do Chinese readers send their eyes? In
R. R. J. Hyona & H. Deubel (Eds.), The mind’s eye: Cognitive and applied
aspects of eye movement research (pp. 159-176). Amsterdam, the Netherlands:
Elsevier.
Vanyukov, P. M., Warren, T., Wheeler, M. E., & Reichle, E. D. (2012). The emergence
of frequency effects in eye movements. Cognition, 123, 185-189.
Wei, W., Li, X., & Pollastsek, A. (2013). Word properties of a fixated region affect
outgoing saccade length in Chinese reading. Vision Research, 80, 1-6.
25
Williams, C. C., & Pollatsek, A. (2007). Searching for an O in an array of Cs: Eye
movements track moment-to-moment processing in visual search. Perception
and Psychophysics, 69, 372-381.
Williams, C. C., Pollatsek, A., & Reichle, E. D. (2014). Examing eye movements in
visual search through clusters of objects in a circular array. Journal of
Cognitive Psychology, 26, 1-14.
Yan, G., Tian, H., Bai, X., & Rayner, K. (2006). The effect of word and character
frequency on the eye movements of Chinese readers. British Journal of
Psychology, 97, 259-268.
Yan, M., Kliegl, R., Richter, E., Nuthmann, A., & Shu, H. (2010). Flexible saccade
target selection in Chinese reading. The Quarterly Journal of Experimental
Psychology, 63, 705-725.
Yang, H., & McConkie, G. W. (1999). Reading Chinese: Some basic eye-movement
characteristics. In J. Wang, A. W. Inhoff, & H.-C. Chen (Eds.), Reading
Chinese script (pp. 207-222). Mahwah, NJ: Erlbaum.
Zang, C., Liversedge, S.P., Bai, X., & Yan, G. (2011). Eye movements during Chinese
reading. In S.P. Liversedge, I.D. Gilchrist, & S. Everling. (Eds). Oxford
Handbook on Eye Movements (pp. 961-978). Oxford University Press.
26
Footnote
1. For the purposes of brevity and to draw parallels between our paradigm and the
actual reading of Chinese text (which is of theoretical interest), we will refer to the
Landolt-squares used in our experiment as “characters” and to the clusters of such
characters as “words” throughout the remainder of this article. (Accordingly, we will
also drop the use of scare quotes.)
27
Table 1. Chi-square tests for the uniformity of all forward fixation locations as a
function of word size (i.e., number of characters).
Word Size
1-character
2-character
3-character
4-character
χ2
0.08
1.19
0.92
1.74
df
1
3
5
7
p
0.778
0.755
0.969
0.973
Note: The null hypothesis is uniformity of all forward fixation locations.
28
Table 2. LMM analyses of all forward fixation locations (aligned to the left of fixated
word, in characters).
Variables
Intercept
Practice (# of trials)
Word N-1
# of characters
gap size (pixels)
Word N
# of characters
gap size (pixels)
Word N+1
# of characters
gap size (pixels)
b
0.01
0.0001
-0.01
-0.001
0.99
0.01
0.01
-0.01
Model
SE
0.07
0.0001
0.01
0.01
0.01
0.01
0.01
0.01
t
0.13
0.40
-1.20
-0.15
97.27
1.01
0.92
-1.24
p
0.894
0.689
0.232
0.878
< 0.001
0.315
0.358
0.214
-
-
2.47
2.46
Min.
2.48
2.47
Values (characters)
Max.
2.45
2.46
2.45
2.48
2.48
2.45
0.99
2.45
3.95
2.48
Notes: For number of characters, min. = 1 character, max. = 4 characters; for gap size,
min. = 2 pixels, max. = 8 pixels; for practice, min. = 1st trial, max. = 256th trial.
Estimates of predicted variable values were calculated while fixing the values of the
other variables equal to their mean values.
29
Table 3. LMM analyses of forward saccade length (in characters).
Variables
Intercept
Practice (# of trials)
Word N-1
# of characters
gap size (pixels)
Word N
# of characters
gap size (pixels)
Word N+1
# of characters
gap size (pixels)
b
2.88
0.0001
0.01
0.01
0.10
0.01
0.04
0.02
SE
0.19
0.0004
0.01
0.003
0.01
0.01
0.01
0.003
Model
t
15.28
0.16
2.43
1.72
9.88
2.23
5.05
5.58
P
< 0.001
0.877
0.015
0.085
< 0.001
0.037
< 0.001
< 0.001
-
-
3.44
3.42
Min.
3.40
3.40
Values (characters)
Max.
3.36
3.37
3.44
3.43
3.48
3.47
3.27
3.38
3.56
3.46
Notes: For number of characters, min. = 1 character, max. = 4 characters; for gap size,
min. = 2 pixels, max. = 8 pixels; for practice, min. = 1st trial, max. = 256th trial.
Estimates of predicted variable values were calculated while fixing the values of the
other variables equal to their mean values.
30
Table 4. LMM analyses of first-fixation durations (ms).
Model
Variables
Intercept
Practice (# of trials)
Word N-1
# of characters
gap size (pixels)
Word N
# of characters
gap size (pixels)
Word N+1
# of characters
gap size (pixels)
b
319.70
-0.03
0.63
-0.07
-1.35
-1.16
1.46
-0.07
SE
10.80
0.02
0.49
0.34
0.67
0.33
0.93
0.32
t
29.61
-1.64
1.28
-0.20
-2.01
-3.50
1.57
-0.23
p
< 0.001
0.117
0.200
0.847
0.051
0.002
0.131
0.819
-
308
-
315
310
312
Values (ms)
Min. Max.
309
312
313
311
312
311
313
315
309
308
Notes: For number of characters, min. = 1 character, max. = 4 characters; for gap size,
min. = 2 pixels, max. = 8 pixels; for practice, min. = 1st trial, max. = 256th trial.
Estimates of predicted variable values were calculated while fixing the values of the
other variables equal to their mean values.
31
Table 5. LMM analyses of gaze durations (ms).
Variables
Intercept
Practice (# of trials)
Word N-1
# of characters
gap size (pixels)
Word N
# of characters
gap size (pixels)
Word N+1
# of characters
gap size (pixels)
Model
t
8.68
0.59
p
< 0.001
0.560
-3.34
-1.37
< 0.001
0.187
b
246.30
0.04
-4.03
-1.39
SE
28.39
0.07
1.21
1.02
135.70
-9.75
10.91
0.72
12.44
< 0.001
-13.47 < 0.001
-6.84
0.55
1.53
0.60
-4.46
0.92
< 0.001
0.357
-
528
-
517
529
527
Values (ms)
Min. Max.
533
522
513
525
517
519
322
552
730
494
Notes: For number of characters, min. = 1 character, max. = 4 characters; for gap size,
min. = 2 pixels, max. = 8 pixels; for practice, min. = 1st trial, max. = 256th trial.
Estimates of predicted variable values were calculated while fixing the values of the
other variables equal to their mean values.
32
Table 6. LMM analyses of total-viewing times (ms).
Model
Variables
Intercept
Practice (# of trials)
Word N-1
# of characters
gap size (pixels)
Word N
# of characters
gap size (pixels)
Word N+1
# of characters
gap size (pixels)
b
306.80
-0.08
-5.06
-1.66
158.70
-13.16
-6.53
-0.42
SE
31.93
0.08
1.41
1.11
3.66
0.85
1.41
0.70
t
9.61
-1.00
-3.59
-1.49
p
< 0.001
0.328
< 0.001
0.151
43.37
< 0.001
-15.44 < 0.001
-4.62
-0.60
< 0.001
0.548
-
576
-
598
595
592
Values (ms)
Min. Max.
597
589
578
586
580
582
349
627
825
548
Notes: For number of characters, min. = 1 character, max. = 4 characters; for gap size,
min. = 2 pixels, max. = 8 pixels; for practice, min. = 1st trial, max. = 256th trial.
Estimates of predicted variable values were calculated while fixing the values of the
other variables equal to their mean values.
33
Figure Caption
Figure 1. Examples of two Chinese sentences and their respective translations. In the
first, the sequence of four underlined characters correspond to two words. In the
second, the same characters correspond to a single word.
Figure 2. Examples of experimental materials, with the target (i.e., a square) and one
of the “word” clusters being rendered in gray for illustrative purposes.
Figure 3. The distribution of initial fixations on words that received one or more
fixations (panels a-b), of first-of-multiple fixations on words that received two or
more fixations (panels c-d), of single fixations on words that received only one
fixation (panels e-f), and of all forward fixations on words during first-pass scanning
(panels g-h). All of the panels show the fixation-location distributions as a function of
the number of characters within a word. Panels a, c, e, and g (left column) also show
the fixation-location distributions as a function of the gap size within a word, whereas
panels b, d, f, and g (right column) also show the fixation-location distributions as a
function of practice (ordinal block number). Error bars represented the standard errors
of the means.
34
Figure 1.
35
Figure 2.
36
Figure 3.
37
|
1511.00235 | 1 | 1511 | 2015-11-01T11:46:33 | Broadband Macroscopic Cortical Oscillations Emerge from Intrinsic Neuronal Response Failures | [
"q-bio.NC",
"cond-mat.stat-mech",
"physics.bio-ph"
] | Broadband spontaneous macroscopic neural oscillations are rhythmic cortical firing which were extensively examined during the last century, however, their possible origination is still controversial. In this work we show how macroscopic oscillations emerge in solely excitatory random networks and without topological constraints. We experimentally and theoretically show that these oscillations stem from the counterintuitive underlying mechanism - the intrinsic stochastic neuronal response failures. These neuronal response failures, which are characterized by short-term memory, lead to cooperation among neurons, resulting in sub- or several- Hertz macroscopic oscillations which coexist with high frequency gamma oscillations. A quantitative interplay between the statistical network properties and the emerging oscillations is supported by simulations of large networks based on single-neuron in-vitro experiments and a Langevin equation describing the network dynamics. Results call for the examination of these oscillations in the presence of inhibition and external drives. | q-bio.NC | q-bio | Broadband Macroscopic Cortical Oscillations
Emerge from Intrinsic Neuronal Response Failures
Amir Goldental1,†, Roni Vardi2,†, Shira Sardi1,2, Pinhas Sabo1 and Ido Kanter1,2,*
1Department of Physics, Bar-Ilan University, Ramat-Gan 52900, Israel
2Gonda Interdisciplinary Brain Research Center and the Goodman Faculty of Life Sciences, Bar-
Ilan University, Ramat-Gan, Israel
†These authors contributed equally to this work.
*Correspondence: [email protected]
Abstract
Broadband spontaneous macroscopic neural oscillations are rhythmic cortical firing which were extensively
examined during the last century, however, their possible origination is still controversial. In this work we show
how macroscopic oscillations emerge in solely excitatory random networks and without topological constraints.
We experimentally and theoretically show that these oscillations stem from the counterintuitive underlying
mechanism - the intrinsic stochastic neuronal response failures. These neuronal response failures, which are
characterized by short-term memory, lead to cooperation among neurons, resulting in sub- or several- Hertz
macroscopic oscillations which coexist with high frequency gamma oscillations. A quantitative interplay
between the statistical network properties and the emerging oscillations is supported by simulations of large
networks based on single-neuron in-vitro experiments and a Langevin equation describing the network
dynamics. Results call for the examination of these oscillations in the presence of inhibition and external drives.
1. INTRODUCTION
The most widespread cooperative activity of neurons within the cortex is spontaneous macroscopic
oscillations(Silva et al., 1991;Gray, 1994;Contreras et al., 1997;Buzsaki and Draguhn, 2004;Chialvo,
2010), which range between sub- and hundred- Hertz(Başar et al., 2001;Brovelli et al., 2004;Buzsaki and
Draguhn, 2004;Grillner et al., 2005;Giraud and Poeppel, 2012). The high cognitive functionalities of these
oscillations are still controversial(Klimesch, 1996;1999;Başar et al., 2001;Wiest and Nicolelis,
2003;Buzsaki, 2006;Kahana, 2006;Fries, 2009;Giraud and Poeppel, 2012;Thivierge et al., 2014) and are
typically attributed to transitory binding activities among indirect macroscopic distant cortical
regions(Gray, 1994;Başar et al., 2001;Buzsaki and Draguhn, 2004;Roxin et al., 2004;Fries, 2009). In
Cooperative oscillations emerge from response failures
addition, it was found that the theta rhythms(Klimesch, 1999;Buzsaki and Draguhn, 2004), oscillations in
the range of 4-10 Hz, play a key role in the formation and retrieval of episodic and spatial
memory(Hasselmo, 2005). This theta rhythm is usually accompanied by high frequency oscillations in the
range of 30-80 Hz, known as gamma oscillations(Colgin and Moser, 2010). Gamma oscillations are also
related to sensory stimulations and induce neuronal ensemble synchrony by generating a narrow window
for effective excitation(Cardin et al., 2009).
There are several suggested mechanisms for the formation of such rhythms on the network level(Wang,
2010). Most of the proposed mechanisms are based on the existence of inhibitory synapses(Wilson and
Cowan, 1972;Jirsa and Haken, 1996;Brunel and Wang, 2003), especially for high frequency
oscillations(Brunel and Wang, 2003;Colgin and Moser, 2010;Wang, 2010). For illustration, assume that a
fast excitation increases neural firing in an excitatory short-delayed feedback loop. Consequently, neuronal
populations along the excitatory feedback loop will fire at high rates that will cause a slower response of
the inhibitory neurons. As a result, the inhibitory neurons will depress the activity within the excitatory
population. This will then depress the excitation of the inhibitory population. Finally, the depression of the
inhibitory neurons allows a repeated fast excitation of the excitatory population.
In this work we show how extra-cellular potential oscillations, synchronized rhythmic firing of neurons,
emerge in random networks without inhibitory synapses. Our findings are based on an experimental
observations of neuronal plasticity in the form of intrinsic neuronal response failures(Vardi et al., 2015).
Using simulations of large networks, based on single-neuron in-vitro experiments, we show that this type
of neuronal plasticity leads to the coexistence of both theta and gamma oscillations. Results are supported
by a quantitative approach based on a Langevin equation, which describes the network dynamics.
2. MATERIALS AND METHODS
2.1 In-vitro experiments
2.1.1 Animals
All procedures were conducted in accordance with the National Institutes of Health Guide for the Care and
Use of Laboratory Animals and Bar-Ilan University Guidelines for the Use and Care of Laboratory Animals
in Research and were approved and supervised by the Institutional Animal Care and Use Committee.
2.1.2 Culture preparation
Cortical neurons were obtained from newborn rats (Sprague-Dawley) within 48 h after birth using
mechanical and enzymatic procedures. The cortical tissue was digested enzymatically with 0.05% trypsin
solution in phosphate-buffered saline (Dulbecco’s PBS) free of calcium and magnesium, and supplemented
with 20 mM glucose, at 37◦C. Enzyme treatment was terminated using heat-inactivated horse serum, and
cells were then mechanically dissociated. The neurons were plated directly onto substrate-integrated multi-
2
Cooperative oscillations emerge from response failures
electrode arrays (MEAs) and allowed to develop functionally and structurally mature networks over a time
period of 2-3 weeks in-vitro, prior to the experiments. Variability in the number of cultured days in this
range had no effect on the observed results. The number of plated neurons in a typical network was in the
order of 1,300,000, covering an area of about 380 mm2. The preparations were bathed in minimal essential
medium (MEM-Earle, Earle's Salt Base without L-Glutamine) supplemented with heat-inactivated horse
serum (5%), glutamine (0.5 mM), glucose (20 mM), and gentamicin (10 g/ml), and maintained in an
atmosphere of 37◦C, 5% CO2 and 95% air in an incubator as well as during the electrophysiological
measurements.
2.1.3 Synaptic blockers
All experiments were conducted on cultured cortical neurons that were functionally isolated from their
network by a pharmacological block of glutamatergic and GABAergic synapses. For each culture 20 l of
a cocktail of synaptic blockers was used, consisting of 10 μM CNQX (6-cyano-7-nitroquinoxaline-2,3-
dione), 80 μM APV (amino-5-phosphonovaleric acid) and 5 μΜ bicuculline. This cocktail did not block
the spontaneous network activity completely, but rather made it sparse. At least one hour was allowed for
stabilization of the effect.
2.1.4 Stimulation and recording
An array of 60 Ti/Au/TiN extracellular electrodes, 30 μm in diameter, and spaced 500 μm from each other
(Multi-Channel Systems, Reutlingen, Germany) was used. The insulation layer (silicon nitride) was pre-
treated with polyethyleneimine (0.01% in 0.1 M Borate buffer solution). A commercial setup (MEA2100-
2x60-headstage, MEA2100-interface board, MCS, Reutlingen, Germany) for recording and analyzing data
from two 60-electrode MEAs was used, with integrated data acquisition from 120 MEA electrodes and 8
additional analog channels, integrated filter amplifier and 3-channel current or voltage stimulus generator
(for each 60 electrode array). Mono-phasic square voltage pulses typically in the range of [-800, -500] mV
and [60, 400] μs were applied through extracellular electrodes. Each channel was sampled at a frequency
of 50k samples/s, thus the changes in the neuronal response latency were measured at a resolution of 20 s.
2.1.5 Cell selection
Each node was represented by a stimulation source (source electrode) and a target for the stimulation – the
recording electrode (target electrode). These electrodes (source and target) were selected as the ones that
evoked well-isolated, well-formed spikes and reliable response with a high signal-to-noise ratio. This
examination was done with a stimulus intensity of -800 mV with duration of 200 μs using 30 repetitions at
a rate of 5 Hz, followed by 1200 repetitions at a rate of 10 Hz.
2.1.6 In-vitro experiment with feedback loops and neural circuits
The activity of all source and target electrodes was collected and action potentials were detected on-line by
threshold crossing, and entailed stimuli were delivered in accordance with the circuit's connectivity, as
3
Cooperative oscillations emerge from response failures
described below. A successful response was defined as a spike occurring within a typical time window of
2-10 ms following the beginning of an electrical stimulation.
In Figures 2A,B, after every spike detection two supra-threshold extracellular stimulations were given to
the same neuron, after 600 ms and 630 ms. In case that the timings of the stimulations overlap, only one
stimulation is given.
In Figures 2C,D, after every spike detection supra-threshold extracellular stimulations were given to its
connected neurons. For example, if a spike was detected at the left (green) neuron (Figures 2C), a supra-
threshold extracellular stimulation will be given to the middle (brown) neuron after 330 ms.
2.1.7 Data analysis
Analyses were performed in a Matlab environment (MathWorks, Natwick, MA, USA). The reported results
were confirmed based on at least eight experiments each, using different sets of neurons and several tissue
cultures.
The temporal firing frequency, around stimulation no. i, of the neuron in Figure 1 was calculated using the
following procedure
𝑓𝑖𝑟𝑖𝑛𝑔 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 (𝑖) = 𝑠𝑡𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 ∙ ∑
𝑖+125
m=max (0,𝑖−125)
𝐼𝑠_𝑆𝑝𝑖𝑘𝑒𝑑(𝑚)
𝑖 + 125 − max(0, 𝑖 − 125)
where Is_Spiked(m) = 1 if the neuron responded to stimulation no. m, otherwise Is_Spiked(m) = 0.
2.2 Simulations
Simulations (similar to (Vardi et al., 2015)) consist of a network of N leaky integrate and fire neurons
𝑑𝑉𝑖
𝑑𝑡
= −
𝑉𝑖
𝜏
+ ∑
𝑁
𝑗=1
𝐽𝑗𝑖 ∑ 𝛿(𝑡 − 𝑡′ − 𝐷𝑗𝑖)
𝑡′
where i∈[1,N], =20 ms, Jji and Dji are the connection’s strength and delay from neuron j to i, respectively.
The summation over t' sums all firing times of neuron j, the integration time step is 0.05 ms, and the
threshold is 1. For the nth threshold crossing of a neuron, its probability for a response is [Σ(n-m/c)exp(-
m)]/[Σexp(-m)], where n is the time gap between the nth and (n-1)th threshold crossings, c=1/fc, =1.4
and the sum is over m≥0. A refractory period of 2 ms is imposed after an evoked spike, for response failures
the voltage is set to 0.2. Results were found to be insensitive to initial conditions.
2.2.1 Various forms of p(s) lead to the same <ISI>
The probability for a response, given the last inter-stimulation-interval, p(s), should lead to <ISI>=c
(Figures 1 and 3). One can show that any p(s) satisfying
∞
∫ 𝑝(𝑠𝜏)𝑝(𝜏)𝑑𝜏
0
∞
∫ 𝜏𝑝(𝜏)𝑑𝜏
0
=
1
𝜏𝑐
4
Cooperative oscillations emerge from response failures
where p() is the probability density of an inter-stimulation-interval equals to leads to <ISI>=c. The
numerator on the left hand side stands for the average probability for a successful response, and the
denominator stands for the average inter-stimulation-interval. This ratio is equivalent to the firing rate,
hence equals to 1/c. For instance p(s)=/c, this theoretical curve fits all p() (Figure 3D). In the activity
of some random networks a good approximation for p() is
𝑝(𝜏) =
2
𝜏𝑐
𝑒𝑥𝑝 (−
2𝜏
𝜏𝑐
).
For this p() some of the p(s) solutions, which lead to <ISI>=c, are p(s)=0.5, p(s)=/c and
𝑝(𝑠𝜏) = 1 − 𝑒𝑥𝑝 (−
2𝜏
𝜏𝑐
),
which is similar to Figure 3D.
2.2.2 Fourier analysis of the rate
To perform a Fourier analysis on the activity of the network we define the rate vector:
𝑅(𝑡𝑖𝑚𝑒 = 𝑖 ∙ 𝑤) =
1
𝑁𝑤
(𝑖+0.5)𝑤
∫
(𝑖−0.5)𝑤
∑ 𝛿(𝑡′ − 𝑡𝑠𝑝𝑖𝑘𝑒)
𝑡𝑠𝑝𝑖𝑘𝑒
𝑑𝑡′
where i is an integer, w is a predefined time window and the sum is over all spike times of all N neurons.
Next, a discrete Fourier transform is preformed and the function of resulting amplitudes is normalized and
smoothed using a sliding window of 1 Hz.
3. RESULTS
3.1 Neuronal plasticity: Neuronal response failures
We start with the quantification of the neuronal response latency (NRL)(Vardi et al., 2013a;Marmari et al.,
2014;Vardi et al., 2015), measured as the time-lag between a stimulation and its corresponding evoked
spike (Figure 1). It was recently shown(Vardi et al., 2015) that when a neuron is repeatedly stimulated, its
NRL stretches gradually (Figure 1, upper panel), and when the stimulation frequency is high enough
stochastic neuronal response failure (NRFs) emerge. Specifically, each neuron is characterized by a critical
frequency, fc, typically ranging among neurons between 2 and 30 Hz. Stimulation frequencies above fc
result in NRFs, whereas for supra-threshold stimulations below fc a response is assured. The probability of
the NRFs is such that the neuron functions similar to a low pass filter, saturating its firing rate (Figure 1,
lower panel). Quantitatively, for a stimulation frequency f, the NRF probability is 1-fc/f, i.e. the firing
frequency is saturated at fc. Thus, changing the stimulation frequency will change the probability for NRFs,
while the firing frequency remains bounded, fc.
This observation is demonstrated using a cultured cortical neuron, functionally separated from its network
by synaptic blockers, with above-threshold stimulations (see section 2.1 In-vitro experiments). We examine
a neuron with periodic stimulation trials of 10, 12 and 15 Hz, and NRFs appear after a short transient where
5
Cooperative oscillations emerge from response failures
the neuron exhibits an increase of its NRL (Figure 1, upper panel). Examining the temporal firing rate of
the neuron (Figure 1, lower panel), it is noticeable that the firing frequency of the neuron is saturated at
fc=5 Hz, independent of the stimulation frequency.
The effect of NRFs is first examined on small neuronal circuits using the following experiment: We
stimulate the neuron ones and impose on the neuron two self-feedback delay loops, e.g. 600 ms and 630 ms
(Figure 2A). The neuron is stimulated 600 ms and 630 ms after each evoked spike (see section 2.1.6 In-
vitro experiment with feedback loops and neural circuits). In the case of vanishing probability for NRF, the
neuron should fire every 30 ms (Kanter et al., 2011;Vardi et al., 2012b), i.e. 33.3 Hz. Since fc=5 Hz, some
NRFs appear (Figure 2B) forming bunched firing, separated by ~600 ms, whereas the intra-bunch time-
lags, 30 ms, originated from the difference between the two feedback delay loops (Figure 2B). The
emergence of such firing bunches indicate some dynamical changes of the NRF’s probability, where the
neuron adapts its failure probability as a result of its recent stimulation history.
A more biologically realistic scenario is a neuronal circuit consisting of three artificially connected
neurons (Vardi et al., 2012b), forming the same delay loops (Figure 2C), 600 ms and 630 ms. In addition,
each neuron is identified by different fc and only the central neuron receives the initial stimulation. For the
central neuron, which is characterized by fc=7.2 Hz, NRFs occur, since the frequency of its driven neurons
is greater than its critical frequency, 6.8+4>7.2 Hz, and similarly the outer neurons have response failures
(7.2>6.8, 7.2>4) (Figure 2D). Besides the formation of firing bunches for each neuron, their firing are
correlated at zero- or shifted- time-lags (Figure 2D). The question whether the repeated bunches in such
small neuronal circuits shed light on macroscopic cortical oscillations requires to sail towards large scale
simulations.
3.2 Short-term memory of neuronal plasticity
The observed firing bunches indicate a form of short-term memory of neuronal plasticity, where the NRF
probability is mainly a function of the few preceding stimulations. Our next goal is to experimentally
quantify this neuronal plasticity and then examine its implementation on the dynamics of large neural
networks using large scale simulations.
We first assume a simplified network where each neuron has two above-threshold inputs, two outputs and
the same fc (Figure 3A), hence the statistics of the inter-stimulation-intervals of each neuron is expected to
approximately follow an exponential distribution with 2fc rate, Exp(2fc) (Figure 3B, upper panel). To
quantify the statistics of the NRFs, a long stimulation trail obeying Exp(2fc), underc/8 time resolution,
was given to a cultured neuron with fc~5 Hz (Figure 3B). Next, the conditional probabilities for a successful
response (an evoked spike), p(sii-1), were estimated for events where the current inter-stimulation-
interval equals i and the previous one equals i-1 (Figures 3B,C). It is evident that p(sii-1) is primarily a
function of i, i.e. the probability for a successful response is dictated mainly by the current inter-
6
Cooperative oscillations emerge from response failures
stimulation-interval, i. Hence, the NRFs can be dynamically approximated using p(s) (Figure 3D),
indicating that the neuronal response failure probability might be fairly estimated based solely on the last
inter-stimulation-interval, .
3.3 Neuronal oscillations on the network level
The experimental estimation of p(s) is utilized to simulate a large scale network (Figure 3A) and is
exemplified for 2000 neurons where delays between connected neurons, D, are randomly selected from a
uniform distribution U(10,15) ms. The simulation is initialized with timings of evoked spikes for a subset
of the neurons, however, besides the transient time results were found to be insensitive to the initial
conditions. The response failure is then randomly selected following the experimentally measured p(s),
independently for each neuron and stimulation (Figure 3D). Indeed, the assumption Exp(2fc) (Figure 3B,
upper panel) was confirmed (Figure 3E), the statistics of the inter-stimulation-intervals of each neuron
approximately follows an exponential distribution with 2fc rate. The raster plot of the network firing as well
as the time-dependent firing rate (Figures 4A,B) clearly indicate cooperative oscillation which can be
quantified using the Fourier analysis to fosc~3 Hz (Figure 4C, see section 2.2), and are absent in the Fourier
analysis of the firing of each individual neuron (Figure 3F). Another broadened Fourier peak is centered
at f~80 Hz, gamma oscillations(Brunel and Wang, 2003;Cunningham et al., 2004;Fries, 2009;Minlebaev
et al., 2011;Dugladze et al., 2012), which is attributed to the average delay, <D>=12.5 ms. It reflects the
average firing frequency of each neuron where NRFs vanish and all delays are equal to <D>, as GCD=<D>
for delay loops of such a random network(Kanter et al., 2011;Vardi et al., 2012a). Similar cooperative
oscillations were obtained using a counterpart simulation for the same network (Figures 4D,F) while using
the theoretical form p(s) (Figure 3D). Although the form of p(s) varies among neurons as well as
between the theoretical and the experimental forms (Figure 3D), the cooperative oscillations are found
quantitatively to be only slightly affected by its exact form. The robustness of fosc was also confirmed in
simulations for more realistic scenarios where fc significantly varies among neurons as well as their input
and output connectivity distributions. A more biological realization is exemplified in Figures 4G-I. The
number of connections per neuron is much greater than 2, i.e. more than 50 pre- and 50 post- synaptic
connections, where most of them are sub-threshold and on the average 1.5 of pre- and post-synaptic are
above-threshold. Specifically, each sub-threshold connection produces an excitatory postsynaptic potential,
EPSP, which is equal to 0.03 of the threshold. It is apparent that these additional connections do not
qualitatively change the oscillations. In addition, f was found to be robust to a wider distribution of delays
and followed its center (Figure 4I). Note that without these intrinsic NRFs, i.e. p(s)=1, the firing of each
neuron is only bounded by the refractory period which is in the order of several milliseconds. In this limiting
7
Cooperative oscillations emerge from response failures
case, the abovementioned theta and gamma oscillations disappear, as was confirmed in simulations (not
shown).
3.4 Cortical oscillations versus statistical properties of the network
The origin of the fast oscillations, f, (Figures 4C,F,I) is evident, however, the mechanism underling the
slow cooperative oscillations(Wu et al., 1999;Sanchez-Vives and McCormick, 2000;Bollimunta et al.,
2008;Nir et al., 2008;Crunelli and Hughes, 2010;Bollimunta et al., 2011), fosc, is still unclear. To explore
this mechanism we identify the following two characteristic distances on the network. The first distance,
Path, is the average minimal path between pairs of nodes, and the second distance, Loop, is the average
over the minimal feedback loop of each node, both counted by the number of nodes along the route (Figure
5A). Numerical estimations based on various network topologies indicate that the average values of these
two distances as well as their distributions are almost identical (Figure 5B) and their scaling decrease as
1/ln(<K>), where <K> is the average neuronal input connectivity (Figure 5C). These identities and scaling
(Figures 5B,C), are also supported by the following theoretical argument. Assume a random network
consisting of N neurons and an average connectivity <K>. The quantity Q(m) denotes the number of new
connected neurons to a seed neuron at a distance of m neurons, hence Q is proportional to the probability
(green) in Figure 5B. Start at an arbitrary neuron, Q(0)=1, this neuron is connected (pre-synaptic) to Q(1)
neurons. Using a recursive formula one can approximate
Q(i) = N (1 − exp (−< K >
Q(i−1)
N
)) (1 − ∑
i−1
m=1
Q(m)
N
)
where N(1-exp(-<K>Q(i-1)/N)) stands for the average number of neurons at a distance i, which are
connected from new neurons at distance i-1. The rightmost term is the probability that the neuron at distance
i is a new neuron which was not counted at shorter distances, m<i. This recursive relation is solved
numerically and the normalized Q(i) are presented in Figures 5B. The above analysis is valid for Loops
and Paths, hence their statistics are identical, in agreement with the sampling of these quantities in Figures
5B.
The importance of this distance in the formation of fosc can be understood according to the following
argument. Assume a random subgroup of firing neurons activates another random group of neurons and
vice versa. The minimal delay between pairs of neurons belonging to the two subgroups is Path·<D>.
Consequently, the oscillations are expected to scale with Path·<D>, and indeed results indicate that
fosc∝ln(<K>)∝(Path)-1 (Figures 5C,D). The minimal path is the most reliable one with respect to the NRFs,
however, the effect of longer paths is not negligible as they might maintain the activity of the Path (Figure
2), especially as their entropy is higher. In addition, the NRFs are responsible to limit the firing of all the
8
Cooperative oscillations emerge from response failures
network simultaneously (Figures 4B,E,H). Assume one neuron fires and activates <K> neurons after <D>
ms, hence after m<D> ms, <K>m neurons fire. This exponential firing growth is bound by mPath, as self-
feedback loops (Figures 2A,B) significantly lead to NRFs and to a fast decrease in the firing rate. In
addition, for a given network topology, fosc is found to scale with ln(fc) (Figure 5E), and to be robust for
networks composed of neurons with different fc (Figures 4G,H,I). These predictions might be realized in
further experiments by controlling the network topology either by the neuronal concentration or by
pharmacological manipulations.
3.5 Analytical description of the network oscillations
An analytical description of fosc is also possible, and to simplify the presentation the method is briefly
described for homogeneous networks only. Each node has the same fixed input and output connectivity, K,
and all delays are equal to D ms, hence neurons can fire only at i·D ms, where i is an integer. The fraction,
R(m), of neurons that fire at step m is given by
𝑅(𝑚) = 𝜒(𝑚) ∙ 𝐾 ∙ 𝑅(𝑚 − 1) + 𝜉(𝑚) (1)
where (m) stands for the time-dependent white noise representing the stochastic nature of the response
failures. The function (m) represents the susceptibility of the network, i.e. the fraction of neurons that fires
if all neurons are stimulated at step m, and is explicitly given by
∞
𝑖
𝜒(𝑚) = ∑ 𝑝(𝑠𝑖) ∙ 𝐾 ∙ 𝑅(𝑚 − (𝑖 + 1)) ∏(1 − 𝐾 ∙ 𝑅(𝑚 − 𝑛)) . (2)
𝑖=0
𝑛=2
The first term, p(si), stands for the probability for an evoked spike when the previous stimulation was given
before i steps (Figure 3D). The term KR(m-(i+1)) pinpoints a neuron stimulated before i steps, and the
product, the rightmost term, indicates that the neuron was not stimulated since step (m-i). Equation (2)
indicates that (m) is a function of the variable R(l) only, with l<m. Hence, after the insertion of equation
(2), into equation (1), one finds a recursion relation for R(m) which can be solved numerically given the
initial conditions. The dynamical solution of this recursion relation revealed oscillations which were found
to fit fairly good with those observed in large-scale simulations (Figures 4C,F,I). The equations imply that
the network has some memory of its previous activity, which dictates the responsiveness of the entire
network. This analytical description can be generalized to advanced structured networks, including random
connectivity, distribution for the delays as well as to include variations among neuronal critical frequencies,
fc.
4. Discussion
We have demonstrated that intrinsic neuronal response failures drive a neural network activity towards
oscillations, where high frequency oscillations, gamma, and low frequency oscillations, delta and theta,
9
Cooperative oscillations emerge from response failures
coexist. The high frequency oscillations correspond to the average delay between connected neurons in the
network, while low frequency oscillations are governed by statistical properties of the network, e.g. the
average number of connections per neuron and the average critical frequency of neurons. The coexistence
of high and low frequency oscillations was confirmed in a new type of simulations, based on a single neuron
in-vitro experiments, to evaluate the firing activity of complex networks. Results were also supported by
an analytical description of the stochastic dynamics of the network.
Preliminary results in-vivo support our findings. The NRL increases by several milliseconds under
periodic stimulations and terminates at an intermittent phase(Vardi et al., 2015). This phase is characterized
by fluctuations around a constant NRL and accompanied by NRFs. Results indicate that fc can be below 10
Hz and vary among neurons. However, quantitative measurements of fc and the statistics of the NRFs
require long stimulation trials, i.e. many thousands of high frequency periodic stimulations, which are
currently beyond our experimental capabilities.
The average firing rate of neurons in the network is low, e.g. ~3.6, ~3.9 and ~2.6 Hz in Figures 4B,E,H,
respectively. These network low firing rates are lower than the neuronal critical frequency in Figures 4B,E,
fc=5.7 Hz, and <fc>=6.5 Hz, in Figure 4H. Surprisingly, the neuronal critical firing frequency is not
saturated even when the network is completely excitatory. A biological mechanism that suppresses the
firing frequency of a single neuron below fc is aperiodic time-lags between stimulations (Vardi et al., 2015).
For illustration, assume a slow mode of alternation between stimulation frequencies of 2fc (0.5c) and 2fc/3
(1.5c), such that <>=c. For the high and low frequency mode, the expected probability for a NRF is 0.5
and 0, respectively. Consequently, <ISI>=0.5(1.5c+c)=1.25c, corresponding to a lower firing rate, 0.8fC.
In addition, the firing rates are considerably lower than f indicating that high frequency network
oscillations consist of temporarily synchronized sub-groups of neurons. Indeed, the Fourier spectrum of a
single neuron does not exhibit any dominant peaks (Figure 3F).
Although the formation of broadband network oscillations is usually attributed to the existence of
inhibition, it is evident that another possible mechanism is intrinsic neuronal response failures that
dynamically drives neural networks to generate coherent oscillations with low averaged firing rates(Vardi
et al., 2015). These observations raise the question of which functionalities demand synaptic inhibition. It
was shown that inhibition slightly suppresses the network firing frequency even further(Vardi et al., 2015)
and it also might change the amplitude of the oscillations. An additional possible hypothesis is that the role
of inhibitory connections is to allow some neuronal computations which are based on conditional temporal
formation of neuronal firing patterns. This type of functionality is an exclusive property of inhibitory
synapses which probabilistically block an evoked spike of its driven nodes in a given time window(Vardi
et al., 2013b;Goldental et al., 2014). In addition, the coexistence of the network oscillations with neuronal
inhibition is intriguing, and especially the questions whether inhibition induces more modes of oscillations,
10
Cooperative oscillations emerge from response failures
sharpens the existing ones, or suppresses the oscillatory behavior and stabilizes the network activity.
Preliminary results of simulations indicate that inhibition might suppress the amplitude of the oscillations
in the low frequency range and sharpen the oscillations in the gamma range. However, results might be
sensitive to the selected parameters.
The interplay between the presented spontaneous cortical oscillations and external stimulations given to
the network is another intriguing question. Specifically, it is interesting to examine the coexistence and the
interplay between the spontaneous oscillation frequencies determined by the network topology and the
frequencies of the induced external stimulations. The understanding of these dynamics will shed light on
the emerging cortical oscillations among coupled networks characterized by different statistical properties.
ACKNOWLEDGEMENTS
This research was supported by the Ministry of Science and Technology, Israel. The authors declare no
competing financial interests.
AUTHOR CONTRIBUTION
R.V. and S.S. prepared and performed the experiments; R.V., S.S. and A.G. analysed the data; P.S. designed
the experimental real-time interface; A.G. performed the simulations and developed the theoretical
framework with help of P.S.; I.K., R.V. and A.G. wrote the manuscript.
11
Cooperative oscillations emerge from response failures
FIGURE 1 Neuronal plasticity – in-vitro experiments. Upper panel: The NRL of a neuron stimulated
at 10, 12 and 15 Hz (light blue, green and purple dots, respectively). Response failures are denoted by
NRL<2 ms. Lower panel: The firing frequency calculated from the averaged ISI using sliding windows of
250 stimulations, or the maximal available one for Stim<250.
12
Cooperative oscillations emerge from response failures
FIGURE 2 Firing bunches stem from neuronal response failures – in-vitro experiments. (A)
Schematic of the neuron in Figure 1 characterized by fc=5 Hz, with 600 and 630 ms self-feedback loops.
(B) A 3.5 s snapshot of the experimental results of (A) where stimulations (blue lines) and their
corresponding evoked spikes (blue dashed lines) were recorded after an offset of to=21 s and a preparation
at 5 Hz stimulation frequency over 300 s to settle the neuron at the intermittent phase. (C) Schematic of a
circuit consisting of three different neurons with fc=6.8 (green), 7.2 (brown) and 4.0 (purple) Hz and 600
and 630 ms delay loops, similar to (A), (different neurons than the one in (A)). (D) A 5 s snapshot of the
experimental results of (C) where ten initial stimulations were given to the central neuron (brown) at 4 Hz
and to=25 s. Stimulations given to the colored neurons (brown/green/purple lines, respectively) and their
corresponding evoked spikes (brown/green/purple dashed lines, respectively). A zoom-in of the gray area
is presented (bottom).
13
Cooperative oscillations emerge from response failures
FIGURE 3 The short-term memory of the stochastic neuronal response failures – in-vitro
experiments and simulations. (A) Schematic of a prototypical examined excitatory network, where each
neuron has two pre- and two post- synaptic connections and the same fc. Inset: The time-lags between
neuronal stimulations is expected to approximately follow an exponential distribution, Exp(2fc). (B) Upper
panel: The stimulation scheme where a neuron is stimulated under the resolution of c/8, such that the
discrete differences between two consecutive stimulations, , follows Exp(2fc). Middle panel: Experimental
NRL of a cultured neuron with fc~5 Hz under a long trial of stimulations following Exp(2fc), response
failures are denoted at NRL=3.5 ms. Lower panel: Zoom-in of the middle panel (gray area) and schematic
of the conditional probability p(sii-1), measuring the probability of a successful response, spike, given
that the current inter-stimulation-interval equals i and the previous one equals i-1. (C) The probabilities
14
Cooperative oscillations emerge from response failures
p(sii-1) obtained from the experimental data in (B) for time>3500 (middle panel). Each colored line
presents p(sii-1) for a given i-1 in c/8 time units (legend). (D) The experimentally measured p(si)
(black), measuring the probability of a successful response, spike, given that the current inter-stimulation-
interval equals i, and the theoretically predicted one (green) using the simplified assumption, p(si)=i/c
for i<c. For both curves, the average ISI~c is preserved. (E) The probability density function of inter-
stimulation-intervals, , for all neurons of Figure 4A (blue). (F) Typical Fourier amplitude of spike timings
of a randomly chosen neuron, taken from Figure 4A.
15
Cooperative oscillations emerge from response failures
FIGURE 4 Cooperative cortical oscillations on a network level. (A) Raster plot of the evoked spikes
(blue dots) obtained in the simulation of a network of 2000 neurons with fc=5.7 Hz. Each neuron has two
randomly selected pre- and post- synaptic connections, and the simulation is based on the experimentally
obtained p(s), (Figure 3D). Delays are randomly selected from U(10,15) ms. The contrast of the raster
was enhanced using a dilution of a constant amount of randomly chosen points in each sliding window of
23 ms, with a step of 0.23 ms. The average dilution is ~60% of the points. (B) The average firing rate per
neuron as a function of time, calculated for windows of 1 ms. (C) The normalized Fourier amplitude, using
a sliding window of 1 Hz, of the entire firing of all neurons over a time slot of 30 s, indicating fosc~3.6 Hz
and f~80 Hz. Inset: The normalized Fourier amplitudes in the range [0,30] Hz obtained from R(m),
equation (1), D=12.5 ms (red) and from the simulation, ((B), blue). (D-F) Similar to (A-C) where each
neuron has on average two pre- and post- randomly chosen synaptic connections and utilizing the theoretical
p(s), (Figure 3D). fosc~4.0 Hz and f~80 Hz at (F). (G-I) Simulation of a network of 2000 neurons where
each neuron has on the average 1.5 pre- and 1.5 post- synaptic above-threshold connections, and 50 pre-
and 50 post- synaptic sub-threshold connections with a strength of 0.03, relative to a threshold of 1. p(si)
16
Cooperative oscillations emerge from response failures
is generalized to an exponential decay function of the neuronal stimulation history, Σi-m/c)exp(-
m)/[Σexp(-m)], =1.4 and the sum is over stimulation history, m≥0. Delays are randomly selected from
U(12.5,20) ms and fc from U(3,10) Hz. fosc~2.6 Hz and f~60 Hz at (I), and the inset is similar to (C) and
(F), but with K=1.5, fc=6.5 Hz and D=16.25 ms.
17
Cooperative oscillations emerge from response failures
FIGURE 5 Scaling properties of fosc and f. (A) Schematic of an excitatory network where a self-
feedback loop (light-red line) and the minimal self-feedback loop (red line) for a given neuron (filled red
circle) are denoted. Similarly, a path (light-blue line) between two neurons (filled blue circles) and the
minimal path (blue line) are denoted. (B) The distribution and its average (vertical lines) for the minimal
path (Path) and for the minimal loop (Loop) obtained in simulations for networks as in (Figure 3A) with
18
Cooperative oscillations emerge from response failures
N=4000, error bars are comparable with the circles. The analytical estimation is shown in green. (C) The
scaling of the averaged quantities in (B) as a function of the average connectivity, <K>. (D) Simulation
results indicate foscln<K>, where N=4000, fc is randomly chosen for each neuron from U(5,15) Hz and
delays are randomly chosen from U(10,15) ms. The probability for a connection between two neurons is
<K>/N. Error bars indicate the standard deviation. (E) Simulation results indicate foscln(fc), for networks
as in (D), with <K>=2, but fc is the same for all neurons. (F) Simulation results indicate f1/<D>, for
networks as in (D), with <K>=2, but delays are randomly chosen from U(<D>-2.5,<D>+2.5) ms.
19
Cooperative oscillations emerge from response failures
REFERENCES
Başar, E., Başar-Eroglu, C., Karakaş, S., and Schürmann, M. (2001). Gamma, alpha, delta, and theta
oscillations govern cognitive processes. International Journal of Psychophysiology 39, 241-248.
Bollimunta, A., Chen, Y., Schroeder, C.E., and Ding, M. (2008). Neuronal mechanisms of cortical alpha
oscillations in awake-behaving macaques. The Journal of neuroscience 28, 9976-9988.
Bollimunta, A., Mo, J., Schroeder, C.E., and Ding, M. (2011). Neuronal mechanisms and attentional
modulation of corticothalamic alpha oscillations. The Journal of Neuroscience 31, 4935-4943.
Brovelli, A., Ding, M., Ledberg, A., Chen, Y., Nakamura, R., and Bressler, S.L. (2004). Beta oscillations in a
large-scale sensorimotor cortical network: directional influences revealed by Granger causality.
Proceedings of the National Academy of Sciences of the United States of America 101, 9849-9854.
Brunel, N., and Wang, X.-J. (2003). What determines the frequency of fast network oscillations with
irregular neural discharges? I. Synaptic dynamics and excitation-inhibition balance. Journal of
neurophysiology 90, 415-430.
Buzsaki, G. (2006). Rhythms of the Brain. Oxford University Press.
Buzsaki, G., and Draguhn, A. (2004). Neuronal oscillations in cortical networks. Science 304, 1926-1929.
Cardin, J.A., Carlén, M., Meletis, K., Knoblich, U., Zhang, F., Deisseroth, K., Tsai, L.-H., and Moore, C.I.
(2009). Driving fast-spiking cells induces gamma rhythm and controls sensory responses. Nature
459, 663-667.
Chialvo, D.R. (2010). Emergent complex neural dynamics. Nature Physics 6, 744-750.
Colgin, L.L., and Moser, E.I. (2010). Gamma oscillations in the hippocampus. Physiology 25, 319-329.
Contreras, D., Destexhe, A., Sejnowski, T.J., and Steriade, M. (1997). Spatiotemporal patterns of spindle
oscillations in cortex and thalamus. J Neurosci 17, 1179-1196.
Crunelli, V., and Hughes, S.W. (2010). The slow (< 1 Hz) rhythm of non-REM sleep: a dialogue between
three cardinal oscillators. Nature neuroscience 13, 9-17.
Cunningham, M.O., Whittington, M.A., Bibbig, A., Roopun, A., Lebeau, F.E., Vogt, A., Monyer, H., Buhl,
E.H., and Traub, R.D. (2004). A role for fast rhythmic bursting neurons in cortical gamma
oscillations in vitro. Proceedings of the National Academy of Sciences of the United States of
America 101, 7152-7157.
Dugladze, T., Schmitz, D., Whittington, M.A., Vida, I., and Gloveli, T. (2012). Segregation of axonal and
somatic activity during fast network oscillations. Science 336, 1458-1461.
Fries, P. (2009). Neuronal gamma-band synchronization as a fundamental process in cortical computation.
Annual review of neuroscience 32, 209-224.
Giraud, A.-L., and Poeppel, D. (2012). Cortical oscillations and speech processing: emerging computational
principles and operations. Nature neuroscience 15, 511-517.
Goldental, A., Guberman, S., Vardi, R., and Kanter, I. (2014). A computational paradigm for dynamic logic-
gates in neuronal activity. Frontiers in computational neuroscience 8.
Gray, C.M. (1994). Synchronous oscillations in neuronal systems: mechanisms and functions. J Comput
Neurosci 1, 11-38.
Grillner, S., Markram, H., De Schutter, E., Silberberg, G., and Lebeau, F.E. (2005). Microcircuits in action–
from CPGs to neocortex. Trends in neurosciences 28, 525-533.
Hasselmo, M.E. (2005). What is the function of hippocampal theta rhythm?—Linking behavioral data to
phasic properties of field potential and unit recording data. Hippocampus 15, 936-949.
Jirsa, V.K., and Haken, H. (1996). Field theory of electromagnetic brain activity. Physical Review Letters 77,
960.
Kahana, M.J. (2006). The cognitive correlates of human brain oscillations. J Neurosci 26, 1669-1672.
Kanter, I., Kopelowitz, E., Vardi, R., Zigzag, M., Kinzel, W., Abeles, M., and Cohen, D. (2011). Nonlocal
mechanism for cluster synchronization in neural circuits. EPL (Europhysics Letters) 93, 66001.
20
Cooperative oscillations emerge from response failures
Klimesch, W. (1996). Memory processes, brain oscillations and EEG synchronization. International Journal
of Psychophysiology 24, 61-100.
Klimesch, W. (1999). EEG alpha and theta oscillations reflect cognitive and memory performance: a review
and analysis. Brain research reviews 29, 169-195.
Marmari, H., Vardi, R., and Kanter, I. (2014). Chaotic and non-chaotic phases in experimental responses
of a single neuron. EPL (Europhysics Letters) 106, 46002.
Minlebaev, M., Colonnese, M., Tsintsadze, T., Sirota, A., and Khazipov, R. (2011). Early gamma oscillations
synchronize developing thalamus and cortex. Science 334, 226-229.
Nir, Y., Mukamel, R., Dinstein, I., Privman, E., Harel, M., Fisch, L., Gelbard-Sagiv, H., Kipervasser, S.,
Andelman, F., and Neufeld, M.Y. (2008). Interhemispheric correlations of slow spontaneous
neuronal fluctuations revealed in human sensory cortex. Nature neuroscience 11, 1100-1108.
Roxin, A., Riecke, H., and Solla, S.A. (2004). Self-sustained activity in a small-world network of excitable
neurons. Physical Review Letters 92, 198101.
Sanchez-Vives, M.V., and Mccormick, D.A. (2000). Cellular and network mechanisms of rhythmic recurrent
activity in neocortex. Nature neuroscience 3, 1027-1034.
Silva, L.R., Amitai, Y., and Connors, B.W. (1991). Intrinsic oscillations of neocortex generated by layer 5
pyramidal neurons. Science 251, 432-435.
Thivierge, J.-P., Comas, R., and Longtin, A. (2014). Attractor dynamics in local neuronal networks. Frontiers
in neural circuits 8.
Vardi, R., Goldental, A., Guberman, S., Kalmanovich, A., Marmari, H., and Kanter, I. (2013a). Sudden
synchrony leaps accompanied by frequency multiplications in neuronal activity. Frontiers in neural
circuits 7.
Vardi, R., Goldental, A., Marmari, H., Brama, H., Stern, E.A., Sardi, S., Sabo, P., and Kanter, I. (2015).
Neuronal Response Impedance Mechanism Implementing Cooperative Networks with Low Firing
Rates and Microsecond Precision. Frontiers in Neural Circuits 9, 29.
Vardi, R., Guberman, S., Goldental, A., and Kanter, I. (2013b). An experimental evidence-based
computational paradigm for new logic-gates in neuronal activity. EPL (Europhysics Letters) 103,
66001.
Vardi, R., Timor, R., Marom, S., Abeles, M., and Kanter, I. (2012a). Synchronization with mismatched
synaptic delays: A unique role of elastic neuronal latency. EPL (Europhysics Letters) 100, 48003.
Vardi, R., Wallach, A., Kopelowitz, E., Abeles, M., Marom, S., and Kanter, I. (2012b). Synthetic
reverberating activity patterns embedded in networks of cortical neurons. EPL (Europhysics
Letters) 97, 66002.
Wang, X.-J. (2010). Neurophysiological and computational principles of cortical rhythms in cognition.
Physiological reviews 90, 1195-1268.
Wiest, M.C., and Nicolelis, M.A. (2003). Behavioral detection of tactile stimuli during 7–12 Hz cortical
oscillations in awake rats. Nature neuroscience 6, 913-914.
Wilson, H.R., and Cowan, J.D. (1972). Excitatory and inhibitory interactions in localized populations of
model neurons. Biophysical journal 12, 1.
Wu, J.-Y., Guan, L., and Tsau, Y. (1999). Propagating activation during oscillations and evoked responses
in neocortical slices. The Journal of neuroscience 19, 5005-5015.
21
|
1904.02523 | 1 | 1904 | 2019-04-04T12:53:35 | Explaining versus Describing Human Decisions. Hilbert Space Structures in Decision Theory | [
"q-bio.NC",
"cs.AI",
"quant-ph"
] | Despite the impressive success of quantum structures to model long-standing human judgement and decision puzzles, the {\it quantum cognition research programme} still faces challenges about its explanatory power. Indeed, quantum models introduce new parameters, which may fit empirical data without necessarily explaining them. Also, one wonders whether more general non-classical structures are better equipped to model cognitive phenomena. In this paper, we provide a {\it realistic-operational foundation of decision processes} using a known decision-making puzzle, the {\it Ellsberg paradox}, as a case study. Then, we elaborate a novel representation of the Ellsberg decision situation applying standard quantum correspondence rules which map realistic-operational entities into quantum mathematical terms. This result opens the way towards an independent, foundational rather than phenomenological, motivation for a general use of quantum Hilbert space structures in human cognition. | q-bio.NC | q-bio |
Explaining versus Describing Human Decisions
Hilbert Space Structures in Decision Theory
Sandro Sozzo
School of Business and Centre IQSCS
University Road LE1 7RH
Leicester (United Kingdom)
Email address: [email protected]
Abstract
Despite the impressive success of quantum structures to model long-standing human judgement and
decision puzzles, the quantum cognition research programme still faces challenges about its explanatory
power. Indeed, quantum models introduce new parameters, which may fit empirical data without nec-
essarily explaining them. Also, one wonders whether more general non-classical structures are better
equipped to model cognitive phenomena. In this paper, we provide a realistic-operational foundation of
decision processes using a known decision-making puzzle, the Ellsberg paradox, as a case study. Then,
we elaborate a novel representation of the Ellsberg decision situation applying standard quantum corre-
spondence rules which map realistic-operational entities into quantum mathematical terms. This result
opens the way towards an independent, foundational rather than phenomenological, motivation for a
general use of quantum Hilbert space structures in human cognition.
Keywords: Quantum structures; Cognitive science; Decision theory; Ellsberg paradox; Operational real-
ism.
1 Introduction
Traditional cognitive theories systematically apply classical set-theoretic structures to model human judge-
ments and decisions under uncertainty. This is particularly evident in theories of rational decision-making,
like expected utility theory, where Bayesian, or Kolmogorovian [1], models of probability directly follow
from axioms on agents' preferences [2, 3].
However, several cognitive puzzles have been discovered in empirical tests, which provide evidence of
systematic deviations from Kolmogorovian probability structures (see, e.g., [4]). For example, Kahneman
and Tversky identified a conjunction fallacy in human probability judgements, namely, the law of mono-
tonicity of Kolmogorovian probability does not generally hold in this kind of judgements [5]. Also, in
human decision-making, Tversky and Shafir proved that the law of total Kolmogorovian probability does
not hold in the disjunction effect [6], while Allais and Ellsberg indicated that people do not always choose
by maximizing an expected utility with respect to a Kolmogorovian probability measure [7].
As a consequence of the puzzles above, traditional theories using Kolmogorovian structures, though
normatively compelling, are descriptively flawed, which led several authors to elaborate alternative pro-
posals able to more efficiently and realistically represent human behaviour. This was the starting point of
the bounded rationality research programme, initially proposed by Herbert Simon [8] and systematically ap-
plied by Kahneman and Tversky [5, 6] to describe concrete judgements and decisions. Bounded rationality
models give good predictions in a variety of circumstances. However, despite their simplicity and intuitive
1
character, these models lack a unitary methodology, as well as deeper explanations, and thus provide a
very fragmented picture of cognitive phenomena [9].
The quantum cognition research programme has recently attracted the interest of the scientific commu-
nity due the superiority of quantum models over traditional and bounded rationality models to deal with
the puzzles above. Quantum models were successfully applied to a variety of complex cognitive processes,
including human language [10, 11, 12, 13], judgement [4, 14, 15] and decision [15, 16, 17] (see also [18]).
Despite these impressive results, however, quantum cognition still raises doubts regarding its explanatory
power. Indeed, on the one side, quantum cognitive models introduce new parameters, which may fit ex-
perimental data, but do not necessarily explain them. On the other side, one is naturally led to wonder
whether cognitive science really needs the entire mathematical formalism of quantum theory in Hilbert
space or, on the contrary, non-Kolmogorovian non-Hilbertian models of probability are needed (see, e.g.,
[19, 20]).
In the present paper, we present binding motivations towards an independent, foundational, rather
than purely phenomenological, justification of the quantum formalism in human judgement and decision-
making under uncertainty. We start from the realistic and operational axiomatizations of quantum physics
initiated by Jauch [21] and Piron [22] in Geneva and extended by Aerts (see, e.g., [23, 24]) in Brussels.
Efforts have been made in the second part of the last century to derive the mathematical formalism of
quantum theory in Hilbert space from more intuitive and empirically justified axioms, resting on basic
notions directly connected with the operations that are performed in a laboratory. Particularly, in the
Brussels approach, any physical entity is expressed in terms of the basic notions of state, context, property
and mutual relationships between them (SCoP system). The approach is realistic, in the sense that the
state, being the result of an effective preparation procedure, describes aspects of the reality of the entity.
The approach is also operational, in the sense that all basic notions are expressed in terms of well defined
empirical terms, like preparation and registration devices, statistics of outcomes, etc. If suitable "purely
operational" axioms are imposed on a SCoP system, then the Hilbert space representation uniquely arises
for the physical entity.
We believe that the above realistic-operational justification of the quantum Hilbert space formalism in
physics also provides a strong motivation, if not a justification in itself, for the use of quantum Hilbert
space structures in cognition. To this end we particularize in Sect. 4 to a specific decision-making situation,
the Ellsberg paradox situation, used as a case study here, the realistic-operational foundation of cognitive
entities we have recently elaborated [25], in which a cognitive entity is abstractly described in terms of
well defined empirical notions, i.e. state, context, property and outcome probability. Then, the stunning
analogies in the realistic and operational descriptions of entities in physical and cognitive realms, suggest
that the same Hilbert space leading axiomatics should be used for a cognitive, e.g., decision-making, entity
(Sect. 2).
The Ellsberg paradox is reviewed in Sect. 3, where we explain the difficulties of both expected utility
and bounded rationality theories, to accommodate Ellsberg preferences and the results of more general
Ellsberg-like decision situations.
We then elaborate in Sect. 5 a mathematical representation in Hilbert space of the Ellsberg paradox
situation and the ambiguity aversion pattern found in empirical literature. We had already presented
quantum models of various Ellsberg thought experiments, including two-color and three-color urns [16,
17, 26, 27, 28]. The novelty of the mathematical representation developed here consists in the fact that
it follows directly from the canonical quantum representation of the realistic-operational terms of state,
context, property and outcome probability in Hilbert space, which makes the use of quantum mathematics
in this kind of situations more firmly founded and generalizable to other decision situations.
We finally offer some conclusive remarks and considerations in Sect. 6, where we specify that the
realistic-operational foundation of cognitive science can be in principle extended to several other judgement
2
and decision-making situations, which constitutes a strong indication that "possible failures of Hilbert space
modelling" should be searched in other cognitive domains than individual judgements and decisions.
2 Descriptive versus explanatory power of quantum structures
Traditional theories of individual judgement and decision-making use, often implicitly, set-theoretic struc-
tures, that closely resemble the formal operations of classical Boolean logic and Kolmogorovian probability
theory [1]. This is specially evident in rational decision theory, according to which rational agents behave
in such a way to maximize expected utility with respect to a Kolmogorovian probability measure and an
underlying economic model [2, 3].
These theories are normatively compelling, however, the judgement and decision puzzles in Sect. 1 make
them descriptively problematical and suggest alternative more realistic approaches to human behaviour
under uncertainty. A major research programme of this kind was initiated by Herbert Simon who put
forward the bounded rationality project [8]. Boundedly rational agents experience practical limitations in
formulating and solving complex problems and in processing information. They tackle such limitations by
taking mental short-cuts, making subjective evaluations and putting psychological aspects above rational
reasoning.
Within the bounded rationality project, one can cope with cognitive puzzles with judgement heuristics
and reasoning biases, namely, the conjunction fallacy with the representativeness heuristics [5], the Allais
paradox with prospect theory [6], the Ellsberg paradox with cumulative prospect theory [6], the disjunction
effect by uncertainty aversion [6], etc. These approaches undoubtedly provide an intuitive account of
how individuals actually behave in situations of uncertainty. However, the reader recognizes at once that
a rather eclectic methodology or, better, a variety of methodologies, are employed to accommodate the
puzzles above and, while some authors support the hypothesis of an adaptive toolbox to deal with these
problems [29], many psychologists will find the bounded rationality research programme as unsatisfactory,
while many philosophers of science will try to derive these puzzling phenomena from a universal theory
able to overcome the fragmentation of existing approaches.
The quantum cognition research programme reaches both effectiveness and unitarity. Since the nineties,
quantum Hilbert space models have shown impressive superiority over traditional and bounded rationality
approaches in dealing with the puzzles of human cognition and attributing them to genuine quantum effects,
like contextuality, emergence, entanglement, interference and superposition. On the other side, quantum
models introduce new parameters which can be possibly fitted by empirical data, without however nec-
essarily explaining them. Hence, the quantum cognition research programme, though phenomenologically
successful, does not seem to offer a deeper understanding and/or explanation of these puzzles. In addi-
tion, it is reasonable to wonder whether one really needs the entire Hilbert space formalism to represent
cognitive phenomena, and should not better use more general non-Kolmogorovian representations outside
physics (see, e.g., [19, 20]).
In this respect, it should be noted that prospect theory already proposes
non-Kolmogorovian probability models of probability of human decision [6].
It is clear from the considerations above that one needs a deeper justification for the use of the Hilbert
space formalism of quantum theory in cognition and decision and, more important, of its necessity. In this
respect, a crucial result comes from increasing evidence that "judgements and decisions create rather than
record" [4] -- see. e.g., the following quotations.
"There is a growing body of evidence that supports an alternative conception according to
which preferences are often constructed not merely revealed in the elicitation process. These
constructions are contingent on the framing of the problem, the method of elicitation, and the
context of the choice." [30]
3
". . . the process of choice -- and in particular the act of choice -- can make substantial difference
to what is chosen.
. . . , there is a particular necessity to take note of (i) chooser dependence,
and (ii) menu dependence, of preference, even judged from a particular person's perspective."
[31]
". . . valuations are initially malleable but become imprinted after the agent is called upon to
make an initial decision." [32]
It is more and more acknowledged that, in any judgement or decision, a contextual interaction occurs
between the situation that is the object of the evaluation and the individual who takes the decision (agent,
decision-maker), which may affect the situation itself. At the end of this interaction, a result is actualized
among a set of results that were only potential before the interaction [26]. Hence, a judgement/decision
process closely resembles a quantum measurement process, where a contextual interaction occurs between
the quantum particle that is measured and the measurement apparatus, which changes the state of the
quantum particle determining what Heisenberg called "transition from potential to actual".
We believe that these analogies between micro-physics and cognition are a good starting point towards
a foundational justification for the use of Hilbert space quantum formalism in cognition and decision.
In the sixties and seventies of the previous century, several authors wondered whether and how one can
provide an independent justification for the Hilbert space formalism in quantum physics, deriving this for-
malism from physically justified axioms, resting on well defined empirical notions, directly connected with
the operations that are usually performed in a laboratory. One of the well-known approaches to the foun-
dations of quantum physics is the Geneva-Brussels realistic-operational approach, initiated by Jauch [21]
and Piron [22] in Geneva, and successively extended by Aerts in Brussels (see, e.g., [23, 24]). This research
consisted in abstractly describing any physical entity by relevant sets of states, contexts, properties and
statistical connections between these notions (SCoP system). These theoretical notions are directly inter-
pretable on physical operations on macroscopic apparatuses, such as preparation and registration devices,
performed in spatio-temporal domains, such as physical laboratories. Measurements, state transformations,
outcome probabilities and dynamics can then be expressed in terms of these more fundamental notions. If
suitable axioms are imposed on the mathematical structures underlying a SCoP system, then the Hilbert
space structure of quantum theory emerges as a unique mathematical representation, up to isomorphisms
[33]. This justification provides the "fundamental architecture of quantum theory in Hilbert space".
We have recently proved that any cognitive entity Ω, e.g., a concept, a conceptual combination, a
proposition, or a more complex decision-making situation, can be abstractly described by a SCoP system
(Σ, L , C , µ, ν) [25], exactly like in physics. We review the essential elements of a SCoP system in cognition
in the following.
(1) The complex of experimental procedures conceived by the experimenter, the experimental setting
and the cognitive effect that one wants to analyse, define a cognitive entity Ω, and are usually associated
with a preparation procedure of a state of Ω.
(2) Σ is the set of all states of Ω. A state p of Ω is the consequence of a preparation procedure of Ω
and has a cognitive, rather than physical, nature. The state of the cognitive entity is a state of affairs.
It indeed expresses a "reality of the cognitive entity", in the sense that, once prepared in a given state,
such condition is independent of any measurement procedure, and can be confronted with the different
participants in an experiment, leading to outcome data and their statistics.
(3) C is the set of all contexts of Ω. A context e is an element that can provoke a change of state of the
cognitive entity. A special context is the one introduced by a measurement. Indeed, when the cognitive
experiment starts, an interaction occurs between the measured entity Ω under study and a participant in
the experiment, in which the state p of Ω generally changes, being transformed to another state q. This
cognitive interaction is formalized by means of a context e.
4
(4) L is the set of all properties of Ω. A property a of Ω is something Ω "has" independently of any
context influencing the entity. An entity Ω is a given state p has a set of properties that are actual in that
state, the others being potential. A context e may change the status actual/potential of a property, but
cannot change the property itself.
(5) The change function µ : Σ × C × Σ −→ [0, 1] is such that, for every p, q ∈ Σ, e ∈ C , µ(q, e, p) is the
probability, as the large number limit of relative frequencies, that the context e changes the initial state p
of Ω to the final state q.
Once recognizes at once in (1) -- (5) the building blocks of the realistic-operational description of a
physical entity, in the sense that in both physical and cognitive realms, a SCoP system incorporates all
what is needed to study what an entity is, behaves and changes under a context. These impressive analogies
indicate that the axioms generally used to justify the Hilbert space formalism of quantum physics are also
appropriate to represent cognitive entities and processes. This provides an independent foundational clue
and non-phenomenological motivation, if not a justification, that the mathematics of Hilbert space should
be used to represent judgement and decision phenomena.
In Sect. 4, we will provide a realistic-operational description of a specific decision-making situation, the
Ellsberg paradox, setting the grounds for a quantum mathematical representation of it in Sect. 5. In the
next section, we will instead summarize the serious difficulties of both traditional and bounded rationality
approaches to handle such kind of decision-making situations.
3 Rational decision theory and its puzzles
Traditional theories of rational decision-making rest on the tenet that, in situations of uncertainty, indi-
vidual agents choose in such a way to maximize their expected utility, or degree of satisfaction.
In 1944, von Neumann and Morgenstern presented in a seminal work the first axiomatic formulation
of expected utility theory. People continuously take decisions among different options. These decisions
are assumed to reveal underlying preferences. Then, von Neumann and Morgenstern proposed a set of
"reasonable" axioms on human preferences such that, if the decisions are coherent, in the sense that they
reveal axiom satisfying preferences, then the decisions are equivalent to the maximization of an expected
utility functional with respect to a Kolmogorovian probability measure [2].
von Neumann and Morgenstern's formulation of expected utility theory has a major limitation, in
that it only deals with the uncertainty that can be formalized by known probabilities (also referred to
as objective uncertainty, or risk). On the other hand, situations frequently occur in which uncertainty
cannot be formalized by known probabilities (also referred to as subjective uncertainty, or ambiguity)
[34]. The Bayesian approach to probability minimizes the distinction between objective and subjective
uncertainty introducing the notion of subjective probability. Even when probabilities are not known, people
may still construct their own beliefs, or priors (which may differ from one individual to another), and
they maximize expected utility with respect to these priors. Indeed, Leonard Savage presented in 1954 an
axiomatic formulation of expected utility theory which extends the one of von Neumann and Morgenstern
to subjective uncertainty [3].
We summarize in the following the essential definitions and results of Savage's expected utility theory,
together with its major pitfalls. We refer to [35, 36] for detailed reviews of these results.
Savage introduced a set of basic notions, including states of nature, consequences, preferences, and
looked for justified axioms on preferences able to provide a representation theorem in which ordering of
preferences is characterized by maximization of expected utility. This procedure formally resembles the
procedures used in the axiomatizations of quantum physics in Sect. 4.
Let S be the set of all (physical) states of nature, which we assume to be discrete and finite here, for
the sake of simplicity. Let P(S ) be the power set of S and A ⊆ P(S ) be a (Boolean) σ-algebra. An
5
element E ∈ A denotes an event. A Kolmogorovian probability measure over A is a function p : A ⊆
P(S ) −→ [0, 1] satisfying the axioms of Kolmogorov [1].
Then, let X be the set of all consequences, whose elements we assume to denote monetary payoffs,
hence real numbers, here, for the sake of simplicity. In Savage's formulation, a function f : S −→ X
denotes an act. Let F be the set of all acts. Let us endow F with a weak preference relation (cid:37), that is, a
reflexive, symmetric and transitive relation over the Cartesian product F × F . In (cid:37), the relations (cid:31) and
∼ denote strong preference and indifference, respectively, that is, we write f (cid:31) g whenever an individual
strictly prefers act f to act g and f ∼ g whenever the individual is indifferent between f and g.
Next, let u : X −→ (cid:60) be a utility function over X . This function typically expresses the decision-
maker's taste, hence it is assumed to be strictly increasing and continuous, with additional technical
constraints related to the specification of the decision-maker's attitude towards risk.
The mathematical definitions above can be simplified by introducing a set {E1, . . . , En} of mutually
exclusive and exhaustive elementary events, where Ei = {si ∈ S }, i ∈ {1, . . . , n}, which thus form a
partition of S . For every i ∈ {1, . . . , n}, let xi be the utility associated by the act f to the event Ei. Then,
f can be equivalently expressed by the 2n-tuple f = (E1, x1; . . . ; En, xn), meaning that the individual will
get the outcome x1 if the event E1 occurs (i.e. the state of nature s1 realizes), . . . , the outcome xn if the
event En occurs (i.e. the state of nature sn realizes).
Finally, we denote by W (f ) = (cid:80)n
i=1 p(Ei)u(xi) the expected utility associated with the act f with
respect to the Kolmogorovian probability measure p.
In his representation theorem, Savage proved that, if the algebraic structure (F , (cid:37)) satisfies a number
of "reasonable" axioms1 then, for every f, g ∈ F , a unique Kolmogorovian probability measure p and a
unique (up to positive affine transformations) utility function u exist such that f is preferred to g, i.e.
f (cid:37) g, if and only if the expected utility of f is greater than the expected utility of g, i.e. W (f ) ≥ W (g).
For every i ∈ {1, . . . , n}, the utility value u(xi) depends on the decision-maker's risk preferences, while
p(Ei) is interpreted as the subjective probability, expressing the individual's belief that the event Ei occurs
[3].
Savage's result is both compelling at a normative level and testable at a descriptive level. Indeed:
(i) if the axioms are intuitively reasonable and decision-makers agree with them, then they must all
behave as if they were maximizing an expected utility with respect to a single subjective probability
distribution satisfying Kolmogorov's axioms;
(ii) the axioms suggest to design decision-making experiments to test the validity of expected utility
theory, hence of the axioms themselves, in real life situations.
Because of (i), Savage's expected utility formulation is generally accepted to prescribe "how rational
agents should choose". However, one the one side, the theory offers very little about where beliefs come
from and how they should be calculated and, on the other side, regarding (ii), decision-making experiments
have systematically found deviations from that rational behaviour in concrete situations.
In particular, Daniel Ellsberg proved in 1961 in a number of thought experiments that decision-makers
generally prefer acts with known (or objective) probabilities over acts with unknown (or subjective) prob-
abilities [7]. We analyse here the famous Ellsberg three-color example as a paradigmatic example to show
that (i) traditional decision theories do not work, (ii) bounded rationality approaches are not sufficiently
explanatory, (iii) quantum structures are needed.
Consider one urn with 30 red balls and 60 balls that are either yellow or black, the latter in unknown
proportion. One ball will be drawn at random from the urn. Then, free of charge, a person is asked to bet
1One of the axioms is the famous sure thing principle, which is violated in the Ellsberg paradox. The other axioms
are: ordinal event independence, comparative probability, non-degeneracy, small event continuity and dominance, and have a
technical nature [3]. However, these axioms are not relevant to the present purposes, hence we will not dwell on them, for the
sake of brevity.
6
Act
f1
f2
f3
f4
1/3
2/3
Red
$100
$0
$100
$0
Yellow
$0
$0
$100
$100
Black
$0
$100
$0
$100
Table 1. The payoff matrix for the Ellsberg three-color example.
on pairs of the acts f1, f2, f3 and f4 in Table 1. Ellsberg suggested that, when asked to rank these acts, most
individuals will prefer f1 over f2 and f4 over f3. Indeed, f1 and f4 are unambiguous acts, in the sense that
they are associated with events over known probabilities -- the events "a red ball is drawn" and "a yellow or
black ball is drawn" are associated with objective probabilities 1/3 and 2/3, respectively. On the contrary,
f2 and f3 are ambiguous acts, in the sense that they are associated with events over unknown probabilities
-- the events "a yellow ball is drawn" and "a black ball is drawn" are both associated with a probability
ranging from 0 to 2/3. This attitude of decision-makers to prefer "probabilized over non-probabilized
uncertainty" has been known as ambiguity aversion since Ellsberg studies [7].
Several experiments on Ellsberg urns decisions, but also on financial, insurance and medical decisions,
have confirmed the Ellsberg preferences f1 (cid:31) f2 and f4 (cid:31) f3, thus indicating that ambiguity aversion is
a good candidate to explain concrete decisions in this case, and only Slovic and Tversky found ambiguity
seeking patterns (see, e.g., [36] for a review of experimental studies).
In [27], we tested various human decision puzzles, including the Ellsberg three-color example. We
asked 200 people, chosen among colleagues and friends, to fill a questionnaire in which they had to choose
between various options. People had on average a basic knowledge of probability theory, but no specific
training in decision theory. Participants were provided with a questionnaire similar to the one in Figure 1,
in which they had to choose between acts f1 and f2 and, then, between acts f3 and f4 in Table 1. Overall,
125 participants preferred acts f1 and f4, 38 preferred acts f1 and f3, 6 preferred acts f2 and f3, and 31
preferred acts f2 and f4. This means that 163 participants over 200 preferred act f1 over act f2, which
entails a preference weight of 0.815. Also, 156 participants over 200 preferred act f4 over act f3, which
entails a preference weight of 0.780. The inversion rate is 0.655, a pattern that agrees with the Ellsberg
preferences found in the literature and significantly indicates the presence of ambiguity aversion.
Preferences of decision-makers who are sensitive to ambiguity, that is, are ambiguity averse or ambiguity
seeking, cannot be explained within Savage's expected utility theory, because they violate the sure thing
principle, according to which, preferences should be independent of the common outcome. In the specific
case of the three-color example, preferences should not depend on whether the common event "a yellow
ball is drawn" pays off $0 or $100. More technically, Savage's expected utility theory predicts consistency
of preferences, namely, f1 is preferred to f2 if and only if f3 is preferred to f4. A simple calculation
shows that this is impossible within a traditional expected utility framework.
Indeed, if we denote by
pR, pY and pB the probability that a red ball, a yellow ball, a black ball, respectively, is drawn (with
pR = 1/3 = 1 − (pY + pB)), then the expected utilities W (fi), i = 1, 2, 3, 4, are such that W (f1) > W (f2)
if and only if (pR − pB)(u(100)− u(0)) > 0 if and only if W (f3) > W (f4). We can equivalently say that no
assignment of Kolmogorovian probabilities pR, pY and pB reproduces a preference with W (f1) > W (f2)
and W (f4) > W (f3), whence the Ellsberg paradox.
Several extensions of Savage's expected utility theory have been put forward in order to accommodate
the Ellsberg paradox (see, e.g., the reviews in [35] and [36]). One of the major proposals is Tversky and
Kahneman's cumulative prospect theory, mentioned in Sect. 2 and elaborated within a bounded rationality
research programme [6]. In particular, to reproduce Ellsberg preferences, Tversky and Kahneman replaced
7
Figure 1. A sample of the questionnaire related to the decision-making experiment on the Ellsberg
three-color example: choice between acts f1 and f2 in Table 1.
8
(i) the utility function u by a scaling function u(cid:48) reflecting the subjective value of the outcome utility,
and (ii) the subjective probability measure p by a non-additive measure p(cid:48) satisfying the mathematical
properties of a capacity. As we have mentioned in Sect. 2, such bounded rationality models, though
descriptively interesting and easily interpretable intuitively, provide a too fragmented view of decision
theory, hence they are not able to provide a unitary and adequate explanatory framework to understand
the deep aspects of decision processes. In addition, cumulative prospect theory, as well as other major
non-expected utility models, fails to reproduce the empirical results of a recently elaborated variant of the
Ellsberg paradox, the Machina paradox [27, 37, 38].
An innovative aspect of descriptive, like bounded rationality, approaches, is the representation of subjec-
tive probabilities by more general, possibly non-Kolmogorovian, mathematical structures. This is crucial
towards a more satisfactory framework for human decision-making that goes beyond Savage's expected
utility, as we will see in Sect. 5.
In the next section, we intend to elaborate a realistic-operational description of a decision-making
situation, using the Ellsberg three-color example as a case study. We will demonstrate that, once the
Ellsberg paradox situation is formulated in terms of states, contexts, properties and transition probabilities,
then the application of the mathematical formalism of quantum theory directly follows from the canonical
representation of these realistic-operational terms in Hilbert space.
4 A realistic-operational description of a decision-making situation
In this section, we specify the realistic-operational description of cognitive entities in Sect. 2 to the decision-
making situation presented in the Ellsberg three-color example [25]. In it, we explicitly distinguish physical
from cognitive, in this case, decision-making, entities. Analogously, we distinguish physical from cognitive
states of nature, though one can intuitively see that some cognitive states are mapped into the corresponding
physical states.
In the Ellsberg three-color example, the cognitive, i.e. decision-making, entity ΩDM is the urn with
30 red balls and 60 yellow or black balls in unknown proportion. This is what the individual reads in a
questionnaire, interacts with and takes a decision on.
The cognitive entity ΩDM is associated with a defined set ΣDM of states.2 A state p ∈ ΣDM has a
cognitive nature and incorporates aspects of ambiguity. A context e does not pertain to ΩDM but can
interact with it. Let CDM be the set of all contexts of ΩDM . The interaction of ΩDM with a context
e ∈ CDM may determine a change of the state of ΩDM from p to a different state q. The probability of such
a state transition will be denoted by µ(q, e, p), where µ : ΣDM ×CDM ×ΣDM −→ [0, 1]. We might complete
the realistic-operational description of ΩDM defining a set LDM of properties and an actuality relation
connecting properties and states. However, they are not needed in the Ellsberg three-color scenario, hence
we omit specifying these notions here, for the sake of brevity, though they may be relevant in more general
decision situations.
Let us now introduce a color context eC ∈ CDM , which is the context associated with a drawing of a
ball from the urn. As a result of the drawing, we have three possible outcomes, R, Y and B, corresponding
to the colors of the balls, red, yellow and black, respectively. The outcomes R, Y and B are the eigenvalues
of eC and are respectively associated with the final states, or eigenstates, pR, pY and pB of the cognitive
entity ΩDM . These eigenstates are such that µ(pi, eC, pi) = 1, i ∈ {R, Y, B}.
2Some authors identify the notion of "state" with the notion of "belief state" of the individual participating in the cognitive
experiment, e.g., taking the decision (see, e.g., [4, 9, 15]). We instead neatly distinguish states from measurements here. A
state is defined by a preparation procedure of the cognitive entity under investigation. The participant in the experiment acts
as a (measurement) context that interacts with the cognitive entity and changes its state.
9
The color context eC introduces three mutually exclusive and exhaustive elementary events Ei =
(eC,{i}), i ∈ {R, Y, B}, which are such that the subjective probability that the event Ei occurs when the
cognitive entity ΩDM is in the state p is given by the transition probability µ(pi, eC, p) ∈ [0, 1], i ∈ {R, Y, B}.
Then, in analogy with Savage's expected utility theory, we can introduce monetary payoffs x ∈ X ⊆ (cid:60),
utility functions u : X −→ (cid:60) and acts taking the form f = (ER, xR; EY , xY , EB, xB), mapping the event
Ei into the payoff xi, i ∈ {R, Y, B}. In particular, the acts f1, f2, f3 and f4 in Table 1, Sect. 3 are defined
in the way above.
Let us now come to the operational-realistic description of a decision-making process in the Ellsberg
three-color situation. Suppose that, in the absence of any context, the entity ΩDM is in the initial state
p0. This state corresponds to a preparation of the cognitive entity ΩDM and can be set by the information
on the corresponding physical entity. For example, it is reasonable to assume that p0 is such that, for
every i ∈ {R, Y, B}, µ(pi, eC, p0) = 1/3, because of the indifference principle. Whenever an individual
is asked to rank f1 and f2, the individual's attitude towards ambiguity, e.g., ambiguity aversion, can be
described as a new context e1 ∈ CDM acting on ΩDM in the initial state p0 and changing p0 into a new state
p1, characterized by a new probability distribution µ(pi, eC, p1), i ∈ {R, Y, B}. Similarly, whenever the
individual is asked to rank f3 and f4, the individual's attitude towards ambiguity, e.g., ambiguity aversion,
can be described as a new context e2 ∈ CDM acting on ΩDM in the initial state p0 and changing p0 into
a new state p2, characterized by a new probability distribution µ(pi, eC, p2), i ∈ {R, Y, B}. The cognitive
states p1 and p2, and their ensuing probability distributions, are responsible of the inversion of preferences
which occur in the Ellsberg paradox situation.
Finally, a decision process between acts f1 and f2 can be operationally described as a measurement
context e12 ∈ CDM acting on the entity ΩDM in the ambiguity averse state p1, with possible outcomes
"yes" and "no". Similarly, a decision process between acts f3 and f4 can be described as a measurement
context e34 ∈ CDM acting on the entity ΩDM in the ambiguity averse state p2, with possible outcomes
"yes" and "no". These contexts give rise of the statistics of outcomes in a decision-making test on the
Ellsberg three-color urn.
Now, the considerations in Sect. 2 naturally indicate to represent states, contexts, properties and
outcome probabilities of ΩDM by using the canonical quantum representation of states, contexts, properties
and outcome probabilities in Hilbert space. In particular, subjective probabilities will be represented using
the Born rule of quantum probability. This is what we intend to show in the next section where the realistic-
operational terms defined here will be canonically represented using quantum mathematical terms.
5 A novel quantum representation of the Ellsberg paradox
In this section we elaborate a new quantum representation of the three-color example straightly follow-
ing the canonical Hilbert space representation of the realistic and operational notions in Sect. 3. This
representation generalizes and strengthens those in [16, 17, 26, 27].
5.1 Quantum representation of basic notions
The cognitive entity ΩDM is associated with a Hilbert space H . Since, the three-color example involves
three mutually exclusive and exhaustive elementary events, H can be chosen to be isomorphic to the
complex Hilbert space C3 of ordered triples of complex numbers. Let {(1, 0, 0), (0, 1, 0), (0, 0, 1)} be the
canonical orthonormal basis of C3.
A state pv of the entity ΩDM is represented by the unit vector v(cid:105) ∈ H , v(cid:105) =(cid:112)(cid:104)vv(cid:105) = 1.
The elementary event Ei is represented by the one-dimensional orthogonal projection operator Pi =
i(cid:105)(cid:104)i, i ∈ {R, Y, B}, where we choose R(cid:105) = (1, 0, 0), Y (cid:105) = (0, 1, 0) and B(cid:105) = (0, 0, 1). The color context
10
eC is then represented by the spectral family {PR, PY , PB}.
In the canonical basis of C3, the unit vector v(cid:105) can be written as
v(cid:105) = ρReiθRR(cid:105) + ρY eiθY Y (cid:105) + ρBeiθBB(cid:105) = (ρReiθR, ρY eiθY , ρBeiθB )
(1)
We use the Born rule to represent subjective probabilities. Then, the subjective probability that the
elementary event Ei, i ∈ {R, Y, B}, occurs when the entity ΩDM is in the state pv is given by
µv(Ei) = (cid:104)vPiv(cid:105) = (cid:104)iv(cid:105)2 = ρ2
i
(2)
In addition, the subjective probability that the event E, represented by the orthogonal projection operator
PE, occurs when the cognitive entity ΩDM is in the state pv is µv(E) = (cid:104)vPEv(cid:105) = PEv(cid:105)2. Finally, for
every state pv represented by the unit vector v(cid:105), the subjective probability measure
(3)
is a quantum probability measure over the lattice L (C3) of all orthogonal projection operators on the
Hilbert space C3.
Compatibility with the standard three-color situation entails ρ2
R = 1
3 , hence it follows from Eq. (1) that
µv : L (C3) −→ [0, 1]
v(cid:105) = (
where the last equality is obtained from(cid:112)(cid:104)vv(cid:105) = 1.
eiθR, ρY eiθY , ρBeiθB ) = (
1√
3
1√
3
eiθR, ρY eiθY ,
Special states are the state with no black balls represented by
and the state with no yellow balls represented by
vRY (cid:105) = (
1√
3
eiθR,
(cid:114) 2
3
eiθY , 0)
(cid:114) 2
(cid:114) 2
3
− ρ2
yeiθB )
(4)
(5)
(6)
(7)
(8)
(9)
The acts f1, f2, f3 and f4 are represented by the self-adjoint operators
vRB(cid:105) = (
1√
3
eiθR, 0,
eiθB )
3
F1 = u(100)PR + u(0)PY + u(0)PB
F2 = u(0)PR + u(0)PY + u(100)PB
F3 = u(100)PR + u(100)PY + u(0)PB
F4 = u(0)PR + u(100)PY + u(100)PB
(10)
respectively. The utility function u(·) is not given by the theory but it is revealed in a decision test by
concrete choices, in analogy with standard procedures in the literature.
The expected utility Wv(fi), i ∈ {1, 2, 3, 4}, in a generic state pv of the entity ΩDM is
Wv(f1) = (cid:104)v F1v(cid:105) =
1
3
Wv(f2) = (cid:104)v F2v(cid:105) = (
1
3
Wv(f3) = (cid:104)v F3v(cid:105) = (
1
3
Wv(f4) = (cid:104)v F4v(cid:105) =
1
3
u(100) +
2
3
u(0)
+ ρ2
Y )u(0) + (
2
3
+ ρ2
Y )u(100) + (
− ρ2
2
3
Y )u(100)
− ρ2
Y )u(0)
u(0) +
2
3
u(100)
11
(11)
(12)
(13)
(14)
As we can see, the expected utilities Wv(f1) and Wv(f4) do not depend on the cognitive state pv of the
entity ΩDM , in agreement with the fact that f1 and f4 are unambiguous acts. On the contrary, the expected
utilities Wv(f2) and Wv(f3) do depend on the cognitive state pv, in agreement with the fact that f2 and
f3 are ambiguous acts. This also agrees with our assumption that cognitive states provide information on
ambiguity.
5.2 Reproducing Ellsberg preferences with ambiguity averse states
Let us now suppose that, in the absence of any context, the cognitive entity ΩDM is in the initial state p0.
The principle of indifference (see Sect. 4) then suggests that p0 is the state represented by the unit vector
v0(cid:105) =
1√
3
(1, 1, 1)
(15)
leading to uniform probabilities of drawing a red, yellow and black ball. The ambiguity attitude contexts
e1 and e2 will determine a change of state of the entity ΩDM , depending on individual preferences to-
wards ambiguity. For example, two ambiguity seeking states pw1 and pw2 will be such that the following
inequalities hold
Ww1(f2) >
Ww2(f3) >
1
3
1
3
u(100) +
2
3
u(0)
u(0) +
2
3
u(100)
(16)
(17)
We will instead explicitly determine two ambiguity averse states pw1 and pw2 which reproduce Ellsberg
preferences, that is,
Two general cognitive states pw1 and pw2 states are represented by the unit vectors
Ww1(f1) > Ww1(f2)
Ww2(f4) > Ww2(f3)
w1(cid:105) = (
w2(cid:105) = (
1√
3
1√
3
eiθR, ρY eiθY ,
eiφR, τY eiφY ,
− ρ2
yeiθB )
− τ 2
y eiφB )
(cid:114) 2
(cid:114) 2
3
3
(18)
(19)
(20)
(21)
(22)
(23)
respectively. For the sake of simplicity, let us look for states with simple phases, namely, θR = φR = 0 and
θY , θB, φY , φB = 0, π. In particular, one can show that, for every α > 1√
, the unit vectors
3
(cid:114) 2
(cid:114) 2
3
− α2)
3
− α2, α)
w1(cid:105) = (
w2(cid:105) = (
1√
3
1√
3
, α,−
,−
reproduce Ellsberg preferences. However, the ambiguity averse states pw1 and pw2 are not generally or-
thogonal, unless α = ±0.7887. Let us choose the positive sign, so that the orthonormal vectors
w1(cid:105) = (0.5774, 0.7887,−0.2113)
w2(cid:105) = (0.5774,−0.2113, 0.7887)
(24)
(25)
reproduce Ellsberg preferences in the three-color example within a quantum mathematical representation.
12
5.3 Modelling empirical data in Hilbert space
The final step of the quantum representation of the three-color example consists in modelling the experimen-
tal data in Sect. 3. To this aim, we describe the decision between acts f1 and f2 by a measurement context
e12 and represent the latter by the spectral family {M, 1 − M}, where the orthogonal projection operator
M projects onto the one-dimensional subspace generated by the unit vector m(cid:105) = ( 1√
, ρY eiθY , ρBeiθB ),
where ρ2
3
Y + ρ2
B = 2
3 . The one-dimensional projection operator M is then given by
ρBe−iθB
1√
3
ρY ρBei(θY −θB)
M = m(cid:105)(cid:104)m =
ρY e−iθY
ρ2
Y
1√
3
1
3
1√
3
1√
3
ρY eiθY
ρBeiθB
ρY ρBe−i(θY −θB)
ρ2
B
Analogously, we describe the decision between acts f3 and f4 by a measurement context e34 and represent
the latter by the spectral family {N, 1 − N}, where the orthogonal projection operator N projects onto
the one-dimensional subspace generated by the unit vector n(cid:105) = ( 1√
B = 2
3 .
The one-dimensional projection operator N is then given by
, τY eiφY , τBeiφB ), where τ 2
Y + τ 2
3
N = n(cid:105)(cid:104)n =
1
3
1√
3
τY e−iφY
τ 2
Y
τBe−iφB
1√
3
τY τBei(φY −φB)
τY eiφY
τBeiφB
1√
3
1√
3
(1, 1, 1) and w1(cid:105) and w2(cid:105) in Eqs. (24) and (25). It follows that the
τY τBe−i(φY −φB)
(27)
τ 2
B
We refer to the unit vectors v0(cid:105) = 1√
conditions
3
(cid:104)mm(cid:105) = 1
(cid:104)nn(cid:105) = 1
(cid:104)w1Mw1(cid:105) = 0.815
(cid:104)w2Nw2(cid:105) = 0.780
(cid:104)v0Mv0(cid:105) = 0.500
(cid:104)v0Nv0(cid:105) = 0.500
(28)
(29)
(30)
(31)
(32)
(33)
must be satisfied by the real parameters ρY , ρB, θY , θB, τY , τB, φY and φB. Equations (28) and (29)
are determined by normalization conditions, while Eqs. (30) and (31) are determined by empirical data.
Finally, Eqs.
(32) and (33) are determined by the fact that decision-makers who are not sensitive to
ambiguity should overall be indifferent between f1 and f2, as well as between f3 and f4. Hence, on average,
half respondents are expected to prefer f1 (f3) and the other half f2 (f4). To simplify the analysis, let us
set θY = φB = 0. Hence, we are left with a system of 6 equations in 6 unknown variables whose solution is
(26)
(34)
(35)
(36)
ρY = 0.6853
ρB = 0.4438
θB = 105.07◦
τY = 0.4785
τB = 0.6616
φY = 102.87◦
Equivalently, we get
m(cid:105) = (0.5773, 0.6853, 0.4438ei105.07◦
n(cid:105) = (0.5773, 0.4785ei102.87◦
, 0.6616)
)
13
Thus, the orthogonal projection operators in Eqs. (26) and (27) reproducing the experimental data in
Sect. 3 are
0.333
0.396
0.256ei105.07◦
0.333
0.276ei102.87◦
0.382
0.396
0.470
0.304ei105.07◦
0.276e−i102.87◦
0.317e−i102.87◦
0.229
0.256e−i105.07◦
0.304e−i105.07◦
0.197
0.382
0.317ei102.87◦
0.438
M =
N =
(37)
(38)
The construction of a quantum model for the data on the Ellsberg three-color experiment in Sect. 3 is
thus completed. As we can see, the quantum model naturally arises from the canonical Hilbert space
representation of the realistic-operational termsin Sect. 4.
6 Conclusions
Despite its phenomenological success to deal with classically problematical cognitive puzzles, the quan-
tum cognition research programme still poses challenging questions regarding its explanatory power and
necessity.
In this paper, we specialized to the Ellsberg paradox decision situation a realistic-operational foundation
which we have recently extended from physics to cognition. Then, we applied to the Ellsberg three-color
example the canonical quantum representation of realistic-operational terms in Hilbert space.
This result on the Ellsberg paradox situation is paradigmatic, in the sense that one can follow the same
strategy to generally claim that the mathematical representation of human judgements and decision-making
in Hilbert space has now an independent motivation of a foundational, rather than phenomenological,
nature.
To conclude, we agree that quantum theory in Hilbert space is not the ultimate theory in cognition --
recent results on sequential measurements and order effects seem to confirm this conclusion(see, e.g., [25]).
However, we also believe that there are strong theoretical motivations, in addition to its empirical success
and unitary explanation, to continue using Hilbert space structures in cognition.
Acknowledgements
This work was supported by QUARTZ (Quantum Information Access and Retrieval Theory), the Marie
Sk(cid:32)lodowska-Curie Innovative Training Network 721321 of the European Union's Horizon 2020 research and
innovation programme.
References
[1] Kolmogorov, A.N.: Grundbegriffe der Wahrscheinlichkeitrechnung, Ergebnisse Der Mathematik
(1933); translated as: Foundations of Probability. Chelsea Publishing Company, New York (1950)
[2] von Neumann, J., Morgenstern, O.: Theory of Games and Economic Behavior. Princeton University
Press, Princeton (1944)
[3] Savage, L.: The Foundations of Statistics. John Wiley & Sons, New York (1954); revised and enlarged
edition: Dover Publications, New York (1972)
14
[4] Busemeyer, J.R., Bruza, P.D.: Quantum Models of Cognition and Decision. Cambridge University
Press, Cambridge (2012)
[5] Kahneman, D., Slovic, P., Tversky, A.: Judgment Under Uncertainty: Heuristics and Biases. Cam-
bridge University Press, New York (1982)
[6] Kahneman, D., Tversky, A. (Eds.): Choices, Values and Frames. Cambridge University Press, New
York (2000)
[7] Ellsberg, D.: Risk, ambiguity, and the Savage axioms. Q. J. Econ. 75, 643 -- 669 (1961)
[8] Simon, H: A behavioral model of rational choice. Quart. J. Econ. 69, 99-118 (1955)
[9] Blutner, R., beim Graben, P.: Quantum cognition and bounded rationality. Synthese 193, 3239 -- 3291
(2016)
[10] Dalla Chiara, M.L., Giuntini, R., Leporini, R., Sergioli, G.: Holistic logical arguments in quantum
computation. Mathematica Slovaca 66, 313 -- 334 (2006)
[11] Aerts, D.: Quantum structure in cognition. J. Math. Psychol. 53, 314 -- 348 (2009)
[12] Aerts, D., Gabora, L., Sozzo, S.: Concepts and their dynamics: A quantum -- theoretic modeling of
human thought. Top. Cogn. Sci. 5, 737 -- 772 (2013b)
[13] Dalla Chiara, M.L., Giuntini, R., Leporini, R., Negri, E., Sergioli, G.: Quantum information, cognition,
and music. Front. Psychol. doi 10.3389/fpsyg.2015.01583 (2015)
[14] Aerts, D., Broekaert, J., Gabora, L., Sozzo, S.: Quantum structure and human thought. Behav. Brain
Sci. 36, 274 -- 276 (2013a)
[15] Haven, E., Khrennikov, A.Y.: Quantum Social Science. Cambridge University Press, Cambridge (2013)
[16] Aerts, D., Sozzo, S., Tapia, J.: Identifying quantum structures in the Ellsberg paradox. Int. J. Theor.
Phys. 53, 3666 -- 3682 (2014)
[17] Aerts, D., Sozzo, S.: From ambiguity aversion to a generalized expected utility. Modeling preferences
in a quantum probabilistic framework. J. Math. Psychol. 74, 117 -- 127 (2016)
[18] Sozzo, S.: Effectiveness of the quantum-mechanical formalism in cognitive modeling. Soft Comp. 21,
1455 -- 1465 (2017)
[19] Holik, F., Fortin, S., Bosyk, G., Plastino, A.: On the interpretation of probabilities in generalized
probabilistic models. Lect. Not. Comp. Sci. 10106, 194 -- 205 (2016)
[20] Holik, F., Sergioli, G., Freytes, H., Plastino, A.: Pattern recognition in non-Kolmogorovian structures.
Found. Sci. 23, 119 -- 132 (2017)
[21] Jauch, J.M.: Foundations of Quantum Mechanics. Addison Wesley, Reading MA (1968)
[22] Piron, C.: Foundations of Quantum Physics. Benjamin, Reading MA (1976)
[23] Aerts, D.: Foundations of quantum physics: A general realistic and operational approach. Int. J.
Theor. Phys. 38, 289 -- 358 (1999)
15
[24] Aerts, D.: Being and change: Foundations of a realistic operational formalism. In: Probing the
Structure of Quantum Mechanics: Nonlinearity, Nonlocality, Computation and Axiomatics. Aerts, D.,
Czachor, M., Durt, T. (Eds.) World Scientific, Singapore (2002)
[25] Aerts, D., Sassoli de Bianchi, M., Sozzo, S.: On the foundations of the Brussels operational-realistic
approach to cognition. Front. Phys. 4, doi: 10.3389/fphy.2016.00017 (2016)
[26] Aerts, D., Haven, E., Sozzo, S.: A proposal to extend expected utility in a quantum probabilistic
framework. Econ. Theory 65, 1079 -- 1109 (2018)
[27] Aerts, D., Geriente, S., Moreira, C., Sozzo, S.: Testing ambiguity and Machina preferences within a
quantum-theoretic framework for decision-making, J. Math. Econ. doi 10.1016/j.jmateco.2017.12.002
(2018)
[28] Sozzo, S.: Quantum structures in human decision-making: Towards quantum expected utility, Int. J.
Theor. Phys. doi 10.1007/s10773-019-04022-w (2019)
[29] Gigerenzer, G., Selten, R.: Bounded Rationality: An Adaptive Toolbox. MIT Press, Cambridge MA
(2001)
[30] Tversky, A., Simonson, I.: Context-dependent preferences. Manag. Sci. 39, 85 -- 117 (1993)
[31] Sen, A.: Maximization and the act of choice. Econometrica 65, 745 -- 779 (1997)
[32] Ariely, D., Prelec, G., Lowenstein, D.: "Coherent arbitrariness": Stable demand curve without stable
preferences. Quart. J. Econ. 118, 73-103 (2003)
[33] Beltrametti, E.G., Cassinelli, G.: The Logic of Quantum Mechanics. Addison -- Wesley, Reading MA
(1981)
[34] Knight, F.H.: Risk, Uncertainty and Profit. Houghton Mifflin, Boston (1921)
[35] Gilboa, I., Marinacci, M.: Ambiguity and the Bayesian paradigm. In: Advances in Economics and
Econometrics: Theory and Applications. Acemoglu, D., Arellano, M., Dekel, E. (Eds.) Cambridge
University Press, New York (2013)
[36] Machina, M.J., Siniscalchi, M.: Ambiguity and ambiguity aversion. In: Machina, M.J., Viscusi, K.
(Eds.) Handbook of the Economics of Risk and Uncertainty, pp. 729 -- 807. Elsevier, New York (2014)
[37] Machina, M.J.: Risk, ambiguity, and the dark -- dependence axioms. Am. Econ. Rev. 99, 385-392 (2009)
[38] L'Haridon, O., Placido, L.: Betting on Machinas reflection example: An experiment on ambiguity.
Theor. Decis. 69, 375 -- 393 (2010)
16
|
1712.10062 | 1 | 1712 | 2017-12-28T21:26:43 | Multi-timescale memory dynamics in a reinforcement learning network with attention-gated memory | [
"q-bio.NC",
"cs.LG",
"cs.NE",
"stat.ML"
] | Learning and memory are intertwined in our brain and their relationship is at the core of several recent neural network models. In particular, the Attention-Gated MEmory Tagging model (AuGMEnT) is a reinforcement learning network with an emphasis on biological plausibility of memory dynamics and learning. We find that the AuGMEnT network does not solve some hierarchical tasks, where higher-level stimuli have to be maintained over a long time, while lower-level stimuli need to be remembered and forgotten over a shorter timescale. To overcome this limitation, we introduce hybrid AuGMEnT, with leaky or short-timescale and non-leaky or long-timescale units in memory, that allow to exchange lower-level information while maintaining higher-level one, thus solving both hierarchical and distractor tasks. | q-bio.NC | q-bio |
Multi-timescale memory dynamics
in a reinforcement learning network
with attention-gated memory
Marco Martinolli†, Wulfram Gerstner† and Aditya Gilra†
†School of Computer and Communication Sciences,
and Brain-Mind Institute, School of Life Sciences,
École Polytechnique Fédérale de Lausanne, 1015 Lausanne EPFL, Switzerland
Correspondence: [email protected], [email protected]
Abstract
Learning and memory are intertwined in our brain and their relationship is at the core of several recent neural
network models. In particular, the Attention-Gated MEmory Tagging model (AuGMEnT) is a reinforcement
learning network with an emphasis on biological plausibility of memory dynamics and learning. We find that the
AuGMEnT network does not solve some hierarchical tasks, where higher-level stimuli have to be maintained over
a long time, while lower-level stimuli need to be remembered and forgotten over a shorter timescale. To
overcome this limitation, we introduce hybrid AuGMEnT, with leaky or short-timescale and non-leaky or
long-timescale units in memory, that allow to exchange lower-level information while maintaining higher-level
one, thus solving both hierarchical and distractor tasks.
1
Introduction
Memory spans various timescales and plays a
crucial role in human and animal learning [Tetzlaff
et al. 2012]. In cognitive neuroscience, the memory
system that enables manipulation and storage
of information over a period of a few seconds is
called Working Memory (WM), and is correlated
with activity in prefrontal cortex (PFC) and basal
ganglia (BG) [Frank et al. 2001; Mink 1996]. In com-
putational neuroscience, there are not only several
standalone models of WM dynamics [Barak and
Tsodyks 2014; Samsonovich and McNaughton
1997; Compte et al. 2000], but also supervised
and reinforcement learning models augmented
by working memory [Alexander and Brown 2015;
Rombouts et al. 2015; Graves et al. 2014; 2016;
Santoro et al. 2016].
Memory mechanisms can be implemented by
enriching a subset of artificial neurons with slow
time constants and gating mechanisms [Hochreiter
and Schmidhuber 1997; Gers and Schmidhuber
2001; Cho 2014]. More recent memory-augmented
neural network models like Neural Turing Ma-
chine [Graves et al. 2014] and Differentiable Neural
Computer [Graves et al. 2016], employ an address-
able memory matrix that works as a repository of
past experiences and a neural controller that is able
to store and retrieve information from the external
memory to improve its learning performance.
Here, we study and extend the Attention-Gated
MEmory Tagging model or AuGMEnT [Rombouts
et al. 2015]. AuGMEnT is trained with a Reinforce-
ment Learning (RL) scheme, where learning is
based on a reward signal that is released after each
response selection. The representation of stimuli is
accumulated in the memory states and the mem-
ory is reset at the end of each trial (see Methods).
The main advantage of the AuGMEnT network
for the computational neuroscience community
resides in the biological plausibility of its learning
algorithm.
Notably, the AuGMEnT network [Rombouts et al.
1
2015] uses a memory-augmented version of a
biologically plausible learning rule [Roelfsema
and van Ooyen 2005] mimicking backpropagation
(BP). Learning is the result of the joint action of
two factors, neuromodulation and attentional
feedback, both influencing synaptic plasticity. The
former is a global reward-related signal that is
released homogeneously across the network to
inform each synapse of the reward prediction
error after response selection [Schultz et al. 1993;
1997; Waelti et al. 2001]. Neuromodulators such as
dopamine influence synaptic plasticity [Yagishita
et al. 2014; He et al. 2015; Brzosko et al. 2015;
2017; Frémaux and Gerstner 2016]. The novelty of
AuGMEnT compared to three-factor rules [Xie and
Seung 2004; Legenstein et al. 2008; Vasilaki et al.
2009; Frémaux and Gerstner 2016] is to add an
attentional feedback system in order to keep track
of the synaptic connections that cooperated for
the selection of the winning action and overcome
the so-called structural credit assignment problem
[Roelfsema and van Ooyen 2005; Rombouts et al.
2015]. AuGMEnT includes a memory system, where
units accumulate activity across several stimuli in
order to solve temporal credit assignment tasks
involving delayed reward delivery [Sutton 1984;
Okano et al. 2000].
The attentional feedback
mechanism in AuGMEnT works with: a) synaptic
eligibility traces that decay slowly over time, and
b) non-decaying neuronal traces that store the
history of stimuli presented to the network up to
the current time [Rombouts et al. 2015] [Rombouts
et al. 2015]. The AuGMEnT network solves the
Saccade-AntiSaccade task [Rombouts et al. 2015],
which is equivalent to a temporal XOR task [Ab-
bott et al. 2016] (see Supplementary Material).
different decay constants so that they work on
different temporal scales, while the network learns
to weight their usage based on the requirements of
the specific task. In our simulations, we employed
just two subgroups of cells in the memory, where
one half of the memory is non-leaky and the
other is leaky with a uniform decay time constant;
however, more generally,
the hybrid AuGMEnT
architecture may contain several subgroups with
distinct leakage behaviours.
The paper is structured as follows.
Section
2 presents the architectural and mathematical
details of hybrid AuGMEnT. Section 3 describes the
simulated results of the hybrid AuGMEnT network,
the standard AuGMEnT network and a fully leaky
control network, on two cognitive tasks, a non-
hierarchical task involving sequence prediction
[Cui et al. 2015] and a hierarchical task 12AX
[O'Reilly and Frank 2006]. Finally, in Section 4
we discuss our main achievements in comparison
with state-of-the-art models and present possible
future developments of the work.
2 Methods
2.1 Hybrid AuGMEnT network architecture
and operation
The network controls an agent which, in each time
step t, receives a reward in response to the previ-
ous action, processes the next stimulus, and takes
the next action, as in Figure 1B. In each time step,
we distinguish two phases, called the feedforward
pass and feedback pass, depicted in Figure 1C.
However,
in the case of more complex tasks
with long trials and multiple stimuli, like 12AX
[O'Reilly and Frank 2006] depicted in Figure 1A,
we find that the accumulation of information in
AuGMEnT can lead to memory interference and
loss in performance. Hence, we ask the question
whether a modified AuGMEnT model would lead
to a broader applicability of attention-gated rein-
forcement learning. We propose a variant of the
AuGMEnT network, named hybrid AuGMEnT, that
introduces timescales of forgetting or leakage in
the memory dynamics to overcome this kind of
learning limitation. We employ memory units with
2.1.1 Feedforward pass: stimulus to action selec-
tion
In AuGMEnT [Rombouts et al. 2015], information is
processed through a network with three layers, as
shown in the left panel of Figure 1C. Each unit of
the output layer corresponds to an action. There
are two pathways into the output layer: the regular
R branch and the memory M branch.
The regular branch is a standard feedforward
network with one hidden layer. The current stimu-
lus sR
i (t), indexed by unit index i = 1, . . . , S is con-
2
B
A
C
Figure 1. Overview of AuGMEnT network operation. A. Example of trials in the 12AX task, where task symbols
appear sequentially on a screen organized in outer loops, which start with either digit 1 or 2, and a random number
of inner loops (e.g. B-Y, C-X and A-X). Each cue presentation is associated with a Target (R) or Non-Target (L) correct
response. When output and correct response coincide, the agent receives a positive reward (+), otherwise it gets
a negative reward (-). Figure is adapted from Figure 1 of O'Reilly and Frank [2006]. B. AuGMEnT operates in
discrete time steps each comprising the reception of reward (r), input of state or stimulus (s) and action taken (a).
It implements the State-Action-Reward-State-Action (SARSA, in figure s'a'rsa) reinforcement learning algorithm.
In time step t, reward r is obtained for the previous action a' taken in time step t − 1. The network weights are
updated once the next action a is chosen. C. The AuGMEnT network is structured in three layers with different types
of units. Each iteration of the learning process consists of a feedforward pass (left) and a feedback pass (right). In
the feedforward pass, sensory information about the current stimulus in the bottom layer, is fed to regular units
without memory (left branch) and units with memory (right branch) in the middle layer, whose activities in turn
are weighted to compute the Q-values in the top activity layer. Based on the Q-values, the current action is selected
(e.g. green z2). The reward obtained for the previous action is used to compute the reward prediction error, which
modifies the connection weights, that contributed to the selection of the previous action, in proportion to their
eligibility traces. After this, temporal eligibility traces and tags (in green) on the connections are updated to reflect
the correlations between the current pre and post activities. Then, in the feedback pass, spatial eligibility traces (in
red) are updated, attention-gated by the current action (e.g. red z2), via feedback weights.
3
nected to the hidden units (called regular units) in-
dexed by j, via a set of modifiable synaptic weights
ji yielding activity yR
vR
j :
j (t) = σ(cid:0)hR
j
(cid:1) ,
yR
(cid:88)
hR
j =
vR
jisR
i (t),
(1)
i
where σ is the sigmoidal
function σ(x) =
(1 + exp(−x))−1.
Input units are one-hot bi-
nary with values Si ∈ {0, 1} (equal to 1 if stimulus
i is currently presented, 0 otherwise).
The memory branch is driven by transitions be-
tween stimuli, instead of the stimuli themselves.
The sensory input of the memory branch consists
of a set of 2S transient units, i.e. S ON units
l ∈ {0, 1}S that encode the onset of each stimu-
s+
lus, and S OFF units s−
l ∈ {0, 1}S that encode the
offset:
l (t) = [sl(t) − sl(t − 1)]+
s+
s−
l (t) = [sl(t − 1) − sl(t)]+,
(2)
where the brackets signify rectification.
In the
following, we denote the input into the memory
branch with a variable sM
i defined as the concate-
nation of these ON and OFF units:
sM
i (t) =
s+
i (t),
s−
i−S(t),
if i ≤ S
if i > S,
(3)
(cid:40)
(cid:88)
The memory units have to maintain task-relevant
information through time. The transient input
is transmitted via the synaptic connections vM
ji to
the memory layer, where it is accumulated in the
states:
hM
j (t) = ϕjhM
j (t − 1) +
vM
ji sM
i (t).
(4)
i
We introduce the factor ϕj ∈ [0, 1] here, as an
extension to the standard AuGMEnT [Rombouts
et al. 2015], to incorporate decay or forgetting of
j over time. Setting ϕj ≡ 1
the memory state hM
for all j, we obtain non-leaky memory dynamics
as in the original AuGMEnT network [Rombouts
et al. 2015] (Fig.
In our hybrid
AuGMEnT network, each memory cell or subgroup
of memory cells may be assigned different leak
co-efficients ϕj (Fig. 2, right panel). In this way,
the memory is composed of subpopulations of
neurons that cooperate in different ways to solve
2, left panel).
4
a task, allowing at the same time long-time main-
tenance and fast decay of information in memory.
In contrast to the forget gate of Long Short-Term
Memory [Hochreiter and Schmidhuber 1997] or
Gated Recurrent Unit [Cho 2014], our memory
leak co-efficient is not trained and gated, but fixed.
The memory state hM
j
leads to the activation of a
memory unit:
j (t) = σ(cid:0)hM
j (t)(cid:1) .
yM
(5)
The states of the memory units are reset to 0 at the
end of each trial.
Both branches converge onto the output layer.
The activity of an output unit with index k approx-
imates the Q-value of action k = a given the input
s ≡ [si], denoted as Qs,a(t). Q-values are formally
defined as the future expected discounted reward
conditioned on stimulus s(t) and action a(t), that
is:
(cid:12)(cid:12)(cid:12)(cid:12) s = s(t), a = a(t)
(cid:35)
(6)
Qs,a(t) = E
γτ rt+τ +1
(cid:34) ∞(cid:88)
τ =0
where γ ∈ [0, 1] is a discount factor. Numeri-
cally, the vector Q that approximates the Q-values
is obtained by combining linearly the hidden states
from the regular and the memory branches, with
synaptic weights wR
kj and wM
kj :
Qk(t) =
wR
kjyR
j (t) +
wM
kj yM
j (t).
(7)
j
j
Finally, the Q-values of the different actions partic-
ipate in an -greedy winner-take-all competition to
select the response of the network. With probabil-
ity 1 − , the next action a(t) is the one with the
maximal Q-value:
a(t) = argmaxkQk(t).
(8)
With probability , a stochastic policy is chosen
with probability:
(cid:88)
(cid:88)
(cid:80)
pa =
exp(g(t)Qa)
k exp(g(t)Qk)
(9)
π arctan( t
is a weight
function defined as
where g(t)
g(t) = 1 + 10
t∗ ), that gradually increases in
time over a task-specific, fixed time scale t∗. Over
time, this emphasizes the action with maximal
Q-value, improving prediction stability.
AuGMEnT
Hybrid AuGMEnT
Figure 2. Architectures of standard AuGMEnT and hybrid AuGMEnT networks. The difference between the net-
works consist in their memory dynamics: the memory layer of standard AuGMEnT (left) has only conservative units
with ϕj ≡ 1, while hybrid AuGMEnT (right) possesses a memory composed of both leaky ϕj < 1 and non-leaky
ϕj = 1 units.
2.1.2 Feedforward pass: reward-based update of
weights, and correlation-based update of
eligibility traces and tags
AuGMEnT follows the SARSA updating scheme and
updates the Q-values for the previous action a(cid:48)
taken at time t − 1, once the action a at time t
is known (see Fig. 1B). Q-values depend on the
weights via equation (7). The temporal difference
(TD) error is defined as [Wiering and Schmidhuber
1998; Sutton and Barto 1998]:
δ(t) = (r(t) + γQa(t)) − Qa(cid:48)(t − 1),
(10)
where a is the action chosen at current time t, and
r(t) is the reward obtained for the action a(cid:48) taken at
time t − 1. The Temporal Difference (TD) error δ(t)
acts as a global reinforcement signal to modify the
weights of all connections as
vR,M
ji
wR,M
kj
(t + 1) = vR,M
(t + 1) = wR,M
ji
kj
(t) + βeR,M
(t) + βeR,M
ji
kj
(t)δ(t),
(t)δ(t),
(11)
and eR,M
where β is a learning rate and eR,M
are synaptic eligibility traces, defined below.
Superscript R or M denotes the regular or memory
branch respectively. We use the same symbol eR,M
for eligibility traces at the input-to-hidden (i to
kj
ji
5
j) and hidden-to-output (j to k) synapses, even
though these are different, with the appropriate
one clear from context and the convention for
indices.
After the update of weights, a synapse from neu-
ron j in the hidden layer to neuron k in the output
layer updates its temporal eligibility trace
eR
kj(t + 1) = yR
kj (t + 1) = yM
eM
j (t)zk(t) + (1 − α)eR
j (t)zk(t) + (1 − α)eM
kj(t),
kj (t),
(12)
where α ∈ [0, 1] is a decay parameter, zk is a
binary one-hot variable that indicates the winning
action (equal to 1 if action k has been selected, 0
otherwise), and M or R denotes the regular or the
memory branch respectively.
Similarly, a synapse from neuron i in the input
layer to neuron j in the hidden layer sets momen-
tary tags T R,M
as:
ji
j (t)),
i (t) σ(cid:48)(hR
ji (t) σ(cid:48)(hM
ji (t) = sR
T R
ji (t) = X M
T M
where σ(cid:48)(hR,M
) is a nonlinear function of the in-
put potential, defined as the derivative of the gain
j (t)),
(13)
j
function σ, and X M
is a synaptic trace [Pfister and
ji
Gerstner 2006; Morrison et al. 2008] defined as fol-
lows:
i
j
ji
X M
(14)
i (t).
, zk, sR,M
ji (t) = ϕj X M
ji (t − 1) + sM
Note that the tag T R,M
has no memory beyond one
time step, i.e. it is set anew at each time step. Nev-
ertheless, since X M
ji depends on previous times, the
tag T M
ji of memory units can link across time steps.
Since activities yR,M
and input potentials
hR,M
are quantities available at the synapse, a bi-
j
ological synapse can implement the updates of eli-
gibility traces and tags locally. We emphasize that
both eligibility traces and tags can be interpreted
as 'Hebbian' correlation detectors. In the original
AuGMEnT model [Rombouts et al. 2015], all eligibil-
ity traces and tags were said to be updated in the
feedback pass. Here, without changing the order
of operations of the algorithm, we have conceptu-
ally shifted the update of those traces and tags that
depend on the correlations of the activities, to the
last step of the feedforward pass when these activ-
ities are still available. Note that activities could
in principle change via attention-gating during the
feedback pass [Roelfsema et al. 2010; Moore and
Armstrong 2003].
jk
It must be noted that the feedback synapses
w(cid:48)R,M
follow the same update rule as their feed-
forward partner wR,M
Therefore, even if the
initializations of the feedforward and feedback
weights are different,
their strengths become
similar during learning, as suggested by neuro-
physiological findings [Mao et al. 2011].
kj
.
2.2 Deriving the learning rules via gradient
descent
For networks with one hidden layer and one-
hot coding in the output, attentional feedback is
equivalent to backpropagation [Roelfsema and van
Ooyen 2005; Rombouts et al. 2015]. We can show
that the equations for eligibility traces and tagging
along with the weight update equations reduce an
RPE-based loss function E defined as:
E =
1
2
(δ(t))2
(16)
Here we specifically discuss the case of the tagging
equations (13) and (15) and the update rule (11)
associated with the weight vM
ji from sensory input
into memory, as it contains the memory decay
factor ϕj that we introduced, but analogous dis-
cussion holds also for weights vR
kj and wR
kj.
ji, wM
2.1.3 Feedback pass: attention-gated update of
eligibility traces
Proof. We want to show that
After action selection and the updates of weights,
tags, and temporal eligibility traces in the feedfor-
ward pass, the synapses that contributed to the cur-
rently selected action update their spatial eligibil-
ity traces in an attentional feedback step. For the
synapses from the input to the hidden layer, the tag
T R,M
from equation (13) is combined with a spa-
ji
tial eligibility trace which can be interpreted as an
attentional feedback signal [Rombouts et al. 2015].
eR
ji(t + 1) = T R
ji
eM
ji (t + 1) = T M
ji
w(cid:48)R
jkzk + (1 − α)eR
jk zk + (1 − α)eM
w(cid:48)M
ji(t),
ji (t),
(cid:88)
(cid:88)
k
k
∆vM
ji = β eM
ji δt ∝ − ∂E
∂vM
ji
(17)
For simplicity, here we prove (17) neglecting the
temporal decay of the eligibility trace eM
ji (i.e. α =
1), so that
eM
ji = T M
ji
(cid:48)M
jk zk = T M
ji w
(cid:48)M
ja(cid:48)
w
where a(cid:48) is the selected action at time t − 1.
k
(cid:88)
(15)
We first observe that the right-hand side of equa-
tion (17) can be rewritten as:
where feedback weights from the output layer to
the hidden layer have been denoted as w(cid:48)
jk and
zk ∈ {0, 1} is the value of output unit k (one-hot
response vector as defined for equation (12)).
− ∂E
∂vM
ji
= − ∂E
∂Qa(cid:48)
∂Qa(cid:48)
∂vM
ji
= δt
∂Qa(cid:48)
∂vM
ji
Thus, it remains to show that
∂Qa(cid:48)
∂vM
ji
= eM
ji .
6
Similarly to the approach used in backpropaga-
tion, we now apply the chain rule and we focus on
each term separately:
∂Qa(cid:48)
∂vM
ji
=
∂Qa(cid:48)
∂yM
j
∂yM
j
∂hM
j
∂hM
j
∂vM
ji
Parameter
β : Learning parameter
λ : Eligibility persistence
γ : Discount factor
α : Eligibility decay rate
: Exploration rate
Value
0.15
0.15
0.9
1 − γλ
0.025
From equations (5) and (7), we immediately have
that:
∂yM
j
∂hM
j
= σ(cid:48)(hM
j )
∂Qa(cid:48)
∂yM
j
= wM
a(cid:48)j
However, in the feedback step the weight wM
a(cid:48)j is
replaced by its feedback counterpart w(cid:48)M
ja(cid:48) . As dis-
cussed above, this is a valid approximation because
they become similar during learning.
Finally, starting from equation (4) we can write:
Table 1. Parameters for the AuGMEnT network.
memory cells and ϕj = 0.7 for the second half. To
reduce to the standard AuGMEnT [Rombouts et al.
2015] network, we set ϕj ≡ 1 for all j, while for
a leaky control network we set ϕj ≡ 0.7 for all j.
In general, the leak co-efficients may be tuned to
adapt the overall memory dynamics to the specific
task.
(cid:88)
≈(cid:88)
i
t−1(cid:88)
(cid:88)
t(cid:88)
τ =t0
i
ϕt−τ
j
sM
i (τ )
hM
j (t) =
vM
ji (t) sM
i (t) +
vM
ji (t)
i
τ =t0
ϕt−τ
j
vM
ji (τ ) sM
i (τ )
3 Results
where t0 indicates the starting time of the trial and
last approximation derives from the assumption of
slow learning dynamics, i.e. vM
ij (t) for
t0 ≤ τ < t. As a consequence, we have:
ij (τ ) = vM
∂hM
∂vM
j (t − 1)
ji (t − 1)
≈ t−1(cid:88)
τ =t0
ϕt−τ +1
j
sM
i (τ ) = X M
ji (t − 1)
In conclusion, we combine the different terms
and we obtain the desired result:
(cid:48)M
ja(cid:48) = δt eM
ji .
ji ∝ δt X M
ji σ(cid:48)(hM
j ) w
∆vM
2.3 Simulation and tasks
simulation
scripts were written
(https://www.python.org/),
in
All
with
python
rendered using the matplotlib module
plots
(http://matplotlib.org/).
simulation
and plotting scripts are available online at
https://github.com/martin592/hybrid_AuGMEnT.
These
We use the parameters listed in Table 1 for our
simulations.
Further, for the Hybrid AuGMEnT
network, we set ϕj = 1 for the first half of the
7
AuGMEnT [Rombouts et al. 2015] includes a differ-
entiable memory system and is trained in an RL
framework with learning rules based on the joint
effect of synaptic tagging, attentional feedback and
neuromodulation (see Methods). Here, we study
our proposed variant of AuGMEnT, named hybrid
AuGMEnT, that has an additional leak factor in a
subset of memory units, and compare it to the
original AuGMEnT and to a control network with
all leaky memory units.
As a first step, we validated our implementa-
tions of standard and hybrid AuGMEnT networks
on the Saccade-AntiSaccade (S-AS)
task, used
in the reference paper [Rombouts et al. 2015]
(Supplementary Material). We next simulated the
networks on two other cognitive tasks with differ-
ent structure and memory demands: the sequence
prediction task [Cui et al. 2015] and the 12AX task
[O'Reilly and Frank 2006]. In the former, the agent
has to predict the final letter of a sequence depend-
ing only on its starting letter, while in the latter, the
agent has to identify target pairs inside a sequence
of hierarchical symbols. The S-AS task maps to a
temporal XOR task [Abbott et al. 2016], thus the
hidden layer is essential for the task [Minsky and
Papert 1969; Rumelhart et al. 1985]. The 12AX also
resembles an XOR structure, but is more complex
Table 2. Network Architecture Parameters for the Simulations
Network Parameter
Sequence Prediction Task 12AX Task
S : Number of sensory units
R : Number of regular units
M : Number of memory units
A : Number of activity units
(L = sequence length)
L − 1
3
8
2
8
10
20
2
due to an additional dimension and distractors
in the inner loop (Supplementary Figure 3). The
complexity of the sequence prediction task is less
compared to the 12AX task, and can be effectively
solved by AuGMEnT. We will show that hybrid
AuGMEnT performs well on both cognitive tasks,
whereas standard AuGMEnT fails on the 12AX task.
The parameters involving the architecture of the
networks on each task are reported in Table 2. We
now discuss each of the tasks in more detail.
3.1 Task 1: Sequence Prediction
In the sequence prediction task [Cui et al. 2015],
letters appear sequentially on a screen and at the
end of each trial the agent has to correctly predict
the last letter. Each sequence starts either with an A
or with an X, which is followed by a fixed sequence
of letters (e.g. B-C-D-E). The trial ends with the
prediction of the final letter, which depends on the
initial cue: if the sequence started with A, then the
final letter has to be a Z; if the initial cue was an X,
then the final letter has to be a Y. In case of correct
prediction the agent receives a reward of 1 unit,
otherwise he is punished with a negative reward
of −1. A scheme of the task is presented in Figure
3 for sequences of four letters.
The network has to learn the task for a given
sequence length, kept fixed throughout training.
The agent must learn to maintain in memory, the
initial cue of the sequence until the end of the trial,
to solve the task. At the same time, the agent has
to learn to neglect the information coming from
the intermediate cues (called distractors). Thus the
difficulty of the task is correlated with the length
of the sequence.
We studied the performance of the AuGMEnT
network [Rombouts et al. 2015] and our hybrid
Figure 3. Scheme of the sequence prediction task.
Scheme of sequence prediction trials with sequence
length equal to 4 (i.e. 2 distractors): the two possible
sequences are: A-B-C-Z (blue) or X-B-C-Y (red)
variant on the sequence prediction task. The mean
trend of the RPE-based energy function defined
in equation (16) (Fig. 4A) shows that both models
converge in a few hundreds of iterations. As a
control, we also simulated a variant in which
also memory units were leaky. We noticed that
hybrid and standard AuGMEnT networks are more
efficient than the purely leaky control. This is not
surprising because the key point in the sequence
prediction task consists in the maintenance of the
initial stimulus, which is simpler with a non-leaky
memory than with a leaky one. We notice that the
hybrid model has a behaviour similar to AuGMEnT.
We also analyzed the effect of the temporal
length of the sequences on the network perfor-
mance, by varying the number of distractors (i.e.
the intermediate letters) per sequence (Fig. 4B). For
each sequence length, the network was retrained
ab initio. We required 100 consecutive correct
predictions as the criterion for convergence. We
ran 100 simulations starting with different initial-
izations for each sequence length and averaged
8
A
B
Figure 4. Convergence in the sequence prediction task. A. Time course of error of the models on the sequence
prediction task with sequences of five letters (three distractors): the mean squared RPE decays to zero for all
networks but the leaky control network (blue) is much slower than AuGMEnT (green) and Hybrid AuGMEnT (red).
B. Convergence time of the AuGMEnT network and its variants on the sequence prediction task with increasing
number of distractors, i.e. intermediate cues before final prediction.
Figure 5. Memory weights of AuGMEnT networks in sequence prediction task. Memory weight matrices after
convergence for AuGMEnT (left) and Hybrid AuGMEnT (right) networks on the sequence prediction task with se-
quences of length five (first row) and ten (second row). Note that the first two memory units in Hybrid AuGMEnT
are leaky (M1-L and M2-L), while the last ones are conservative or non-leaky (M1-C and M2-C).
9
Table 3. The 12AX task: table of key information
Task feature
Input
Output
Target sequences 1-. . . -A-X or 2-. . . -B-Y.
Details
8 possible stimuli: 1,2,A,B,C,X,Y,Z.
Non-Target (L) or Target (R).
Training dataset
Inner loops
Probability of target sequence is 25%.
Sequence of outer loops starting with 1 or 2.
Maximum number of training samples is 1000000.
Each outer loop contains a random number of inner loops,
between 1 and 4
the convergence time. Again, AuGMEnT and Hy-
brid AuGMEnT show good learning performance,
maintaining an average of about 250 trials to
convergence for sequences containing up to 10
distractors, whereas a network with purely leaky
units is much slower to converge.
The leaky dynamics are not helpful for the
sequence prediction task, because the intermediate
cues are not relevant for the final model perfor-
mance. Therefore, it is sufficient to supress the
weight values in the VM matrix for distractors,
and increase those of the initial A/X letter. This is
confirmed by the structure of the weight matrices
of the memory in all networks shown after con-
vergence (Fig. 5) in simulations of the sequence
prediction task on sequences of five and ten letters.
The weight values are highest in absolute value
for letters A and X, for both the ON (+) and OFF
(−) units.
Finally, we also notice that Hybrid
AuGMEnT employs mainly the conservative or
non-leaky memory units (M1-C and M2-C) rather
than the leaky ones (M1-L and M2-L) to solve the
task, showing that the network is able to focus the
update dynamics on the connections that are more
suitable for the specific task.
3.2 Task 2: 12AX
The 12AX task is a standard cognitive task used
to test Working Memory and diagnose behavioral
and cognitive deficits related to memory dys-
functions [Alexander and Brown 2015]. Basically,
the problem consists in identifying some target
sequences among a group of symbols that appear
on a screen.
The general procedure of the task is schematized
in Figure 1A and details involving the construction
of the 12AX dataset are collected in Table 3. The set
of possible stimuli consists of 8 symbols: two digit
cues (1 and 2), two context cues (A and B), two
target cues (X and Y) and finally two distractors
(C and Z). Each trial (or outer loop) starts with a
digit cue and is followed by a random number
Inner loops are composed
of 1 to 4 inner loops.
of patterns of context-target cues,
like A-X, B-X
or B-Y. The distractors are non task-relevant cues
that can invalidate a subsequence creating wrong
inner loops like A-Z or C-X. The cues are presented
one by one on a screen and the agent has two
possible responses for each of them: Target (R) and
Non-Target (L). There are only two valid Target
cases:
in trials that start with digit 1, the Target
is associated with the target cue X if preceded by
context A (1-. . . -A-X); otherwise, in case of initial
digit 2, the Target occurs if the target cue Y comes
after context B (2-. . . -B-Y). The dots are inserted
to stress that the target inner loop can occur even
a long time after the digit cue, as happens in the
following example sequence: 1-A-Z-B-Y-C-X-A-X
(whose sequence of correct responses is L-L-L-L-
L-L-L-L-R). The variability in the temporal length
of each trial is the main issue in solving the 12AX
task because of the temporal credit assignment
problem. Moreover, since 1-A-X and 2-B-Y are
target sequences, whereas 2-A-X and 1-A-Y are not,
the task can be seen as a generalization of temporal
XOR (Supplementary Fig. 3).
The types of the inner loops are determined
randomly, with a probability of 50% to have
pairs A-X or B-Y. As a result, combined with the
10
probability to have either 1 or 2 as starting digit
of the trial, the overall probability to have target
pair is 25%. Since the Target response R has to
be associated only with an X or Y stimulus that
appears in the correct sequence, the number of
Non-Targets L is generally much larger, on average
8.96 Non-Targets to 1 Target. We rewarded the
correct predictions of a Non-Target with 0.1 and
of Targets with 1, and punished the wrong predic-
tions with reward of −1. In effect, we balanced the
positive reward approximately equally between
Targets and Non-Targets based on their relative
frequencies, which aids convergence.
We simulated Hybrid AuGMEnT network, base
AuGMEnT and leaky control on the 12AX task, in
order to see whether in this case the introduction
of the leaky dynamics improves learning perfor-
mance. Figure 6A shows the evolution of the mean
squared RPE for the three networks. After a sharp
descent, all networks converge to an error level
that is non-zero, indicating that learning of the
12AX task is not completely achieved, possibly
due to memory interference. However, hybrid
AuGMEnT and leaky control saturate at a lower
error value than base AuGMEnT. This difference can
be attributed mostly to the errors in responding to
the Target cues (Fig. 6B) than to Non-Target cues
(Fig. 6C). Note that, since 12AX is a Continuous
Performance Task (CPT), the error is computed
at each iteration – including the more frequent
and trivial Non-Target predictions – and averaged
over 2, 000 consecutive predictions. All networks
quickly learn to recognize the Non-Target cues (1,
2, A, B, C, Z are always Non-Targets) (Fig. 6C).
However, hybrid AuGMEnT and leaky control learn
the more complex identification of Target patterns
within a trial when X or Y are presented to the
network, better than base AuGMEnT (Fig. 6B). The
gap in the mean squared RPE between hybrid
AuGMEnT and leaky control versus base AuGMEnT
is wider when only potential target cues are con-
sidered in the mean-squared RPE as in Figure 6B,
than when only non-targets are considered as in
Figure 6C.
convergence
Adopting the
criterion from
Alexander and Brown [2015] that requires 1, 000
consecutive correct predictions, we show the per-
centage of successful learning over 100 simulations
and the average learning time in Figure 7 for the
three networks. Standard AuGMEnT network was
unable to match the convergence condition during
any simulation (0% success), despite presenting
1,000,000 outer loop trials in each simulation.
However, hybrid AuGMEnT and leaky control
performed 100% consistently,
suggesting that
leaky memory units are necessary for the 12AX
task. Leaky control learned slightly faster (learn-
ing time mean=30, 032.2 and s.d.=11, 408.9) than
hybrid AuGMEnT (learning time mean=34, 263.6
and s.d.=12, 737.3).
8, left panel):
In order to understand how the hybrid memory
works on the 12AX task, we analyzed the weight
structure of the connectivity matrices which be-
long to the memory branch of the hybrid AuGMEnT
network (Figure 8). Unlike in the sequence pre-
diction task, here the hybrid network employs
both the leaky and the non-leaky memory units.
However,
there is an overall separation in the
memory activity between the two groups of cells
(Fig.
the leaky units are mainly
responsible for the storage of the digit information,
having the highest values in absolute value on the
weights associated with 1(±) and 2(±) (e.g. on M4
and M9), while the non-leaky cells emphasize more
the information coming from the potential Target
cues X(±) and Y(±) (e.g. on M14, M17 and M20).
The storage of the initial digit cue is 'assigned'
to leaky units, because the interference from the
following letters is reduced thanks to the gradual
loss of information while the digit information
can survive through sufficient increasing of the
digit-related input weights. In this way, the mem-
ory interference problem is mitigated, because
the crucial digit information is maintained over
time without interference in the leaky units and
the identification of the inner loops is done by
the conservative part of the memory. As a result,
all memory units contribute to the definition of
the activity Q-values (Fig. 8, right panel) and, in
particular, the memory units that are more active
(i.e. the same ones mentioned above) are the ones
that strongly discriminate Non-Targets (L) against
Targets (R), giving positive contribution to one and
negative to the other.
The memory units show an opposing behavior
on activation versus on deactivation of Target cues:
11
A
B
C
Figure 6. Learning convergence of the AuGMEnT variants in the 12AX task. Minimization of the RPE-based
energy function during training on the 12AX task. A. All networks show a good decay of the mean-squared RPE,
but they seem to converge to a non-zero regime and, in particular, the base AuGMEnT network (green) is the one
that maintains a higher mean-squared RPE level when compared to leaky control (blue) and Hybrid AuGMEnT
(red). B. Mean-square RPE associated with only potential target cues X and Y. C. Mean-squared RPE related to
only non-target cues.
Figure 7. Comparative statistics of the AuGMEnT variants on performance on the 12AX task. Barplot description
of the learning behavior of the three networks on the 12AX task according to the convergence criterion given by
Alexander and Brown [2015]. After 100 simulations, we measured the fraction of times that the model satisfies con-
vergence condition (left) and the average number of training trials needed to meet the same convergence criterion
(right). Although training dataset consists of 1, 000, 000 outer loops, the base AuGMEnT network never manages
to satisfy the convergence criterion, while the leaky (blue) and hybrid (red) models have similar convergence
performance with a learning time of about 30, 000 trials.
for instance, if X+ has strong positive weight in-
tensity, then X- shows a contrary negative weight
intensity (see M14 or M17). In this way, the network
tries to reduce the problems of memory interfer-
ence between subsequent inner cycles by adding
to the memory during deactivation, an opposite
amount of information stored during the previous
activation, effectively erasing the memory. Further,
the difference in absolute value between activation
and deactivation is higher in case of the leaky cells,
because the deactivation at the next iteration has
to remove only a lower amount of information
from the memory due to leakage. However, for
the digit cues 1 and 2, the weights for activation
and deactivation have typically the same sign in
order to reinforce the digit signal in memory in
two subsequent timesteps (see on M4 and M9).
12
Figure 8. Memory weights of Hybrid AuGMEnT in the 12AX task. Plot of the weight matrices in the memory
branch of hybrid AuGMEnT network after convergence on the 12AX Task. Left: weights from the transient stimulus
into the 20 memory units (half leaky, half conservative). Right: weights from the memory cells into the output
units.
In conclusion, the conservative dynamics of the
memory in standard AuGMEnT can be a limitation
for the learning ability of the model, especially in
cases of complex tasks with many data to store or
long trials. In fact, even though the complexity of
the 12AX task is limited compared to other typical
RL tasks, the AuGMEnT network fails to maintain
a sufficiently stable performance to satisfy the
required convergence criterion. The introduction
of the leaky co-efficient in Hybrid AuGMEnT leads
to the network solving the 12AX task, overcom-
ing memory interference. However, the loss of
information from the leaky memory does not
improve learning in other tasks with lower risk of
memory interference like the sequence prediction
task. Hybrid AuGMEnTcan be adapted to different
task structures and to different temporal scales
by varying the size and the composition of the
memory,
for example by considering multiple
subpopulations of neurons with distinct memory
timescales, say in a power law distribution.
4 Discussion
A key goal of the computational neuroscience
community is to develop neural networks that are
at the same time biologically plausible and able
to learn complex tasks similar to humans. The
embedding of memory is certainly an important
step in this direction, because memory plays
a central role in human learning and decision
making. Our interest in the AuGMEnT network
[Rombouts et al. 2015] derives mainly from the
biological plausibility of its learning and memory
dynamics.
In particular, the biological setting of
the learning algorithm is based on synaptic tag-
ging, attentional feedback and neuromodulation,
providing a possible biological interpretation to
backpropagation-like methods.
We developed Hybrid AuGMEnT, by introducing
leaky dynamics in the memory system, with the
aim of improving its learning performance and
extending the variety of solvable tasks. Hybrid
AuGMEnT with both leaky and non-leaky units
in its memory system, solves the 12AX task
13
on which standard AuGMEnT fails. Both solve
the simpler saccade-antisaccade and sequence
prediction tasks. Hybrid AuGMEnT inherits the
biological plausibility of base AuGMEnT [Rombouts
et al. 2015]. In addition, consistent learning with
decaying memory units requires that the decay of
synaptic traces in a memory unit, as per equation
(14) [Pfister and Gerstner 2006; Morrison et al.
2008], be at the same timescale as decay of the
unit's memory state as per equation (4).
Despite the improvement with our hybrid
the learning ability of AuGMEnT is
variant,
still
limited compared to other state-of-the-art
memory-augmented networks. For instance, the
Hierarchical Temporal Memory (HTM) network
[Cui et al. 2015] presents a greater flexibility in
sequence learning than what we have experienced
in AuGMEnT on the simple sequence prediction
task. Utilizing a complex column-based architec-
ture and an efficient system of inner inhibitions,
the HTM network is able to maintain a dual neural
activity, both at column level and at unit level, that
allows to have sparse representations of the input
and give multi-order predictions using an unsu-
pervised Hebbian-like learning rule. Nonetheless,
it is unclear how the HTM network can be applied
to reward-based learning,
in particular to tasks
like the 12AX, with variable number of inner loops.
Although the hybrid memory in the AuGMEnT
network remarkably improved its convergence
performance on the 12AX task, its learning effi-
ciency is still lower than the reference Hierarchical
Error Representation model (HER) [Alexander and
Brown 2015; 2016]. In fact, in our simulations, hy-
brid AuGMEnT showed a mean time to convergence
equal to 34, 263.6 outer loops, while the average
learning time of HER on the same convergence
condition is around 750 outer loops. The reason for
this large gap in the learning performance resides
in the gating mechanism of HER network that is
specifically developed for hierarchical tasks and is
used to decide at each iteration whether to store
the new input or maintain the previous content
in memory. Unlike HER model, the memory in
AuGMEnT does not include any gating mechanism,
meaning that the network does not learn when to
store and recall information but the memory dy-
namics are entirely developed via standard weight
modulation. On the other hand, the HER model is
not as biologically plausible as the AuGMEnT net-
work, because, although its hierarchical structure
is inspired on the supposed organization of the
prefrontal cortex, its learning scheme is artificial
and based on standard backpropagation.
In addition,
the recent delta-RNN network
[Ororbia et al. 2017] presents interesting simi-
larities with hybrid AuGMEnT in employing two
timescales, maintaining memory via interpolation
of fast and slow changing inner representations.
The delta-RNN, whose memory dynamics are a
generalization of the gating mechanisms of LSTM
and GRU, outperforms these popular recurrent
architectures. Thus, it likely has a better learning
ability than hybrid AuGMEnT, though it requires a
higher number of parameters and the network is
not based on biological considerations.
The lack of a memory gating system is a great
limitation for AuGMEnT variants, when compared
with networks equipped with a gated memory,
like HER [Alexander and Brown 2015; 2016] or
LSTM [Hochreiter and Schmidhuber 1997; Gers
and Schmidhuber 2001], especially on complex
tasks with high memory demand.
Still, even
though it cannot be properly defined as a gating
system,
the forgetting dynamics introduced in
hybrid AuGMEnT has a similar effect as the activity
of the forget gates in LSTM or GRU. However,
unlike forget gates, the decay coefficients are not
learnable and are not input-dependent for each
memory cell.
The Hybrid AuGMEnT network
could be further developed by adding a gating
control on the leakage:
leak gates could be an
output of the controller branch of the network
and then applied as a gate or decay co-efficient
in the memory branch.
In this way, the gating
value becomes stimulus-dependent and leakage is
adjusted to optimize the model performance. On
the other hand, such a gating system would make
the network more complex, where learning of the
gate variables implies an error backpropagation
through multiple layers, that may compromise the
biological plausibility of the AuGMEnT learning
dynamics (though see [Lillicrap et al. 2016; Baldi
et al. 2016; Guerguiev et al. 2017]).
Alternatively,
inspired by the hierarchical ar-
14
chitecture of HER [Alexander and Brown 2015],
the memory in AuGMEnT could be divided into
multiple levels each with their own memory
dynamics: each memory level could be associated
with distinct synaptic decay and leaky coefficients,
learning rates, or gates, in order to cover different
temporal scales and encourage level specialization.
Compared with hybrid AuGMEnT,
this differ-
entiation in memory will not only involve the
leaky dynamics, but also the temporal dynamics
associated with attentional feedback and synaptic
potentiation.
In the past years, the reinforcement learning
community has proposed several deep RL net-
works, like deep Q-networks [Mnih et al. 2015]
or the AlphaGo model [Chen 2016], that combine
the learning advantages of deep neural networks
with reinforcement learning [Li 2017]. Thus, it
may be interesting to consider a deep version of
the AuGMEnT network with additional hidden
layers of neurons. While conventional error
backpropagation in AuGMEnT may not yield local
synaptic plasticity, locality might be retained with
alternative backpropagation methods [Lillicrap
et al. 2016; Baldi et al. 2016; Guerguiev et al. 2017].
5 Acknowledgements
We thank Vineet
Jain for helpful discussions.
Financial support was provided by the European
Research Council (Multirules, grant agreement no.
268689), the Swiss National Science Foundation
(Sinergia, grant agreement no. CRSII2_147636),
and the European Commission Horizon 2020
Framework Program (H2020)
(Human Brain
Project, grant agreement no. 720270).
References
Abbott, L., De Pasquale, B., and Memmesheimer,
R.-M. (2016). Building functional networks of
spiking model neurons. Nature neuroscience,
19(3):350–355.
Alexander, W. H. and Brown, J. W. (2015). Hier-
archical error representation: A computational
model of anterior cingulate and dorsolateral pre-
frontal cortex. Neural Computation, 27(11):2354–
2410.
Alexander, W. H. and Brown, J. W. (2016). Frontal
cortex function derives from hierarchical predic-
tive coding. bioRxiv, page 076505.
Baldi, P., Sadowski, P., and Lu, Z. (2016). Learn-
ing in the machine: Random backpropaga-
arXiv preprint
tion and the learning channel.
arXiv:1612.02734.
Barak, O. and Tsodyks, M. (2014). Working models
of working memory. Current opinion in neurobiol-
ogy, 25:20–24.
Brzosko, Z., Schultz, W.,
and Paulsen, O.
(2015). Retroactive modulation of spike timing-
eLife, page
dependent plasticity by dopamine.
4:e09685.
Brzosko, Z., Zannone, S., Schultz, W., Clopath, C.,
and Paulsen, O. (2017). Sequential neuromodu-
lation of hebbian plasticity offers mechanism for
eLife, page
effective reward-based navigation.
6:e27756.
Chen, J. X. (2016). The evolution of computing:
Alphago. Computing in Science & Engineering,
18(4):4–7.
Cho, K. e. a. (2014).
Learning phrase repre-
sentations using rnn encoder-decoder for sta-
arXiv preprint
tistical machine translation.
arXiv:1406.1078.
Compte, A., Brunel, N., Goldman-Rakic, P. S., and
Wang, X.-J. (2000). Synaptic mechanisms and
network dynamics underlying spatial working
memory in a cortical network model. Cerebral
Cortex, 10:910–923.
Cui, Y., Surpur, C., Ahmad, S., and Hawkins, J.
(2015). Continuous online sequence learning
with an unsupervised neural network model.
CoRR, abs/1512.05463.
Frank, M. J., Loughry, B., and O'Reilly, R. C. (2001).
Interactions between frontal cortex and basal
ganglia in working memory: a computational
model. Cognitive, Affective, & Behavioral Neuro-
science, 1(2):137–160.
15
Frémaux, N. and Gerstner, W. (2016). Neuromodu-
lated spike-timing-dependent plasticity, and the-
ory of three-factor learning rules. Frontiers in
Neural Circuits, page 9:85.
Gers, F. A. and Schmidhuber, J. (2001). Long short-
term memory learns context free and context
sensitive languages. In Artificial Neural Nets and
Genetic Algorithms, pages 134–137. Springer.
Gottlieb, J. and Goldberg, M. E. (1999). Activity
of neurons in the lateral intraparietal area of the
monkey during an antisaccade task. Nature neu-
roscience, 2(10):906–912.
Graves, A., Wayne, G.,
I.
(2014). Neural turing machines. arXiv preprint
arXiv:1410.5401.
and Danihelka,
Graves, A., Wayne, G., Reynolds, M., Harley,
T., Danihelka, I., Grabska-Barwi ´nska, A., Col-
menarejo, S. G., Grefenstette, E., Ramalho, T.,
Agapiou, J., et al. (2016). Hybrid computing
using a neural network with dynamic external
memory. Nature, 538(7626):471–476.
Guerguiev, J., Lillicrap, T. P., and Richards, B. A.
(2017). Towards deep learning with segregated
dendrites. eLife, page 6:e22901.
He, K., Huertas, M., Hong, S. Z., Tie, X., Hell,
J. W., Shouval, H., and Kirkwood, A. (2015). Dis-
tinct eligibility traces for ltp and ltd in cortical
synapses. Neuron, 88(3):528–538.
Hochreiter, S. and Schmidhuber, J. (1997). Long
computation,
Neural
short-term memory.
9(8):1735–1780.
Legenstein, R., Pecevski, D., and Wolfgang, M.
(2008). A learning theory for reward-modulated
spike-timing-dependent plasticity with appli-
PLOS Comput Biol.,
cation to biofeedback.
4(10):e1000180.
Li, Y. (2017). Deep reinforcement learning: An
overview. arXiv preprint arXiv:1701.07274.
Mao, T., Kusefoglu, D., Hooks, B. M., Huber, D.,
Petreanu, L., and Svoboda, K. (2011). Long-
range neuronal circuits underlying the interac-
tion between sensory and motor cortex. Neuron,
72(1):111–123.
Mink, J. W. (1996). The basal ganglia: focused se-
lection and inhibition of competing motor pro-
grams. Progress in neurobiology, 50(4):381–425.
Minsky, M. and Papert, S. (1969). Perceptrons. MIT
press.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A.,
Veness, J., Bellemare, M. G., Graves, A., Ried-
miller, M., Fidjeland, A. K., Ostrovski, G., et al.
(2015). Human-level control through deep rein-
forcement learning. Nature, 518(7540):529–533.
Moore, T. and Armstrong, K. M. (2003). Selective
gating of visual signals by microstimulation of
frontal cortex. Nature, 421(6921):370–373.
Morrison, A., Diesmann, M., and Gerstner, W.
(2008). Phenomenological models of synaptic
plasticity based on spike timing. Biological Cy-
bernetics, 98(6):459–478.
Okano, H., Hirano, T., and Balaban, E. (2000).
Learning and memory. Proceedings of the National
Academy of Sciences, 97(23):12403–12404.
O'Reilly, R. C. and Frank, M. J. (2006). Making
working memory work: a computational model
of learning in the prefrontal cortex and basal gan-
glia. Neural computation, 18(2):283–328.
Ororbia, A. G., Mikolov, T., and Reitter, D. (2017).
Learning simpler language models with the dif-
ferential state framework. Neural Computation,
29(12):3327–3352.
Pfister, J.-P. and Gerstner, W. (2006). Triplets of
spikes in a model of spike timing-dependent
Journal of Neuroscience, 26(38):9673–
plasticity.
9682.
Lillicrap, T. P., Cownden, D., Tweed, D. B., and Ak-
erman, C. J. (2016). Random synaptic feedback
weights support error backpropagation for deep
learning. Nature communications, 7.
Roelfsema, P. R. and van Ooyen, A.
(2005).
Attention-gated reinforcement learning of inter-
nal representations for classification. Neural com-
putation, 17(10):2176–2214.
16
Roelfsema, P. R., van Ooyen, A., and Watanabe, T.
(2010). Perceptual learning rules based on rein-
forcers and attention. Trends in cognitive sciences,
14(2):64–71.
Waelti, P., Dickinson, A., and Schultz, W. (2001).
Dopamine responses comply with basic as-
sumptions of formal learning theory. Nature,
412(6842):43–48.
Wiering, M. and Schmidhuber, J. (1998). Fast online
q(λ). Machine Learning, 33(1):105–115.
Xie, X. and Seung, H. S. (2004). Learning in neural
networks by reinforcement of irregular spiking.
Phys Rev E., 69(4):041909.
Yagishita, S., Hayashi-Takagi, A., Ellis-Davies,
G. C., Urakubo, H., Ishii, S., and Kasai, H. (2014).
A critical time window for dopamine actions on
the structural plasticity of dendritic spines. Sci-
ence, 345(6204):1616–1620.
Rombouts, J. O., Bohte, S. M., and Roelfsema, P. R.
(2015). How attention can create synaptic tags for
the learning of working memories in sequential
tasks. PLOS Computational Biology, 11(3):1–34.
Rumelhart, D. E., Hinton, G. E., and Williams,
R. J. (1985). Learning internal representations by
error propagation. Technical report, California
Univ San Diego La Jolla Inst for Cognitive Sci-
ence.
Samsonovich, A. and McNaughton, B. L. (1997).
Path integration and cognitive mapping in a con-
tinuous attractor neural network model. Journal
of Neuroscience, 17(15):5900–5920.
Santoro, A., Bartunov, S., Botvinick, M., Wierstra,
D., and Lillicrap, T. (2016). One-shot learn-
ing with memory-augmented neural networks.
arXiv preprint arXiv:1605.06065.
Schultz, W., Apicella, P., and Ljungberg, T. (1993).
Responses of monkey dopamine neurons to re-
ward and conditioned stimuli during successive
steps of learning a delayed response task. Journal
of neuroscience, 13(3):900–913.
Schultz, W., Dayan, P., and Montague, P. R. (1997).
A neural substrate of prediction and reward. Sci-
ence, 275(5306):1593–1599.
Sutton, R. S. (1984). Temporal Credit Assignment in
Reinforcement Learning. PhD thesis. AAI8410337.
Sutton, R. S. and Barto, A. G. (1998). Reinforcement
learning: An introduction, volume 1. MIT press
Cambridge.
Tetzlaff, C., Kolodziejski, C., Markelic, I., and
Wörgötter, F. (2012). Time scales of memory,
learning, and plasticity. Biological Cybernetics,
106(11):715–726.
Vasilaki, E., Frémaux, N., Urbanczik, R., Senn, W.,
and Gerstner, W. (2009). Spike-based reinforce-
ment learning in continuous state and action
space: When policy gradient methods fail. PLOS
Computational Biology, 5(12):e1000586.
17
Supplementary Material
Model Validation of AuGMEnT Network:
The Saccade-Antisaccade Task
The Saccade-AntiSaccade Task (S-AS), presented
in the AuGMEnT paper [Rombouts et al. 2015], is
inspired by cognitive experiments performed on
monkeys to study the memory representations of
visual stimuli in the Lateral Intra-Parietal cortex
(LIP). The structure of each trial covers different
phases in which different cues are presented on a
screen and at the end of each episode the agent has
to respond accordingly in order to gain a reward.
Actually, following a shaping strategy, monkeys
received also an intermediate smaller reward
when they learnt to fixate on the task-relevant
marks at the center of the screen. The details on the
procedure of the trials and the experimental results
are discussed in Gottlieb and Goldberg [1999].
The agent has to look either to 'Left' (L) or to
'Right' (R) in agreement with a sequence of marks
that appear on a screen at each episode. The
response, corresponding to the direction of the
eye movement (called saccade), depends on the
combination of the location and fixation marks.
The location cue is a circle displayed either at the
left side (L) or at the right side (R) of the screen,
while the fixation mark is a square presented at the
center that indicates whether the final move has to
be concordant with the location cue (Prosaccade -
P) or in the opposite direction (Antisaccade - A).
As a consequence, there are four types of trials
corresponding to the four cue combinations.
As can be seen in Figure 1, each trial is struc-
tured in five phases: a) start, where the screen
is initially empty, b) fix, when the fixation mark
appears c) cue, where the location cue is added on
the screen, d) delay, in which the location circle
disappears for two timesteps, e) go, when the
fixation mark vanishes as well and the agent has
to give the final response to get the reward. Since
the action is given at the end of the trial when the
screen is completely empty, the task can be solved
only if the network stores and maintains both the
Figure 1. Structure of the Saccade-AntiSaccade task.
Structure of the trials in all the possible modalities: P-L
and A-R have final response L (green arrow), while trials
P-R and A-L lead to take action R (red arrow). Figure
taken from publication [Rombouts et al. 2015].
stimuli in memory in spite of the delay phase. In
addition, the shaping strategy mentioned above
is applied in the fix phase of the experiment, by
giving an intermediate reward if the agent gazes
at the fixation mark for two consecutive timesteps,
to ensure that he observes the screen during the
whole trial and that the go response is not random
but consequential to the cues. So, the reward for
the final response is equal to 1.5 units in case of
correct response, 0 otherwise, but the intermediate
reward for the shaping strategy is smaller, equal
to 0.2 units. The most important details about the
trial structure are summarized in Table 1.
The mean trend of the prediction error during
training shows that learning of the S-AS task
is achieved by the of AuGMEnT network also in
our simulation (Figure 2).
In particular, in order
to compare it with the reference performance
in Rombouts et al. [2015] we applied the same
convergence condition,
for which training on
the S-AS task is considered to be successful if
the accuracy for each trial type is higher than
18
Table 1. Table of the S-AS task
Task feature
Task Structure
Inputs
Outputs
Trial Types
Details
5 phases: start, fix, cue, delay, go
Fixation mark: Pro-saccade (P) or Anti-saccade (A)
Location mark: Left (L) or Right (R)
Eye movement: Left (L), Front (F) or Right (R)
1. P+L=L
3. A+L=R
2. P+R=R
4. A+R=L
Training dataset Maximum number of trials is 25, 000.
Each trial type has equal probability.
Correct saccade at go (1.5 units)
Fixation of the screen in fix (0.2 units)
Rewards
Figure 2. Validation of learning perfomance on the saccade task. Decay of the mean predicition error during
training of AuGMEnT networks on the S-AS task, computed as the mean number of errors in 50 trials and averaged
over 100 simulations.
hybrid AuGMEnT and leaky control, proving that
they also solve the S-AS task. However, the time to
convergence is higher (especially for leaky control)
because non-leaky memory is better suited to this
simple task.
90% in the last 50 trials.
In the original paper,
convergence of AuGMEnT is achieved 99.45% of the
times, with a training of around 4, 100 trials.
In
our simulations, the network reaches convergence
every time, with a mean time of 2, 063 trials
(s.d.= 837.7). The slightly better performance in
our simulations could be due to minor differences
in the interpretation of the task structure of S-AS
or of the convergence criterion, but in any case the
error plot in Figure 2 confirms a sharp decrease in
the variability of the prediction error after 4, 000
iterations, compliant with the convergence results
in the reference paper.
In addition, we also show the performance of
19
A Temporal XOR Task
B
S-AS Task
C
12AX Task
Figure 3. Spatial representation of temporal XOR, S-AS and 12AX tasks. Schematic representation of the task
structures indicating for each cue combination the correct response. Temporal XOR (A) and S-AS (B) tasks have
analogous input-output maps and they both have an output space that is not linearly separable. The map of 12AX
task (C) is three-dimensional because its structure is based on three hierarchical levels of inputs instead of two.
However, for better visualization we add at right the sections of the structure space with respect to the digit inputs
that start the outer loops. The output space is still not linearly separable in 3D.
20
|
1607.06563 | 1 | 1607 | 2016-07-22T06:06:46 | Leaky Integrate-and-Fire Neuron under Poisson Stimulation | [
"q-bio.NC"
] | We consider a single Leaky integrate-and-fire neuron stimulated with Poisson process. We develop a method, which allows one to obtain the first passage time probability density function without any additional approximations. | q-bio.NC | q-bio | International Young Scientists Forum on Applied Physics and Engineering
YSF-2016
Leaky Integrate-and-Fire Neuron
under Poisson Stimulation.
Kseniia Kravchuk
Synergetics department
Bogolyubov institute for Theoretical physics
Kyiv, Ukraine
[email protected]
Abstract -We consider a single Leaky integrate-and-fire
neuron stimulated with Poisson process. We develop a method,
which allows one to obtain the first passage time probability
density function without any additional approximations.
Keywords- Leaky
Integrate-and-Fire neuron; Poissson
process; First Passage Time; rigorous solution
I.
INTRODUCTION
If the input stimulation of a neuron were known, what
would be the output? This problem has been addressed many
times during the last 5 decades, see e.g. [1]–[3]. A widely used
way to deal with this problem is to apply the diffusion
approximation. In the diffusion approximation, it is assumed,
that each input impulse, received by a neuron, has an
infinitesimally small effect on the neuron's state (on its
membrane voltage). If so, the number of input impulses needed
to trigger the neuron is infinite. Nevertheless, a careful
consideration of experimental data shows, that the number of
input impulses, needed to trigger a real neuron, is not always
sufficiently large to support the diffusion approximation [4]–
[6]. Moreover, the diffusion approximation does not allow a
consistent theoretical consideration of a neural network, when
the output impulses of some neuron are sent to the input of
another neuron.
That
is why, one should cast out
the diffusion
approximation and consider a more realistic situation, when
each input impulse causes a significant change in neuron's
membrane voltage, and the number of input impulses, needed
to trigger the neuron, is finite. Recently, the first steps were
made in this direction for the Binding neuron model [7] and for
the Leaky integrate-and-fire (LIF) neuron model [8] stimulated
with Poisson stream of input impulses. In [8], the first passage
time probability density function (p.d.f.) is derived for the
particular case, when 2 impulses are enough to trigger the
neuron. Unfortunately, the approach developed in [8] is too
complicated and cannot be extended to the general case. In this
paper, we develop a general method for the LIF model, which
allows to obtain exact analytical results for the first passage
time p.d.f. without any additional approximations.
II. PROBLEM STATEMENT
Consider a separate neuron, stimulated by a sequence of
input impulses. Intervals between input impulses are stochastic
variables. As
input
interspike intervals we take Poisson p.d.f. with intensity .
the probability distribution for
the
Assume that any input impulse has a significant effect on
neuron's state, and that the amount of impulses needed to
trigger the neuron is finite. The neuron is modelled with the
LIF model. We find rigorously a probability density P(t) for
the first passage time of this system. Namely, we assume that
the neuron starts from its resting state at 0 moment of time, and
find the probability P(t)dt that the threshold will be reached for
the first time at the moment t, with precision dt (t is called the
first passage time).
A. Leaky Integrate-and-Fire neuronal model
At any moment of time, the state of LIF neuron is
completely determined by a single real variable V, which
resembles the membrane voltage of a real neuron. The resting
state is associated with V=0. Each input impulse raises the
membrane voltage by a fixed quantity h:
V(ti
+) = V(ti
–)+ h,
(1)
were ti denotes an arrival time of i-th input impulse, i=1,2,....
Between
the membrane voltage decays
exponentially with a character time :
impulses,
input
V(ti + t) = V(ti
+) e– t/,
(2)
were t is a time interval free of input impulses.
LIF neuron fires an output spike, when the membrane
voltage reaches its threshold value V0, and then the neuron
comes back to its resting state V=0.
III. THE METHOD
A. Zones of membrane voltage values
Depending on its parameters, LIF neuron may require
different number of input impulses to reach the threshold and
fire an output spike. Let us denote with N the minimal number
of input impulses, needed to trigger the neuron. An expression
for N can be easily derived from neuron parameters:
N = [V0/h] + 1,
(3)
where the square brackets denote an integer part of the
corresponding expression. In particular realizations of an input
stochastic process, the number of input impulses, which trigger
the neuron, may be different, depending on impulses' arrival
times, but never less than N.
Let us divide the range of membrane voltage values,
V[0;V0], into zones of height h, beginning from the top, see
October 10-14, 2016 Kharkiv, Ukraine
1
International Young Scientists Forum on Applied Physics and Engineering
YSF-2016
let us introduce li,i – 1(t) as a flow of probability across the
zones' borders due to the exponential decay. Namely, li,i – 1(t) dt
gives the probability that at the moment t with precision dt the
trajectory of membrane voltage will cross the border of i-th and
(i – 1)-th zones due to the exponential decay. In calculation of
li,i – 1(t), only trajectories which have never reached the
threshold should be accounted.
Now, all the transitions between zones can be written as the
following system of ODEs:
Fig. 1. Zones of membrane voltage values.
Fig.1. We obtain (N–1) zones of height h and one more zone of
a smaller height at the bottom. The zones can be assigned with
numbers, e.g. as at the Fig. 1:
1-st zone: V (0; V0 – (N–1)h),
i-th zone: V [V0 – (N–i+1)h; V0 – (N–i)h), i = 2,3,…,N.
(4)
The resting state, V=0, should be considered as a separate
zone, because of nonzero probability, associated with it. So, let
us define a 0 zone as follows:
0 zone: V = 0.
(5)
Let us introduce probabilities p(i,t), i = 0,1,…,N to have the
membrane voltage within i-th zone at the moment t, given the
neuron didn't reach the threshold before t. We would like to
emphasize, that we are not going to find probabilities of the
specific values of V at different moments of time – just the
probabilities to have the membrane voltage within the certain
limits, see (4). Such simplification can be made without any
loose of precision in P(t). Indeed, let us imagine, that at some
moment of time the membrane voltage takes its value within
N-th zone: V[V0 – h; V0). If the neuron obtains an input
impulse at this moment, then the membrane voltage will reach
its threshold, and an output spike will be generated. The
generation of an output impulse will not depend on the exact
value of V, since V belongs to the N-th zone. Therefore, output
ISI probability density can be found as a product of p(N,t) and
the probability to obtain an input impulse within time interval
[t;t+dt), which is dt for a Poisson process:
P(t)dt = p(N,t) dt,
(6)
were is the intensity of input Poisson process. So, the first
passage time problem reduces to finding p(N,t).
B. Equations
It seems, that the probability p(N,t) should be found
together with probabilities p(i,t) for the rest of zones, because
the membrane voltage may get across the zone's borders. The
mechanisms of such crossings are
jumps,
exponential decay.
The first one is activated with the arrival of input impulse, and
guarantees, that V will move for one zone up. The second one
is activated permanently, and provides the possibility to move
for one zone down. In order to account the second mechanism,
Here, the terms like p(i,t) in the right-hand side correspond to
the jumping out of the i-th zone upward with the arrival of
input impulse (see the first mechanism, above). The terms like
li,i – 1(t) describe the probability flow across the zones' borders
due
leak. This system should be
supplemented with initial conditions:
the exponential
to
p(0,t)t=0 = 1,
p(i,t)t=0 = 0,
i = 1,2,…,N.
(8)
In order to solve (7) and to obtain p(N,t), one should first
find expressions for all li,i – 1(t).
C. The flow of probability across the zones borders
The crossing of a given zones' border at definite time t may
happen in different trajectories. Such trajectories may contain a
different number k of input impulses, obtained within time
interval (0;t), and the arrival times of such impulses may differ
as well. That is why, it is natural to find li,i – 1(t) as a sum by k:
where functions lk
i,i–1(t) correspond to trajectories with k input
impulses within interval (0;t). Namely, lk
i,i–1(t) dt gives the
probability of crossing the border of i-th and (i-1)-th zones at
time t with precision dt due to exponential decay (only
trajectories which never reached the threshold should be
accounted).
D. Minimum and maximum number of input impulses
The number k of input impulses which contribute to li,i – 1(t)
in (9) should be limited both from the top and from the bottom.
Indeed, k should be sufficiently large to let the membrane
voltage reach the zones' border (which in accordance with (4)
is determined by equality V = V0 – (N – i +1) h):
(kmin)i = [(V0 – (N – i + 1)h) / h]+ 1 = i – 1,
(10)
where the definition (3) is taken into account.
October 10-14, 2016 Kharkiv, Ukraine
2
International Young Scientists Forum on Applied Physics and Engineering
YSF-2016
j is a time interval needed to the membrane voltage to decay from
Fig. 2. Ti
the value V=jh to the border of i-th and (i-1)-th zones, V=V0–(N–i +1)h.
On the other hand, k cannot be arbitrarily large, because
otherwise the threshold will be reached before the moment t.
This imposes an upper limit for the possible values of k. In
order to find such limit, let us first define two sets of time
intervals with the following expressions:
j is a time interval, needed for the membrane voltage to decay
Ti
from the level V=jh to the border of i-th and (i-1)-th zones (see
Fig. 2), and
is the time interval which it takes to the
membrane voltage to decay from V0 to the same border.
Now, consider the closest packing of input impulses within
time t, which does not cause triggering and allows crossing the
zones border afterwards, see Fig.3. The first (N–1) impulses
may arrive arbitrarily close in time to one another, and they
would not be able to trigger the neuron. And the others should
be separated with time intervals, needed to the membrane
voltage to decay to the level V0–h, to avoid triggering. After the
(kmax)i-th impulse there should be a time interval
to let the
– to the level V = V0 – (N –
membrane voltage drop from V = V0
i +1) h of zones' border, see Fig. 3.
Therefore, one obtains
be rewritten using (11) as follows:
, which can
Fig. 3. The closest packing of (N+2) input impulses, which does not cause
triggering. * – the first (N-1) impulses may arrive arbitrarily close to zero
time moment.
If t < Ti
i-1, the probability flow across the zones' border is
zero:
li,i–1(t)=0,
t < Ti
i-1,
(14)
because any trajectory will not have enough time to reach the
i-1, the trajectories may manage to cross the
border. If t ≥ Ti
zones' border and the probability flow emerges. Therefore, in
this case, one should obtain all the lk
i,i–1(t) in order to find
li,i–1(t).
E. Functions lk
i,i-1(t)
Trajectories contribute to lk
i,i–1(t) if the following conditions
are satisfied:
1.
2.
time interval (0;t) contains k input impulses;
trajectory crosses the border between i-th and (i-1)-th
zones at time t with precision dt;
3. trajectory haven't reached the threshold until the
moment t.
Then lk
i,i–1(t) will be a measure of trajectories which satisfy
conditions (1)–(3), above. This measure can be found as
follows.
Let us denote with t1, t2,…, tk the arrival times of input
impulses in a trajectory which contributes to lk
i,i–1(t). The
probability to obtain from the Poisson stream k input impulses
within time interval (0;t) at the moments t1, t2,…, tk equals
ke–t. In order to obtain lk
i,i–1(t) one should integrate the this
expression over all possible values of arrival times tj:
Similarly, if
, (kmax)i can be obtained as
(kmax)i = i –1 + j, t [Ti
i-1+j; Ti
i+j )j, j = 0,1,…, N – i –1
(13)
Finally, we have obtained, that (kmax)i increases with the
increase in t and changes its value by unity at the points
and the corresponding values of (kmax)i are i–1, i,…, N–1, N,
N+1, N+2, N+3, …
where the integration domain T is defined by conditions 2 and
3, above. In (15), the integration over t1 is missing, because the
variables t1, t2,…, tk are not independent. Indeed, condition 2
ensures that one of the arrival times can be found using all the
others, e.g. one can find t1 as a function of t2,…, tk and t.
Domain T in the space with coordinates t2,…, tk has a
complicated structure, due to the conditions 2 and 3, above.
This problem may be simplified by appropriate variable
change. Let us introduce variables V2 , V3 ,…, Vk, where Vj ,
j=2,…,k, denotes the membrane voltage just after the arrival of
j-th input impulse:
October 10-14, 2016 Kharkiv, Ukraine
3
International Young Scientists Forum on Applied Physics and Engineering
YSF-2016
Vj V(tj
+),
j = 2, 3, …, k.
(16)
were
is defined by the following expressions:
The mapping t2,…,tk V2,…,Vk is one-to-one, so we can
perform the corresponding variable change in (15):
where the integration domain V in a space with coordinates
V2,…, Vk corresponds to the domain T in a space with
coordinates t2,…, tk and is defined by the conditions 2 and 3,
above. The Jacobean in (17), can be easily found using (1), (2)
and (16):
Such variable change makes it possible to find upper limits
of integration easily. Indeed, the first N–1 impulses may be
located arbitrarily close to each other, therefore, Vj < jh for j =
2,3,…, N – 1. Then, after the N-th input impulse, the threshold
can be reached and the condition 3 should be accounted. In
terms of the membrane voltage, this condition can be written as
the simple inequality Vj < V0,
j = N, N+1,…, k. Finally,
one can write formally:
were the upper limits are
and the lower limits Vj are still to be found using the
conditions 2 and 3, above.
In order to find Vj, one should rewrite the condition 2 in
terms of membrane voltage variables Vj, using (1) and (2):
The lower limit Vj is achieved, if all the previous impulses, m =
1,…, j–1, are located as early as possible, given the conditions
, m = 1,2,…,j–1, the
1–3, above. Let us denote with
membrane voltage after the arrival of m-th input impulse in
such arrangement. The first (N–1) impulses may be located as
close as possible to the moment 0,
tm = 0, m=1,2,…, N–1,
(22)
and the others should be separated with time intervals, needed
to membrane voltage to decay to the level V0–h, as at Fig. 3.
That is why,
So, finally, in order to obtain the first passage time p.d.f.,
one should 1) find all the lk
i,i–1(t), using expressions (17)–(20),
(24) and (25); 2) for any fixed i, sum up all the lk
i,i–1(t) and
obtain the probability flow li,i–1(t) across the zones' border,
using eqs. (9), (10), (12) and (13); 3) solve the system (7) of
ODEs for the probabilities p(i,t) to have the membrane voltage
within i-th zone; 4) having p(N,t), obtain expression for the
first passage time p.d.f., using (6).
IV. CONCLUSSIONS
In a present paper, we have considered a single leaky
integrate-and-fire neuron stimulated with Poisson stream of
input impulses. We have developed a method, which allows
one to obtain rigorous expressions for the first passage time
p.d.f. for such neuron. The expressions obtained will be
different for different relations between neuron's threshold and
the altitude of input impulse in the LIF model. In a future, we
intend to apply the developed method to a particular case of
such relation and to obtain specific expressions for the first
passage time p.d.f. in this case.
REFERENCES
[1] P.I.M. Johannesma, "Diffusion models for the stochastic activity of
neurons" in Neural networks, R. Caianiello, Ed. Berlin: Springer,
pp. 116-144, 1968.
[2] L.M. Ricciardi "Diffusion Processes and Related Topics in Biology" in
Lecture Notes in Biomathematics, Vol. 14, Berlin: Springer, 1977.
[3] S. Ditlevsen, P. Lansky, "Parameters of stochastic diffusion processes
estimated from observations of first-hitting times: Application to the
leaky integrate-and-fire neuronal model", Phys. Rev. E., Vol. 76,
p. 041906, 2007.
[4] R. Miles, "Synaptic excitation of inhibitory cells by single CA3
hippocampal pyramidal cells of the guinea-pig in vitro", J. Physiol., vol.
428, pp. 61-77, September 1990.
[5] B. Barbour, "Synaptic currents evoked by purkinje cells by stimulating
individual granule cells", Neuron, vol. 11, no. 4, pp. 759-769, October
1993.
[6] P. Andersen, "Synaptic integration in hyppocampal neurons" in Fidia
Research Foundation Neuroscience Award Lectures, New York: Raven
Press Ltd., 1991, pp. 51-71.
Substituting values of
from (23) and t1=0 from (22) to
(21), one obtains the lower limit Vj as a function of Vj+1,…,Vk:
October 10-14, 2016 Kharkiv, Ukraine
[7] A.K. Vidybida, "Output stream of a binding neuron", Ukrainian
Mathematical Journal, Springer US, vol. 59, no. 12, pp. 1819-1839,
2008.
[8] A.K. Vidybida, "Output stream of leaky integrate-and-fire neuron",
Reports of the NAS of Ukraine, vol. 12, pp. 18-23, 2014.
4
|
1209.3411 | 1 | 1209 | 2012-09-15T14:28:54 | A Computational Model of the Effects of Drug Addiction on Neural Population Dynamics | [
"q-bio.NC",
"cs.SI"
] | Reward processing and derangements thereof, such as drug addiction, involve the coordinated activity of many brain areas. Prior work has identified many behavioral, molecular biological and single neuron changes throughout the mesocorticolimbic system that reflect and drive addictive behavior. Subpopulations in the ventral tegemental area (VTA) encode positive reward prediction error, negative reward prediction error, and the magnitude of the reward. Phasic activity in VTA dopaminergic neurons correlates with hedonic value. Tonic activity of groups in the dorsomedial prefrontal cortex (dmPFC) can encode antidepressant states. However, little is known about how drug addiction might affect population encoding across larger brain regions. Here, we compare the information content associated with network patterns in naive, acutely intoxicated and chronically addicted states in a plastic attractor network. We found that addiction decreases the network's ability to store and discriminate among patterns of activity. Altered dopaminergic tone flattens the energy landscape and decreases the entropy associated with each network pattern. Altered dmPFC activity produces signal-to-noise deficits similar to computational models of schizophrenia. Our results provide a conceptual framework for interpreting altered neural population dynamics in psychopathological states based on information theory. They also suggest a view of the subtypes of depression as on a continuum of combinations of cortical and subcortical dysfunction. This suggests that patients who suffer from depression with psychotic features will have more cortical than mesolimbic dysfunction. Furthermore, our framework can be applied to other psychiatric illnesses and so may help us, in general, quantitatively understand psychiatric illnesses as disorders in the representation and processing of information by distributed brain networks. | q-bio.NC | q-bio |
Model
Evolution of Voltage over Time
Consider a group of interacting populations of neurons. Equation 1 describes the evolution over time of the membrane
voltages for the ith ensemble. Table 1 explains the remaining variables.
τi
d~vi
dt
= −~vi + Mii~vi + X
i6=j
Mij~vj + Wi~u
Symbol Meaning
Mij
Wi
~u
τ
Connections from jth ensemble to ith
Weight of input for ith ensemble
Input
Time constant
Table 1: Meaning of symbols in Equation 1
To represent n ensembles, we may combine n versions of Equation 1, as Equation 2 describes.
τv
d~v
dt
= −~v + M~v + W~u
(1)
(2)
~v =
~vi
~vj
...
~vn
, M =
Mii Mij
Mji Mjj
...
Mni
. . . Min
. . .
Mnn
If the real part of the eigenvalues of M are less than one, then the system will evolve to the voltage that Equation 3
describes.
~v∞ = KW~u K = (I − M)−1
(3)
Islands of Steady State Behavior
M need not be symmetric. But, if some block matrices within it are, the corresponding subpopulations can approach
their own steady state even if the network is still unstable. If, furthermore, the combination of recurrent and feedforward
input to those subpopulations is a saturating function, then the system's Lyapunov function is bounded and fixed points
for that subpopulation must exist (Cohen and Grossberg, 1983).
One reasonable way to create a symmetric weight matrix for Mk, the block matrix that describes the kth ensemble
is to assume that it recognizes one of N memory patterns, (cid:8)v1, v2, . . . , vm, . . . , vN(cid:9). Assume that a subpopulation with
n neurons signals its recognition of any memory pattern, ~vm, by displaying a voltage vector c~vm. One matrix that
accomplishes this is:
M =
λ
c2αN (1 − α)
N
X (~v − αcn) ⊗ (~v − αcn) −
n ⊗ n
αN
(4)
Longer-term Effects on Neural Population Dynamics
Assume that two additional processes occur as defined by Equations 5 and 6. Both are much slower than Equation
2. One modifies W and an even slower one modifies M.
τW
dW
dt
= h~v~ui − α h~v~vi W α > 0
τM
dM
dt
= − (W~u) ~v + K−1
(5)
(6)
Replacing ~v with ~v∞in Equations 5 and 6 yields Equations 7 and 8. Note the appearance of the autocorrelation
matrices for the stimulus, Q, and network activity, R.
τW
dW
dt
= KWQ − αR∞W Q = h~u~ui , R∞ = h ~v∞ ~v∞i
(7)
1
τM
dM
dt
= − (W~u) (KW~u) + K−1
(8)
Equations 7 and 8 describe somewhat contrasting behaviors. Equation 7 aligns the correlation structure of the
network activity with the strength of its recurrent connections. Equation 8 urges the outputs to be decorrelated. A faster
correlating influence and slower decorrelating one allow oscillations in the correlation of network activity. According to
this model, interestingly, those oscillations are dependent on input but not directly on the correlation structure of the
input.
Remarks on the Structure of the Model
There is an interesting concordance between R and Equation 4. If Rmem = h~vmem~vmemi, then M ∝ P Rmem. This
further highlights how interrelated Equations 7 and 8 are.
Effect of Correlation Structure on the Dynamics of the Feedforward Weights
If the stimulus is random, that is Q = I, then the feedforward weights, W, stop changing only when K = αR∞. Com-
bining the definition of K in Equation 3 with the observation that R∞ must haverank 1 we note that such stimuli prevent
this system from recognizing any pattern that requires a distribution of activities over the network. Moreover, consider-
ing the rank-nullity theorem and ranks of Q and R, one can see that this result holds for any stimulus autocorrelation
matrix, Q, that results from the outer product of a vector with itself.
It is next natural to consider how this system responds to many superimposed stimuli that each have different corre-
lation structures. That is, consider aQ that results from the sum of i correlation matrices. Each of those autocorrelation
matrices results from the outer product of the ith activity pattern with itself, as Equation 9 describes.
Q = X
i
h~ui~uii
(9)
If we assume that all the input patterns are pairwise independent, then the rank of Q becomes equal to i, whose
upper bound we assume to be the number of neurons. This stands in contrast to the result of the previous paragraph.
Effects of Correlation Structure on the Dynamics of Network Connections
By studying Equation 8 we can gain some insight into how ~u and W influence M. If ~u lies in the nullspace of the
columns of W, then M approaches the identity matrix. Said another way, when ~u and W jam each other,M tries to
span the largest basis it can.
The contrary effects of
Nonrandom Constant Stimulus
Consider any brief nonrandom stimulus such that, Q 6= I but is constant. ThenW only stops changing if the columns
of Q are in the nullspace of the rows of W. BecauseQ results from the tensor product of a vector with itself, it has rank 1
and so it may lie in the nullspace of W.
Brief Pulse
If a stimulus is presented to the system for a brief period of time, it will only cause sustained activity in the system if
the activity that it induces is itself a fixed point of the system. That is, for a stimulus, ~u, and its response, ~v, Equation
must hold.
Extension to Drug Addiction
Craving .
Similar response to variable reinforcement and huge rewards. Let the ith subpopulation of Equation 2, in analogy with
a proposed role for the dopaminergic neurons in ventral tegmental area (VTA) represent the positive expecrted reward
of a stimulus. Let another population, the jth one, in analogy with the GABAergic neurons in the VTA represent the
negative expected reward.
2
|
1610.00161 | 3 | 1610 | 2017-04-07T18:45:30 | Towards deep learning with segregated dendrites | [
"q-bio.NC"
] | Deep learning has led to significant advances in artificial intelligence, in part, by adopting strategies motivated by neurophysiology. However, it is unclear whether deep learning could occur in the real brain. Here, we show that a deep learning algorithm that utilizes multi-compartment neurons might help us to understand how the brain optimizes cost functions. Like neocortical pyramidal neurons, neurons in our model receive sensory information and higher-order feedback in electrotonically segregated compartments. Thanks to this segregation, the neurons in different layers of the network can coordinate synaptic weight updates. As a result, the network can learn to categorize images better than a single layer network. Furthermore, we show that our algorithm takes advantage of multilayer architectures to identify useful representations---the hallmark of deep learning. This work demonstrates that deep learning can be achieved using segregated dendritic compartments, which may help to explain the dendritic morphology of neocortical pyramidal neurons. | q-bio.NC | q-bio |
Towards deep learning with segregated dendrites
Jordan Guergiuev1,2, Timothy P. Lillicrap4, and Blake A. Richards1,2,3,*
1 Department of Biological Sciences, University of Toronto Scarborough, Toronto, ON,
Canada
2 Department of Cell and Systems Biology, University of Toronto, Toronto, ON, Canada
3 Learning in Machines and Brains Program, Canadian Institute for Advanced Research,
Toronto, ON, Canada
4 DeepMind Technologies Inc., London, UK
* Corresponding author, email: [email protected]
Abstract
Deep learning has led to significant advances in artificial intelligence, in part, by adopting strategies
motivated by neurophysiology. However, it is unclear whether deep learning could occur in the real brain.
Here, we show that a deep learning algorithm that utilizes multi-compartment neurons might help us to
understand how the brain optimizes cost functions. Like neocortical pyramidal neurons, neurons in our
model receive sensory information and higher-order feedback in electrotonically segregated compartments.
Thanks to this segregation, the neurons in different layers of the network can coordinate synaptic weight
updates. As a result, the network can learn to categorize images better than a single layer network.
Furthermore, we show that our algorithm takes advantage of multilayer architectures to identify useful
representations-the hallmark of deep learning. This work demonstrates that deep learning can be
achieved using segregated dendritic compartments, which may help to explain the dendritic morphology
of neocortical pyramidal neurons.
Introduction
Deep learning refers to an approach in artificial intelligence (AI) that utilizes neural networks with
multiple layers of processing units. Importantly, deep learning algorithms are designed to take advantage
of these multi-layer network architectures in order to generate hierarchical representations wherein each
successive layer identifies increasingly abstract, relevant variables for a given task (Bengio and LeCun,
2007; LeCun et al., 2015). In recent years, deep learning has revolutionized machine learning, opening
the door to AI applications that can rival human capabilities in pattern recognition and control (Mnih
et al., 2015; Silver et al., 2016; He et al., 2015). Interestingly, the representations that deep learning
generates resemble those observed in the neocortex, particularly in higher-order sensory areas (Kubilius
et al., 2016; Khaligh-Razavi and Kriegeskorte, 2014; Cadieu et al., 2014), suggesting that something akin
to deep learning is occurring in the mammalian brain (Yamins and DiCarlo, 2016; Marblestone et al.,
2016).
Yet, a large gap exists between deep learning in AI and our current understanding of learning and
memory in neuroscience. In particular, unlike deep learning researchers, neuroscientists do not yet have a
solution to the "credit assignment problem" (Rumelhart et al., 1986; Lillicrap et al., 2016; Bengio et al.,
2015). The credit assignment problem refers to the fact that the behavioral consequences of synaptic
changes in early layers of a neural network depend on the current connections in the downstream layers.
For example, consider the behavioral effects of synaptic changes, i.e. long-term potentiation/depression
1/41
(LTP/LTD), occurring between different sensory circuits of the brain. Exactly how these synaptic
changes will impact behavior and cognition depends on the downstream connections between the sensory
circuits and motor or associative circuits (Figure 1A). Hence, learning to optimize some behavioral or
cognitive function requires a method for identifying what a given neuron's place is within the wider
neural network, i.e. it requires a means of assigning "credit" (or "blame") to neurons in early sensory
areas for their contribution to the final behavioral output (LeCun et al., 2015; Bengio et al., 2015).
Despite its importance for real-world learning, the credit assignment problem has received little attention
in neuroscience.
The lack of attention to credit assignment in neuroscience is, arguably, a function of the history of
biological studies of synaptic plasticity. Due to the well-established dependence of LTP and LTD on
presynaptic and postsynaptic activity, current theories of learning in neuroscience tend to emphasize
Hebbian learning algorithms (Dan and Poo, 2004; Martin et al., 2000), that is, learning algorithms where
synaptic changes depend solely on presynaptic and postsynaptic activity. Hebbian learning models can
produce representations that resemble the representations in the real brain (Zylberberg et al., 2011; Leibo
et al., 2017) and they are backed up by decades of experimental findings (Malenka and Bear, 2004; Dan
and Poo, 2004; Martin et al., 2000). But, current Hebbian learning algorithms do not solve the credit
assignment problem, nor do global neuromodulatory signals used in reinforcement learning (Lillicrap
et al., 2016). As a result, deep learning systems from AI that can do credit assignment outperform existing
Hebbian models of perceptual learning on a variety of tasks (Yamins and DiCarlo, 2016; Khaligh-Razavi
and Kriegeskorte, 2014). This suggests that a critical, missing component in our current models of the
neurobiology of learning and memory is an explanation of how the brain solves the credit assignment
problem.
However, the most common solution to the credit assignment problem in AI is to use the back-
propagation of error algorithm (Rumelhart et al., 1986). Backpropagation assigns credit by explicitly
using current downstream synaptic connections to calculate synaptic weight updates in earlier layers,
commonly termed "hidden layers" (LeCun et al., 2015) (Figure 1B). This technique, which is sometimes
referred to as "weight transport", involves non-local transmission of synaptic weight information between
layers of the network (Lillicrap et al., 2016; Grossberg, 1987). Weight transport is clearly unrealistic
from a biological perspective (Bengio et al., 2015; Crick, 1989). It would require early sensory processing
areas (e.g. V1, V2, V4) to have precise information about billions of synaptic connections in downstream
circuits (MT, IT, M2, EC, etc.). According to our current understanding, there is no physiological
mechanism that could communicate this information in the brain. Some deep learning algorithms utilize
purely Hebbian rules (Scellier and Bengio, 2016; Hinton et al., 2006). But they depend on feedback
synapses that are symmetric with feedforward synapses to solve the credit assignment problem (Scellier
and Bengio, 2016; Hinton et al., 2006), which is essentially a version of weight transport. Altogether,
these artificial aspects of current deep learning solutions to credit assignment have rendered many
scientists skeptical to the proposal that deep learning occurs in the real brain (Crick, 1989; Grossberg,
1987; Harris, 2008; Urbanczik and Senn, 2009).
Recent findings have shown that these problems may be surmountable, though. Lillicrap et al. (2016),
Lee et al. (2015) and Liao et al. (2015) have demonstrated that it is possible to solve the credit assignment
problem even while avoiding weight transport or symmetric feedback weights. The key to these learning
algorithms is the use of feedback signals from output layers that are transmitted to calculate a local
error signal, which then guides synaptic updates in hidden layers (Lee et al., 2015; Lillicrap et al., 2016;
Liao et al., 2015). These learning algorithms can take advantage of multi-layer architectures, leading
to performance that rivals backpropagation (Lee et al., 2015; Lillicrap et al., 2016; Liao et al., 2015).
Hence, this work has provided a significant breakthrough in our understanding of how the real brain
might do credit assignment.
Nonetheless, the models of Lillicrap et al. (2016), Lee et al. (2015) and Liao et al. (2015) involve
some problematic assumptions. Specifically, although it is not directly stated in all of the papers, there
is an implicit assumption that there is a separate feedback pathway for transmitting the signals that
drive synaptic updates in hidden layers (Figure 2A). Such a pathway is required in these models because
the synaptic weight updates in the hidden layers depend on the difference between feedback that is
2/41
generated in response to a purely feedforward propagation of sensory information, and feedback that
is guided by a teaching signal (Lillicrap et al., 2016; Lee et al., 2015; Liao et al., 2015). In order to
calculate this difference, sensory information must be transmitted separately from the feedback signals
that are used to drive learning. In single compartment neurons, keeping feedforward sensory information
separate from feedback signals is impossible without a separate pathway. At face value, such a pathway
is possible. But, closer inspection uncovers a couple of difficulties with such a proposal.
First, the error signals that solve the credit assignment problem are not global error signals (like
neuromodulatory signals used in reinforcement learning). Rather, they are cell-by-cell error signals. This
would mean that the feedback pathway would require some degree of pairing, wherein each neuron in
the hidden layer is paired with a feedback neuron (or circuit). That is not impossible, but there is no
evidence to date of such an architecture in the neocortex. Second, the error that is communicated to
the hidden layer is signed (i.e. it can be positive or negative), and the sign determines whether LTP
or LTD occur in the hidden layer neurons (Lee et al., 2015; Lillicrap et al., 2016; Liao et al., 2015).
Communicating signed signals with a spiking neuron can theoretically be done by using a baseline firing
rate that the neuron can go above (for positive signals) or below (for negative signals). But, in practice,
such systems are difficult to operate because as the error gets closer to zero any noise in the spiking of
the neuron can switch the sign of the signal, which switches LTP to LTD, or vice versa. This means
that as learning progresses the network's ability to communicate error signs gets worse. Therefore, the
real brain's specific solution to the credit assignment problem is unlikely to involve a separate feedback
pathway for cell-by-cell, signed signals to instruct plasticity.
However, segregating the integration of feedforward and feedback signals does not require a separate
pathway if neurons have more complicated morphologies than the point neurons typically used in artificial
neural networks. Taking inspiration from biology, we note that real neurons are much more complex
than single-compartments, and different signals can be integrated at distinct dendritic locations. Indeed,
in the primary sensory areas of the neocortex, feedback from higher-order areas arrives in the distal
apical dendrites of pyramidal neurons (Manita et al., 2015; Budd, 1998; Spratling, 2002), which are
electrotonically very distant from the basal dendrites where feedforward sensory information is received
(Larkum et al., 1999, 2007, 2009). Thus, as has been noted by previous authors (Kording and Konig,
2001; Spratling, 2002; Spratling and Johnson, 2006), the anatomy of pyramidal neurons may actually
provide the necessary segregation of feedforward and feedback information to calculate local error signals
and perform deep learning in biological neural networks.
Here, we show how deep learning can be implemented if neurons in hidden layers contain segregated
"basal" and "apical" dendritic compartments for integrating feedforward and feedback signals separately
(Figure 2B). Our model builds on previous neural networks research (Lee et al., 2015; Lillicrap et al.,
2016) as well as computational studies of supervised learning in multi-compartment neurons (Urbanczik
and Senn, 2014; Kording and Konig, 2001; Spratling and Johnson, 2006). Importantly, we use the distinct
basal and apical compartments in our neurons to integrate feedback signals separately from feedforward
signals. With this, we build a local target for each hidden layer that coordinates learning across the layers
of the network. We demonstrate that even with random synaptic weights for feedback into the apical
compartment, our algorithm can coordinate learning to achieve classification of the MNIST database of
hand-written digits better than can be achieved with single layer networks. Furthermore, we show that
our algorithm allows the network to take advantage of multi-layer structures to build hierarchical, abstract
representations, one of the hallmarks of deep learning (LeCun et al., 2015). Our results demonstrate
that deep learning can be implemented in a biologically feasible manner if feedforward and feedback
signals are received at electrotonically segregated dendrites, as is the case in the mammalian neocortex.
Results
A network architecture with segregated dendritic compartments
Deep supervised learning with local weight updates requires that each neuron receive signals that
can be used to determine its "credit" for the final behavioral output. We explored the idea that the
3/41
cortico-cortical feedback signals to pyramidal cells could provide the required information for credit
assignment. In particular, we were inspired by four observations from both machine learning and biology:
1. Current solutions to credit assignment without weight transport require segregated feedforward
and feedback signals (Lee et al., 2015; Lillicrap et al., 2016).
2. In the neocortex, feedforward sensory information and higher-order cortico-cortical feedback are
largely received by distinct dendritic compartments, namely the basal dendrites and distal apical
dendrites, respectively (Spratling, 2002; Budd, 1998).
3. The distal apical dendrites of pyramidal neurons are electrotonically distant from the soma, such
that passive transmission from the distal dendrites to the soma is significantly attenuated. Instead,
apical communication to the soma depends on active propagation through the apical dendritic
shaft driven by voltage-gated calcium channels. These non-linear, active events in the apical shaft
generate prolonged upswings in the membrane potential, known as "plateau potentials", which can
drive burst firing at the soma (Larkum et al., 1999, 2009).
4. Plateau potentials driven by apical activity can guide plasticity in vivo (Bittner et al., 2015).
With these considerations in mind, we hypothesized that the computations required for credit
assignment could be achieved without any separate pathways for feedback signals. Instead, they could
be achieved by having two distinct dendritic compartments in each hidden layer neuron: a "basal"
compartment, strongly coupled to the soma for integrating feedforward information, and an "apical"
compartment for integrating feedback information that would only drive activity at the soma when
"plateau potentials" occur (Figure 3A).
As an initial test of this concept we built a network with a single hidden layer. Although this network
is not very "deep", even a single hidden layer can improve performance over a one-layer architecture if
the learning algorithm solves the credit assignment problem (Bengio and LeCun, 2007; Lillicrap et al.,
2016). Hence, we wanted to initially determine whether our network could take advantage of a hidden
layer to reduce error at the output layer.
The network architecture is illustrated in Figure 3A. An image from the MNIST data set is used to
set the spike rates of (cid:96) = 784 Poisson point-process neurons in the input layer (one neuron per image
pixel, rates-of-fire determined by pixel intensity). These project to a hidden layer with m = 500 neurons.
The neurons in the hidden layer are composed of three distinct compartments with their own voltages:
the apical compartments (with voltages described by the vector A(t) = [A1(t), ..., Am(t)]), the basal
compartments (with voltages B(t) = [B1(t), ..., Bm(t)]), and the somatic compartments (with voltages
C(t) = [C1(t), ..., Cm(t)]). (Note: for notational clarity, all vectors and matrices in the paper are in
boldface.) The voltages in the dendritic compartments are calculated as weighted sums of a postsynaptic
potential kernel convolved with the incoming spike train (see Materials and Methods, equations (10) and
(11)), while the somatic voltages are calculated as leaky integrators of the dendritic inputs, i.e. for the
ith hidden layer neuron:
dCi(t)
dt
= −gLCi(t) + gB(Bi(t) − Ci(t)) + gA(Ai(t) − Ci(t))
(1)
where gL, gB and gA represent the leak conductance, the conductance from the basal dendrites,
and the conductance from the apical dendrites, respectively. Note, for mathematical simplicity we are
assuming a resting membrane potential of 0 V and a membrane capacitance of 1 F (these values do not
affect the results). We implement electrotonic segregation in the model by altering the gA value-low
values for gA lead to electrotonically segregated apical dendrites. In the initial set of simulations we
set gA = 0, but we relax this hard constraint in later simulations. The somatic compartments generate
spikes using Poisson processes. The instantaneous rates of these processes are described by the vector
4/41
λC(t) = [λC
m(t)], which is in units of spikes/s or Hz. These rates-of-fire are determined by a
non-linear sigmoid function, σ(·), applied to the somatic voltages, i.e. for the ith hidden layer neuron:
1 (t), ..., λC
λC
i (t) = λmaxσ(Ci(t))
= λmax
1
1 + e−Ci(t)
(2)
where λmax is the maximum rate-of-fire for the neurons.
Spiking inputs from the input layer arrive at the basal compartments via the m × (cid:96) synaptic weight
matrix W 0, hidden layer somatic spikes are projected to the output layer neurons via the n× m synaptic
weight matrix W 1, and spiking inputs from the output layer arrive at the apical compartments via the
m × n synaptic weight matrix Y (Figure 3A).
The output layer neurons consist of n = 10 two compartment neurons (one for each image category),
similar to those used in a previous model of dendritic prediction learning (Urbanczik and Senn, 2014). The
output dendritic voltages (V (t) = [V1(t), ..., Vn(t)]) and somatic voltages (U (t) = [U1(t), ..., Un(t)]) are
updated in a similar manner to the hidden layer basal compartment and soma, respectively (see Materials
and Methods, equations (14) and (15)). As well, like the hidden layer neurons, the output neurons
spike using Poisson processes whose instantaneous rates, λU (t) = [λU
n (t)], are determined by
the somatic voltages, i.e. λU
i (t) = λmaxσ(Ui(t)). Unlike the hidden layer, however, the output layer
neurons also receive a set of "teaching signals", specifically, excitatory and inhibitory conductances that
push their voltages towards target values (see Materials and Methods, equations (15) and (16)). These
teaching signals are similar to the targets that are used for training in deep artificial neural networks
(Bengio and LeCun, 2007; LeCun et al., 2015). Whether any such teaching signals exist in the real brain
is unknown, though there is evidence that animals can represent desired behavioral outputs with internal
goal representations (Gadagkar et al., 2016).
1 (t), ..., λU
Critically, we define two different modes of integration in the hidden layer neurons: "transmit"
and "plateau". During the transmit mode, the apical compartment is considered to be electrotonically
segregated from the soma so it affects the somatic voltage minimally (depending on gA), and most of
the conductance to the soma comes from the basal compartment (Figure 3B, left). In contrast, during
the plateau mode, the apical voltage is averaged over the most recent 20-30 ms period and the sigmoid
non-linearity is applied to it, giving us "plateau potentials" (see equation (3) below). These non-linear
versions of the apical voltages are then transmitted to the somatic and basal compartments for synaptic
updates (Figure 3B, right). The intention behind this design was to mimic the non-linear transmission
from the apical dendrites to the soma that occurs during a plateau potential driven by calcium spikes in
the apical dendritic shaft (Larkum et al., 1999). Furthermore, the temporal averaging was intended to
mimic, in part, the temporal dynamics introduced by NMDA plateaus (Schiller et al., 2000), which are
particularly good for driving apical calcium spikes (Larkum et al., 2009).
To train the network we alternate between two phases. First, we present an image to the input layer
without any signals to the output layer during the "forward" phase, which occurs between times t0 to t1.
At t1 a plateau potential occurs in all the hidden layer neurons and the "target" phase begins. During
this phase the image continues to drive the input layer, but now the output layer also receives teaching
conductances, which force the output units closer to the correct answer. For example, if an image of a '9'
is presented, then over the time period t1-t2 the '9' neuron in the output layer receives strong excitatory
conductances from the teaching signal, while the other neurons receive strong inhibitory conductances
(Figure 3C). This phase lasts from t1 to t2, and at t2 another set of plateau potentials occurs in the
hidden layer neurons. The result is that we have plateaus potentials in the hidden layer neurons for both
the forward (αf = [αf
m]) and target (αt = [αt
1 , ..., αf
1, ..., αt
m]) phases:
5/41
(cid:90) t1
(cid:90) t2
t1−tp
t2−tp
αf
i = σ(
αt
i = σ(
Ai(t)dt)
Ai(t)dt)
(3)
where tp is the integration time for the plateau potential (see Materials and Methods). Note that
since the plateau potentials are generated using the same sigmoid non-linearity that we use to calculate
the rate-of-fire at the soma, they are analogous to the firing rates of the hidden layer neurons. This allows
us to use the plateau potentials to define target firing rates for the neurons (see below). The network is
simulated in near continuous-time (except that each plateau is considered to be instantaneous), and the
intervals between plateaus are randomly sampled from an inverse Gaussian distribution (Figure 3D, top).
As such, the specific amount of time that the network is presented with each image and teaching signal
is stochastic, though usually somewhere between 50-60 ms of simulated time (Figure 3D, bottom). This
stochasticity was not necessary, but it demonstrates that although the system operates in phases, the
specific length of the phases is not important as long as they are sufficiently long to permit integration
(see Lemma 1). In the data presented in this paper, all 60,000 images in the MNIST training set were
presented to the network one at a time, and each exposure to the full set of images was considered an
"epoch" of training. At the end of each epoch, the network's classification error rate on a separate set of
10,000 test images was assessed with a single forward phase (see Materials and Methods).
It is important to note that there are many aspects of this design that are not physiologically accurate.
Most notably, stochastic generation of synchronized plateau potentials across a population is not an
accurate reflection of how real pyramidal neurons operate, since apical calcium spikes are determined
by a number of concrete physiological factors in individual cells, including back-propagating action
potentials, spike-timing and inhibitory inputs (Larkum et al., 1999, 2007, 2009). However, we note that
calcium spikes in the apical dendrites can be prevented from occurring via the activity of distal dendrite
targeting inhibitory interneurons (Murayama et al., 2009), which can synchronize pyramidal activity
(Hilscher et al., 2017). Furthermore, distal dendrite targeting neurons can themselves can be rapidly
inhibited in response to temporally precise neuromodulatory inputs (Pi et al., 2013; Pfeffer et al., 2013;
Karnani et al., 2016; Hangya et al., 2015; Brombas et al., 2014). Therefore, it is entirely plausible that
neocortical micro-circuits would generate synchronized pyramidal plateaus at punctuated periods of time
in response to disinhibition of the apical dendrites governed by neuromodulatory signals that determine
"phases" of processing. Alternatively, oscillations in population activity could provide a mechanism for
promoting alternating phases of processing and synaptic plasticity (Buzs´aki and Draguhn, 2004). But,
complete synchrony of plateaus in our hidden layer neurons is not actually critical to our algorithm-only
the temporal relationship between the plateaus and the teaching signal is critical (see below). This
relationship itself is arguably plausible given the role of neuromodulatory inputs in dis-inhibiting the
distal dendrites of pyramidal neurons (Karnani et al., 2016; Brombas et al., 2014). Of course, we are
engaged in a great deal of speculation here. But, the point is that our model utilizes anatomical and
functional motifs that are analogous to what is observed in the neocortex. Importantly for the present
study, the key issue is the use of segregated dendrites and distinct transmit and plateau modes to solve
the credit assignment problem.
Credit assignment with segregated dendrites
To solve the credit assignment problem without using weight transport, we had to define a local error
function for the hidden layer that somehow takes into account the impact that each hidden layer neuron
has on the output at the final layer. In other words, credit assignment could only be achieved if we knew
that changing a hidden layer synapse to reduce the local error would also help to reduce error at the
output layer. To obtain this guarantee, we defined local targets for the output and the hidden layer, i.e.
desired firing rates for both the output layer neurons and the hidden layer neurons. Learning is then
6/41
a process of changing the synaptic connections to achieve these target firing rates across the network.
Importantly, we defined our hidden layer targets in a similar manner to Lee et al. (2015). These hidden
layer targets coordinate with the output layer by incorporating the feedback information contained in
the plateau potentials. In this way, credit assignment is accomplished through mechanisms that are
spatially local to the hidden layer neurons.
Specifically, to create the output and hidden layer targets, we use the average rates-of-fire and average
voltages during different phases. We define:
y
X
= [X1
y
, ..., Xk
y
]
(cid:90) t1
(cid:90) t2
t1−tp
t2−tp
f
Xi
=
t
Xi
=
Xi(t)dt
Xi(t)dt
(4)
where y ∈ {f, t}, k ∈ {(cid:96), m, n} and Xi(t) can be a voltage variable (e.g. Ci(t) or Ui(t)) or a rate-of-
i (t)). Using the averages over different phases, we then define the target
n ], to be the average rates-of-fire during the target
fire variable (e.g. λC
rates-of-fire for the output layer, λU = [λU
phase:
i (t) or λU
1 , ..., λU
t
λU
i = λU
i
(5)
For the hidden layer we define the target rates-of-fire, λC = [λC
m], using the average rates-of-fire
during the forward phase and the difference between the plateau potentials from the forward and transmit
phase:
1 , ..., λC
f
λC
i = λC
i
+ αt
i − αf
i
(6)
The goal of learning in the hidden layer is to change the synapses W 0 to achieve these targets in
response to the given inputs.
More generally, for both the output layer and the hidden layer we then define error functions, L1 and
L0, respectively, based on the difference between the local targets and the activity of the neurons during
the forward phase (Figure 4A):
L1 = λU − λmaxσ(U
L0 = λC − λmaxσ(C
f
f
2
)2
)2
2
(7)
Note that the loss function for the hidden layer, L0, will, on average, reduce to the difference
αt − αf2
2. This provides an intuitive reason for why these targets help with credit assignment.
Specifically, a hidden layer neuron's error is small only when the output layer is providing similar
feedback to it during the forward and target phases. Put another way, hidden layer neurons "know" that
they are "doing well" when they receive the same feedback from the output layer regardless of whether
or not the teaching signal is present. Thus, hidden layer neurons have access to some information about
how they are doing in helping the output layer to achieve its targets.
More formally, it can be shown that the output and hidden layer error functions will, on average,
agree with each other, i.e. if the hidden layer error is reduced then the output layer error is also reduced.
Specifically, similar to the proof employed by Lee et al. (2015), it can be shown that if the output layer
error is sufficiently small and the synaptic matrices W 1 and Y meet some conditions, then:
7/41
λU − λmaxσ(kDW 1 λC)2
2 < λU − λmaxσ(E[U
(8)
where kD is a conductance term and E[·] denotes the expected value. (See Theorem 1 for the proof
and a concrete description of the conditions). In plain language, equation (8) says that when we consider
the difference between the output target and the activity that the hidden target would have induced at
the output layer, it is less than the difference between the output target and the expected value of the
output layer activity during the forward phase. In other words, the output layer's error would have, on
average, been smaller if the hidden layer had achieved its own target during the forward phase. Hence,
if we reduce the hidden layer error, it should also reduce the output layer error, thereby providing a
guarantee of appropriate credit assignment.
2
f
])2
With these error functions, we update the weights in the network with local gradient descent:
∆W 1 ∝ ∂L1
∂W 1
∆W 0 ∝ ∂L0
∂W 0
(9)
where ∆W i refers to the update term for weight matrix W i (see Materials and Methods, equations
(19), (20), (22) and (23) for details of the weight update procedures). Given the coordination implied by
equation (8), as the hidden layer reduces its own local error with gradient descent, the output layer's
error should also be reduced, i.e. hidden layer learning should imply output layer learning.
To test that we were successful in credit assignment for the hidden layer, and to provide empirical
support for the proof, we compared the local error at the hidden layer to the output layer error across all
of the image presentations to the network. We observed that, generally, whenever the hidden layer error
was low, the output layer error was also low. For example, when we consider the errors for the set of '2'
images presented to the network during the second epoch, there was a Pearson correlation coefficient
between L0 and L1 of r = 0.61, which was much higher than what was observed for shuffled data,
wherein output and hidden activities were randomly paired (Figure 4B). Furthermore, these correlations
were observed across all epochs of training, with most correlation coefficients for the hidden and output
errors falling between r = 0.2 - 0.6, which was, again, much higher than the correlations observed for
shuffled data (Figure 4C).
Interestingly, the correlations between L0 and L1 were smaller on the first epoch of training. This
suggests that the coordination between the layers may only come into full effect once the network has
engaged in some learning. Therefore, we inspected whether the conditions on the synaptic matrices that
are assumed in the proof were, in fact, being met. More precisely, the proof assumes that the feedforward
and feedback synaptic matrices (W 1 and Y , respectively) produce forward and backward transformations
between the output and hidden layer that are approximate inverses of each other (see Proof of Theorem
1). Since we begin learning with random matrices, this condition is almost definitely not met at the
start of training. But, we found that the network learned to meet this condition. Inspection of W 1
and Y showed that during the first epoch the forward and backwards functions became approximate
inverses of each other (Figure 4, Supplement 1). This means that during the first few image presentations
the network was actually learning to do credit assignment. This result is very similar to previous
models examining feedback alignment (Lillicrap et al., 2016), and shows that the feedback alignment
(Lillicrap et al., 2016) and difference target propagation (Lee et al., 2015) algorithms are intimately
linked. Furthermore, our results suggest that very early development may involve a period of learning
how to assign credit appropriately. Altogether, our model demonstrates that credit assignment using
random feedback weights is a general principle that can be implemented using segregated dendrites.
Deep learning with segregated dendrites
Given our finding that the network was successfully assigning credit for the output error to the hidden
layer neurons, we had reason to believe that our network with local weight-updates would exhibit deep
8/41
learning, i.e. an ability to take advantage of a multi-layer structure (Bengio and LeCun, 2007). To test
this, we examined the effects of including hidden layers. If deep learning is indeed operational in the
network, then the inclusion of hidden layers should improve the ability of the network to classify images.
We built three different versions of the network (Figure 5A). The first was a network that had no
hidden layer, i.e. the input neurons projected directly to the output neurons. The second was the
network illustrated in Figure 3A, with a single hidden layer. The third contained two hidden layers,
with the output layer projecting directly back to both hidden layers. This direct projection allowed us
to build our local targets for each hidden layer using the plateaus driven by the output layer, thereby
avoiding a "backward pass" through the entire network as has been used in other models (Lillicrap
et al., 2016; Lee et al., 2015; Liao et al., 2015). We trained each network on the 60,000 MNIST training
images for 60 epochs, and recorded the percentage of images in the 10,000 test image set that were
incorrectly classified. The network with no hidden layers rapidly learned to classify the images, but it
also rapidly hit an asymptote at an average error rate of 8.3% (Figure 5B, gray line). In contrast, the
network with one hidden layer did not exhibit a rapid convergence to an asymptote in its error rate.
Instead, it continued to improve throughout all 60 epochs, achieving an average error rate of 4.1% by
the 60th epoch (Figure 5B, blue line). Similar results were obtained when we loosened the synchrony
constraints and instead allowed each hidden layer neuron to engage in plateau potentials at different
times (Figure 5, Supplement 1). This demonstrates that strict synchrony in the plateau potentials is
not required. But, our target definitions do require two different plateau potentials separated by the
teaching signal input, which mandates some temporal control of plateau potentials in the system.
Interestingly, we found that the addition of a second hidden layer further improved learning. The
network with two hidden layers learned more rapidly than the network with one hidden layer and achieved
an average error rate of 3.2% on the test images by the 60th epoch, also without hitting a clear asymptote
in learning (Figure 5B, red line). However, it should be noted that additional hidden layers beyond
two did not significantly improve the error rate (data not shown), which suggests that our particular
algorithm could not be used to construct very deep networks as is. Nonetheless, our network was clearly
able to take advantage of multi-layer architectures to improve its learning, which is the key feature of
deep learning (Bengio and LeCun, 2007; LeCun et al., 2015).
Another key feature of deep learning is the ability to generate representations in the higher layers of
a network that capture task-relevant information while discarding sensory details (LeCun et al., 2015;
Mnih et al., 2015). To examine whether our network exhibited this type of abstraction, we used the
t-Distributed Stochastic Neighbor Embedding algorithm (t-SNE). The t-SNE algorithm reduces the
dimensionality of data while preserving local structure and non-linear manifolds that exist in high-
dimensional space, thereby allowing accurate visualization of the structure of high-dimensional data
(Maaten and Hinton, 2008). We applied t-SNE to the activity patterns at each layer of the two hidden
layer network for all of the images in the test set after 60 epochs of training. At the input level, there
was already some clustering of images based on their categories. However, the clusters were quite messy,
with different categories showing outliers, several clusters, or merged clusters (Figure 5C, bottom). For
example, the '2' digits in the input layer exhibited two distinct clusters separated by a cluster of '7's:
one cluster contained '2's with a loop and one contained '2's without a loop. Similarly, there were
two distinct clusters of '4's and '9's that were very close to each other, with one pair for digits on a
pronounced slant and one for straight digits (Figure 5C, bottom, example images). Thus, although there
is built-in structure to the categories of the MNIST dataset, there are a number of low-level features
that do not respect category boundaries. In contrast, at the first hidden layer, the activity patterns were
much cleaner, with far fewer outliers and split/merged clusters (Figure 5C, middle). For example, the
two separate '2' digit clusters were much closer to each other and were now only separated by a very
small cluster of '7's. Likewise, the '9' and '4' clusters were now distinct and no longer split based on the
slant of the digit. Interestingly, when we examined the activity patterns at the second hidden layer the
categories were even better segregated with only a bit of splitting or merging of category clusters (Figure
5C, top). Therefore, the network had learned to develop representations in the hidden layers wherein the
categories were very distinct and low-level features unrelated to the categories were largely ignored. This
abstract representation is likely to be key to the improved error rate in the two hidden layer network.
9/41
Altogether, our data demonstrates that our network with segregated dendritic compartments can engage
in deep learning.
Coordinated local learning mimics backpropagation of error
The backpropagation of error algorithm (Rumelhart et al., 1986) is still the primary learning algorithm
used for deep supervised learning in artificial neural networks (LeCun et al., 2015). Previous work has
shown that learning with random feedback weights can actually match the synaptic weight updates
specified by the backpropagation algorithm after a few epochs of training (Lillicrap et al., 2016). This
fascinating observation suggests that deep learning with random feedback weights is not so much a
different algorithm than backpropagation of error, but rather, networks with random feedback connections
learn to approximate credit assignment as it is done in backpropagation (Lillicrap et al., 2016). Hence, we
were curious as to whether or not our network was, in fact, learning to approximate the synaptic weight
updates prescribed by backpropagation. To test this, we trained our one hidden layer network as before,
but now, in addition to calculating the vector of hidden layer synaptic weight updates specified by our
local learning rule (∆W 0 in equation (9)), we also calculated the vector of hidden layer synaptic weight
updates that would be specified by non-locally backpropagating the error from the output layer, (∆W 0
BP ).
We then calculated the angle between these two alternative weight updates. In a very high-dimensional
BP ≈ 90◦).
space, any two independent vectors will be roughly orthogonal to each other (i.e. ∆W 0(cid:54) ∆W 0
BP < 90◦),
If the two synaptic weight update vectors are not orthogonal to each other (i.e. ∆W 0(cid:54) ∆W 0
then it suggests that the two algorithms are specifying similar weight updates.
As in previous work (Lillicrap et al., 2016), we found that the initial weight updates for our network
were orthogonal to the updates specified by backpropagation. But, as the network learned the angle
dropped to approximately 65◦, before rising again slightly to roughly 70◦ (Figure 6A, blue line). This
suggests that our network was learning to develop local weight updates in the hidden layer that were in
rough agreement with the updates that explicit backpropagation would produce. However, this drop in
orthogonality was still much less than that observed in non-spiking artificial neural networks learning
with random feedback weights, which show a drop to below 45◦ (Lillicrap et al., 2016). We suspected
that the higher angle between the weight updates that we observed may have been because we were
using spikes to communicate the feedback from the upper layer, which could introduce both noise and
bias in the estimates of the output layer activity. To test this, we also examined the weight updates that
our algorithm would produce if we propagated the spike rates of the output layer neurons, λU (t), back
directly through the random feedback weights, Y . In this scenario, we observed a much sharper drop in
BP angle, which reduced to roughly 35◦ before rising again to 40◦ (Figure 6A, red line).
the ∆W 0(cid:54) ∆W 0
These results show that, in principle, our algorithm is learning to approximate the backpropagation
algorithm, though with some drop in accuracy introduced by the use of spikes to propagate output layer
activities to the hidden layer.
To further examine how our local learning algorithm compared to backpropagation we compared
the low-level features that the two algorithms learned. To do this, we trained the one hidden layer
network with both our algorithm and backpropagation. We then examined the receptive fields (i.e. the
synaptic weights) produced by both algorithms in the hidden layer synapses (W 0) after 60 epochs of
training. The two algorithms produced qualitatively similar receptive fields (Figure 6C). Both produced
receptive fields with clear, high-contrast features for detecting particular strokes or shapes. To quantify
the similarity, we conducted pair-wise correlation calculations for the receptive fields produced by the
two algorithms and identified the maximum correlation pairs for each. Compared to shuffled versions of
the receptive fields, there was a very high level of maximum correlation (Figure 6B), showing that the
receptive fields were indeed quite similar. Thus, the data demonstrate that our learning algorithm using
random feedback weights into segregated dendrites can in fact come to approximate the backpropagation
of error algorithm.
10/41
Conditions on feedback weights
Once we had convinced ourselves that our learning algorithm was, in fact, producing deep learning similar
to that produced by backpropagation of error, we wanted to examine some of the constraints on learning.
First, we wanted to explore the structure of the feedback weights. In our initial simulations we used
non-sparse, random (i.e. normally distributed) feedback weights. We were interested in whether learning
could still work with sparse weights, given that neocortical connectivity is sparse. As well, we wondered
whether symmetric weights would improve learning, which would be expected given previous findings
(Lillicrap et al., 2016; Lee et al., 2015; Liao et al., 2015). To explore these questions, we trained our one
hidden layer network using both sparse feedback weights (only 20% non-zero values) and symmetric
weights (Y = W 1T ) (Figure 7A,C). We found that learning actually improved slightly with sparse
weights (Figure 7B, red line), achieving an average error rate of 3.7% by the 60th epoch, compared to
the average 4.1% error rate achieved with fully random weights. But, this result appeared to depend
on the magnitude of the sparse weights. To compensate for the loss of 80% of the weights we initially
increased the sparse synaptic weight magnitudes by a factor of 5. However, when we did not re-scale the
sparse weights learning was actually worse (Figure 7, Supplement 1). This suggests that sparse feedback
provides a signal that is sufficient for credit assignment, but only if it is of appropriate magnitude.
Similar to sparse feedback weights, symmetric feedback weights also improved learning, leading to
a rapid decrease in the test error and an error rate of 3.6% by the 60th epoch (Figure 7D, red line).
However, when we added noise to the symmetric weights this advantage was eliminated and learning was,
in fact, slightly impaired (Figure 7D, blue line). At first, this was a very surprising result: given that
learning works with random feedback weights, why would it not work with symmetric weights with noise?
However, when we considered our previous finding that during the first epoch the feedforward weights,
W 1, learn to match the inverse of the feedback function (Figure 4, Supplement 1) a possible answer
becomes clear. In the case of symmetric feedback weights the synaptic matrix Y is changing as W 1
changes. This works fine when Y is set to W 1T
, since that artificially forces weight alignment. But, if the
feedback weights are set to W 1T
plus noise, then the system can never align the weights appropriately,
since Y is now a moving target. This would imply that any implementation of feedback learning must
either be very effective (to achieve the right feedback) or very slow (to allow the feedforward weights to
adapt).
Learning with partial apical attenuation
Another constraint that we wished to examine was whether total segregation of the apical inputs as we
had done was necessary, given that real pyramidal neurons only show an attenuation of distal apical
inputs to the soma (Larkum et al., 1999). To examine this, we re-ran our two hidden layer network, but
now, we allowed the apical dendritic voltage to influence the somatic voltage by setting gA = 0.05. This
value gave us twelve times more attenuation than the attenuation from the basal compartments (since
gB = 0.6). This difference in the levels of attenuation is in-line with experimental data (Larkum et al.,
1999, 2009) (Figure 8A). When we compared the learning in this scenario to the scenario with total apical
segregation, we observed very little difference in the error rates on the test set (Figure 8B, gray and red
lines). Importantly, though, we found that if we decreased the apical attenuation to the same level as
the basal compartment (gA = gB = 0.6) then the learning was significantly impaired (Figure 8B, blue
line). This demonstrates that although total apical attenuation is not necessary, partial segregation of
the apical compartment from the soma is necessary. This result makes sense given that our local targets
for the hidden layer neurons incorporate a term that is supposed to reflect the response of the output
neurons to the feedforward sensory information (αf ). Without some sort of separation of feedforward
and feedback information, as is assumed in other models of deep learning (Lillicrap et al., 2016; Lee
et al., 2015), this feedback signal would get corrupted by recurrent dynamics in the network. Our data
show that electrontonically segregated dendrites is one potential way to achieve the required separation
between feedforward and feedback information.
11/41
Discussion
Deep learning has radically altered the field of AI, demonstrating that parallel distributed processing
across multiple layers can produce human/animal-level capabilities in image classification, pattern
recognition and reinforcement learning (Hinton et al., 2006; LeCun et al., 2015; Mnih et al., 2015; Silver
et al., 2016; Krizhevsky et al., 2012; He et al., 2015). Deep learning was motivated by analogies to the real
brain (LeCun et al., 2015; Cox and Dean, 2014), so it is tantalizing that recent studies have shown that
deep neural networks develop representations that strongly resemble the representations observed in the
mammalian neocortex (Khaligh-Razavi and Kriegeskorte, 2014; Yamins and DiCarlo, 2016; Cadieu et al.,
2014; Kubilius et al., 2016). In fact, deep learning models can match cortical representations even more
so than some models that explicitly attempt to mimic the real brain (Khaligh-Razavi and Kriegeskorte,
2014). Hence, at a phenomenological level, it appears that deep learning, defined as multilayer cost
function reduction with appropriate credit assignment, may be key to the remarkable computational
prowess of the mammalian brain (Marblestone et al., 2016). However, the lack of biologically feasible
mechanisms for credit assignment in deep learning algorithms, most notably backpropagation of error
(Rumelhart et al., 1986), has left neuroscientists with a mystery. How can the real brain solve the
credit assignment problem (Figure 1)? Here, we expanded on an idea that previous authors have
explored (Kording and Konig, 2001; Spratling, 2002; Spratling and Johnson, 2006) and demonstrated that
segregating the feedback and feedforward inputs to neurons, much as the real neocortex does (Larkum
et al., 1999, 2007, 2009), can enable the construction of local targets to assign credit appropriately to
hidden layer neurons (Figure 2). With this formulation, we showed that we could use segregated dendritic
compartments to coordinate learning across layers (Figure 3 and Figure 4). This enabled our network to
take advantage of multiple layers to develop representations of hand-written digits in hidden layers that
enabled better levels of classification accuracy on the MNIST dataset than could be achieved with a
single layer (Figure 5). Furthermore, we found that our algorithm actually approximated the weight
updates that would be prescribed by backpropagation, and produced similar low-level feature detectors
(Figure 6). As well, we showed that our basic framework works with sparse feedback connections (Figure
7) and more realistic, partial apical attenuation (Figure 8). Therefore, our work demonstrates that deep
learning is possible in a biologically feasible framework, provided that feedforward and feedback signals
are sufficiently segregated in different dendrites.
Perhaps the most biologically unrealistic component of backpropagation is the use of non-local
"weight-transport" (Figure 1B) (Grossberg, 1987). Our model builds on recent neural networks research
demonstrating that weight transport is not, in fact, required for credit assignment (Lillicrap et al., 2016;
Liao et al., 2015; Lee et al., 2015). By demonstrating that the credit assignment problem is solvable using
feedback to generate spatially local weight updates, these studies have provided a major breakthrough in
our understanding of how deep learning could work in the real brain. However, this previous research
involved an implicit assumption of separate feedforward and feedback pathways (Figure 2A). Although
this is a possibility, there is currently no evidence for separate feedback pathways to guide learning in the
neocortex. In our study, we obtained separate feedforward and feedback information using electrotonically
segregated dendrites (Figure 2B), in analogy to neocortical pyramidal neurons (Larkum et al., 1999).
This segregation allowed us to perform credit assignment using direct feedback pathways between layers
(Figures 3-5). Moreover, we found that even with partial attenuation of the conductance from the apical
dendrites to the soma our network could engage in deep learning (Figure 8). It should be recognized,
though, that although our learning algorithm achieved deep learning with spatially local update rules,
we had to assume some temporal non-locality. Our credit assignment system depended on the difference
between the target and forward plateaus, αt − αf , which occurred at different times (roughly a 50-60
ms gap). Although this is not an ideal solution, this small degree of temporal non-locality for synaptic
update rules is in-line with known biological mechanisms, such as slowly decaying calcium transients or
synaptic tags (Redondo and Morris, 2011). Hence, our model exhibited deep learning using only local
information contained within the cells.
In this work we adopted a similar strategy to the one taken by Lee et al.'s (2015) difference target
propagation algorithm, wherein the feedback from higher layers is used to construct local activity targets
12/41
at the hidden layers. One of the reasons that we adopted this strategy is that it is appealing to think that
feedback from upper layers may not simply be providing a signal for plasticity, but also a modulatory
signal to push the hidden layer neurons towards a "better" activity pattern in real-time. This sort of
top-down modulation could be used by the brain to improve sensory processing in different contexts and
engage in inference (Bengio et al., 2015). Indeed, framing cortico-cortical feedback as a mechanism to
modulate incoming sensory activity is a more common way of viewing feedback signals in the neocortex
(Larkum, 2013; Gilbert and Li, 2013; Zhang et al., 2014; Fiser et al., 2016). In light of this, it is interesting
to note that distal apical inputs in somatosensory cortex can help animals perform sensory discrimination
tasks (Takahashi et al., 2016; Manita et al., 2015). However, in our model, we did not actually implement
a system that altered the hidden layer activity to make sensory calculations-we simply used the feedback
signals to drive learning. In-line with this view of top-down feedback, two recent papers have found
evidence that cortical feedback can indeed guide feedforward sensory plasticity (Thompson et al., 2016;
Yamada et al., 2017). Yet, there is no reason that feedback signals cannot provide both top-down
modulation and a signal for learning (Spratling, 2002). In this respect, a potential future advance on our
model would be to implement a system wherein the feedback actively "nudges" the hidden layers towards
appropriate activity patterns in order to guide learning while also shaping perception. This proposal
is reminiscent of the approach taken in previous computational models (Urbanczik and Senn, 2014;
Spratling and Johnson, 2006; Kording and Konig, 2001). Future research could study how top-down
modulation and a signal for credit assignment can be combined in deep learning models.
It is important to note that there are many aspects of our model that are not biologically realistic.
These include (but are not limited to) synaptic inputs that are not conductance based, synaptic weights
that can switch from positive to negative (and vice-versa), and plateau potentials that are considered
instantaneous and are not an accurate reflection of calcium spike dynamics in real pyramidal neurons
(Larkum et al., 1999, 2009). However, the intention with this study was not to achieve total biological
realism, but rather to demonstrate that the sort of dendritic segregation that neocortical pyramidal
neurons exhibit can subserve credit assignment. Although we found that deep learning can still work
with realistic levels of passive conductance from the apical compartment to the somatic compartment
(Figure 8B, red line) (Larkum et al., 1999), we also found that learning was impaired if the basal and
apical compartments had equal conductance to the soma (Figure 8B, blue line). This is interesting,
because it suggests that the unique physiology of neocortical pyramidal neurons may in fact reflect
nature's solution to deep learning. Perhaps the relegation of feedback from higher-order brain regions to
the electrotonically distant apical dendrites (Budd, 1998; Spratling, 2002), and the presence of active
calcium spike mechanisms in the apical shaft (Larkum et al., 2009), are mechanisms for coordinating local
synaptic weight updates. If this is correct, then the inhibitory interneurons that target apical dendrites
and limit active communication to the soma (Murayama et al., 2009) may be used by the neocortex
to control learning. Although this is speculative, it is worth noting that current evidence supports the
idea that neuromodulatory inputs carrying temporally precise salience information (Hangya et al., 2015)
can shut off interneurons to disinhibit the distal apical dendrites (Pi et al., 2013; Karnani et al., 2016;
Pfeffer et al., 2013; Brombas et al., 2014), and presumably, promote apical communication to the soma.
Recent work suggests that the specific patterns of interneuron inhibition on the apical dendrites are very
spatially precise and differentially timed to motor behaviours (Munoz et al., 2017), which suggests that
there may well be coordinated physiological mechanisms for determining when cortico-cortical feedback is
transmitted to the soma. Future research should examine whether these inhibitory and neuromodulatory
mechanisms do, in fact, open the window for plasticity in the basal dendrites of pyramidal neurons, as
our model, and some recent experimental work (Bittner et al., 2015), would predict.
An additional issue that should be recognized is that the error rates which our network achieved were
by no means as low as can be achieved with artificial neural networks, nor at human levels of performance
(LeCun et al., 1998; Li et al., 2016). As well, our algorithm was not able to take advantage of very deep
structures (beyond two hidden layers, the error rate did not improve). In contrast, increasing the depth
of networks trained with backpropagation can lead to performance improvements (Li et al., 2016). But,
these observations do not mean that our network was not engaged in deep learning. First, it is interesting
to note that although the backpropagation algorithm is several decades old (Rumelhart et al., 1986), it
13/41
was long considered to be useless for training networks with more than one or two hidden layers (Bengio
and LeCun, 2007). Indeed, it was only the use of layer-by-layer training that initially led to the realization
that deeper networks can achieve excellent performance (Hinton et al., 2006). Since then, both the use
of very large datasets (with millions of examples), and additional modifications to the backpropagation
algorithm, have been key to making backpropagation work well on deeper networks (Sutskever et al.,
2013; LeCun et al., 2015). Future studies could examine how our algorithm could incorporate current
techniques used in machine learning to work better on deeper architectures. Second, we stress that our
network was not designed to match the state-of-the-art in machine learning, nor human capabilities.
To test our basic hypothesis (and to run our leaky-integration and spiking simulations in a reasonable
amount of time) we kept the network small, we stopped training before it reached its asymptote, and
we did not implement any add-ons to the learning to improve the error rates, such as convolution and
pooling layers, initialization tricks, mini-batch training, drop-out, momentum or RMSProp (Sutskever
et al., 2013; Tieleman and Hinton, 2012; Srivastava et al., 2014). Indeed, it would be quite surprising if a
relatively vanilla, small network like ours could come close to matching current performance benchmarks
in machine learning. Third, although our network was able to take advantage of multiple layers to
improve the error rate, there may be a variety of reasons that ever increasing depth didn't improve
performance significantly. For example, our use of direct connections from the output layer to the hidden
layers may have impaired the network's ability to coordinate synaptic updates between hidden layers.
As well, given our finding that the use of spikes produced weight updates that were less well-aligned to
backpropagation (Figure 6A) it is possible that deeper architectures require mechanisms to overcome the
inherent noisiness of spikes.
One aspect of our model that we did not develop was the potential for learning at the feedback
synapses. Although we used random synaptic weights for feedback, we also demonstrated that our
model actually learns to meet the mathematical conditions required for credit assignment, namely the
feedforward weights come to approximate the inverse of the feedback (Figure 4, Supplement 1). This
suggests that it would be beneficial to develop a synaptic weight update rule for the feedback synapses
that made this aspect of the learning better. Indeed, Lee et al. (2015) implemented an "inverse loss
function" for their feedback synapses which promoted the development of feedforward and feedback
functions that were roughly inverses of each other, leading to the emergence of auto-encoder functions in
their network. In light of this, it is interesting to note that there is evidence for unique, "reverse" spike-
timing-dependent synaptic plasticity rules in the distal apical dendrites of pyramidal neurons (Sjostrom
and Hausser, 2006; Letzkus et al., 2006), which have been shown to produce symmetric feedback weights
and auto-encoder functions in artificial spiking networks (Burbank and Kreiman, 2012; Burbank, 2015).
Thus, it is possible that early in development the neocortex actually learns cortico-cortical feedback
connections that help it to assign credit for later learning. Our work suggests that any experimental
evidence showing that feedback connections learn to approximate the inverse of feedforward connections
could be considered as evidence for deep learning in the neocortex.
In summary, deep learning has had a huge impact on AI, but, to date, its impact on neuroscience has
been limited. Nonetheless, given a number of findings in neurophysiology and modeling (Yamins and
DiCarlo, 2016), there is growing interest in understanding how deep learning may actually be achieved by
the real brain (Marblestone et al., 2016). Our results show that by moving away from point neurons, and
shifting towards multi-compartment neurons that segregate feedforward and feedback signals, the credit
assignment problem can be solved and deep learning can be achieved. Perhaps the dendritic anatomy of
neocortical pyramidal neurons is important for nature's own deep learning algorithm.
Materials and Methods
Code for the model can be obtained from a GitHub repository (https://github.com/jordan-g/Segregated-
Dendrite-Deep-Learning) (Guergiuev, 2017). For notational simplicity, we describe our model in the case
of a network with only one hidden layer. We describe how this is extended to a network with multiple
layers at the end of this section. As well, at the end of this section in Table 1 we provide a table listing
14/41
the parameter values we used for all of the simulations presented in this paper.
Neuronal dynamics
The network described here consists of an input layer with (cid:96) neurons, a hidden layer with m neurons,
and an output layer with n neurons. Neurons in the input layer are simple Poisson spiking neurons
whose rate-of-fire is determined by the intensity of image pixels (ranging from 0 - λmax). Neurons in the
hidden layer are modeled using three functional compartments-basal dendrites with voltages B(t) =
[B1(t), B2(t), ..., Bm(t)], apical dendrites with voltages A(t) = [A1(t), A2(t), ..., Am(t)], and somata with
voltages C(t) = [C1(t), C2(t), ..., Cm(t)]. Feedforward inputs from the input layer and feedback inputs
from the output layer arrive at basal and apical synapses, respectively. At basal synapses, presynaptic
spikes are translated into post-synaptic potentials P SP B(t) = [P SP B
(cid:96) (t)] given
by:
2 (t), ..., P SP B
1 (t), P SP B
P SP B
j (t) =
κ(t − s)
(10)
(cid:88)
s∈Xj
(cid:96)(cid:88)
n(cid:88)
j=1
where Xj is the set of pre-synaptic spike times at synapses with presynaptic input neuron j.
κ(t) = (e−t/τL − e−t/τs)Θ(t)/(τL − τs) is the response kernel, where τs and τL are short and long
time constants, and Θ is the Heaviside step function. The post-synaptic potentials at apical synapses,
P SP A(t) = [P SP A
n (t)], are modeled in the same manner. The basal and apical
dendritic potentials for neuron i are then given by weighted sums of the post-synaptic potentials at
either its basal or apical synapses:
2 (t), ..., P SP A
1 (t), P SP A
Bi(t) =
Ai(t) =
W 0
ijP SP B
j (t) + b0
i
YijP SP A
j (t)
(11)
where b0 = [b0
m] are bias terms, W 0 is the m × (cid:96) matrix of feedforward weights for neurons
in the hidden layer, and Y is the m× n matrix of their feedback weights. The somatic voltage for neuron
i evolves with leak as:
2, ..., b0
1, b0
j=1
dCi(t)
dt
= −gLCi(t) + gB(Bi(t) − Ci(t)) + gA(Ai(t) − Ci(t))
(12)
where gL is the leak conductance, gB is the conductance from the basal dendrite to the soma, and
gA is the conductance from the apical dendrite to the soma. Equation (12) is identical to equation (1)
in Results. Note that for simplicity's sake we are assuming a resting potential of 0 V and a membrane
capacitance of 1 F, but these values are not important for the results.
The instantaneous firing rates of neurons in the hidden layer are given by λC(t) = [λC
2 (t), ..., λC
m(t)],
i (t) is the result of applying a nonlinearity, σ(·), to the somatic potential Ci(t). We chose σ(·)
1 (t), λC
where λC
to be a simple sigmoidal function, such that:
λC
i (t) = λmaxσ(Ci(t)) = λmax
1
1 + e−Ci(t)
(13)
Here, λmax is the maximum possible rate-of-fire for the neurons, which we set to 200 Hz. Note that
equation (13) is identical to equation (2) in Results. Spikes are then generated using Poisson processes
with these firing rates. We note that although the maximum rate was 200 Hz, the neurons rarely achieved
15/41
m(cid:88)
anything close to this rate, and the average rate of fire in the neurons during our simulations was 24 Hz,
in-line with neocortical firing rates (Steriade et al., 2001).
Units in the output layer are modeled using only two compartments, dendrites with voltages
V (t) = [V1(t), V2(t), ..., Vn(t)] and somata with voltages U (t) = [U1(t), U2(t), ..., Un(t)]. Vi(t) is given by:
Vi(t) =
W 1
ijP SP V
j (t) + b1
i
(14)
j=1
where P SP V (t) = [P SP V
m (t)] are the post-synaptic potentials at synapses
that receive feedforward input from the hidden layer, and are calculated in the manner described by
equation (10). Ui(t) evolves as:
2 (t), ..., P SP V
1 (t), P SP V
dUi(t)
dt
= −gLUi(t) + gD(Vi(t) − Ui(t)) + Ii(t)
(15)
where gL is the leak conductance, gD is the conductance from the dendrite to the soma, and
I(t) = [I1(t), I2(t), ..., In(t)] are somatic currents that can drive output neurons toward a desired somatic
voltage. For neuron i, Ii is given by:
Ii(t) = gEi(t)(EE − Ui(t)) + gIi(t)(EI − Ui(t))
(16)
where gE(t) = [gE1 (t), gE2(t), ..., gEn (t)] and gI (t) = [gI1 (t), gI2(t), ..., gIn (t)] are time-varying exci-
tatory & inhibitory nudging conductances, and EE and EI are the excitatory and inhibitory reversal
potentials. In our simulations, we set EE = 8 V and EI = −8 V. During the target phase only, we set
gIi = 1 and gEi = 0 for all units i whose output should be minimal, and gEi = 1 and gIi = 0 for the unit
whose output should be maximal. In this way, all units other than the "target" unit are silenced, while
the "target" unit receives a strong excitatory drive. In the forward phase, I(t) is set to 0. The Poisson
spike rates λU (t) = [λU
n (t)] are calculated as in equation (13).
2 (t), ..., λU
1 (t), λU
Plateau potentials
At the end of the forward and target phases, we calculate plateau potentials αf = [αf
αt = [αt
m] for apical dendrites of hidden layer neurons, where αf
2, ..., αt
i and αt
1, αt
1 , αf
2 , ..., αf
i are given by:
m] and
(cid:90) t1
(cid:90) t2
t1−tp
t2−tp
αf
i = σ(
αt
i = σ(
Ai(t)dt)
Ai(t)dt)
(17)
where t1 and t2 are the end times of the forward and target phases, respectively, and tp = ti+1 − (ti +
∆ts) is the integration time for the plateau potential. ∆ts = 30 ms is the settling time for the voltages.
Note that equation (17) is identical to equation (3) in Results. These plateau potentials are used by
hidden layer neurons to update their basal weights.
Weight updates
All feedforward synaptic weights are updated at the end of each target phase. Output layer units update
their synaptic weights W 1 in order to minimize the loss function L1 given in equation (7). All average
voltages are calculated after a delay ∆ts from the start of a phase, which allows for the network to
16/41
reach a steady state before averaging begins. In practice this means that the average somatic voltage for
output layer neuron i in the forward phase, Ui
, has the property
f
(cid:0) m(cid:88)
(cid:1)
f ≈ kDVi
f
Ui
= kD
W 1
ijP SP V
j
f
+ b1
i
where kD = gD/(gL + gD). Thus,
j=1
∂L1
∂W 1 ≈ −kDλmax(λU − λmaxσ(U
∂b1 ≈ −kDλmax(λU − λmaxσ(U
∂L1
f
))σ(cid:48)(U
f
) ◦ P SP V
f
f
))σ(cid:48)(U
f
)
(18)
(19)
The dendrites in the output layer use this approximation of the gradient in order to update their
weights using gradient descent:
W 1 → W 1 − η1P 1 ∂L1
∂W 1
b1 → b1 − η1P 1 ∂L1
∂b1
(20)
where η1 is a learning rate constant, and P 1 is a scaling factor used to normalize the scale of the
rate-of-fire function.
In the hidden layer, basal dendrites update their synaptic weights W 0 by minimizing the loss function
L0 given in equation (7). We define the target rates-of-fire λC = [λC
1 , λC
2 , ..., λC
m] such that
f
λC
i = λC
i
+ αt
i − αf
i
(21)
1 , αf
where αf = [αf
2 , ..., αf
m] and αt = [αt
1, αt
2, ..., αt
m] are forward and target phase plateau potentials
given in equation (17). Note that equation (21) is identical to equation (6) in Results. These hidden
layer target firing rates are similar to the targets used in difference target propagation (Lee et al., 2015).
Using the fact that λC
i
i − λmaxσ(Ci
), we can say that λC
f ≈ λmaxσ(Ci
i , and hence:
i − αf
) ≈ αt
f
f
∂L0
∂W 0 ≈ −kB(αt − αf )λmaxσ(cid:48)(C
∂b0 ≈ −kB(αt − αf )λmaxσ(cid:48)(C
∂L0
f
) ◦ P SP B
f
f
)
where kB = gB/(gL + gB + gA). Basal weights are updated in order to descend this gradient:
W 0 → W 0 − η0P 0 ∂L0
∂W 0
b0 → b0 − η0P 0 ∂L0
∂b0
(22)
(23)
Importantly, this update rule is fully local for the hidden layer neurons. It consists essentially of three
terms, (1) the difference in the plateau potentials for the target and forward phases (αt − αf ), (2) the
derivative of the spike rate function (λmaxσ(cid:48)(C
). All
three of these terms are values that a real neuron could theoretically calculate using some combination
of molecular synaptic tags, calcium currents, and back-propagating action potentials.
)), and (3) the postsynaptic potentials (P SP B
f
f
17/41
Multiple hidden layers
In order to extend our algorithm to deeper networks with multiple hidden layers, our model incorporates
direct synaptic connections from the output layer to each hidden layer. Thus, each hidden layer receives
feedback from the output layer through its own separate set of fixed, random weights. For example, in a
network with two hidden layers, both layers receive the feedback from the output layer at their apical
dendrites through backward weights Y 0 and Y 1. The local targets at each layer are then given by:
t
λU = λU
λC1
λC0
= λC1 t
= λC0 t
+ α1t − α1f
+ α0t − α0f
where the superscripts 0 and 1 denote the first and second hidden layers, respectively.
The local loss functions at each layer are:
L2 = λU − λmaxσ(U
L1 = λC1 − λmaxσ(C1
L0 = λC0 − λmaxσ(C0
)2
)2
)2
2
f
f
f
2
2
(24)
(25)
(26)
(27)
where L2 is the loss at the output layer. The learning rules used by the hidden layers in this scenario
are the same as in the case with one hidden layer.
Learning rate optimization
For each of the three network sizes that we present in this paper, a grid search was performed in order to
find good learning rates. We set the learning rate for each layer by stepping through the range [0.1, 0.3]
with a step size of 0.02. For each combination of learning rates, a neural network was trained for one
epoch on the 60, 000 training examples, after which the network was tested on 10,000 test images. The
learning rates that gave the best performance on the test set after an epoch of training were used as a
basis for a second grid search around these learning rates that used a smaller step size of 0.01. From
this, the learning rates that gave the best test performance after 20 epochs were chosen as our learning
rates for that network size.
In all of our simulations, we used a learning rate of 0.19 for a network with no hidden layers, learning
rates of 0.21 (output and hidden) for a network with one hidden layer, and learning rates of 0.23 (hidden
layers) and 0.12 (output layer) for a network with two hidden layers. All networks with one hidden layer
had 500 hidden layer neurons, and all networks with two hidden layers had 500 neurons in the first
hidden layer and 100 neurons in the second hidden layer.
Training paradigm
For all simulations described in this paper, the neural networks were trained on classifying handwritten
digits using the MNIST database of 28 pixel × 28 pixel images. Initial feedforward and feedback weights
were chosen randomly from a uniform distribution over a range that was calculated to produce voltages
in the dendrites between −6 - 12 V.
Prior to training, we tested a network's initial performance on a set of 10,000 test examples. This
set of images was shuffled at the beginning of testing, and each example was shown to the network
in sequence. Each input image was encoded into Poisson spiking activity of the 784 input neurons
representing each pixel of the image. The firing rate of an input neuron was proportional to the brightness
of the pixel that it represents (with spike rates between [0 - λmax]. The spiking activity of each of the
18/41
784 input neurons was received by the neurons in the first hidden layer. For each test image, the network
underwent only a forward phase. At the end of this phase, the network's classification of the input
image was given by the neuron in the output layer with the greatest somatic potential (and therefore
the greatest spike rate). The network's classification was compared to the target classification. After
classifying all 10,000 testing examples, the network's classification error was given by the percentage of
examples that it did not classify correctly.
Following the initial test, training of the neural network was done in an on-line fashion. All 60,000
training images were randomly shuffled at the start of each training epoch. The network was then shown
each training image in sequence, undergoing a forward phase ending with a plateau potential, and a
target phase ending with another plateau potential. All feedforward weights were then updated at the
end of the target phase. At the end of the epoch (after all 60,000 images were shown to the network),
the network was again tested on the 10,000 test examples. The network was trained for up to 60 epochs.
Simulation details
For each training example, a minimum length of 50 ms was used for each of the forward and target
phases. The lengths of the forward and target training phases were determined by adding their minimum
length to an extra length term, which was chosen randomly from a Wald distribution with a mean of 2
ms and scale factor of 1. During testing, a fixed length of 500 ms was used for the forward transmit
phase. Average forward and target phase voltages were calculated after a settle duration of ∆ts = 30 ms
from the start of the phase.
For simulations with randomly sampled plateau potential times (Figure 5, Supplement 1), the time at
which each neuron's plateau potential occurred was randomly sampled from a folded normal distribution
(µ = 0, σ2 = 3) that was truncated (max = 5) such that plateau potentials occurred between 0 ms and 5
ms before the start of the next phase. In this scenario, the average apical voltage in the last 30 ms was
averaged in the calculation of the plateau potential for a particular neuron.
The time-step used for simulations was dt = 1 ms. At each time-step, the network's state was updated
bottom-to-top beginning with the first hidden layer and ending with the output layer. For each layer,
dendritic potentials were updated, followed by somatic potentials, and finally their spiking activity. Table
1 lists the simulation parameters and the values that were used in the figures presented.
All code was written using the Python programming language version 2.7 (RRID: SCR 008394) with
the NumPy (RRID: SCR 008633) and SciPy (RRID: SCR 008058) libraries. The code is open source
and is freely available at https://github.com/jordan-g/Segregated-Dendrite-Deep-Learning (Guergiuev,
2017). The data used to train the network was from the Mixed National Institute of Standards and
Technology (MNIST) database, which is a modification of the original database from the National
Institute of Standards and Technology (RRID: SCR 006440) (LeCun et al., 1998). The MNIST database
can be found at http://yann.lecun.com/exdb/mnist/. Some of the simulations were run on the SciNet
High-Performance Computing platform (Loken et al., 2010).
Proofs
Theorem for loss function coordination
The targets that we selected for the hidden layer (see equation (6)) were based on the targets used in
Lee et al. (2015). The authors of that paper provided a proof showing that their hidden layer targets
guaranteed that learning in one layer helped reduce the error in the next layer. However, there were
a number of differences between our network and theirs, such as the use of spiking neurons, voltages,
different compartments, etc. Here, we modify the original Lee et al. (2015) proof slightly to prove
Theorem 1 (which is the formal statement of equation (8)).
One important thing to note is that the theorem given here utilizes a target for the hidden layer that
is slightly different than the one defined in equation (6). However, the target defined in equation (6) is a
19/41
t
f
+ σ(Y λU
numerical approximation of the target given in Theorem 1. After the proof of we describe exactly how
these approximations relate to the targets given here.
Theorem 1. Consider a neural network with one hidden layer and an output layer. Let λC∗
λC
σ(·) is a differentiable function. Assume that U
be the target firing rates
for the output layer. Also, for notational simplicity, let β(x) ≡ λmaxσ(kDW 1x) and γ(x) ≡ σ(Y x).
Theorem 1 states that if λU − λmaxσ(E[U
]) is sufficiently small, and the Jacobian matrices Jβ and Jγ
satisfy the condition that the largest eigenvalue of (I − JβJγ)T (I − JβJγ) is less than 1, then
])) be the target firing rates for neurons in the hidden layer, where
) − σ(Y λmaxσ(E[U
. Let λU = λU
f ≈ kDV
=
f
f
f
t
λU − λmaxσ(kDW 1 λC∗
)2
2 < λU − λmaxσ(E[U
f
])2
2
We note again that the proof for this theorem is essentially a modification of the proof provided in
f
Lee et al. (2015) that incorporates our Lemma 1 to take into account the expected value of P SP V
given that spikes in the network are generated with non-stationary Poisson processes.
,
Proof.
λU − λmaxσ(kDW 1 λC∗
) ≡ λU − β(λC∗
= λU − β(λC
f
)
t
+ γ(λU
Lemma 1 shows that λmaxσ(E[U
ficiently large averaging time window. Assume that λmaxσ(E[U
Then,
]) = λmaxσ(E[kDW 1P SP V
f
f
) − γ(λmaxσ(E[U
])))
]) ≈ λmaxσ(kDW 1λC
]) = λmaxσ(kDW 1λC
f
f
f
) given a suf-
) ≡ β(λC
f
).
f
λU − β(λC∗
) = λU − β(λC
f
+ γ(λU
t
) − γ(β(λC
f
)))
Let e = λU
t − β(λC
f
). Applying Taylor's theorem,
λU − β(λC∗
) = λU − β(λC
f
+ Jγe + o(e2))
where o(e2) is the remainder term that satisfies lime→0 o(e2)/e2 = 0. Applying Taylor's theorem
again,
λU − β(λC∗
) = λU − β(λC
f
) − Jβ(Jγe + o(e2))
− o((Jγe + o(e2)2)
Then,
λU − β(λC∗
f
= λU − β(λC
= (I − JβJγ)e − o(e2)
) + JβJγe − o(e2)
)2
2 = ((I − JβJγ)e − o(e2))T ((I − JβJγ)e − o(e2))
= eT (I − JβJγ)T (I − JβJγ)e − o(e2)T (I − JβJγ)e
− eT (I − JβJγ)T o(e2) + o(e2)T o(e2)
= eT (I − JβJγ)T (I − JβJγ)e + o(e2
2)
≤ µe2
2 + o(e2
2)
20/41
where µ is the largest eigenvalue of (I−JβJγ)T (I−JβJγ). If e is sufficiently small so that o(e2
then
2)) < (1 − µ)e2
2,
λU − λmaxσ(kDW 1 λC∗
)2
2 ≤ e2
2 = λU − λmaxσ(E[U
f
])2
2
Note that the last step requires that µ, the largest eigenvalue of (I − JβJγ)T (I − JβJγ), is below 1.
Clearly, we do not actually have any guarantee of meeting this condition. However, our results show
that even though the feedback weights are random and fixed, the feedforward weights actually learn to
meet this condition during the first epoch of training (Figure 4, Supplement 1).
Hidden layer targets
Theorem 1 shows that if we use a target λC∗
)) for the
hidden layer, there is a guarantee that the hidden layer approaching this target will also push the upper
layer closer to its target λU , if certain other conditions are met. Our specific choice of λC defined in the
Results (equation (6)) approximates this target rate vector using variables that are accessible to the
hidden layer units.
) − σ(Y λmaxσ(kDW 1λC
+ σ(Y λU
= λC
f
f
t
Because neuronal units calculate averages after the network has reached a steady state, λmaxσ(U
f
f
f
f
] ≈ kDW 1λC
) ≈
. If we assume that
f
λU
U
and A
f ≈ E[U
f
f ≈ Y λU
], which is true on average, then:
. Using Lemma 1 and equation (4), E[U
αf = σ(A
f
) ≈ σ(Y λU
f
) ≈ σ(Y λmaxσ(U
f
)) ≈ σ(Y λmaxσ(kDW 1λC
f
))
and:
αt = σ(A
t
) ≈ σ(Y λU
t
)
(28)
(29)
Therefore, λC ≈ λC∗
Thus, our hidden layer targets ensure that our model employs a learning rule similar to difference
.
target propagation that approximates the necessary conditions to guarantee error convergence.
Lemma for firing rates
Theorem 1 had to rely on the equivalence between the average spike rates of the neurons and the
postsynaptic potentials. Here, we prove a lemma showing that this equivalence does indeed hold as long
as the integration time is long enough relative to the synaptic time constants ts and tL.
Lemma 1. Let X be a set of presynaptic spike times during the time interval ∆t = t1 − t0, distributed
according to an inhomogeneous Poisson process. Let N = X denote the number of presynaptic spikes
during this time window, and let si ∈ X denote the ith presynaptic spike, where 0 < i ≤ N . Finally, let
λ(t) denote the time-varying presynaptic firing rate (i.e. the time-varying mean of the Poisson process),
and P SP (t) be the postsynaptic potential at time t given by equation (10). Then, during the time window
∆t, as long as ∆t (cid:29) 2τ 2
s λ2/(τL − τs)2(τL + τs),
Lτ 2
E[P SP (t)] ≈ λ
21/41
Proof. The average of P SP (t) over the time window ∆t is
P SP =
=
Since Θ(t − s) = 0 for all t < s,
P SP =
=
1
∆t
1
∆t
1
∆t
1
∆t
EX [P SP ] = EX
(cid:34)
1
∆t
EX [N ]
(cid:90) t1
(cid:88)
t0
P SP (t)dt
(cid:90) t1
s∈X
t0
s
s∈X
(cid:90) t1
(cid:88)
(cid:32)
N −(cid:88)
(cid:32)
N −(cid:88)
(cid:34) N(cid:88)
s∈X
s∈X
− 1
∆t
e−(t−s)/τL − e−(t−s)/τs
τL − τs
Θ(t − s)dt
e−(t−s)/τL − e−(t−s)/τs
τL − τs
dt
τLe−(t1−s)/τL − τse−(t1−s)/τs
τL − τs
(cid:33)
(cid:33)(cid:35)
τLe−(t1−s)/τL − τse−(t1−s)/τs
τL − τs
(cid:32)
τLe−(t1−si)/τL − τse−(t1−si)/τs
(cid:33)(cid:35)
The expected value of P SP with respect to X is given by
g(s)fsi(s)ds
22/41
Since the presynaptic spikes are an inhomogeneous Poisson process with a rate λ, EX [N ] =(cid:82) t1
∆t
i=0
τL − τs
EX
=
λdt.
t0
Thus,
(cid:90) t1
t0
1
∆t
EX [P SP ] =
(cid:35)
g(si)
(cid:34) N(cid:88)
(cid:35)
i=0
EX
λdt − 1
∆t
(cid:34) N(cid:88)
= λ − 1
∆t
EX
g(si)
i=0
where we let g(si) ≡ (τLe−(t1−si)/τL − τse−(t1−si)/τs)/(τL − τs). Then, the law of total expectation gives
Letting fsi (s) denote P (si = s), we have that
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)N
(cid:35)(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)N = n
(cid:35)
(cid:33)
· P (N = n)
EX
g(si)
= EN
g(si)
(cid:34) N(cid:88)
(cid:35)
i=0
=
(cid:34) N(cid:88)
i=0
EX
g(si)
i=0
EX
(cid:34)
(cid:32)
(cid:34) N(cid:88)
(cid:34) N(cid:88)
∞(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)N = n
(cid:35)
EX
i=0
=
n=0
g(si)
n(cid:88)
n(cid:88)
i=0
EX [g(si)]
(cid:90) t1
=
i=0
t0
Then,
i=0
(cid:35)
g(si)
=
=
(cid:34) N(cid:88)
i=0
EX
fsi(s) =
=
(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)N = n
λ(s)
λ(t)dt
(cid:82) t1
t0
λ(s)
λ∆t
(cid:90) t1
n(cid:88)
(cid:90) t1
i=0
t0
1
λ∆t
n
=
=
λ∆t
t0
(cid:34) N(cid:88)
EX
g(si)
g(s)λ(s)ds
(cid:32)(cid:90) t1
t0
(cid:32)
(cid:32)(cid:90) t1
λ∆t
n
∞(cid:88)
n=0
1
λ∆t
t0
g(s)λ(s)ds
· P (N = n)
g(s)λ(s)ds
n · P (N = n)
(cid:33)
(cid:33)
g(s)λ(s)ds
(cid:33)
(cid:33)(cid:32) ∞(cid:88)
n=0
−(cid:82) t1
t0
λ(t)dt
[(cid:82) t1
P (N = n) =
=
t0
λ(t)dt]ne
n!
[λ∆t]ne−(λ∆t)
Now, for an inhomogeneous Poisson process with time-varying rate λ(t),
Since Poisson spike times are independent, for an inhomogeneous Poisson process:
for all s ∈ [t0, t1]. Since Poisson spike times are independent, this is true for all i. Thus,
Thus,
Then,
n!
(cid:32)(cid:90) t1
(cid:32)(cid:90) t1
t0
t0
(cid:33)(cid:32) ∞(cid:88)
(cid:33)
n=0
g(s)λ(s)ds
(cid:33)
n
[λ∆t]n
n!
g(s)λ(s)ds
(λ∆t)eλ∆t
(cid:34) N(cid:88)
i=0
EX
(cid:35)
g(si)
=
=
=
e−(λ∆t)
λ∆t
e−(λ∆t)
λ∆t
(cid:90) t1
g(s)λ(s)ds
t0
(cid:32)(cid:90) t1
t0
(cid:33)
g(s)λ(s)ds
EX [P SP ] = λ − 1
∆t
The second term of this equation is always greater than or equal to 0, since g(s) ≥ 0 and λ(s) ≥ 0 for all
s. Thus, EX [P SP ] ≤ λ. As well, the Cauchy-Schwarz inequality states that
23/41
(cid:90) t1
t0
g(s)λ(s)ds ≤
=
(cid:115)(cid:90) t1
(cid:115)(cid:90) t1
t0
t0
(cid:115)(cid:90) t1
(cid:112)
t0
g(s)2ds
λ(s)2ds
g(s)2ds
λ2∆t
(cid:90) t1
t0
g(s)2ds =
(cid:90) t1
t0
(cid:16) τLe−(t1−s)/τL − τse−(t1−s)/τs
(cid:17)2
ds
(cid:32)
4
(cid:33)
τL − τs
τ 2
Lτ 2
s
τL + τs
1
2(τL − τs)2
2τ 2
Lτ 2
s
≤
=
(τL − τs)2(τL + τs)
(cid:115)
g(s)λ(s)ds ≤
(cid:90) t1
t0
(cid:112)
λ2∆t
2τ 2
Lτ 2
s
(τL − τs)2(τL + τs)
(cid:115)
2τ 2
Lτ 2
s λ2
(τL − τs)2(τL + τs)
√
∆t
=
(cid:115)
√
∆t
EX [P SP ] ≥ λ − 1
∆t
(cid:115)
= λ −
2τ 2
Lτ 2
s λ2
(τL − τs)2(τL + τs)
2τ 2
Lτ 2
s λ2
∆t(τL − τs)2(τL + τs)
(cid:115)
λ −
2τ 2
Lτ 2
s λ2
∆t(τL − τs)2(τL + τs)
≤ EX [P SP ] ≤ λ
where
Thus,
Therefore,
Then,
Thus, as long as ∆t (cid:29) 2τ 2
s λ2/(τL − τs)2(τL + τs), EX [P SP ] ≈ λ.
Lτ 2
What this lemma says, effectively, is that the expected value of P SP is going to be roughly the
average presynaptic rate of fire as long as the time over which the average is taken is sufficiently long in
comparison to the postsynaptic time constants and the average rate-of-fire is sufficiently small. In our
simulations, ∆t is always greater than or equal to 50 ms, the average rate-of-fire is approximately 20 Hz,
and our time constants τL and τs are 10 ms and 3 ms, respectively. Hence, in general:
2τ 2
Lτ 2
s λ2/(τL − τs)2(τL + τs) = 2(10)2(3)2(0.02)2/(10 − 3)2(10 + 3)
≈ 0.001
(cid:28) 50
Thus, in the proof of Theorem 1, we assume EX [P SP ] = λ.
24/41
Parameter Units
Value
Description
dt
λmax
τs
τL
∆ts
gB
gA
gD
gL
P0
P1
ms
Hz
ms
ms
ms
S
S
S
S
–
–
1
200
3
10
30
0.6
Time step resolution
Maximum spike rate
Short synaptic time constant
Long synaptic time constant
Settle duration for calculation of average voltages
Hidden layer conductance from basal dendrite to the soma
0, 0.05, 0.6 Hidden layer conductance from apical dendrite to the soma
0.6
0.1
20/λmax
20/λ2
max
Output layer conductance from dendrite to the soma
Leak conductance
Hidden layer error signal scaling factor
Output layer error signal scaling factor
Table 1. List of parameter values used in our simulations.
25/41
Acknowledgments
We would like to thank Douglas Tweed, Joao Sacramento, Walter Senn, and Yoshua Bengio for helpful
discussions on this work. This research was supported by two grants to B.A.R.: a Discovery Grant from
the Natural Sciences and Engineering Research Council of Canada (RGPIN-2014-04947) and a Google
Faculty Research Award. The authors declare no competing financial interests. Some simulations were
performed on the gpc supercomputer at the SciNet HPC Consortium. SciNet is funded by: the Canada
Foundation for Innovation under the auspices of Compute Canada; the Government of Ontario; Ontario
Research Fund - Research Excellence; and the University of Toronto.
References
Bengio, Y. and LeCun, Y. (2007). Scaling learning algorithms towards AI. Large-scale kernel machines,
34(5):1–41.
Bengio, Y., Lee, D.-H., Bornschein, J., and Lin, Z. (2015). Towards biologically plausible deep learning.
arXiv preprint arXiv:1502.04156.
Bittner, K. C., Grienberger, C., Vaidya, S. P., Milstein, A. D., Macklin, J. J., Suh, J., Tonegawa, S.,
and Magee, J. C. (2015). Conjunctive input processing drives feature selectivity in hippocampal CA1
neurons. Nat Neurosci, 18(8):1133–1142.
Brombas, A., Fletcher, L. N., and Williams, S. R. (2014). Activity-Dependent Modulation of Layer 1
Inhibitory Neocortical Circuits by Acetylcholine. The Journal of Neuroscience, 34(5):1932.
Budd, J. M. (1998). Extrastriate feedback to primary visual cortex in primates: a quantitative analysis of
connectivity. Proceedings of the Royal Society of London B: Biological Sciences, 265(1400):1037–1044.
Burbank, K. S. (2015). Mirrored STDP Implements Autoencoder Learning in a Network of Spiking
Neurons. PLoS Comput Biol, 11(12):e1004566.
Burbank, K. S. and Kreiman, G. (2012). Depression-Biased Reverse Plasticity Rule Is Required for
Stable Learning at Top-Down Connections. PLoS Comput Biol, 8(3):e1002393.
Buzs´aki, G. and Draguhn, A. (2004). Neuronal oscillations in cortical networks. Science, 304(5679):1926–
1929.
Cadieu, C. F., Hong, H., Yamins, D. L. K., Pinto, N., Ardila, D., Solomon, E. A., Majaj, N. J., and
DiCarlo, J. J. (2014). Deep neural networks rival the representation of primate IT cortex for core
visual object recognition. PLoS Comput Biol, 10(12):e1003963.
Cox, D. and Dean, T. (2014). Neural Networks and Neuroscience-Inspired Computer Vision. Current
Biology, 24(18):R921–R929.
Crick, F. (1989). The recent excitement about neural networks. Nature, 337(6203):129–132.
Dan, Y. and Poo, M.-M. (2004). Spike timing-dependent plasticity of neural circuits. Neuron, 44(1):23–30.
Fiser, A., Mahringer, D., Oyibo, H. K., Petersen, A. V., Leinweber, M., and Keller, G. B. (2016).
Experience-dependent spatial expectations in mouse visual cortex. Nat Neurosci, 19(12):1658–1664.
Gadagkar, V., Puzerey, P. A., Chen, R., Baird-Daniel, E., Farhang, A. R., and Goldberg, J. H. (2016).
Dopamine neurons encode performance error in singing birds. Science, 354(6317):1278.
Gilbert, C. D. and Li, W. (2013). Top-down influences on visual processing. Nat Rev Neurosci,
14(5):350–363.
26/41
Grossberg, S. (1987). Competitive learning: From interactive activation to adaptive resonance. Cognitive
science, 11(1):23–63.
Guergiuev, J. (2017). Segregated-dendrite-deep-learning. Github. https://github.com/jordan-g/
Segregated-Dendrite-Deep-Learning. 23f2c66.
Hangya, B., Ranade, S., Lorenc, M., and Kepecs, A. (2015). Central cholinergic neurons are rapidly
recruited by reinforcement feedback. Cell, 162(5):1155–1168.
Harris, K. D. (2008). Stability of the fittest: organizing learning through retroaxonal signals. Trends in
Neurosciences, 31(3):130–136.
He, K., Zhang, X., Ren, S., and Sun, J. (2015). Delving deep into rectifiers: Surpassing human-level
performance on imagenet classification. In Proceedings of the IEEE International Conference on
Computer Vision, pages 1026–1034.
Hilscher, M. M., Leao, R. N., Edwards, S. J., Leao, K. E., and Kullander, K. (2017). Chrna2-
Martinotti Cells Synchronize Layer 5 Type A Pyramidal Cells via Rebound Excitation. PLOS
Biology, 15(2):e2001392.
Hinton, G. E., Osindero, S., and Teh, Y.-W. (2006). A fast learning algorithm for deep belief nets.
Neural Computation, 18(7):1527–1554.
Karnani, M. M., Jackson, J., Ayzenshtat, I., Hamzehei Sichani, A., Manoocheri, K., Kim, S., and
Yuste, R. (2016). Opening holes in the blanket of inhibition: Localized lateral disinhibition by VIP
interneurons. The Journal of Neuroscience, 36(12):3471–3480.
Khaligh-Razavi, S.-M. and Kriegeskorte, N. (2014). Deep supervised, but not unsupervised, models may
explain IT cortical representation. PLoS Comput Biol, 10(11):e1003915.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classification with deep convolutional
neural networks. In Advances in neural information processing systems, pages 1097–1105.
Kubilius, J., Bracci, S., and Op de Beeck, H. P. (2016). Deep neural networks as a computational model
for human shape sensitivity. PLoS Comput Biol, 12(4):e1004896.
Kording, K. P. and Konig, P. (2001). Supervised and Unsupervised Learning with Two Sites of Synaptic
Integration. Journal of Computational Neuroscience, 11(3):207–215.
Larkum, M. (2013). A cellular mechanism for cortical associations: an organizing principle for the
cerebral cortex. Trends in Neurosciences, 36(3):141–151.
Larkum, M. E., Nevian, T., Sandler, M., Polsky, A., and Schiller, J. (2009). Synaptic Integration in Tuft
Dendrites of Layer 5 Pyramidal Neurons: A New Unifying Principle. Science, 325(5941):756–760.
Larkum, M. E., Waters, J., Sakmann, B., and Helmchen, F. (2007). Dendritic spikes in apical dendrites
of neocortical layer 2/3 pyramidal neurons. The Journal of Neuroscience, 27(34):8999–9008.
Larkum, M. E., Zhu, J. J., and Sakmann, B. (1999). A new cellular mechanism for coupling inputs
arriving at different cortical layers. Nature, 398(6725):338–341.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521(7553):436–444.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 86(11):2278–2324.
Lee, D.-H., Zhang, S., Fischer, A., and Bengio, Y. (2015). Difference target propagation. In Joint
European Conference on Machine Learning and Knowledge Discovery in Databases, pages 498–515.
Springer.
27/41
Leibo, J. Z., Liao, Q., Anselmi, F., Freiwald, W. A., and Poggio, T. (2017). View-Tolerant Face
Recognition and Hebbian Learning Imply Mirror-Symmetric Neural Tuning to Head Orientation.
Current Biology, 27(1):62–67.
Letzkus, J. J., Kampa, B. M., and Stuart, G. J. (2006). Learning rules for spike timing-dependent
plasticity depend on dendritic synapse location. Journal of Neuroscience, 26(41):10420–10429.
Li, Y., Li, H., Xu, Y., Wang, J., and Zhang, Y. (2016). Very deep neural network for handwritten digit
recognition. In International Conference on Intelligent Data Engineering and Automated Learning,
pages 174–182. Springer.
Liao, Q., Leibo, J. Z., and Poggio, T. (2015). How Important is Weight Symmetry in Backpropagation?
arXiv preprint arXiv:1510.05067.
Lillicrap, T. P., Cownden, D., Tweed, D. B., and Akerman, C. J. (2016). Random synaptic feedback
weights support error backpropagation for deep learning. Nature Communications, 7:13276.
Loken, C., Gruner, D., Groer, L., Peltier, R., Bunn, N., Craig, M., Henriques, T., Dempsey, J., Yu, C.-H.,
Chen, J., Dursi, L. J., Chong, J., Northrup, S., Pinto, J., Knecht, N., and Zon, R. V. (2010). Scinet:
Lessons learned from building a power-efficient top-20 system and data centre. Journal of Physics:
Conference Series, 256(1):012026.
Maaten, L. v. d. and Hinton, G. (2008). Visualizing data using t-SNE. Journal of Machine Learning
Research, 9(Nov):2579–2605.
Malenka, R. C. and Bear, M. F. (2004). LTP and LTD. Neuron, 44(1):5–21.
Manita, S., Suzuki, T., Homma, C., Matsumoto, T., Odagawa, M., Yamada, K., Ota, K., Matsubara, C.,
Inutsuka, A., Sato, M., Ohkura, M., Yamanaka, A., Yanagawa, Y., Nakai, J., Hayashi, Y., Larkum,
M., and Murayama, M. (2015). A top-down cortical circuit for accurate sensory perception. Neuron,
86(5):1304–1316.
Marblestone, A., Wayne, G., and Kording, K. (2016). Towards an integration of deep learning and
neuroscience. arXiv preprint arXiv:1606.03813.
Martin, S. J., Grimwood, P. D., and Morris, R. G. M. (2000). Synaptic plasticity and memory: an
evaluation of the hypothesis. Annual Review of Neuroscience, 23(1):649–711.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller,
M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H.,
Kumaran, D., Wierstra, D., Legg, S., and Hassabis, D. (2015). Human-level control through deep
reinforcement learning. Nature, 518(7540):529–533.
Murayama, M., Perez-Garci, E., Nevian, T., Bock, T., Senn, W., and Larkum, M. E. (2009). Dendritic
encoding of sensory stimuli controlled by deep cortical interneurons. Nature, 457(7233):1137–1141.
Munoz, W., Tremblay, R., Levenstein, D., and Rudy, B. (2017). Layer-specific modulation of neocortical
dendritic inhibition during active wakefulness. Science, 355(6328):954.
Pfeffer, C. K., Xue, M., He, M., Huang, Z. J., and Scanziani, M. (2013).
Inhibition of inhibition
in visual cortex: the logic of connections between molecularly distinct interneurons. Nat Neurosci,
16(8):1068–1076.
Pi, H.-J., Hangya, B., Kvitsiani, D., Sanders, J. I., Huang, Z. J., and Kepecs, A. (2013). Cortical
interneurons that specialize in disinhibitory control. Nature, 503(7477):521–524.
Redondo, R. L. and Morris, R. G. M. (2011). Making memories last: the synaptic tagging and capture
hypothesis. Nat Rev Neurosci, 12(1):17–30.
28/41
Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986). Learning representations by back-propagating
errors. Nature, 323:9.
Scellier, B. and Bengio, Y. (2016). Towards a biologically plausible backprop.
arXiv preprint
arXiv:1602.05179.
Schiller, J., Major, G., Koester, H. J., and Schiller, Y. (2000). NMDA spikes in basal dendrites of cortical
pyramidal neurons. Nature, 404(6775):285–289.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J.,
Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner,
N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis, D. (2016).
Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):484–489.
Sjostrom, P. J. and Hausser, M. (2006). A Cooperative Switch Determines the Sign of Synaptic Plasticity
in Distal Dendrites of Neocortical Pyramidal Neurons. Neuron, 51(2):227–238.
Spratling, M. W. (2002). Cortical Region Interactions and the Functional Role of Apical Dendrites.
Behavioral and Cognitive Neuroscience Reviews, 1(3):219–228.
Spratling, M. W. and Johnson, M. H. (2006). A feedback model of perceptual learning and categorization.
Visual Cognition, 13(2):129–165.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. (2014). Dropout: A
simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research,
15(1):1929–1958.
Steriade, M., Timofeev, I., and Grenier, F. (2001). Natural Waking and Sleep States: A View From
Inside Neocortical Neurons. Journal of Neurophysiology, 85(5):1969–1985.
Sutskever, I., Martens, J., Dahl, G. E., and Hinton, G. E. (2013). On the importance of initialization
and momentum in deep learning. ICML (3), 28:1139–1147.
Takahashi, N., Oertner, T. G., Hegemann, P., and Larkum, M. E. (2016). Active cortical dendrites
modulate perception. Science, 354(6319):1587.
Thompson, A., Picard, N., Min, L., Fagiolini, M., and Chen, C. (2016). Cortical Feedback Regulates
Feedforward Retinogeniculate Refinement. Neuron, 91(5):1021–1033.
Tieleman, T. and Hinton, G. (2012). Lecture 6.5-rmsprop: Divide the gradient by a running average of
its recent magnitude. COURSERA: Neural Networks for Machine Learning, 4(2).
Urbanczik, R. and Senn, W. (2009). Reinforcement learning in populations of spiking neurons. Nature
Neuroscience, 12(3):250.
Urbanczik, R. and Senn, W. (2014). Learning by the dendritic prediction of somatic spiking. Neuron,
81(3):521–528.
Yamada, Y., Bhaukaurally, K., Madar´asz, T. J., Pouget, A., Rodriguez, I., and Carleton, A. (2017).
Context- and Output Layer-Dependent Long-Term Ensemble Plasticity in a Sensory Circuit. Neuron,
93(5):1198–1212.e5.
Yamins, D. L. K. and DiCarlo, J. J. (2016). Using goal-driven deep learning models to understand
sensory cortex. Nat Neurosci, 19(3):356–365.
Zhang, S., Xu, M., Kamigaki, T., Hoang Do, J. P., Chang, W.-C., Jenvay, S., Miyamichi, K., Luo, L.,
and Dan, Y. (2014). Long-range and local circuits for top-down modulation of visual cortex processing.
Science, 345(6197):660–665.
29/41
Zylberberg, J., Murphy, J. T., and DeWeese, M. R. (2011). A Sparse Coding Model with Synaptically
Local Plasticity and Spiking Neurons Can Account for the Diverse Shapes of V1 Simple Cell Receptive
Fields. PLOS Computational Biology, 7(10):e1002250.
30/41
Figures
Figure 1. The credit assignment problem in multi-layer neural networks.
(A) Illustration of the credit assignment problem. In order to take full advantage of the multi-circuit architecture
of the neocortex when learning, synapses in earlier processing stages (blue connections) must somehow receive
"credit" for their impact on behavior or cognition. However, the credit due to any given synapse early in a
processing pathway depends on the downstream synaptic connections that link the early pathway to later
computations (red connections). (B) Illustration of weight transport in backpropagation. To solve the credit
assignment problem, the backpropagation of error algorithm explicitly calculates the credit due to each synapse
in the hidden layer by using the downstream synaptic weights when calculating the hidden layer weight changes.
This solution works well in AI applications, but is unlikely to occur in the real brain.
31/41
Figure 2. Potential solutions to credit assignment using top-down feedback.
(A) Illustration of the implicit feedback pathway used in previous models of deep learning. In order to assign
credit, feedforward information must be integrated separately from any feedback signals used to calculate error
for synaptic updates (the error is indicated here with δ). (B) Illustration of the segregated dendrites proposal.
Rather than using a separate pathway to calculate error based on feedback, segregated dendritic compartments
could receive feedback and calculate the error signals locally.
32/41
Figure 3. Illustration of a multi-compartment neural network model for deep learning.
(A) Left: Diagram of our network architecture. An input image is represented by the spiking of input units which
propagate to the hidden layer through weights W 0. Hidden layer activities arrive at the output layer through
weights W 1. Feedback from the output layer is sent back to the hidden layer through fixed, random weights Y .
Right: Illustration of unit i in the hidden layer. The unit has a basal dendrite, soma, and apical dendrite with
membrane potentials Bi, Ci and Ai, respectively. The apical dendrite communicates to the soma predominantly
using non-linear signals σ(·). The spiking output of the cell, Si(t), is a Poisson process with rate λmaxσ(Ci(t)).
(B) Illustration of neuronal dynamics in the two modes of processing. In the transmit mode, the apical dendrite
accumulates a measure of its average membrane potential from t1 − tp to t1. In the plateau mode, an apical
plateau potential equal to the nonlinearity σ(·) applied to the average apical potential travels to the soma and
basal dendrite. (C) Illustration of the sequence of network phases that occur for each training example. The
network undergoes a forward phase (transmit & plateau) and a target phase (transmit & plateau) in sequence.
In the target phase, a teaching signal is introduced to the output layer, shaping its activity. (D) Illustration of
phase length sampling. For each training example, the lengths of forward & target phases are randomly drawn
from a shifted inverse Gaussian distribution with a minimum of 50 ms.
33/41
Figure 4. Co-ordinated errors between the output and hidden layers.
(A) Illustration of output error (L1) and local hidden error (L0). For a given test example shown to the network
in a forward phase, the output layer error is defined as the squared norm of the difference between target firing
rates λU and the sigmoid function applied to the average somatic potentials U
across all output units. Hidden
layer error is defined similarly, except the target is λC (as defined in the text). (B) Plot of L1 vs. L0 for all of
the '2' images after one epoch of training. There is a strong correlation between hidden layer error and output
layer error (real data, black), as opposed to when output and hidden errors were randomly paired (shuffled data,
gray). (C) Plot of correlation between hidden layer error and output layer error across training for each category
of images (each dot represents one category). The correlation is significantly higher in the real data than the
shuffled data throughout training. Note also that the correlation is much lower on the first epoch of training (red
oval), suggesting that the conditions for credit assignment are still developing during the first epoch.
f
34/41
Figure 5. Improvement of learning with hidden layers.
(A) Illustration of the three networks used in the simulations. Top: a shallow network with only an input layer
and an output layer. Middle: a network with one hidden layer. Bottom: a network with two hidden layers.
Both hidden layers receive feedback from the output layer, but through separate synaptic connections with
random weights Y 0 and Y 1. (B) Plot of test error (measured on 10,000 MNIST images not used for training)
across 60 epochs of training, for all three networks described in A. The networks with hidden layers exhibit deep
learning, because hidden layers decrease the test error. Right: Spreads (min – max) of the results of repeated
weight tests (n = 20) after 60 epochs for each of the networks. Percentages indicate means (two-tailed t-test,
1-layer vs. 2-layer: t38 = 197.11, P38 = 2.5 × 10−58; 1-layer vs. 3-layer: t38 = 238.26, P38 = 1.9 × 10−61; 2-layer
vs. 3-layer: t38 = 42.99, P38 = 2.3 × 10−33, Bonferroni correction for multiple comparisons). (C) Results of
t-SNE dimensionality reduction applied to the activity patterns of the first three layers of a two hidden layer
network (after 60 epochs of training). Each data point corresponds to a test image shown to the network. Points
are color-coded according to the digit they represent. Moving up through the network, images from identical
categories are clustered closer together and separated from images of different catefories. Thus the hidden layers
learn increasingly abstract representations of digit categories.
35/41
Figure 6. Approximation of backpropagation with local learning rules.
(A) Plot of the angle between weight updates prescribed by our local update learning algorithm compared to
those prescribed by backpropagation of error, for a one hidden layer network over 10 epochs of training (each point
on the horizontal axis corresponds to one image presentation). Data was time-averaged using a sliding window of
100 image presentations. When training the network using the local update learning algorithm, feedback was
sent to the hidden layer either using spiking activity from the output layer units (blue) or by directly sending
the spike rates of output units (red). The angle between the local update ∆W 0 and backpropagation weight
BP remains under 90◦ during training, indicating that both algorithms point weight updates in
updates ∆W 0
a similar direction. (B) Examples of hidden layer receptive fields (synaptic weights) obtained by training the
network in A using our local update learning rule (left) and backpropagation of error (right) for 60 epochs. (C)
Plot of correlation between local update receptive fields and backpropagation receptive fields. For each of the
receptive fields produced by local update, we plot the maximum Pearson correlation coefficient between it and
all 500 receptive fields learned using backpropagation (Regular). Overall, the maximum correlation coefficients
are greater than those obtained after shuffling all of the values of the local update receptive fields (Shuffled).
36/41
Figure 7. Conditions on feedback synapses for effective learning.
(A) Diagram of a one hidden layer network trained in B, with 80% of feedback weights set to zero. The remaining
feedback weights Y (cid:48) were multiplied by 5 in order to maintain a similar overall magnitude of feedback signals.
(B) Plot of test error across 60 epochs for our standard one hidden layer network (gray) and a network with
sparse feedback weights (red). Sparse feedback weights resulted in improved learning performance compared to
fully connected feedback weights. Right: Spreads (min – max) of the results of repeated weight tests (n = 20)
after 60 epochs for each of the networks. Percentages indicate mean final test errors for each network (two-tailed
t-test, regular vs. sparse: t38 = 16.43, P38 = 7.4 × 10−19). (C) Diagram of a one hidden layer network trained
in D, with feedback weights that are symmetric to feedforward weights W 1, and symmetric but with added
noise. Noise added to feedback weights is drawn from a normal distribution with variance σ = 0.05. (D) Plot of
test error across 60 epochs of our standard one hidden layer network (gray), a network with symmetric weights
(red), and a network with symmetric weights with added noise (blue). Symmetric weights result in improved
learning performance compared to random feedback weights, but adding noise to symmetric weights results in
impaired learning. Right: Spreads (min – max) of the results of repeated weight tests (n = 20) after 60 epochs
for each of the networks. Percentages indicate means (two-tailed t-test, random vs. symmetric: t38 = 18.46,
P38 = 4.3 × 10−20; random vs. symmetric with noise: t38 = −71.54, P38 = 1.2 × 10−41; symmetric vs. symmetric
with noise: t38 = −80.35, P38 = 1.5 × 10−43, Bonferroni correction for multiple comparisons).
37/41
Figure 8. Importance of dendritic segregation for deep learning.
(A) Left: Diagram of a hidden layer neuron with an unsegregated apical dendrite that affects the somatic potential
Ci. gA represents the strength of the coupling between the apical dendrite and soma. Right: Example traces of
the apical voltage Ai and the somatic voltage Ci in response to spikes arriving at apical synapses. Here gA = 0.05,
so the apical activity is strongly attenuated at the soma. (B) Plot of test error across 60 epochs of training on
MNIST of a two hidden layer network, with total apical segregation (gray), strong apical attenuation (red) and
weak apical attenuation (blue). Apical input to the soma did not prevent learning if it was strongly attenuated,
but weak apical attenuation impaired deep learning. Right: Spreads (min – max) of the results of repeated
weight tests (n = 20) after 60 epochs for each of the networks. Percentages indicate means (two-tailed t-test,
total segregation vs. strong attenuation: t38 = −4.00, P38 = 8.4 × 10−4; total segregation vs. weak attenuation:
t38 = −95.24, P38 = 2.4 × 10−46; strong attenuation vs. weak attenuation: t38 = −92.51, P38 = 7.1 × 10−46,
Bonferroni correction for multiple comparisons).
38/41
Figure Supplements
Figure 4, Supplement 1. Weight alignment during first epoch of training.
(A) Plot of the maximum eigenvalue of (I − JβJγ)T (I − JβJγ) over 60,000 training examples for a one hidden
layer network, where Jβ and Jγ are the mean feedforward and feedback Jacobian matrices for the last 100 training
examples. The maximum eigenvalue of (I − JβJγ)T (I − JβJγ) drops below 1 as learning progresses, satisfying
the main condition for the learning guarantee described in Theorem 1 to hold. (B) The product of the mean
feedforward and feedback Jacobian matrices, JβJγ, for a one hidden layer network, before training (left) and
after 1 epoch of training (right). As training progresses, the network updates its weights in a way that causes
this product to approach the identity matrix, meaning that the two matrices are roughly inverses of each other.
39/41
Figure 5, Supplement 1. Learning with stochastic plateau times.
(A) Left: Raster plot showing plateau potential times during presentation of two training examples for 100
neurons in the hidden layer of a network where plateau potential times were randomly sampled for each neuron
from a folded normal distribution (µ = 0, σ2 = 3) that was truncated (max = 5) such that plateau potentials
occurred between 0 ms and 5 ms before the start of the next phase. In this scenario, the apical potential over the
last 30 ms was integrated to calculate the plateau potential for each neuron. (B) Plot of test error across 60
epochs of training on MNIST of a one hidden layer network, with synchronized plateau potentials (gray) and
with stochastic plateau potentials (red). Allowing neurons to undergo plateau potentials in a stochastic manner
did not hinder training performance.
40/41
Figure 7, Supplement 1. Importance of weight magnitudes for learning with sparse weights.
Plot of test error across 20 epochs of training on MNIST of a one hidden layer network, with regular feedback
weights (gray), sparse feedback weights that were amplified (red), and sparse feedback weights that were not
amplified (blue). The network with amplified sparse feedback weights is the same as in Figure 7A & B, where
feedback weights were multiplied by a factor of 5. While sparse feedback weights that were amplified led to
improved training performance, sparse weights without amplification impaired the network's learning ability.
Right: Spreads (min – max) of the results of repeated weight tests (n = 20) after 20 epochs for each of the networks.
Percentages indicate means (two-tailed t-test, regular vs. sparse, amplified: t38 = 44.96, P38 = 4.4 × 10−34;
regular vs. sparse, not amplified: t38 = −51.30, P38 = 3.2 × 10−36; sparse, amplified vs. sparse, not amplified:
t38 = −100.73, P38 = 2.8 × 10−47, Bonferroni correction for multiple comparisons).
41/41
|
1801.05017 | 1 | 1801 | 2018-01-15T21:17:37 | Hypergraph based Subnetwork Extraction using Fusion of Task and Rest Functional Connectivity | [
"q-bio.NC"
] | Functional subnetwork extraction is commonly used to explore the brain's modular structure. However, reliable subnetwork extraction from functional magnetic resonance imaging (fMRI) data remains challenging due to the pronounced noise in neuroimaging data. In this paper, we proposed a high order relation informed approach based on hypergraph to combine the information from multi-task data and resting state data to improve subnetwork extraction. Our assumption is that task data can be beneficial for the subnetwork extraction process, since the repeatedly activated nodes involved in diverse tasks might be the canonical network components which comprise pre-existing repertoires of resting state subnetworks. Our proposed high order relation informed subnetwork extraction based on a strength information embedded hypergraph, (1) facilitates the multisource integration for subnetwork extraction, (2) utilizes information on relationships and changes between the nodes across different tasks, and (3) enables the study on higher order relations among brain network nodes. On real data, we demonstrated that fusing task activation, task-induced connectivity and resting state functional connectivity based on hypergraphs improves subnetwork extraction compared to employing a single source from either rest or task data in terms of subnetwork modularity measure, inter-subject reproducibility, along with more biologically meaningful subnetwork assignments. | q-bio.NC | q-bio |
Hypergraph based Subnetwork Extraction using
Fusion of Task and Rest Functional Connectivity
Chendi Wang, Rafeef Abugharbieh
Biomedical Signal and Image Computing Lab, UBC, Canada
[email protected], [email protected]
Abstract
Functional subnetwork extraction is commonly used to explore the brain's
modular structure. However, reliable subnetwork extraction from functional
magnetic resonance imaging (fMRI) data remains challenging due to the pro-
nounced noise in neuroimaging data. In this paper, we proposed a high order
relation informed approach based on hypergraph to combine the information
from multi-task data and resting state data to improve subnetwork extrac-
tion. Our assumption is that task data can be beneficial for the subnetwork
extraction process, since the repeatedly activated nodes involved in diverse
tasks might be the canonical network components which comprise pre-existing
repertoires of resting state subnetworks [1]. Our proposed high order relation
informed subnetwork extraction based on a strength information embedded
hypergraph, (1) facilitates the multisource integration for subnetwork extrac-
tion, (2) utilizes information on relationships and changes between the nodes
across different tasks, and (3) enables the study on higher order relations
among brain network nodes. On real data, we demonstrated that fusing task
activation, task-induced connectivity and resting state functional connectivity
based on hypergraphs improves subnetwork extraction compared to employing
a single source from either rest or task data in terms of subnetwork modularity
measure, inter-subject reproducibility, along with more biologically meaning-
ful subnetwork assignments.
Keywords: Brain Subnetwork Extraction, Multisource Fusion, Functional
Connectivity, Hypergraph
1
Introduction
The human brain can be regarded as being a network where units, or nodes, represent
different specialized regions, and edges represent communication pathways. Brain
1
network analysis methods for connectome studies include an important branch of
brain subnetwork identification. Given brain connectivity matrices, brain networks
can be quantitatively examined for certain commonly used network measures. The
modular structure (community structure) is of particular interest; it is from this
structure that we can infer information about brain subnetworks. The modular
structure is extracted by subdividing a network into groups of nodes with the maxi-
mal possible within-group links and minimal between-group links using community
detection methods [2].
Most existing functional subnetwork extraction methods focus on resting state
function connectivity data [3, 4], using functional homogeneity clustering, Independent
Component Analysis (ICA), or graph community detection. However, resting state
functional connectivity is inherently with low signal-to-noise ratio (SNR) and prone
to false positive correlations [5]. Such noisy resting state functional connectivity in-
formation leads to unreliable subnetwork extraction results. Given the resemblance
between resting state and task functional subnetworks [6] and high order nodal re-
lations reflected from multi-task data, we here aim to incorporate information from
task data into the subnetwork extraction based on multilayer network. We explore if
this integration can improve the subnetwork extraction by exploiting the mechanism
of how groups of nodes collaborate together to execute a function and how these
groups communicate with each other.
1.1 Related Work - Relationship between Task and Resting
Functional Connectivity
Recent studies indicate that resting state functional activity actually persists dur-
ing task performance [7], and similar network architecture is present across task and
rest, which is supported by the existence of similar multi-task Functional Connec-
tivity (FC) and resting-state FC matrices that were averaged across subjects [8].
Studies have also shown that there is a strong resemblance between rest and task
subnetworks [6, 9]. The spatial overlap between resting-state functional subnetworks
and task-evoked activities has been discovered [10, 11].
Based on the close relationship between the two, resting state data have been
used to predict the task activities, by using group ICA to discover repertories of
canonical network components that will be recruited in tasks [1]; by applying the
graphical connectional topology of brain regions at rest to predict functional activity
of them during task [11]; or based on a voxel-matched regression method to estimate
the magnitude of task-induced activity [12].
On the other hand, aggregating brain imaging data from thousands of task re-
lated studies allowed the construction of 'co-activation networks', whose major com-
2
ponents and overall network topology strongly resembled functional subnetworks
derived from resting-state recordings [13, 14, 15].
It has been suggested that networks involved in cognition are a subset of net-
works embedded in spontaneous activity [6, 16], and a number of canonical network
components in the pre-existing repertoires of intrinsic subnetworks are selectively
and dynamically recruited for various cognitions [17, 1].
1.2 Related Work - Multilayer Brain Network Analysis
Multilayer network has recently been used to model and analyze complex high or-
der data, such as multivariate and multiscale information within the human brain
[18]. Different layers can represent relationships across different temporal variations
[19], reflect different imaging modalities (such as task and rest) [18], or different
frequency bands [20], etc. Hypergraph is a type of multilayer graphs, in which
edges can link any number of nodes [21]. Hypergraphs have been used to identify
non-random structure in structural connectivity of the cortical microcircuits [22],
identify high order brain connectome biomarkers for disease prediction [23], and
study relationships between functional and structural connectome data [24].
2 High Order Relation Informed Subnetwork Ex-
traction
Our assumption is that task data can be beneficial for subnetwork extraction since
the repeatedly activated nodes in different tasks could be the canonical network
components in the spontaneous resting state subnetworks. At the same time, the
multilayer structure of repeatedly activated nodes across multi-task can be elegantly
presented as a hypergraph. We propose a high order relation informed subnetwork
extraction model, which (1) facilitates multisource integration of task and rest data
for subnetwork extraction, (2) utilizes information from the relationship between
groups of activated nodes across different tasks, and (3) enables the study on higher
order relations among brain network nodes.
2.1 Framework
We propose a high order relation informed approach based on hypergraph to in-
tegrate both resting state and task information for brain subnetwork extraction.
We firstly construct a brain graph based on a certain parcellation atlas. Secondly,
we detect activation of brain nodes from task data to define the nodes for mul-
tiple layers in the hypergraph, and define the connection strength between nodes
3
using task-induced connectivity. Thirdly, we construct the multitask hypergraph
and incorporate resting state FC strength information when setting the weights of
hyperedges. Fourthly, we fuse task and rest FC using weighted combination model
before performing graphcut on the constructed graph.
2.2 Notation Overview of Hypergraph
2.2.1 Notations
We here follow most of the notations presented in [21]. Let V denote a set of
nodes, and E denote a family of subsets e of V such that ∪e ∈ E = V . Then we
define G = (V ; E) a hypergraph with the vertex set V and the hyperedge set E. A
hyperedge containing just two nodes is a simple graph edge. A hyperedge e is said
to be incident with a node v when v ∈ e. Two nodes are connected if they both
belong to the same hyperedge. Two hyperedges are connected if the intersection of
them is not an empty set, ei ∩ ej (cid:54)= ∅. Given an arbitrary set X, let X denote the
cardinality of X. A hypergraph G can be represented by a V ×E incidence matrix
H with entries h(v, e) = 1 if v ∈ e and 0 otherwise, see an example in Figure 1. A
weighted hypergraph, G = (V ; E; w), is a hypergraph that has a positive number
w(e) associated with each hyperedge e, called the weight of hyperedge e. Next, we
define four important measures of hypergraph properties.
the number of nodes that exist in the hyperedge.
2. We further define the hyperdegree of a hyperedge as the number of hyperedges
hyperedge degrees. Take Figure 1 as an example, δ(e1) = 3, and δ(e2) = 2.
For a hyperedge e ∈ E:
1. We follow [21] to define its degree as d h(e) = δ(e) := e, which counts
If one uses the incidence ma-
trix, δ(e) := (cid:80){v∈V } h(v, e). Let De denote the diagonal matrices containing the
connected to it, denoted as d hH(e) :=(cid:80){ei∈E,ei(cid:54)=e} e∩ ei. For example, d hH(e1) =
3. We follow [21] to define its degree by d(v) = (cid:80){e∈Ev∈e} w(e).
the incidence matrix, d(v) =(cid:80){e∈E} w(e)h(v, e). When all w(e) = 1, d(v) counts
the number of hyperedges which include this node: d(v) =(cid:80){e∈Ev∈e} 1, or d(v) =
(cid:80){e∈E} h(v, e). Let Dv denote the diagonal matrices containing the node degrees.
4. We then define the hyperdegree of a node as d H(v) :=(cid:80){v∈ee∈E} δ(e), which
3, d hH(e3) = 2, and d hH(e4) = 0 in Figure 1.
For a node v ∈ V :
If one uses
counts the number of nodes connected to a particular node across all hyperedges.
For example, d H(v2) = 5, d H(v3) = 6, d H(v5) = 3 in Figure 1.
version will be estimating the strength between the connected node pairs.
Its weighted
Next, let W denote the diagonal matrix containing the weights w(e) of hyper-
edges. Correspondingly, the adjacency matrix A of hypergraph G is defined as:
4
(a) Toy example of a hypergraph
(b) Simple graph
(c) Incidence matrix
H
Left:
Figure 1: Hypergraph and its corresponding simple graph and incidence ma-
an hyperedge set E = {e1, e2, e3, e4} and a node set V =
trix.
{v1, v2, v3, v4, v5, v6, v7}. Middle: the corresponding simple graph. Right: the in-
cidence matrix H of the hypergraph on the left, with the entry (vi, ej) being set to
1 if vi is in ej, and 0 otherwise.
A = HWHT − Dv,
(1)
where HT is the transpose of H.
2.2.2 Graphcut of the Hypergraph
One can group the nodes into subsets using graph partitioning methods, i.e., graph-
cut. The intuition is to find a partition of the graph such that the edges within a
adjacency matrix A(X, Y ) :=(cid:80)
subset have high weights (strong intra-class connections), and the edges between dif-
ferent subsets have low weights (weak inter-class connections). Let S ∈ V denote a
subset of nodes and Sc denote the complement of S. Follow the notations in [25], the
i∈X,j∈Y aij. For a given number M of subnets, the
Mincut approach [26] implements the graphcut by generating a partition S1, . . . , SM
which minimizes
cut(S1, . . . , SM ) :=
1
2
A(Si, Sc
i ).
(2)
M(cid:88)
i=1
To solve the problem of separating individual nodes as a subset in Mincut, Ra-
tioCut [27] and Normalized cuts (Ncuts) [28] have been proposed to encode the
information of the size of a subset.
RatioCut(S1, . . . , SM ) :=
M(cid:88)
i=1
1
2
A(Si, Sc
i )
Si
=
M(cid:88)
i=1
cut(Si, Sc
i )
Si
,
where S measures the number of nodes in S.
Ncut(S1, . . . , SM ) :=
1
2
M(cid:88)
i=1
A(Si, Sc
i )
vol(Si)
=
5
M(cid:88)
i=1
cut(Si, Sc
i )
vol(Si)
,
(3)
(4)
attached to the nodes as vol(S) := (cid:80)
where vol(S) measures the volume of S by summing over the weights of all edges
v∈S ds(v), and node strength ds(v) is the
weighted version of node degree d(v).
Ncuts has been widely used in image segmentation and brain study commu-
nity, since it utilizes the weight information. In the following, we show that Ncuts
approach can be generalized from simple graphs to hypergraphs, which has been
proven in [21].
For a hypergraph G = (V ; E; w), a cut is a partition of V into two parts S and Sc.
A hypergraph e is cut when it is incident with the nodes in S and Sc at the same time.
The hyperedge boundary of S is defined as ∂S := {e ∈ Ee∩S (cid:54)= ∅, e ∈ Ee∩Sc (cid:54)= ∅},
which is a hyperedge set consisting of the hyperedges which are cut [21]. The
nodes in S, vol(S) := (cid:80)
definition of the volume in a hypergraph vol(S) is the sum of the degrees of the
v∈S d(v). Each hyperedge is essentially a fully connected
subgraph, then the edges in a subgraph is called subedges, being assigned with the
same weight w(e)/δ(e). When a hyperdege e is cut, there are e∩ Se∩ Sc subedges
are cut. Hence, the volume of ∂S is defined by
vol(∂S) :=
e ∩ Se ∩ Sc
δ(e)
,
w(e)
(5)
(cid:88)
e∈∂S
which is the sum of weights over the subedges being cut. By this definition, we have
vol(∂S) = vol(∂Sc). Similar to the simple graphs, Normalized hypergraph cut is to
keep the high intra-class connection and low inter-class connection with a partition
S1, . . . , SM by minimizing the cut as below:
M(cid:88)
i=1
argmin
∅(cid:54)=S1,...,SM⊂V
vol(∂Si)
vol(Si)
.
(6)
2.3 Task Activation Detection - Node Definition in the Hy-
pergraph
In order to construct the multiple layers in the hypergraph, we apply the activation
detection technique on the task data to define the nodes that are contained in
different hyperedges. The standard way of activation detection is to use a General
Linear Model (GLM) where statistics, such as t-values, reflect the degree of the
similarity between the stimulus and voxel time courses. The estimated statistics
produce an activation statistics map (t-map), followed by a thresholding of the
map to identify the activated voxels [29]. Due to the pronounced noise in the
Functional Magnetic Resonance Imaging (fMRI) data, activation detection at the
individual level could be inaccurate [30].
In order to derive more reliable task-
induced activation, we have chosen a group activation detection over the individual
6
based approach. First, to compute the intra-subject activation patters, a standard
GLM is applied as below [29]:
Yi = Xiβi + Ei,
(7)
where Yi is a t×N matrix of the task-induced fMRI time courses of N brain regions
from subject i, βi is a d× N activation matrix to be estimated, Ei is a t× N residual
confounds] is a t × d matrix. Xtask is the task regressors
matrix, and Xi = [XtaskXi
and Xi
confounds is the confound regressors. Next, we combine the activation results
across subjects to assemble a group activation map, which is used to define nodes
for each layer of the hypergraph. Specifically, we apply a max-t permutation test
[31] on βi aggregated from all the subjects, which implicitly accounts for multiple
comparisons and control over false detections [32]. Group activation is declared at
a p-value threshold of 0.05.
2.4 Strength Informed Weighted Multi-task Hypergraph
In the beginning of section 2, we argued that multi-task information can be pre-
sented as a hypergraph, with the hyperedges being different tasks, and the nodes
in each hyperedge being the brain regions activated in a certain task. In the tradi-
tional definition of hypergraph, nodes are connected to each other binarily, i.e., the
edge weights between a node pair are 1 if they are connected, or 0 otherwise. We
here propose a strength informed weighted hypergraph model by incorporating the
strength information from the connections between nodes. We further determine the
hyperedge weight w(e) using the graphical measures defined in subsubsection 2.2.1.
2.4.1 Pairwise Nodal Connection Strength Estimation
In order to estimate the strength of the connections between two nodes, we use
the Pearson's correlations between time courses from pairs of brain regions. We
denote the resting state connectivity matrix as Crest. To produce the task-induced
connectivity matrix Ctask, we use the task-induced time course information. We
follow the strategy in [8] to remove all inter-block rest periods from all regions' time
courses, before computing the pairwise Pearson's correlations across all concatenated
block/event duration time courses within a task. To keep the consistency when
combining information from the nodes across different layers, we keep all the Ctask
having the same dimension of N × N as the Crest, then set the rows and columns of
non-activated nodes to zero.
7
2.4.2 Proposed Strength Informed Weighted Hypergraph
We present a modified hypergraph cut criteria formulation based on Equation 5
to incorporate pairwise nodal connection strength information from C as below in
Equation 8. The symbol . indicates the usage of strength information.
(cid:80)
(cid:88)
e∈∂S
vol(∂S) :=
w(e)
i∈{e∩S},j∈{e∩Sc} Ce
ij
δ(e)
,
(8)
where vol(∂S) is a strength informed version of vol(∂S) in Equation 5, Ce is the
connectivity matrix derived from the task corresponding to the layer e, and w(e)
is the modified weight item in the hypergraph. We propose here to incorporate
strength information from the connectivity matrix and utilize the four hypergraph
measures defined in subsubsection 2.2.1 to determine w(e), whose nature is the
importance of the hyperedge in the hypergraph. Based on the definition of the
four hypergraph measures, we exploit their corresponding biological meanings to set
vol(∂S) and w(e) as below:
1. The degree of a hyperedge δ(e) counts the number of brain regions that are
activated in a task. To avoid the bias of the hyperedge size, vol(∂S) should be
normalized by δ(e).
2. The hyperdegree of a hyperedge is defined as the number of hyperedges that
are connected to it. Higher value indicates that more frequently activated patterns
in the brain activities exist in this hyperedge. Thus, w(e) should be proportional to
d hH(e), i.e., w(e) ∝ d hH(e).
3. The degree of a node counts the number of hyperedges that contain this node,
and the biological equivalence is the number of different tasks in which one node is
activated. A node with a higher degree is similar to the definition of the connector
hubs residing within different subnetworks. Hence, w(e) should be proportional to
some statistics derived from d(v) of the nodes in a hyperedge e. We denote the
statistics computation method as stat here and it can be widely used statistics such
as average value (mean), median value (median) and maximum value (max). Thus,
w(e) ∝ stat(d(v)).
4. The hyperdegree of a node reflects the number of all other nodes that are
connected to it across all layers, which equals the number of connections from other
co-activated nodes to it across multiple tasks. The biological meaning of a node
with a high value coincides with the definition of hubs. Hence, w(e) should be
proportional to some statistics derived from d H(v) of the nodes in a hyperedge e,
i.e., w(e) ∝ stat(d H(v)). Here, in order to incorporate strength information, we
apply the weighted version of d H(v), the strength of the node d Hs(v) as defined
8
in Equation 9, i.e., w(e) ∝ stat(d Hs(v)).
(cid:88)
(cid:88)
{v∈ee∈E}
u∈e
d Hs(v) :=
Ce
uv,
(9)
where Ce is the task-induced connectivity matrix for the eth task.
In order to utilize strength information and hypergraph measures, we propose
the w(e) formulation as below:
w(e) := w1 · d hH(e) + w2 · stat(d(v)) + w3 · stat(d Hs(v)),
(10)
where w1, w2, w3 are free parameters to control the contributions of each measure
to the hyperedge.
2.5 Multisource Integration of Rest and Task fMRI
Given the close correspondence between task and rest connectivity architecture
and subnetworks, we further extend the multi-task hypergraph model to integrate
Resting State Functional Connectivity based on MRI (rs-fcMRI) information. To
do that, we use Crest for the pairwise nodal connection strength computation in
Equation 9 as below:
d Hs(v) :=
Crest
uv ,
(11)
(cid:88)
(cid:88)
{v∈ee∈E}
u∈e
Furthermore, we explicitly combine the two sources of task and rest data for sub-
network extraction. We firstly fuse the multiple layers of the multi-task hypergraph
into one single layer, and secondly combine it with a resting state connectivity layer.
Given that the hypergraph cut criterion (Equation 5) is to evaluate the aggregated
sum of the cuts across all the pairwise subedges (nodal connections) in the hyper-
gragh, we propose to aggregate the strength information between node pairs across
all the layers. To do that, we transform the multiple pairwise nodal connections
across task layers (Equation 8) into one single nodal connection as below:
T(cid:88)
k=1
¯Ctask
ij =
1
T
w(ek)
δek Cek
ij ,
(12)
where the subscript k = 1, . . . , T is the indicator for tasks, T is the total number
of tasks available, and ek is the hyperedge in the kth layer of the hypergraph. Cek
is the connectivity matrix derived using the time courses in the task k using the
procedure described in subsubsection 2.4.1.
We next explicitly combine the two sources by a linear weighted combination
between the aggregated multi-task connectivity matrix from above (Equation 12)
and the resting state connectivity matrix in Equation 13 as below:
Ct-r := γ ¯Ctask + (1 − γ)Crest,
(13)
9
where γ a free parameter, which can be optimized by cross-validation, or deter-
mined by the number of the tasks available. Our linear model for combining two
sources, which are both derived from functional modality, was motivated by the
study indicating a largely linear superposition of task-evoked signal and resting
state modulations in the brain [7]. We also explore combining the two by applying
a multislice community detection approach [33], which extends modularity quality
function based on the stability of communities under Laplacian dynamics with a
coupling parameter ω to control over interslice correspondence of communities.
3 Results
We first investigated the similarity of connectivity between resting state and task-
general and task-specific connectivity. To evaluate our proposed approaches, we
assessed the graphical metric modularity Q value, the inter-subject reproducibil-
ity and examined the biological meaning of subnetwork assignments. We applied
subnetwork extraction on (1) resting state FC alone, (2) task-induced FC alone,
(3) multi-task hypergraph, (4) multi-task hypergraph integrated with resting state
connectivity strength, (5) weighted combination of (4) and resting state FC, (6) com-
bination of (4) and resting state FC using multislice community detection method
[33].
3.1 Materials
We used the resting state fMRI and task fMRI scans of 77 unrelated healthy subjects
from the Human Connectome Project (HCP) dataset [34]. Two sessions of resting
state fMRI with 30 minutes for each session, and 7 sessions of task fMRI data were
available for multisource integration. The seven tasks are working memory (total
time: 10:02), gambling (6:24), motor (7:08), language (7:54), social cognition (6:54),
relational processing (5:52) and emotion processing (4:32). Preprocessing already
applied to the HCP fMRI data includes gradient distortion correction, motion cor-
rection, spatial normalization to Montreal Neurological Institute (MNI) space with
nonlinear registration based on a single spline interpolation, and intensity normal-
ization [35]. Additionally, we regressed out motion artifacts, mean white matter
and cerebrospinal fluid confounds, and principal components of high variance voxels
using compCor [36]. Next, we applied a bandpass filter with cutoff frequencies of
0.01 and 0.1 Hz for resting state fMRI data. For task fMRI data, we performed
similar temporal processing, except a high-pass filter at 1/128 Hz was used. The
data were further demeaned and normalized by the standard deviation. We then
used the Harvard-Oxford (HO) atlas [37], which has 112 region of interest (ROI)s,
10
to define the brain region nodes. We chose the well-established HO atlas because
it sampled from every major brain system, and consists of the highest number of
subjects with both manual and automatic labelling technique compared to other
commonly used anatomical atlases. Voxel time courses within ROIs were averaged
to generate region time courses. The region time courses were demeaned, normalized
by the standard deviation. Group level time courses were generated by concatenat-
ing the time courses across subjects. The Pearson's correlation values between the
region time courses were taken as estimates of FC matrices. Negative elements in all
connectivity matrices were set to zero due to the currently unclear interpretation of
negative connectivity [38]. For task activation, we applied the activation detection
on the seven tasks available following the steps described in subsection 2.3.
We summarize here the annotation of the graphs for six methods being evaluated
for subnetwork extraction. (1) Resting state FC matrix Crest is used. (2) The task
general FC Ctask was generated by concatenating the time courses across all tasks
In (3), we use task-specific FC in Equation 9
before the Pearson's correlation.
and Equation 10 for each hyperedge, denoted as Chyper-task. We implement (4) by
using resting state FC in Equation 9 and Equation 10 as described in subsection 2.5,
denoted as Chyper-t-r. For (5), we first generate ¯Ctask
ij by using task-specific FC as Cek
ij ,
and resting state Crest to compute w(ek) based on Equation 9 and Equation 10. We
next applied our proposed local thresholding [39] on resting state FC Crest to match
with the graph density of ¯Ctask at 0.2765, which lies within the normal range of
thresholding before subnetwork extraction between [0.2, 0.3] [3]. We then estimate
Ct-r using Equation 13. We set free parameters w1, w2, w3 to one, and the stat
to median value based on inner cross-validation. For (6), we generated the ¯Ctask
and thresholded Crest as the same way as in (5), then the multisource integration
is implemented using a multislice approach [33], denoted as Ct-r-multislice. We set
ij
the weighting for multisource integration γ or coupling parameter ω from 0.01 to
1 at an interval of 0.01. In order to perform fair comparison, Crest in method (1)
and Ctask in method (2) have also been local thresholded at the graph density of
0.2765. Method (1) to (5) used Ncuts and (6) used generalized Louvain as the
graph partitioning approach. The number of subnetworks was set to seven given
that there are seven tasks available to examine if subnetwork assignments can be
related to tasks. We note that setting the number of subnetworks is non-trivial
as discussed in the previous section that we leave as future work. All statistical
comparisons are based on the Wilcoxon signed rank test with significance declared
at an α of 0.05 with Bonferroni correction.
11
3.2 Similarity of FC between Resting state and Task data
We observed a similarity at Dice Similarity Coefficient (DSC) = 0.7845 between
resting state FC and task general FC, which was generated by concatenating the
time courses across all different tasks. For seven specific tasks, the corresponding
DSC between task-specific FC and task general FC are 0.8971 for emotion process-
ing, 0.8557 gambling, 0.8676 for language, 0.9043 for motor, 0.8594 for relational
processing, 0.8307 for social cognition, and 0.8751 for working memory. This high
similarities confirms the findings in [8] that a set of small but consistent changes
common across tasks suggests the existence of a task-general network architecture
distinguishing task states from rest.
When resting state FC is compared to task-specific FC, the DSC are 0.7193
for emotion processing, 0.7689 for gambling, 0.7390 for language, 0.7067 for motor,
0.7533 for relational processing, 0.7659 for social cognition and 0.7118 for working
memory, respectively. The variation of similarities between task-specific and resting
state FC around a relatively high average level further confirms that the brain's
functional network architecture during task is configured primarily by an intrinsic
network architecture which can be present during rest, and secondarily by changes
in evoked task-general (common across tasks) and task-specific network [8].
These findings confirms the close relationship between task and rest, and the
support for integrating multitask information into resting state based subnetwork
extraction.
3.3 Modularity Q Value
Modularity Q value has been used to assess a graph partitioning through reflecting
the intra- and inter- subnetwork connection structure of a network [9]. We observe
that Q values of group level subnetwork extraction for method (1)-(6) are 0.1401,
0.1282, 0.1624, 0.1711, 0.2290 and 0.1905 when γ and ω were selected at the highest
inter-subject reproducibility.
At the subject-wise level, the modularity Q values estimated from the subnetwork
extraction using method (1)-(6) are 0.1397±0.0142, 0.1234±0.0159, 0.2072 ± 0.0199,
0.2094±0.0189, 0.2183±0.0192, and 0.2089±0.0165 respectively, Figure 2.
We show that the modularity estimated from subnetworks extracted based on
simply concatenating task time courses is lower than using resting state data. Using
hypergraph framework (3) Chyper-task and (4) Chyper-t-r achieves statistically higher
modularity values than using either resting state data or simple concatenation of
task data. Moreover, incorporating resting state information into the hypergraph
framework (5) Ct-r can increase modularity compared to hypergraph method. Mul-
tislice integration (6) Ct-r-slice results in a lower modularity than (5) the linear model;
12
Figure 2: Subject-wise level modularity Q values using Method (1)-(6). For method
(5) and (6), parameter γ and ω were selected at the highest inter-subject repro-
ducibility.
however, it still outperforms all the other uni-source methods. Overall, incorporat-
ing resting state information explicitly using a weighted combination strategy, i.e.,
method (5) gives a statistically higher modularity than all contrasted methods at
p < 10−4 based on Wilcoxon signed rank test. We note that the Q values derived
here are around 0.2, when the number of the subnetworks was set to seven, i.e., the
number of tasks. It is relatively low due to the inherent resolution limit of Q, i.e.,
Q decreases when the number of subnetworks increases. We explored this direc-
tion by achieving the similar level of Q values around 0.3-0.4 when the number of
subnetworks decreases to 4 as in [13].
3.4 Inter-subject Reproducibility of Subnetwork Extraction
We assessed the inter-subject reproducibility by comparing the subnetwork extrac-
tion results using subject-wise data against the group level data. The average
DSC between subject-wise and group level subnetworks across 77 subjects based on
methods (1)-(6) are 0.6362±0.0828, 0.5704±0.0872, 0.7083±0.1094, 0.7258±0.1201,
0.7561±0.1199, and 0.7406±0.0725, Figure 3. We noticed that the reproducibility
using resting state FC Crest is higher than simple concatenation of task time courses
data Ctask. It could be that there exist great differences in reaction to stimuli from
different subjects, and simple concatenation is hard to discover the higher order
relationship between canonical network components. On the other hand, analyz-
ing multi-task information using hypergraph (3) Chyper-task achieved much higher
stability in subnetwork extraction, and incorporating resting information implicitly
within the hypergraph (4) Chyper-t-r, or explicit weighted combination (5) Ct-r can
even further enhance reproducibility. We note that the weighted combination out-
performs multislice integration (6) Ct-r-slice, which is still better than all the other
13
(1)Crest(2)Ctask(3)Chyper-task(4)Chyper-t-r(5)Ct-r(6)Ct-r-multislice00.050.10.150.20.250.30.35 Figure 3: Subject-wise level inter-subject reproducibility of subnetwork extraction
using Method (1)-(6). For method (5) and (6), parameter γ and ω were selected at
the highest inter-subject reproducibility.
uni-source methods. The reason could be that a simple linear model suffices the
fusion of task and rest data. Overall, the inter-subject reproducibility derived by
(5) Ct-r is statistically higher than all contrasted methods at p < 10−4 based on
Wilcoxon signed rank test.
3.5 Biological Meaning
We next examined the biological meaning of the subnetworks extracted from method
(1) - (6), where γ was set to 0.5 to report the results when resting state and hy-
pergraph based multitask information are equally combined as an example. Seven
subnetworks were extracted based on the number of tasks available. Method (1) de-
tects most of the traditional resting state subnetworks with several false positive and
negative detection. The results of method (2) oftentimes combined some important
regions from different subnetworks, which lacks biological justifications. Method (3)
and (4) generate similar results and both improve the results of method (2) greatly
when bringing task dynamics into the subnetwork extraction. Overall, method (5)
detects brain regions, which are more biologically meaningful, by combining the in-
trinsic network architecture from resting state data and the task dynamics based on
high-order hypergraph. We report our findings in details as the following and the
visualization of subnetwork extraction results can be found in Figure 4.
Using method (1) based on resting state FC alone, subnetwork 1 and 6 are de-
tected as left and right side of a combination of Executive Control Network (ECN)
and frontoparietal network, which include superior frontal gyrus, middle frontal
gyrus, inferior frontal gyrus, posterior supramarginal gyrus, angular gyrus, frontal
orbital cortex, and frontal operculum cortex. Method (1) mistakenly classified left
inferior lateral occipital cortex and left anterior supramarginal gyrus into the Left
14
(1)Crest(2)Ctask(3)Chyper-task(4)Chyper-t-r(5)Ct-r(6)Ct-r-multislice0.30.40.50.60.70.80.91 (a) Method (1) Crest
(b) Method (2) Ctask
(c) Method (3) Chyper-task
(d) Method (4) Chyper-t-r
(e) Method (5) Ct-r
(f) Method (6) Ct-r-slice
Figure 4: Visualization of subnetworks extraction using methods (1)-(6). The mass
center of each ROI is plotted in the MNI space and colorcoded by the membership
of seven subnetworks.
15
Executive Control Network (LECN). Anterior supramarginal gyrus is part of the so-
matosensory association cortex, which interprets tactile sensory data and is involved
in perception of space and limbs location or language processing, thus it should be
included in Default Mode Network (DMN) instead of ECN [40]. On the other hand,
our proposed method (5) detects both the left and right sides of most of the anterior
portion of ECN and posterior supramarginal gyri for subnetwork 1. Using method
(5), the left inferior lateral occipital cortex was not include in ECN, which is more
accurate. Besides, method (5) clustered anterior supramarginal gyrus symmetrically
into subnetwork 6, which includes both sides of Posterior Cingulate Cortex (PCC),
precuneus, and angular gyrus, comprising most of the posterior portion of DMN de-
fined in [40]. As for method (2), the simple concatenation of multitask time courses,
subnetwork 1 consists of frontal medial cortex and only the left side of frontal orbital
cortex, and subnetwork 6 consists of most of the anterior portion of ECN, angular
gyrus and only the left posterior supramarginal gyrus, which should be symmetri-
cally included in DMN. Besides, there are two other ROIs, left subcallosal cortex
and left caudate, included in subnetwork 6, which lacks biological meaning. Subnet-
work 1 derived from method (3) and (4) both consist of most of the anterior portion
of ECN, except that method (3) has two more one-sided frontal areas, which makes
(4) more biological meaningful (with symmetric results). Subnetwork 6 of method
(3) and (4) both consist of one isolate area:
left anterior parahippocampal gyrus,
which further indicates that there is need to incorporate resting state information
into the multitask based on hypergraph framework.
Subnetwork 2 of method (1) includes both sides of Anterior Cingulate Cortex
(ACC), caudate, thalamus, putamen and accumbens. Method (5) includes all the
same brain regions as method (1) plus one other region, the insula. This subnetwork
should be related to the gambling task and emotional processing, which expect to
activate ACC [41, 42], ventral striatum (such as thalamus [42] and accumbens [43]),
and insula [44]. Usually insula is part of the salience network and has been found to
play key roles in emotional processing [45]. However, using method (1), the insula
was clustered into subnetwork 5 (mostly motor system). Method (2) included right
ACC and both sides of PCC, precuneous, left side of supracalcarine cortex, and
accumbens inside subnetwork 2, which seems like a mixture of part of DMN, one-
sided region from motor system, and one region from gambling system. As for
method (3) and (4), they both extracted similar regions for subnetwork 2 as using
method (5), except that they missed thalamus and falsely included left frontal medial
cortex.
Subnetwork 3 derived from method (1) includes superior lateral occipital cortex,
frontal medial cortex, left subcallosal cortex, PCC, precuneous, parahippocampal
gyrus, temporal fusiform cortex, brain stem, hippocampus and amygdala. This as-
16
signment does not make too much sense by clustering regions from visual, auditory,
emotion circuit and frontal system together. Meanwhile, the results using method
(5) consists mostly of emotion circuit and social processing, which includes brain
stem [46], hippocampus and parahippocampal gyrus [47], amygdala [48], and sub-
callosal cortex [49]. Method (5) also detected regions related to auditory functions
such as temporal pole, which is reasonable since the negative emotion was induced
by listening to stories. Subnetwork 3 detected by method (2) includes right anterior
parahippocampal gyrus, temporal fusiform cortex and brain stem, which still lacks
important brain regions in the emotion circuit. Method (3) detects more biologically
meaningful regions than (2), such as hippocampus and amygdala. Using method (4)
can even detect more related regions than method (3), such as frontal orbital cortex
[50].
Method (1) and (5) detected almost the same brain regions for subnetwork 4,
which is the visual system, except that method (5) detected one more region of the
inferior lateral occipital cortex, making the results more symmetric. This subnet-
work includes inferior lateral occipital cortex, intracalcarine cortex, cuneal cortex,
lingual gyrus, occipital fusiform gyrus, temporal occipital fusiform cortex, occipital
pole, and supracalcarine cortex. Method (2) detected most of the visual regions
except for cuneal cortex and the right supracalcarine cortex. Method (3) and (4)
detected extra regions in right ECN and auditory system besides all the regions
found using (5) in the visual system.
Subnetwork 5 derived from method (1) comprises of the motor system, including
precentral gyrus, postcentral gyrus, only the right side of anterior supramarginal
gyrus, juxtapositional lobule cortex; and the frontoparietal network including left
central opercular cortex, superior parietal lobule, and parietal operculum cortex.
Method (5) generated similar results as method (1), only that the results are more
symmetric, which include both sides of anterior supramarginal gyrus (part of so-
matosensory association cortex); and more accurate in terms of frontoparietal net-
work, which includes frontal operculum cortex instead of central opercular cortex.
Both method (3) and (4) generated similar regions for subnetwork 5 as well, which
includes motor system and frontoparietal network, except that they both included
brain stem into this subnetwork. However, method (2) mis-classified insula, puta-
men and thalamus into the motor and frontal parietal networks. We note that the
motor system and frontoparietal network are clustered together, it could be that the
working memory tasks recruited both the motor system and frontoparietal network.
As for the subnetwork 7, both method (1) and (5) detected brain regions cor-
responding to language task and related auditory regions, such as anterior superior
temporal gyrus, planum temporale, planum polare, and Heschls gyrus (includes H1
and H2) [51]. Different from method (1), method (5) included central opercular
17
cortex, which can be explained by how fronto-opercular is related to language [52].
Method (2) detected some false positive brain regions in the language system such as
parahippocampal gyrus, hippocampus and amygdala. Method (3) and (4) correctly
clustered all the brain regions into the language network as method (5).
Method (6) generated similar results compared to method (5), only a couple
regions in subnetworks 2 and 5 were switched, a couple regions in subnetwork 6
and 7 were switched, and a couple regions in 1 and 6 were switched. Overall, The
subnetwork results derived by method (5) Ct-r have more biological meaning than
contrasted methods.
4 Discussion
4.1 Hypergraph encodes higher order nodal relationship
Subnetwork results derived from methods based on hypergraph achieved higher mod-
ularity, higher inter-subject reproducibility, and more reasonable biological meaning
than traditional connectivity analysis of pairwise correlation between nodes. These
results indicate that hypergraph, which is a natural presentation of multitask acti-
vation, can be explored to study higher order relations among the network nodes.
The proposed strength informed version of automatic weight setting of the hyper-
edge incorporates connectivity information to reveal more accurate higher order
relationship among nodes rather than just using binary information.
4.2 Multisource Integration Improves Subnetwork Extrac-
tion
We have proved that multisource integration of task and rest information can im-
prove subnetwork extraction compared to using a single source in terms of graphical
metrics, inter-subject reproducibility, along with biologically meaningful subnetwork
assignments. We note that the implicit integration of rest information into multi-
task hypergraph achieved less improvements as the explicit integration based on the
linear combination. The reason could be that the limited number of tasks available
restricts the comprehensive representation of the brain using the hypergraph. Thus,
by integrating rest data to compensate possible missing information resulted in over-
all better outcomes. Another observation is that the linear combination outperforms
the multislice community detection, which still performs better than uni-source ap-
proaches. Our assumption is that rest and task FC are both derived from a single
functional modality, which complements each other by revealing the two sides of
FC, i.e., the resting intrinsic side and the activated evoked side. Thus, a simple
18
linear weighted combination would suffice this situation, which outperforms other
alternative combination approach in practice.
4.3 Limitations and Future Directions
There are several limitations in our present work. First, our study investigated only
seven available tasks with high quality data and decent amount of data per task.
This sample of seven tasks is not enough. A possible solution is to have access to
both task and rest data from previous task studies or co-activation studies, which
covers much wider variety of tasks. At the same time, with much more information
from a greater amount of task data, we can devise a reliable automatic manner to
determine the integration weighting parameter γ. The underlying rationale is that
with more tasks available, we can rely more on the hypergraph based multitask
source, hence the higher γ.
Secondly, we set the number of the subnetworks to be seven, which corresponds
to the number of tasks available. The reason is simply to see if we can associate the
subnetwork results to different tasks and gain insights from the findings based on
task-induced functions. In the future, a finer scale of subnetwork extraction using
multi-scale hierarchical approach would improve the interpretation of the findings.
5 Conclusion
We proposed a high order relation informed approach based on hypergraph to com-
bine the information from multi-task data and resting state data to improve subnet-
work extraction. We demonstrated that fusing task activation, task-induced con-
nectivity and resting state functional connectivity based on hypergraphs improves
subnetwork extraction compared to employing a single source from either rest or
task data in terms of subnetwork modularity measure, inter-subject reproducibility,
along with more biologically meaningful subnetwork assignments.
6 List of Acronyms
ACC Anterior Cingulate Cortex
DMN Default Mode Network
DSC Dice Similarity Coefficient
ECN Executive Control Network
FC Functional Connectivity
19
fMRI Functional Magnetic Resonance Imaging
GLM General Linear Model
HCP Human Connectome Project
HO Harvard-Oxford
ICA Independent Component Analysis
LECN Left Executive Control Network
MNI Montreal Neurological Institute
Ncuts Normalized cuts
PCC Posterior Cingulate Cortex
ROI region of interest
rs-fcMRI Resting State Functional Connectivity based on MRI
SNR signal-to-noise ratio
References
[1] Park, B., Kim, D.S., Park, H.J.: Graph independent component analysis reveals
repertoires of intrinsic network components in the human brain. PloS one 9(1)
(2014) e82873
[2] Girvan, M., Newman, M.E.: Community structure in social and biological
networks. Proceedings of the National Academy of Sciences 99(12) (2002)
7821 -- 7826
[3] Van Den Heuvel, M., Mandl, R., Pol, H.H.: Normalized cut group clustering
of resting-state fmri data. PloS one 3(4) (2008) e2001
[4] Nicolini, C., Bifone, A.: Modular structure of brain functional networks: break-
ing the resolution limit by surprise. Scientific reports 6 (2016)
[5] Murphy, K., Birn, R.M., Bandettini, P.A.: Resting-state fmri confounds and
cleanup. Neuroimage 80 (2013) 349 -- 359
[6] Smith, S.M., Fox, P.T., Miller, K.L., Glahn, D.C., Fox, P.M., Mackay, C.E.,
Filippini, N., Watkins, K.E., Toro, R., Laird, A.R., et al.: Correspondence of
the brain's functional architecture during activation and rest. Proceedings of
the National Academy of Sciences 106(31) (2009) 13040 -- 13045
20
[7] Fox, M.D., Snyder, A.Z., Vincent, J.L., Raichle, M.E.:
Intrinsic fluctuations
within cortical systems account for intertrial variability in human behavior.
Neuron 56(1) (2007) 171 -- 184
[8] Cole, M.W., Bassett, D.S., Power, J.D., Braver, T.S., Petersen, S.E.: Intrinsic
and task-evoked network architectures of the human brain. Neuron 83(1) (2014)
238 -- 251
[9] Sporns, O., Betzel, R.F.: Modular brain networks. Annual review of psychology
67 (2016) 613 -- 640
[10] Tavor, I., Jones, O.P., Mars, R., Smith, S., Behrens, T., Jbabdi, S.: Task-free
mri predicts individual differences in brain activity during task performance.
Science 352(6282) (2016) 216 -- 220
[11] Chan, M.Y., Alhazmi, F.H., Park, D.C., Savalia, N.K., Wig, G.S.: Resting-
state network topology differentiates task signals across the adult life span.
Journal of Neuroscience 37(10) (2017) 2734 -- 2745
[12] Mennes, M., Kelly, C., Zuo, X.N., Di Martino, A., Biswal, B.B., Castellanos,
F.X., Milham, M.P.:
Inter-individual differences in resting-state functional
connectivity predict task-induced bold activity. Neuroimage 50(4) (2010) 1690 --
1701
[13] Crossley, N.A., Mechelli, A., V´ertes, P.E., Winton-Brown, T.T., Patel, A.X.,
Ginestet, C.E., McGuire, P., Bullmore, E.T.: Cognitive relevance of the com-
munity structure of the human brain functional coactivation network. Proceed-
ings of the National Academy of Sciences 110(28) (2013) 11583 -- 11588
[14] Bertolero, M.A., Yeo, B.T., D'Esposito, M.: The modular and integrative func-
tional architecture of the human brain. Proceedings of the National Academy
of Sciences 112(49) (2015) E6798 -- E6807
[15] Bassett, D.S., Sporns, O.: Network neuroscience. Nature neuroscience 20(3)
(2017) 353
[16] Laird, A.R., Fox, P.M., Eickhoff, S.B., Turner, J.A., Ray, K.L., McKay, D.R.,
Glahn, D.C., Beckmann, C.F., Smith, S.M., Fox, P.T.: Behavioral interpre-
tations of intrinsic connectivity networks. Journal of cognitive neuroscience
23(12) (2011) 4022 -- 4037
[17] Mesulam, M.M.: From sensation to cognition. Brain: a journal of neurology
121(6) (1998) 1013 -- 1052
21
[18] De Domenico, M.: Multilayer modeling and analysis of human brain networks.
Giga Science 6(5) (2017) 1 -- 8
[19] Muldoon, S.F., Bassett, D.S.: Network and multilayer network approaches
to understanding human brain dynamics. Philosophy of Science 83(5) (2016)
710 -- 720
[20] De Domenico, M., Sasai, S., Arenas, A.: Mapping multiplex hubs in human
functional brain networks. Frontiers in neuroscience 10 (2016)
[21] Zhou, D., Huang, J., Scholkopf, B.: Learning with hypergraphs: Clustering,
classification, and embedding. In: Advances in neural information processing
systems. (2007) 1601 -- 1608
[22] Dotko, P., Hess, K., Levi, R., Nolte, M., Reimann, M., Scolamiero, M., Turner,
K., Muller, E., Markram, H.: Topological analysis of the connectome of digital
reconstructions of neural microcircuits. arXiv preprint arXiv:1601.01580 (2016)
[23] Zu, C., Gao, Y., Munsell, B., Kim, M., Peng, Z., Zhu, Y., Gao, W., Zhang,
D., Shen, D., Wu, G.: Identifying high order brain connectome biomarkers via
learning on hypergraph. In: International Workshop on Machine Learning in
Medical Imaging, Springer (2016) 1 -- 9
[24] Munsell, B.C., Wu, G., Gao, Y., Desisto, N., Styner, M.:
Identifying rela-
tionships in functional and structural connectome data using a hypergraph
learning method. In: International Conference on Medical Image Computing
and Computer-Assisted Intervention, Springer (2016) 9 -- 17
[25] Von Luxburg, U.: A tutorial on spectral clustering. Statistics and computing
17(4) (2007) 395 -- 416
[26] Stoer, M., Wagner, F.: A simple min-cut algorithm. Journal of the ACM
(JACM) 44(4) (1997) 585 -- 591
[27] Hagen, L., Kahng, A.B.: New spectral methods for ratio cut partitioning and
clustering. IEEE transactions on computer-aided design of integrated circuits
and systems 11(9) (1992) 1074 -- 1085
[28] Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions
on pattern analysis and machine intelligence 22(8) (2000) 888 -- 905
[29] Friston, K.J., Holmes, A.P., Worsley, K.J., Poline, J.P., Frith, C.D., Frack-
owiak, R.S.:
Statistical parametric maps in functional imaging: a general
linear approach. Human brain mapping 2(4) (1994) 189 -- 210
22
[30] Ng, B., Hamarneh, G., Abugharbieh, R.: Modeling brain activation in fmri
using group mrf. IEEE transactions on medical imaging 31(5) (2012) 1113 --
1123
[31] Nichols, T., Hayasaka, S.: Controlling the familywise error rate in functional
neuroimaging: a comparative review. Statistical methods in medical research
12(5) (2003) 419 -- 446
[32] Yoldemir, B.: Multimodal fusion for assessing functional segregation and inte-
gration in the human brain. PhD thesis, University of British Columbia (2016)
[33] Mucha, P.J., Richardson, T., Macon, K., Porter, M.A., Onnela, J.P.: Commu-
nity structure in time-dependent, multiscale, and multiplex networks. science
328(5980) (2010) 876 -- 878
[34] Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugur-
bil, K., Consortium, W.M.H., et al.: The wu-minn human connectome project:
an overview. Neuroimage 80 (2013) 62 -- 79
[35] Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fischl, B., An-
dersson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J.R., et al.: The
minimal preprocessing pipelines for the human connectome project. Neuroim-
age 80 (2013) 105 -- 124
[36] Behzadi, Y., Restom, K., Liau, J., Liu, T.T.: A component based noise correc-
tion method (compcor) for bold and perfusion based fmri. Neuroimage 37(1)
(2007) 90 -- 101
[37] Desikan, R.S., S´egonne, F., Fischl, B., Quinn, B.T., Dickerson, B.C., Blacker,
D., Buckner, R.L., Dale, A.M., Maguire, R.P., Hyman, B.T., et al.: An auto-
mated labeling system for subdividing the human cerebral cortex on mri scans
into gyral based regions of interest. Neuroimage 31(3) (2006) 968 -- 980
[38] Skudlarski, P., Jagannathan, K., Calhoun, V.D., Hampson, M., Skudlarska,
B.A., Pearlson, G.: Measuring brain connectivity: diffusion tensor imaging
validates resting state temporal correlations. Neuroimage 43(3) (2008) 554 --
561
[39] Wang, C., Ng, B., Abugharbieh, R.: Modularity reinforcement for improving
brain subnetwork extraction. In: International Conference on Medical Image
Computing and Computer-Assisted Intervention, Springer (2016) 132 -- 139
23
[40] Heinonen, J., Numminen, J., Hlushchuk, Y., Antell, H., Taatila, V., Suomala,
J.: Default mode and executive networks areas: Association with the serial
order in divergent thinking. PloS one 11(9) (2016) e0162234
[41] Charpentier, C.J., Martino, B.D., Sim, A.L., Sharot, T., Roiser, J.P.: Emotion-
induced loss aversion and striatal-amygdala coupling in low-anxious individuals.
Social cognitive and affective neuroscience 11(4) (2015) 569 -- 579
[42] Koehler, S., Ovadia-Caro, S., van der Meer, E., Villringer, A., Heinz, A.,
Romanczuk-Seiferth, N., Margulies, D.S.:
Increased functional connectivity
between prefrontal cortex and reward system in pathological gambling. PLoS
One 8(12) (2013) e84565
[43] Limbrick-Oldfield, E.H., Mick, I., Cocks, R., McGonigle, J., Sharman, S., Gold-
stone, A.P., Stokes, P., Waldman, A., Erritzoe, D., Bowden-Jones, H., et al.:
Neural substrates of cue reactivity and craving in gambling disorder. Transla-
tional psychiatry 7(1) (2017) e992
[44] Leong, J.K., Pestilli, F., Wu, C.C., Samanez-Larkin, G.R., Knutson, B.: White-
matter tract connecting anterior insula to nucleus accumbens correlates with
reduced preference for positively skewed gambles. Neuron 89(1) (2016) 63 -- 69
[45] Cauda, F., D'agata, F., Sacco, K., Duca, S., Geminiani, G., Vercelli, A.: Func-
tional connectivity of the insula in the resting brain. Neuroimage 55(1) (2011)
8 -- 23
[46] Venkatraman, A., Edlow, B.L., Immordino-Yang, M.H.: The brainstem in
emotion: A review. Frontiers in neuroanatomy 11 (2017)
[47] Ohmura, Y., Izumi, T., Yamaguchi, T., Tsutsui-Kimura, I., Yoshida, T., Yosh-
ioka, M.: The serotonergic projection from the median raphe nucleus to the
ventral hippocampus is involved in the retrieval of fear memory through the
corticotropin-releasing factor type 2 receptor. Neuropsychopharmacology 35(6)
(2010) 1271
[48] Zald, D.H.: The human amygdala and the emotional evaluation of sensory
stimuli. Brain Research Reviews 41(1) (2003) 88 -- 123
[49] Laxton, A.W., Neimat, J.S., Davis, K.D., Womelsdorf, T., Hutchison, W.D.,
Dostrovsky, J.O., Hamani, C., Mayberg, H.S., Lozano, A.M.: Neuronal cod-
ing of implicit emotion categories in the subcallosal cortex in patients with
depression. Biological psychiatry 74(10) (2013) 714 -- 719
24
[50] Levens, S.M., Devinsky, O., Phelps, E.A.: Role of the left amygdala and right
orbital frontal cortex in emotional interference resolution facilitation in working
memory. Neuropsychologia 49(12) (2011) 3201 -- 3212
[51] Noesselt, T., Shah, N.J., Jancke, L.: Top-down and bottom-up modulation of
language related areas -- an fmri study. BMC neuroscience 4(1) (2003) 13
[52] Meyer, M., Alter, K., Friederici, A.: Functional mr imaging exposes differential
brain responses to syntax and prosody during auditory sentence comprehension.
Journal of Neurolinguistics 16(4) (2003) 277 -- 300
25
|
1503.07552 | 1 | 1503 | 2015-03-25T20:54:56 | Making a swim central pattern generator out of latent parabolic bursters | [
"q-bio.NC",
"nlin.PS"
] | We study the rhythmogenesis of oscillatory patterns emerging in network motifs composed of inhibitory coupled tonic spiking neurons represented by the Plant model of R15 nerve cells. Such motifs are argued to be used as building blocks for a larger central pattern generator network controlling swim locomotion of sea slug Melibe leonina. | q-bio.NC | q-bio | September 25, 2018
10:41 parabolicbursters
International Journal of Bifurcation and Chaos
© World Scientific Publishing Company
5
1
0
2
r
a
M
5
2
]
.
C
N
o
i
b
-
q
[
1
v
2
5
5
7
0
.
3
0
5
1
:
v
i
X
r
a
MAKING A SWIM CENTER PATTERN GENERATOR OUT
OF LATENT PARABOLIC BURSTERS
Department of Mathematics and Statistics, Georgia State University, Atlanta 30303, USA.
DENIZ ALAC¸ AM
Email: [email protected].
ANDREY SHILNIKOV
Neuroscience Institute and Department of Mathematics and Statistics, Georgia State University, Atlanta
Institute for Information Technology, Mathematics and Mechanics, Lobachevsky State University of
Nizhni Novgorod, Nizhni Novgorod, 603950, Russia.
30303, USA. Email: [email protected].
Received (to be inserted by publisher)
We study the rhythmogenesis of oscillatory patterns emerging in network motifs composed of
inhibitory coupled tonic spiking neurons represented by the Plant model of R15 nerve cells. Such
motifs are argued to be used as building blocks for a larger central pattern generator network
controlling swim locomotion of sea slug Melibe leonina.
Keywords: Plant model, parabolic bursting, half-center oscillation, central pattern generator,
swim locomotion, see slug
1. Introduction
A plethora of vital rhythmic motor behaviors, such as heartbeat, respiratory functions and locomotion are
produced and governed by neural networks called central pattern generators (CPGs) [Selverston, 1985; Bal
et al., 1988; Marder & Calabrese, 1996; Frost & Katz, 1996; Kristan et al., 2005; Katz & Hooper, 2007].
A CPG is a microcircuit of interneurons whose mutually synergetic interactions autonomously generate
an array of multi-phase bursting rhythms underlying motor behaviors. There is a growing consensus in
the community of neurophysiologists and computational researchers that some basic structural and func-
tional elements must be shared by CPGs in invertebrate and vertebrate animals. As such, we should first
understand these elements, find the universal principles, and develop efficient mathematical and computa-
tional tools for plausible and phenomenological models of CPG networks. Pairing experimental studies and
modeling studies has proven to be key to unlocking insights into operational and dynamical principles of
CPGs [Gillner & Wallen, 1985; Kopell & Ermentrout, 2004; Matsuoka, 1987; Kopell, 1988; Canavier et al.,
1994; Skinner et al., 1994a; Dror et al., 1999; Prinz et al., 2003a]. Although various circuits and models of
specific CPGs have been developed, it still remains unclear what makes the CPG dynamics so robust and
flexible [Best et al., 2005; Belykh & Shilnikov, 2008; Sherwood et al., 2010; Koch et al., 2011; Calabrese
et al., 2011; Marder, 2012]. It is also unclear what mechanisms a multi-functional motor system can use
to generate polyrhythmic outcomes to govern several behaviors [Kristan, 2008; Briggman & Kristan, 2008;
Wojcik et al., 2014]. Our goal is to gain insight into the fundamental and universal rules governing pattern
formation in complex networks of neurons. To achieve this goal, we should identify the rules underlying
1
September 25, 2018
10:41 parabolicbursters
2 Ala¸cam and Shilnikov
(a)
(b)
Fig. 1.
(a) Melibe leonina lateral swim style. (b) Network bursting in swim interneurons (Si) of the Melibe swim CPG halts
when Si3R is hyperpolarized, thus its counterpart Si3L begins tonic spiking; the photographs and in-vitro recording provided
courtesy of A. Sakurai [Sakurai et al., 2014]
the emergence of cooperative rhythms in simple CPG networks.
Recently, a great deal of computational studies have been focused on a range of 3-cell motifs of bursting
neurons coupled by chemical (inhibitory and excitatory) and electrical synapses to disclose the role of
coupling in generating sets of coexisting rhythmic outcomes, see [Shilnikov et al., 2008; Wojcik et al., 2011,
2014; Schwabedal et al., 2014, 2015; Collens et al., 2015] and references therein. These network structures
reflect the known physiological details of various CPG networks in real animals. Next, we would like to
explore dynamics and stability of some identified CPG circuits constituted by 4-cells [Jalil et al., 2013].
Examples of such sub-networks can be found in the cerebral crustacean stomatogastric ganglion (STG)
[Selverston, 1985; Prinz et al., 2003b, 2004; Marder, 2012], as well as in the swim CPGs of the sea slugs
-- Melibe leonina (depicted during swimming in Fig. 1(a)) and Dendronotus iris [Newcomb et al., 2012;
Sakurai et al., 2011; Sakurai & Katz, 2011]. Our greater goal is to create dynamical foundations for the
onset, morphogenesis and structural robustness of rhythmic activity patterns produced by swim CPGs in
these animals. A pilot mathematical model of the Melibe swim CPG will be discussed in this paper. The
circuitry shown in Fig. 2(a) depicts only some core elements identified in the biological CPG; its detailed
diagram can be found in [Sakurai et al., 2014].
(a)
(b)
(a) A core circuitry of the biological Melibe swim CPG with inhibitory (●), excitatory (▼) and electrical ((cid:131)~(cid:131)~) synapses
Fig. 2.
[Sakurai et al., 2014]. (b) In-vitro voltage activity recordings from identified swim interneurons, Si2L and Si3L/R, of the Melibe
swim CPG with the characteristic 3
4 -phase lag between the HCO2 and HCO3; intracellular recording provided courtesy of
A. Sakurai [Sakurai et al., 2014].
Being inspired by experimental studies of voltage activity recorded from the swim CPGs of the sea
slugs Melibe leonina and Dendronotus iris, we would like to develop an assembly line for CPG constructors
September 25, 2018
10:41 parabolicbursters
CPGs out of latent parabolic bursters
3
Fig. 3.
CPG. Recording provided courtesy of A. Sakurai and time series analysis by A. Kelley.
(a) Parabolic distribution of spike frequency within bursts produced by networked interneurons in the Melibe swim
made of coupled biophysically plausible models. Our first simplifying assumption is that CPGs are made
of universal building blocks -- half center oscillations (HCOs) [Hill et al., 2003]. Loosely speaking, a HCO
is treated as a pair of interneurons interacting with each other through reciprocally inhibitory synapses
and exhibiting anti-phase bursting. The interneurons of a HCO can be endogenous bursters, tonic spiking
or quiescent ones, which exhibit alternating bursting only when they inhibit each other. Theoretical stud-
ies [Wang & Rinzel, 1985] have indicated that formation of an anti-phase bursting rhythm is always based
on slow subsystem dynamics. There are three basic mechanisms to generate alternating bursting in the
HCO: release, escape, and post-inhibitory rebound (PIR). The first mechanism is typical for endogenously
bursting neurons [Jalil et al., 2010, 2012]. The other two mechanisms underlie network bursting in HCOs
comprised of neurons, which are hyperpolarized quiescent in isolation [Perkel & Mulloney, 1974; Skinner
et al., 1994b; Angstadt et al., 2005; Kopell & Ermentrout, 2002]. Our second assumption is that the swim
CPG interneurons are intrinsic tonic spikers that become network bursters only when externally driven
or coupled by inhibitory synapses, as recent experimental studies suggest [Sakurai et al., 2014]. The third
assumption is that network bursting in the Melibe swim CPG is parabolic, i.e. the spike frequency within a
burst increases at the middle, and decreases at the ends, as one can observe from Fig. 3. This observation
indicates the type of neuronal models to be employed to describe network cores. Our model of choice for
parabolic bursting is the Plant model [Plant & Kim, 1975, 1976; Plant, 1981]. The Plant model has been
developed to accurately describe the voltage dynamics of the R15 neuron in a mollusk Aplysia Californica,
which has turned out to be an endogenous burster [Levitan & Levitan, 1988]. Most dynamical properties
of the R15 neuron have been modeled and studied in detail [Canavier et al., 1991; Bertran, 1993; Butera
et al., 1995; Butera, 1998; Sieling & Butera, 2011; Ji et al., 2013].
2. Methods: the Plant model of parabolic bursting
The conductance based Plant model [Plant, 1981] for the R15 neuron [Sieling & Butera, 2011] located in
the abdominal ganglion of a slug Aplysia Californica is given by the following set of ordinary differential
equations derived within the framework of the Hodgkin-Huxley formalism to describe the dynamics of the
fast inward sodium [Na], outward potassium [K], slow TTX-resistant calcium [Ca] and an outward calcium
sensitive potassium [KCa] currents:
Cm V =−IN a− ICa− IKCa− Ileak− Iext− Isyn.
(1)
The last three currents are the generic ohmic leak Ileak, external constant Iext and synaptic Isyn currents
flowing from a pre-synaptic neuron. The full details of the representation of the currents employed in the
model are given in the Appendix below.
September 25, 2018
10:41 parabolicbursters
4 Ala¸cam and Shilnikov
(a)
(b)
(c)
There are two bifurcation parameters in the individual model. The first one is the constant external
introduced in the slowest equation:
Ca= ρ(Kcx(VCa− V + ∆)− Ca)
(2)
Fig. 4.
(a) Endogenous bursting in the Plant model as alternations of tonic spiking activity and quiescent periods. (b) Single
burst featuring a characteristic spike frequency increase in the middle of each burst. (c) Parabolic shape of the frequency
distribution of spikes within a burst is a feature of this kind of bursting. The parameters are ρ= 0.00015ms−1, Kc= 0.00425ms−1
and τx= 9400ms.
current, Iext, which is set Iext = 0. Following [Shilnikov, 2012], the other bifurcation parameter, ∆, is
deviation from a mean value of the reversal potential VCa = 140mV evaluated experimentally for the
At ∆= 0, the neuron is an endogenous burster, see Fig. 4. According to [Rinzel & Lee, 1987], this
calcium current in the R15 cells. As such, this makes ∆ a bifurcation parameter. Secondly its variations
are not supposed to alter the topology of the slow motion manifolds in the 5D phase space, which are called
tonic spiking and quiescent in the mathematical neuroscience context, as they are made of, respectively,
round periodic orbits and equilibrium states [of the slow subsystem] of the model (Fig. 5).
describing the concentration of the intracellular calcium in the Plant model. By construction, ∆ is a
type of bursting is termed parabolic. The reason for this term is that the spike frequency within bursts is
maximized in the middle of bursts and minimized at the beginning and the end (see Fig. 4c). The parabolic
structure of a burst is due to the calcium-activated potassium current. Its magnitude is determined by
the intracellular calcium concentration. As the intracellular calcium concentration increases, the calcium
Fig. 5. Bursting (green) orbit recursively switching between two slow -- motion critical manifolds: tonic spiking, Mlc, with
a characteristic fold and originating through a sub-critical Andronov-Hopf (AH) bifurcation from a depolarized equilibrium
state, and quiescent, Meq (orange curve), projected onto the (h, V ) and slow Ca variables of the of the Plant model; a plane
represents the synaptic threshold, Θsyn= 0mV .
September 25, 2018
10:41 parabolicbursters
CPGs out of latent parabolic bursters
5
Fig. 6. Responses of the bursting neuron (∆= 0mV ) on the synaptic drive Isyn= gsyn(V − Vrev). (a) Excitatory synaptic
drive with gsyn= 0.002nS and Vrev= 40mV applied at t= 80sec switches the neuron from bursting to tonic spiking activity. (b)
The inhibitory drive with gsyn= 0.005nS and Vsyn=−80mV halts bursting and makes the neuron hyperpolarized quiescent.
dependent potassium current gets activated, which causes an increase of the inward potassium current. As
the membrane potential increases over a threshold value, the intracellular calcium concentration decreases,
as well as the inward potassium current (see Eq. (20) in the Appendix). The parabolic distribution of
spikes within bursts is shown in Fig. 4. The instant frequency value is calculated by the reciprocal of each
inter-spike interval. Panels b and c of Fig. 4 clearly disclose the parabolic inter-spike structure of bursts.
It was shown in [Rinzel & Lee, 1987] that the mechanism underlying a transition between quiescent and
tonic spiking of bursting in the Plant model is due to a homoclinic bifurcation of a saddle-node equilibrium
state [Shilnikov, 1963; Afraimovich et al., 2014]. This bifurcation occurs in the fast 3D(V, h, n)-subspace
of the model and is modulated by the 2D slow dynamics in the(Ca, x)-variables, which are determined
by slow oscillations of the intracellular calcium concentration [Plant & Kim, 1975, 1976]. The unfolding
of this codimension-one bifurcation includes an onset of a stable equilibrium, which is associated with a
hyperpolarized phase of bursting, and on the other end, an emergent stable periodic orbit that is associated
with tonic spiking phase of bursting. The period of this stable orbit decreases, as it moves further away
from the saddle-node equilibrium mediated by decreasing calcium concentration. The period of the tonic
spiking orbit grows with no upper bound as it approaches the homoclinic loop of the saddle-node [Shilnikov
et al., 1998,2001].
Variations of ∆ change the duty cycle of bursting, which is a ratio of the active tonic spiking phase of
bursting to its period. Decreasing ∆ reduces the inactive, quiescent phase of bursting, i.e. increases its duty
cycle. Zero duty cycle is associated with the homoclinic saddle-node bifurcation that makes the neuron
hyperpolarized quiescent. This corresponds to an emergence of stable equilibrium state for all dynamical
variables of the model (1). In other words, decreasing ∆ makes the active phase longer, so that below a
threshold ∆=−32mV the neuron switches to tonic spiking activity. Tonic spiking activity is associated with
the emergence of a stable periodic orbit in the fast(V, h, n)-subspace, while the(Ca, x)-variables of the slow
due to relaxation of periodic oscillations in the 2D (Ca, x)-subspace, which slowly modulates fast tonic
spiking oscillations in the(V, h, n) variables. The relaxation limit cycles emerge from one and collapse into
the other equilibrium state in the(Ca, x)-plane through Andronov-Hopf bifurcations, which can be sub-
subspace converge to a stable equilibrium state. As such, bursting occurs in the Plant and similar models
or super-critical. At the transitions between bursting and tonic spiking, and bursting and hyperpolarized
quiescence, the neuron can produce chaotic dynamics, which are basically due to the membrane potential
oscillatory perturbations of plain canards at the folds of the relaxation cycle.
(a) (b) September 25, 2018
10:41 parabolicbursters
6 Ala¸cam and Shilnikov
Fig. 7. Tonic spiking neuron 1 at ∆=−34mV near the bifurcation transition between tonic spiking and bursting is forced to
syn= 0.001nS, from the pre-synaptic neuron 2 at
t= 60sec. Halting the inhibitory drive restores tonic spiking activity in the targeted neuron (not shown).
become a network burster with an application of an inhibitory drive with ginh
3. Endogenous and network bursting. Inhibitory and excitatory drives
A half-center oscillator is a network of two neurons coupled by reciprocally inhibitory synapses that robustly
produces bursting in alternation, or anti-phase bursting. Such a network can be multistable, i.e. produce
other bursting rhythms as well, such as synchronous bursting [Jalil et al., 2010] and rhythmic outcomes
with slightly shifted phase lags between the endogenously bursting neurons [Jalil et al., 2012].
In this study, the synaptic current Isyn is modeled through the fast threshold modulation (FTM)
approach [Kopell & Somers, 1993]. The synapses are assumed to be fast and non-delayed, which is true for
the swim CPG in both sea slugs under consideration. The synaptic current is given by
Isyn= gsyn(Vpost− Esyn)
1+ e−k(Vpre−Θsyn) ,
1
(3)
where gsyn is the maximal conductance of the current, which is used as a bifurcation parameter of the
networked model; Vpost(t) and Vpre(t) are the voltages on the post-synaptic (driven) and pre-synaptic
(driving) neurons; Esyn is the synaptic reversal potential. To make Isyn excitatory, we set Esyn= 40mV ,
while in the inhibitory case we set Esyn=−80mV . In Eq. (3), the second term is a Boltzmann coupling
function that quickly, (k= 100), turns the synaptic current on and off as soon the voltage, Vpre, of the
(driving) pre-synaptic cell(s) raises above and falls below the synaptic threshold, here Θsyn= 0mV (Fig. 5).
To model the constant synaptic drive onto the post-synaptic neuron, we assume that Vpre > Θsyn.
This allows us to calibrate the state of the post-synaptic neuron, and to determine the drive threshold
that separates the qualitatively distinct states of the individual and networked neurons. This statement is
illustrated in Fig. 6 by simulating responses of the endogenous parabolic burster to network perturbation.
Figure 6(a) shows, with a properly adjusted excitatory drive, that the endogenous burster switches into
tonic spiking activity. On the other hand, bursting in the networked neuron can be halted when it receives
a sufficient inhibitory drive from the pre-synaptic neuron of the network (Figure 6(b)). Eliminating either
drive makes the post-synaptic neuron return to its natural state, i.e. these experiments de-facto prove that
the neuron is mono-stable for the given parameter values.
A HCO, in the canonical Brown definition [Graham-Brown, 1911], is a pair of neurons bursting in
anti-phase when they are networked by inhibitory synapses. In isolation, such neurons are not endogenous
bursters but tonic spikers instead, or remain quiescent [Marder & Calabrese, 1996]. There are multiple
mechanisms underlying such anti-phase bursting, or, more accurately, anti-phase oscillations in HCOs and
CPGs made of relaxation oscillators [Kopell & Ermentrout, 2002; Daun et al., 2009]. The list includes the
well studied mechanisms of post-inhibitory rebound and escape for quiescent neurons [Perkel & Mulloney,
1974; Wang & Rinzel, 1985; Skinner et al., 1994b; Destexhe et al., 1994; Matveev et al., 2007], as well as
less-known mechanisms of HCOs constituted by intrinsically spiking neurons. Such networks utilizing the
Plant models are discussed below.
into the tonic spiking mode. This is done by setting the bifurcation parameter, ∆=−34mV , see Fig. 7.
To construct such a HCO with relatively weak inhibitory coupling, the Plant model must be first set
September 25, 2018
10:41 parabolicbursters
CPGs out of latent parabolic bursters
7
Fig. 8. Anti-phase network bursting produced by a HCO of two Plant neurons as soon as the inhibition is turned on. Blocking
the inhibition restores tonic spiking activity in both neurons, and vice versa. Here, the the network parameters are ginh
syn =
0.008nS and Esyn =−80mV , and the parameters of the individual neurons are the following: ∆=−60mV, ρ= 0.0003ms−1,
Kc= 0.0085ms−1, τx= 235ms and x∞(V)= 1~(1+ e−0.15(V+50)).
drive of gsyn= 0.001nS from the post-synaptic neuron 2 at t= 60sec. The inhibitory drive is sufficient to
Next, we consider a unidirectional network where the tonic spiking neuron 1 starts receiving, an inhibitory
shift the post-inhibitory neuron over the bifurcation transition back into bursting activity. The minimal
inhibitory drive must be increased proportionally to make the targeted neuron a network burster whenever
it stays further away from the bifurcation transition between tonic spiking and bursting in isolation.
4. Forming a half-center oscillator
In this section, we discuss the dynamics of half-center oscillators made of two tonically spiking Plant neurons
reciprocally coupled by inhibitory synapses. As before, we describe such synapses within the framework
of the fast threshold modulation (FTM) paradigm using Eq. (3) to match the shape and magnitude of
inhibitory postsynaptic potentials (IPSPs) in the post-synaptic neurons. IPSPs are the indicators of the
type and the strength of synapses in the network.
We perform simulations in a fashion that is analogous to the dynamic clamp technique used in neu-
rophysiological experiments. The approach involves the dynamic block, restoration and modulation of
synaptic connections during simulation. These modeling perturbations should closely resemble the exper-
imental techniques of drug-induced synaptic blockade, modulation, wash-out, etc. Restoring the chemical
synapses during a simulation makes the HCO regain network bursting activity with specific phase charac-
teristics. Depending on the coupling strength as well as the way the tonically spiking neurons are clamped,
the network bursting may change phase-locked states, i.e. be potentially multi-stable. Experimental ob-
servations also suggest specific constraints on the range of coupling strengths of the reciprocal inhibition,
such that the networks stably and generically achieve the desired phase-locking.
are initiated in tonic spiking mode. After turning on the reciprocally inhibitory synapses gsyn= 0.008nS,
Figure 8 demonstrates the stages of anti-phase bursting formation in the HCO. The uncoupled neurons
the HCO quickly transitions to the regime of robust anti-phase bursting. Turning off the synapses restores
the native tonic spiking activity in both neurons. Turning on the reciprocal synapses makes the HCO
regain the network bursting. Note that the length of transients from tonic spiking to network bursting
depends on the strength of the synaptic coupling for the fixed parameters of the individual Plant neurons.
By comparing the magnitude of IPSPs in the voltage traces represented in Figs. 8 and 9, one can conclude
that the coupling in the later case is weaker. This is why, the onset of network bursting in the HCO is less
pronounced.
Our modeling studies agree well with experimental recordings from the identified interneurons in the
Melibe swim CPG which suggests that the observed bursting is due to synergetic interactions of interneurons
of the network [Sakurai et al., 2014]. One can see from Fig. 1(b) that network bursting in the biological
HCO formed by two Si3 interneurons of the Melibe swim CPG is seized as soon as the right one, Si3R,
September 25, 2018
10:41 parabolicbursters
8 Ala¸cam and Shilnikov
Fig. 9. Onset of emergent network anti-phase bursting in the HCO with reciprocally inhibitory. (Esyn=−80mV ) synapses
syn= 0.0073nS.
at ginh
receives a negative current pulse that makes it hyperpolarized quiescent, while its left bursting counterpart,
Si3L, turns into tonic spiking activity instead. Moreover, one can deduct from the wiring diagram of the
CPG depicted in Fig. 2(a) and the analysis of voltage traces represented in Fig. 1(b) that the interneuron
Si2L becomes a tonic-spiker as soon as the pre-synaptic interneuron Si3R stops inhibiting it (compare
with Fig. 8.) This further supports the assertion that the swim CPG is made of intrinsically tonic spiking
interneurons.
To test the robustness of network anti-phase bursting to perturbation and to calibrate the necessary
influx of reciprocal inhibition generated by the Plant neurons, we consider a HCO with excitatory autapses.
The objective here is to determine an equivalent amount of excitatory drive to be projected onto the post-
inhibitory network burster to cancel out the inhibitory drive and shift it back to the initial tonic-spiking
mode.
An autapse is a synapse of a neuron onto itself, where the axon of the neuron ends on its own dendrite.
After their discovery [Van der Loos & Glaser, 1972] autapses have been observed in a range of nervous
systems. The autapses are arguably to be responsible for tuning of neural networks. This particular con-
figuration of the HCO depicted in Fig. 10 is formally motivated by the swim CPG circuitry, see Fig. 2(a).
One can see from it that the interneurons of the bottom HCO receive excitatory drives from the top in-
terneurons forming the top HCO. We would like to find the threshold over which the neurons no longer
form a stably bursting HCO. This would allow us to calibrate and quantify the relative strengths of the
mixed synaptic connections in the swim CPG models.
Both autapses are introduced to the model using the FTM approach with Eaut= 40mV . In this experiment,
In this HCO configuration, each neuron inhibits its counterpart and self-excites through the autapse.
Fig. 10. Turning on the excitatory autapses at gexc
bursting.
aut= 0.016nS in the HCO with ginh
syn= 0.0073nS halts pronounced network
September 25, 2018
10:41 parabolicbursters
CPGs out of latent parabolic bursters
9
Fig. 11. Assembly line of the Melibe swim CPG model out of four intrinsically tonic spiking Plant neurons. First the reciprocal
inhibition between Si3R and Si3L is turned on, followed by turning on the reciprocal inhibition between Si2R and Si2L, and next
simultaneous turning on unidirectional cross-lateral inhibition from Si3R(L) projected onto Si2L(R), and bi-lateral excitation
originating from Si2R(L) down onto Si3R(L). After a short transient, the CPG model exhibits the desired 3/4 phase shift lag
between Si2L and Si3L. Compare with voltage traces of the biological CPG in Fig. 2(b).
the conductance values for inhibitory synapses are set at ginh
generate robust anti-phase bursting as seen in Fig. 9. Next, we add the autapses along with inhibition and
gradually increase gexc
the network stops exhibiting anti-phase bursting. We note that unlike a permanent excitatory drive from
pre-synaptic neurons, an introduction of the excitatory autapse, acting only when the self-driving neuron
is above the synaptic threshold, is effectively perturbation equivalent for the calibration purpose.
aut proportionally increase the delay. At gexc
aut. We found that increasing gexc
syn= 0.0073nS. This is sufficient for the HCO to
aut= 0.016nS,
5. Assembly line of a Melibe swim CPG
In this final section, we put together a pilot model of the Melibe swim CPG according to a circuitry based
on identified interneurons and synapses; its wiring diagram is sketched in Fig. 2(a). This network model
is made of the two HCOs constituted by tonic spiking Plant neurons. We would like to find out whether
this sample CPG model can already produce phase lags similar to those between bursting interneurons in
the biological CPG. For the sake of simplicity, we do not include Si4R/L interneurons in the model and
we also omit electrical synapses. It is known from experimental studies [Sakurai et al., 2014] that blocking
chemical, inhibitory and excitatory synapses between the interneurons may be sufficient to break down
the motor pattern by the network. Figure 2(b) points out that the interneurons of either HCO burst in
anti-phase and there is the characteristic 3/4 phase lag between the burst initiation in the neurons Si2L
and Si3L, as well as between Si2R and Si3L. This phase lag is repeatedly observed in both adult and
juvenile animals.
As before, we use the Plant neurons initiated in the tonic spiking mode, relatively close to the transition
to bursting. Initial conditions of the neurons are randomized. After letting the neurons settled down to tonic
spiking activity, the network connections are turned on. As Fig. 11 shows, with the reciprocal inhibition
being first turned on, the bottom interneurons Si3L and Si3R become anti-phase network bursters, and
so do Si2R and Si2L as soon as the reciprocal inhibition between them is turned them on, too. At this
stage, the CPG model is formed by two uncoupled HCOs. A few seconds later, they become coupled by
September 25, 2018
10:41 parabolicbursters
10 Ala¸cam and Shilnikov
simultaneous turning on the unidirectional cross-lateral inhibition from Si3R(L) projected onto Si2L(R),
and bi-lateral excitation from Si2R(L) down onto Si3R(L). One can see from this figure that all four
interneurons of the CPG model exhibit network bursting with the desired phase lags. These are 0.5 (half
period) between the interneurons of each HCO, and 3/4 (a fraction of the network period) between the
HCOs, or between the corresponding reference interneurons Si2L and Si3L. We note that such a phase
shift was reported in a similar Melibe swim CPG constituted by endogenous bursters; that model also
incorporated electrical synapses [Jalil et al., 2013]. There is a great room for improvement of CPG network
models to include other identified interneurons and to incorporate additional electrical synapses to find
out whether additions of new elements can stabilize or desynchronize the desired bursting pattern as it
was done using the Poincar´e return maps for endogenous bursters [Wojcik et al., 2014]. Of our special
interest are various problems concerning structural stability of the network, and its robustness (Lyapunov
stability) for bursting outcomes subjected to perturbations by pulses of the external current, as well as
reductions to return maps between burst initiations in constituent neurons. These questions are beyond
the scopes of the given examination and will be addressed in full detail in our forthcoming publications
soon. The question about a possible linking of the characteristic 3
4 phase lag and the Melibe leonina lateral
swim style is the paramount one among them.
6. Summary
We have discussed a basic procedure for building network bursting CPGs made of intrinsically tonic spiking
neurons. As a model for such networks, we have employed the biophysically plausible Plant model that
was originally proposed to describe endogenous bursting R15-cells in the Aplysia mollusk. Such bursting
was intracellularly recorded, and identified as parabolic, from the known interneurons in the swim CPGs of
two sea slugs: Melibe leonina and Dendronotus iris. There is experimental evidence that bursting in these
swim CPGs is due to synergetic interactions of all constituent neurons that are intrinsic tonic-spikers in
isolation. To model the Melibe swim CPG, we have first examined dynamical and structural properties of
the Plant model and its responses to perturbations. These perturbations include inhibitory and excitatory
inputs from pre-synaptic neurons in the network. We have identified the transition boundary beyond
which the bursting Plant model becomes a tonic-spiker and shifted it slightly over the threshold using an
introduced bifurcation parameter. We have shown that the perturbed/calibrated Plant neuron, exhibiting
intrinsically tonic spiking activity, becomes a network burster when it receives an inhibitory drive from
a pre-synaptic neuron. By combining two such neurons, we have created a genuine half-center oscillator
robustly producing anti-phase bursting dynamics. We have also considered a HCO configuration with two
excitatory autapses to assess the robustness of anti-phase bursting with respect to excitatory perturbations.
Finally, we have employed all necessary components to assemble a truncated model of the Melibe swim CPG
with the characteristic 3/4-phase lags between the bursting onsets in the four constituent interneurons.
In future studies, we plan to examine the dynamics of the CPG models with all synaptic connections,
including electrical, as well as incorporating additional identified interneurons. We will also explore their
structural stability, robustness and potential multi-stability of their bursting outcomes with various phase
lags. An additional goal is to find out whether the motor pattern with the 3/4-phase lags will persist
in networks with interneurons represented by other mathematical models including phenomenologically
reduced ones. Potentially, these findings shall provide a systematic basis for comprehension of plausible
biophysical mechanisms for the origination and regulation of rhythmic patterns generated by various CPGs.
Our goal is to extend and generalize the dynamical principles disclosed in the considered networks for other
neural systems besides locomotion, such as olfactory cellular networks.
7. Acknowledgments
A.S. also acknowledges the support from GSU Brain and Behaviors pilot grant, RFFI 11-01-00001, RSF
grant 14-41-00044 at the Lobachevsky University of Nizhny Novgorod and the grant in the agreement of
August 27, 2013 N 02.B.49.21.0003 between the Ministry of education and science of the Russian Federation
and Lobachevsky State University of Nizhni Novgorod (Sections 2-4), as well as NSF BIO-DMS grant
XXXXX. We thank A. Sakurai and P.S. Katz for sharing experimental data and suggestions, and the
September 25, 2018
10:41 parabolicbursters
acting and associated members of the NEURDS (Neuro Dynamical Systems) lab at GSU: A. Kelley, K.
Pusuluri, S. Pusuluri, T.Xing, M. Fen, F. Fen, J. Collens, D. Knapper, A. Noriega, P. Ozluk, B. Chung
and E. Latash for helpful discussions, and R. Clewley for his guidance on the PyDSTool package [Clewley
et al., 2006] used in simulations.
CPGs out of latent parabolic bursters
11
September 25, 2018
10:41 parabolicbursters
12 Ala¸cam and Shilnikov
8. Appendix: the conductance based Plant model
Cm V =−IN a− IK− ICa− IKCa− Ileak− Isyn,
The model in this study is adopted from [Plant, 1981]. The dynamics of the membrane potential, V , is
governed by the following equation:
where Cm= 1µF~cm2 is the membrane capacitance, IN a is the N a+ current, IK is the K+ current, ICa
is the Ca+2 current, IKCa is the Ca2+ activated K+ current, Ileak is the leak current, Isyn is the synaptic
where the reversal potential VN a = 30mV and the maximum N a+ conductance value gN a = 4nS. The
current. The fast inward sodium current is given by
(4)
(5)
instantaneous activation variable is defined as
IN a= gN am3∞(V)h(V − VN a),
αm(V)
m∞(V)=
αm(V)+ βm(V) ,
50− Vs
βm(V)= 4 exp((25− Vs)~18),
,
while the dynamics of inactivation variable h is given by
The fast potassium current is given by the equation
where the reversal potential is VK=−75mV and the maximum K+ conductance value is gK= 0.3nS.The
(12)
dynamics of inactivation gating variable is described by
(6)
(7)
(8)
(9)
(10)
(11)
(13)
(14)
(15)
(16)
where
where
with
where
where
with
αm(V)= 0.1
exp((50− Vs)~10)− 1
h= h∞(V)− h
τh(V)
and τh(V)=
αh(V)= 0.07 exp((25− Vs)~20) and βh(V)=
αh(V)
αh(V)+ βh(V)
h∞(V)=
,
12.5
αh(V)+ βh(V) ,
exp((55− Vs)~10)+ 1
1
,
105
mV.
Vs= 127V + 8265
IK= gKn4(V − VK),
n= n∞(V)− n
τn(V)
and τn(V)=
and βn(V)= 0.125 exp((45− Vs)~80).
αh(V)+ βh(V) ,
12.5
,
ICa= gCax(V − VCa),
αh(V)
n∞(V)=
αh(V)+ βh(V)
55− Vs
αn(V)= 0.01
exp((55− Vs)~10)− 1
The TTX-resistant calcium current is given by
September 25, 2018
10:41 parabolicbursters
REFERENCES 13
where
governed by
,
1
x∞(V)=
dynamics of the slow activation variable is described by
where the reversal potential is VCa= 140mV and the maximum Ca2+ conductance is gCa= 0.03nS. The
x= x∞(V)− x
τx(V)
and τx(V)= 9400ms.
exp(−0.3(V + 40))+ 1
The outward Ca2+ activated K+ current is given by
[Ca]i
(V − VK),
IKCa= gKCa
0.5+[Ca]i
where the reversal potential is VCa = 140mV . The dynamics of intracellular calcium concentration is
Ca= ρ[Kcx(VCa− V)−[Ca]i] ,
where the reversal potential is VCa = 140mV , and the constant values are ρ= 0.00015mV−1 and Kc =
0.00425mV−1. The leak current is given by
Ileak= gL(V − VL),
where the reversal potential VL=−40mV and the maximum conductance value gL= 0.0003nS. The synaptic
Isyn= gsyn(Vpost− Erev)
1+ e−k(Vpre−Θsyn)
with the synaptic reversal potential Vpost=−80mV for inhibitory synapses and Vpost= 40mV for excitatory
synapses and the synaptic threshold Θsyn= 0mV , and k= 100.
(20)
(21)
current is defined as
(17)
(18)
(19)
(22)
References
Afraimovich, V., Gonchenko, S., Lerman, L., Shilnikov, A. & Turaev, D. [2014] "Scientific heritage of L.P.
Shilnikov. Part 1." Regular and Chaotic Dynamics 19, 435 -- 460.
Angstadt, J., Grassmann, J., Theriault, K. & Levasseur, S. [2005] "Mechanisms of postinhibitory rebound
and its modulation by serotonin in excitatory swim motor neurons of the medicinal leech," Journal of
Comparative Physiology A-Neuroethology, Sensory, Neural and Behavioral Physiology 191, 715 -- 732,
doi:10.1007/s00359-005-0628-6.
Bal, T., Nagy, F. & Moulins, M. [1988] "The pyloric central pattern generator in crustacea: a set of
conditional neural oscillators," Journal of Comparative Physiology A 163, 715 -- 727.
Belykh, I. & Shilnikov, A. [2008] "When weak inhibition synchronizes strongly desynchronizing networks
of bursting neurons." Phys Rev Lett 101, 078102.
Bertran, R. [1993] "A computational study of the effects of serotonin on a molluscan burster neuron," Biol.
Cybern. 69, 257 -- 267.
Best, J., Borisyuk, A., Rubin, J., Terman, D. & M., W. [2005] "The dynamic range of bursting in a model
respiratory pacemaker network," SIAM J. Appl. Dyn. Syst. 4, 1107 -- 1139.
Briggman, K. L. & Kristan, W. B. [2008] "Multifunctional pattern-generating circuits." Annu Rev Neurosci
31, 271 -- 294, doi:10.1146/annurev.neuro.31.060407.125552, URL http://dx.doi.org/10.1146/annurev.
neuro.31.060407.125552.
Butera, R. [1998] "Multirhythmic bursting," Chaos 8, 274 -- 284.
Butera, R., Clark, J. W. J., Canavier, C. C., Baxter, D. A. & Byrne, J. H. [1995] "Analysis of the effects of
modulatory agents on a modeled bursting neuron: Dynamic interactions between voltage and calcium
dependent systems," J Comput. Neuroscience 2, 19 -- 44.
Calabrese, RL., Norris, BJ., Wenning, A. & Wright, TM. [2011] "Coping with variability in small neuronal
networks," Integrative and Comparative Biology 51, 845 -- 855, doi:10.1093/icb/icr074.
September 25, 2018
10:41 parabolicbursters
14 REFERENCES
Canavier, C., Clark, J. & Byrne, J. [1991] "Simulation of the bursting activity of neuron R15 in Aplysia:
Role of ionic currents, calcium balance, and modulatory transmitters" J Neurophysiology 66, 2107 --
2124.
Canavier, C., Baxter, DA., Clark, JW. & Byrne, JH. [1994] "Multiple modes of activity in a model neuron
suggest a novel mechanism for the effects of neuromodulator" J Neurophysiology 72, 872 -- 882.
Clewley, R., Sherwood, W., LaMar, M. & Guckenheimer, J. [2006] "Pydstool: an integrated simulation,
modeling, and analysis package for dynamical systems." Tech. rep., http://pydstool.sourceforge.net.
Collens, J., Kelley, A., Alacam, D., Xing, T., D., K., Schwabedal, J. & Shilnikov, A. [2015] "Intrinsic
mechanisms for pattern generation in three-node networks," submitted .
Daun, S., Rubin, J. E. & Rybak, I. A.
[2009] "Control of oscillation periods and phase durations
in half-center central pattern generators: a comparative mechanistic analysis," Journal of Com-
putational Neuroscience 27, 3 -- 36, doi:10.1007/s10827-008-0124-4, URL http://dx.doi.org/10.1007/
s10827-008-0124-4.
Destexhe, A., Contreras, D., Sejnowski, T. & Steriade, M. [1994] "A Model of Spindle Rhythmicity in the
Isolated Thalamic Reticular Nucleus," Journal of Neurophysiology 72, 803 -- 818.
Dror, R. O., Canavier, C. C., Butera, R. J., Clark, J. W. & Byrne, J. H. [1999] "A mathematical criterion
based on phase response curves for stability in a ring of coupled oscillators." Biol Cybern 80, 11 -- 23,
doi:10.1007/s004220050501, URL http://dx.doi.org/10.1007/s004220050501.
Frost, W. N. & Katz, P. S. [1996] "Single neuron control over a complex motor program," Proc.Nat.Acad.
Sc. 93, 422 -- 426.
Gillner, S. & Wallen, P. [1985] "Central pattern generators for locomotion, with special references to
vertebrates," Ann. Rev. Neurosci. 8, 233 -- 261.
Graham-Brown, T. [1911] "The intrinsic factors in the act of progression in the mammal," Lond B Biol
Society 84, 308 -- 319.
Hill, A., Van Hooser, S. & Calabrese, R. [2003] "Half-center oscillators underlying rhythmic movements,"
The Handbook of Brain Theory and Neural Networks, ed. Arbib, M. A. (The MIT Press).
Jalil, S., Allen, D., Youker, J. & Shilnikov, A. [2013] "Toward robust phase-locking in melibe swim central
pattern generator models," Chaos 23, 046105, doi:http://dx.doi.org/10.1063/1.4825389, URL http:
//scitation.aip.org/content/aip/journal/chaos/23/4/10.1063/1.4825389.
Jalil, S., Belykh, I. & Shilnikov, A. [2010] "Fast reciprocal inhibition can synchronize bursting neurons,"
Physical Review E 81, doi:{10.1103/PhysRevE.81.045201}.
Physical Review E 85, doi:{10.1103/PhysRevE.85.036214}.
Jalil, S., Belykh, I. & Shilnikov, A. [2012] "Spikes matter for phase-locked bursting in inhibitory neurons,"
Ji, L., Zhang, J., Lang, X. & Zhang, X. [2013] "Coupling and noise induced spiking-bursting transi-
tion in a parabolic bursting model," Chaos: An Interdisciplinary Journal of Nonlinear Science 23,
013141, doi:http://dx.doi.org/10.1063/1.4795281, URL http://scitation.aip.org/content/aip/journal/
chaos/23/1/10.1063/1.4795281.
Katz, P. S. & Hooper, S. [2007] "Invertebrate central pattern generators," Invertebrate Neurobiology, eds.
North, G. & R. Greenspan, R. (Cold Spring Harbor Laboratory Press, NY, New York).
Koch, H., Garcia, A. J. & Ramirez, J.-M. [2011] "Network reconfiguration and neuronal plasticity in
rhythm-generating networks," Integrative and Comparative Biology 51, 856 -- 868, doi:10.1093/icb/
icr099.
Kopell, N. [1988] "Toward a theory of modelling central pattern generators," Neural Control of Rhythmic
Movements in Vertebrates, eds. Cohen, A., Rossingol, S. & Grillner, S. (Wiley, New York).
Kopell, N. & Ermentrout, B. [2004] "Chemical and electrical synapses perform complementary roles in the
synchronization of interneuronal networks," Proc. Natl. Acad, Sci. 101, 15482 -- 15487, doi:10.1073/
pnas.0406343101.
Kopell, N. & Ermentrout, G. [2002] "Mechanisms of phase-locking and frequency control in pairs of coupled
neural oscillators," Handbook on Dynamical Systems 2, 3 -- 54.
Kopell, N. & Somers, D. [1993] "Rapid synchronization through fast threshold modulation," Biol. Cybern.
68, 5.
Kristan, W. [2008] "Neuronal decision-making circuits." Curr Biol 18, R928 -- R932, doi:10.1016/j.cub.2008.
September 25, 2018
10:41 parabolicbursters
REFERENCES 15
07.081, URL http://dx.doi.org/10.1016/j.cub.2008.07.081.
Kristan, W., Calabrese, R. & Friesen, W. [2005] "Neuronal control of leech behavior," Prog. Neurobiol. 76,
279.
Levitan, E. & Levitan, I. B. [1988] "Serotonin acting via cyclic amp enhances both the hyperpolarizing
and depolarizing phases of bursting pacemaker activity in the aplysia neuron R15," J Neuroscience
8, 1152 -- 1161.
Marder, E. [2012] "Neuromodulation of neuronal circuits: back to the future," Neuron 76, 1.
Marder, E. & Calabrese, R. [1996] "Principles of rhythmic motor pattern generation," Physiological Reviews
76, 687 -- 717.
Matsuoka, K. [1987] "Mechanisms of frequency and pattern control in the neural rhythms generators,"
Matveev, V., Bose, A. & Nadim, F. [2007] "Capturing the bursting dynamics of a two-cell inhibitory
network using a one-dimensional map," Journal of Computational Neuroscience 23, 169 -- 187, doi:
Biol. Cybernetics 1, 1.
{10.1007/s10827-007-0026-x}.
Newcomb, J., Sakurai, A., Lillvis, J., Gunaratne, C. & Katz, P. S. [2012] "Homology and homoplasy of
swimming behaviors and neural circuits in the nudipleura (mollusca, gastropoda, opistho-branchia),"
Proc. Natl. Acad. Sci. 109, 10669 -- 76.
Perkel, D. & Mulloney, B. [1974] "Mechanism of postinhibitory rebound in molluscan neurons," Science
185, 181 -- 183, doi:10.1126/science.185.4146.181.
Plant, R. [1981] "Bifurcation and resonance in a model for bursting nerve cells," J. Math. Biol 11, 15 -- 32.
Plant, R. & Kim, M. [1975] "On the mechanism underlying bursting in the aplysia abdominal ganglion r15
cell," Math. Biosci. 26, 357 -- 375.
Plant, R. E. & Kim, M. [1976] "Mathematical description of a bursting pacemaker neuron by a modification
of the hodgkin-huxley equations," Biophys J 16, 227 -- 244.
Prinz, A., Billimoria, C. & Marder, E. [2003a] "Alternative to hand-tuning conductance-based models:
construction and analysis of databases of model neurons," J.Neurophysiol. 90, 3998 -- 4015, URL PM:
12944532.
Prinz, A. A., Bucher, D. & Marder, E. [2004] "Similar network activity from disparate circuit parameters,"
Nature neuroscience 7, 1345 -- 1352.
Prinz, A. A., Thirumalai, V. & Marder, E. [2003b] "The functional consequences of changes in the strength
and duration of synaptic inputs to oscillatory neurons," The Journal of neuroscience 23, 943 -- 954.
Rinzel, J. & Lee, Y. S. [1987] "Dissection of a model for neuronal parabolic bursting," Journal of Mathe-
matical Biology 25, 653 -- 675.
Sakurai, A., Gunaratne, C. A. & Katz, P. S. [2014] "Two interconnected kernels of reciprocally inhibitory
interneurons underlie alternating left-right swim motor pattern generation in the mollusk melibe
leonina," Journal of Neurophysiology 112, 317 -- 1328.
Sakurai, A. & Katz, P. S. [2011] "Distinct neural circuit architectures produce analogous rhythmic behaviors
in related species," Soc. Neurosci. Abstr. 37.918.04.
Sakurai, A., Newcomb, J., Lillvis, J. & Katz, P. S. [2011] "Synaptic patterning of left-right alternation in
a computational model of the rodent hindlimb central pattern generator," Curr. Biol. 21, 1036.
Schwabedal, J., Knapper, D. & Shilnikov, A. [2015] "Qualitative and quantitative stability of polyrhythmic
circuits," submitted .
Schwabedal, J. T. C., Neiman, A. B. & Shilnikov, A. L. [2014] "Robust design of polyrhythmic neural
circuits," Phys. Rev. E 90, 022715, doi:10.1103/PhysRevE.90.022715, URL http://link.aps.org/doi/
10.1103/PhysRevE.90.022715.
Selverston, A. (ed.) [1985] Model Neural Networks and Behavior (Springer, Berlin).
Sherwood, W., Harris-Warrick, R. & Guckenheimer, J. [2010] "Synaptic patterning of left-right alternation
in a computational model of the rodent hindlimb central pattern generator," J. Comput Neuroscience
30, 323.
Shilnikov, A. [2012] "Complete dynamical analysis of a neuron model," Nonlinear Dynamics 68, 305 -- 328,
doi:{10.1007/s11071-011-0046-y}.
Shilnikov, A., Gordon, R. & Belykh, I. [2008] "Polyrhythmic synchronization in bursting networking mo-
September 25, 2018
10:41 parabolicbursters
16 REFERENCES
tifs," Chaos 18, doi:{10.1063/1.2959850}.
Shilnikov, L. [1963] "Some cases of generation of periodic motion from singular trajectories." Math. USSR
Sbornik 61, 443 -- 466.
Shilnikov, L., Shilnikov, A., Turaev, D. & Chua, L. [1998,2001] Methods of Qualitative Theory in Nonlinear
Dynamics, Parts I and II (World Scientific Publ.), ISBN 9789810240721.
Sieling, F. & Butera, R. [2011] "Aplysia R15 neuron," Scholarpedia 6, 4181, URL doi:10.4249/scholarpedia.
4181.
Skinner, F., Kopell, N. & Marder, E. [1994a] "Mechanisms for oscillation and frequency control in networks
of mutually inhibitory relaxation oscillators," Comput. Neurosci. 1, 69.
Skinner, F., Kopell, N. & Marder, E. [1994b] "Mechanisms for oscillation and frequency control in recip-
rocally inhibitory model neural networks," Journal of Computational Neuroscience 1, 69 -- 87.
Van der Loos, H. & Glaser, E. [1972] "Autapses in neocortex cerebri: synapses between a pyramidal cell's
axon and its own dendrites," Brain Research 48.
Wang, X.-J. & Rinzel, J. [1985] "Alternating and synchronous rhythms in reciprocally inhibitory model
neurons," Ann. Rev. Neurosci. 8, 233 -- 261.
Wojcik, J., Clewley, R., Schwabedal, J. & Shilnikov, A. [2014] "Key bifurcations of bursting polyrhythms
in 3-cell central pattern generators," PLoS ONE 9.
Wojcik, J., Clewley, R. & Shilnikov, A. [2011] "Order parameter for bursting polyrhythms in multifunctional
central pattern generators," Phys Rev E 83, 056209 -- 6.
Wojcik, J., Schwabedal, J., Clewley, R. & Shilnikov, A. [2014] "Key bifurcation of bursting polyrhythms
in 3-cell central pattern generators," PLoS ONE 9, e92918.
|
1107.2521 | 3 | 1107 | 2013-06-01T14:49:02 | Network algorithmics and the emergence of synchronization in cortical models | [
"q-bio.NC"
] | When brain signals are recorded in an electroencephalogram or some similar large-scale record of brain activity, oscillatory patterns are typically observed that are thought to reflect the aggregate electrical activity of the underlying neuronal ensemble. Although it now seems that such patterns participate in feedback loops both temporally with the neurons' spikes and spatially with other brain regions, the mechanisms that might explain the existence of such loops have remained essentially unknown. Here we present a theoretical study of these issues on a cortical model we introduced earlier [Nathan A, Barbosa VC (2010) Network algorithmics and the emergence of the cortical synaptic-weight distribution. Phys Rev E 81: 021916]. We start with the definition of two synchronization measures that aim to capture the synchronization possibilities offered by the model regarding both the overall spiking activity of the neurons and the spiking activity that causes the immediate firing of the postsynaptic neurons. We present computational results on our cortical model, on a model that is random in the Erd\H{o}s-R\'enyi sense, and on a structurally deterministic model. We have found that the algorithmic component underlying our cortical model ultimately provides, through the two synchronization measures, a strong quantitative basis for the emergence of both types of synchronization in all cases. This, in turn, may explain the rise both of temporal feedback loops in the neurons' combined electrical activity and of spatial feedback loops as brain regions that are spatially separated engage in rhythmic behavior. | q-bio.NC | q-bio |
1
Network Algorithmics and the Emergence of Synchronization in
Cortical Models
Andre Nathan, Valmir C. Barbosa∗
Programa de Engenharia de Sistemas e Computa¸cao, COPPE, Universidade Federal do
Rio de Janeiro, Rio de Janeiro, RJ, Brazil
∗ Corresponding author, e-mail: [email protected]
Abstract
When brain signals are recorded in an electroencephalogram or some similar large-scale record of brain
activity, oscillatory patterns are typically observed that are thought to reflect the aggregate electrical
activity of the underlying neuronal ensemble. Although it now seems that such patterns participate
in feedback loops both temporally with the neurons' spikes and spatially with other brain regions, the
mechanisms that might explain the existence of such loops have remained essentially unknown. Here we
present a theoretical study of these issues on a cortical model we introduced earlier [Nathan A, Barbosa
VC (2010) Network algorithmics and the emergence of the cortical synaptic-weight distribution. Phys
Rev E 81: 021916]. We start with the definition of two synchronization measures that aim to capture
the synchronization possibilities offered by the model regarding both the overall spiking activity of the
neurons and the spiking activity that causes the immediate firing of the postsynaptic neurons. We present
computational results on our cortical model, on a model that is random in the Erdos-R´enyi sense, and
on a structurally deterministic model. We have found that the algorithmic component underlying our
cortical model ultimately provides, through the two synchronization measures, a strong quantitative basis
for the emergence of both types of synchronization in all cases. This, in turn, may explain the rise both
of temporal feedback loops in the neurons' combined electrical activity and of spatial feedback loops as
brain regions that are spatially separated engage in rhythmic behavior.
Introduction
Current technology allows the recording of electromagnetic brain activity at several different spatial scales,
ranging from invasive recordings that capture the signals from neuronal ensembles comprising thousands
to millions of cells to those that are noninvasive (like the electroencephalogram) and capture the signals
from large-scale cortical areas [1]. Invariably the recorded signals appear as oscillatory patterns which,
depending on the scale at which the recording is performed, occur in different frequency ranges and
provide distinct interpretive possibilities. At the smallest spatial scales, for example, the recorded signals
occur in the range of a few kilohertz and are thought to reflect the combined spiking (or firing) activity
of the group of neurons in question. The signals recorded at the largest scales, on the other hand, occur
in the range of a few to 100 hertz and are thought to reflect the so-called local field potentials (LFPs),
which in turn are purported to reflect the aggregate effects of all the electrical activity taking place in
the corresponding area [2].
LFPs have acquired great prominence recently, owing mainly to the fact that LFPs related to different
brain areas appear to combine with one another in such ways as to correlate significantly with the brain's
sensory-motor mechanisms and other, higher-level functions as well (such as memory, attention, and
others; cf. [3 -- 5] and references therein). The ways in which LFPs combine seem to involve forms of cross-
frequency coupling that cover wide temporal and spatial ranges and ultimately promote the integration
of activities with different temporal and spatial characteristics [5]. Decoding the various forms of LFP
coupling may one day hold the key to understanding how computation and communication take place in
the brain.
2
As it happens, though, many aspects of the nature of LFPs have remained elusive. In particular, their
precise origin and relation to the underlying firing activity of the neurons seem to depend crucially on
the brain region being considered. However, notwithstanding this indefiniteness that still surrounds the
specifics of LFP emergence and interaction with other LFPs, some of the crucial points are beginning
to be clarified [2, 6]. One of them is the relation with the combined firing activity of groups of neurons.
Although LFPs seem to derive largely from the accumulation of potential at the neurons' membranes
that eventually leads to neuronal firing, it appears that the firing patterns themselves have a role to play
as well. As a consequence, a picture that seems to be emerging is that of a feedback loop in which the
larger-scale LFPs both influence and get influenced by the smaller-scale firing patterns. Another point
is more related to the spatial characteristics of how LFPs interact with one another: It now seems clear
that feedback loops also occur involving the LFPs of distinct brain areas.
While a complete resolution of these issues will certainly require considerable further research, a
detailed understanding will undoubtedly benefit from a clear picture of firing synchronization at the
neuronal level. By synchronization we mean both the convergence of multiple action potentials onto
distinct cells within a relatively narrow temporal window, and also the firing due to such potentials
within a similar window. As we mentioned above, both phenomena are closely related to the rise of
LFPs, and consequently the rise of the higher-level functions that LFPs are thought to support. In our
view, accounting for the possibilities the brain affords for the appearance of such synchronized behavior
depends both on the anatomical properties of how neurons interconnect and on the individual firing
behavior of each neuron. Our aim in this paper is the study of these types of synchronism on a graph-
theoretic model of the cortex.
Our study is preceded by a few others [7 -- 9], all of which have modeled neuronal behavior as a
continuous-time process in which the relevant signals obey certain differential equations and feed some
measure of synchronous activity. Invariably these studies have modeled the synchronization of membrane
potentials directly, and have for this reason stayed apart from explicitly modeling most of the relevant
details that characterize neuronal firings, like the so-called local histories of each neuron [10]. Our ap-
proach is to take a different course and focus on the extent to which certain events at the various neurons
can be said to be synchronized. These events are, in essence, the arrival of action potentials from other
neurons and the eventual creation of new action potentials. These, of course, are precisely the events
that make up membrane potentials, but we have found in previous studies of related problems that the
additional level of detail pays off by providing unprecedented insight. Examples here are the emergence of
the synaptic-weight distribution, in which case we demonstrated that experimentally observed distribu-
tions [11] can be reproduced [12], and a more tractable version [13] of the integrated-information theory
for the emergence of consciousness [14].
Our model is based on what is generally called network algorithmics.
In our case this refers to
the combination of a structural component to represent the anatomical properties of the cortex at the
neuronal level, and a distributed-algorithmic component to represent the traffic of action potentials among
the neurons as they fire. In general, a distributed algorithm is a collection of sequential procedures, each
run by an agent in a distributed system with provisions for messages to be sent to other agents for
communication [15]. The distributed algorithm of interest here is inherently asynchronous, meaning that
the behavior of each agent (each neuron) is purely reactive to the arrival of messages (action potentials),
which can in turn undergo unpredictable delays before delivery.
It follows that independent runs of
the same algorithm from the same initial conditions cannot in general be guaranteed to generate the
same sequence of message arrivals at any agent, and therefore not the same behavior. Lately, this
inherent source of unpredictability has led asynchronous distributed algorithms to be recognized as a
powerful abstraction in the modeling of biological systems [16, 17], since they allow the incorporation
into the model of small behavioral variations that might otherwise be simply smoothed out and rendered
irrelevant.
3
An asynchronous distributed algorithm never makes any reference to a global-time entity.
In our
model, therefore, every property emerging from neuronal interactions is the result of strictly local ele-
ments. The presence of such strong asynchronism may seem contradictory with the search for signs of
synchronism. However, we have found that observing the system from the outside can reveal surprising
synchronization possibilities that could never be known within the scope of individual neurons, provided
we uncover the causal relationship that binds those local elements on a global scale.
Using network algorithmics as the basis of our model makes it substantially more detailed than those
that target membrane potentials as the basic analytical units, in the sense that a neuron's behavior
can now be described as its reaction to the arrival of a new action potential given its current state.
Clearly, though, there would be room for considerably more detail, even down to the molecular level.
Our approach, therefore, resides in the middle ground between two extremes and as such is related to
what has become known as the artificial-life approach [18, 19], whose "life as it could be" metaphor is
close in spirit to our own view. Our approach is also part of a growing collection of efforts that attempt to
tackle eminent problems of neuroscience from the perspective of graph-based and other interdisciplinary
methods [20 -- 31]. As an enterprise that eventually seeks to be able to handle realistically large problem
instances from a graph-theoretic perspective, the present study is moreover in line with the many others
that in the last decade have addressed the so-called complex networks [32 -- 34].
Our model is defined on a strongly connected directed graph S, i.e., a directed graph in which it is
possible to reach every node from every other node through directed paths only. We assume that S has
N nodes, each representing a neuron, so each one can be either excitatory or inhibitory in the proportion
of 4 to 1, respectively [35 -- 37]. In S, an edge exists directed from node i to node j to indicate a synapse
from i's axon to one of j's dendrites, but no edge can connect two inhibitory nodes [35]. The results we
describe in the remainder of the paper for our cortical model are predicated upon S being drawn from
the following random-graph model on n ≥ N nodes in such a way that S is the giant strongly connected
component, GSCC [38], of the directed graph D that is actually drawn. In D a randomly chosen node,
say i, has out-degree k > 0 with probability proportional to k−1.8.1 Following [50, 51], each of these k
nodes is precisely another randomly chosen node, say j, with probability proportional to e−dij , where dij
is the Euclidean distance between nodes i and j when all n nodes are placed uniformly at random on a
radius-1 sphere. We have observed this policy for edge placement to lead to N ≈ 0.9n on average (i.e.,
we expect the GSCC of D, and hence S, to comprise roughly 90% of the n nodes) [13].
Methods
We start with the specification of the asynchronous distributed algorithm to be executed by the nodes
of graph S. This algorithm prescribes a common procedure for each node to react to the arrival of a
message and to possibly send messages out as a consequence. Such a procedure runs atomically (i.e.,
without being interrupted), so actually implementing the algorithm requires provisions for the queuing
of incoming messages at each node. Each run of the distributed algorithm can be described formally by
specifying how the various message arrivals at the different nodes relate to one another. As we do this a
precise notion of causality emerges and we use it to define our two synchronization measures.
1Originally the choice of this scale-free distribution [39] was inspired by related work that observed such a power law at
a higher level of detail in the cortical architecture [40, 41], but recent measurements regarding the distribution of neuronal
(undirected) degrees [42] seem to support the choice we made. Specifically, ignoring edge directions in our model leads to
an exponential degree distribution [13], as observed in [42]. We also note that our adoption of a power law to describe the
distribution of out-degrees represents a significant departure from earlier modeling attempts [35,43], which used randomness
in the Erdos-R´enyi sense [44] directly. We note, moreover, that cortices do exhibit many other scale-free properties [29,45,46],
including some that evince their small-world nature [47], the existence of hubs (nodes with very large out-degrees), and
others [48, 49].
4
Asynchronous distributed algorithm
All nodes in S share the same rest and threshold potentials, denoted by v0 and vt, respectively, such
that v0 < vt. We use vj to represent the potential of node j and wij to represent the synaptic weight
of the edge directed from i to j. At all times, these are constrained in such a way that vj ∈ [v0, vt] and
wij ∈ [0, 1]. An asynchronous distributed algorithm, henceforth referred to as A, underlies the node-
potential and synaptic-weight dynamics of our cortical model. At node j, algorithm A is built around
the "firing" operation which consists of sending a message to each out-neighbor of j in S and setting vj
to the rest potential v0. Algorithm A is purely reactive, that is, it only acts at any node when the node
receives a message. Evidently there must be exceptions to this purely reactive character, since at least
one node must send messages out spontaneously to start the algorithm. We call such nodes initiators and
let m ≤ N be the number of initiators in a given run of algorithm A. When node j acts as an initiator
(this happens at most once in a run for that node), it simply fires with probability 1. When it reacts to
the reception of a message, say from node i, node j performs the following steps:
1. If i is excitatory, then set vj to min{vt, vj + wij}. If it is inhibitory, then set vj to max{v0, vj − wij}.
2. Fire with probability (vj − v0)/(vt − v0).
3. If firing did occur during step 2, then set wij to min{1, wij + δ}. If it did not occur but the previous
message received by node j from any of its in-neighbors did cause firing to occur, then set wij to
(1 − α)wij .
The parameters δ and α appearing in step 3 are constrained by δ ≤ α [12]. They are meant to
reproduce, although to a limited extent, the spike-timing-dependent plasticity principles [52, 53], which
roughly dictate that wij is to increase if firing occurs and decrease otherwise, always taking into account
how close in time the relevant firings by nodes i and j are.2 Moreover, increases in wij are to occur by
a fixed amount, decreases by proportion [54 -- 56]. Every run of algorithm A is guaranteed to eventually
reach a point at which no more message traffic exists. It is then said to have terminated.
Modeling causality
The problem of handling synchronization during a run of an asynchronous distributed algorithm such as
algorithm A is akin to that of handling simultaneity in special relativity. This is so because, in either
context, establishing the notion of simultaneity is dependent upon how signals propagate and also on
locality considerations [57]. A convenient framework within which to tackle it is then the event-based
formalism originally laid down in [58]. An event in the present context is said to have occurred either
when an initiator fires (and thereby sends a message to each of its out-neighbors in graph S) or when a
node reacts to the reception of a message by executing steps 1 -- 3 (which includes the possibility of firing).
In either occasion the event involves the execution of an atomic set of actions by the node in question.
A run of algorithm A can then be regarded as a set R of events. An event r ∈ R is described by the
4-tuple r = hi, ti, mi, Mii, where i identifies the node of S at which event r occurred, ti ≥ 1 indicates that
r was the tith event to occur at node i since the run began, mi is the message (if any) that triggered the
occurrence of the event, and Mi is the set of messages sent by node i if it fired during the event. Many
of the events in R are implicitly related to one another. We make such a relation explicit by defining the
binary relation B ⊆ R2. Given events r as above and r′ = hj, tj, mj, Mji, we say that the ordered pair
(r, r′) ∈ B if and only if:
(a) Either i and j are the same node and ti < tj with no intervening event between r and r′;
2In Hebbian terms, learning depends on how the synapses incoming to j compete with one another for the firing of j.
Strengthening the synapse from i, in particular, depends on whether the arrival of an action potential from i succeeds in
making j fire, either by itself or combined with other recent arrivals that did not cause firing [53].
5
(b) Or j is an out-neighbor of i and mj ∈ Mi (i.e., the message that triggered r′ was sent in connection
with the occurrence of r).
We interpret (r, r′) ∈ B as meaning that r happened at node i immediately before r′. If (a) is the
case, then this use of "before" seems natural because both ti and tj are relative to the same (local) time
basis. Extending this interpretation to case (b), though apparently less natural, is still consistent because
there is no reference in this case to either ti or tj. Closing B under transitivity yields another binary
relation on R, here denoted by B+, such that B ⊆ B+ ⊆ R2. This relation can be interpreted in such
a way that (r, r′) ∈ B+ means that r happened before r′ regardless of how close the specific nodes at
which they occurred are to each other in graph S. Through B+, therefore, relation B is the core of a
characterization of how the events of run R relate to one another causally. In particular, if r and r′ are
such that neither (r, r′) ∈ B+ nor (r′, r) ∈ B+, then the situation is analogous to the space-like separation
of events in special relativity (since neither could the occurrence of r influence that of r′ during run R,
nor conversely).
Relation B is also instrumental in our characterization of synchronization in run R. The first step
is to recognize that it naturally gives rise to a directed graph, call it P , whose node set is R (the set of
events) and whose edge set is B. This graph is necessarily acyclic (no directed cycles are formed) and
allows the definition of event r's depth, denoted by depth(r), as follows. Given a directed path in graph P
between two events, let the path's length be defined as the number of edges on the path that correspond
to messages. That is, the edges contributing to the path's length are those that fall under case (b) above
in the definition of relation B. We define depth(r) as the length of the lengthiest path leading to r in
graph P . Intuitively, depth(r) is the size of the longest causal chain of messages leading up to event r
during run R. These notions are illustrated in Figure 1.
Measures of synchronization
We use two measures of synchronization. They are both based on the premise that, if the sequence
of events at two neurons (nodes of S) can be aligned with each other so that sufficiently many event
pairs of the same depth become matched in the alignment, then there is more synchronism between
the two sequences than there would be with fewer event pairs aligned. We do this sequence alignment
as described next. Let i and j be the two neurons in question and let their event sequences during
j , e2
run R be e1
j , respectively, where Li and Lj are the sequences' sizes. Let
µij = max{depth(eLi
j )}, i.e., µij is the depth of the last event of the two sequences that is deep-
est. We do the alignment of the two sequences by creating two new size-µij sequences, viz. t1, t2, . . . , tµij
and u1, u2, . . . , uµij for the first measure, and x1, x2, . . . , xµij and y1, y2, . . . , yµij for the second mea-
sure, each with defining characteristics that depend on the particular measure of synchronization under
consideration.
and e1
i ), depth(eLj
j , . . . , eLj
i , e2
i , . . . , eLi
i
The first measure aims to capture the synchronization that may exist in the overall flow of messages
and, through it, in the accumulation of potential at the neurons' membranes. As such, it is based on
positioning, relative to the t sequence, those of the Li events of neuron i that promoted the accumulation
of potential, and likewise the Lj events of neuron j relative to the u sequence. For k = 1, 2, . . . , µij , the
t sequence is defined recursively as follows. If k = 1:
• tk = 1, if depth(e1
i ) = 1;
• tk = 0, otherwise.
If k > 1:
• tk = k, if depth(eℓ
i ) = k for some ℓ ∈ {1, 2, . . . , Li};
• tk = tk−1, otherwise.
6
The u sequence is defined entirely analogously for neuron j.
If, for example, the events in the two original sequences have depths 2, 3, 3, 7, 8, 9, 11 and 1, 3, 4, 5, 5, 9
for i and j, respectively, then the t sequence is
and the u sequence is
t = h0, 2, 3, 3, 3, 3, 7, 8, 9, 9, 11i
u = h1, 1, 3, 4, 5, 5, 5, 5, 9, 9, 9i.
(1)
(2)
Thus, the kth position of sequence t or u equals k′ such that 0 < k′ ≤ k if and only if the corresponding
neuron received at least one message of depth k′ during the run and, for k′ < k, never since did the
neuron receive a message whose depth is in the interval [k′ + 1, k]. It equals 0 otherwise.
The first measure of synchronization is denoted by ρ−
ij for neurons i and j. It is given by
ρ−
ij =
1
µij
µij
Xk=1
min{tk, uk}
max{tk, uk}
(3)
(assuming 0/0 ≡ 1). Clearly, ρ−
sequences we get ρ−
ij = 1. The example above yields
ij ∈ [0, 1] and grows with the similarity of sequences t and u. For identical
ρ−
ij =
1
11 (cid:18) 0
1
+
1
2
+
3
3
+
3
4
+
3
5
+
3
5
+
5
7
+
5
8
+
9
9
+
9
9
+
9
11(cid:19) ≈ 0.69.
(4)
The second measure addresses the synchronization possibilities that may exist as the neurons fire.
In this case only those of the Li events of neuron i that entailed firing are positioned relative to the x
sequence, and likewise the Lj events of neuron j with respect to the y sequence. For k = 1, 2, . . . , µij ,
the x sequence is such that:
• xk = k, if depth(eℓ
i ) = k for some ℓ ∈ {1, 2, . . . , Li} such that neuron i fired at the occurrence of eℓ
i ;
• xk = 0, otherwise.
The y sequence is defined analogously for neuron j.
Following up on the examples given above, and assuming that neuron i fired at one of its depth-3
events and neuron j fired at all its events except those of depth 5, the x sequence is
and the y sequence is
x = h0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0i
y = h1, 0, 3, 4, 0, 0, 0, 0, 9, 0, 0i.
(5)
(6)
Thus, the kth position of sequence x or y equals k > 0 if and only if the corresponding neuron fired upon
receiving a depth-k message. It equals 0 otherwise.
For neurons i and j, the second measure of synchronization is denoted by ρ+
ij and given by
ρ+
ij =
1
µij
µij
Xk=1
min{xk, yk}
max{xk, yk}
(7)
(assuming 0/0 ≡ 1). As in the previous case, ρ+
y. Identical sequences yield ρ+
ij = 1. For the example above we get
ij ∈ [0, 1] and grows with the similarity of sequences x and
ρ+
ij =
1
11 (cid:18) 0
1
+
0
0
+
3
3
+
0
4
+
0
0
+
0
0
+
0
0
+
0
0
+
0
9
+
0
0
+
0
0(cid:19) ≈ 0.73.
(8)
ij and ρ+
7
For a fixed pair i, j of neurons, both ρ−
ij seek to characterize the possibility of synchronized
behavior during run R. They do this by quantifying the extent to which the events occurring at the two
neurons could be said to happen synchronously if a time basis existed common to neurons i and j and
time according to this basis elapsed along increasing event depth. Of course, our cortical model per se
has no need for such a time basis and is algorithmically correct regardless of any timing assumptions one
may make concerning the delay for action-potential propagation down the axons. However, the kind of
synchronized behavior we are investigating occurs at a variety of temporal and spatial scales, therefore
some time-related assumption is inevitable if we are to extract any meaning out of the multitude of
neuronal events. Our choice seems reasonable because, conceptually, it contrasts with the algorithm's
inherent asynchronism only minimally: Every single pair of neurons is given its own time basis for the
computation of ρ−
ij and this is reflected on the notationally explicit dependency of µij on the i, j
pair. Furthermore, viewing increasing event depth as time may even have some biological plausibility to
it. In fact, it appears that the delay for an action potential to reach the various synapses connecting out
of the same axon is independent of how much of the axon actually has to be traversed [59, 60]. In these
terms, what our assumption does is to generalize this independence for a group of axons.
ij and ρ+
Notwithstanding this common underlying feature of the two synchronization measures, they are also
markedly different in how they use event depth to assess the similarity of two event sequences. In the case
of ρ+
ij, this is done rather stringently, since only same-depth firing events and the 0/0 ratios contribute
to it. In the case of ρ−
ij, these continue to be some of the strongest contributors, but now they are joined
by any pair of same-depth events (not necessarily firing ones). Moreover, now an event's depth lingers
until a greater-depth event occurs and in the meantime continues to influence ρ−
ij .
Results
Our computational results are based on the same methodology we followed in [13], of which we now
provide a brief review for the reader's benefit. We use v0 = −15, vt = 0, δ = 0.0002, and α = 0.04 in
all runs of algorithm A on S. Each run starts at m = 50 initiators chosen uniformly at random and
progresses until termination. A run is implemented as a sequential program that, initially, selects an
initiator randomly out of the m that were chosen for the run, lets it fire, and queues up the messages
it sends for reception by the destination nodes. Subsequently, after all initiators have had a chance to
proceed in this way, a list is maintained containing all nodes with nonempty input-message queues. One
of them is selected at random and the processing of its head-of-queue message is carried out (again with
the possible queuing of messages for consumption by other nodes). This is repeated until all queues are
empty.
The directed graph S can be of one of types (i) -- (iii), as explained next. In order to keep the com-
putational effort reasonably bounded with current technology, the directed graph D of which S is the
GSCC has n = 100 nodes.
(i) In this case D is sampled from the cortical model described above. As explained, the corresponding
S is expected to have N ≈ 90 nodes. The expected in- or out-degree in D is 3.7.
(ii) In this case D is sampled from the generalized Erdos-R´enyi model of directed graphs [61]. Given
the expected in- or out-degree z, an edge is placed from each node i to each node j 6= i with
probability z/(n − 1). For consistency with type-(i) graphs we use z = 3.7, in which case S is such
that N ≈ 100 with high probability.
(iii) In this case D has a deterministic structure and is by construction strongly connected. So S = D
and N = 100. The structure of D is that of the directed circulant graph [62] with in- or out-degree
equal to ⌈3.7⌉ = 4. If we number the n nodes 0, 1, . . . , n − 1, then in D every node i has nodes i + 1,
i + 2, i + 3, and i + 4 (modulo n) as out-neighbors.
8
For each fixed S, 0.2N nodes are chosen uniformly at random to be inhibitory, so long as no two of these
nodes are connected by an edge [note that, if S is of type (iii), then all 20 inhibitory nodes are in fact
placed deterministically, since necessarily they are found at equal intervals as we traverse the nodes in
the order 0, 1, . . . , n − 1]. Moreover, for each S node potentials and synaptic weights are chosen uniformly
at random from the intervals [v0, vt] and [0, 1], respectively. For type-(iii) graphs this is the only source
of nondeterminism.
We use 50 S instances of each type. For each fixed S we use 50 000 run sequences of algorithm A, each
sequence comprising 10 000 runs. The first run in a sequence starts from the node potentials and synaptic
weights chosen for the graph. Each subsequent run in the sequence starts from the node potentials and
synaptic weights at which the previous run ended. Along each sequence we observe the behavior of ρ−
ij
and ρ+
ij for each pair i, j of distinct nodes at six checkpoints. The first of these occurs before any run
actually takes place. Each of the remaining five occurs after 2 000 additional runs have elapsed. The
observation that takes place at a checkpoint is based on 100 additional runs, called side runs, each starting
at its own set of m randomly chosen initiators and from the node potentials and synaptic weights that
are current at the checkpoint. At the end of the side runs, the main sequence of runs is resumed from
these same node potentials and synaptic weights.
It is important to note that this computational setup would entail a considerable amount of processing
even if the side runs were excluded, since for fixed S algorithm A would be run to completion 5×108 times.
These runs are grouped into sequences so that the cumulative action of the model's weight-update rule
can be effected, but we also need many independent sequences to account for the inherent nondeterminism
of our model's asynchronous setting. As defined, however, our synchronization measures are properties
of a single run, so they too need to be averaged out over many runs. We might have chosen to do so
directly over the runs that correspond to checkpoints, that is, over 50 000 runs per checkpoint (one for
each sequence). Each of these runs, however, starts at the set of node potentials and synaptic weights
that are current in its sequence, so for each such set only one run would take place. Introducing side runs
has been a means to increase this number to 100, with the consequence of elevating the overall number
of runs for fixed S to 5.3 × 108.
The contribution of each side run concerning a fixed pair i, j of distinct nodes is to record the values
ij and ρ+
of ρ−
ij at the end of the run, as well as tag them with the labels δmin = min{δij, δji} and
δmax = max{δij, δji}, where δij and δji are the directed distances between the two nodes in S (from i
to j and from j to i, respectively). After all 2.5 × 108 side runs for a graph type have been completed
at a checkpoint, we calculate the average values of ρ−
ij over all node pairs having the same tags.
These averages are henceforth denoted by ρ− and ρ+, respectively.
ij and ρ+
Our results are presented in Figures 2 and 3 for type-(i) graphs, Figures 4 and 5 for type-(ii) graphs,
and Figures 6 and 7 for type-(iii) graphs. The former figure in each pair refers to ρ−, the latter figure
to ρ+. Each figure comprises six panels, each panel for each of the six observational checkpoints. The A
panels refer to the first checkpoints, the B panels to the second checkpoints, and so on. Each panel is
organized as a two-dimensional array and gives the ρ− or ρ+ averages as a function of the node pairs' δmin
values (as abscissas) and δmax values (as ordinates). We display these averages by means of a color code
that assigns different colors to different intervals inside [0, 1] suitably. The hues we use vary from a dark
shade of red to a dark shade of purple, indicating the lowest interval and the highest one, respectively. We
note that this choice of colors is the same through all the figures and that the colors always correspond
to the same intervals.
Each average displayed in these figures refers to directed cycles in the S graphs whose length is
δmin + δmax for the particular δmin and δmax values in question. Because of the strongly connected nature
of S, every two nodes belong to at least one common directed cycle. By the definitions of δmin and δmax,
each average plotted in the figures is therefore relative to all node pairs for which the shortest of these
cycles has the same length. In particular, traversing any of the array diagonals for which δmin + δmax
is a constant merely shifts the relative positions of the two nodes involved in each pair on the shortest
9
directed cycle that they share. We henceforth refer to the value of δmin + δmax for a certain node pair as
the pairs' girth.3 Moreover, we refer to the pair as being more or less balanced on the directed cycle of
length δmin + δmax, depending respectively on whether δmax − δmin is close to 0 or not (i.e., close to the
δmin = δmax diagonal of the array).
All panels in Figures 2 through 7 display their data inside the upper triangular region relative to the
δmin = δmax diagonal of the array. All blank spots inside this region refer to δmin, δmax pairs that never
occurred in any of the S instances we used. This can be verified by resorting to Figure 8, where the
probability distributions for the occurrence of these pairs are shown for type-(i) and type-(ii) graphs (in
panels A and B respectively, through the use of color codes similar to those of the previous figures). As
for type-(iii) graphs, it follows easily from their definition that either δmin + δmax = N/4 or δmin + δmax =
N/4 + 1, respectively 25 or 26 for N = 100, so in this case these fixed girth values are the constraints
determining the appearance of blank spots (that is, they appear when the constraints are violated).
Figure 8 is also useful in helping elucidate which of the various possible girth values in type-(i) and
type-(ii) graphs are the most common. Readily, girths of about 18 or less are by far the most common in
type-(i) graphs. This value becomes about 12 for type-(ii) graphs. In the forthcoming discussion, we use
these rough delimiters to characterize what happens to most node pairs (i.e., those whose girth values
are overwhelmingly the most common).
Discussion
The results for ρ−, given in Figures 2, 4, and 6, provide a clear picture of what is to be expected regarding
the overall synchronization that may be present in the flow of messages as they get received at the neurons.
In the case of type-(i) graphs (Figure 2), already at the first checkpoint most node pairs i, j have ρ−
ij
values in the interval (0.7, 0.85]. Four thousand runs later (that is, at the third checkpoint) this holds
for the interval (0.9, 0.95]. At the sixth and last checkpoint, the interval is (0.9, 1]. A closely analogous
conclusion holds in the case of type-(ii) graphs (Figure 4), now with the intervals (0.6, 0.85] for the first
checkpoint, (0.9, 0.95] for the third, and (0.9, 1] for the last one. As for type-(iii) graphs (Figure 6),
their rigidly constrained girth values lead ρ−
ij to be concentrated inside the interval (0.7, 0.75] at the first
checkpoint for nearly all node pairs. Similarly, already at the third checkpoint the situation that we
observe in the last checkpoint has been established and ρ−
ij is concentrated in the interval (0.9, 0.95] for
nearly all node pairs.
It is curious to observe for type-(iii) graphs that, at all checkpoints, all node pairs of girth 26 for which
δmin = 1 have ρ−
ij values occupying intervals one or two notches above the intervals we gave for nearly
all pairs, that is, (0.8, 0.85] for the first checkpoint and (0.95, 1] for the third checkpoint and onwards.
Revisiting Figures 2 and 4, we see that a similar conclusion holds also for type-(i) and type-(ii) graphs:
Node pairs for which δmin is near 1 tend to have a slightly higher ρ−
ij value if their girth is sufficiently
large.
The results for ρ+, which relates to how much synchronization may be present as neurons fire, tell a
story that differs from that of ρ− in important ways. The first of these differences is clear from Figures 3,
5, and 7: For all graph types the range of occurring ρ+ values is wider by roughly 50 -- 100% than that
of the ρ− values. In fact, the ρ+ values are now scattered inside the (0.4, 1] interval for all graph types
at the earliest checkpoints and inside (0.5, 1] at the latest ones. Readily, then, according to our two
measures of synchronization there appear to be substantially fewer synchronization possibilities in the
firing of neurons than in the accumulation of potential as reflected by the reception of messages. This
can be quantified by examining the data more closely, as follows.
For type-(i) graphs (Figure 3), at the first checkpoint most node pairs have ρ+
ij values in the interval
(0.4, 0.8]. At the last checkpoint, if we ignore all node pairs for which δmin = 1 for the time being then
3In a free extension, to a node pair, of the homonymous notion in graph theory that concerns the entire graph in the
undirected case [63].
10
nearly all node pairs have ρ+
ij values in the interval (0.9, 1] [the single exception is that of δmin = δmax = 2,
for which the interval is (0.85, 0.9]]. The case of type-(ii) graphs (Figure 5), still disregarding all δmin = 1
entries, is closely analogous to that of type-(i) graphs, the differences being that now most node pairs
span the larger interval (0.4, 1] already at the first checkpoint and that, without exception, at the last
checkpoint all node pairs hit the interval (0.9, 1]. Finally, if we go on focusing on node pairs for which
δmin > 1 exclusively, then for type-(iii) graphs (Figure 7) all node pairs have ρ+
ij values in the interval
(0.45, 0.5] at the first checkpoint. At the last checkpoint, on the other hand, this interval becomes
(0.85, 0.9] if δmin = 2, (0.9, 0.95] if δmin > 2.
As Figures 3, 5, and 7 demonstrate, the δmin = 1 case sets itself apart from the others for all graph
types at nearly all checkpoints. For example, if we concentrate on the last checkpoint, at which we believe
the model's dynamics to have already settled into some sort of stationary regime [12], then for both type-
(i) and type-(ii) graphs it holds that ρ+
ij tends to increase from some value in the interval (0.55, 0.6] when
the girth is 2 to some value in the interval (0.9, 1] when the girth is close to the rough upper bound we set
earlier for declaring most node pairs to have been counted [i.e., roughly 18 for type (i), roughly 12 for type
(ii)]. Naturally, increasing the girth while δmin is held fixed at 1 implies considering node pairs that are
progressively more imbalanced, since δmax grows with the girth. The case of type-(iii) graphs is different
in this respect, as for δmin = 1 all node pairs have ρ+
ij values in the interval (0.6, 0.65], regardless of which
value of δmax is in question (either 24 or 25). These pairs, however, are of course highly imbalanced as
well.
As we mentioned above, a notion that is becoming increasingly central to the study of synchronization
in the brain is that of feedback loops, both in a temporal sense (as LFPs and neuronal firing patterns
exert influence on one another) and in a spatial sense (as the LFPs of spatially separated areas affect
one another). The results we have obtained with our algorithmic model of a cortex lend support to this
notion both in the temporal and in the spatial sense.
In the temporal sense we have found ample evidence that our distributed algorithm A is capable
of promoting abundant opportunities both for potential to be accumulated in a synchronized way as
messages are received at the neurons and for neurons to fire in a synchronized manner. We have found
that this holds across all three graph types, from the cortical model first introduced in [12], to an Erdos-
R´enyi directed graph, to the completely deterministic and tautly shaped structure of a directed circulant
graph. In our view, this independence from the graph's structural characteristics points at an inherent
ability of algorithm A at providing some of the elements that help give rise to brain synchronization.
Notwithstanding this, our results also do shed some light on the role played by graph structure. As it
happens, of the nondeterministic graph types only type-(i) graphs provide the opportunity of long-distance
(in the sense of graph distances) synchronization in the two senses we have studied, since distances in
type-(ii) graphs are significantly shorter.
In the spatial sense there are two important issues to be highlighted. The first one is that, although
by Figure 8 node pairs having higher-than-18 girth are rare, they do occur and have yielded high ρ−
and ρ+ values at all the observational checkpoints. The exact significance this may have in the case
of real cortices is unknown, to the best of our knowledge, since spatial feedback loops are known only
for very small graph distances [2, 3]. So the fact that they may also occur at significantly larger graph
distances remains a tantalizing possibility. The second important issue is the presence of such strong
dependence of ρ+ on a node pair's girth when δmin = 1 as we observed. Our results indicate that in
this case the synchronization of neuronal spikes is favored on feedback loops involving highly imbalanced
pairs of neurons (i.e., node pairs for which δmax ≫ 1). As with the first issue, the significance this may
have for real cortices is unknown and merits special attention as further data are obtained.
All our results depend strongly on the measures of synchronization we gave in Equations (3) and (7).
They also depend on the model summarized above, but that is now backed up by interesting validating
finds [12, 13] and has therefore proven its usefulness as an artificial-life abstraction. It then seems that
furthering our study of emerging synchronization properties depends on validating our two measures in
11
a way that ties our causality-based definitions to real data as tightly as possible. We expect to be able
to do this as further insight into real cortices becomes available.
Acknowledgments
We acknowledge partial support from CNPq, CAPES, and a FAPERJ BBP grant.
References
1. Nunez PL, Srinivasan R (2006) Electric Fields of the Brain: The Neurophysics of EEG. New York,
NY: Oxford University Press.
2. Berens P, Logothetis NK, Tolias AS (2012) Local field potentials, BOLD and spiking activity:
In: Kriegeskorte N, Kreiman G, editors, Visual
Relationships and physiological mechanisms.
Population Codes: Towards a Common Multivariate Framework for Cell Recording and Functional
Imaging, Cambridge, MA: The MIT Press. pp. 599 -- 623.
3. Fries P (2005) A mechanism for cognitive dynamics: Neuronal communication through neuronal
coherence. Trends Cogn Sci 9: 474 -- 480.
4. Canolty RT, Ganguly K, Kennerley SW, Cadieu CF, Koepsell K, et al. (2010) Oscillatory phase
coupling coordinates anatomically dispersed functional cell assemblies. Proc Natl Acad Sci USA
107: 17356 -- 17361.
5. Canolty RT, Knight RT (2010) The functional role of cross-frequency coupling. Trends Cogn Sci
14: 506 -- 515.
6. Katzner S, Nauhaus I, Benucci A, Bonin V, Ringach DL, et al. (2009) Local origin of field potentials
in visual cortex. Neuron 61: 35 -- 41.
7. Huerta R, Bazhenov M, Rabinovich MI (1998) Clusters of synchronization and bistability in lattices
of chaotic neurons. Europhys Lett 43: 719 -- 724.
8. Lago-Fern´andez LF, Huerta R, Corbacho F, Siguenza JA (2000) Fast response and temporal co-
herent oscillations in small-world networks. Phys Rev Lett 84: 2758 -- 2761.
9. Masuda N, Aihara K (2004) Global and local synchrony of coupled neurons in small-world networks.
Biol Cybern 90: 302 -- 309.
10. Barbour B, Brunel N, Hakim V, Nadal JP (2007) What can we learn from synaptic weight distri-
butions? Trends Neurosci 30: 622 -- 629.
11. Song S, Sjostrom PJ, Reigl M, Nelson S, Chklovskii DB (2005) Highly nonrandom features of
synaptic connectivity in local cortical circuits. PLoS Biol 3: e68.
12. Nathan A, Barbosa VC (2010) Network algorithmics and the emergence of the cortical synaptic-
weight distribution. Phys Rev E 81: 021916.
13. Nathan A, Barbosa VC (2011) Network algorithmics and the emergence of information integration
in cortical models. Phys Rev E 84: 011904.
14. Balduzzi D, Tononi G (2008) Integrated information in discrete dynamical systems: Motivation
and theoretical framework. PLoS Comput Biol 4: e1000091.
12
15. Barbosa VC (1996) An Introduction to Distributed Algorithms. Cambridge, MA: The MIT Press.
16. Fisher J, Henzinger TA (2007) Executable cell biology. Nat Biotechnol 25: 1239 -- 1249.
17. Fisher J, Harel D, Henzinger TA (2011) Biology as reactivity. Commun ACM 54(10): 72 -- 82.
18. Forbes N (2000) Life as it could be: Alife attempts to simulate evolution. IEEE Intell Syst 15(6):
2 -- 7.
19. Lindley D (2010) Brains and bytes. Commun ACM 53(9): 13 -- 15.
20. Sporns O, Chialvo DR, Kaiser M, Hilgetag CC (2004) Organization, development and function of
complex brain networks. Trends Cogn Sci 8: 418 -- 425.
21. Sporns O, Tononi G, Kotter R (2005) The human connectome: A structural description of the
human brain. PLoS Comput Biol 1: 245 -- 251.
22. Achard S, Salvador R, Whitcher B, Suckling J, Bullmore E (2006) A resilient, low-frequency, small-
world human brain functional network with highly connected association cortical hubs. J Neurosci
26: 63 -- 72.
23. Bassett DS, Bullmore E (2006) Small-world brain networks. Neuroscientist 12: 512 -- 523.
24. He Y, Chen ZJ, Evans AC (2007) Small-world anatomical networks in the human brain revealed
by cortical thickness from MRI. Cereb Cortex 17: 2407 -- 2419.
25. Honey CJ, Kotter R, Breakspear M, Sporns O (2007) Network structure of cerebral cortex shapes
functional connectivity on multiple time scales. Proc Natl Acad Sci USA 104: 10240 -- 10245.
26. Reijneveld JC, Ponten SC, Berendse HW, Stam CJ (2007) The application of graph theoretical
analysis to complex networks in the brain. Clin Neurophysiol 118: 2317 -- 2331.
27. Sporns O, Honey CJ, Kotter R (2007) Identification and classification of hubs in brain networks.
PLoS ONE 2: e1049.
28. Stam CJ, Reijneveld JC (2007) Graph theoretical analysis of complex networks in the brain. Non-
linear Biomed Phys 1: 3.
29. Yu S, Huang D, Singer W, Nikoli´c D (2008) A small world of neuronal synchrony. Cereb Cortex
18: 2891 -- 2901.
30. Chialvo DR (2010) Emergent complex neural dynamics. Nat Phys 6: 744 -- 750.
31. Modha DS, Ananthanarayanan R, Esser SK, Ndirango A, Sherbondy AJ, et al. (2011) Cognitive
computing. Commun ACM 54(8): 62 -- 71.
32. Bornholdt S, Schuster HG, editors (2003) Handbook of Graphs and Networks. Weinheim, Germany:
Wiley-VCH.
33. Newman M, Barab´asi AL, Watts DJ, editors (2006) The Structure and Dynamics of Networks.
Princeton, NJ: Princeton University Press.
34. Bollob´as B, Kozma R, Mikl´os D, editors (2009) Handbook of Large-Scale Random Networks. Berlin,
Germany: Springer.
35. Abeles M (1991) Corticonics: Neural Circuits of the Cerebral Cortex. Cambridge, UK: Cambridge
University Press.
13
36. Ananthanarayanan R, Modha DS (2007) Anatomy of a cortical simulator. In: Proc. of the 2007
ACM/IEEE Conf. on Supercomputing. New York, NY: ACM, p. 3.
37. Ananthanarayanan R, Esser SK, Simon HD, Modha DS (2009) The cat is out of the bag: Cortical
In: Proc. of the Conf. on High Performance
simulations with 109 neurons and 1013 synapses.
Computing Networking, Storage and Analysis. New York, NY: ACM, p. 63.
38. Dorogovtsev SN, Mendes JFF, Samukhin AN (2001) Giant strongly connected component of di-
rected networks. Phys Rev E 64: 025101(R).
39. Newman MEJ (2005) Power laws, Pareto distributions and Zipf's law. Contemp Phys 46: 323 -- 351.
40. Egu´ıluz VM, Chialvo DR, Cecchi GA, Baliki M, Apkarian AV (2005) Scale-free brain functional
networks. Phys Rev Lett 94: 018102.
41. van den Heuvel MP, Stam CJ, Boersma M, Hulshoff Pol HE (2008) Small-world and scale-free
organization of voxel-based resting-state functional connectivity in the human brain. Neuroimage
43: 528 -- 539.
42. Modha DS, Singh R (2010) Network architecture of the long-distance pathways in the macaque
brain. Proc Natl Acad Sci USA 107: 13485 -- 13490.
43. Bullmore E, Sporns O (2009) Complex brain networks: Graph theoretical analysis of structural
and functional systems. Nat Rev Neurosci 10: 186 -- 198.
44. Erdos P, R´enyi A (1959) On random graphs. Publ Math (Debrecen) 6: 290 -- 297.
45. Freeman WJ, Kozma R, Bollob´as B, Riordan O (2009) Scale-free cortical planar networks. In:
Bollob´as et al. [34], pp. 277 -- 324.
46. Sporns O (2011) Networks of the Brain. Cambridge, MA: The MIT Press.
47. Amaral LAN, Scala A, Barth´elemy M, Stanley HE (2000) Classes of small-world networks. Proc
Natl Acad Sci USA 97: 11149 -- 11152.
48. Freeman WJ (2007) Scale-free neocortical dynamics. Scholarpedia 2(2): 1357.
49. Honey CJ, Sporns O, Cammoun L, Gigandet X, Thiran JP, et al. (2009) Predicting human resting-
state functional connectivity from structural connectivity. Proc Natl Acad Sci USA 106: 2035 --
2040.
50. Kaiser M, Hilgetag CC (2004) Modelling the development of cortical systems networks. Neuro-
computing 58 -- 60: 297 -- 302.
51. Kaiser M, Hilgetag CC (2004) Spatial growth of real-world networks. Phys Rev E 69: 036103.
52. Abbott LF, Nelson SB (2000) Synaptic plasticity: Taming the beast. Nat Neurosci 3: 1178 -- 1183.
53. Song S, Miller KD, Abbott LF (2000) Competitive Hebbian learning through spike-timing-
dependent synaptic plasticity. Nat Neurosci 3: 919 -- 926.
54. Bi GQ, Poo MM (1998) Synaptic modifications in cultured hippocampal neurons: Dependence on
spike timing, synaptic strength, and postsynaptic cell type. J Neurosci 18: 10464 -- 10472.
55. Bi GQ, Poo MM (2001) Synaptic modification by correlated activity: Hebb's postulate revisited.
Annu Rev Neurosci 24: 139 -- 166.
14
56. Kepecs A, van Rossum MCW (2002) Spike-timing-dependent plasticity: Common themes and
divergent vistas. Biol Cybern 87: 446 -- 458.
57. Henriksen RN (2011) Practical Relativity: From First Principles to the Theory of Gravity. Chich-
ester, UK: Wiley.
58. Lamport L (1978) Time, clocks, and the ordering of events in a distributed system. Commun ACM
21: 558 -- 565.
59. Innocenti GM, Lehmann P, Houzel JC (1994) Computational structure of visual callosal axons.
Eur J Neurosci 6: 918 -- 935.
60. Salami M, Itami C, Tsumoto T, Kimura F (2003) Change of conduction velocity by regional
myelination yields constant latency irrespective of distance between thalamus and cortex. Proc
Natl Acad Sci USA 100: 6174 -- 6179.
61. Karp RM (1990) The transitive closure of a random digraph. Random Struct Algor 1: 73 -- 93.
62. Lonc Z, Parol K, Wojciechowski JM (2001) On the number of spanning trees in directed circulant
graphs. Networks 37: 129 -- 133.
63. Bollob´as B (1998) Modern Graph Theory. New York, NY: Springer.
15
k1
k3
k2
k4
A
B
n
o
r
u
e
N
k1
k2
k3
k4
0
1
2
3
4
5
6
7
8
9
10
Depth
Figure 1. Directed graphs S and P . A run of algorithm A on graph S (A) started by the single
initiator k1 may result in the set of events shown in panel B as the node set of graph P . Events are
shown in panel B either as filled circles, in the case of firing events, or as empty circles, otherwise. Each
event is positioned against a background grid that highlights the node of S at which it occurred (this is
given by the grid's rows) and its depth (the grid's columns). An edge of P either joins consecutive
events occurring at the same node of S or corresponds to a message. We omit the former from panel B
for clarity. An event's depth is the greatest number of message edges on a directed path arriving at it.
Every event (except the first one at the initiator) has exactly one incoming message edge. If it is a firing
event, then it also has as many outgoing message edges as the corresponding node in graph S has
out-neighbors. Whenever two or more events occurring at the same node of S have the same depth,
they are shown as close to the correct grid point as possible.
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
x
a
m
A
C
E
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
d min
5
10
15
20
5
10
15
20
5
10
15
20
16
B
D
F
5
10
15
20
5
10
15
20
5
10
15
20
Figure 2. Average value of ρ−
given for the first checkpoint (A) through the sixth (F).
ij for type-(i) graphs as a function of δmin and δmax. Data are
d
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
x
a
m
A
C
E
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
d min
5
10
15
20
5
10
15
20
5
10
15
20
17
B
D
F
5
10
15
20
5
10
15
20
5
10
15
20
Figure 3. Average value of ρ+
given for the first checkpoint (A) through the sixth (F).
ij for type-(i) graphs as a function of δmin and δmax. Data are
d
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
x
a
m
A
C
E
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
d min
5
10
15
20
5
10
15
20
5
10
15
20
18
B
D
F
5
10
15
20
5
10
15
20
5
10
15
20
Figure 4. Average value of ρ−
given for the first checkpoint (A) through the sixth (F).
ij for type-(ii) graphs as a function of δmin and δmax. Data are
d
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
x
a
m
A
C
E
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
d min
5
10
15
20
5
10
15
20
5
10
15
20
19
B
D
F
5
10
15
20
5
10
15
20
5
10
15
20
Figure 5. Average value of ρ+
given for the first checkpoint (A) through the sixth (F).
ij for type-(ii) graphs as a function of δmin and δmax. Data are
d
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
x
a
m
A
C
E
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
0.4 < r − £ 0.45
0.45 < r − £ 0.5
0.5 < r − £ 0.55
0.55 < r − £ 0.6
0.6 < r − £ 0.65
0.65 < r − £ 0.7
0.7 < r − £ 0.75
0.75 < r − £ 0.8
0.8 < r − £ 0.85
0.85 < r − £ 0.9
0.9 < r − £ 0.95
0.95 < r − £ 1
d min
5
10
15
20
5
10
15
20
5
10
15
20
20
B
D
F
5
10
15
20
5
10
15
20
5
10
15
20
Figure 6. Average value of ρ−
given for the first checkpoint (A) through the sixth (F).
ij for type-(iii) graphs as a function of δmin and δmax. Data are
d
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
x
a
m
A
C
E
25
20
15
10
5
25
20
15
10
5
25
20
15
10
5
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
0.4 < r + £ 0.45
0.45 < r + £ 0.5
0.5 < r + £ 0.55
0.55 < r + £ 0.6
0.6 < r + £ 0.65
0.65 < r + £ 0.7
0.7 < r + £ 0.75
0.75 < r + £ 0.8
0.8 < r + £ 0.85
0.85 < r + £ 0.9
0.9 < r + £ 0.95
0.95 < r + £ 1
d min
5
10
15
20
5
10
15
20
5
10
15
20
21
B
D
F
5
10
15
20
5
10
15
20
5
10
15
20
Figure 7. Average value of ρ+
given for the first checkpoint (A) through the sixth (F).
ij for type-(iii) graphs as a function of δmin and δmax. Data are
d
22
B
(0.0001, 0.0003]
(0.0003, 0.001]
(0.001, 0.003]
(0.003, 0.01]
(0.01, 0.03]
(0.03, 0.1]
25
20
15
10
5
x
a
m
A
(0.0001, 0.0003]
(0.0003, 0.001]
(0.001, 0.003]
(0.003, 0.01]
(0.01, 0.03]
(0.03, 0.1]
25
20
15
10
5
5
10
15
20
5
10
15
20
d min
Figure 8. Probability distribution of δmin, δmax pairs for the type-(i) (A) and type-(ii) (B)
graphs we used.
d
|
1509.05556 | 1 | 1509 | 2015-09-18T09:16:12 | Firing dynamics of an autaptic neuron | [
"q-bio.NC",
"physics.bio-ph"
] | Autapses are synapses that connect a neuron to itself in the nervous system. Previously, both experimental and theoretical studies have demonstrated that autaptic connections in the nervous system have a significant physiological function. Autapses in nature provide self-delayed feedback, thus introducing an additional time scale to neuronal activities and causing many dynamic behaviors in neurons. Recently, theoretical studies have revealed that an autapse provides a control option for adjusting the response of a neuron: e.g., an autaptic connection can cause the electrical activities of the Hindmarsh-Rose neuron to switch between quiescent, periodic, and chaotic firing patterns; an autapse can enhance or suppress the mode-locking status of a neuron injected with sinusoidal current; and the firing frequency and interspike interval distributions of the response spike train can also be modified by the autapse. In this paper, we review recent studies that showed how an autapse affects the response of a single neuron. | q-bio.NC | q-bio |
Firing dynamics of an autaptic neuron
Hengtong Wang1 and Yong Chen2, ∗
1School of Physics and Information Technology, Shaanxi Normal University, Xi'an 710119, China
2Center of soft matter physics and its application, Beihang University, Beijing 100191, China
School of Physics and Nuclear Energy Engineering, Beihang University, Beijing 100191, China
Autapses are synapses that connect a neuron to itself in the nervous system. Previously, both
experimental and theoretical studies have demonstrated that autaptic connections in the nervous
system have a significant physiological function. Autapses in nature provide self-delayed feedback,
thus introducing an additional time scale to neuronal activities and causing many dynamic behaviors
in neurons. Recently, theoretical studies have revealed that an autapse provides a control option for
adjusting the response of a neuron: e.g., an autaptic connection can cause the electrical activities
of the Hindmarsh-Rose neuron to switch between quiescent, periodic, and chaotic firing patterns;
an autapse can enhance or suppress the mode-locking status of a neuron injected with sinusoidal
current; and the firing frequency and interspike interval distributions of the response spike train
can also be modified by the autapse. In this paper, we review recent studies that showed how an
autapse affects the response of a single neuron.
Keywords: Autapse, Self-delay feedback, single neuron, firing pattern
PACS: 87.19.ll, 87.19.ls, 87.18.Sn, 87.19.lg
I.
INTRODUCTION
The nervous system is a complex network of neurons,
synapses, and other specialized cells, and through this
system, neurons receive and transmit information be-
tween different parts of the body [1]. The responses of
neurons and the function of synapses have received con-
siderable attention because of their status as building
blocks of the nervous system [2 -- 4]. Nearly a century
ago, neuroscientists found a special synaptic structure,
the autapse, which is a synapse between different parts
of the same neuron [5 -- 7]. Because of its odd structure,
the self-synaptic connection has remained poorly under-
stood. However, many recent studies found that autapses
are much more widespread in the nervous system than
previously thought. Autapses have been reported in var-
ious brain areas, such as the neocortex, cerebellum, hip-
pocampus, striatum, and substantia nigra [6, 8 -- 12]. In-
terestingly, about 80% of cortical pyramidal neurons have
autaptic connections [8]. Many studies have revealed that
autapses are not merely curiosities, but play an authen-
tic physiological role in the nervous system. Autapses
can maintain persistent activity in the nervous system
by mediating positive feedback [13].
Recent experiments also demonstrated that autapses
are very important for the processing function of the
brain [14]. Bacci et al.
recorded the activity of fast
spiking interneurons in acute brain slices of juvenile rats
and found that autaptic transmission increased the spike-
timing precision [15]. A somatic spike can evoke ex-
citatory postsynaptic currents with various amplitudes
through an autapse [11]. However, the function of au-
tapses and their contribution to information processing
∗ [email protected]
are still unclear [14, 16]. Thus, understanding the effect
of autaptic activities on the responses of a neuron is a
fundamental step to elucidating the process of informa-
tion transfer in the nervous system [17 -- 20].
In fact, natural autapses are self-feedback connections
in the nervous system and may allow a unique type of
self-control, similar to the feedback in other systems [14].
Actually, feedback creates a circuit, or loop, that con-
nects a system to itself and commonly occurs in many
systems, e.g., gene regulatory networks [21, 22], popula-
tion dynamics [23], and climate systems [24]. Feedback
is also used extensively to control the state of a non-
linear system, such as to stabilize periodic orbits or to
control coherence resonance [25, 26]. Feedback is also
very commonly used in the design of electric circuit el-
ements [27].
In these systems, information about the
past or the present influences the same phenomena in the
present or future, respectively. Moreover, feedback has a
marked effect on the dynamics of nonlinear systems. The
frequency at the onset of the oscillation of a nonlinear
system can be modified by the feedback loop [28]. Fur-
thermore, self-delayed feedback with a fixed delay time
can suppress chaos and also control steady states [29, 30].
In a simulation study, Popovych et al. showed that time-
delayed feedback has the ability to desynchronize groups
of model neurons [31].
A neuron with an autaptic connection can provide a
distinctive physiological self-feedback model with many
dynamic properties. Rusin et al. performed a time-
delayed feedback stimulation of a group of cultured neu-
rons and experimentally demonstrated that time-delayed
feedback could cause synchronization of the action poten-
tials of a group of neurons [32]. Prager et al. found that
time-delayed feedback can facilitate the noise-induced os-
cillation of a neuronal system [33]. Autaptic delayed
feedback was found to modify the bursting in the dis-
tribution of interspike intervals of a stochastic Hodgkin-
Huxley (HH) neuron [34]. Such autaptic delayed feed-
back can also reduce the spontaneous spiking activity of
a stochastic neuron at the characteristic frequencies. The
effect of an autapse on the spike rate of a single neuronal
system shows dependence on the duration of the autapse
activity [35]. Modulation of autaptic delay feedback can
cause the dynamic behaviors of a Hindmarsh-Rose (HR)
neuron to switch between quiescent, periodic, and chaotic
firing patterns [36].
In this article, we review the current stat of knowledge
of the effect of an autapse on single neurons.
II. AUTAPTIC CONNECTION
Scientists have long investigated the functions and
properties of neurons and synapses. Neurons vary widely
with different sizes and forms, and neuron responses are
also involved in many dynamic behaviors. A synapse is a
structure in the nervous system that permits a neuron (or
nerve cell) to transmit an electrical or chemical signal to
another cell (neural or otherwise) [37]. The self-synapse,
or the autapse, was first described as a synapse between
the axon of a pyramidal cell and its own dendrites by Van
der Loos and Glaser in 1972 [5]. Before the report by
Van der loos et al., other terms, such as 'self-excitation'
and 'self-sensing', were used to describe the self-synaptic
structure. For a long time, autapses in the nervous sys-
tem seemed like anatomical curiosities with questionable
functional significance. However, recent experiments be-
gan to reveal how autapses might play an important role
in brain function. Some reports also suggested that au-
tapses are connected to some neural diseases [38]. Neu-
robiology experiments have always focused on excitatory
(glutamate-releasing) and inhibitory (GABA-releasing)
autapses. To date, autaptic structures have been found
in almost all parts of the nervous system [6]. Fig. 1 shows
photos of the many autapses of a single rat hippocampal
neuron.
In nature, autaptic connections enable self-feedback of
a neuron. Thus, in almost all current autapse studies,
the model of a neuron with an autaptic connection con-
tains a single compartment that exhibits a self-delayed
feedback mechanism, as shown in Fig. 2. The delay time
represents the elapsed time associated with the axonal
propagation prior to the reintroduction of the signal into
the neuron. The delay time is also believed to be one of
the important properties of autaptic connections. Math-
ematical models of an autapse in previous studies of the
effect of self-synapses on neurons fall into two main cat-
egories: one has linear self-coupling and describes a gap
junction, and the other has nonlinear self-coupling and
describes a chemical synapse, which may be an excitatory
or inhibitory synapse. The specific mathematical form of
the autapse depends on the chosen neuronal model.
The autaptic connection of a simplified neuron model
(such as the HR neuron, FitzHugh-Nagumo neuron,
2
FIG. 1. A single rat hippocampal neuron (arrowed,
left)
grown in culture on a 'microdot' of glia (flat gray cells),
showing abundant autapses labeled with an antibody (small
dark spots, right) [14].
Dendritic tree
Axon
Soma
Autapse
axonal tree
FIG. 2. Schematic drawing of a neuron with autaptic feed-
back. The structure in the circle represents an autaptic con-
nection between dendritic and axonal terminals.
Izhikevich neuron and others) can often be described by
one of the following formulas [35, 36, 39]:
• Linear coupling gives as electrical diffusive-type:
Iaut = gaut[u(t − τ ) − u(t)].
(1)
gaut is the autaptic conductivity. τ is the delay
time. This form of autaptic current is proportional
to the difference between the membrane potential
at t and that at an earlier time point t − τ .
• The chemical synapse function is modeled using
the so-called fast threshold modulation (FTM)
scheme [40, 41]:
Iaut(t) = −gaut(u(t) − Vsyn)S(t − τ ),
S(t − τ ) = 1/{1 + exp[−k(u(t − τ ) − θ)]},
(2)
(3)
where gaut is the autaptic intensity, and Vsyn is the
synaptic reversal potential. Vsyn=2 and Vsyn=-2
correspond to excitatory and inhibitory autapses,
respectively.
For a conductance-based neuron model (such as the
HH neuron, Morris-Lecar neuron, Connor-Stevens neu-
ron and others), the mathematic model of the autapse is
often given as follows [34, 35, 42, 43]
• Linear coupling is also always set as electrical
diffusive-type and has the same form as Eq. 1.
• Conductance-based chemical autapses can be de-
scribed using a bi-exponential function, which has
been reported to fit well with experiment re-
sults [44],
Iaut(t) = G(t − tdelay)(V − Esyn),
(4)
where G(t − tdelay) is the autaptic conductance,
and Esyn is the autaptic reversal potential. In this
equation, we used Esyn = 0 mV for excitatory neu-
rons and -80 mV for inhibitory neurons. The au-
taptic conductance is defined as
G(t) = gautσ(t − tf ire)
σ(t) =
exp(−t/td) − exp(−t/tr)
td − tr
(5)
(6)
with presynaptic spikes occurring at tf ire. The pa-
rameters td and tr represent the decay and rise
times of the function, respectively, and these pa-
rameters determine the duration of the response.
The autaptic rise and decay times were set to
tr = 0.1 ms and td = 3 ms, respectively.
In the following sections, we will review the two main
forms of an autapse. Studies have shown that autapses
that exhibit delayed feedback provide a control option for
adjusting neuron responses.
III. FIRING PATTERN TRANSITION IN A
BURSTING NEURON
Experimental observations have indicated that action
potentials can occur with different firing patterns, and
the primary firing patterns are spiking and bursting [45 --
50]. Bursting is an extremely diverse general phe-
nomenon in firing patterns exhibited by neurons in the
central nervous system and spinal cord [51, 52]. Previ-
ous studies on a bursting neuron with an autapse showed
the novel dynamics and the transition of firing patterns
induced by an autapse [36].
The simplified model of Hindmarsh and Rose has
turned out to capture the features of experimentally mea-
sured electrical data quite accurately, particularly for
studies of the spiking-bursting behavior of neuron mem-
brane potentials.
Without an autapse, HR neurons exhibit many dy-
namic behaviors, including quiescent, regular spiking, pe-
riodic, and chaotic bursting firing patterns (see Fig. 1 in
Ref. [36] and Ref. [53]).
3
(a)
2
0
u
−2
5000
(d)
2
0
u
5200
t
5400
(b)
2
0
u
−2
5000
(e)
2
0
u
5200
t
5400
(c)
2
0
u
−2
5000
(f)
2
0
u
5200
t
5400
−2
5000
5200
t
5400
−2
5000
5200
t
5400
−2
5000
5200
t
5400
FIG. 3. Time courses of the membrane potentials in response
to different Iext with gaut = 0.5 [36]. The blue curves repre-
sent the time course of the HR neuron with an autapse, and
the red curves indicate that without autapse.
The presence of an autapse completely changes the fir-
ing patterns of the original HR neuron. The firing pat-
tern can be adjusted from a periodic or chaotic pattern to
another periodic pattern or to a chaotic bursting pattern
as the autaptic parameters change, independently of the
original firing pattern. Fig. 3 shows the time courses of
the membrane potentials of an HR neuron with an elec-
trical autapse as an example. Importantly, the maximum
action potential is increased by the electrical autapse.
The ISI plot bifurcation diagrams of the HR neuron
clearly show the transition between periodic and aperi-
odic firing. For an electrical autapse, the periodic state
transits to the chaotic state exhibiting an alternating be-
havior as the time delay increased. With higher autap-
tic intensity, this alternating behavior is more noticeable
and occurs more frequently. For shorter delay times, the
firing pattern of the HR neuron showed periodic spik-
ing independent of the external DC input. The neuron
with an excitatory chemical autapse exhibits chaotic fir-
ing patterns in a larger area of gaut-τ space than that of
the neuron with an electrical autapse. For strong exter-
nal stimuli, the chaotic region of the combinational pa-
rameters in the gaut-τ space is enlarged. The excitatory
autapse plays a positive role in generating and enhancing
chaos as a whole. As neurons with an inhibitory autaptic
connection, the chaotic spiking of the HR neuron can be
decreased and suppressed. With the proper inhibitory
autaptic parameters, the HR neuron could be driven to
a resting state. Fig. 4 shows the bifurcation of the ISIs
of an HR neuron for the three types of autapses.
Without an autapse, the firing pattern of an HR neu-
ron switches from silent to periodic bursting and then to
chaotic firing with period-doubling bifurcation as the DC
current increased [54 -- 57]. For neurons with an autapse,
the firing pattern can switch into any other firing pat-
tern, regardless of the original pattern. As a whole, there
are two main approaches to transition from a periodic to
a chaotic firing pattern: discontinuous and continuous.
The interspike interval return map is a useful approach
to characterizing the transition to chaos. Fig. 5 shows an
4
1
+
n
I
S
I
80
60
40
20
0
0
80
60
40
20
0
0
(a)
τ=2.182
20
40
60
80
(e)
τ=41.97
20
40
60
80
80
60
40
20
0
0
80
60
40
20
0
0
(b)
τ=2.183
20
40
60
80
(f)
τ=41.99
(c)
τ=2.184
20
40
60
80
(g)
τ=42.00
80
60
40
20
0
0
80
60
40
20
20
40
60
80
0
0
ISI
n
20
40
60
80
80
60
40
20
0
0
80
60
40
20
0
0
(d)
τ=2.185
20
40
60
80
(h)
τ=42.22
20
40
60
80
FIG. 5. An example of the two ways of transition (the dis-
continuous transition [upper] and the continuous transition
[bottom]) into chaos from periodic firing [36]. External stim-
uli is set as Iext=2.67.
4
3
2
(a)
t
e
a
r
g
n
i
r
i
f
l
a
n
o
(b)
15
t
e
a
r
g
n
i
r
i
f
l
a
n
o
10
5
0
0
i
t
a
r
1
i
t
a
r
0
0
10
20
t
delay
30
40
(c)
15
e
t
a
r
g
n
i
r
i
f
l
a
n
o
i
t
a
r
10
5
0
0
(d)
15
e
t
a
r
g
n
i
r
i
f
l
a
n
o
i
t
a
r
10
5
0
0
10
20
t
delay
30
40
10
20
t
delay
30
40
f=9
f=15
f=37
f=87
f=105
f=167
10
20
t
delay
30
40
FIG. 6. Mode-locking of responses of an HH autaptic neuron
with different delay times [42]. The lines with different mark-
ers denote the firing rate of the neuron excited by different
input frequencies.
tem. Thus, the neuron-autapse system may exhibit very
complex dynamics, due to the interplay between autaptic
delayed feedback, an external periodic stimulus, and the
intrinsic activity of the neuron.
In previous studies, the model neuron-autapse system
often contained an HH neuron and an autapse. The
HH model is a conductance-based model that describes
how action potentials in neurons are initiated and prop-
agated [62]. In this section, we review the mode-locking
firing pattern of an HH neuron with an electrical autapse
that has the same mathematical form as Eq. (1).
Without an autapse, the mode-locking behaviors de-
pend on the values of the stimulus frequencies and am-
plitudes. As shown in Fig. 1 in Ref [42], the Arnold
tongues in the frequency and amplitude space show the
overall features of various phase-locked states and provide
the different p : q (denoting output action potentials per
input spikes) mode-locking regions. Ref [58] also shows
FIG. 4. Bifurcation diagram of the interspike interval of
an HR neuron with an autapse versus the autaptic delay
time [36]. From left to right, the panels show the results
of the electrical autapse, excitatory chemical autapse, and in-
hibitory chemical autapse.
example of the two approaches for the transition of a neu-
ron to chaos. In the discontinuous transition, the neuron
has a periodic bursting firing pattern because the delay
time is short. When the delay time increases, the system
suddenly enters a chaotic firing state. In the continuous
transition, the neuron first displays a periodic firing pat-
tern. As the autaptic parameters change, the number
of spikes in the single burst as well as the period of the
periodic bursting increase almost continuously, and the
neuron transitions into the chaotic state.
IV. MODE-LOCKING BEHAVIOR OF A
REGULAR SPIKING NEURON
When excited by a periodic stimulus, neurons respond
with various mode-locking firing patterns and quasi-
periodic states [42, 58 -- 61]. When an autapse is present,
the mode-locking firing of a neuron can be switched to
another state, depending on the chosen autapse param-
eters. An autapse provides self-feedback and contributes
an additional time scale to the dynamic neuronal sys-
the bifurcation mechanisms that create the boundaries of
the complex mode-locking structure.
The presence of an autapse substantially modifies the
mode-locking patterns of the neuron. For the same si-
nusoidal stimulus, the neuron with an autapse can fire
more or fewer action potentials than that of a neuron
without an autapse. The activities of the neuron can
also be driven into sub-threshold oscillation. The electri-
cal autapse displays a regulating function that controls
the mode-locking firing of the neuron. The time courses
of the membrane potentials of the neuron with and with-
out an autapse are shown in Fig. 2 of Ref. [42]. With
an autapse, the modification of the mode-locking firing
pattern depends on the autaptic conductivity and the de-
lay time. When the synaptic conductivity is small, the
mode-locking p : q is similar to that without an autaptic
connection. When the synaptic conductivity is large, the
neuron displays very complex mode-locking firing. The
p : q value of mode-locked firing increases from zero to a
very large value when the frequency of the sinusoidal cur-
rent increases from zero to the "threshold" frequency (the
minimal input frequency for which the neuron fires an ac-
tion potential given a fixed input amplitude). For further
increases in the frequency of the sinusoidal current, the
p : q value of mode-locking firing decreases. In the case
of extremely strong autaptic conductivity [see Fig. 3 (d)
in Ref. [42]], the autaptic activities completely disrupt
the original mode-locking firing of the neuron without an
autapse.
For continuous changes in the delay time, the model-
locked firing of the neuron is also quite interesting.
Fig. 6 shows the dependence of the mode-locking state
on the delay time for different input frequencies and au-
taptic conductivities. For weak autaptic conductivities
[Fig. 6(a)], the p : q values of mode-locking are similar to
those without self-feedback for the short delay time. As
the delay time increases, the values of p : q mode-locking
begin to fluctuate. The smaller the input frequency is,
the larger the resulting fluctuation. For strong autap-
tic intensities [shown in Fig. 6(b-d)], the p : q values
of mode-locking decrease with increasing delay time in
a stepwise manner. With further increases in the delay
time, the p : q values of mode-locking suddenly jump to a
very large value (larger than that without autaptic self-
feedback) and then decrease smoothly. As the delay time
increases, the p : q value of mode-locking of a neuron with
a high autaptic intensity exhibits periodic behavior.
With autaptic self-feedback, the responses of an HH
neuron to a sinusoidal stimulus can have higher or lower
p : q mode-locking responses than that of a neuron with-
out an autapse. These dynamical behaviors depend on
the autaptic intensity and the delay time. Moreover, the
presence of an autapse increases the range of values for
which the HH neuron spiking is locked with the sinusoidal
current. When the autaptic intensity is weak, the mode-
locking behaviors show no marked changes. That is, an
autapse with a weak intensity will have a negligible ef-
fect on the response of the neuron. For stronger autaptic
5
120
110
100
y
c
n
e
u
q
e
r
f
g
n
i
r
i
f
90
80
70
60
10
20
t
delay
30
(ms)
40
50
120
(a)
1
)
2
m
c
/
S
m
0.8
0.6
(
g
t
u
a
0.4
0.2
0
(c)
1
0.8
0.6
)
2
m
c
/
S
m
(
g
t
u
a
0.4
y
c
n
e
u
q
e
r
f
g
n
i
r
i
f
110
100
90
80
70
60
0.2
0
10
20
t
delay
30
(ms)
40
50
FIG. 7. (Color online) Firing frequencies ( (a),(c)) and ISI
distribution ((b), (d)) of a neuron with a chemical autapse
(upper panels: excitatory autapse; lower panels:
inhibitory
autapse) excited by a DC current [43]. Dotted lines gives
the odd multiples of half the intrinsic period of a single HH
neuron excited by the corresponding DC current.
intensities, the p : q value of the mode-locking increases
and then decreases as the delay time increases, exhibiting
nearly periodic behaviors. Thus, for sufficiently strong
autaptic intensity, changing the delay time provides bet-
ter regulation of mode-locking than does changing the
autaptic intensity. The autaptic connection may also pro-
vide a control option for adjusting mode-locked firing in
a neural information process.
V. RESPONSES OF A REGULAR SPIKING
NEURON
A. Response to DC currents
When injected with a DC current, an HH neuron with-
out an autapse displays regular spiking when the current
intensity is larger than a critical current value. As for
most neuronal models, increases in the current will in-
crease the firing rate of the neuron, and the HH neu-
ron undergoes a Hopf bifurcation.
In the presence of
an autapse, the output frequency can be higher or lower
than the intrinsic frequency (the firing frequency of an
HH neuron without an autapse) when the autaptic con-
ductance and delay time are changed. Interestingly, the
output frequency, as well as the output ISI distribution,
shows periodic behaviors as the autaptic delay time in-
creases. This periodicity of the changes in the output
frequency (ISI distributions) in response to changes in
the delay time is also similar to the intrinsic period of
the neuron. When an HH neuron is connected to an elec-
trical or excitatory chemical autapse, the firing rate and
the ISI distribution of the response spike trains can both
be either amplified or depressed. When the delay time
approaches the odd multiples of the intrinsic half period
of the firing of an HH neuron with an electrical or excita-
tory chemical autapse, the ISIs of the spike train are very
different from the intrinsic ISIs. When the delay time
approaches a multiple of the intrinsic period, the ISIs of
the spike train are similar to the intrinsic ISIs. However,
the inhibitory chemical autapse can only suppress the
firing responses of the neuron. The ISIs of the response
spike train of a neuron with an inhibitory autapse are al-
ways greater than the intrinsic period, whereas the corre-
sponding response frequency is not less than the intrinsic
response frequency. Current neurobiological experiments
also reveal that autapses in the nervous system can be ei-
ther self-inhibitory or self-excitatory, depending on their
location on the neuron [6, 63 -- 65].
Because autaptic pulses perturb neural
spiking
through a process that is similar to that of an oscillating
system with a fixed-delay feedback, the phase-response
curve (PRC) theory can provide useful insight into the
phenomena of the neuron-autapse system [66 -- 69]. For an
oscillating system, the PRC describes the phase shift that
occurs in response to a brief external stimulus [70, 71].
Fig. 8 gives the phase-response curves of the three types
of autapses. The PRCs of a neuron with an electrical
autapse or an excitatory chemical autapse can be nega-
tive or positive, depending on the stimulus phase (delay
time). Thus, an electrical autapse or excitatory chemical
autapse could delay or advance the next spike. Therefore,
the firing rate of a neuron with an electrical autapse or
an excitatory chemical autapse can be higher or smaller
than that of the neuron without an autapse. For the in-
hibitory chemical autapse, however, the PRC is always
positive, because the delay time increased. That is, an in-
hibitory autapse always postpones the next spike. Thus,
the firing rate of an HH neuron with an inhibitory au-
tapse is not larger than that of the neuron without an
autapse.
When the DC current increases, the HH neuron with-
out an autapse undergoes a Hopf bifurcation from a
quiescent state to a periodic spiking state. Although
the limit cycle and the oscillation period of the neuron
have been disturbed by the autaptic current, the neuron-
autapse system still undergoes the Hopf bifurcation, gen-
erating a stable periodic orbit. Interestingly, an HH neu-
ron with some special autaptic parameters does not fire
regular action potentials and is attracted to the quies-
cent state. This result reveals the phenomenon of spiking
death that is induced by an autaptic connection, which is
more clearly shown in the membrane potential and autap-
tic current traces in Fig. 9. With an upper-threshold DC
current, the neuron fires action potentials at first. After
the delay time, the subsequent autaptic pulse reaches the
neuron and drives the neuron to a fixed point. Moreover,
the spiking death induced by the autapse is independent
of the initial conditions of the neuron. As shown in the
inset of Fig. 9 (a), the neuron-autapse system can return
(a)
e
v
r
u
c
e
s
n
o
p
s
e
r
e
s
a
h
P
0.5
0
−0.5
0
0.5
phase
(c)
0.05
0
e
v
r
u
c
e
s
n
o
p
s
e
r
e
s
a
h
P
−0.05
−0.1
−0.15
g
aut
g
aut
g
aut
g
aut
g
aut
=0.1
=0.3
=0.5
=0.7
=0.9
−0.2
0
0.5
phase
(b)
0.2
0.1
0
e
v
r
u
c
e
s
n
o
p
s
e
r
e
s
a
h
P
1
−0.1
0
0.5
phase
6
1
1
FIG. 8.
(Color online) The phase-response curve of three
types of autapses with different autaptic conductivity gaut,
(a) electrical autapse, (b) excitatory chemical autapse, and
(c) inhibitory chemical autapse. The external DC stimuli is
set as 26 µA/cm2.
(a)
40
0
−40
−80
)
v
m
(
V
490
320
150
−20
0
100 200 300 400 500
(b)
1
m
0.8
0.6
0.4
0.2
0
−100
time (ms)
(c)
40
0
−40
−80
)
v
m
(
V
(d)
0.18
490
320
0.16
m
150
−20
0.14
0.12
−50
0
V (mv)
50
0
10
20
30
time (ms)
40
50
−58
−56
V (mv)
−54
FIG. 9.
(Color online) (a) Traces of the membrane poten-
tial (dashed lines) and autaptic current (solid lines) [43]. (b)
The phase portrait corresponds to (a). (c) and (d) are the
enlarged plots of (a) and (b), respectively. The different col-
ors indicate different integral initial conditions (The specific
set of parameters can be found in Ref. [43]). The inset in
(a): The membrane potential of HH neuron-autapse system
perturbed by a brief step pause (black solid line).
to the fixed point even when the input is a brief exter-
nal perturbation. The figure also shows that the spiking
death induced by the autapse is stable. The Ref. [72] also
shows a similar phenomenon: an HH neuron transitions
from a limit cycle to a fixed point when the neuron is
perturbed by an excitatory chemical synaptic pulse.
B. Response to a random synaptic pulse-like input
The assumed α form of the postsynaptic current model
that is used in many works is perfect for generating pulse-
like currents that are similar to the synaptic pulses ob-
served experimentally and also enables easy modifica-
tion of the ISI of the input current to investigate the
input-output properties conveniently [60]. Injecting this
type of synaptic pulse-like signal with different ISIs into
the HH neuron, Hideo et al. have reported that the
mean firing frequency decreases as the mean input ISI
increases [73, 74]. When a neuron is injected with a
synaptic-like pulsed current with random ISIs, the neu-
ron displays interesting autapse-induced response behav-
iors [43]. As the delay time increases, the frequency of
the neuron with a sufficient electrical or excitatory chem-
ical autapse shows nearly periodic behaviors. Such peri-
odicity is not changed when the mean ISIs of the input
synaptic current changes. When given the input synaptic
pulses with a large mean value of ISIs, however, the re-
sponse frequency of a neuron with an inhibitory autapse
does not show periodic behavior when the autaptic delay
time increases.
Autaptic activities also influence the detailed response
of the ISI distribution. For a short autaptic delay time,
the output ISIs are distributed in an area [blue bars in
Fig. 10] that is smaller than the input area (red lines in
Fig. 10). When the delay time increases, the region in
which the output ISIs are distributed does not change
much compared with that for the neuron without an au-
tapse (green lines in Fig. 10). For further increases in the
delay time, the distribution of the output ISIs suddenly
decreases almost to a single point (tdelay = 8.5 ms in
Fig. 10). For these autaptic parameters, the responding
spike train is strongly regulated. Then, the size of the
distribution of ISIs increases slowly as the delay time is
increased further.
When the size of the distribution of the output ISIs in-
creases to its maximum (this maximum distribution area
is smaller than the area for the low delay time condition
tdelay < 5.5ms), the distribution suddenly shrinks nearly
to a point again (see tdelay = 20.5 ms in Fig. 10). For the
entire range of delay times, the distribution of the output
ISIs periodically exhibits the above behaviors when the
delay time increases. Thus, the delayed feedback activi-
ties of an autapse can act as a regulator that adjusts the
ISIs of the output spike train. For some specific autaptic
parameters, the resulting spike train of a neuron can be
modified to obtain an almost regular spiking, even for a
very random ISI input.
Without an autapse, the neuron will filter the spikes
with a short ISI and thus show low-pass filtering behav-
ior when the neuron is injected with a random synaptic
pulse-like current [73]. For a neuron with an autapse,
long ISI pulses can be filtered in addition to the short ISI
spikes. Thus, the neuron-autapse system displays com-
plex filtering behaviors, including low-pass filtering and
band-filtering behaviors.
7
t
delay
=1.5
1
0.5
t
delay
=3.5
1
0.5
t
delay
=5.5
0
0
1
20
40
t
delay
=8.5
20
40
t
delay
=13.5
0
0
1
20
40
t
delay
=17.5
0.5
0.5
20
40
t
delay
=20.5
0
0
1
0.5
20
0
0
40
20
ISI (ms)
0
0
1
20
40
t
delay
=24.5
20
40
t
delay
=28.5
0.5
0
0
40
20
40
)
t
i
n
u
.
b
r
a
(
s
m
a
r
g
o
t
s
H
i
1
0.5
0
0
1
0.5
0
0
1
0.5
0
0
FIG. 10. (Color online) Histograms of output (blue bars) ISIs
of a neuron with an electrical autapse [43]. The green lines
gives the output ISI of a neuron without an autapse and the
red lines give the ISI of the random input pulse train.
When the delay time is short, an HH neuron with an
electrical or excitatory chemical autapse acts as a low-
pass filter similar to a neuron without an autapse and
removes the spikes with short ISIs. When the autap-
tic delay time is long enough, the neuron displays band-
pass filtering behavior and removes both the short- and
long-ISI spikes. The cut-off value for the ISIs (either the
minimum or the maximum ISIs that will be retained in
the output spike train) can be changed by changing the
autaptic parameters. For some specific autaptic param-
eters, the neuron can filter most of the input pulse, and
the output spike train can be altered into a nearly regular
spike train, even for an input with highly random ISIs.
More interestingly, a neuron with an inhibitory chemi-
cal autapse can only act as a low-pass filter. Thus, the
inhibitory chemical autapse does not have a significant
effect on the frequency of the output spike train.
It can be conjectured that the filtering properties de-
pend on the intrinsic properties of both the neuron and
the autapse. For electrical or excitatory chemical au-
tapses, the delay time will act as a border or as a cut-off
value for the ISI filtering of synaptic pulses with a long
ISI. The intrinsic filtering property of neurons removes
the short input ISI pulses, and the time delay of the elec-
trical or excitatory chemical autapse removes the long
input ISI pulses. However, an inhibitory autapse delays
the spikes of a neuron and thus filters the spikes that have
short ISI pulses. Thus, the firing frequency of a neuron
with an inhibitory autapse is not higher than that of a
neuron without an autapse.
VI. EFFECT OF NOISE ON THE FIRING
DYNAMICS OF A NEURON
capability of the HH neuron is strongly dependent on the
cell size and autaptic strength.
8
Neuronal noise is random electrical fluctuations within
neuronal networks and affects the patterns of neural ac-
tivity in a determinant way [75 -- 78]. In the presence of
noise, autapses can also substantially affect the response
of a neuron. In this section, we review previous studies
on the interplay of autapses and noise on the firing of an
autaptic neuron.
Considering the subthreshold dynamics of a neuron
with interaction between autaptic-delayed feedback and
noise, Masoller et al.
investigated the firing patterns of
an HH model of a thermoreceptor neuron with an elec-
trical autaptic feedback in the presence of a Gaussian
white noise [79].
In their studies, the neuron displays
only subthreshold oscillations in the absence of feedback
and noise. Their results show that the interaction among
weak autaptic feedback, noise, and the subthreshold in-
trinsic activity is nontrivial. The subthreshold oscillation
amplitude can be enhanced by the autapse in the pres-
ence of external noise, and this enhancement is more pro-
nounced for certain delay values. For negative autaptic
delay feedback, the firing rate can be lower than that of
the noise-free situation, depending on the delay. This is
because noise inhibits feedback-induced spikes by driving
the neuronal oscillations away from the firing threshold.
For positive autaptic delay feedback, there are regions of
delay values where the noise-induced spikes are inhibited
by the feedback; in this case, the autaptic feedback drives
the neuronal oscillations away from the threshold.
Li and his co-workers analyzed the effects of electri-
cal autapses on the spiking dynamics of a stochastic HH
neuron [34], considering the stochastic gating of ion chan-
nels or the so-called intrinsic channel noise is considered.
They found that the delayed feedback manifests itself
in the occurrence of bursting and a rich multimodal in-
terspike interval distribution, exhibiting a delay-induced
reduction in the spontaneous spiking activity at charac-
teristic frequencies. For small numbers of ion channels,
the channel noise is sizable and the excitatory dynam-
ics remain practically unaffected by the delay. However,
smaller noise levels and stronger autaptic intensities in-
duce different synchronization phenomena between the
delay time and the intrinsic time scales. The delay time
and the intrinsic time scales determine the number of
spikes that will be induced and become subsequently
locked during one delay epoch.
Recently, Yilmaz and Ozer showed that the electrical
autaptic delayed feedback either enhances or suppresses
the weak signal detection, depending on the parameters
of autapse and channel noise [80]. When the delay time
is close to integer multiples of the period of the intrinsic
oscillations, the autapse enhances the weak periodic sig-
nal detection for the optimal values of the intrinsic noise
and the autaptic intensity. The system response also ex-
hibits stochastic resonance behavior, depending on the
autaptic intensity. Moreover, the weak signal detection
VII. CONCLUSION AND OUTLOOK
Self-delayed feedback significantly affects nonlinear dy-
namic systems because such self-feedback loops introduce
a new time scale into the dynamics of the system [81, 82].
The system of a neuron with an autapse contains two
time scales: the intrinsic time scale of the neuron and
the time scale of the autapse. However, the situation is
more complex in the nervous system, because there are
many more time scales, including the time scales of the
neurons, the synapse, the environment, and the autaptic
connections. This complexity raises the question of how
the activity time of an autapse influences the activities
of coupled neurons in the nervous system.
Autapses provide self-feedback circuits that are com-
mon in the nervous system. Since the naming of au-
tapses, many experimental studies have revealed that au-
tapses play important roles in brain function [6, 11, 14,
15, 17, 18, 20]. The current theoretical studies on the
topic of autaptic connections also reported the impor-
tance of autapses and the many novel dynamic behaviors
induced by the autapse [32, 34 -- 36, 42, 43]. Autapses
provide a control option that can sufficiently adjust the
firing behaviors of a neuron for any form of input stimu-
lus, regardless of the neuron type.
Autapses offer a new mechanism for switching between
quiescent, periodic and chaotic firing patterns in bursting
neurons [36]. Additionally, the nervous system responds
rapidly to an external stimulus based on the transition
between neuron firing patterns [40, 83]. Thus, the results
of the firing pattern transition induced by an autapse also
indicate that an autapse could act as an efficient tool for
controlling the transition among different relevant neu-
ronal activities in the nervous system.
For an HH neuron with an autapse, the firing fre-
quency and interspike interval distributions of the out-
put spike train show periodic behavior when the delay
time is increased [43]. When specific autaptic parameters
are chosen, the response spike trains are nearly regular,
and the ISI distribution covers a small area, even with a
highly random input. This phenomenon emphassises the
autapse-induced filtering behaviors of the neuron. These
results about the autapse are useful for studying the con-
trol of a nonlinear system.
In neuronal systems, an autapse can take the form of
recurrent excitation, i.e., the discharge of a neuron, pos-
sibly after passing through the axon, can subsequently
induce an excitatory response in the same neuron. This
self-excitation mechanism is important for maintaining
persistent activity, particularly the feeding behavior in
sea slugs. Leonel et al. also reported that such processes
play functional roles in amplifying activity in neuronal
assemblies, causing reverberating activity, inducing some
form of memory, or generating rhythmic patterns, such
as those in central pattern generators [84]. The previous
experimental data also indicate that the brain tissue ex-
presses a novel form of self inhibition, namely autaptic
inhibitory transmission in Fast-Spike (FS) cells slow self
activities in interneurons [85]. Ma et al. also investi-
gated the activities of a two-dimensional neural network
containing autapses with different time delays and found
that the autapses induced many novel collective behav-
iors [86]. The effect of autapse on the synchronization
of neural network have also been conducted in a group
of HH neurons with small world network structure [87].
It was found that the neurons exhibit synchronization
transitions as autaptic delay feedback is varied, and fine
synchronized network activities occur when an optimal
autaptic strength is chosen. Alberto Bacci and his col-
leagues studied the autaptic self-inhibition of basket cells
and the role of the autaptic connections of disinhibition
within cortical circuits and argued that autaptic feed-
back could have a dual function in temporally coordi-
nating parvalbumin basket cells during cortical network
9
activity [88]. Another study also indicated that fast spik-
ing neurons with autapses showed the strongest asyn-
chronous release in brain slices obtained from patients
with intractable epilepsy, and that such discharges may
be involved in generating and regulating network activi-
ties, including epileptic activity [38].
Although there are many studies on the topic of the
autapse, the precise function of autapses and their con-
tribution to information processing are remain unclear.
There are also many open questions both in the exper-
imental and theoretical areas: 1) What is the role of
autapses in coordinating network activities? i.e., FS cell
autapses in adjusting fast network synchrony. 2) What
are the molecular mechanisms underlying autaptic asyn-
chronous release and what is its relevance during physio-
logical and pathological network activities? 3) What are
the functions of autapses in the information propagation
in the neural circuit and neural network? Addressing
these questions will also facilitate our understanding of
the fundamental mechanisms governing several core func-
tions of cortical activities.
[1] Kandel E R, Schwartz J H, Jessell T M, Siegelbaum S
A and Hudspeth A J 2000 Principles of Neural Science
(New York: McGraw-Hill Medical)
[2] Bartos M, Vida I and Jonas P 2007 Nat. Rev. Neurosci.
[23] Cabrera J L and Milton J G 2002 Phys. Rev. Lett. 89
158702
[24] Bonan G B 2008 Science 320 1444
[25] Postlethwaite C M and Silber M 2007 Phys. Rev. E 76
8 45
056214
[3] Bennett M V L and Zukin R S 2004 Neuron 41 495
[4] Connors B W and Long M A 2004 Annu. Rev. Neurosci.
27 393
[5] Van der Loos H and Glaser E M 1972 Brain Res. 48 355
[6] Bekkers J M 1998 Curr. Biol. 8 R52
[7] Yamaguchi K 2008 Autapse Encyclopedia of Neuroscience
(M. D. Binder ed.), N Hirokawa and U Windhorst
(Springer Berlin Heidelberg) pp 229-32
[26] Sethia G C, Kurths J, and Sen A 2007 Phys. Lett. A 364
227
[27] Chen W-K 2005 Circuit Analysis and Feedback Amplifier
Theory (Bocan Raton, FL: CRC Press)
[28] Gaudreault M, Drolet F and Vials J 2012 Phys. Rev. E
85 056214
[29] Ahlborn A and Parlitz U 2004 Phys. Rev. Lett. 93 264101
[30] Balanov A G, Janson N B and Scholl E 2005 Phys. Rev.
[8] Lubke J, Markram H, Frotscher M and Sakmann B 1996
E 71 016222
J. Neurosci. 16 3209
[31] Popovych O V, Hauptmann C and Tass P A 2006 Biol.
[9] Flight M H 2009 Nat. Rev. Neurosci. 10 316
Cybern. 95 69
[10] Branco T and Staras K 2009 Nat. Rev. Neurosci. 10 373
[11] Kimura F, Otsu Y and Tsumoto T 1997 J. Neurophysiol.
[32] Rusin C G, Johnson S E, Kapur J and Hudson J L 2011
Phys. Rev. E 84 066202
77 2805
[33] Prager T, Lerch H P, Schimansky-Geier L and Scholl E
[12] Tam´as G, Buhl E H and Somogyi P 1997 J. Neurosci. 17
2007 J. Phys. A 40 11045
6352
[34] Li Y, Schmid G, Hnggi P and Schimansky-Geier L 2010
[13] Saada R, Miller N, Hurwitz I and Susswein A J 2009
Phys. Rev. E 82 061907
Curr. Biol. 19 479
[35] Hashemi M, Valizadeh A and Azizi Y 2012 Phys. Rev. E
[14] Ikeda K and Bekkers J M 2006 Curr. Biol. 16 R308
[15] Bacci A and Huguenard J R 2006 Neuron 49 119
[16] Bekkers J M 2009 Curr. Biol. 19, R296
[17] Salinas E and Sejnowski T J 2001 Nat. Rev. Neurosci. 2
539
85 021917
[36] Wang H, Ma J, Chen Y and Chen Y 2014 Commun.
Nonlinear Sci. Numer. Simul. 19 3242
[37] Schacter D L, Gilbert D T, Wegner D M and Nock M K
2014 Psychology (New York, NY: Worth Publishers)
[18] Chen Y, Yu L and Qin S M 2008 Phys. Rev. E 78 051909
[19] Chen Y, Zhang H, Wang H, Yu L and Chen Y 2013 PLoS
[38] Jiang M, Zhu J, Liu Y, Yang M, Tian C, Jiang S, Wang Y,
Guo H, Wang K and Shu Y 2012 PLoS Biol. 10 e1001324
ONE 8 e56822
[20] Quiroga R Q and Panzeri S 2009 Nat. Rev. Neurosci. 10
[39] Jun Ma H Q 2015 Int. J. of Mod. Phys. B 29 1450239
[40] Belykh I, de Lange E and Hasler M 2005 Phys. Rev. Lett.
173
94 188101
[21] Davidson E. and Levin M. 2005 Proc. Natl. Acad. Sci.
[41] Buri´c N, Todorovi´c K and Vasovi´c N 2008 Phys. Rev. E
U.S.A. 102 4935
78 036211
[22] Zhang H, Chen Y and Chen Y 2012 PLoS ONE 7 e51840
[42] Wang H, Sun Y, Li Y and Chen Y 2014 J. Theor. Biol.
10
358 25
[43] Wang H, Wang L, Chen Y and Chen Y 2014 Chaos 24
033122
[44] Connelly W M and Lees G 2010 J. Physiol. 588, 2047
[45] Cocatre-Zilgien JH, Delcomyn F 1992 J. Neurosci. Meth.
Lett. A 125, 119
[67] Glass L and Zeng W-Z 1990 Ann.NY Acad.Sci. 591 316
[68] Lewis J E, Glass L, Bachoo M, and Polosa C 1992 J.
Theor. Biol. 159 491
[69] Kunysz A M, Shrier A, and Glass L 1997 Am. J. Physiol.
41 19
273 C331
[46] Mainen Z F and Sejnowski T J 1996 Nature 382 363
[47] Izhikevich E M 2000 Int. J. Bifurcat. Chaos 10 1171
[48] Kepecs A, Wang X-J and Lisman J 2002 J. Neurosci. 22
9053
[49] Gu Hua-Guang,Zhu Zhou,Jia Bing. 2011 Acta Physica
Sinica 60 100505
[70] Smeal R M, Ermentrout G B, and White J A 2010 Trans.
R. Soc. B 365 2407
[71] Krogh-Madsen T, Butera R, Ermentrout G B, and Glass
L, in Phase Response Curves in Neuroscience, edited by
N.W. Schultheiss, Prinz A A., and Butera R J (Springer
New York, 2012), pp. 33
[50] Yu H-T, Wang J, Deng B and Wei X-L 2013 Chin. Phys.
[72] Wang S-J, Xu X-J, Wu Z-X, Huang Z-G, and Wang Y-H
B 22 018701
[51] Wagenaar D A, Pine J and Potter S M 2006 BMC Neu-
rosci. 7 11
[52] Yu H, Wang J, Deng B, Wei X, Wong Y K, Chan W L,
2008 Phys. Rev. E 78 061906
[73] Hasegawa H 2000 Phys. Rev. E 61 718
[74] Borkowski L S Phys. Rev. E 80 051914
[75] White J A, Rubinstein J T and Kay A R 2000 Trends in
Tsang K M, Ziqi Y. et al. 2011 Chaos 21 013127
Neurosci. 23 131
[53] Tang G, Xu K and Jiang L 2011 Phys. Rev. E 84 046207
[54] Wang X-J 1993 Physica D 62 263
[55] Jm G-M 2003 Chaos 13 845
[56] Innocenti G, Morelli A, Genesio R and Torcini A 2007
[76] Jacobson G A, et al. 2005 J. Physiol. 564 145
[77] Wang H, Yu L and Chen Y 2009 Acta Phys. Sin. 58 5070
[78] Destexhe A and Rudolph-Lilith M 2012 Neuronal Noise
(New York: Springer)
Chaos 17 043128
[79] Masoller C, Torrent M C and Garcia-Ojalvo J 2008 Phys.
[57] Innocenti G and Genesio R 2009 Chaos 19 023124
[58] Lee S-G and Kim S 2006 Phys. Rev. E 73 041924
[59] Che Y-Q, Wang J, Si W-J and Fei X-Y 2009 Chaos, Soli-
tons & Fractals 39 454
Rev. E 78 041907
[80] Yilmaz E and Ozer M 2015 Physica A 421 455
[81] Soriano M C, Garcia-Ojalvo J, Mirasso C R and Fischer
I 2013 Rev. Mod. Phys. 85 421
[60] Wang H, Wang L, Yu L and Chen Y 2011 Phys. Rev. E.
[82] Park J-H, Huh S-H, Kim S-H, Seo S-J and Park G-T
83 021915
[61] Fellous J-M, Houweling, A R, Modi, R H, Rao, R P N,
Tiesinga P H E and Sejnowski T J2001 J Neurophysiol.
85 1782
2005 IEEE Trans. Neural. Netw. 16 414
[83] Erichsen R J, Brunnet L G. 2008 Phys. Rev. E 78 061917
[84] G´omez L, Budelli R and Pakdaman K 2001 Phys. Rev.
E 64 061910
[62] Hodgkin A L and Huxley A F 1952 The J. of Physiol.
[85] Pawelzik H, Hughes D I and Thomson A M 2003 J Phys-
117 500
iol 546 701
[63] Bekkers J M 2003 Curr. Biol. 13, R433
[64] Gulledge A T and Stuart G J 2003 Neuron 37, 299
[65] Saada-Madar R, Miller N and Susswein A J 2012 J. Mol.
Histol. 43 431
[86] Qin H, Wu Y, Wang C and Ma J 2015 Commun. Non-
linear. Sci. Numer. Simulat. 23 164
[87] Wu Y, Gong Y and Wang Q 2015 Chaos 25 043113
[88] Deleuze C, Pazienti A and Bacci A 2014 Curr. Opin.
[66] Lewis J, Bachoo M, Glass L, and Polosa C 1987 Phys.
Neurobiol. 26 64
|
1602.08881 | 1 | 1602 | 2016-02-29T09:35:07 | Brain State Control by Closed-Loop Environmental Feedback | [
"q-bio.NC"
] | Brain state regulates sensory processing and motor control for adaptive behavior. Internal mechanisms of brain state control are well studied, but the role of external modulation from the environment is not well understood. Here, we examined the role of closed-loop environmental (CLE) feedback, in comparison to open-loop sensory input, on brain state and behavior in diverse vertebrate systems. In fictively swimming zebrafish, CLE feedback for optomotor stability controlled brain state by reducing coherent neuronal activity. The role of CLE feedback in brain state was also shown in a model of rodent active whisking, where brief interruptions in this feedback enhanced signal-to-noise ratio for detecting touch. Finally, in monkey visual fixation, artificial CLE feedback suppressed stimulus-specific neuronal activity and improved behavioral performance. Our findings show that the environment mediates continuous closed-loop feedback that controls neuronal gain, regulating brain state, and that brain function is an emergent property of brain-environment interactions. | q-bio.NC | q-bio | Brain State Control by Closed-Loop Environmental Feedback
Christopher L. Buckley1,2,*, Satohiro Tajima1, Toru Yanagawa1, Kana Takakura1, Yasuo
Nagasaka1, Naotaka Fujii1, and Taro Toyoizumi1,*
1 RIKEN Brain Science Institute, Japan
2 University of Sussex, UK
Summary
Brain state regulates sensory processing and motor control for adaptive behavior. Internal
mechanisms of brain state control are well studied, but the role of external modulation from
the environment is not well understood. Here, we examined the role of closed-loop
environmental (CLE) feedback, in comparison to open-loop sensory input, on brain state and
behavior in diverse vertebrate systems. In fictively swimming zebrafish, CLE feedback for
optomotor stability controlled brain state by reducing coherent neuronal activity. The role of
CLE feedback in brain state was also shown in a model of rodent active whisking, where
brief interruptions in this feedback enhanced signal-to-noise ratio for detecting touch. Finally,
in monkey visual fixation, artificial CLE feedback suppressed stimulus-specific neuronal
activity and improved behavioral performance. Our findings show that the environment
mediates continuous closed-loop feedback that controls neuronal gain, regulating brain state,
and that brain function is an emergent property of brain-environment interactions.
Introduction
The repertoire of animal behavior involves both passive and active interactions of the brain
with the environment. Passive interactions are driven by the environment and convey signals
for passive sensing and alert. Active interactions, in contrast, are bidirectional, mediating
movement and goal-directed behaviors. During active interactions sensory input from the
environment is shaped by motor actions, and sensory percepts inform future motor
commands, forming a closed-loop between action and perception. This closed-loop
environmental (CLE) feedback is central to the production of motor control behaviors
(Wolpert and Ghahramani, 2000; Kawato, 1999; Scott, 2004) and active sensing behaviors
1
such as saccading and sniffing (Smith 1962; Von Holst, 1954; Crapse and Sommer, 2008).
Yet, how this CLE feedback impacts on brain state, i.e., brain-wide neuronal dynamics and
processing, has received relatively little attention.
Brain state, typically characterized by the degree of synchronously fluctuating neuronal
activity as reported by electroencephalography (EEG), local field potentials (LFP),
electrocortiography (ECoG), membrane potential, and population spiking activity, is strongly
modulated by behavioral context (Buzsáki et al., 2012; Harris and Thiele, 2011). One of the
earliest experimental examples of this phenomenon was the demonstration that alpha
frequency power (7.5–12.5Hz) in EEG recordings is enhanced when subjects close their eyes
or during periods of drowsiness (Berger, 1929). More generally recent experiments in
rodents have demonstrated that synchronous low-frequency fluctuations (a synchronized
state) are typically suppressed (moving to a desynchronized state) when animals transition
from quiet attentive behaviors, where animals are largely passive, to active behaviors such
as running or whisking (Crochet and Petersen, 2006; Niel and Stryker, 2010; Schneider et
al., 2014; Otazu et al., 2009; Zagha et al., 2013).
Brain states not only alter spontaneous brain dynamics, but also sensory representations,
presumably allowing brain function to adapt to behavioral context. For example, it has been
shown that the onset of running amplifies cortical responses to visual and auditory stimuli
(Niell and Stryker, 2010, Fu et al., 2014; Schneider et al., 2014; McGinley and McCormick,
2014). In addition, cortical responses to punctuate stimuli are larger during synchronized
states (Poulet and Petersen, 2008; Fanselow and Nicolelis, 1999; Castro-Alamancos, 2004,
Otazu et al., 2009). These brain state changes are typically thought to happen across the
whole brain (Berger, 1929, Steriade, 2001; Fu et al., 2014; McGinley and McCormick, 2014)
and are cross-modal (Otazu et al., 2009).
2
Several internal mechanisms have been implicated in these changes including
neuromodulatory pathways (Goard and Dan, 2009; Polack et al., 2012; Pinto et al., 2013; Fu
et al., 2014), thalamo-cortical projections (Poulet et al., 2012), and intracortical feedback
(Zagha et al., 2014; Schneider et al., 2014). Changes in the external sensory environment
also play an important role. Under passive sensing conditions sensory stimulation can shift
neural activity in the visual cortex between asynchronous and synchronous regimes (Tan et
al., 2014) and more generally sensory stimuli quench neural variability (Churchland et al.,
2010). More relevant to the present study, brain state transitions often coincide with the
onset of active behaviors, which are characterized by the presence of reafferent input
(sensory input resulting from one's own action, mediated by the environment). It is well
known that reafferent input strongly influences neural activity (Curtis and Kleinfeld, 2009;
Urbain et al., 2015; Brooks et al., 2015) and modulate sensory responses (Diamond et al.,
2008). However, the role played by reafferent input in regulating brain state is not well
understood. In this paper, we combine theoretical and experimental approaches to dissect
the effect CLE feedback mediated by reafferent input, as oppose to sensory input per se, has
on modulating brain state. In doing so we elucidate a novel function that CLE feedback plays
in sensing and behavior.
Results
During active behavior, sensory input to the brain is directly shaped by motor actions, and
reciprocally, motor actions are informed by prior sensations forming a closed-loop between
the brain and the environment. Under such conditions, it is possible to distinguish two
sources of sensory input (Von Holst, 1954), see Figure 1A. First, exafferent input describes
sensory signals that originate in the environment, but which are completely external to the
brain (Von Holst, 1954). Second, reafferent input describes sensory signals, which, while
mediated by the environment, are a direct consequence of an animal's own actions and
constitute a sensory feedback signal to the brain (Von Holst, 1954). While it is known that
3
sensory input in general can influence brain state variability (Churchland et al., 2010), it is not
clear whether exafferent and reafferent input play distinct roles in determining these
changes, or whether only total synaptic input is of importance. To examine the effect of each
type of input on brain state, we distinguish three types of brain-environment interaction. First,
an open-loop condition, where sensory input is not coupled to motor output and thus the
brain receives sensory input independent of its own activity (Figure 1B). In this condition,
reafferent input is absent and the brain receives only exafferent input. Second, a closed-loop
condition, where the brain interacts with the environment in a closed-loop thus coupling
motor action and sensory perception (Figure 1C). In this condition, the brain receives
reafferent input in addition to any possible exafferent input. Lastly, we define a replay
condition, where again the brain operates in an open-loop but where reafferent input
experienced during a prior closed-loop condition is recorded and replayed as exafferent input
(Figure 1D). In this condition, the total sensory input is identical to that in the closed-loop
condition, but reafferent input is absent. Specifically, the sensory input during the replay
condition is not a consequence of motor actions within that condition. Therefore, this
condition serves as a strong control allowing us to distinguish precisely the contribution made
by reafferent input rather than that of total sensory input, to the brain state.
To understand the possible effects that CLE feedback could have on brain activity we
introduce a simple idealized model. We start by describing collective neuronal activity, e.g.,
an EEG signal, of an open-loop brain (denoted as 𝐵𝑜(𝑡)), in terms of a first-order differential
equation,
𝑑𝐵𝑜(𝑡)
𝑑𝑡
= −
𝐵𝑜(𝑡)
𝜏
+ 𝜉𝑜(𝑡),
where 𝜉𝑜 is white noise of instantaneous variance 𝜎2 generated inside the brain, 𝑡 is time,
and 𝜏 is the time constant of the system. The autocorrelation peak (instantaneous variance)
of variable 𝐵𝑜 is given by Peak𝑜 = 𝜎2𝜏/2 (see, Figure 1B for open-loop time trace). To
4
describe neuronal activity in the closed-loop condition (denoted as 𝐵𝑐(𝑡)), we approximate
reafferent input as a linear function of the brain variable, i.e.,
𝑑𝐵𝑐(𝑡)
𝑑𝑡
= −
𝐵𝑐(𝑡)
𝜏
+ 𝑤𝐵𝑐(𝑡) + 𝜉𝑐(𝑡),
where 𝑤 scales the strength of reafferent input and 𝜉𝑐 is again white noise of instantaneous
variance 𝜎2. (Later we also explore a more realistic reafferent input that involves filtering and
delay). In this condition, the continuous cycles of reafferent input constitutes a CLE feedback
signal to the brain. The presence of this CLE feedback changes the effective time constant to
𝜏eff = 𝜏/(1 − 𝑤𝜏) or, equivalently, it the changes the system's gain. We can quantify this
change by using an expression for the autocorrelation peak of the system, as follows,
Peak𝑐 = Peak𝑜/(1 − 𝑤𝜏). In particular, we find that CLE feedback suppresses fluctuations
within the brain if this feedback is negative (𝑤 < 0; see Figure 1C for closed-loop time trace).
In contrast, we can describe neuronal activity in the replay condition (denoted as 𝐵𝑟(𝑡)) as:
𝑑𝐵𝑟(𝑡)
𝑑𝑡
= −
𝐵𝑟(𝑡)
𝜏
+ 𝑤𝐵𝑐(𝑡) + 𝜉𝑟(𝑡),
where 𝜉𝑟 is again white noise of instantaneous variance 𝜎2. Here, the brain receives the
same total sensory input as in the closed-loop condition, 𝑤𝐵𝑐(𝑡), but as exafferent input, i.e., it
depends on the dynamics of 𝐵𝑐(𝑡), and not on 𝐵𝑟(𝑡). Again we can calculate the
autocorrelation peak of the brain variable in this condition as Peak𝑟 = Peak𝑐 + Peak𝑜 ∗
2𝑤𝜏/(𝑤𝜏 − 2). Notably, even though the brain receives exactly the same total sensory input
in the replay and closed-loop conditions the amplitudes of fluctuations are not the same. In
particular, if CLE feedback is negative, we obtain peak𝑐 < peak𝑜 < peak𝑟 (see, Figure 1D for
replay time trace). In summary, this simple model suggests that CLE feedback constituted by
continuous reafferent input could have a profound effect on neuronal fluctuations and thus
5
could be implicated in the modulation of brain state. In the next section, we examine this
hypothesis using experimental data.
Negative CLE feedback suppresses synchronous neural fluctuations
We tested our predictions by analysing two-photon imaging data recorded from larval zebrafish
behaving in a virtual flow simulator (Ahrens et al., 2012), a setup that allowed us to quantify the
differences between the closed-loop and replay conditions directly (Figure 1). In this setup, bouts of
activity recorded from motor nerves along the spine of a paralyzed fish were translated into a
backward drift of a visual grating, simulating forward swimming. This allows the fish to
maintain its perceived horizontal position by swimming against water flow (Ahrens et al.,
2012), (Figure 2A). In a transgenic fish expressing the calcium indicator GCaMP2 brain-wide
calcium activity was monitored using a two-photon microscope to scan single planes in the
brain. In a closed-loop condition, the fish actively maintain their position in the virtual
environment. In a replay condition, the same fish receives a replay of the closed-loop visual
stimulus without real-time visual feedback (Figure 2A), (Ahrens et al., 2012). We found that
neural dynamics between the two conditions were significantly different despite the fact that
both the closed-loop and replay conditions involve identical sensory input (Figure 2B).
Notably, the only information the fish received about oncoming flow was visual, i.e., there
was no proprioceptive input as the fish was paralyzed (Ahrens et al., 2012). Individual
neurons were heterogeneous across the whole brain, but on average low frequency (0.01 -
0.15 Hz) fluctuations were suppressed (p = 0.046, sign test) and neurons were decorrelated
(p = 0.005, sign test) under the closed-loop condition compared to the replay condition
(Figure 2B). Changes in the geometric mean of low frequency fluctuations and correlation
were highly correlated in each pair of cells (r=0.69, p<10-10, Spearman's rank correlation),
suggesting a common cause (see Figure 2C). The decorrelation effect was not an artifact of
measurement noise, which may dominate at high frequency because the result was robust to
the removal of low-level calcium activity by thresholding (see Supplementary Figure S1A).
This change aligns with brain state transitions at the onset of active behavior that has been
reported in other species (see Poulet et al., 2012 or Harris and Thiele, 2011 for a review). This
6
suppression of low frequency power and correlation could be caused by an efference copy
signal (signal encoding intended motor action) that suppresses synchronous neuronal
fluctuations (Zagha et al., 2014; Schneider et al., 2014). However, while motor activity levels
were higher in the closed-loop condition than the replay condition, this did not explain the
difference in neuronal fluctuations. Specifically, the motor activity level positively correlated
with the changes in low frequency power (Supplementary Figure S1B, top panel; r = 0.18, p
< 10-2, Spearman's rank correlation) and was not significantly correlated with changes in
pairwise correlations between cells (Supplementary Figure S1B, bottom panel r = 0.03, p >
0.5, Spearman's rank correlation).
The only difference between the closed-loop and replay conditions was the presence of CLE
feedback; thus it can be assumed that this must play a causal role in these changes. To
investigate this further, we quantified CLE feedback by first estimating how fish behavior
affects neural activity, and conversely, how neural activity drives their behavior. To this end
we calculated linear filters that characterize dynamic interactions between single neuron
activity (the brain variable: B) and swimming power as quantified by the activity of motor
neurons (a putative environmental variable: E) (see Methods). Figure 3A (inset)
schematically shows their interaction under each condition. In the closed-loop condition B
and E interact mutually; this has the potential to confound calculation of independent causal
filters. Consequently, we calculated filters based on the replay condition (however, see below
for filters computed based on the closed-loop condition). In this condition, the brain variable
B' is driven by the recorded CLE variable E, which in turn generates its own environmental
variable E', without involving CLE feedback. Specifically, for each observed neuron, we
computed a linear filter that predicts B' from E and a linear filter that predicts E' from B'.
While these filters are neuron-dependent, the average normalized filters showed clear net
effects (Figure 3A). On average, across cells, the interaction from the brain to the
environment, B'→E' (Figure 3A, purple solid line), was strongly positive, indicating that an
7
increase in neural activity positively drove motor behavior. However, we found that
interaction from the environment to the brain, E→B' (Figure 3A, pink solid line), was net
negative, indicating that an increase in motor behavior on average suppressed neural
activity. It is reasonable to assume that the same filters (circuits) operate in the closed-loop
condition. If this is the case, CLE feedback for each cell in the closed-loop condition is
characterized by the convolution of the two filters, i.e., by the convolution of the B'→E' and
E→B' interactions computed for each cell. The convolution of these filters was on average
also negative (Figure 3A, cyan dashed line), peaking at about 1 s. This suggests, that on
average, increases in neural activity self suppressed after 1 s due to negative CLE feedback.
Notably, the negative CLE feedback interactions were also confirmed by behavioral analysis,
in which the the E→E' filter was estimated directly (see Supplementary Figure S1C), which
supported the robustness of our results.
Next, we considered what the consequence of this negative CLE feedback on each cell's
activity would be. Our conceptual model, Figure 1, suggests that fluctuations in neural
activity should decrease if a cell receives negative CLE feedback. In this case the estimated
CLE feedback in each cell should predict to what extent fluctuations in low frequency power
should change in the closed-loop condition relative to the replay condition (see Methods). As
theoretically expected, Figure 3B demonstrates that the predicted degree to which a neuron's
activity was suppressed during the closed-loop condition relative to the replay condition was
highly correlated with what was actually observed (r = 0.39, p < 10-8, Spearman's rank
correlation). Note that neural activity in the closed-loop condition was not used to fit each
filter. Although the majority of cells across the fish brain were suppressed by the behavior,
the top 10 percentile of cells that were both strongly suppressed, and strongly involved in,
negative CLE feedback were clustered in the cerebellum (Figure 3C), a brain area implicated
in sensory-motor planning and coordination (Scott, 2004). This supports the idea that the
8
cerebellum plays a central role mediating negative closed-loop interaction between the brain
and the environment by converting sensation into action in fish during optic flow stabilization.
We then investigated whether the dynamic relation between neuronal activity and motor
activity were also affected by CLE feedback. To do this, we quantified the brain/environment
dynamics for each neuron by naively evaluating the E→B filter in the closed-loop condition,
and compared this to the corresponding E→B' filter in the replay condition (Figure 3D). These
two filters were generally distinct in the observed neurons, but were particularly so for those
cells that were strongly stabilized by CLE feedback (Figure 3D; Inset). In these neurons,
while the E→B' filter from the replay condition was approximately causal, as expected, the
closed-loop filter E→B had an additional acausal component. We suggest that this latter filter
reflects the closed-loop interaction between the brain B and the environment E. To test if the
closed-loop interaction could explain this discrepancy, we theoretically predicted the E→B
filter in the closed-loop condition based on data from the replay condition, i.e., using both the
E→B' and B'→E' filters (see Methods). We found that, on average, this closed-loop effect
could account for the difference between E→B and E→B' as calculated for the closed-loop
stabilized cells (Figure 3D; Inset). To quantify this for each cell, we calculated the mean
square error between the predicted and E→B' filters and between the E→B and E→B' filter.
The ratio of these two error terms then quantifies the fraction of the mean square error that
was explained by the prediction. This ratio was significantly less than one (median = 0.80, p
< 10-11, sign test), indicating that the prediction was more accurate when the closed-loop
effect was included, and the ratio was positively correlated (r=0.25, p<10-13, Spearman's rank
correlation) with the degree to which individual cells were stabilized in the closed-loop
9
condition (Figure 3D). Altogether, these results indicated that neuronal dynamics, as well as
its relation to sensory stimulus and behavior, not only depend on brain circuits, but are also
dynamically shaped by the mutual interaction between the brain and the environment.
CLE feedback explains the difference between active and passive sensing
Behaviors not only induce brain state transitions, but also differentiate active sensing from
passive sensing. To investigate whether these differences could be accounted for by CLE
feedback, we examined a well-studied model of a brain state transition in the rodent whisker
system. Specifically, we consider CLE feedback to the rodent brain mediated by whisking
vibrissa. Although more precisely we should refer to this as closed-loop body/environmental
feedback, we retain CLE acronym by generalizing the notion of environment to refer to all
processes outside of the brain highlighting the generality of our theory. Whisking not only
changes brain state (Crochet and Petersen, 2006; Poulet and Petersen, 2008), but also
reduces the sensitivity of the cortex to passive whisker deflections (Crochet and Petersen,
2006; Crochet et al., 2011). However, interestingly, robust responses are maintained for
active contact events when the whisker collides with an object placed in the whisk field
(Crochet and Petersen, 2006; Crochet et al., 2011). Furthermore, in an active touch condition
where the whisker repeatedly collides with an object both intra-neuronal correlations and low
frequency power of membrane potential recovers close to their passive state values (Crochet
et al., 2011). We theoretically investigated whether these phenomena could be explained by
the hypothesis that vibrissa dynamics mediates negative CLE feedback to the brain,
influencing cortical dynamics in a way analogous to fish swimming behavior. Notably, while
multiple different mechanisms are involved in brain state transition (Goard and Dan, 2009;
Polack et al., 2012; Pinto et al., 2013; Fu et al., 2014; Poulet et al., 2012; Zagha et al., 2014;
Schneider et al., 2014), for which sensory input is not always necessary (Poulet and Petersen,
2008; Poulet et al., 2012), this does not exclude the role of CLE feedback under physiological
conditions. Indeed, although brain state transitions can happen at whisking onset even after
the sensory nerve is cut (Poulet and Petersen, 2008), our analyses revealed that the latency
of brain state transition, measured by whisking-related reduction in low frequency power of
10
cortical LFP or in thalamic spiking rate, was delayed under this condition as compared to the
control condition (Supplementary Section S2). This supports a physiological role for sensory
input in the brain state transition.
Here, we construct a simple model of the vibrissa system to investigate the possible role of
CLE interaction. The model comprises excitatory and inhibitory cortical populations
dynamically interacting with a single vibrissa (Figure 4A; see Methods for details). In our
model, synchronous membrane potential fluctuations arise endogenously from interplay
between the build up of excitatory cortical activity by recurrent connections and their eventual
suppression by adaptation after ca. 1 s in each cell (see Methods). To model the brain/body
interaction, we assumed that cortical excitatory neurons drive whisker motor behavior and
both excitatory and inhibitory neurons receive the sensory input that reflects whisker angle.
Based on our theory, Figure 1, we hypothesized that synchronous neural fluctuations could
be reduced during whisking by negative CLE feedback. We modeled this by assuming that
cortical excitatory neurons are activated by whisker retraction but this activation drives
whisker protraction (see Figure 4A). (Note the specific biological implementation of the
negative CLE feedback is not central to our claims, and there are several other plausible
schemes, see Discussion and Supplementary Section 6.) We also model a central pattern
generator (CPG, likely located in the brain stem) (Hill et. al, 2011) that, once turned on,
rapidly cycles the position of the vibrissa back and forth at around 10Hz.
Sensory input during contact events involves both contact-related and whisker angle-related
signals (Diamond et al., 2008). To capture this sensory input we modeled a single whisker as
two stiff mass-less sections connected by hinges at the center and the base, which are
constrained by muscles (simple torsion springs). Whisking is implemented by driving the
equilibrium position of the base spring. The center spring has an equilibrium value of zero
angular displacement and thus tends to align both sections (see Supplemental Information
Section 3). We start with a simple non-bending stiff whisker but later vary the stiffness
11
parameter (see Methods). A horizontal solid wall is placed above the whisker and, as the
whisker collides with the wall, the whisker tip stops, see Figure 4A, interrupting CLE
feedback about whisker angle. Whisker contact not only involves a modification of CLE
feedback but also invokes exafferent input, or contact-detection signal, that results from the
stereotypical response of pressure sensitive cells in the trigeminal ganglion (Szwed et al.,
2003). We model this as a brief square wave pulses (ca. 25 ms) triggered by each contact
event (see Methods).
Our key hypothesis, i.e., that whisker position mediates negative CLE feedback to cortical
neurons, allowed us to reproduce rodent brain state findings that are observed under three
behavioral conditions (the quiet condition, the whisking condition, and the active touch
condition; Figure 4B [Crochet et al., 2011]). Specifically, the synchronous low frequency
membrane potential fluctuations generated during the quiet condition (open-loop: the whisker
position is fixed at 0) were significantly suppressed during the whisking condition (closed-
loop: the whisker is driven by the CPG and cortex) (Figure 4C). In effect, the negative CLE
feedback reduced the gain of the cortical system, see conceptual model Figure 1, and
replaced prominent (ca. 1 Hz) synchronous fluctuations of the membrane potential with fast
(ca. 10 Hz), but weak, fluctuations locked into the whisking cycle. As a result, the average
inter-neuronal correlation of membrane potential for pairs of neurons was also suppressed
(Figure 4D).
During the active touch condition, occasional whisker touch events stopped the whisker tip
on the object's surface (Figure 4B; touch events marked by red bars). Both the low frequency
power and cross-correlation of membrane potential fluctuations were partially recovered
during the active touch condition in agreement with experimental results (Crochet et al., 2011)
(Figure 4C and D, red lines). We hypothesized that this recovery was a consequence of
intermittent interruption of negative CLE feedback during whisker contact events. To verify
this, we examined whether this result depended on the stiffness of the whisker. We also
12
hypothesized that, if the whisker is very flexible, then the base angle would change
continuously, despite contact of the tip, fully preserving CLE feedback. In contrast, a very stiff
whisker would come to complete rest on the wall during each contact maximally interrupting
CLE feedback (see also Supplementary Figure S3B). We found that the recovery of 1-5 Hz
power was indeed predominantly caused by the interruption of CLE feedback and not touch-
evoked exafferent input (see Supplementary Figure S3C) demonstrating that these transient
interruptions of CLE interaction can rapidly switch the active brain state to a passive brain
state.
This model also accounted for observed brain state dependent changes in sensory
processing (Crochet et al., 2011) without assuming additional mechanisms (Lee et al., 2008;
Nguyen and Kleinfeld, 2005). In agreement with experimental results we found that passive
exafferent input evoked large sensory responses in the quiet condition but markedly smaller
responses during the whisking condition (Figure 4E) (Crochet et al., 2011). Again, this was
because negative CLE feedback decreases the gain of the cortical circuit (Supplemental
Section 4). However, in agreement with experimental data, cortical neurons exhibited more
reliable responses to active touch events (Figure 4B and E). This was because the negative
CLE feedback, that suppressed neural fluctuations during whisking, was transiently removed
during active touch events allowing endogenous brain dynamics to amplify the cortical
response to exafferent input (Figure 4E and F). Consequently, active touch events combined
large sensory evoked responses (signal) and low background fluctuations (noise).
To quantify this effect, we computed a discriminability index (see Methods) that measures
the separation between the distributions of membrane potentials in the presence or absence
of sensory events. The value of the index was similar for perturbations in both quiet and
whisking conditions, i.e., although the deflection-evoked response (signal) was greater in the
quiet condition, so were background fluctuations of membrane potential (noise) (see Figure
4E and F). In contrast, the discriminability index was greater for active touch events (Figure
13
4E and F). Hence, the model suggests that cortical neurons are selectively sensitive to
interruption of the animal's own active sensing. All these results are intuitively reproduced in
a simplified model of CLE interaction, demonstrating the robustness of our findings
(Supplemental Section 4).
Neurofeedback modulates brain dynamics and behavior
We further tested a core prediction of our theory by examining the impact of an artificially
constructed CLE feedback on brain dynamics. Specifically, we investigated whether we could
use an ECoG-based fast neurofeedback technique to modulate brain dynamics and
subsequently behavioral performance in primates. This setup allowed us to observe the
impact of CLE interaction without involving efference copy signal (which can induce brain
state transition; Zagha et al., 2014; Schneider et al., 2014) and to investigate the possible
clinical relevance of CLE feedback. We implemented real-time neurofeedback between the
visual areas (occipitotemporal and parietal cortices) of a fixating macaque monkey (equipped
with a 128-channel ECoG on the cortical surface, Supplementary Figure S5A) and a visual
stimulus (equivalent to the environment in the fish study or vibrissa feedback), while the
monkey was fixating (Figure 5A). During the presentation of a visual grating stimulus (Figure
5A), we decoded orientation (vertical or horizontal grating), in real-time, from ECoG activity
with a support-vector machine. During neurofeedback sessions, a computer monitored the
output of the classifier (the decision value), a nonlinear projection of the ECoG activity, and
then modified the stimulus presented on the screen in real-time. A large and positive (or
negative) decision value indicated that ECoG activity was likely driven by the horizontal (or
vertical) grating stimulus, respectively. When the decision value exceeded a positive
threshold, the computer showed a vertical grating for the subsequent 100 ms period; vice
versa, exceeding a negative threshold triggered the presentation of a horizontal grating (see
Methods). If the decision value fell between these two thresholds, a grey screen was
presented. This protocol effectively approximates negative feedback (the closed-loop
condition) on the decision value dynamics of the animal (however see Discussion). To
14
distinguish between the influence of neurofeedback from the visual stimuli alone, we used a
control condition in which the monkey was presented with a recording of the visual stimuli
that emerged during the closed-loop condition (a replay condition). Thus, as in our
conceptual mode (Figure 1A), and analogous to the fish experiment (see Figure 2A),
corresponding closed-loop and replay conditions share identical visual stimuli and only
differed by the presence of neurofeedback. Again, this allowed us to examine the exact effect
of CLE feedback, rather than sensory input on brain activity and behavior.
We observed that neurofeedback immediately altered the dynamics of the decision value and
thus brain dynamics, without requiring any prior training or habituation of animals. The power
spectrum of the decision value below 5 Hz significantly differed under the two conditions
(Figure 5B). While the 0-1 and 3-4 Hz (summarized in Figure 5C) frequency ranges were
suppressed, the 1-3 Hz frequency range was enhanced during the closed-loop condition
relative to the replay condition. This oscillatory modulation of the power spectrum could have
arisen from delay (ca. 100 ms) in negative CLE feedback, but it would also depend on
various factors, including the kinetics of the decision value, and early visual stream
processing. However, the exact mechanisms underlying the changes in the power spectrum
are not the focus of this study. We found that not only were the dynamics of the decision
value altered by neurofeedback, but also that there was significant improvement in the
monkey's fixation performance (Figure 5D). Specifically, the distance of the eye position from
the fixation point was quantified and compared under the two conditions. The fixation
performance at 1600-2100 ms after the appearance of the fixation point (about 300-900 ms
after typical grating onset; Supplementary Figure S5C) in the closed-loop condition was
enhanced (Figure 5E). To establish a quantitative relationship between changes in neural
dynamics and behavior, we analyzed the trial-to-trial correlation between the power spectrum
of the decision value and fixation performance. The reduction of fluctuation in the decision
value at specific frequencies correlated with improved fixation performance (Figure 5F). In
particular, the neurofeedback-induced reduction of decision value fluctuations in the 3-4 Hz
15
frequency range (estimated from data before the behavioral improvement; see
Supplementary Figure S5D) showed a significant correlation with improved fixation
performance (Figure 5G). Taken together, these results indicate that in primates artificial
neurofeedback can change brain dynamics related to sensing and task-related behavior,
which was consistent with our fish and rodent studies. Although the causal chain of events
underlying these changes may also involve internal brain mechanisms for state transitions
(e.g., neuromodulators; see Discussion), we assert that neurofeedback constitutes the
primary cause of these changes as it is the only difference between each condition.
Discussion
Brain state transitions triggered by the onset of active interactions between the brain and environment,
represent a major neuronal mechanism shaping sensing and behavior. In this study, we use both theory
and experiment to support the idea that negative CLE feedback inhibits network gain, which in turn,
suppresses synchronous neuronal fluctuations and sharpens sensory responses. We generalize and
support the theoretical framework in three diverse animal model systems summarized in Figure 6. In
each system, we show that negative CLE feedback regulated real-time brain state and animal behavior.
Specifically, CLE feedback quantitatively predicted cell-specific suppression of neural oscillations in
zebrafish, enhanced signal-to-noise ratio for active sensing in rodent, and enhanced task performance
in primate vision.
The importance of using naturalistic sensory stimuli to study and manipulate brain state
dynamics is widely demonstrated (Felsen and Dan, 2005). However, an important prediction
of our theory (Figure 1, Supplemental Information Section 7), supported by our experimental
findings (Figures 2, 3, and 5), is that brain dynamics during active sensing cannot be fully
recapitulated or re-encoded, even if the same sensory input is precisely recorded and
replayed back into a passive brain. These results provide evidence that brain state
neurometrics and behavioral psychometrics during active behaviors can only be accurately
understood by a quantitative account of ongoing brain-environment interactions (O'Regan
16
and Noë, 2001).
Neuronal gain control by CLE feedback
The formal component of our theory, i.e., that CLE feedback can modulate a system's gain, is well
documented in dynamical systems theory and control theory (e.g., Aströmand Murray, 2010). This
gain control occurs even though the instantaneous effect of the pathways mediating feedback is purely
additive (c.f. Figure 1) because the effect of repeated cycles of feedback accumulates over time. For
example, input 𝐼 to a linear dynamical system with feedback strength 𝑏 makes a first direct
contribution 𝐼 to its response, but then also makes subsequent contributions as this initial response
cycles around a feedback loop-a contribution of 𝑏𝐼 after one cycle, 𝑏2𝐼 after two cycles and so on.
The cumulative sum of these contributions I + 𝑏𝐼 + 𝑏2𝐼 +. . . = 𝐼/(1 − 𝑏) is equivalent to divisively
scaling the input magnitude by a factor that depends on the feedback strength, i.e., effectively
changing the system's gain. Thus, a constitutively active closed-loop feedback that mediates multiple
action-perception cycles is essential for the form of gain control we propose.
This means that discrete and intermittent involvement of reafferent input does not imply gain
modulation. For example, the classical reafference principle explains neuronal responses by a one-time
detection of the mismatch between an efference copy (predicted) and reafferent (actual) (Von Holst,
1954). However, this situation is likely an inaccurate idealization to describe the closed-loop systems
studied here. For example, in the zebrafish system, swim bouts typically occur every 700 ms and this
interval closely overlapped with the peak of the estimated environmental feedback interaction (Figure
3A and Supplementary Figure S1C). Hence, the neural responses in the fish experiment suggest a
more dynamic system, where neural activity evoked by many cycles of action and sensation are
continuously and mutually interacting.
Neuronal mechanisms of negative CLE feedback
The presence of continuous negative CLE feedback during active behavior is fundamental for our
theory. Although, as a higher order theory, it is agnostic to the detail of the neural implementation, we
17
discuss below the implications for each system. In zebrafish, the presence of negative CLE feedback
during fish swimming behavior would seem a priori necessary for optic-flow stabilization behavior
because the fish must act in opposition to perceived optic flow in order to minimize horizontal
displacement (Wolf et al. 1992, Fry et al., 2009). Interestingly, neurons that received strong negative
CLE feedback and were substantially stabilized were located in the cerebellum (Figure 3C) consistent
with a theoretical viewpoint that the cerebellum mediates this optomotor response by converting
sensation into action; e.g., by completing the action-perception cycle (Ito 1972, Lisberger et al. 1987,
Kawato 1999).
Our rodent whisker model explores the overall effects of negative CLE feedback mediated by the
cortical-whisker circuit during active whisking. As in the fish study, the theoretical assumption of
negative CLE feedback is consistent with the idea that the barrel cortex is involved in control.
Specifically, the barrel cortex comprises a nested set of servo control loops that regulate various
aspects of whisker dynamics (Ahissar and Kleinfeld, 2003). At the level of the whole vibrissa system,
multiple parallel and nested feedback loops, both positive and negative (Ahissar and Kleinfeld, 2003)
most likely exist. While our model is abstract in terms of the known complexities of vibrissa system
anatomy and exact concordance of the model with vibrissa system anatomy is beyond the scope of this
paper, we provide several possible schemes for experimentalists to examine in Supplemental
Information Section 6.
The artificial CLE feedback in the primate experiment implements negative feedback, although this is
less obvious than in our other studies and simple models. A system is defined to be under negative
feedback if this feedback causes perturbations to the system's state to decay faster in time and
suppresses fluctuations due to noise. Hence, neurofeedback in the monkey experiment is a hybrid of
positive and negative feedback-the feedback is negative in the 0-1 Hz and 3-4 Hz frequency bands
and positive in the 1-3 Hz frequency band, because the feedback reduces or increases fluctuations of
the decision value in these frequency bands, respectively (Figure 5B). In this experiment, the system's
state was quantified by the decoder's decision value, with positive and negative values indicating
18
vertical-like and horizontal-like brain activity patterns, respectively. Due to experimental constraints
of delays in neurofeedback processes and unidentified dynamic properties of the brain, imposing
purely negative neurofeedback in this experiment was not possible. However, negative CLE feedback
in the 3-4 Hz frequency band is strongly correlated with enhanced fixation performance (Figure 5F),
supporting its behavioral significance.
Comparison to internal mechanisms of brain state transitions
The proposed mechanism of brain state transitions by CLE feedback is qualitatively different from the
many other mechanisms governing brain state that operate inside the brain (Goard and Dan, 2009;
Polack et al. 2012; Pinto et al., 2013; Poulet et al., 2012; Zagha et al., 2014). Specifically, the CLE
feedback mechanism involves the dynamic coordination between brain activity and
body/environmental dynamics (c.f. Figure 3D), a continuous reciprocal interaction that is critical for
many forms of goal-directed behavior. In this regard, our theory directly links a mechanism of brain
state transitions to behavior and is closed from an explanatory view, i.e., without assumptions about
upstream causes.
In our zebrafish study, an alternative interpretation of our findings is that correlated neural fluctuations
are suppressed directly by an efference copy signal (Zagha et al., 2014; Schneider et al. 2014) due to
greater motor commands in the closed-loop condition (Supplementary Figure S2). However, in
opposition to this view, we found that motor activity correlated positively with fluctuations in each
cell, and were not associated with pairwise neuronal correlations (Supplementary Figure S2). Hence,
we cannot account for the results by any known functions of efference copy signals. Moreover, the
primate neurofeedback allowed us to examine the impact of CLE feedback in the absence of the overt
involvement of the motor system, i.e., feedback was only dependent on readout from visual areas. This
allows us to conclude that CLE feedback alone, in the absence of noticeable motor activity, is
sufficient to modulate brain activity and behavior.
19
Importantly, brain state control by CLE feedback is not mutually exclusive with other mechanisms,
such as thalamo-cortical input (Poulet et al., 2012) or neuromodulation (Goard and Dan, 2009; Polack
et al. 2012; Pinto et al., 2014; Fu et al. 2014) because these may also be involved in mediating the
action-perception cycle. Furthermore, brain state transitions also occur in the absence of CLE
feedback, such as open-loop behaviors (e.g., the onset of running that does not change the visual
screen or grooming) (Niell and Stryker, 2010, Fee et al. 1997), during sleep (Vyazovskiy et. al 2011;
Steriade, 2005), or under dissection of the sensory nerve (Poulet and Petersen, 2008; Poulet et al.,
2012). Mechanisms underlying brain state transitions are likely to be highly redundant even in the
absence of essential mechanisms, such as thalamo-cortical input (Poulet et al., 2012) or corollary
discharge (Schneider et al. 2014), albeit involving further delay (see also Supplementary Section 2).
Such functional redundancy may help to maintain the stability of brain state (Fu et al. 2014; McGinley
et. al, 2015; Otazu et al. 2009; although with exceptions, see Vyazovskiy et. al 2011). Furthermore, the
relative importance of internal and external mechanisms might adaptively change in an experience-
dependent manner (Nachestedt et. al 2014).
Reafferent feedback in active sensing and revision of the reafference principle
In the whisking model, we proposed that the regulation of cortical gain by CLE feedback could
explain enhanced active touch. Specifically, negative CLE feedback during whisking reproduces
suppressed fluctuations and reduces responses to passive whisker stimulation (see Figure 4F).
However, robust neuronal response to active touch events could be explained by the interruption of
CLE feedback when the whisker is driven into an external object. Such events interrupt CLE feedback,
transiently releasing the cortex from a low gain state and enhancing sensory responses to salient
sensory stimuli (Figure 4E, Supplemental Information Section 4). This mechanism for active touch
contrasts with the account of sensory processing suggested by the reafference principle (Von Holst,
1954), which postulates that motor efference is discounted from sensory input, allowing animals to
sense exafferent signals (externally caused sensory input) without being confounded by the
consequences of their own motor actions. In contrast, our theory suggests that the sensory system is
insensitive to pure exafference during active sensing (Figure 4F), but sensitive to the interruption of
20
reafference which may allow animals to focus attention on the consequences of their own motor
actions.
While it is straightforward to generalize this idea to other tactile systems, its implication for other
sensory modalities is less clear. However, in theory, CLE feedback could be interrupted anywhere
along the action-perception cycle, thus dynamically regulating neuronal gain. The timely interruption
of this feedback could serve as a general mechanism for temporarily accentuating neuronal responses
against a background of reduced noise (Scott, 2004, Hafed et al., 2011). For example, cerebellum
neurons, which are strongly involved in the sensory-motor cycle, could be suppressed in anticipation
of salient sensory events by a relevant brain area, such as the reticular formation (Kinomura et al.
1996).
Rapid neurofeedback and virtual reality-based behavioral enhancement
The primate findings demonstrate that neurofeedback can modulate brain dynamics (Figure
5B and C) and enhance task-related behavior (Figure 5D, E, F, and G). These results portend
the use of CLE feedback as an interventional tool for behavioral enhancement in closed-loop feedback
therapies. Unlike more conventional neurofeedback techniques (Zoefel et al., 2011, Shibata et
al., 2011) that require training periods to affect behavior, our technique enhanced fixation
performance within seconds and required little supervision. This rapid neurofeedback was
achieved by crucial differences in the design between this technique and conventional
neurofeedback. Conventional protocols require conscious/unconscious human learning that
is supported by an appropriate neurofeedback signal. In contrast, our protocol directly
intervened on the fast time scale of the ongoing sensory stream. Our approach has the
potential to be more stable, with more rapid, user-independent, and effective behavioral
control than current conventional decoded neurofeedback methods.
The neurophysiological reason underlying the improved fixation performance is not clear. However,
the high-contrast grating stimulus likely serves as a distractor for the monkey and makes fixation more
21
difficult. During the neurofeedback protocol, the 3-4 Hz frequency component of this distractor signal
was suppressed by the feedback, possibly inhibiting the formation of a grating perception. If so, it is
possible that such CLE feedback may help to allocate attention in physiological settings. Specifically,
our model predicts that, animals may show less attention to sensory modalities that provide ongoing
negative CLE feedback. A full investigation of these issues will require further work.
These findings emphasize the importance of the development of natural and virtual reality systems that
provide well-controlled quantitative measurements of CLE interactions. Such novel systems would be
useful for studying how healthy and disease brain physiology emerges from real-time brain-body-
environment interactions and may suggest new methods for manipulating or enhancing brain
physiology and behavior with neurofeedback technology.
In sum, our findings suggest that context-dependent brain function and flexible behavior may only
emerge from context specific CLE interactions. This supports the idea that CLE interactions may have
a broad impact on our cognition (Clark, 2008) and may also shape social interactions (Froese et al.,
2014).
Materials and Methods
Simple Conceptual Model: The time traces in Figure 1 were calculated by Euler's method,
integrating the equations presented in the text with 𝑤 = −10, 𝜏 = 1, and time step 𝑑𝑡 = 0.01.
Zebrafish experiment: We analyzed the calcium signal (∆F/F) at various sample
frequencies (ca. 23 Hz) across 1908 cells in 32 fish, see Ahrens et al. (2012) for full
description of experimental method. In addition to calcium sources (putative neurons), swim
vigor was measured by taking electrical recordings from motor neurons in the fish's tail. In
the original study, the gain (i.e., the multiplicative factor between fictive swim power, and the
speed of visual feedback) was alternated between a high and low gain condition every 30 s.
22
We studied cells analyzed in Ahrens et al. (2012), which showed modulations of the mean
∆F/F, depending on the gain conditions. This gain alternating protocol is not relevant to the
current study. To reduce this variability in data, we subtracted the mean activity level in each
gain setting in our analysis (from both brain and behavior variables). Notably, our main
results were qualitatively the same, even without such subtraction of the means.
We analyzed data taken from a 6-min recording of 16 prominent calcium sources per fish,
putative neurons, across 600 trials. In the first 3 min, the fish performed the closed-loop
optomotor behavior. For the subsequent 3 min, each fish was presented with an open-loop
stimulus (no longer actively controlling their environment), which is a repeat of what the
animal experienced in the previous 3 min. In this period, sensory input to the animals was
identical and the only difference was the absence of closed-loop dynamics, allowing us to
study how reafferent signal affected neural activity.
In Figure 2B, low frequency power was calculated as the mean log power of a neuron in the
range of 0.010.15 Hz, averaged over all recorded neurons. Correlation was calculated as
the average pairwise correlation (6589 pairs of cells were analyzed) between neurons in the
same fish, averaged over all trials. Figure 2C plots the change in this pairwise correlation
against the change in the log low frequency power (averaged for the pair) between the
closed-loop and the replay condition. Note: for fish data, non-parametric tests were used and
we did not assume normality. In addition, both results reported in Figure 2B were highly
significant under a paired t-test, with p values <10-5.
We fitted the data with linear filters that describe the interaction between individual neurons
and the environmental variable. Note: here we define the environmental variable (E) as the
activity of motor neurons, whose history uniquely determines the visual stimulus. We
calculated the linear filter that minimizes the mean square error between a driven variable
𝑦(𝑡) and the convolution of a driving variable 𝑥(𝑡) and filter 𝐹(𝑡) over time. Filters were
23
constructed as a superposition of Laguere functions. We use Laguere functions up to the
order that stopped the optimization process by choosing the best Akaike Information
Criterion (Akaike, 1987). Almost all filters had an order that was mid-range between 1 and
15, indicating the robustness of the chosen order.
We inferred two filters for each cell. We determined how the environment drives the brain,
𝐹(𝑡), by directly calculating the filter between the closed-loop environment and brain in the
replay condition 𝐸 → 𝐵′, see Figure 3A (magenta line). We determined how the brain drives
the environment, 𝐺(𝑡), by first calculating the residual variability of brain in the open-loop
condition that cannot be accounted for by the closed-loop environment, i.e., 𝑅𝐵′(𝑡) =
𝐵′(𝑡) − 𝐹 ∗ 𝐸(𝑡) and subsequently calculating how 𝑅𝐵′(𝑡) drives the environment in the
open-loop condition, effectively determining 𝐵′ → 𝐸′. Self-feedback is then straightforwardly
estimated as the convolution of both filters 𝐻(𝑡) = 𝐹 ∗ 𝐺(𝑡).
In Figure 2B, we calculated the low frequency power as the mean log power of a neuron in
the range (0.010.15 Hz) and took the ratio between the closed-loop and replay condition. To
analytically calculate this low frequency power ratio from the estimated filters, we can first
write the dynamics of the brain in the closed- and open-loop conditions in the frequency
domain as
𝐵(𝜔) = 𝐻(𝜔)𝐵(𝜔) + 𝜖(𝜔) = (1 − 𝐻(𝜔))
−1
𝜖(𝜔)
𝐵′(𝜔) = 𝐻(𝜔) 𝐵(𝜔) + 𝜖′(𝜔),
where 𝜖(𝜔) and 𝜖′(𝜔) are the Fourier transforms of the noise in the closed- and open-loop
conditions that we assume as having the same frequency spectrum. The ratio of the power
between each condition is simply
𝐵(𝜔)2
𝐵′(𝜔)2 =
1
𝐻(𝜔)2+1−𝐻(𝜔)2,
24
where 𝐻(𝜔) is the estimated combined filter in the frequency domain. From this curve we can
then straightforwardly calculate the mean log power of a neuron in the range (0.010.15 Hz).
In Figure 3D, we calculated the 𝐸 → 𝐵′ filter in the replay condition and the 𝐸 → 𝐵 filter in the
closed-loop condition as above, but using the Hermite rather than the Laguere functions to
capture the acausal (t<0) side of the filter. Notably, the 𝐸 → 𝐵 filter in the closed-loop
condition generally has an acausal component, because the brain 𝐵 and the environment 𝐸
are mutually interacting (see below). On the other hand, the 𝐸 → 𝐵′ filter in the replay
condition is identical to the 𝐹(𝑡) filter defined in the previous paragraph (although with minor
differences due to the use of the Hermite- rather than Laguere-based functions). Hence, the
two filters (𝐸 → 𝐵′ and 𝐸 → 𝐵) are generally different and this difference originates from the
presence of CLE feedback (𝐵 → 𝐸 interaction) in the closed-loop condition. Here, we quantify
this difference using a simple linear model. In the closed-loop condition, the system is
described by
𝐵(𝜔) = 𝐹(𝜔)𝐸(𝜔) + 𝑅𝐵(𝜔)
𝐸(𝜔) = 𝐺(𝜔)𝐵(𝜔) + 𝑅𝐸(𝜔)
in the Fourier domain. Note that we again assumed the same interactions 𝐹(𝜔) and 𝐺(𝜔) in
the closed-loop and replay conditions. Based on the first equation, the second equation is
also described as 𝐸(𝜔) = (1 − 𝐻(𝜔))
−1
(𝐺(𝜔)𝑅𝐵(𝜔) + 𝑅𝐸(𝜔)). Therefore, the 𝐸 → 𝐵 filter in
the closed-loop condition is
𝐵(𝜔)𝐸∗(𝜔)
𝐸(𝜔)𝐸∗(𝜔)
= 𝐹(𝜔) + (
∗
∗ (𝜔)
𝐸(𝜔)𝑅𝐵
)
𝐸(𝜔)𝐸∗(𝜔)
= 𝐹(𝜔) + (
𝐺(𝜔)
∗ 𝑅𝐵(𝜔)2
)
𝐸(𝜔)2
1 − 𝐻(𝜔)
where * describes complex conjugate. Hence, this filter is different from 𝐹(𝜔) by the second
term. To predict the second term in the absence of knowing B, we assume 𝑅𝐵(𝜔)2 ≈
𝑅𝐵′(𝜔)2, where the latter spectrum is based on the residual 𝑅𝐵′ computed in the replay
condition.
25
Whisker model: We model a whisker comprised of two sections that are connected by
hinges at the center and the base, which are constrained by muscles (simple torsion
springs), see Supplementary Figure 3A. Whisking is implemented by driving the equilibrium
position of the base spring. The center spring has an equilibrium value of zero angular
displacement and thus tends to align both sections. A horizontal solid wall is placed above
the whisker and, as the whisker collides with the wall, it deforms accordingly (Supplementary
Figure 3B). By adjusting the relative stiffness of each torsion spring, we can control the
degree to which the base angle is affected by contact events, e.g., if the whisker is very
flexible, the base angle will change continuously, despite contact of the tip, see
Supplementary Figure 3B. Note: the feedback in this model could be equally mediated by
velocity or even acceleration, rather than whisker position, but neither choice would make a
significant difference to our conclusions, as long as the CLE feedback constitutes a net
negative feedback. We fixed the base spring, k1 = 1, and identified the stiffness of the
whisker with k2. For Figures 4AE, we considered a rigid whisker, but relaxed this
assumption in Figure 4F and Figures 3B and C. We simulated the dynamics of the whisker
by minimizing a Lagrangian description of the configurational energy of the massless whisker
under the constraint of the solid wall.
We model a cortical circuit comprising N excitatory and N inhibitory neurons that interact with
a single whisker. Dynamic activity of neuron i (𝑖 = 1, . . . , 𝑁 are excitatory and 𝑖 = 𝑁 + 1, . . . ,2𝑁
are inhibitory, with 𝑁 = 100) is modeled as a linear dynamical system by
𝑥𝑖 = −𝑥𝑖 + ∑
2𝑁
𝑗=1
𝜔𝑖𝑗𝑥𝑗
− 𝑎𝑖 − 𝜔𝜃𝑦 𝜃 + 𝜉𝑖 + 𝐼
Differential equations are solved by the forward Euler integration method with time-bin dt =
0.5 ms. Hereafter, all time derivatives are taken to represent single-step differences divided
by dt (e.g. 𝑥(𝑡) = [𝑥(𝑡 + 𝑑𝑡) − 𝑥(𝑡)]/𝑑𝑡), but we omit the ms time unit. 𝜔𝑖𝑗 is the synaptic
strength from neuron j to i, 𝑎𝑖 is an adaptation current that produces low frequency (ca. 1 Hz)
up/down-like oscillations (Compte et al., 2003; Gentet et al., 2010; Curto et al., 2009) in the
26
absence of neuron/whisker interactions, 𝜃 is whisker angle (positive: protracted and negative:
retracted) interacting with neurons with weight 𝜔𝜃𝑦 = 0.002, 𝐼 is exafferent input that takes a
non-zero value upon whisker stimulation, and 𝜉𝑖 is independent white noise of unit variance
added to each neuron. Here, we interpret 𝑥𝑖 as both the firing rate and membrane potential,
assuming a roughly linear relationship between the two. Entries in the connectivity matrix are
assigned as 𝜔𝑖𝑗 = 𝑏𝑖𝑗𝐽 + 𝑏′𝑖𝑗𝑔 for excitatory synapses (𝑗 = 1, . . . , 𝑁) and 𝜔𝑖𝑗 = −𝑏′′𝑖𝑗𝑔 for
inhibitory synapses (𝑗 = 𝑁 + 1, . . . ,2𝑁), where 𝑏𝑖𝑗, 𝑏′𝑖𝑗, 𝑏′′𝑖𝑗 are all random binary values that
take 𝑏0 with probability 𝑝 = 0.1 and 0 with probability 1 − 𝑝, respectively. The weights are
scaled by 𝐽 =
1
𝑝𝑁
and 𝑔 =
𝑔0
√2𝑁𝑝(1−𝑝)
, so that dynamics are insensitive to the parameter values
of 𝑝 and 𝑁. Note: all excitatory and inhibitory neurons behave similarly but adding sparse
recurrent connections between randomly selected pairs neurons can account for inter-neural
variability (Renart et. al 2010). The parameter 𝑔0 = 0.05 controls inter-neural variability, and
a value less than 1 reproduces highly synchronized up/down-like fluctuations during the quiet
state. Finally, the scaling of the connectivity matrix 𝑏0is determined such that the lead
eigenvalue of the connectivity matrix is close to unity (≈ 0.975) and the dynamics are close
to instability. The adaptation current is integrated as
𝑎 𝑖 = −0.07 𝑎𝑖 + 0.008𝑥𝑖
Over time, the adaptation variable slowly builds upon neural activity and suppresses
neurons, resulting in the ca. 1-Hz oscillation. Consequently, in the absence of whisking,
implemented by setting 𝜔𝜃𝑦 = 0, this simple network reproduces the power spectrum and
cross-correlogram of neurons in the cortex (Figure 4).
The whisker protraction of the base (𝜃1, see Supplementary Figure 3A) is driven by the sum
of activity in the excitatory population and an external CPG activity, 𝑢; i.e., it is modeled as
𝜃1 = −0.93 𝜃1 +
𝜔𝑦𝜃
𝑁
𝑁
∑ 𝑥𝑖
𝑖=1
+ 𝑢
27
and is driven by the sum of activity of the excitatory population. 𝜔𝑦𝜃 = 0.085 describes the
relative strength of the cortex vs. the CPG in driving the whisker variable. With this
parameter, the whisker is mainly driven by the CPG and is modulated by cortical activity.
During whisking, the whisker rhythm is imposed by the CPG. The rhythm is generated by a
simple stochastic oscillator, given by
𝑢 = −.98𝑢 + 2𝜋 𝐹𝑤ℎ𝑖𝑠𝑘𝑣 + 𝜉𝑢
𝑣 = −.98𝑣 − 2𝜋 𝐹𝑤ℎ𝑖𝑠𝑘𝑢 + 𝜉𝑣,
where 𝐹𝑤ℎ𝑖𝑠𝑘 = 10 Hz is the frequency of the oscillator and 𝜉𝑢 , 𝜉𝑣 are independent Gaussian
white noise.
Aside: In the current model, most excitatory neurons respond to whisker retraction and drive
whisker protraction. Adding a separate counterpart population that responds to whisker
protraction and drives whisker retraction does not change the model's behavior. In the
current model, the tuning of cortical neurons to whisker position (Diamond et. al 2008) is
mainly inherited from thalamocortical input.
Passive whisker stimulations are simulated by injecting exafferent input 𝐼 = 0.035 to the
cortical neurons for 25 ms. The magnitude of the exafferent input is selected such that it
approximately matches the evoked change over the standard deviation of the membrane
potential (𝛥𝑉𝑚 / 𝜎(𝑉𝑚)) in response to magnetic whisker deflection during the whisking
condition (Crochet and Petersen, 2006).
During active touch events, we injected exafferent input (𝐼 = 0.035) to the cortical neurons on
contact, for the duration of the contact event, but for no longer than 25 ms. The model was
run for 200 s in the closed loop, open-loop, and sustained period of active touch to calculate
all quantitative measures. The power in Figure 4C is averaged over all neurons and a
quadratic spline fitted to the data.
28
Chernoff discrimination: To calculate signal-to-noise-ratios, we calculated the Chernoff
distance (Cover, 2012) between two probability distributions, 𝑝𝐼(𝑥) and 𝑝0(𝑥), in the
presence or absence of a sensory event, respectively. Specifically,
𝛹(𝑝𝐼𝑝0) ≡ −𝑚𝑖𝑛 0<𝜆<1𝑙𝑜𝑔 ∫ 𝑝𝐼
𝜆(𝑥) 𝑝0
1−𝜆(𝑥) 𝑑𝑥
For our model, the probability distribution for each condition is well described by a Gaussian
distribution
𝑝(𝑥) = 2𝜋𝐶−1/2𝑒𝑥𝑝(−
(𝑥 − 𝜇)𝐶−1(𝑥 − 𝜇)),
1
2
where C and 𝜇 are covariance matrix and vector of means, respectively. By substituting this
into the expression for Chernoff distance and employing the Gaussian integral identity and
expressing the Chernoff distance in terms of 𝐶0, 𝐶𝐼 and 𝜇0, 𝜇𝐼, we calculate the covariance
and mean between three neurons, randomly selected from the network described in the first
section. We calculate covariances across ensembles of 500 networks every 10 ms for a
period of 1 s, starting at the onset of the sensory event. Minimization with respect to 𝜆 is
computed numerically.
Monkey experiment: ECoG activity was recorded from 128 subdural electrodes implanted
in a macaque monkey. All experimental and surgical procedures were performed in
accordance with experimental protocols (No. H24-2-203(4)) approved by the RIKEN ethics
committee and the recommendations of the Weatherall report, "The use of non-human
primates in research".
Subject:
One monkey (male Macaca mulatta, aged 12 years, wild-type) was used in the experiment after
magnetic resonance imaging. Before the implantation of subdural ECoG electrodes, the monkey was
familiarized with the experimental settings and trained with a fixation task. During the fixation task,
the monkey sat in a primate chair with the head in a fixed position using a helmet custom-made for the
29
monkey. The monkey was housed individually in the room with a 12-h lightdark cycle (lights on at
8:00 AM), and participated in the experiment in the daytime. The monkey had also participated in
other experiments involving fixation and voluntary saccadic eye movements more than half a year
previous to this study.
Electrode Implantation:
Subdural electrodes were surgically implanted after the completion of fixation training. To anesthetize
the monkey, we administered ketamine (5 mg/kg, intramuscular), atropine (0.05 mg/kg), and
pentobarbital (20 mg/kg, intravenous). We adjusted the dose of pentobarbital based on monkey's
response to pain and heart rate. We chronically implanted a customized 128-channel ECoG electrode
array in the subdural space (Unique Medical, Tokyo, Japan; Nagasaka et al., 2011).
The monkey was rewarded only in trials for which fixation was maintained from FP onset until
its disappearance. ECoG signals were recorded at a sampling rate of 1 kHz using a Cerebus
data acquisition system (Blackrock, UT, USA) and down-sampled to 200 Hz. Behavioral
variables, such as pupil size and eye position, were recorded with a custom eye-tracking
system (Nagasaka et al., 2011). Visual stimuli were constructed in Pyschtoolbox and
feedback implemented by streaming ECoG data in real-time into Matlab. We used LIBSVM
(Chang and Lin, 2011) under Matlab to implement the SVM.
During a 10-min training session, the monkey passively viewed consecutive 500-ms blocks
of either vertical or horizontal gratings in each trial (total of about 200 trials). We found that
the stimulus was most reliably decoded by applying SVM to inter-electrode correlations. We
collected 100-ms runs of ECoG data from 50 electrodes for each stimulus condition and
constructed the SVM vector by calculating the correlation between electrodes and flattening
half (upper right triangle) of the resulting matrix to form a vector. We trained an SVM to
classify vertical and horizontal grating on all collected vectors.
30
Classification performance strongly depended on the time interval used for the classifier,
typically reaching > 90% cross-validated classification accuracy with 400-ms-long intervals.
In order to reduce temporal delay for neurofeedback, we classified stimuli using short 100-ms
intervals. This typically gave > 60% cross-validated classification accuracy which is still
markedly higher than chance level. A better classification performance was also achieved
from grating orientations separated by 90, rather than 180 degrees, motivating our choice for
horizontal and vertical gratings for the neurofeedback.
We then estimated the decision thresholds, by calculating a distribution of the decision
values for each stimulus condition from the training data. These were roughly Gaussian with
a positive and negative mean for the horizontal and vertical grating stimuli, respectively. We
estimated the threshold for the decision value above (𝜃ℎ𝑜𝑟) and below (𝜃𝑣𝑒𝑟𝑡) which vertical
and horizontal gratings are shown, respectively (see main text), by calculating the absolute
value difference between the distributions and selecting the peaks of maximal difference. No
gratings were presented for at least 1 s after the fixation point onset, to avoid presenting
stimulus before the monkey fixated his eyes.
During testing, we consecutively sampled 100-ms runs of ECoG data, constructed an SVM
vector, and used the SVM kernel to calculate the decision variable. The Matlab processing
time for classifications was shorter than 100 ms. We were able to record 50108 trials for
each of the closed-loop and replay conditions per experiment. We present data from four
recording days in the text. We recorded a minimum of 326 trials each under the closed-loop and
replay conditions. Together with the calibration of the decoder, this added to a reasonable
experimental length for the subject to maintain his attention. In addition, the number of trials was
comparable to that of typical experiments performed in visual psychophysics.
The power spectrum of the decision value (Figure 5B) was calculated from the ECoG signal
within 10002800 ms after the fixation point onset. Trials in which the monkey failed to fixate
31
his eyes for at least 2800 ms were eliminated from the data set. The improvement in the
performance was measured by the difference between the positions in each condition
(Replay - Closed) relative to the average position (Replay + Closed) / 2 within each time-bin.
Normality for applying the t-test (Figure 5C) was examined using the KolmogorovSmirnov
goodness-of-fit hypothesis test (P > 0.07).
The performance improvement for each trial pair was summarized by root-mean-square of
this quantity measured over an interval of 16002100 ms. We computed the trial-to-trial
correlation between the fixation performance improvement and the power spectrum of the
decision value dynamics in the interval of 10002100 ms (Figures 5F and 5G): the power
spectrum within this time period reflects the brain dynamics that precede the behavioral
change. We also confirmed that the overall trend of the power spectrum of the decision value
within this period (Supplementary Figure 5D) had not changed from that derived from the
whole stimulation period shown in Figure 5B.
Acknowledgements: We would like to thank the authors of Ahrens et. al. (2012) for kindly
providing the zebrafish data and Poulet et. al. (2012) providing the rodent whisking data. We
would like to thank Charles Yokoyama and Alexandra V Terashima for comments on the
manuscript, and Misha Ahrens, Hideaki Shimazaki, Hokto Kazama, Masanori Murayama,
Sylvain Crochet and L.F. Abbott for helpful discussions.
Author Contributions: C.L.B and T.T. developed the theory. C.L.B. simulated the whisker
model and analyzed the zebrafish data, in discussion with T.T. C.L.B. and T.T. planned the
monkey experiment. C.L.B., S.T., T.Y., K.T. and Y. N. conducted the monkey experiment
with the help of N.F. S.T. analyzed the monkey data based on the initial account by C.L.B.
C.L.B. and T.T. wrote the manuscript.
References
Ahissar, E., & Kleinfeld, D. (2003). Closed-loop neuronal computations: focus on vibrissa
somatosensation in rat. Cerebral Cortex, 13(1), 53-62.
32
Ahrens, M. B., Li, J. M., Orger, M. B., Robson, D. N., Schier, A. F., Engert, F., et al., (2012).
Brain-wide neuronal dynamics during motor adaptation in zebrafish. Nature, 485(7399),
Akaike, H. (1987). Factor analysis and AIC. Psychometrika, 52(3), 317-332.
Berger, H. (1929). Über das elektrenkephalogramm des menschen. European Archives of
Psychiatry and Clinical Neuroscience, 87(1), 527-570.
Bruno, R. M., & Simons, D. J. (2002). Feedforward mechanisms of excitatory and inhibitory cortical
receptive fields. The Journal of neuroscience, 22(24), 10966-10975.
Buzsáki, G., Anastassiou, C. A., and Koch, C. (2012). The origin of extracellular fields and
currents-EEG, ECoG, LFP and spikes. Nature reviews neuroscience, 13(6), 407-420.
Castro-Alamancos, M. A. (2004). Absence of rapid sensory adaptation in neocortex during
information processing states. Neuron, 41(3), 455-464.
Chang, C., & Lin, C. (2011). LIBSVM: A library for support vector machines. ACM
Transactions on Intelligent Systems and Technology (TIST), 2(3), 27.
Churchland, M. M., Byron, M. Y., Cunningham, J. P., Sugrue, L. P., Cohen, M. R., Corrado,
G. S., et al., (2010). Stimulus onset quenches neural variability: a widespread cortical
phenomenon. Nature neuroscience, 13(3), 369-378.
Clark, A. (2008). Supersizing the mind: Embodiment, action, and cognitive extension: Embodiment,
action, and cognitive extension. Oxford University Press.
Compte, A., Sanchez-Vives, M. V., McCormick, D. A., & Wang, X. (2003). Cellular and
network mechanisms of slow oscillatory activity (< 1 Hz) and wave propagations in a cortical
network model. Journal of neurophysiology, 89(5), 2707-2725.
Constantinople, C. M., & Bruno, R. M. (2011). Effects and mechanisms of wakefulness on
local cortical networks. Neuron, 69(6), 1061-1068.
Crochet, S., & Petersen, C. C. (2006). Correlating whisker behavior with membrane potential
in barrel cortex of awake mice. Nature neuroscience, 9(5), 608-610.
Crochet, S., Poulet, J. F., Kremer, Y., & Petersen, C. C. (2011). Synaptic mechanisms
underlying sparse coding of active touch. Neuron, 69(6), 1160-1175.
Crapse, T. B., & Sommer, M. A. (2008). Corollary discharge across the animal kingdom. Nature
Reviews Neuroscience, 9(8), 587-600.
Curtis, J. C., & Kleinfeld, D. (2009). Phase-to-rate transformations encode touch in cortical
neurons of a scanning sensorimotor system. Nature neuroscience, 12(4), 492-501.
Curto, C., Sakata, S., Marguet, S., Itskov, V., & Harris, K. D. (2009). A simple model of
cortical dynamics explains variability and state dependence of sensory responses in
urethane-anesthetized auditory cortex. The Journal of Neuroscience, 29(34), 10600-10612.
deBettencourt, T. M., Cohen, J. D., Lee, R. F., Norman, K. A., & Turk-Browne, N. B. (2015).
Closed-loop training of attention with real-time brain imaging. Nature neuroscience, 18(3),
470-475.
33
Diamond, M. E., von Heimendahl, M., Knutsen, P. M., Kleinfeld, D., & Ahissar, E. (2008).
'Where'and'what'in the whisker sensorimotor system. Nature Reviews Neuroscience, 9(8),
601-612.
Eliades, S. J., & Wang, X. (2008). Neural substrates of vocalization feedback monitoring in
primate auditory cortex. Nature, 453(7198), 1102-1106.
Eliades, S. J., & Wang, X. (2008). Neural substrates of vocalization feedback monitoring in
primate auditory cortex. Nature, 453(7198), 1102-1106.
Fanselow, E. E., & Nicolelis, M. A. (1999). Behavioral modulation of tactile responses in the
rat somatosensory system. The Journal of neuroscience, 19(17), 7603-7616.
Fee, M. S., Mitra, P. P., & Kleinfeld, D. (1997). Central versus peripheral determinants of
patterned spike activity in rat vibrissa cortex during whisking. Journal of neurophysiology,
78(2), 1144-1149.
Felsen, G., & Dan, Y. (2005). A natural approach to studying vision. Nature neuroscience,
8(12), 1643-1646.
Froese, T., Iizuka, H., & Ikegami, T. (2014). Embodied social interaction constitutes social
cognition in pairs of humans: a minimalist virtual reality experiment. Scientific reports, 4.
Fry, S. N., Rohrseitz, N., Straw, A. D., & Dickinson, M. H. (2009). Visual control of flight
speed in Drosophila melanogaster. Journal of Experimental Biology, 212(8), 1120-1130.
Fu, Y., Tucciarone, J. M., Espinosa, J. S., Sheng, N., Darcy, D. P., Nicoll, R. A., et al.,
(2014). A cortical circuit for gain control by behavioral state. Cell, 156(6), 1139-1152.
Gentet, L. J., Avermann, M., Matyas, F., Staiger, J. F., & Petersen, C. C. (2010). Membrane
potential dynamics of GABAergic neurons in the barrel cortex of behaving mice. Neuron,
65(3), 422-435.
Goard, M., & Dan, Y. (2009). Basal forebrain activation enhances cortical coding of natural
scenes. Nature neuroscience, 12(11), 1444-1449.
Hafed, Z. M., Lovejoy, L. P., & Krauzlis, R. J. (2011). Modulation of microsaccades in
monkey during a covert visual attention task. The Journal of Neuroscience, 31(43), 15219-
15230.
Harris, Kenneth D, and Alexander Thiele. "Cortical state and attention." Nature Reviews
Neuroscience 12.9 (2011): 509-523.
Hill, D. N., Curtis, J. C., Moore, J. D., & Kleinfeld, D. (2011). Primary motor cortex reports
efferent control of vibrissa motion on multiple timescales. Neuron, 72(2), 344-356.
Ito, M. (1972). Neural design of the cerebellar motor control system. Brain research, 40(1),
81-84.
Aström, K. J., & Murray, R. M. (2010). Feedback systems: an introduction for scientists and
engineers. Princeton university press.
Kawato, M. (1999). Internal models for motor control and trajectory planning. Current opinion
in neurobiology, 9(6), 718-727.
34
Keller, G. B., & Hahnloser, R. H. (2009). Neural processing of auditory feedback during vocal
practice in a songbird. Nature, 457(7226), 187-190.
Keller, G. B., Bonhoeffer, T., & Hübener, M. (2012). Sensorimotor mismatch signals in
primary visual cortex of the behaving mouse. Neuron, 74(5), 809-815.
Kinomura, S., Larsson, J., Gulyas, B., & Roland, P. E. (1996). Activation by attention of the
human reticular formation and thalamic intralaminar nuclei. Science, 271(5248), 512-515.
Kleinfeld, D., Ahissar, E., & Diamond, M. E. (2006). Active sensation: insights from the rodent
vibrissa sensorimotor system. Current opinion in neurobiology, 16(4), 435-444.
Lee, S., Carvell, G. E., & Simons, D. J. (2008). Motor modulation of afferent somatosensory
circuits. Nature neuroscience, 11(12), 1430-1438.
Lisberger, S. G., Morris, E., & Tychsen, L. (1987). Visual motion processing and sensory-
motor integration for smooth pursuit eye movements. Annual review of neuroscience, 10(1),
97-129.
Lovick, T. (1972). The behavioural repertoire of precollicular decerebrate rats. The Journal of
physiology, 226(2), 4P-6P.
Matyas, F., Sreenivasan, V., Marbach, F., Wacongne, C., Barsy, B., Mateo, C., et al., (2010).
Motor control by sensory cortex. Science, 330(6008), 1240-1243.
McGinley, M. J., David, S. V., & McCormick, D. A. (2015). Cortical membrane potential
signature of optimal states for sensory signal detection. Neuron, 87(1), 179-192.
Nguyen, Q., & Kleinfeld, D. (2005). Positive feedback in a brainstem tactile sensorimotor
loop. Neuron, 45(3), 447-457.
Niell, C. M., & Stryker, M. P. (2010). Modulation of visual responses by behavioral state in
mouse visual cortex. Neuron, 65(4), 472-479.
O'Regan, J. K., & Noë, A. (2001). A sensorimotor account of vision and visual
consciousness. Behavioral and brain sciences, 24(05), 939-973.
Otazu, G. H., Tai, L., Yang, Y., & Zador, A. M. (2009). Engaging in an auditory task
suppresses responses in auditory cortex. Nature neuroscience, 12(5), 646-654.
Pierret, T., Lavallée, P., & Deschênes, M. (2000). Parallel streams for the relay of vibrissal
information through thalamic barreloids. The Journal of Neuroscience, 20(19), 7455-7462.
Pinto, L., Goard, M. J., Estandian, D., Xu, M., Kwan, A. C., Lee, S., et al., (2013). Fast
modulation of visual perception by basal forebrain cholinergic neurons. Nature neuroscience,
16(12), 1857-1863.
Polack, P., Friedman, J., & Golshani, P. (2013). Cellular mechanisms of brain state-
dependent gain modulation in visual cortex. Nature neuroscience, 16(9), 1331-1339.
Poulet, J. F., & Petersen, C. C. (2008). Internal brain state regulates membrane potential
synchrony in barrel cortex of behaving mice. Nature, 454(7206), 881-885.
Poulet, J. F., Fernandez, L. M., Crochet, S., & Petersen, C. C. (2012). Thalamic control of
35
cortical states. Nature neuroscience, 15(3), 370-372.
Renart, A., de la Rocha, J., Bartho, P., Hollender, L., Parga, N., Reyes, A., et al., (2010). The
asynchronous state in cortical circuits. science, 327(5965), 587-590.
Schneider, D. M., Nelson, A., & Mooney, R. (2014). A synaptic and circuit basis for corollary
discharge in the auditory cortex. Nature.
Scott, S. H. (2004). Optimal feedback control and the neural basis of volitional motor control.
Nature Reviews Neuroscience, 5(7), 532-546.
Semba, K., & Komisaruk, B. (1984). Neural substrates of two different rhythmical vibrissal
movements in the rat. Neuroscience, 12(3), 761-774.
Shibata, K., Watanabe, T., Sasaki, Y., & Kawato, M. (2011). Perceptual learning incepted by
decoded fMRI neurofeedback without stimulus presentation. science, 334(6061), 1413-1415.
Smith, K. U. (1962). Delayed sensory feedback and behavior. W. B. Saunders, Philadelphia
Steriade, M. (2001). Impact of network activities on neuronal properties in corticothalamic
systems. Journal of neurophysiology, 86(1), 1-39.
Steriade, M. (2001). Impact of network activities on neuronal properties in corticothalamic
systems. Journal of neurophysiology, 86(1), 1-39.
Steriade, M. (2005). Cellular substrates of brain rhythms. Electroencephalography: Basic
principles, clinical applications, and related fields, 5, 31-83.
Steriade, M. M., & McCarley, R. W. (2013, March 9). Brainstem control of wakefulness and
sleep. Springer Science & Business Media.
Szwed, M., Bagdasarian, K., & Ahissar, E. (2003). Encoding of vibrissal active touch.
Neuron, 40(3), 621-630.Eliades, S. J., & Wang, X. (2008). Neural substrates of vocalization
feedback monitoring in primate auditory cortex. Nature, 453(7198), 1102-1106.
Tan, A. Y., Chen, Y., Scholl, B., Seidemann, E., & Priebe, N. J. (2014). Sensory stimulation
shifts visual cortex from synchronous to asynchronous states. Nature, 509(7499), 226-229.
Urbain, N., Salin, P. A., Libourel, P., Comte, J., Gentet, L. J., & Petersen, C. C. (2015).
Whisking-Related Changes in Neuronal Firing and Membrane Potential Dynamics in the
Somatosensory Thalamus of Awake Mice. Cell reports, 13(4), 647-656.
Von Holst, E. (1954). Relations between the central nervous system and the peripheral
organs. The British Journal of Animal Behaviour, 2(3), 89-94.
Vyazovskiy, V. V., Olcese, U., Hanlon, E. C., Nir, Y., Cirelli, C., & Tononi, G. (2011). Local
sleep in awake rats. Nature, 472(7344), 443-447.
Welker, W. (1964). Analysis of Sniffing of the Albino Rat 1). Behaviour, 22(3), 223-244.
Wolf, R., Voss, A., Hein, S., Heisenberg, M., & Sullivan, G. (1992). Can a Fly Ride a
Bicycle?[and Discussion]. Philosophical Transactions of the Royal Society B: Biological
Sciences, 337(1281), 261-269.
36
Wolpert, D. M., & Ghahramani, Z. (2000). Computational principles of movement
neuroscience. nature neuroscience, 3, 1212-1217.
Y. Nagasaka, K. Shimoda, N. Fujii. Multidimensional recording (MDR) and data sharing: an
ecological open research and educational platform for neuroscience. PloS one, 6(7), e22561
(2011).
Zagha, E., Casale, A. E., Sachdev, R. N., McGinley, M. J., & McCormick, D. A. (2013). Motor
cortex feedback influences sensory processing by modulating network state. Neuron, 79(3),
567-578.
Figures
Figure 1 A simple model of the brain-environment interaction. (A) A schematic description
of brain-body-environment interactions during closed-loop behavior. The brain receives two
types of sensory input: exafferent input that originates from the environment and reafferent
input, which, while mediated by the environment, results from the consequences of an
animal's own actions. (B-D) Schematic diagrams (Top) and the model's representative brain
activity traces (Bottom) under the following three conditions. In the open-loop condition (B),
the brain receives no reafferent input and exhibits collective activity that spontaneously
fluctuates. In the closed-loop condition (C), reafferent input constitutes a CLE feedback to the
brain. If this feedback is negative, the gain of the brain is reduced and fluctuations are
suppressed. In the replay condition (D), the brain receives a replay of the reafferent input in
37
the closed-loop condition as exafferent input. Any differences from the closed-loop condition
are caused by the absence of CLE feedback because the sensory input is identical to that in
the closed-loop condition. In this condition, the gain of the brain is not suppressed and
fluctuations are much larger than in the closed-loop condition with negative feedback.
Figure 2. CLE feedback suppresses neural fluctuations and correlations.
(A) Photograph of a paralyzed larval zebrafish (left) in the experimental setup (right),
supported by pipettes that record motor activity. (B) Population averages of logarithmic low
frequency power (mean over interval of [0.01 0.15] Hz) (left) and pairwise intra-neural
correlations (right) were both suppressed under the closed-loop condition relative to the
38
replay condition. (C) These changes in pairwise correlations and low frequency power
(replay – closed) were highly correlated in the recorded neurons.
Figure 3. CLE feedback predicts suppression of neural fluctuations and correlations. (A)
Dynamic interactions were estimated for each neuron by fitting linear filters, whose
population averages, after normalizing to peak amplitudes, are summarized. Schematic
interactions of the brain (a single neuron) and the environment (motor neuron activity) are
shown under the closed-loop and replay conditions (inset). On average, the brain positively
drove the environment (B'→E', purple line); the environment negatively drove the brain
39
(E→B', pink line), and, by combining these two effects, we found that self-feedback
(E→B'→E', cyan line) was negative. (B) These filters were then used to predict changes of
neural fluctuations under the two conditions. The predicted changes in each neuron based
on the filters exhibited strong correlation with the actual changes. Some neurons (top 10%,
red dots) exhibited strong negative CLE feedback and were stabilized under the closed-loop
condition as predicted by our theory. (C) The location of these neurons are overlaid with the
morphology of a reference zebrafish brain (colors as in B). Top panel, side view; bottom
panel, top view. Neurons that have strong negative CLE feedback and are strongly stabilized
were predominantly located in the cerebellum. (D) The dynamic relation of neuronal activity
and motor activity for each neuron in the closed-loop condition (quantified by the E→B filter,
naively computed) was qualitatively different from that in the replay condition (the E→B' filter)
(inset). For the closed-loop-stabilized cells in (C, red dots), this difference could be explained
well by the B'→E' filter from the replay condition (Inset: E→B, black line; E→B', blue line;
E→B estimated taking into account of the B'→E' filter from the replay condition, cyan line).
The degree to which the difference was accounted for by the B'→E' filter was computed for
each cell as a fraction of the mean square error explained. This quantity was positively
correlated with the log-power ratio, indicating that the difference in dynamic relation of
neuronal activity and motor activity between the closed-loop and replay conditions was
greater in the closed-loop stabilized cells.
40
Figure 4 A model of the rodent brain state transition. (A) A schematic of the model: 100
excitatory (Exc) and 100 inhibitory (Inh.) neurons receive CLE feedback via a single whisker
driven by a central pattern generator (CPG). Triangle and circles represent excitatory and
inhibitory synapse respectively. Onset of whisking occurs when the CPG is switched on. CLE
feedback is negative overall because the neurons that elicit whisker protraction are assumed
41
to drive whisker retraction. (B) Membrane potential of cortical neurons (gray lines for
individual neurons and black line for population average) and whisker position (blue line)
during quiet attentive (Q) whisking (W), and periods of active touch (T). Large and
synchronous fluctuations of membrane potential were suppressed during whisking. Active
touch elicited reliable responses in these neurons. The vertical dotted line marks the onset of
whisking and the vertical red lines mark onset of individual touch event. The power spectrum
(C; inset for variance) and cross-correlation (D; inset for correlation matrix of randomly
sampled neurons-color warmth indicates the degree of correlation) of membrane potential
are averaged over cortical neurons and shown for each condition. Low frequency fluctuations
and inter-neural correlation are suppressed during whisking but are recovered during the
period of active touch. (E) Membrane potential traces for the Q, W, and T conditions.
Sensory events begin at time 0. (F) The discriminability of each type of sensory event.
Discrimination performance was similar under Q and W because both signal (exafferent
evoked response) and noise (spontaneous fluctuations) were large under Q and both are
small during W. Discrimination performance for active touch events were improved relative to
the Q and W conditions unless the whisker was too flexible. Discrimination performance was
improved with increasing whisker stiffness, reflecting the degree to which CLE feedback was
stopped during touch events.
42
Figure 5 Real-time visual neurofeedback altered brain activity and improved fixation
performance. (A) The experimental setup. An estimate of the current visual stimulus
(horizontal or vertical grating) is decoded (decision value) every 100 ms from electrodes
distributed across the visual cortex. Whenever decision values indicate a high confidence for
either grating stimuli the opposite grating stimulus is presented. The stimulus is a gray
screen for small decision values (low confidence). The box at the bottom shows the
representative decision value dynamics (red trace) and sampling points (circles) used during
neurofeedback. The shaded areas show the periods during which the visual stimulus were
43
presented; the dotted lines, show the decision thresholds (purple: horizontal; green: vertical).
Time 0 indicates appearance of the fixation point. (B) Difference of power spectrum of the
decoder decision value (Replay-Closed). Average and SEM of whole trials. (C) Summary of
the 3-4 Hz amplitude of the difference of decision value in (B). (D) The improvement of
fixation performance was quantified by the relative difference between the eyes deviation
from fixation point (Replay-Closed) / (Replay+Closed)×2. The mean and SEM of whole trials
were used for the analysis. Fixation performance was quantified by the root mean square of
the distance from fixation point. (E) Average fixation performance during the time interval
[1600-2100 ms] in (D). (F) Trial-to-trial correlation coefficient between the differential fixation
performance and the differential amplitude for individual frequencies (bin width: 0.25 Hz). (G)
Summary of correlation coefficient for 3-4Hz. The errorbar shows a bootstrapped SEM.
Asterisk indicates statistical significance (*: P < 0.05; **: P < 0.005).
Figure 6 A summary of the three experimental systems studied (Top). In all systems, CLE
feedback plays a critical role in determining brain dynamics and behavior, as summarized in
the table (Bottom).
44
SUPPLEMENTARY INFORMATION
Contents
S1 Supplementary fish data
S2 A role for CLE feedback in the brain state transition
S3 Whisker model
S4 A simple conceptual model of CLE feedback
S5 Decoding of visual stimulus based on ECoG signals
S6 Alternate schemes for CLE feedback in the whisker system
S1 Supplementary fish data
45
Figure S1 (A) The decorrelation effect is not an artifact of measurement noise. Changes in
pairwise correlations and change in low frequency power were highly correlated in the
recorded neurons even when the each calcium traces (both cells and motor neurons) were
thresholded (the threshold was equal to the mean plus one standard deviation of the calcium
signal measured over both replay and closed loop conditions) (Spearman's rank correlation
=0.57, p<10-8). (B) The increase of motor activity in the closed-loop condition does not
explain reduction in neural fluctuations and correlation. (B,top): Changes in mean motor
activity(replay-closed) and change in low frequency power (mean over interval [0.01 0.15]
Hz) per trial (low frequency power is averaged over all cells within a given trial) are positively
46
correlated (r=0.18, p<10-2, Spearman's rank correlation). B(bottom): Changes in mean motor
activity and changes in pairwise correlations (averaged over all pairwise interactions in a
given trial) are not significantly correlated (r=0.03, p>0.5, Spearman's rank correlation). (C)
Filter describing behavioral feedback. A linear filter that describes behavioral feedback
(E→E') is also strongly negative. Following the kernel method outlined in the methods
section we calculated the behavioral feedback as a direct filter between the closed loop
environment and environment in the replay condition (E→E'). This also indicates that the
CLE feedback to the is strongly negative. Note: the magnitude of the CLE feedback is much
greater than the cell self-feedback reflecting the fact that cellular variability was much greater
than variability across animals.
S2 A role for CLE feedback in the brain state transition
In this section, we analyze if sensory input through infraorbital nerve (ION) plays a role in
coordinating whisking behavior, thalamic spiking activity, and cortical local field potential
(LFP). Previous results based on simultaneous recording from whisker, thalamus, and cortex
exhibited that thalamic spiking rate increased and low frequency power of cortical LFP decreased
during whisking behavior (Poulet et. al 2012). Here, we reanalyze this data set and quantify temporal
coordination between (1) 5-20Hz power of the whisker position, denoted by Whisker; (2) thalamic
spiking rate computed with 20ms averaging window, denoted by Thalamus; (3) and 1-20Hz cortical
LFP power, denoted by Cortex, recorded from ION-intact animals (n=22) and ION-cut animals
(n=19). Raw recordings and these processed traces are shown in Supplementary Figure 2A for an
example animal. We chose the 1-20 Hz range for the analysis of cortical LFP power because notable
brain-state-dependent changes were previously observed in this range (\cite{poulet14}). The
spectrogram was computed using 2s window to reliably estimate the predominant 1Hz power in
cortical LFP and the window was gradually shifted in 20ms steps.
Next, we computed cross-correlation functions between these 3 quantities: Whisker-Thalamus,
Whisker-Cortex, and Thalamus-Cortex. While the resulting cross-correlation functions were noisy in
each animal, a mean cross-correlation function averaged over each animal group exhibited clear
47
common properties. In both ION-intact and ION-cut animals, whisking behavior lead correlated
increase in the thalamic activity and decrease in the cortical slow oscillations. Consistent with this
result, the thalamic activity was negatively correlated with the low-frequency cortical LFP fluctuations
(Supplementary Figure 2B).
However, the position of the mean cross-correlation peaks was significantly shifted in ION-cut
animals relative to the ION-intact animals (Supplementary Figure 2B). Specifically, the peak of the
Whisker-Thalamus cross-correlation was delayed for 400 ms (p=0.02, bootstrap test) and the peak of
Whisker-Cortex cross-correlation was delayed for 200 ms (p=0.03, bootstrap test) in ION-cut animals.
On the other hand, the temporal relationship between the thalamic spiking activity and the low-
frequency cortical LFP power was not significantly altered as assessed by the Thalamus-Cortex
correlation function (p>0.05, bootstrap test). The corresponding bootstrap statistics were computed by
randomly resampling animals from the two groups, assuming a null hypothesis that the two animal
groups are the same (see, the inset panels for the bootstrap statistics about the difference of the cross-
correlation peaks).
These analyses suggest that, after a whisking onset, the brain state transition was delayed in ION-cut
animals relative to ION-intact animals. Thus, while sensory input is not necessary for the brain state
transition, it was necessary for inducing short-latency brain state transitions.
48
Figure S2 (A) A simultaneous recording of whisker position (Top), thalamic spikes (Middle),
and cortical LFP (Bottom) in an example animal (Poulet et. al 2012). .Based on these raw
traces (black), brain-state-relevant quantities (red) are computed and shown in each panel:
5-20 Hz power of the whisker position (Top), thalamic spiking rate (Middle), 1-20 Hz power of
the cortical LFP (Bottom). (B) A cross-correlation function between Whisker and Thalamus
(Top), Whisker and Cortex (Middle), and Thalamus and Cortex (Bottom) for ION-intact
animals (blue) and ION-cut animals (red), where the inset panels show the bootstrap
statistics about the difference of the cross-correlation peaks. The Whisker-Thalamus and the
Whisker-Cortex correlation functions were significantly shifted by the ION cut.
S3 Whisker model
49
Figure S3: (A) A whisker comprising of two sections with a joint angle
and base angle
.
The base and joint are constrained by two springs with spring constants k1, and k2
respectively. Whisking is implemented by driving the equilibrium position of the base spring.
The center spring is in equilibrium at zero angular displacement and tends to align the
50
whisker sections. The whisker is length L and massless but constrained by a solid wall
placed at y, where y<L. (B) The relationship between the joint (red) and base (blue) angle in
a flexible, k2 =.1, (middle) and stiff, k2 =10, (bottom) whisker. (C) The dependence of cortical
fluctuations and responses on whisker stiffness. Model data from quiet attentive (Q, green),
whisking (W,blue) and during contact events (red) for whiskers of increasing stiffness.
(C,top): average cross correlation of membrane between neurons during Q, W and period of
active touch for different stiffnesses (k2/(k1=1)). (C,bottom): same as top but for low
frequency power (average over the range [0.5, 2]Hz).
S4 A simple conceptual model of CLE feedback
We illustrate our whisker theory with an extension of simple conceptual model presented in
the text, Figure 1. It describes the dynamic interaction between brain activity (here
membrane potential of neurons) and the body (here whisker position) (Supplementary Figure
4). Specifically a simplified brain-environment system was modeled as a stochastic 1-
dimensional first order linear ordinary differential equation as follows,
𝐵𝑐 (𝑡) = (𝜔𝑠 − 1)𝐵𝑐(𝑡)– 𝜔𝐸(𝑡) + 𝜉(𝑡) + 𝐼(𝑡)
𝐸(𝑡) = 𝐵𝑐(𝑡)
integrated with a Euler step 𝑑𝑡 = 0.01, where 𝐵𝑐 is a dimensionless dynamical variable
representing the collective activity of neurons (e.g. local field potential) and 𝐸 is the
environmental variable (e.g. whisker position). Note: for simplicity we assume that the
dynamics of the environment variable are much faster compared with the brain dynamics
and, thus, E rapidly converges to current B. The negative CLE feedback robustly reduces
neural fluctuations unless feedback delay is too large. The parameter, 𝜔𝑠 = 0.95, is the self-
coupling within the brain, 𝜔, is the magnitude of the environmental influence on the brain, 𝐼 is
external input and, 𝜉 is unit variance white noise of mean 0. An open-loop condition, i.e., the
absence of environmental coupling, is modeled by setting, 𝜔 = 0 and the closed-loop
condition by setting 𝜔 = 0.4. Here we have set 𝜔𝑠 to be close to criticality (𝜔𝑠 = 1) to show
that even weak CLE feedback can significantly reduce neural fluctuations. However, this
51
choice is not central to our theory because strong CLE feedback works likewise when the
system is away from criticality. Passive sensory stimulation was modeled by setting, 𝐼 = 1.5
for a time period of 20. Interruption of the reafferent signal was implemented by fixing the
environmental variable for the same period, i.e, setting 𝐸(𝑡′) = 𝐸(𝑡) for 𝑡 < 𝑡′ < 𝑡 + 20.
The replay condition is constructed by first recording the environmental variable, 𝐸, in the
closed-loop condition and subsequently replaying this recording to the brain at a different
time. The brain activity in the replay condition, 𝐵𝑟, is described by
𝐵𝑟 (𝑡) = (𝜔𝑠 − 1)𝐵𝑟(𝑡) − 𝜔𝐸(𝑡) + 𝜁(𝑡)
where 𝜔 = 0.4 as in the closed-loop condition and 𝜁 is white noise during the replay
condition.
Like our simple conceptual model under these conditions CLE feedback reduces neuronal
gain and suppresses neural fluctuations in the closed-loop condition (Supplementary Figure.
4E, blue lines). Furthermore, the responses of neurons to exafferent input are also
suppressed (shaded black interval in Supplementary Figure. 4E). However, large responses
during the active state are recovered if exafferent input coincides with brief interruptions of
CLE feedback (e.g., in Supplementary Figure. 4E the environmental variable is briefly fixed --
on the surface of an object for whisker touch events -- after the onset of stimulation for a
period indicated by the shaded red interval; see Supplementary Figure. 4 for a schematic
diagram). Specifically, this brief interruption temporarily increases neuronal gain and
enhances the brain's response to exafferent input.
52
Figure S4 CLE feedback explains the changes in brain state at the onset of active behavior.
A-D: Schematics of the interaction of a brain (B) and environment (E) variable. In the open-
loop condition (A), the brain receives passive exafferent input only. In the closed-loop
condition (B), an environment variable mediates negative CLE feedback to the brain. In the
reafferent-cut condition, (C), this feedback is briefly interrupted. In the replay condition (D),
the environmental input to the brain in the closed-loop condition is recorded (denoted by Er)
and played back to an identical brain (albeit with different noise) at a later time. In all
conditions, in addition to the reafferent input, the brain receives a short pulse of exafferent
input. (E): Traces of the brain and environment variables under different conditions. In the
open loop condition (green) the brain shows large spontaneous fluctuations and large
response to exafferent stimulation, which are suppressed during the closed-loop condition
(red). The gray bar denotes periods of stimulation. In contrast, the brain variable exhibits
large and reliable response to the same input when combined with a brief interruption of the
53
environment (the pink bar denotes the period of stimulation and reafference interruption). In
the replay condition (magenta), the brain behaves similarly to the open-loop condition despite
receiving the same input from the environment as the closed-loop condition.
S5 Decoding of visual stimulus based on ECoG signals
54
Figure S5 Decoding of visual stimulus based on ECoG signal. (A) ECoG electrodes used in
the visual neurofeedback experiment. Small black circles indicate the loci of electrodes. The
electrodes surrounded by red circles were used for decoding by SVM. (B) Decision value
distribution for a trained SVM, in the visual neurofeedback experiment. The purple and the
green curves indicate cumulative probability distribution of the decision value during
presentation of vertical and horizontal gratings, respectively. The dashed lines show the
criteria for presenting feedback stimulus. (C) Averaged temporal evolution of stimulus
presentation frequency within a trial during the visual neurofeedback experiment. The visual
stimulus started to appear typically 1300 ms after the onset of fixation point. To avoid
presenting stimulus before monkey fixated its eyes, no stimulus was presented 0-1000 ms
after the appearance of fixation point. (D) Difference of power spectrum of the decoder
decision value (Replay-Closed) during early period (before the end of behavioral
improvement, 1000-2100 ms). The trend was not changed from the result for whole stimulus-
presentation period (1000-2800 ms), including the significant reduction of the 3-4 Hz power
by the neurofeedback. Convention follows Figure 4B. The black bar indicates the 3-4 Hz,
used in the analysis of behavioral correlation in Figure 4f and 4G.
S6 Alternate schemes for sensory feedback in the whisker system
A
55
B
56
ParameniscalCortexVPMPOmFNPrVSpViRetractionThalamusBrainstemExcitation InhibitionCPGEIProtractionEfferenceReafferenceExafferenceLeminiscalParameniscal C
Figure S6: Three models for whisker circuits mediating a negative CLE feedback. (A) Net
activation of the modeled cortical population drives neurons in the facial nucleus (FN) to drive
57
ParameniscalCortexVPMPOmFNPrVSpViThalamusBrainstemExcitation InhibitionCPGEIProtractionEfferenceReafferenceExafferenceLeminiscalParameniscalExcitation InhibitionCortexVPMPOmFNSpViRetractionThalamusBrainstemCPGEIProtractionEfferenceReafferenceExafferencePrVTop Down ControlLeminiscalParameniscal whisker protraction. This in turn reduces excitatory sensory input to the modeled cortical
population because this is driven by retraction. Here negative CLE feedback is mediated
implicitly at the periphery. (B) Protraction information could be conveyed along the full
pathway but net inhibitory input to the modeled cortical population result because POm
inhibits the cortex. In either model circuit, the initial activation of cortical neurons causes
subsequent suppression of their activity by feedback through the whisker circuit, constituting
a negative sensory feedback loop. These two hypotheses are testable but not necessarily
mutually exclusive. (C) An alternative hypothesis is the whisker feedback is completed in the
brain stem and changes in cortical activity are driven by activity changes in the thalamus.
Our whisker model remains abstract in terms of known vibrissa system anatomy and, in
particular, the relay stations between the cortex and a whisker. The exact concordance of the
model with known vibrissa system anatomy is beyond the scope of this paper, but we provide
a more detailed to demonstrate a possible anatomical explanation of our model and provide
a means for the research community to experimentally examine CLE feedback in specific
biological circuits.
In our model (Supplementary Figure 6A), we assume that projections between regions are
largely excitatory (c.f. Ahissar 2010). Importantly, we distinguish two subcortical pathways
that signal afferent input to cortical neurons - one for transmitting reafferent input and the
other for transmitting exafferent input. This distinction could reflect the separation between a
parameniscal pathway i.e., via thalamic POm, conveying reafferent signals, and a lemniscal
pathway, i.e., via thalamic VPM, conveying exafferent input (Pierret et al., 2000; Urbain et al.,
2015). Accordingly, we modeled exafferent input to cortical neurons by using a stereotypical
pulse upon each whisker contact and brief deflection, and reafferent input proportional to
whisker angle reflecting motor efference (Szwed et al., 2003). Regardless of how the
properties of reaffererent input - whisking phase, absolute position, or their temporal
58
derivatives - are encoded by the pathway they do not change the main conclusion of our
model as long as the CLE feedback constitutes net negative feedback.
In agreement with the anatomy, we assume that cortically generated motor signals modulate
whisking behavior by acting on the facial nucleus (FN) (Ahissar, 2010). Because whisking
behavior persists after sensory denervation (Welker, 1964), cortical ablation (Semba and
Komisaruk, 1984), or decerebration (Lovick, 1972), we explicitly modeled a central pattern
generator (CPG) that autonomously generates whisking patterns locating exogenous to the
cortical-whisker loop (Hill et al., 2011). Thus, the FN receives input from both the cortical
population and CPG and moves the whisker in the reafference model.
We cannot rule out other biological pathways for negative CLE feedback. For example,
negative feedback can also be mediated by dominant cortical inhibition to the modeled
population of cortical neurons (Supplementary Figure 6B). In agreement, thalamocortical
connections strongly innervate fast spiking neurons and consequently implement strong
feedforward inhibition to the cortex (Bruno and Simons 2002). Other potential models arise
from heterogeneity in cortical populations. For example, negative CLE feedback can be
mediated by neurons in the barrel cortex that directly drive whisker retraction with extremely
short latencies (Matyas et al., 2010).
Alternatively, the dominant negative feedback loop could be subcortcial, Supplementary
Figure 6C. Here the dynamics of the cortex only indirectly reflects the stabilization of the
thalamus. This scheme is also consistent with reduced thalamic activity during whisking
(Poulet et al., 2012). It is important to note that this implementation also fundamentally relies
on stabilisation of neuronal activity by negative CLE feedback.
59
At the level of the whole vibrissa system, there are likely multiple parallel and nested
feedback loops, both positive and negative (Ahissar et al., 2003). However, we assume that
the overall or net feedback mediated by the cortical-whisker circuit during corresponding
behavior is negative in sign which we empirically demonstrate the presence of negative CLE
feedback in zebrafish active sensing. These models are experimentally testable. For
example, a group of neurons that encode aspects of reafferent input, such as whisker
protraction, could be genetically labeled and used for anatomical tracing studies. We can
moreover study the physiological role of these neurons by optogenetically silencing them
during active whisking and study how brain state as well as the animal's behavior may be
altered. This specific neuronal population could also be optogenetically activated and the
ensuing behaviors and changes in vibrissae information processing pathways studied.
60
|
1304.5674 | 3 | 1304 | 2013-07-26T21:28:03 | Generalized cable theory for neurons in complex and heterogeneous media | [
"q-bio.NC"
] | Cable theory has been developed over the last decades, usually assuming that the extracellular space around membranes is a perfect resistor. However, extracellular media may display more complex electrical properties due to various phenomena, such as polarization, ionic diffusion or capacitive effects, but their impact on cable properties is not known. In this paper, we generalize cable theory for membranes embedded in arbitrarily complex extracellular media. We outline the generalized cable equations, then consider specific cases. The simplest case is a resistive medium, in which case the equations recover the traditional cable equations. We show that for more complex media, for example in the presence of ionic diffusion, the impact on cable properties such as voltage attenuation can be significant. We illustrate this numerically always by comparing the generalized cable to the traditional cable. We conclude that the nature of intracellular and extracellular media may have a strong influence on cable filtering as well as on the passive integrative properties of neurons. | q-bio.NC | q-bio |
Generalized cable theory for neurons in complex and heterogeneous media
Claude Bedard and Alain Destexhe
Unit´e de Neuroscience, Information et Complexit´e (UNIC),
CNRS, Gif-sur-Yvette, France
(Dated: October 30, 2018)
Cable theory has been developed over the last decades, usually assuming that the extracellular space around
membranes is a perfect resistor. However, extracellular media may display more complex electrical properties
due to various phenomena, such as polarization, ionic diffusion or capacitive effects, but their impact on cable
properties is not known.
In this paper, we generalize cable theory for membranes embedded in arbitrarily
complex extracellular media. We outline the generalized cable equations, then consider specific cases. The
simplest case is a resistive medium, in which case the equations recover the traditional cable equations. We show
that for more complex media, for example in the presence of ionic diffusion, the impact on cable properties such
as voltage attenuation can be significant. We illustrate this numerically always by comparing the generalized
cable to the traditional cable. We conclude that the nature of intracellular and extracellular media may have a
strong influence on cable filtering as well as on the passive integrative properties of neurons.
PACS numbers:
I.
INTRODUCTION
Cable theory, initially developed by Rall [1], is one of the most significant contributions of theoretical neuroscience
and has been extremely useful to explain a large range of phenomena (reviewed in [2]). However, cable theory makes a
number of assumptions, one of which is that the extracellular space around neurons can be modeled by a resistance, or
in other words, that the medium around neurons is resistive or ohmic. While some measurements seem to confirm this
assumption [3], other measurements revealed a marked frequency dependence of the extracellular resistivity [4, 5], which
indicates that the medium is non-resistive. Indirect measurements of the extracellular impedance also show evidence for
deviations from resistivity [6 -- 9], which could be explained by the influence of ionic diffusion [10]. Despite such evidence
for non-resistive media, the possible impact on cable properties has not been evaluated.
The effect of non-resistive media can be investigated by integrating this effect in the impedance of the extracellular
medium, Ze, and in particular, through its frequency dependence. For example, it can be shown that Ze ∼ 1/ω for
capacitive effects or electric polarization [11], Ze ∼ 1/ √ω for ionic diffusion (also called the "Warburg impedance" [12]),
while Ze would be constant for a perfectly resistive medium. To integrate such effects in a given formalism, such as the
genesis of extracellular potentials, our approach has been to integrate a general frequency-dependent function Ze(ω) in the
formalism, and then consider specific cases [10, 12].
In the present paper, we follow this approach and generalize cable equations for media with arbitrarily complex
frequency-dependent impedance. With numerical simulations, we consider specific cases such as resistive media, ionic
diffusion, capacitive media, etc. We evaluate a number of possible consequences on the variation of the membrane poten-
tial along the cable, and how such effects could be measured experimentally.
II. METHODS
All simulations were done using MATLAB. To simulate the cable structure of the models, a classic compartmental
model strategy was used for simulations (see Fig. 3F), but was different from the one used in common simulator programs
such as NEURON (Hines and Carnevale, [13]). Each cylindric compartment is connected to intracellular and extracellular
resistances or impedances, and these are normally used to solve the cable equations. In the present paper, we used another,
where Vm and ii are respectively
equivalent method which consists of defining an auxiliary impedance, given by Za =
Vm
ii
Bedard&Destexhe(October30,2018)
2
the transmembrane potential and the axial current per unit length at the point where Za is connected (see Fig. 1). This
auxiliary impedance allows to take into account the influence of other compartments, including the soma, over the axial
current and transmembrane potential. It is mathematically equivalent to consider the continuity conditions on axial current
and transmembrane potential.
The electric and geometric parameters are considered constant in each compartment, but are allowed to vary between
compartments. In these conditions, Vm and ii are solution of partial differential equations (cable equations) and thus
depend on spatial coordinates.
FIG. 1: (Color online) Convention used to calculate the input impedance and transfer function. A cable segment of length l is repre-
sented, with an impedance Za in series, at the end of the cable. This "auxiliary impedance" Za takes into account the influence of the
other compartments on the axial current ii and transmembrane potential Vm in a compact form. Za = Vm(l)
ii(l) where ii(l) is the current per
unit length and Vm(l) is the transmembrane voltage at coordinate x = l.
The cable equations simulated in this article are generalized to allow one to include media with complex electrical
properties. We have designed a MATLAB code that simulates such generalized cable structures, using different types of
linear density of complex impedances ([Ω/m]) and specific impedances ([Ω.m]) in each compartment. See Results for
details of this method.
All computations were made in Fourier space. We have applied the theory to four different types of media to evidence
their effect on the spatial and frequency profile of the membrane potential. These models are called SC, FC, FO and NIC,
respectively (see Table I). The SC model is the "standard model" as defined by Tuckwell [14]; the FC model corresponds
to a model similar to the standard model (based on a closed circuit), but the cytoplasm and extracellular media impedances
can be frequency-dependent. The FO type model is the same, with an open circuit (no return current). The NIC model
includes a non-ideal capacitance similar to a previous study [15]. See Results for details of these models.
All numerical simulations were made using a "continuous ball-and-stick" model, consisting of a single cylindric com-
partment, described as a continuum (see Results), and a spherical soma. The dendritic compartment has a radius of 2 µm
and a membrane time constant of 5 ms, which corresponds to typical values of in vivo conditions. The has a radius of
7.5 nm and the specific capacitance was of 0.01 F/m2. These parameters represent typical values used in a number of
previous studies [2, 14, 16, 17].
III. RESULTS
We start by generalizing the cable equations for membranes embedded within extracellular media of arbitrarily complex
electrical properties. Next, we consider a few specific cases and numerical simulations.
A. Generalized cable equations
In this section, we redefine the cable equations taking into account the presence of complex and/or heterogeneous prop-
erties of extracellular and intracellular media. Because electrically complex or heterogeneous media can display charge
accumulation, one cannot apply the usual (free-charge) current conservation law. One needs to use a more general con-
Bedard&Destexhe(October30,2018)
3
T ypes
Model
z(m)
e
SC
Standard cable
(closed-circuit)
z(m)
e = −
rmre
(ri+re)(1+iωτm)
FC Frequency-dependent cable
(closed-circuit)
z(m)
e = −
rmze
(zi+ze)(1+iωτm)
λ2
rm
ri+re
rm
zi+ze
λ = 1+iωτm
κ2
λ2
(ri+re)(1+iωτm)
rm
(zi+ze)(1+iωτm)
rm
FO Frequency-dependent cable
z(m)
e
(open-circuit)
rm
zi
[1 +
z(m)
e
rm
(1 + iωτm)]
zi(1+iωτm)
z(m)
e
rm (1+iωτm)]
rm[1+
NIC
Non-ideal cable
(closed-circuit)
z(m)
e = −
ω2rm τmτM
[1+iω(τm+τM)][1+iωτm]
rm
zi
[ (1+iωτm)(1+iωτM)
1+iω(τm+τM)
]
zi
rm
[1 + i ωτm
1+iωτM
]
TABLE I: Summary of dendritic cable types and parameters. The table gives the parameters z(m)
, λ2 and κλ for different model types.
e
The standard model (SC) is the cable model as given by Rall, Koch and Tuckwell [1, 14, 17]. The "frequency-dependent model"
(FC) correspond to a standard cable (closed circuit), but where the parameters zi and ze are allowed to be frequency dependent. In
the "frequency-dependent open-circuit model" (FO), the current in the extracellular medium is "perpendicular" to the membrane (see
Fig. 6). The "non-ideal cable" model (NIC) is similar to the standard model, but the capacitance of the membrane is non-ideal, as
developed previously [15]. zi (see Eq. 10) and ze are respectively the impedance per unit length of the cytoplasm and of the extracellular
medium, respectively, for FC type models. We write ri and re when the parameters zs do not depend on frequency (SC type model).
The parameter z(m)
e
(see Eq. 18) is used in FO type models.
servation law based on the generalized current. In Section III A 1 below, we derive this generalized current conservation
law, while in Section III A 2, we use this generalized conservation law to derive the generalized cable equations.
1. Generalized current conservation law in heterogeneous media
In this section, central to our theory, we show that the free-charge current conservation law (~j f ) does not apply to
systems with complex electrical properties. Another, more general, conservation law must be used, the generalized
current conservation law. We derive here the conservation law for the membrane current in arbitrarily complex media,
starting from first principles.
Maxwell theory of electromagnetism postulates that the following relation is always valid for any medium:
∇ × ~H = ~j f +
∂ ~D
∂t
,
(1)
where ~H is the magnetic field, and ~j f is the current density of free charges, and ∂ ~D
∂t is the displacement current density.
We define the generalized current density ~j g as:
~j g = ~j f +
∂ ~D
∂t
= ~j f + ~j d ,
(2)
where ~j d is the displacement current density.
It is important to note that the term ∂ ~D
field varies in time).
∂t = εo
∂ ~E
∂t
is different from zero, even in the vacuum (assuming that the electric
Bedard&Destexhe(October30,2018)
4
The interest of using the generalized current, is that it is always conserved in any given volume, for any type of medium,
as we explain below (see also Appendix IV).
In the case of an electric field in a homogeneous and locally neutral medium, we have ∇ · ~j f = − ∂ρ
= 0 because
there cannot be charge accumulation anywhere. Because the relation ∇ · ~j g = 0 applies to any type of medium, we also
have ∇ · ( ∂ ~D
∂t ) = 0. Thus, in a homogeneous locally-neutral medium, we have two independent current conservation laws:
one law applies to the free-charge current i f and another one applies to the displacement current i d . Note that in a
homogeneous medium i d is not necessarily negligible, but the application of the current conservation law on i f can be
done independently of the existence of i d because the two laws are independent.
f
∂t
This is the framework assumed in the standard cable theory, in which the extracellular medium is resistive and homo-
geneous, the displacement current i d is negligible, and there cannot be charge accumulation inside the dendrites nor in
the extracellular medium. We will see below that these assumptions do not hold for complex extracellular media. If the
medium is heterogeneous, then charge accumulation will necessarily appear in the presence of an applied electric field.
Capacitive effects is an example of such charge accumulation. In such a case, the two current conservation laws on i f and
i d do not apply to every region of space (see Appendix B). However, the generalized current conservation on i g is still
valid in all cases.
Thus, to derive cable equations in heterogeneous media, one must use the generalized current conservation law, as done
in the next section.
2. Application of the generalized conservation law to cable equations
To start, we consider a small portion of membrane surface and build a domain in the intracellular side, which is limited
by the interior surface of the membrane, while the other surfaces of the domain are located inside the cytoplasm (see
Fig. 2).
FIG. 2: (Color online) Definition of a domain D inside the cytoplasm and adjacent to the membrane. Due to conductance variations
in the membrane, or due to charged currents, the total charge in domain D varies. One cannot consider that the displacement current
across the surface of domain D is zero, because this would be in contradiction with Maxwell-Gauss law (see Appendix B). Black
circles represent negative charges on the interior surface of the membrane, as well as in the cytoplasm, while blue circles indicate
positive charges at the exterior side of the membrane.
Bedard&Destexhe(October30,2018)
5
Using such a definition, in resting conditions, the intracellular side has an excess of negative charges, which are adjacent
to the membrane. In such a state, we can calculate the free charge density in this domain from Maxwell-Gauss law:
~D · n dS = cst ,
Q(t) = (cid:9)
∂D
where ∂D is the surface of the considered domain[34]
Now, suppose that a conductance variation occurs in the domain (for example following the opening of an ion channel).
This will induce a charged current in domain D and therefore, there will be a variation of the total charge included within
domain D, which implies a non-zero displacement current i d across the surface ∂D surrounding domain D (without this
current, the system would be in contradiction with Maxwell-Gauss law; see Appendix B).
(3)
In such conditions, we have:
i d =
dQ
dt
= (cid:9)
∂D
∂ ~D
∂t
· n dS , 0
(4)
Can we neglect this current to study the variations of the membrane potential along the cable? Because it is difficult to
give a rigorous answer to this question [18, 19], in particular when i d is non-zero, we consider the generalized current i g
because this current is conserved independently of i d (see previous section). This will allow us to treat cable equations
without making any hypothesis about charge accumulation inside or outside of the cable.
Moreover, to stay as general as possible, we include a frequency and space dependence of the electric parameters,
which will allow us to simulate the effect of media of different electric properties, such as capacitive or diffusive [10 -- 12].
In this context, the linking equations must be expressed in their most general form [10]:
−∞
~D(~x, t) = R +∞
~j f (~x, t) = R +∞
−∞
εi(~x, t − τ) ~E(x, τ) dτ
[σe
i (~x, t − τ) ~E(~x, τ)
According to this scheme, the generalized current density ~j g
i
inside the cytoplasm obeys:
i (~x, t) = Z +∞
~j g
−∞
[σe
i (~x, t − τ) ~E(~x, τ) + εi(~x, t − τ)
∂ ~E
∂t
(~x, τ)] dτ
(5)
(6)
where σe
i (~x, t) is the intracellular electric conductivity function and εi(~x, t) is the intracellular electric permittivity function.
The first term in the integral accounts for energy dissipation phenomena, such as calorific dissipation (Ohm's differential
law) and diffusion phenomena. The second term represents the effect of charge density variations in the volume elements.
In Fourier frequency space, Eq. 6 becomes algebraic.
g
i (~x, ω) = [σe
~
i (~x, ω) + iωεi(~x, ω)] ~E(~x, ω)
(7)
Moreover, we have ∇ × ~E = 0, which implies ~E = −∇V because electromagnetic induction is negligible in biological
tissue (in the absence of magnetic stimulation[35]).
If we now consider a one-dimensional cylindric cable of constant radius a (Fig. 3A), the generalized current at a position
x of the cable can be written as:
i g
i (x, ω) = ~j g
i (x, ω) · (πa2 n) = −πa2[σe
i (x, ω) + iωεi(x, ω)] ·
∂Vi
∂x
(x, ω)
(8)
where Vi is the intracellular voltage difference with respect to a given reference (which can be far away). In the following
of the text, we will call "compartment" a cylindric cable with constant radius and with uniform electric parameters (see
Fig. 1). It is important to note that this compartment does not need to be isopotential, and the membrane potential will
depend on the position on the compartment (see scheme in Fig. 1).
Bedard&Destexhe(October30,2018)
6
FIG. 3: (Color online) Compartments and equivalent electrical circuits of the membrane and cable segments. A and C depict different
configurations in a cable of constant diameter, with their respective equivalent electrical circuits shown in B and D. E is the equivalent
electrical circuit of a membrane compartment of the cable, and F is the equivalent circuit obtained for three compartments. Vi is
the intracellular potential relative to the reference, Ve is the extracellular potential relative to the same reference, zi is the cytoplasm
impedance, rm/dx and z(m)
e /dx are respectively the impedances of ion channels and the input impedance of the extracellular medium as
seen by the transmembrane current. ie is the output current of a cable element in the extracellular medium, and ii is the axial current.
The membrane potential Vm j equals Vi j − Ve j and may vary according to the position x j.
If we assume that the impedance (per unit length) of cytoplasm zi can be expressed as:
then the axial current can be written as:
zi =
1
πa2[σe
i (x, ω) + iωεi(x, ω)]
,
i g
i (x, ω) = −
1
zi
∂V i
∂x
(x, ω)
(9)
(10)
This expression is similar to the traditional cable equation [2, 14, 16], with the exception that the parameter zi is complex
(with units of [Ω/m])[36]. In addition, the transmembrane current i⊥m over a cable length dx can be expressed as:
i⊥m(x, t) = im (x, t) dx = 2πadx [ Cm
∂Vm(x, t)
∂t
+
σe
m
e
(Vm(x, t) − Em) ] = dx [ cm
∂Vm(x, t)
∂t
+
(Vm(x, t) − Em)
rm
]
(11)
where Vm is the transmembrane voltage, Em is the resting membrane potential, Cm is the specific membrane capacitance
(in F/m2), cm is the membrane capacitance per unit length (in F/m), σe
m is the electric conductivity (in S /m), 1/rm is the
Bedard&Destexhe(October30,2018)
7
linear density of membrane conductance (in S /m), e is the membrane thickness (in m) and im is the transmembrane current
per unit length (in A/m) (Figs 3C-E). Applying the inverse Fourier transform, we obtain:
i⊥m(x, 0) = im (x, 0) dx = 1
rm
[Vm(x, 0) − 2πEmδ(0)] dx ω = 0
i⊥m(x, ω) =
im (x, ω) dx = [iωcm + 1
rm
]Vm(x, ω) dx
ω , 0
(12)
Note that we assume here that the resting membrane potential Em does not depend on time nor on position in the cable.
Thus, we can see that the Fourier transform of Eq 10 generates a Dirac delta function for null frequency. In the following
of the text, we consider frequencies different from zero, because the zero-frequency component of i⊥m is zero for a signal
of finite duration, which is always the case in reality.
In the model above, the expression of the transmembrane current is identical to the generalized membrane current for
frequencies different from zero. In this case, the generalized current is given by:
i g
m = A · j g
m = −2πa dx(σe
m + iωεm)∇V = 2πa dx(σe
m + iωεm)
Vm
l
= dx (
1
rm
+ iωcm)Vm = dx im = i⊥m
(13)
where l is the membrane thickness and A is the membrane surface.
Assuming that the charge variations inside the channels is negligible, then the generalized current conservation law can
apply to point B in the equivalent scheme (see Fig. 3 C), and we can write
i g
i (x + dx, ω) = i g
i (x, ω) − i⊥m(x, ω)
i (x, ω) − i g
m (x, ω) = i g
It follows that:
Using Eqs. 6 and 11, we obtain:
di g
i (x, ω) =
∂i g
i
∂x
dx = −i⊥m(x, ω) = −im (x, ω) dx
πa2 ∂
∂x
[(σe
i (x, ω) + iωεi(x, ω))
∂Vi
∂x
(x, ω)] = [iωcm +
1
rm
]Vm(x, ω)
Applying the partial derivative on the lefthand term, and dividing by πa2(σe
i + iωεi), one obtains:
∂2Vi
∂x2 +
1
(σe
i + iωεi)
∂(σe
i + iωεi)
∂x
∂Vi
∂x
·
=
1
πa2(σe
i + iωεi)
[iωcm +
1
rm
]Vm = ziim
(14)
(15)
(16)
Note that if the righthand term was zero, then this equation would be identical to the equation describing the electric
∂x in one dimension, in which case the right
potential outside of the sources [10, 12, 20], because the ∇ operator equals ex
would be equal to ∇2Vi + ∇γi
· ∇Vi where γi = σe
i + iωεi.
γi
∂
We can simplify Eq. 14 if the cytoplasm is quasi-homogeneous (assuming the scale considered is large compared to
inhomogeneities due to subcellular organelles), in which case we can consider that the electric parameters of the cytoplasm
are independent of position x: σe
i (ω) and εi(x, ω) = εi(ω). This leads to the following expression:
i (x, ω) = σe
1
zi
∂2V i
∂x2 (x, ω) =
1
πa2(σe
i + iωεi)
∂2V i
∂x2 (x, ω) = [iωcm +
1
rm
]Vm(x, ω)
(17)
in Fourier space.
If we now assume that the extracellular medium can also be considered as homogeneous (which will be valid at scales
larger than the typical size of the cellular elements), then we can model the variations of the membrane potential caused
by the transmembrane current i⊥m. We can model this effect from the notion of impedance, without making any hypothesis
on the current field in the extracellular medium. In this case, one can associate to each cable segment dx the specific
impedance of the extracellular medium, z(m)
, as seen by the transmembrane current. z(m)
has a similar physical meaning
e
e
as rm, except that it is a complex number in general. In Section III B, we will see that z(m)
e depends on the direction of the
current field in the extracellular medium.
Bedard&Destexhe(October30,2018)
Without any loss of generality, we can write in Fourier space:
Vi(x, ω) = Vm(x, ω) + z(m)
e
(ω)im(x, ω) .
By substituting this last expression in Eq. 15, we obtain
rm
zi
[1 +
z(m)
e
rm
(1 + iωτm)]
∂2Vm
∂x2 = [1 + iωτm]Vm
where τm = rmcm.
Thus, we can write the system in a form similar to the standard cable equation:
where
8
(18)
(19)
(20)
λ2 ∂2Vm(x, ω)
∂x2
= κ2Vm(x, ω)
λ2 = rm
zi · [1 +
z(m)
e
rm
(1 + iωτm)] = rm
¯zi
κ2 =
1 + iωτm
for a cylindric compartment (see Eq. C8 in Appendix C). It follows that the general solution of this equation in Fourier
space ω , 0 is given by:
Vm(x, ω) = A+(ω)e κ(l−x)
λ + A−(ω)e −κ(l−x)
λ
(21)
for each cylindric compartment of length l and with constant diameter (see Fig. 1 for a definition of coordinates). For a
given frequency, we have a second order differential equation with constant coefficients.
In general, one can apply Eq. 21 for different cylindric compartments, as in Fig. 3F. In this case, one must adjust the
∂Vm
∂x (see Eq. C4 in
different compartments to their specific limit conditions (continuity of Vm and of the current i g
Appendix C).
i = − 1
¯zi
Note that Eq. 21 is exact for a cylindric compartment of constant diameter. Thus, it is possible to use this property to
simulate exactly the full cylindric compartment as a continuum with no need of spatial discretization into segments, as
usually done in numerical simulators. This is only possible if the cylindric compartment has a constant diameter. This
leads to an efficient method to simulate the cable equations. We will refer to this approach as "continuous compartment"
in the following.
As mentioned above, the mathematical forms of Eqs. 19 and 21 are identical to that of the standard cable model, but
with different definitions of λ. Thus, we directly see that the nature of the extracellular medium will change the value of
these parameters, which become frequency dependent. In particular, we see from Eq. 19 that changing these parameters
will impact on the spatial profile of the variations of Vm, if the frequency dependence of the ratio κλ = κ
λ is affected by the
nature of the medium. Thus, experimental measurement of the spatial variations of Vm will be able to identify effects of
the extracellular impedance only if the ratio κλ is affected.
In the next section, we derive expressions to calculate the input impedance Zin(P) =
Vm(P,ω)
ii(P,ω) and the transfer function of
Vm(Pb,ω)
Vm(Pa,ω) between two positions in the cable (as a function of the ratio κλ). Later in
the transmembrane voltage FT (ω) =
Section III B, we will see that it is necessary to know these quantities to calculate the spatial variation of Vm and compare
the standard model with the cable model embedded into complex extracellular media.
3. Method to solve the generalized cable
In this section, we present the theoretical expressions which will allow us to calculate the input impedances needed for
computing the membrane voltage on a cable with varying diameter. We consider the input impedance of the membrane,
Bedard&Destexhe(October30,2018)
9
as well as the impedance of the extracellular medium, both of which are needed to calculate the spatial profile of the Vm
in a given cable segment.
We proceed according to the following steps:
1. In the previous section, we saw that it is necessary to calculate the ratio Zin(P) =
source, to calculate the Vm produced at that point.
Vm(P,ω)
ii(P,ω) at the position of the current
FIG. 4: (Color online) Branching cables. The panels A and B respectively represent a branched cable where a dendrite separates into
two daughter branches, and its equivalent electrical circuit. The equivalent impedance of segment 1 is equal to the input impedances of
segments 2 and 3 (zout 2 and zout 3 ) taken in parallel.
One strategy is, in a first step, to separate the cable into a series of continuous compartments of constant diameter,
where parameters a (Eq. 8), zi (Eq. 10), rm (Eq. 11) and z(m)
(Eq 18) are constant and specific to each compartment. In a
e
Vm(0)
second step, one calculates the (transmembrane) input impedance Zn+1
ii(0) at the begin of each compartment by taking
into account the auxiliary impedance at the end of this compartment, Za = Zn+1
in (see Fig. 1) if there is no
branching point. At the branching points, the auxiliary impedances are simply equal to the equivalent input impedance
of n dendritic branches in parallel (where n is the number of "daughter" branches; see Fig. 4). Thus, because the input
impedance at one end is equal to the input impedance of the other compartment connected to this end, one obtains a
recursive relation (see Eq. C9 in Appendix C):
Vm(ln+1)
ii(ln+1) = Zn
in =
out =
Zn+1
in
[Zn
in] =
¯zin
κλn
(κλnZn
(κλnZn
in + ¯zin) e2κλn ln + (κλnZn
in + ¯zin) e2κλn ln − (κλnZn
in − ¯zin)
in − ¯zin)
(22)
where
Thus, we can write
¯zi =
1 +
z(m)
e
rm
zi
(1 + iωτm)
Zn+1
in = F [Zn
in ; ¯zin , κλn , ln]
This leads to the following expression to relate the first to the nth segment:
in = F [...F [F [Z1
Zn+1
in ; ¯zi1 , κλ1 , l1]; ¯zi2 , κλ2 , l2]...; ¯zin , κλn, ln]
(23)
Note that this algorithm is a generalization of that used to calculate the equivalent resistance for resistances in series.
Indeed, for resistance in series we have req = F(...F(r1; r2); rn) where F(ra; rb) = ra + rb. The difference between this
recurrence function and that of Eq. 25 essentially comes from the fact that there is no current leak in a resistance, while
there is one in a dendritic compartment.
2. To calculate the profile of Vm along the cable, one must use the spatial transfer function Vm(Pn+1,ω)
Vm(Pn,ω) on a continuous
cylindric compartment of arbitrary length, and calculate the product of the transfer functions between each connected
compartment. This leads to (see Appendix D and Eq. D3):
FT (l, ω; Zn
out) =
κλZn
out
κλZn
out cosh(κλl) + ¯zisinh(κλl)
(24)
Bedard&Destexhe(October30,2018)
Vm(Pn, ω)
Vm(P1, ω)
=
n−1
Yi=1
Vm(Pi+1, ω)
Vm(Pi, ω)
10
(25)
3. To evaluate zproximal we must calculate the first impedance Z1
in which enters the recursive relation 24. This impedance
corresponds to the impedance of the soma, which is given by:
Z1
in = Zs + Zcs,
(26)
where Zs is the soma membrane impedance and Zcs is the cytoplasm impedance inside the soma. This relation is obtained
under the hypothesis that the soma is isopotential, and the application of the generalized current conservation law implies
where Vi and Ve are the electric potentials at both sides of the membrane, inside and outside,
i g =
respectively relative to a reference located far-away.
Zs+Zcs ≈ Vm
Vi−Ve
Zs+Zcs
The impedance of the bilipidic membrane is approximated by a parallel RC circuit where R = Rm is the resistance and
τm = RmCm is the membrane time constant. Thus, Z1
in can be written as:
Z1
in = Zs + Zcs =
Rm
1 + iωτm
+ Zcs
(27)
Finally, to evaluate zdistal, we use the "sealed end" boundary condition Z1
coth(κλ1l1) (see Eq. 22). In the case of a single dendritic branch, we can write:
in = ∞. In this condition, we have Z2
in =
¯zi1
κλ1
Zdistal
in
=
¯zi
κλ
coth(κλl) ,
(28)
where l is the total length of the cable.
In the next section, we turn to numerical simulations to investigate passive cable properties in the presence of complex
media. We consider the most general case, where both the impedance of the extracellular medium and that of cytoplasm
can be frequency dependent, and determine the respective impact on the spatial profile and frequency content of the
transmembrane voltage at the level of the proximal and distal ends of the cable.
B. Numerical simulations
The goal of the numerical simulations is here to show how the physical nature of extracellular and intracellular media
can influence the spatial and frequency profiles of the transmembrane potential. We present simulations of a "continuous
ball and stick" model, which consists of a continuous cylindric compartment (described by Eq. 21), connected to a
spherical soma. In this case, the impedance Za of the continuous cylindric compartment is the soma impedance (see
Fig. 1). We do not investigate here the effect of complex dendritic structures, which is left for future studies. Note
that what we call a "continuous cylindric compartment" actually represents an infinite number of compartments each
represented by a resistance in series with a parallel RC circuit (see Fig. 3F).
In a first step, we list the different types of models of intracellular and extracellular media that were used. In a second
step, we present the results of numerical simulations.
1. Different types of cable models
We now explain the parameters used for the simulations of the cable presented in Section III C.
Because the cable equation (Eq. 19) is completely determined by the value of κλ for a given frequency, the spatial and
frequency profiles of the transmembrane voltage are completely determined if the geometry and boundary conditions are
Bedard&Destexhe(October30,2018)
set. And because κλ is a function of 4 parameters (rm, τm, zi, z(m)
)(Eq. 20) for a given frequency, we have a four-dimensional
e
parameter space where the two last parameters (zi, z(m)
) can be frequency dependent. We will limit our exploration of this
e
parameter space by only varying the physical nature of these impedances for realistic values of rm and τm, because the
influence of these parameters has been largely characterized in previous studies [1, 16, 17]. Furthermore, with τm and ω
fixed, the relation κλ =
depends only on λ, and thus, like the classic studies on cable equations, we will use this
parameter as a main determinant of the cable properties.
1+iωτm
11
λ
We will explore the generalized cable equations by considering several typical cases:
Closed−circuit model
Open−circuit model
iR
Re
iz
ze
ze
FIG. 5: (Color online) Two different cable models for neurons. Left: Closed-circuit model. This is the standard cable model which
forms a closed system (all inward and outward currents are balanced) and can be described by an equivalent circuit (bottom; shown
here for a two-compartment model; Re and Ri are the extracellular and intracellular resistances, respectively). In this model, the current
flows parallel to the neuron. Right: Open-circuit model. In this more general model, the current is allowed to flow between neighboring
neurons, or between the neuron and extracellular space, with no necessary condition of local balance (top). In this case, the neuron is
modeled by an open circuit (bottom), and the current flows "perpendicular" to the membrane. The equivalent circuit is modeled more
generally with impedances (Ze extracellular, Zi intracellular)
Standard cable model The first type of model that we will consider is the "standard cable model" (model SC in
Table I), identical to that considered by Rall, Koch and Tuckwell [1, 14, 17]. In this model, the neuron is a closed system,
where the inward and outward currents are balanced, forming a closed circuit (see Fig. 6, left). The extracellular current
flows parallel to the dendrite, as noted previously [14]. This model is equivalent to consider that the field produced by the
neuron corresponds to an electric dipole configuration. In addition, this model considers that the extracellular medium is
resistive, or in other words, that the extracellular impedance is a constant.
In this standard model, the extracellular impedance z(m)
e
is either zero (no extracellular resistivity) as in Rall's and Koch's
Bedard&Destexhe(October30,2018)
12
formulations [1, 2, 17], or is equal to a constant, which is equivalent to model the extracellular medium by a resistance, as
in other formulations [14, 21]. Besides its physical non-sense (the extracellular medium considered as a supraconductor),
using a zero-resistance is usually justified from the fact that the extracellular resistivity is much smaller than the membrane
impedance. We will see that this justification does not hold if the medium is frequency dependent, in which case for some
frequency range the extracellular resistivity may be determinant. Thus, to obtain the general expression of λ and κλ for
the standard model, we set z(m)
(ri+re)(1+iωτm) in Eq. 20 (see Table I).
rmre
e = −
60
50
40
30
20
10
0
0
B
250
500
ν (Hz)
750
1,000
)
g
e
d
(
]
λ
κ
[
Φ
250
500
ν (Hz)
750
1,000
C
150
x 104
2.5
A
)
2
−
m
(
λ
κ
2
1.5
1
0.5
0
0
T
F
1
0.8
0.6
0.4
0.2
0
100
5
5
50
100
50
5
100
50
150 150
150
10−4
L (m)
10−3
FIG. 6: (Color online) Spatial and frequency profile of the membrane potential in the cable model with resistive media. A and B
respectively show the modulus κλ and the phase Φ[κλ] of κλ as a function of frequency ν for a continuous ball-and-stick model. C.
Modulus of the transfer function FT as a function of distance L in the dendritic compartment, for frequencies equal to 5, 50, 100 and
150 Hz (see corresponding frequencies in A and B). The blue curves in − · − correspond to a standard cable model (FC, closed-circuit),
with ri = 28 × 109 Ω/m and re = 18 × 109 Ω/m . The red curves correspond to the same model but in an open-circuit configuration (FO
model), with ri = 28 × 109 Ω/m and z(m)
e = 0.01 τm/2πaCm = 0.4 × 103 Ω.m. The black curves in −− show a non-ideal cable (NIC)
model with τM = 0.01τm, ri = 28 × 109 Ω/m and re = 0 Ω/m.
Frequency-dependent cable model The second type of model is an extension of the standard model, where the intra-
cellular and extracellular impedances (zi and z(m)
, respectively) are allowed to depend on frequency. This "frequency-
e
dependent cable model" (model FC in Table I) can account for example for a neuron embedded in capacitive or
diffusive[37] extracellular media, or if the intracellular medium has such properties, or both. In such cases, the appropriate
frequency-dependent profiles for the impedances must be used.
In this frequency-dependent model, if τm is fixed, the quantity ze + zi completely determines the spatial and frequency
profiles of the Vm, and how they deviate from the standard model (see Table I). To explore the effect of the impedances
zi + ze, we consider three typical cases: "resistive", "capacitive" (which is in fact resistive and capacitive in parallel) and
"diffusive" (which is equivalent to a Warburg type impedance). Such impedances have also been considered in previous
studies [8, 12, 22].
Bedard&Destexhe(October30,2018)
13
Note that, in order to simulate the standard model, one must necessarily assume that the real part of z(m)
e
is negative[38],
is not a passive impedance per unit length, but is active, and thus requires a source of energy, as
which implies that z(m)
e
pointed previously [23, 24]. This point will be further considered in the Discussion.
A
B
)
2
−
m
(
λ
κ
3000
2000
1000
0
T
F
1
0.9
0.8
0.7
0
)
g
e
d
(
]
λ
κ
[
Φ
200 400 600 800 1000
ν (Hz)
C
50
0
−50
−100
5
200 400 600 800 1000
ν (Hz)
50
5
150
100
5
50
100
150
50
100
150
1
2
3
L (m)
4
5
6
x 10−4
FIG. 7: (Color online) Spatial and frequency profiles of the membrane potential for a model with resistive extracellular medium and
diffusive cytoplasm. A and B: modulus κλ and phase Φ[κλ] of κλ as a function of frequency, for a continuous ball-and-stick model. C.
Modulus of the transfer function FT as a function of distance for different frequencies (same arrangement as Fig. 6). The red curves
correspond to a model with zero extracellular resistance. The blue curves (− · −) show models with open-circuit configuration (FO
model with z(m)
(1+i) √w Ω/m). The black curves (−−) show the same
model with closed-circuit configuration with a resistive extracellular medium (FC model re = 18 × 109 Ω/m). Note that for the FC
model, FT progressively increases from 5 to 50 Hz, then decreases between 50 and 100 Hz.
e = 0.5 τm/2πaCm = 20 × 103 Ω.m) and diffusive cytoplasm (zi = 28×109
Open-circuit model
In a third type of model, the "Open-circuit" model (FO in Table I), we use a different approach.
Instead of considering the neuron as a closed system, where all outward currents must return to the neuron, we make
no hypothesis about the return currents, and allow for example that neighboring neurons exchange currents[39]. In this
case, one does not need to describe each neuron by a closed circuit, but all neurons are open circuits are are connected
together (through the extracellular space). Figure 6 shows the current fluxes of the two models, the standard model is
a closed circuit where the outward currents loop into the inward currents (Fig. 6A), while in the open-circuit model, all
currents are exchanged with the surrounding medium (Fig. 6B). These two models correspond to different equivalent
circuits (Fig. 11 in Appendix E).
Note that the Open-circuit cable model is practically equivalent to the traditional (closed-circuit) cable model for an
isolated neuron, if the impedance of the extracellular medium is negligible compared to the membrane impedance. Indeed,
if z(m)
e
and ze tend to 0, then we have (see Table I):
λ2
FO =
lim
z(m)
e →0
rm
zi
λ2
FC
= lim
ze→0
(29)
Similar to the frequency-dependent cable model, we will consider the three types of impedances discussed above
(resistive, capacitive and diffusive) in the simulations of the Open-circuit model. In this case, we separately consider
the two quantities zi and z(m)
because these two parameters directly determine the value of λ in models of FO type (see
e
Table I). Note that in the Open-circuit model, the real part of z(m)
is always positive, so there is no need of any additional
e
energy source (see Discussion).
Non-ideal cable model The fourth type of model considered here is the "non-ideal cable model" introduced previ-
ously [15]. This model postulated that the membrane capacitance is non-ideal, through the use of an additional resistance
Bedard&Destexhe(October30,2018)
14
)
2
−
m
(
λ
κ
2
1.5
1
0.5
0
x 104
A
)
g
e
d
(
]
λ
κ
[
Φ
200
400
600
ν (Hz)
800 1000
C
60
40
20
0
−20
B
200
400
600
ν (Hz)
800 1000
T
F
10−0.1
10−0.9
0
1
2
150
4
3
L (m)
5
50
100
5
6
x 10−4
FIG. 8: (Color online) Spatial and frequency profiles for a model with resistive cytoplasm and diffusive extracellular medium. Same
arrangement of panels as for Figs. 6 and 7, but for different media. The black curves (−−) show the behavior of a closed-circuit (FC)
type model with resistive cytoplasm (ri = 28 × 109 Ω/m) and diffusive extracellular space with Warburg impedance (ze = 18×109
(1+i) √w Ω/m).
The red curves correspond to a closed-circuit (FC) type model with zi = 28 × 109 Ω/m and ze = 0 Ω/m. The blue curves (− · −)
correspond to an open-circuit (FO) type model (ri = 28 × 109 Ω/m, z(m)
(1+i) √w = 20×103
(1+i) √w Ω.m).
e = τm
2πaCm
0.5
A
B
)
2
−
m
(
λ
κ
3000
2000
1000
20
0
T
F
1
0.9
0.8
0.7
0
)
g
e
d
(
]
λ
κ
[
Φ
200 400 600 800 1000
ν (Hz)
C
1
2
3
L (m)
40
20
0
−20
−40
200 400 600 800 1000
ν (Hz)
100
150
50
5
50
100
150
5
5
4
6
x 10−4
FIG. 9: (Color online) Spatial and frequency profiles for fully diffusive cable models. Same arrangement of panels as for Figs. 6 -- 8,
but using a continuous ball-and-stick model where both cytoplasmic and extracellular impedances are of diffusive (Warburg) type.
The black curves (−−) correspond to a closed-circuit (FC) type model with zi = 28×109
(1+i) √w Ω/m. The red curves
correspond to a closed-circuit (FC) type model with zi = 28×109
(1+i) √w Ω/m and ze = 0 Ω/m. The blue curves (− · −) correspond to a closed-
circuit (FO) type model with zi = 28×109
e = 20×103
(1+i) √w Ω.m. Note that for both types of models (FO and FC), FT increases
between 5 and 50 Hz, then decreases between 50 and 100 Hz.
(1+i) √w Ω/m and ze = 18×109
(1+i) √w Ω/m and z(m)
at the arms of the capacitor; this resistance models the fact that there is some inertia time to charge movement (or equiva-
lently, a friction). Such a non-ideal capacitance resulted in a shallower frequency scaling, that is a higher capacity of the
dendritic tree to propagate high-frequency events [15]. Note that in this model, the extracellular medium is modeled as a
Bedard&Destexhe(October30,2018)
15
resistance, so in this respect, the non-ideal cable model is equivalent to the standard model. Mathematically, the non-ideal
cable appears through the use of z(m)
(see Table I), which can therefore be viewed as a particular case of an influence of the
e
extracellular medium on cable properties. Indeed, the non-ideal cable can be shown to be equivalent to -- or a particular
case of -- the open-circuit model, where the Vm corresponds to Vi with a far-away reference (see Appendix E). We keep
this model here for comparison.
C. Simulation of the different models
In this section, we present the results of numerical simulations of the models presented in the previous section (see
Methods). The goal of these simulations is not to be exhaustive in considering all possible combinations of models, but
present a few typical configurations. The central question is whether the nature of the extracellular medium can have
determinant impact on cable properties, and for what type of configuration or parameter values does it happen ?
)
2
−
m
(
λ
κ
2000
1500
5
6
8
10
1000
20
40
500
5
6
8
10
20
40
20
2
2
3
4
3
4
)
g
e
d
(
]
λ
κ
[
Φ
40
60
ν (Hz)
80 100 120
25
20
15
10
5
0
−5
−10
−15
−20
−25
40
2
200
400
600
ν (Hz)
800 1000
FIG. 10: (Color online) Parameter κλ as a function of frequency for fully diffusive models. The black curves −− correspond to FO
and the red curves to FC type models with a time constant of τm = 2, 3, 4, 5, 6, 8, 10, 20 and 40 ms . The FC type model was with
zi = 28×109
(1+i) √w Ω.m (see Table II
for the corresponding resonance frequencies).
(1+i) √w Ω/m . For the FO type model, zi = 28×109
(1+i) √w Ω/m and ze = 18×109
(1+i) √w Ω/m and z(m)
e = τm
2πaCm
0.5
(1+i) √w = 20×103
Analysis of the spatial profiles of Vm variations
In this section, we investigate analytically and numerically different particular cases of extracellular and intracellular
media to determine how the nature of these media affects the spatial and frequency profile of the membrane potential. We
consider the transfer functions as defined in Table I. The analyses presented here are limited to a ball-and-stick model,
which allows a better interpretation of the effect of the physical nature of the different media. The effect of complex
dendritic tree morphology will be the subject of a future study. To compare the results from the different models, we have
considered models with identical geometry (see Methods for parameters).
Resistive models We first considered the "standard model" with resistive intracellular and extracellular media, as well
as the non-ideal cable model [15]. In Figure 7, we can see that the nature of the cable model (closed-circuit or open-
circuit; non-ideal) influences the modulus and the phase of κλ, as well as the spatial profile of the transfer function FT.
The modulus of the transfer function depends more strongly on frequency in the FC model compared to the two other
cases (Fig.7C), as observed previously [15]. Note that the parameters of the FO and NIC models were chosen such that
they are equivalent (see Appendix B).
Capacitive models Next, we considered models where the cytoplasm and extracellular medium are both of capacitive
(RC-circuit) type. Note that we considered capacitive effects without ionic diffusion, because if both are combined, the
Bedard&Destexhe(October30,2018)
16
resulting impedance is of Warburg type. This type of model will be considered next. With purely capacitive media,
we observed effects very similar to the resistive model shown in Fig. 7, with slight differences only visible for large
frequencies (greater than about 200 Hz (not shown). The small dimension of organelles (<< 1 µm2) within cells, as well
as the distance between neighboring cells (∼ 30 nm on average) [25, 26] imply that the capacitance values of the media
are necessary small compared to the membrane capacitance, and thus the purely capacitive effects (without diffusion) are
likely to be negligible.
τm (ms) νr (Hz)
2
3
4
5
6
8
10
20
83
54
40
30
25
20
18
8
TABLE II: Resonance frequencies of fully diffusive models for different membrane time constants. The resonance frequencies of κλ
as a function of the membrane time constant τm (see Fig. 10).
Resistive models with diffusive cytoplasm We next considered models where the extracellular medium was resistive
as above, but where the intracellular medium (cytoplasm) was diffusive, and described by a Warburg impedance. Figure 7
shows the spatial and frequency behavior of this model. We can see that the open-circuit (FO) model shows less attenuation
with distance compared to the closed-circuit (FC) model. Note that these two models give opposite variations when the
extracellular medium has a zero resistance: in FC type models, FT attenuates more steeply as a function of distance
when the extracellular impedance increases, whereas in FO type models, the attenuation becomes less steep. However,
the spatial profile of FT also attenuates less with a diffusive cytoplasm compared to a resistive cytoplasm. The latter
result is expected, because the higher the frequency the more the impedance "short-cuts" the membrane in this case. Note
that the Warburg impedance used in all diffusive models considered here was applied for frequencies larger than 5 Hz.
It is interesting to note that in the FC model, a resonance appears around 24 Hz in the modulus of the transfer function
κλ (Fig. 7A). In contrast, the FO model does not display a resonance.
Resistive cytoplasm with diffusive extracellular medium Next, we considered the opposite configuration as previously,
namely a resistive model for the cytoplasm, but a diffusive extracellular medium. Three sets of parameters were chosen for
the extracellular space. First, a FO type model with a resistive cytoplasm and a diffusive extracellular medium described
by a Warburg type impedance (black curve in Fig. 9), and second, a FC type model with similar parameters (blue curve in
Fig. 8). These two models can be justified if one takes into account the Debye layer at the edge of the membrane [8 -- 10]).
The case with a zero extracellular resistance (short-cut) is also shown for comparison (red curve in Fig. 9). The latter
model represents the same limit case for both FO and FC models, and therefore constitutes the frontier between the two
families of curves.
Fully diffusive cable models Next, we have considered the case where both intracellular and extracellular media are
diffusive. Figure 9 shows the frequency and spatial profiles of the Vm for such fully diffusive models. Taking the FO
model with low extracellular impedance (ze = re at 1 Hz) leads to large differences with the FC model (Fig. 9, black)
compared to the FO model (blue) or the FC model with zero extracellular resistance (red).
We can see that, in FC type models, the larger ze, the steeper the transfer function attenuates with distance. In contrast,
in FO type models, larger zm
e lead to less attenuation. This paradoxical result can be explained as follows: in FC models,
ze plays as similar role as zi, such that for large values of their real part, thermal diffusion attenuates the signal; in FO
models, zm
e limit the leak membrane current, reducing the attenuation
with distance. Thus, for large zm
e , the dendrites become more "democratic" in the sense that the effect of a given input
will be less dependent on its position on the dendrite. This is only the case for FO models, however.
e plays a similar role as rm, and large values of zm
As above, the model with zero extracellular resistance represents the same limit case for both FO and FC models, and
therefore constitutes the frontier between the two models.
Bedard&Destexhe(October30,2018)
17
∂x2 = κλVm, so that the quantity ∂2Vm
Resonances with diffusive models One interesting finding is that resonances appear in several models using diffusive
extracellular impedances (Figs. 7 and 9). This type of resonance was studied further in Fig. 10, where one can see that a
resonance in κλ also implies a resonance in FT: The Vm still attenuates with distance independently of the frequency, so
that we always have ∂Vm∂x < 0. In addition, Eq. 19 shows that ∂2Vm
∂x2 increases when
κλ increases with frequency, which implies that ∂Vm∂x becomes more negative because this derivative is always negative.
It follows that FT attenuates more steeply with distance when κλ increases with frequency. Using a similar reasoning,
one can show that FT attenuates less steeply with distance when κλ diminishes with frequency. We conclude that the
rate of variation of FT with frequency is always opposed to that of κλ. Consequently, the resonance frequency must be
the same for κλ and dFTd f = 0 because we have dFT d f ≤ 0 when dκλd f ≥ 0 and dFT d f ≥ 0 when dκλd f ≤ 0.
We also see that the peak frequency of the resonance continuously depends on the membrane time constant (not shown).
For example, for τm = 5 ms, the resonance is at about 24 Hz, and for τm = 20 ms, the resonance is at about 8 Hz (for more
details see Fig. 10 and Table II). It is interesting to note that we have observed resonances only in FC type models with
resistive extracellular media and diffusive cytoplasm (see Fig. 7), but resonances are present in the two types of models
(FO and FC) when they are fully diffusive.
IV. DISCUSSION
In this paper, we have introduced a generalization of cable equations to membranes within media with complex or
heterogeneous electrical properties. We have shown that generalized cable equations can treat a number of problems
presently not treatable by the traditional cable equations. We have shown that the nature of the extracellular medium has
a significant influence on fundamental neuronal properties, such as voltage attenuation with distance, and the spectral
profile of the transmembrane potential. We enumerate below the consequences and predictions of this work, as well as
outline directions for future studies.
A first main result of this paper is to generalize cable equations to describe membranes in complex and heterogeneous
media. To solve this problem, we have introduced the concept of generalized current, and show that the generalized current
is conserved in all situations. This stands in contrast with the free-charge current, which is conserved only in special cases.
For example, if the medium is electrically non-homogeneous (with conductive and non-conductive domains), there will
be charge accumulation and non-conservation of the free-charge current. Thus the traditional cable formalism, which is
based on the free-charge current, cannot treat this problem. With the generalized current, however, this problem can be
treated in a physically plausible way, in accordance with Maxwell equations.
One drawback of generalized cable equations is that they cannot be solved with available neural simulation environ-
ments, such as NEURON [13], which implements the traditional cable formalism. Consequently, we have developed a
specific method for the numerical simulation of generalized cables. This method is implementable with traditional simu-
lation programs, such as MATLAB. Further work would be needed to determine if generalized cable equations could be
included in neural simulators, as a special case.
Note that specialized models different from the standard model were introduced relatively recently [22, 27] to include
aspects which cannot be treated by the standard model.
In [22], the cytoplasm was considered as non-resistive but
capacitive, and was modeled by a RC circuit. It was estimated that this capacitive aspect is important to understand the
nature of thermal noise in thin dendritic branches. [27] considers the case of the interaction between closely located
dendritic branches. In this case, the authors study the phenomenon of surface polarization (see also [11]) and evaluate the
magnitude of the Maxwell-Wagner time of the effective impedance of the extracellular medium, needed to have significant
influences over the attenuation profile of the Vm. These two studies show that the physical nature of the intracellular or
extracellular media can have significant influences on cable properties. However, they do represent very particular cases,
which motivated the present study where we have attempted to consider a broad range of cases, including both intracellular
and extracellular media, as well as ionic diffusion, which was not treated previously. Thus, the present study generalizes
those prior studies.
A second main result of this paper was to also generalize the electrical circuit representing neuronal membranes. Instead
of considering the neuron as a closed system, where all outward currents return to the neuron, we have considered the
Bedard&Destexhe(October30,2018)
18
more general case which allows current exchange between neighboring neurons, and thus each is represented by an open
circuit. We have systematically compared open-circuit (FO) models with the traditional closed-circuit (FC) models, and
found some important differences. FO models have a transfer function that depends much less on frequency and space,
compared to FC models (see Figs. 7 and 8).
We also showed that a previously introduced model of non-ideal cable [15] is equivalent to a traditional cable with
appropriately scaled extracellular resistances (for frequencies smaller than 100 Hz; see Figs. 7 and 8 in [15], as well as
the discussion in that paper).
One of the most important result of this paper is the finding that the nature of extracellular or intracellular media can
have a strong impact on cable properties such as voltage attenuation with distance. We have observed that the nature of
the extracellular medium has an opposite impact on distance attenuation on FO and FC models. In FO models, larger
extracellular impedances lead to less attenuation and electrotonically more compact dendrites. The attenuation can be
remarkably diminished for fully resistive FO models, with only a few percent attenuation (Fig. 9), whereas for FC type
models, the opposite was seen, the dendrites become more compact for low extracellular impedances. We can say that in
these cases, the effect of distal inputs is close to that of proximal inputs, and thus the dendrite is more "democratic". It
may be that this remarkable property is present in some types of neurons to reduce the attenuation of distal inputs, which
constitutes another interesting direction to explore in future work.
Another interesting observation is that diffusive extracellular impedances can give rise to resonance frequencies (see
Figs. 7 and 9), which also appears as a resonance in κλ (Fig. 10). The resonant frequency depends on the membrane time
constant, and is in the range of 5-40 Hz, which is well within the frequency range of brain oscillations such as theta, alpha,
beta or gamma rhythms [28]. It is therefore possible that this resonance plays a role in the genesis of oscillatory activity
by single neurons.
Interestingly, we observed that the input impedance of the extracellular medium (z(m)
e
) must necessarily be negative
in the standard model where the medium is resistive. In a closed-circuit configuration, this means that one must neces-
sarily assume a source of energy, such as an electromotive force. This source of energy can be simulated by a negative
impedance. This important point was pointed in previous work, where it was called "anomalous impedance" [23, 24].
Interestingly, this constraint disappears in the open-circuit configuration. If the current field is open in the extracellular
medium, then it is not necessary to assume that z(m)
e
is negative, and there is no need of such a source of energy.
Finally, while our analysis shows that the nature of the extracellular or intracellular media may be influential on single-
neuron behavior, we can also foresee consequences at the network level. First, the resonance found for some of the
media may introduce a bias in the genesis of oscillatory behavior by populations of neurons. The fact that the resonance
frequency only depends on membrane parameters, but not on structural parameters such as cell size, suggests that dif-
ferent neurons in the network will have the same resonance frequency. It is thus conceivable that population oscillatory
activity may occur at this resonance frequency. Second, the fact that the diffusive properties of media were found to
be particularly impactful on the attenuation of distal inputs suggests that any regulation of these properties could have
drastic consequences at the network level. If diffusive properties are modified, for example by glial cells who are known
to regulate extracellular ionic concentrations [29, 30], it may affect the voltage attenuation of all cells in the network and
therefore change network behavior.
In conclusion, we think that the generalized cable equations allow one to treat the problem of how neuronal membranes
behave in complex extracellular and heterogeneous media. Given the possible strong impact of such media as found here,
future studies should evaluate in more depth whether such media are indeed influential. A possible approach would be
to find "signatures" of the extracellular medium from the power spectral density of experimentally observable variables,
such as the membrane potential (for a related approach, see [8]). The direct measurement of the extracellular impedance,
at present bound to contradictory experimental results [3 -- 5], should give a definite indication whether the generalized
cable is a necessary approach to accurately model neurons.
Bedard&Destexhe(October30,2018)
19
Appendices
Appendix A: Generalized current and charge conservation
In this appendix, we derive the charge conservation laws for different definitions of currents (see Eqs. 1 and 2). Consider
a domain D delimited by a closed surface ∂D. If we assume that the medium and the field are sufficiently regular, then
the divergence theorem applies in D, and we have:
(cid:9)
∂D
∇ × ~H · n dS ≡ $
D
∇ · (∇ × ~H) dv ≡ 0
because the following equality always applies: ∇ · (∇ × ~H) ≡ 0 [40]
From Eqs. 1, 2 and A1, we have the following identity:
(cid:9)
∂D
~j g · n dS ≡ $
∇ · ~j g dv = 0 ,
D
(A1)
(A2)
which is valid for an arbitrary domain D.
One can distinguish three different types of current, the generalized current i g, the current due to free charges i f , and
the displacement current i d. These currents can be defined across an arbitrary surface S, according to:
i g def
def
= !S
= !S
= !S
i f
i d def
~j g · n dS
~j f · n dS
∂ ~D
∂t
· n dS
(A3)
Within these definitions, we can write that the generalized current i g is conserved at every time and independently of the
nature of the medium. At every time, the inward current entering a given domain D is always equal to the outward current
exiting that domain, independently of the homogeneous or heterogeneous nature of the medium. It is also independent of
the fact that there may be charge accumulation in some elements of volume, because Eq. A2 always applies.
Note that this generalized current conservation law does not express anything new on a physical point of view, but is the
charge conservation law expressed as a function of currents. Indeed, taking into account Maxwell-Gauss law (∇· ~D = ρ f ),
the definition of ~j g (Eq. 2) and the identity given by Eq. A2, we obtain the differential charge conservation law:
∇ · ~j g = ∇ · ~j f + ∇ ·
∂ ~D
∂t
= ∇ · ~j f +
∂ ∇ · ~D
∂t
= ∇ · ~j f +
f
∂ρ
∂t
= 0
(A4)
Appendix B: Displacement current, free current and charge accumulation
In this appendix, we show explicitly that the displacement current i d can be used to formally calculate the charge
variation in a given domain D. Moreover, we show that the displacement current across a closed surface ∂S which
surrounds a given domain D is zero when there is no charge variation inside the domain.
Bedard&Destexhe(October30,2018)
By definition, the density of displacement current (Eq. A3) in frequency space is given by:
~j d(~x, ω) = iωε(~x, ω) ~E(~x, ω)
where ω = 2π f . By applying the divergence on ~j d and taking into account Maxwell-Gauss law, we obtain:
∇ · ~j d = iω∇ · (ε ~E) = iωρ f
20
(B1)
(B2)
Thus, we can calculate the amount of free charges in a given domain D from the density of displacement current in
frequency space. To do this, we have
Q f (ω) = $
D
ρ f (~x, ω) dv =
1
iω $
D
∇ · ~j d(~x, ω) dv ≡
~j d · n dS =
i d(ω)
iω
1
iω (cid:9)
∂D
(B3)
where i d is the displacement current flowing across surface ∂S. Applying the inverse Fourier transform, we obtain the
rate of free charge variation in domain D:
f
dQ
dt
(t) = i d(t)
(B4)
Therefore, one can say that the charge in the considered volume does not vary if the displacement current across surface
∂D is zero. Finally, because the differential conservation law for free charges implies:
f
dQ
dt
(t) = $
D
∂ρ f (~x, t)
∂t
dv = −$
D
∇ · ~j f (~x, t) dv ≡ −(cid:9)
∂D
~j f · n dS = −i f (t) ,
we can then write:
i g(t) = i d(t) + i f (t) = 0
(B5)
(B6)
when the surface is closed and when the free charge conservation law applies.
Thus, the generalized current entering a given closed surface ∂D is always equal at every time to the generalized current
exiting ∂D, even if there is free charge accumulation inside ∂D. However, this equality does not allow one to deduce if
there are variations of free charge density inside ∂D, because the displacement current must necessary be zero across ∂D
to have dQ
= 0 (see Eq. B4). In other words, it is necessary that the displacement current entering ∂D is equal to the
dt
displacement current exiting ∂D to have a constant charge inside ∂D. Note that in any given circuit, Kirchhoff's current
law always applies to the generalized current, even if there is charge accumulation inside the circuit, whereas it applies to
the free charge current only assuming there is no charge accumulation inside the circuit.
f
Appendix C: Input impedance of a cable segment in series with an arbitrarily complex impedance
In this appendix, we calculate the input impedance of a cable segment of length l when this segment is connected to an
arbitrary impedance Za (see Fig. 4).
By definition, we have in x = 0:
Zl
in[Za] =
Vm (0, ω)
i g
i (0, ω)
Applying Eq. 21 allows us to directly express Vm as a function of the cable parameters. We have
Vm(0, ω) = A+(ω) eκλl + A−(ω) e−κλl,
Similarly, applying Eqs. 10, 13 and 18, we obtain:
i g
i = −
1
zi
[1 +
z(m)
e
rm
((1 + iωτm)]
∂Vm
∂x
(C1)
(C2)
Bedard&Destexhe(October30,2018)
This last expression allows us to express the current at coordinate x = 0 as a function of the cable parameters:
where
i g
i (0, ω) =
κλ
¯zi
[A+(ω) eκλl − A−(ω) e−κλl]
¯zi =
1 +
z(m)
e
rm
zi
(1 + iωτm)
Thus, the expression for the input impedance Zl
in is given by:
Zl
in[Za] =
¯zi
κλ ·
( A+
A− ) · e2κλl + 1
( A+
A− ) · e2κλl − 1
21
(C3)
(C4)
(C5)
We can then evaluate the ratio A+
Applying Eqs. 21 and 10 to that point gives:
A− by using the conditions of continuity of the current and of the voltage at point x = l.
Thus, we have
and we can write
Vm(l, ω) =
A+(ω) + A−(ω)
(a)
i g
i (l, ω) = κλ
¯zi
[A+(ω) − A−(ω)] (b)
Za =
Vm(l, ω)
i g
i (l, ω)
=
κλ
¯zi
A+
A− + 1
[ A+
A− − 1]
,
A+
A−
=
κλZa + ¯zi
κλZa − ¯zi
It follows that the input impedance Zl
in is given by:
where
Zl
in[Za] =
¯zi
κλ
(κλZa + ¯zi) e2κλl + (κλZa − ¯zi)
(κλZa + ¯zi) e2κλl − (κλZa − ¯zi)
¯zi =
1 +
z(m)
e
rm
zi
(1 + iωτm)
(C6)
(C7)
(C8)
(C9)
Note that Zl
in[Za] → ¯zi
κλ
when l → ∞, and Zl
in[Za] → ¯zi
κλ
coth(κλl) when Za → ∞.
Appendix D: Calculation of the transfer function FT
In this appendix, we calculate the transfer function FT (l, ω; Za) =
for Appendix C.
Applying Eq. C5a gives:
Vm(l,ω)
Vm(0,ω) using the same conditions and conventions as
Thus, we have
Vm(0, ω) = A+(ω) eκλl + A−(ω) e−κλl
Vm(l, ω) =
A+(ω)
+
A−(ω)
FT (l, ω; Za) =
A+(ω) + A−(ω)
A+(ω) eκλl + A−(ω) e−κλl
(D1)
(D2)
Bedard&Destexhe(October30,2018)
Applying Eq. C7 gives the transfer function:
where
FT (l, ω; Za) =
κλZa
κλZa cosh(κλl) + ¯zisinh(κλl)
¯zi =
1 +
z(m)
e
rm
zi
(1 + iωτm)
Note that FT (l, ω; 0) = 0 and FT (l, ω; ∞) =
1
cosh(κλl) .
22
(D3)
FIG. 11: (Color online) Equivalence of the electrical circuits of open-circuit and non-ideal cable models. The circuits A and B are
equivalent when the ratio V12(ω)
I(ω) of the voltage difference between points 1 and 2 and the input current between these points is invariant,
and when the correspondence between the elements of these circuits are independent of frequency. Note that the values of the elements
between the two circuits are related by a transformation law which is independent of frequency; this equivalence also applies to the
temporal domain. In other words, according to this equivalence, the two circuits are equivalent when it is impossible to distinguish
their topology from external measurements. The circuit A corresponds to the non-ideal capacitance model introduced previously [15],
while circuit B corresponds to a "standard cable model" with a short-cut (zero extracellular resistivity).
Appendix E: A new interpretation of the non-ideal cable
In this appendix, we show that the non-ideal capacitance model introduced previously [15] is equivalent to an open-
circuit resistive model if we assume that the circuits A and B in Fig. 11 are linked by the following transformation:
ra = rm − rmrsc
rm+rsc
rm = ra + rb
rb =
rmrsc
rm+rsc
rsc = rb +
r2
b
ra
ca =
(rm+rsc)
rm rsc
rm−
rm +rsc
cm
cm =
r2
a
(ra+rb)2 ca
(E1)
We show that the Vm in the non-ideal cable model corresponds to Vi in an open-circuit (FO) type resistive model, with a
reference located far-away. According to circuits A and B in Fig. 11, we have:
circuit A (rsc ⊕ cm) k rm
circuit B (ra k ca) ⊕ rb
(E2)
Bedard&Destexhe(October30,2018)
It follows that the impedances of circuits A and B are equal if we have:
V12(ω)
i(ω)
=
rm + iωrmrsccm
1 + iωcm(rm + rsc)
=
ra + rb + iωcararb
1 + iωcara
23
(E3)
We see that the ratio V12(ω)
is a homographic transform of variable ω. Consequently, ∀ω we have the following relation
i(ω)
V12(ω)
i(ω) = aA+bAω
1+dAω = aB+bBω
1+dBω when the two circuits are equivalent. The only way to guarantee that the equivalence is
independent of frequency is to assume that the corresponding coefficient of the transformations are equal. We can thus set
aA = aB, bA = bB and dA = dB. This gives us 3 equations which link the 3 parameters of circuit A to those of circuit B.
The solution is the transformation law (Eqs. E1). Thus, on a physical point of view, one cannot distinguish the topology
of circuits A and B if we would perform external measurements. Moreover, because the functions rm = fm(ra, rb.rc),
rsc = fsc(ra, rb.rc), rcm = fcm(ra, rb.rc) do not depend on frequency, their equivalence will be also valid for all frequencies.
We can deduce that the two circuits will behave identically as a function of time .
It follows that the Vm (between points 1 and 2) in circuit A (non-ideal capacitance) corresponds to the Vi relative to a far-
away reference in circuit B (see Table I). Therefore, a model with non-ideal capacitance and zero extracellular resistance
should produce a Vm equivalent to the Vi of a model with ideal capacitance and resistive extracellular medium. Thus, the
frequency-scaling behavior of the Vm obtained in a previous non-ideal cable model [15] also applies to the resistive FO
model, but only if one studies the intracellular potential Vi.
Acknowledgments
Research supported by the CNRS, the ANR (Complex-V1 project) and the European Union (BrainScales FP7-269921
and Magnetrodes FP7-600730).
[1] Rall, W. (1962) Electrophysiology of a dendritic neuron model. Biophys J. 2: 145-167.
[2] Rall, W. (1995) The theoretical foundations of dendritic function. MIT Press, Cambridge, MA.
[3] Logothetis N.K., Kayser, C. and Oeltermann, A. (2007) In vivo measurement of cortical impedance spectrum in monkeys :
implications for signal propagation. Neuron 55: 809-823.
[4] Gabriel, S., Lau, R.W. and Gabriel, C. (1996a) The dielectric properties of biological tissues : I. Literature survey. Phys. Med.
Biol.. 41 : 2231-2249.
[5] Gabriel, S., Lau, R.W. and Gabriel, C. (1996b) The dielectric properties of biological tissues : II. Measurements in the frequency
range 10 Hz to 20 GHz. Phys. Med. Biol.. 41: 2251-2269.
[6] B´edard, C., H. Kroger, and A. Destexhe. 2006. Does the 1/f frequency scaling of brain signals reflect self-organized critical states ?
Physical Review Lett. 97: 118102.
[7] Bazhenov M, Lonjers P, Skorheim P, Bedard C and Destexhe A. (2011) Non-homogeneous extracellular resistivity affects the
current-source density profiles of up-down state oscillations Phil Trans R Soc A 369: 3802-3819.
[8] B´edard, C., Rodrigues, S., Roy, N., Contreras, D. and Destexhe, A. (2010) Evidence for frequency-dependent extracellular
impedance from the transfer function between extracellular and intracellular potentials. J. Computational Neurosci. 29: 389-403.
[9] Dehghani, N., B´edard, C., Cash, S.S., Halgren, E. and Destexhe, A. (2010) Comparative power spectral analysis of simultane-
ous elecroencephalographic and magnetoencephalographic recordings in humans suggests non-resistive extracellular media. J.
Computational Neurosci. 29: 405-421.
[10] B´edard, C. and Destexhe, A. (2011) A generalized theory for current-source density analysis in brain tissue. Physical Review E
84: 041909.
[11] B´edard, C., Kroger, H. and Destexhe, A., (2006b) Model of low-pass filtering of local field potentials in brain tissue. Phys. Rev.
E 73:051911.
[12] B´edard, C. and Destexhe, A. (2009) Macroscopic models of local field potentials and the apparent 1/f noise in brain activity.
Biophys. J. 96: 2589-2603.
[13] Hines, M.L. and Carnevale, N.T (1997) The NEURON simulation environment. Neural Computation 9: 1179-1209.
[14] Tuckwell, H.C. (1988) Introduction to Theoretical Neurobiology: Linear cable theory and dendritic structure. Cambridge Uni-
versity Press, Cambdridge, UK.
Bedard&Destexhe(October30,2018)
24
[15] B´edard, C. and Destexhe, A. (2008) A modified cable formalism for modeling neuronal membranes at high frequencies. Biophys.
J. 94: 1133-1143.
[16] Johnston, D. and Wu, S.M. (1995) Foundations of cellular neurophysiology. MIT Press, Cambridge, MA.
[17] Koch, C. (1999) Biophysics of Computation. Oxford University press, Oxford, UK.
[18] Destexhe, A. and B´edard, C. (2012) Do neurons generate monopolar current sources ? J. Neurophysiol. 108: 953-955.
[19] Riera, J.J., Ogawa, T., Goto, T., Sumiyoshi, A., Nonaka, H., Evans, A., Miyakawa, H. and Kawashima, R. (2012) Pitfalls in the
dipolar model for the neocortical EEG sources. J. Neurophysiol 108: 956-975.
[20] B´edard, C., Kroger, H. and Destexhe, A. (2004) Modeling extracellular field potentials and the frequency-filtering properties of
extracellular space. Biophys. J. 64: 1829-1842.
[21] Wang K, Riera JJ, Enjieu-Kadji H and Kawashima R. (2013) The role of the extracellular conductivity profiles in compartmental
models for neurons: Particulars for layer 5 pyramidal cells. Neural Computation 25: 1807-1852.
[22] Poznanski, R. (2010) Thermal noise due to surface-charge effects within the Debye layer of endogenous structures in dendrites.
Physical Review E 81: 021902.
[23] Conciauro, G. and Puglisi, M. (1981) Meaning of the negative impedance. NASA STI/Recon Technical Report 82: 14458.
[24] Mauro, A. (1961) Anomalous impedance, a phenomenological property of time-variant resistance: An analytic review. Biophys.
J. 1: 353-372.
[25] Braitenberg, V. and A. Shutz. (1998) Cortex: Statistics and Geometry of Neuronal Connectivity. (2nd ed.), Springer-Verlag, Berlin,
Germany.
[26] Nicholson, C.. (2005) Factors governing diffusing molecular signals in brain extracellular space. J. Neural Transm. 112: 29-44.
[27] Hiromu, M, Inoue, M., Miyakawa, H. and Aonishi, T. (2012) Low-Frequency dielectric dispersion of brain tissue due to electri-
cally long neurites. Physical Review E 86: 061911.
[28] Buzsaki G. (2006) Rhythms of the Brain., Oxford University Press, Oxford UK.
[29] Walz W. (1989) Role of glial cells in the regulation of the brain ion microenvironment. Progress Neurobiol. 33: 309-333.
[30] Kettenmann H. and Ransom BR. (1995) Neuroglia, Oxford University Press, Oxford, UK.
[31] George, M., Lisanby, S., and Sackeim, H. (1999) Transcranial magnetic stimulation : Applications in neuropsychiatry. Arch. Gen.
Psychiatry, 56 (4) :300311.
[32] Landau, L.D. and ifshitz, E.M. (1984) Electrodynamics of Continuous Media. Pergamon Press, Moscow, Russia.
[33] Purcell, E.M. (1985) Electricity and Magnetism. Berkeley Physics Course Vol.2 chap. 3.
[34] Note that if the negative charges are exclusively on the membrane (neglecting Debye layers), then the surface integral simplifies
to the side of the domain in contact with the membrane. In this case, the electric displacement is different from zero only in the
latter portion of the surface.
[35] In the presence of magnetic stimulation, such as for example transcranial magnetic stimulation, we have [31 -- 33]: ∇× ~E = − ∂~B
, 0
[36] This law can be simplified if one assumes that the cytoplasm is ohmic and homogeneous (zi = ri) for negligible la permittivity. In
∂t
this case, we have:
i = i f
i g
i = ~ji
f
· (πa2 n) = −πa2σe
i
∂Vi
∂x
1
ri
∂Vi
∂x
= −
[37] A medium is said to be "diffusive" when ionic diffusion is non-negligible in the presence of an electric field.
[38] For example, if ze = re and zi = ri, then we have R(z(m)
e
[39] This will be the case for example if two neighboring dendrites have current sources of opposite sign, there will be a direct
current flow between them. If they belong to different neurons, this configuration necessarily requires an open-circuit model to be
accounted for.
) = − rmre
< 0.
1
ri+re
1+ω2τ2
m
[40] We use the symbol ≡ for a mathematical identity while the symbol = will mark an equality or a physical law.
|
1902.06352 | 1 | 1902 | 2019-02-17T23:47:34 | Metabolic basis of brain-like electrical signalling in bacterial communities | [
"q-bio.NC",
"q-bio.CB"
] | Information processing in the mammalian brain relies on a careful regulation of the membrane potential dynamics of its constituent neurons, which propagates across the neuronal tissue via electrical signalling. We recently reported the existence of electrical signalling in a much simpler organism, the bacterium Bacillus subtilis. In dense bacterial communities known as biofilms, nutrient-deprived B. subtilis cells in the interior of the colony use electrical communication to transmit stress signals to the periphery, which interfere with the growth of peripheral cells and reduce nutrient consumption, thereby relieving stress from the interior. Here we explicitly address the interplay between metabolism and electrophysiology in bacterial biofilms, by introducing a spatially-extended mathematical model that combines the metabolic and electrical components of the phenomenon in a discretised reaction-diffusion scheme. The model is experimentally validated by environmental and genetic perturbations, and confirms that metabolic stress is transmitted through the bacterial population via a potassium wave. Interestingly, this behaviour is reminiscent of cortical spreading depression in the brain, characterised by a wave of electrical activity mediated by potassium diffusion that has been linked to various neurological disorders, calling for future studies on the evolutionary link between the two phenomena. | q-bio.NC | q-bio |
Metabolic basis of brain-like electrical signalling in
bacterial communities
Rosa Martinez-Corral1, Jintao Liu2,*, Arthur Prindle3,*, Gurol M. Suel4, and
Jordi Garcia-Ojalvo1
1Department of Experimental and Health Sciences, Universitat Pompeu Fabra, Barcelona Biomedical
Research Park, 08003 Barcelona, Spain
2Center for Infectious Diseases Research and Tsinghua-Peking Center for Life Sciences, School of
Medicine, Tsinghua University, 100084 Beijing, China
3Department of Biochemistry and Molecular Genetics, Feinberg School of Medicine, Northwestern
University, Chicago, IL 60611, USA; Center for Synthetic Biology, Northwestern University, Evanston,
4Division of Biological Sciences, San Diego Center for Systems Biology, and Center for Microbiome
Innovation, University of California San Diego, California 92093, USA
IL 60208, USA.
*Equal contribution
Keywords -- membrane potential, electrical signalling, potassium waves, bacterial biofilms,
cellular excitability
Abstract
Information processing in the mammalian brain relies on a careful regulation of the mem-
brane potential dynamics of its constituent neurons, which propagates across the neuronal
tissue via electrical signalling. We recently reported the existence of electrical signalling
in a much simpler organism, the bacterium Bacillus subtilis. In dense bacterial communi-
ties known as biofilms, nutrient-deprived B. subtilis cells in the interior of the colony use
electrical communication to transmit stress signals to the periphery, which interfere with
the growth of peripheral cells and reduce nutrient consumption, thereby relieving stress
from the interior. Here we explicitly address the interplay between metabolism and elec-
trophysiology in bacterial biofilms, by introducing a spatially-extended mathematical model
that combines the metabolic and electrical components of the phenomenon in a discretised
reaction-diffusion scheme. The model is experimentally validated by environmental and ge-
netic perturbations, and confirms that metabolic stress is transmitted through the bacterial
population via a potassium wave.
Interestingly, this behaviour is reminiscent of cortical
spreading depression in the brain, characterised by a wave of electrical activity mediated by
potassium diffusion that has been linked to various neurological disorders, calling for future
studies on the evolutionary link between the two phenomena.
1
1
Introduction
Local interactions between individual cells are known to generate complex emergent behaviour in
multicellular organisms, which underlie the functionalities of tissues, organs, and physiological
systems [1]. The human brain provides one of the finest examples of such phenomena, with
its complexity ultimately emerging from the interactions between neuronal cells [2]. Similarly,
unicellular organisms can also self-organise into communities with complex community-level
phenotypes [3]. Bacterial biofilms, in particular, provide a good model system for the study of
collective behaviour in biological systems [4].
We have previously shown that biofilms of the bacterium B. subtilis can display community-
level oscillatory dynamics [5]. These oscillations ensure the viability, under low-nitrogen (low-
glutamate) conditions, of the cells in the centre of the community, which are crucial for com-
munity regrowth upon external sources of stress like antibiotics.
In our experimental setup
(Fig. 1A), growing two-dimensional biofilms that reach a certain size develop oscillations [6],
such that peripheral cells periodically stop their growth (see black dotted line in Fig. 1B,C).
This behaviour was attributed to a metabolic codependence between central and peripheral
cells mediated by ammonium [5] (see introduction to Section 2 below).
Besides exhibiting metabolic oscillations in growth rate, our biofilms display periodic changes
in extracellular potassium (Fig. 1B) and cellular membrane potential (Fig. 1C), as reported
by the fluorescent dyes APG-4 and Thioflavin-T (ThT), respectively. The membrane potential
changes reported by ThT are caused by the periodic release of intracellular potassium ions (Ref.
[7] and Fig. 1D). Potassium propagates from the centre of the biofilm to its periphery according
to the following scenario: potassium released by glutamate-deprived (and thus metabolically
stressed) cells in the centre diffuses towards neighbouring cells, causing depolarisation, stress
and subsequent potassium release (Fig. 2-right), which thereby keeps propagating along the
biofilm in a bucket-brigade manner [7].
Notably, potassium and glutamate are the two most abundant ions in living cells [8]. In
addition to their their roles in osmoregulation, pH maintenance and nitrogen metabolism, they
are essential for information propagation in animal nervous systems. In particular, potassium
is one of the key ions involved in the modulation of the membrane potential in neurons, and
glutamate is a central excitatory neurotransmitter. Besides the presence of the same ions in
the two scenarios, there are also phenomenological links.
In a normally functioning brain,
electrical signalling typically propagates along neurons through action potentials mediated by
sodium and potassium that can travel at speeds up to meters per second [9]. In addition, under
pathological conditions, electrical activity can also propagate over the brain on a much slower
timescale, on the order of millimetres per minutes (reviewed in Ref. [10]), similar to the biofilm
potassium waves. This occurs during the phenomenon of cortical spreading depression, a wave of
intense depolarisation followed by inhibition of electrical activity that entails a strong metabolic
disturbance of neuronal function. Cortical spreading depression has been shown to occur in the
brains of multiple species, from insects [11, 12] to mammals [13], where it has been regarded as
2
Figure 1: Biofilms of B. subtilis display oscillations in growth rate and electrical signalling
activity. A) Scheme of the microfluidics device used to grow biofilms. Biofilms grow around the
central pillar or attached to the wall, as in the cartoon. Biofilm dynamics can be represented
as a kymograph, where each row represents a one dimensional cross-section of the biofilm area
at a given movie frame. B-D) Experimental kymographs of a cross-section of a biofilm, as
depicted in A, for different channels. Each pixel is the average of a region of 5.2 × 5.2 µm. The
dotted black line denotes the limit of the biofilm as established from the phase-contrast images.
B) Oscillations in extracellular potassium, reported by APG-4. C) Oscillations in membrane
potential, reported by ThT. D) Extracellular potassium and ThT time traces corresponding
to the position denoted by the white line in panels B,C. Data has been smoothed with an
overlapping sliding window of size 3 frames.
3
A3 mmspacespacekymograph3 mmspacetime 3 mmspacet=0t=1t=n...spacet=0t=1t=n3 mm3 mmmedia flow0200400600space (m)48121620time (h)Bextracellular potassium0200400600space (m)48121620time (h)Cmembrane potential6101418time (h)0.30.40.50.60.7APG-4 (a.u.)D0.00.20.40.60.81.0APG-4(a.u.)0.00.20.40.60.81.0ThT (a.u.)0.200.250.300.350.400.45ThT (a.u.) an evolutionary conserved process [10] that has been linked to human pathologies like migraine,
and to the propagation of brain damage during ischemic stroke. The phenomenon has been
modelled as a reaction-diffusion process [14, 15] according to the following mechanism: an initial
stimulus leads to an increase in extracellular potassium, causing the membrane potential to rise
beyond a critical threshold that triggers a self-amplifying release of potassium, a major ionic
redistribution of sodium, calcium and chloride, and release of neurotransmitters like glutamate.
Diffusion of potassium and glutamate to neighbouring cells leads to subsequent depolarisation
beyond a threshold, thus causing ionic release and self-propagation of the wave [10, 16, 17].
The widespread importance of potassium and glutamate, and the links between electrical
signalling in bacteria and neuronal dynamics led us to investigate in more detail the roles
of those two ions in the oscillations exhibited by B. subtilis.
In this work, we unify in a
discretised reaction-diffusion scheme the metabolic and electrical components of the system.
Assuming a homogeneous population of cells, the model reveals the spontaneous emergence of
two phenotypically-distinct populations of cells as a consequence of the spatiotemporal dynamics
of the system, which causes oscillations to start beyond a critical size. We begin by developing a
model that accommodates the interactions considered in previous works both for the metabolic
and electrical components of the phenomenon. Then we show that the details of the ammonium
metabolism are not required to explain the key aspects of the dynamics. The behaviour can be
explained by a model with minimal interactions between glutamate metabolism and potassium
signalling, thus clarifying the process in bacteria and providing further insight into the roles of
these ubiquitous biological ions.
2 A unified spatially-extended model of biofilm oscillations
Conceptually, biofilm oscillations can be understood to arise as result of a delayed, spatially-
extended negative feedback of metabolic stress [6]. A first molecular model was proposed
in Ref. [5] (left half of Fig. 2), based on the fact that cells need to create glutamine from
glutamate and ammonium for biomass production. Since ammonium is not present in the MSgg
medium used in the experiments, it must be synthesised by the cells from glutamate. The model
assumed that only cells in the interior of the biofilm produce ammonium (through the glutamate
dehydrogenase enzyme GDH, H in Fig. 2), part of which diffuses away to the periphery of the
biofilm, where it is used by the peripheral cells to grow. In turn, peripheral growth reduces
glutamate availability in the centre, and subsequently ammonium availability in the periphery.
As a result, peripheral cells stop growing, thus allowing the centre to recover glutamate and
produce ammonium again, leading to a new oscillation cycle. The model considered two different
cell populations -- interior and peripheral -- , and was subsequently implemented on a continuous
spatial domain in Ref. [18], with still pre-defined interior and peripheral cell types.
This model can readily generate oscillations, which crucially rely on the assumption of a
strongly nonlinear activation of GDH by glutamate in the interior population of cells (modelled
with a Hill function with large Hill coefficient). However, it does not account for the electrical
4
Figure 2: Metabolic (left) and electric (right) processes involved in the response of the biofilm to
a gradient of nutrient-limitation-drive stress from the interior to the periphery. The left-hand
side schematises the interactions considered in the metabolic model of Ref.
[5], whereas the
right-hand side schematises the electrical signalling propagation model proposed in Ref. [7]. In
the diagram on the left, H represents the enzyme GDH. In the diagram on the right, EK stands
for excess extracellular potassium, V for the membrane potential, and nk for the gating variable
of the potassium channel.
signalling component of the phenomenon, which we incorporate next.
2.1 Full model description
We consider a horizontal cross-section through the middle of a biofilm (as depicted in Fig. 1A)
which allows us to simplify the system into a one-dimensional lattice, as shown in Fig. 3.
The left-most lattice sites are 'biofilm' (shaded squares, 'b'), followed by 'non-biofilm' sites
(white squares, 'n') on the right. We begin by a more detailed model (Fig. 3) that includes all
the aforementioned metabolic interactions [5] (left half of Fig. 2) in addition to the electrical
signalling propagation [7] (right half of Fig. 2). The details of the equations can be found in
the Supplementary, and we next describe the main points of the model (see also Table 1).
'Non-biofilm' sites have extracellular variables only: extracellular glutamate (Ge), ammo-
nium (A, assumed to have the same concentration inside and outside cells, following [5]), and
extracellular potassium Ke. These variables diffuse between neighboring lattice sites. In ad-
dition, we also account for the fact that media is flowing constantly through the microfluidics
chamber such that their concentration tends to become close to that in the media (Supplemen-
tary Eqs. (S1), (S2), (S17)).
5
30303030interiorperipherygrowthammoniumglutamateammoniumglutamateMetabolic model (Liu et al, 2015)Electric signalling model (Prindle et al, 2015)V stress (S)nkEKEKEKVEKintracellularextracellulardiffusionbiofilmH Figure 3: Scheme of the unified metabolic-electric model introduced in this work.
In the
lattice scheme on the left, the labels 'b' and 'n' inside the sites denote 'biofilm' and 'non-
biofilm' lattice sites, respectively.
In the network diagram on the right, G, K, h, H and A
represent glutamate, potassium, inactive GDH, active GDH and ammonium, respectively. The
quantity r represents biomass-producing biomolecules such as ribosomal proteins. Rg denotes
the glutamate receptor, V the membrane potential, nk the gating variable of the potassium
channel, and S the stress. h stands for inactive GDH. The subindices 'i' and 'e' label intracellular
and extracellular quantities, respectively. The two thick flat arrows highlight the links between
the metabolic and electrical components of the system.
'Biofilm' sites correspond to small biofilm regions, such that variable values would corre-
spond to averages across multiple cells. As schematised in Fig. 3, in this sites there are also the
extracellular variables (also subject to diffusion and media flow) in addition to the intracellular
ones that take part in the various biochemical reactions. We model explicitly the dynamics of
extracellular (Ge) and intracellular (Gi) glutamate in the biofilm (Supplementary Eqs. (S5)-
(S6)), where we consider glutamate uptake into the cells and the conversion of glutamate into
ammonium by the action of the enzyme GDH. In turn, we model GDH synthesis in its inactive
form (h) and subsequent activation (H) (Supplementary Eqs. (S8)-(S9)). We assume that high
concentrations of glutamate inhibit GDH synthesis. This follows from the original observation
that GDH overexpression in the periphery kills biofilm oscillations [5], suggesting that GDH
expression is restricted to the centre, where glutamate levels are low. Moreover, we assume
that there is a threshold for GDH activation by glutamate, as in the original metabolic model.
Glutamate and ammonium are assumed to be used for the production of biomass-producing
biomolecules such as ribosomal proteins, denoted by r (Supplementary Eq. (S11)). We also ac-
count explicitly for the glutamate transporter concentration (Rg), which we assume to saturate
for large enough Ge. In order to accommodate the assumption that cells with a higher metabolic
activity consume more glutamate, we consider that Rg is subject to an inducible activation that
depends on the presence of biomass-producing biomolecules (Supplementary Eq. (S7)).
In [7], the YugO potassium channel, which has a TrkA gating domain, was determined to
be responsible for potassium release in this phenomenon. TrkA proteins are potassium chan-
nel regulatory proteins with nucleotide-binding domains that bind NAD/NADH [19, 20] and
ADP/ATP [21], thus coupling potassium translocation to the metabolic state of the cell. Glu-
tamate is a central amino acid at the cross-road between carbon and nitrogen metabolism. Our
6
GiRg SnkKiKerAhHVintracellular extracellulardiffusiondiffusionbbbbnnGe assumption is that low glutamate levels lead to imbalances in the aforementioned nucleotides
that are sensed by TrkA and trigger potassium release [7]. Therefore, we assume that high
levels of intracellular glutamate inhibit the production of stress-related biomolecules, resulting
in the following equation for stress dynamics:
1 +(cid:0) Gi
S0
(cid:1)ns − γs S
(1)
dS
dt
=
Gs0
In turn, S enhances the opening probability of the potassium channel, given by nk. This results
in a link from the metabolic component of the phenomenon to the electrical counterpart.
As the potassium channel opens, potassium is released to the extracellular environment.
Conversely, potassium uptake is assumed to be governed by homeostatic processes that tend
to keep its intracellular concentration at a fixed value, in addition to depend on the cellular
metabolic state (glutamate level), to account for the energy demand of the process (Supplemen-
tary Eqs. (S13)-(S14)). Extracellular and intracellular potassium concentrations determine the
membrane potential (V ), whose dynamics is described by a Hodgkin-Huxley-like conductance-
based model (Supplementary Eq. (S18)). Finally we include the ThT reporter T downstream
of the membrane potential, increasing when the cells become hyperpolarised due to potassium
release (Supplementary Eq. (S20)).
In B. subtilis, glutamate is imported by pumps such as the symporter GltP [22], which uses
the proton motive force as a source of energy, and this is in turn influenced by the membrane
potential [23]. We therefore consider that glutamate transport into the cell is modulated by the
membrane potential V , such that depolarisation reduces entry, and hyperpolarisation enhances
it. This provides the link from the electrical to the metabolic component of the phenomenon,
and is represented with the following switch-like import-modulation term:
F(V ) =
1
1 + exp(gv (V − V0))
(2)
The biofilm is assumed to expand proportionally to the biomass-producing biomolecules. In
order to simulate growth, we consider that 'non-biofilm' lattice sites neighbouring biofilm sites
become occupied with cells (and thus become a lattice site of type 'biofilm') with probability
Pgrow dt r, with dt being the simulation time step. Importantly, each new biofilm lattice site
inherits the intracellular variables of the 'mother' site. Due to this particularity, the continuum
approximation is unsuitable, and we simulate the system as a coupled map lattice.
2.2 Simulations of the model
The diffusion coefficient in water for ions and small molecules such as potassium and glutamate is
of the order of ∼ 106µm2/h. We thus fixed the diffusion coefficient of potassium and ammonium
in the media to this value, and assumed the diffusion coefficient of glutamate to be half that
value, due to its larger molecular weight. Regular glutamate concentration in the media is
7
Model assumption
Full model
(section 2)
Simplified model
(section 3)
Extracellular glutamate flow and diffusion
Ammonium flow and diffusion
Extracellular potassium flow and diffusion
Decay of flow and diffusion rates towards the
biofilm interior as a consequence of cellular
density
Glutamate uptake favoured by
hyperpolarisation and disfavoured by
depolarisation
Yes
Yes
Yes
Yes
Yes
Yes
No (no ammonium
present)
Yes
Yes
Yes
Glutamate uptake enhanced in metabolically
active cells
Through the
effect of biomass-
producing
biomolecules
Directly linked to
intracellular
glutamate
Saturable glutamate uptake
GDH dynamics regulated by glutamate
Ammonium produced from intracellular
glutamate by GDH
Yes
Yes
Yes
Growth favored by high levels of intracellular
glutamate
Through
intermediate
biomass-
producing
biomolecules.
Yes
No (no GDH
present)
No (no ammonium
present)
Directly
modulated by
intracellular
glutamate
Metabolic stress inhibited by intracellular
glutamate
Potassium channel opening probability (nk)
increased by stress
Potassium released as a result of the opening
probability (nk) and membrane potential
Potassium uptake governed by homeostatic
processes and the metabolic state of the cells
(intracellular glutamate)
Membrane potential determined by
intracellular and extracellular potassium
concentrations
ThT as a reporter of membrane potential
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Table 1: Main assumptions and comparison between the full and simplified models.
8
30 mM and potassium concentration is 8 mM. Since there is no ammonium in the medium,
Am = 0. We fixed the resting membrane potential to −150 mV. Furthermore, oscillations
experimentally start at a biofilm size of around 600 µm with a period of around 2 hours [6],
and potassium and ThT signals should be correlated [7]. With these constraints, we manually
adjusted the rest of the parameters (Table S1) in order to reproduce the experimentally observed
oscillatory dynamics.
Figure 4 shows kymographs for extracellular potassium and ThT (variables that can be
experimentally monitored as explained above, see Fig. 1), as well as for extracellular glutamate,
stress, active GDH and ammonium, which have not been experimentally quantified. Oscillations
with a period of about two hours emerge in all variables once the biofilm becomes large enough.
Moreover, in agreement with the experimental data, the top left plot shows that potassium
peaks are likely to coincide with periods of no growth.
Figure 4: Oscillations in the full metabolic-electrophysiological model. Kymographs of the
model simulation results, showing oscillations in various model variables.
In contrast to the original model of Ref. [5] (see also Ref. [18]), where two populations of cells
(central and peripheral) were pre-defined regarding GDH production, in the current model an
a priori separation of cell types is no longer required, but emerges spontaneously. This can be
seen in the bottom left kymograph of Fig. 4, which shows that a central population with higher
levels of GDH activity and stress emerges over time. Notably, this model does not require strong
cooperativity in GDH activation for the oscillations to emerge (nH = 2, in contrast to the value
of 7 in Ref. [5] and 12 in Ref. [18]). This is likely to be the result of the electrical component
9
4812162024time (h)extracellularpotassiumThTextracellular glutamate050010001500space (m)4812162024time (h)active GDH050010001500space (m)ammonium050010001500space (m)stress8910111213141516Ke(mM)0.000.250.500.751.001.251.501.75ThT(µM)5101520Ge(mM)0.000.050.100.150.200.250.300.350.40H(µM)0.30.40.50.60.70.80.9A(mM)0.0000.0050.0100.0150.0200.025S(µM) of the model.
In order to test the relevance of this component of the oscillations, we next
simplify the metabolic details and consider only the interplay between glutamate metabolism
and electrical signalling.
3 Oscillations emerge from the interplay between glutamate
and electrical signalling
We now simplify the metabolic aspects of the full model introduced above, eliminating GDH,
ammonium and biomass-producing biomolecules (Supplementary Eqs. (S8)-(S11)), and suppos-
ing instead that intracellular glutamate directly increases the glutamate transporter and directs
growth (Fig. 5A). With this, the equations governing the metabolic part of the phenomenon
become:
dGe
dt
dGi
dt
dRg
dt
= −αg F(V ) Rg
= αg F(V ) Rg
Ge
kg + Ge
Ge
+ Λφ φ (Gm − Ge) + ΛD Dg∇2 Ge
− δg Gi
Gnr
i
kg + Ge
= αR − δR Rg + αr
knr
r + Gnr
i
(3)
(4)
(5)
in addition to considering glutamate diffusion in the 'non-biofilm' sites (Supplementary Eq. S1).
In this case, the probability with which the biofilm grows depends directly on the concentra-
tion of intracellular glutamate at the biofilm boundary according to Pgrow dt Gi. The dynamics
of the electrical part of the model remain unchanged (Supplementary Eqs. (S12)-(S20)).
The kymographs in Fig. 5B show that oscillations beyond a critical size also emerge in this
case. The results suggest the following mechanism for the oscillations: once the biofilm reaches
a critical size, glutamate in the centre of the biofilm is too low and leads to metabolic stress. As
a result, cells release potassium, that actively propagates towards the edge due to depolarisation
and subsequent metabolic stress. Depolarisation leads to a drop in glutamate consumption levels
in the peripheral region, due to the immediate effect on the transport efficiency, and a slightly
delayed effect through the reduction in the glutamate transporter. As a result, glutamate can
diffuse inwards to allow stress relief in the centre, and a new oscillation cycle can start.
According to this, oscillation onset requires stress in the centre to surpass a threshold that
leads to potassium release. As we have recently described [6], oscillations can be triggered in
experimental biofilms by transiently stopping the media flow of the microfluidics chamber. This
is also the case in this model, where oscillations can be triggered by transiently setting the
flow-rate parameter (φ) to zero. If the biofilm is sufficiently large, such a perturbation leads
to a sudden reduction in the glutamate availability and stress increase, with the subsequent
potassium release and oscillatory onset (Fig. S1).
Moreover, the model predicts that depolarising any region of the biofilm should be sufficient
to trigger a self-propagating wave of stress and potassium release, with the associated changes
in membrane potential. This is reproduced experimentally as shown in Fig. 5C: a short increase
10
Figure 5: Oscillations in a simplified model, with glutamate and the electrical signalling compo-
nents. A) Scheme of the model. B) Kymographs of various variables of the system. C) Electrical
signalling wave (reported by ThT) in an experimental biofilm triggered by an increase of potas-
sium in the media from 8 mM to 300 mM during 3 minutes (denoted by the orange mark). The
wave propagates inwards, from the periphery (right) to the centre (left).
in the concentration of potassium in the media triggers a wave in experimental biofilms, which
propagates inwards from the original (peripheral) depolarisation site.
4 Glutamate metabolism and stress release determine oscilla-
tion onset size and period
According to the aforementioned mechanism for the oscillations, if a biofilm increases its gluta-
mate consumption, or if the glutamate concentration in the media is reduced, oscillations should
start earlier because the centre becomes stressed at a smaller biofilm size. This expectation is
fulfilled when simulations are performed with halved glutamate concentration in the media
(Gm), in good agreement with the experimental data (compare the first two conditions in the
left panel of Fig. 6, see also Refs. [5, 6]). Similarly, bacteria that cannot synthesise glutamate
due to a deletion in the gltA gene are expected to consume more glutamate from the media. We
model this condition as an increase in both glutamate uptake rate (αg) and degradation rate
(δg), as a result of higher glutamate demand. Also in this case oscillations start earlier both in
the model and the experiments (third condition in the left panel of Fig. 6).
11
GiRg SnkKiKeVintracellular extracellulardiffusionbbbnnGeA050100150200space (m)02040time (min)Response to K shock0.00.20.40.60.81.0ThT (a.u.)Cbiofilmnon-biofilm050010001500space (m)4812162024time (h)extracellularpotassium050010001500space (m)membranepotential050010001500space (m)stress81012141618Ke(mM)0.000.250.500.751.001.251.501.75ThT(µM)0.000.010.020.030.040.050.06Stress(µM)B Figure 6: Biofilm size and oscillation period at onset. Light grey: Experimental Data. Dark
grey: Model. The data is normalised with respect to the average of each WT condition (model
and experiments separately). Boxplots extend between the first and third quartiles of the
normalised data, with a line (orange) at the median. The extreme of the whiskers denote
the range of the data within 1.5x the interquartile range, with outliers plotted individually.
Experimental data contains a minimum of 12 data points per condition. See Methods for
details on the simulations.
In these two cases of reduced onset size, the onset period is also smaller than in the reference
situation (first three conditions in the right panel of Fig. 6). As we have described in previous
work [6], larger biofilms tend to have higher periods, the explanation being that the larger a
biofilm is, the more time it takes for the stress signal to propagate from centre to periphery,
and thus the longer the oscillation period.
However, according to the oscillation mechanism that this model suggests, the period of the
oscillations must also depend on the metabolism dynamics and the stress relieving capabilites
of the cells due to the potassium effects. The ∆trkA mutant strain was characterised in Ref.
[7] to be deficient in electrochemical signalling, due to the deletion of the gating domain of the
YugO channel. The mutation can be interpreted to render a channel with reduced conductance
and increased basal leak [7]. This can be modelled with increased a0 (to simulate a leaky
channel), and reduced gK (to simulate reduced conductance once the channel is open). The
model simulations predict that oscillations start at the same size as with the basal parameters,
but with a reduced period, in good agreement with the experimental data (last condition in
Fig. 6). The unaffected onset size is consistent with the fact that glutamate metabolism is largely
unperturbed in this mutant, whereas the altered potassium channel leads to less effective stress
relief during wave propagation (Figs. S2), and thus facilitates a subsequent cycle. Therefore, the
relationship between biofilm size and oscillation period is not universal, but depends upon the
metabolic and electrical signalling capabilities of the cells and thus on the genetic background
of the strain.
12
WT(1xG)WT(0.5xG)gltAtrkA0.60.81.01.2normalised onset size onset sizeWT(1xG)WT(0.5xG)gltAtrkA0.51.01.5normalised onset period onset periodExperimentModel 5 Discussion
The oscillations in growth rate and membrane potential exhibited by B. subtilis biofilms is an
instance of self-organising, emergent behaviour similar to the complex collective dynamics that
characterizes neuronal tissue. Here we have proposed a unified conceptual framework for these
bacterial oscillations that combines glutamate metabolism with potassium wave propagation.
Our discrete reaction-diffusion scheme assumes an initially homogeneous population of cells, and
exhibits a spontaneous emergence of oscillations beyond a critical biofilm size. Our work shows
that oscillations are triggered by a decrease in glutamate levels in the biofilm centre, that lead
to metabolic stress and potassium release. Potassium diffusion to the neighbors interferes with
glutamate metabolism, thus causing a self-propagating wave of membrane potential changes
that allows glutamate recovery in the centre and thus explains the oscillation cycle.
With some notable recent exceptions in multicellular eukaryotes [24], electrophysiology has
only been considered relevant so far for excitable cells such as those in cardiac and neural
tissue. The resemblance, both at the molecular and functional level, between excitable wave
propagation in these tissues and the communication of metabolic stress among bacterial cells
calls for an evolutionary perspective on the phenomenon of electrical signalling. Animal nervous
systems have evolved intricate electrochemical circuits that allow sensing and responding to
both internal and external stimuli, with high integrative and computational capabilities clearly
evidenced by the human brain. How this complexity arose is still a matter of active investigation
[25, 26, 27, 28] but it is tempting to speculate that the bacterial oscillations reported here may
represent an ancient instance of electrically-mediated information transmission.
Primitive prokaryotic potassium channels are regarded as the ancestors of the animal cation
channels whose diversification has been linked to the evolution of nervous systems [29, 30].
Given that high intracellular concentrations of potassium are essential for bacterial pH main-
tenance [31, 8], it is natural to expect that potassium channels were used in early bacteria for
osmoregulation [27]. The biofilm oscillations studied here suggest that potassium also acts as
a signalling intermediate in modern bacteria, communicating information among cells of both
the same and also different bacterial species [32].
The link between metabolism and electrical signalling described here is also reminiscent of
the relationship between cellular energy and neuronal function [33]. Neuronal dynamics is highly
dependent upon cellular metabolism and energy state: the high energy requirements of the ion
pumps that maintain proper electrochemical gradients are well known [34]. In addition, some
animal voltage-gated ion channels directly respond to energy-related metabolites like NAD(H)
and ADP/ATP through direct ligand binding [35], as in the bacterial YugO channel.
While the mechanisms that allow information processing in bacterial biofilms share similar
chemical and electrical processes with animal brains, the latter compute at much faster speeds
and thus in much more complex ways. This can be partly ascribed to the reliance of animal
brains on faster voltage-gated ion channels, such as those dependent on sodium, which evolved
only when rapid responses were increasingly beneficial (due for instance to the appearance of
13
predators) [27]. Nevertheless, slow propagation of electrical activity is still present in modern
animal brains. During the phenomenon of cortical spreading depression, for instance, a slow
wave of electrical activity propagates over brain tissue, with massive ionic redistribution involv-
ing, critically, potassium and glutamate accompanied by a strong metabolic disturbance. This
phenomenon has been regarded as intrinsic to neurons and is evolutionarily conserved across
animal species [10]. In the light of these facts, it is tempting to see the potassium-mediated
transmission of stress among bacterial cells as a precursor of more recent electrically-mediated
information propagation tasks in animal brains. The study of membrane potential dynamics and
electrical signalling processes in other bacterial species will be important to further illuminate
this issue.
Methods
5.1 Experimental data
Strains and Plasmids. All experiments were performed using Bacillus subtilis strain NCIB
3610. For the ∆trkA mutant, we deleted the C-terminal portion of yugO (amino acids 117328),
leaving only the N-terminal ion channel portion of YugO (amino acids 1116). For the ∆gltA
mutant, the gltA gene was replaced by kanamycin resistant gene.
Growth conditions. Biofilms were grown using the standard MSgg biofilm-forming medium
[36]. In this media glutamate is the only nitrogen source for the bacteria. We explored biofilm
dynamics at glutamate concentrations of 1× (30 mM) and 0.5× (15 mM). For microfluidics,
We used the CellASIC ONIX Microfluidic Platform and the Y04D microfluidic plate (EMD
Millipore). Details can be found in our previous work [5, 37].
Time-Lapse Microscopy. Biofilms were monitored using time-lapse microscopy, using Olym-
pus IX81 and IX83 inverted microscopes. To image entire biofilms, 10× lens objectives were
used. Images were taken every 10 min. We tracked membrane potential dynamics using the
fluorescent dyes Thioflavin T (10 µM) and APG-4 (2µM), from TEFLabs [7].
5.2 Modeling
Simulation methods. Simulations were performed using custom-code in C, and analysis and
plotting was done in Python. We used a first-order approximation of the discrete Laplacian
operator [38] and the time evolution of the variables was obtained using a 4th order Runge-
Kutta method. Boundary conditions were reflective: the outside neighbours of the lattice sites
at the boundary of the system are the boundary sites themselves (such that the discrete normal
derivatives of the variables at the boundary are zero). The simulation time step was set to
5 × 10−6 h, and each lattice site corresponds to 10 µm. The system was allowed to relax
towards the steady state at the beginning of the simulation by preventing expansion during 1
14
hour of simulated time. The initial conditions for the intracellular variables are S = 0 µM ,
nk = 0, V = −156 mM, Ki = 300 mM, T hT = 0 µM , Rg = 1 µM , Hi = 0.5, H = 0.25,
R = 1, and Gi is log-normally distributed around 3 mM. In the biofilm cells, Ke is log-normally
distributed around 8 mM, and A = 1 mM. In the non-biofilm cells, A = 0. Initial Ge = 30 mM
(or 15 mM in reduced glutamate concentrations).
Numerical perturbations.
In order to study the effect of halved glutamate in the media,
and the trkA and gltA deletions, we assumed the following parameter values with respect to
the WT: ∆gltA mutant: δg = 1.5×, αgt = 1.3×; ∆trkA mutant: a0 = 2×, gK = 0.5×. We
performed 30 simulations per condition starting from a random initial radius between 150 and
250 µm. Potassium peaks were defined based on a threshold of 9.5 mM (pulse begins when
crossing the threshold). Onset period is defined to be the time between the beginnings of the
first two peaks, and onset size that of the first beginning.
Funding
This work was supported by the Spanish Ministry of Economy and Competitiveness and FEDER
(project FIS2015-66503-C3-1-P), and by the Generalitat de Catalunya (project 2017 SGR 1054).
R.M.C. acknowledges financial support from La Caixa foundation. J.G.O. acknowledges support
from the ICREA Academia programme and from the "Mar´ıa de Maeztu" Programme for Units
of Excellence in R&D (Spanish Ministry of Economy and Competitiveness, MDM-2014-0370).
G.M.S. acknowledges support for this research from the San Diego Center for Systems Biology
(NIH grant P50 GM085764), the National Institute of General Medical Sciences (grant R01
GM121888), the Defense Advanced Research Projects Agency (grant HR0011-16-2-0035), and
the Howard Hughes Medical Institute-Simons Foundation Faculty Scholars program. A.P. is
supported by a Simons Foundation Fellowship of the Helen Hay Whitney Foundation and a
Career Award at the Scientific Interface from the Burroughs Wellcome Fund.
References
[1] J. D. Murray, Mathematical biology. II Spatial models and biomedical applications. Springer,
2001.
[2] W. Freeman, Neurodynamics: an exploration in mesoscopic brain dynamics. Springer,
2000.
[3] A. F. Mar´ee and P. Hogeweg, "How amoeboids self-organize into a fruiting body: multi-
cellular coordination in dictyostelium discoideum," Proceedings of the National Academy
of Sciences, vol. 98, no. 7, pp. 3879 -- 3883, 2001.
[4] H. C. Flemming, J. Wingender, U. Szewzyk, P. Steinberg, S. A. Rice, and S. Kjelleberg,
15
"Biofilms: An emergent form of bacterial life," Nature Reviews Microbiology, vol. 14, no. 9,
pp. 563 -- 575, 2016.
[5] J. Liu, A. Prindle, J. Humphries, M. Gabalda-Sagarra, M. Asally, D.-y. D. Lee, S. Ly,
J. Garcia-Ojalvo, and G. M. Suel, "Metabolic co-dependence gives rise to collective oscil-
lations within biofilms," Nature, vol. 523, no. 7562, pp. 550 -- 554, 2015.
[6] R. Martinez-Corral, J. Liu, G. M. Suel, and J. Garcia-Ojalvo, "Bistable emergence of
oscillations in growing Bacillus subtilis biofilms," Proceedings of the National Academy of
Sciences, vol. 115, pp. E8333 -- E8340, 2018.
[7] A. Prindle, J. Liu, M. Asally, S. Ly, J. Garcia-Ojalvo, and G. M. Suel, "Ion channels enable
electrical communication in bacterial communities," Nature, vol. 527, no. 7576, pp. 59 -- 63,
2015.
[8] J. Gundlach, F. M. Commichau, and J. Stulke, "Perspective of ions and messengers: an in-
tricate link between potassium, glutamate, and cyclic di-AMP," Current Genetics, vol. 64,
no. 1, pp. 191 -- 195, 2018.
[9] J. E. Hall and A. C. Guyton, Guyton and Hall textbook of medical physiology. Elsevier
Saunders, 2011.
[10] C. Ayata and M. Lauritzen, "Spreading Depression, Spreading Depolarizations, and the
Cerebral Vasculature," Physiological Reviews, vol. 95, no. 3, pp. 953 -- 993, 2015.
[11] C. I. Rodgers, G. A. B. Armstrong, K. L. Shoemaker, J. D. LaBrie, C. D. Moyes, and
R. M. Robertson, "Stress Preconditioning of Spreading Depression in the Locust CNS,"
PLoS ONE, vol. 2, no. 12, p. e1366, 2007.
[12] K. E. Spong, E. C. Rodr´ıguez, and R. M. Robertson, "Spreading depolarization in the brain
of Drosophila is induced by inhibition of the Na+/K+-ATPase and mitigated by a decrease
in activity of protein kinase G," Journal of Neurophysiology, vol. 116, no. 3, pp. 1152 -- 1160,
2016.
[13] M. Lauritzen, J. P. Dreier, M. Fabricius, J. A. Hartings, R. Graf, and A. J. Strong, "Clinical
Relevance of Cortical Spreading Depression in Neurological Disorders: Migraine, Malignant
Stroke, Subarachnoid and Intracranial Hemorrhage, and Traumatic Brain Injury," Journal
of Cerebral Blood Flow & Metabolism, vol. 31, no. 1, pp. 17 -- 35, 2011.
[14] M. Dahlem and S. Muller, "Reaction-diffusion waves in neuronal tissue and the window of
cortical excitability," Annalen der Physik, vol. 13, no. 78, pp. 442 -- 449, 2004.
[15] J. C. Chang, K. C. Brennan, D. He, H. Huang, R. M. Miura, P. L. Wilson, and J. J. Wylie,
"A Mathematical Model of the Metabolic and Perfusion Effects on Cortical Spreading
Depression," PLoS ONE, vol. 8, no. 8, pp. 1 -- 9, 2013.
16
[16] Y. A. Dahlem, M. A. Dahlem, T. Mair, K. Braun, and S. C. Muller, "Extracellular potas-
sium alters frequency and profile of retinal spreading depression waves," Experimental
Brain Research, vol. 152, no. 2, pp. 221 -- 228, 2003.
[17] R. Enger, W. Tang, G. F. Vindedal, V. Jensen, P. Johannes Helm, R. Sprengel, L. L.
Looger, and E. A. Nagelhus, "Dynamics of Ionic Shifts in Cortical Spreading Depression,"
Cerebral Cortex, vol. 25, pp. 4469 -- 76, 2015.
[18] F. Bocci, Y. Suzuki, M. Lu, and J. N. Onuchic, "Role of metabolic spatiotemporal dynamics
in regulating biofilm colony expansion," Proceedings of the National Academy of Sciences,
vol. 115, no. 16, pp. 4288 -- 4293, 2018.
[19] A. Schlosser, A. Hamann, D. Bossemeyer, E. Schneider, and E. P. Bakker, "NAD+binding
to the Escherichia coli K+uptake protein TrkA and sequence similarity between TrkA and
domains of a family of dehydrogenases suggest a role for NAD+in bacterial transport,"
Molecular Microbiology, vol. 9, no. 3, pp. 533 -- 543, 1993.
[20] T. P. Roosild, S. Miller, I. R. Booth, and S. Choe, "A mechanism of regulating transmem-
brane potassium flux through a ligand-mediated conformational switch," Cell, vol. 109,
no. 6, pp. 781 -- 791, 2002.
[21] Y. Cao, Y. Pan, H. Huang, X. Jin, E. J. Levin, B. Kloss, and M. Zhou, "Gating of the TrkH
ion channel by its associated RCK protein TrkA," Nature, vol. 496, no. 7445, pp. 317 -- 322,
2013.
[22] F. Kunst et al., "The complete genome sequence of the gram-positive bacterium Bacillus
subtilis," Nature, vol. 390, pp. 249 -- 256, 1997.
[23] B. Tolner, T. Ubbink-Kok, B. Poolman, and W. N. Konings, "Characterization of the
proton/glutamate symport protein of Bacillus subtilis and its functional expression in Es-
cherichia coli," Journal of Bacteriology, vol. 177, no. 10, pp. 2863 -- 2869, 1995.
[24] M. Levin, G. Pezzulo, and J. M. Finkelstein, "Endogenous bioelectric signaling networks:
exploiting voltage gradients for control of growth and form," Annual Review of Biomedical
Engineering, vol. 19, pp. 353 -- 387, 2017.
[25] L. L. Moroz, "On the independent origins of complex brains and neurons," Brain, Behavior
and Evolution, vol. 74, no. 3, pp. 177 -- 190, 2009.
[26] D. Bucher and P. A. V. Anderson, "Evolution of the first nervous systems - what can we
surmise?," Journal of Experimental Biology, vol. 218, no. 4, pp. 501 -- 503, 2015.
[27] W. B. Kristan, "Early evolution of neurons," Current Biology, vol. 26, no. 20, pp. R949 --
R954, 2016.
17
[28] L. L. Moroz and A. B. Kohn, "Independent origins of neurons and synapses:
insights
from ctenophores," Philosophical Transactions of the Royal Society B: Biological Sciences,
vol. 371, no. 1685, p. 20150041, 2016.
[29] P. A. V. Anderson and R. M. Greenberg, "Phylogeny of ion channels: Clues to structure
and function," Comparative Biochemistry and Physiology - B Biochemistry and Molecular
Biology, vol. 129, no. 1, pp. 17 -- 28, 2001.
[30] H. H. Zakon, "Adaptive evolution of voltage-gated sodium channels: The first 800 million
years," Proceedings of the National Academy of Sciences, vol. 109, pp. 10619 -- 10625, 2012.
[31] C. M. Armstrong, "Packaging Life: The Origin of Ion-Selective Channels," Biophysical
Journal, vol. 109, no. 2, pp. 173 -- 177, 2015.
[32] J. Humphries, L. Xiong, J. Liu, A. Prindle, F. Yuan, H. A. Arjes, L. Tsimring, and
G. M. Suel, "Species-Independent Attraction to Biofilms through Electrical Signaling,"
Cell, vol. 168, no. 1-2, pp. 200 -- 209.e12, 2017.
[33] L. Yu and Y. Yu, "Energy-efficient neural information processing in individual neurons
and neuronal networks," Journal of Neuroscience Research, vol. 95, no. 11, pp. 2253 -- 2266,
2017.
[34] P. J. Magistretti and I. Allaman, "A Cellular Perspective on Brain Energy Metabolism and
Functional Imaging," Neuron, vol. 86, no. 4, pp. 883 -- 901, 2015.
[35] P. J. Kilfoil, S. M. Tipparaju, O. A. Barski, and A. Bhatnagar, "Regulation of ion channels
by pyridine nucleotides," Circulation Research, vol. 112, no. 4, pp. 721 -- 741, 2013.
[36] S. S. Branda, J. E. Gonz´alez-Pastor, S. Ben-Yehuda, R. Losick, and R. Kolter, "Fruiting
body formation by bacillus subtilis," Proceedings of the National Academy of Sciences,
vol. 98, no. 20, pp. 11621 -- 11626, 2001.
[37] J. Liu, R. Martinez-Corral, A. Prindle, D. L. Dong-yeon, J. Larkin, M. Gabalda-Sagarra,
J. Garcia-Ojalvo, and G. M. Suel, "Coupling between distant biofilms and emergence of
nutrient time-sharing," Science, vol. 356, pp. 638 -- 642, 2017.
[38] F. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W.
Clark, B. R. Miller, and B. V. Saunders, "NIST Digital Library of Mathematical Func-
tions." Release 1.0.18, 2018.
18
SUPPLEMENTARY MATERIAL
1 Model equations and description
As explained in the main text, we consider a one-dimensional array of simulation lattice sites
where biochemical species react and diffuse. Here we describe in detail the model equations.
1.1 Metabolic component
In the 'non-biofilm' sites, we consider media flow and diffusion of glutamate (G) and ammonium
(A) according to the following equations:
dGe
dt
dA
dt
= φ (Gm − Ge) + Dg∇2 Ge
= φ (Am − A) + Da∇2 A
(S1)
(S2)
In the microfluidics chamber, media is flowing constantly. Thus, we simulate the effect of the
media flow with the first term of each equation, such that with some rate φ, the concentration
of the chemical species tends to equate that in the medium (Xm). The second term models
diffusion.
In the biofilm, we assume that the diffusion coefficient and flow rate of glutamate decay
exponentially with the distance to the biofilm edge de, due to the extracellular matrix and high
cell density, according to the following functions:
Λφ =
ΛD =
exp(−γφ de) + aφ
1 + aφ
exp(−γD de) + aD
1 + aD
(S3)
(S4)
These functional forms are chosen such that at the edge Λx = 1, and flow and diffusion match
those in the media. The coefficients decay towards the biofilm interior, tending asymptotically
to a/(1 + a) in the centre, which ensures some remaining flow and diffusion. We follow the
original assumption [5] that ammonium diffusion over the biofilm is very fast, and do not apply
these reduction terms to this variable.
The dynamics of extracellular (Ge) and intracellular (Gi) glutamate in the biofilm are mod-
elled with the following dynamical equations:
dGe
dt
dGi
dt
= −αg F(V ) Rg
= αg F(V ) Rg
kg + Ge
Ge
kg + Ge
Ge
+ Λφ φ (Gm − Ge) + ΛD Dg∇2 Ge
− αa H Gi − δg Gi r ,
(S5)
(S6)
where, as mentioned, Ge is affected by media flow and diffusion. The first term in the right-hand
side of the two equations represents glutamate transport into the cells. As explained in the main
text, we consider that glutamate transport into the cell is modulated by the membrane potential
V , such that depolarisation reduces entry, and hyperpolarisation enhances it, according to the
functional form given in Eq. (2) of the main text.
19
In addition, we assume that glutamate is imported into the cells through the glutamate
transporter Rg, which saturates for large enough Ge, with half-maximum concentration kg. We
describe explicitly the dynamics of Rg by:
dRg
dt
= αR − δR Rg + βR
rnR
knR
R + rnR
.
(S7)
We thus assume that Rg is produced at a basal rate αR and degraded at a rate δR. The
last term accounts for the higher glutamate uptake by metabolically active cells, such that the
presence of biomass-producing biomolecules, such as ribosomal proteins, denoted by r, enhances
Rg synthesis via a Hill function with exponent nR.
Equation (S6) also assumes that intracellular glutamate concentration decays due to ammo-
nium production via the GDH enzyme (represented by H in the αa-term at the right-hand side
of the equation) and through various metabolic tasks including in particular biomass production
(δg-term).
The production of ammonium is regulated by the activity of the enzyme GDH. We describe
the dynamics of the inactive and active forms of this enzyme, h and H respectively, by the
equations:
dh
dt
dH
dt
αh
=
1 + (Gi/kh)nh
GnH
i
= αH h
H + GnH
knH
i
− αH h
GnH
i
knH
H + GnH
− γH H
i
− γh h + γH H
(S8)
(S9)
such that we account for synthesis and degradation of inactive GDH and its conversion into
the active form (αH -term in the two equations). As explained in the main text, we assume
that high concentrations of glutamate inhibit GDH synthesis, whereas activation is positively
regulated by glutamate via a Hill function with exponent nH . We also consider deactivation at
a constant rate γH .
Ammonium dynamics is affected by production from glutamate by active GDH, consumption
for various metabolic processes such as biomass production, and diffusion:
= αa H Gi − δa A r + Da∇2 A
dA
dt
(S10)
Finally, biomass production is considered to increase with ammonium and intracellular gluta-
mate, and to be subject to linear decay:
= βr A Gi − γr r
dr
dt
(S11)
1.2 Electrical signalling
Next we incorporate an adapted version of the electrical model introduced in [7]. As explained
in the main text, we consider an inhibitory effect of intracellular glutamate on a stress variable
S, whose production rate is modelled with an inhibitory Hill function:
1 +(cid:0) Gi
S0
Gs0
(cid:1)ns − γs S
dS
dt
=
(S12)
20
We explicitly consider both extracellular (Ke) and intracellular (Ki) potassium, whose dy-
namics are given by:
dKe
dt
dKi
dt
= F gK n4
= −F gK n4
k (V − VK) − Dp Gi Ke(Ki0 − Ki) + Λφ φ (Km − Ke) + ΛDDk∇2 Ke
k (V − VK) + Dp Gi Ke(Ki0 − Ki)
(S13)
(S14)
Potassium uptake is assumed to be governed by homeostatic processes that tend to keep its
intracellular concentration at a fixed value Ki0, described by the second term in the right-hand
side of Eqs. (S13)-(S14). Uptake is also made to depend on the metabolic state (glutamate
level), to account for the energy demand of the process. In addition, extracellular potassium
diffuses and is subject to the media flow in the chamber.
Potassium flow through its ion channel (first term in the right-hand side of the Ke and Ki
equations) is governed by the corresponding Nernst potential:
(cid:18) Ke
(cid:19)
VK = VK0 ln
,
Ki
(S15)
(S16)
and depends on the opening probability of the potassium channel, nk:
dnk
dt
=
a0 S
Sth + S
(1 − nk) − b nk
In the media lattice sites, as in the case of glutamate and ammonium [Eqs. (S1) and (S2)], the
extracellular potassium dynamics is affected by diffusion and by the media flow:
dKe
dt
= φ (Km − Ke) + Dk∇2 Ke
(S17)
The membrane potential dynamics is described by a Hodgkin-Huxley-like conductance-based
model containing potassium flux through the ion channel and a leak current [7]:
dV
dt
= −gK n4
k (V − VK) − gL (V − VL) ,
(S18)
where the leak potential VL is assumed to depend on the extracellular potassium [7] in a
threshold-linear manner, such that when Ke is larger than its basal level in the medium, Km,
the leak potential VL grows linearly (and the cell depolarizes), while when Ke < Km the leak
potential stays at its basal level:
VL = VL0 + dL
Ke − Km
1 − e−(Ke−Km)/σ
(S19)
Finally, we include the ThT reporter T downstream of the membrane potential, increasing when
the cells become hyperpolarised due to potassium release:
dT
dt
=
αT
1 + exp(gT (V − V0T ))
− γT T
(S20)
21
1.3 Simplified model
In the simplified model, we keep the same dynamics for the electrical part (Eqs. (S12)-(S20)).
The metabolic part is simplified as follows: Eqs. (S8)-(S11) are removed, the equations for
extracellular glutamate dynamics, in both 'biofilm' and 'non-biofilm' sites (Eqs. S1, S5), are
maintained, and Eqs. (S6) and (S7) become:
dGi
dt
dRg
dt
= αg F(V ) Rg
Ge
kg + Ge
= αR − δR Rg + αr
− δg Gi
Gnr
i
knr
r + Gnr
i
(S21)
(S22)
2 Supplementary tables and figures
Parameter
Description
Full model
Simplified
Units
αg
kg
Gm
Dg
αa
δg
αR
δR
βR
kR
nR
αh
kh
γh
nh
αH
kH
nH
γH
δa
Am
Da
glutamate uptake constant
extracellular glutamate concentration at half-
maximal uptake rate
glutamate concentration in the media
glutamate diffusion coefficient
ammonium production constant
glutamate degradation constant
glutamate receptor synthesis rate
glutamate receptor decay constant
maximum rate of glutamate receptor induc-
tion
threshold for Rg induction
Hill coefficient for Rg induction
maximal GDH synthesis rate
36.0
0.75
30.0
4e+06
4.5
0.525
6.75
36.0
45.0
5.0
2.0
0.075
threshold for inhibition of GDH synthesis by
1.5
glutamate
GDH decay rate
0.01
Hill coefficient for GDH synthesis inhibition
2.0
by glutamate
GDH activation constant
intracellular glutamate concentration for half-
maximal GDH activation
3.0
0.4
Hill coefficient for GDH activation by gluta-
2.0
mate
GDH deactivation constant
ammonium consumption constant
ammonium concentration in the media
ammonium diffusion coefficient
5.0
0.135
0.0
7e+06
model
24.0
0.75
30.0
4e+06
-
4.8
4.5
24.0
31.0
2.25
2.0
-
-
-
-
-
-
-
-
-
-
-
mM /(µM h)
mM
mM
µm2/h
µM−1h−1
mM−1h−1
µM/h
h−1
µM/h
mM
-
µM/h
mM
h−1
-
h−1
mM
-
h−1
mM−1h−1
mM
µm2/h
Table S1 -- Continued on next page
22
Parameter
Description
Full model
Simplified
Units
Table S1 -- Continued from previous page
βr
γr
S0
GS0
ns
γs
a0
Sth
b
F
gK
VK0
Dp
Ki0
Dk
Km
gL
dL
VL0
σ
gv
V0
αT
γT
gT
Pgrow
γφ
aφ
γD
aD
φ
model
-
-
1.12
0.2
2.0
2.8
91.0
0.03
34.0
0.05
70.0
100.0
0.12
-
7e+06
8.0
18.0
4.0
mM−1h−1
h−1
µM/h
mM
-
h−1
h−1
µM
h−1
mM/mV
h−1
mV
mM−2h−1
mM
µm2/h
mM
h−1
mV/mM
mV
mM
mV−1
mV
µM/h
h−1
mV−1
h−1
µm
-
µm
h−1
biomass-producing
biomolecules
synthesis
15.0
rate
biomass-producing biomolecules decay rate
maximum stress production rate
threshold for stress inhibition by glutamate
Hill coefficient for stress inhibition by gluta-
mate
stress decay constant
6.0
1.12
0.2
2.0
2.8
maximum rate of increase of the potassium
91.0
channel gating probability
stress level for half-maximal gating activity
opening probability decay constant
membrane capacitance
potassium channel conductance
Nernst potential prefactor
potassium uptake constant
intracellular potassium concentration
potassium diffusion coefficient
potassium concentration in the media
leak conductance
leak slope coefficient
basal leak potential
0.03
21.25
0.05
70.0
100.0
0.12
300.0
7e+06
8.0
18.0
4.0
-156.0
-156.0
leak threshold sharpness coefficient
inverse sensitivity of glutamate uptake to
membrane potential
resting membrane potential
maximal rate of ThT uptake
intracellular ThT decay constant
0.1
1.0
-150.0
20.0
10.0
inverse sensitivity of ThT to membrane poten-
0.3
tial
growth probability
0.3
spatial decay rate of the flow rate within the
0.0085
biofilm
basal flow factor
0.012
spatial decay rate of the diffusion coefficient
0.0085
within the biofilm
basal diffusion factor
flow rate
0.012
5.0
0.1
1.0
-150.0
20.0
10.0
0.3
0.5
0.0085
0.012
0.0085
-
5.0
Table S1: Parameter description and basal values for the two mod-
els.
23
Figure S1: Stop-flow triggers oscillations in the simplified model. Normalised ThT time traces
at the periphery (50 µm from the biofilm edge). A) Reference simulation. Vertical dashed line
indicates the time of stop-flow in B. (This simulation is the same as in Fig. 5B). B) The biofilm
was perturbed with a stop-flow-like perturbation (φ = 0 during 20 min at t = 9 hours). The
growth noise realisation is the same in both simulations, such that the only difference is the
perturbation.
Figure S2: The ∆trkA mutation in the model leads to impaired stress relief. For each simulation,
either the maximum or the minimum of the variable during the oscillations was computed for
the peripheral region (outermost 100 µm). Each dot represents a simulation, from the same
data as in Fig. 6 from the main text.
24
5101520time (h)0.00.51.0ThT (a.u.)5101520time (h)0.00.51.0ThT (a.u.)ABWTtrkA0.0500.0750.1000.125minnkWTtrkA15.017.520.022.5maxKe(mM)WTtrkA0.00060.0008minS(mM) |
1111.3610 | 1 | 1111 | 2011-11-15T19:01:36 | Scaling of brain metabolism and blood flow in relation to capillary and neural scaling | [
"q-bio.NC",
"physics.bio-ph",
"q-bio.CB",
"q-bio.TO"
] | Brain is one of the most energy demanding organs in mammals, and its total metabolic rate scales with brain volume raised to a power of around 5/6. This value is significantly higher than the more common exponent 3/4 relating whole body resting metabolism with body mass and several other physiological variables in animals and plants. This article investigates the reasons for brain allometric distinction on a level of its microvessels. Based on collected empirical data it is found that regional cerebral blood flow CBF across gray matter scales with cortical volume $V$ as $CBF \sim V^{-1/6}$, brain capillary diameter increases as $V^{1/12}$, and density of capillary length decreases as $V^{-1/6}$. It is predicted that velocity of capillary blood is almost invariant ($\sim V^{\epsilon}$), capillary transit time scales as $V^{1/6}$, capillary length increases as $V^{1/6+\epsilon}$, and capillary number as $V^{2/3-\epsilon}$, where $\epsilon$ is typically a small correction for medium and large brains, due to blood viscosity dependence on capillary radius. It is shown that the amount of capillary length and blood flow per cortical neuron are essentially conserved across mammals. These results indicate that geometry and dynamics of global neuro-vascular coupling have a proportionate character. Moreover, cerebral metabolic, hemodynamic, and microvascular variables scale with allometric exponents that are simple multiples of 1/6, rather than 1/4, which suggests that brain metabolism is more similar to the metabolism of aerobic than resting body. Relation of these findings to brain functional imaging studies involving the link between cerebral metabolism and blood flow is also discussed. | q-bio.NC | q-bio | Scaling of brain metabolism and blood flow in relation to
capillary and neural scaling
1
1
0
2
v
o
N
5
1
]
.
C
N
o
i
b
-
q
[
1
v
0
1
6
3
.
1
1
1
1
:
v
i
X
r
a
Jan Karbowski
Institute of Biocybernetics and Biomedical Engineering,
Polish Academy of Sciences, 02-109 Warsaw, Poland
Abstract
Brain is one of the most energy demanding organs in mammals,
and its total metabolic rate scales with brain volume raised to a power
of around 5/6. This value is significantly higher than the more com-
mon exponent 3/4 relating whole body resting metabolism with body
mass and several other physiological variables in animals and plants.
This article investigates the reasons for brain allometric distinction
on a level of its microvessels. Based on collected empirical data it
is found that regional cerebral blood flow CBF across gray matter
scales with cortical volume V as CBF ∼ V −1/6, brain capillary diam-
eter increases as V 1/12, and density of capillary length decreases as
V −1/6. It is predicted that velocity of capillary blood is almost in-
variant (∼ V ǫ), capillary transit time scales as V 1/6, capillary length
increases as V 1/6+ǫ, and capillary number as V 2/3−ǫ, where ǫ is typ-
ically a small correction for medium and large brains, due to blood
viscosity dependence on capillary radius. It is shown that the amount
of capillary length and blood flow per cortical neuron are essentially
conserved across mammals. These results indicate that geometry and
dynamics of global neuro-vascular coupling have a proportionate char-
acter. Moreover, cerebral metabolic, hemodynamic, and microvascular
variables scale with allometric exponents that are simple multiples of
1/6, rather than 1/4, which suggests that brain metabolism is more
similar to the metabolism of aerobic than resting body. Relation of
these findings to brain functional imaging studies involving the link
between cerebral metabolism and blood flow is also discussed.
1
Keywords: Brain metabolism; Cerebral blood flow; Capillary; Scaling; Mammals.
Email: [email protected]; [email protected]
2
Introduction
It is well established empirically that whole body metabolism of resting mammals
scales with body volume (or mass) with an exponent close to 3/4, which is known as
Kleiber's law [1, 2, 3, 4]. The same exponent or its simple derivatives govern the scalings
of respiratory and cardiovascular systems in mammals and some other physiological
parameters in animals and plants [2, 3, 5]. Because of its almost ubiquitous presence,
the quarter power has often been described as a general law governing metabolism and
blood circulation, and several formal models explaining its origin have been proposed
that still cause controversy [6, 7, 8, 9]. However, as was found by the author [10], the
brain metabolism at rest seems to follow another scaling rule. Total brain metabolic
rate (both oxygen and glucose) scales with brain volume with an exponent ≈ 0.85, or
close to 5/6 [10]. Consequently, the volume-specific cerebral metabolism decreases with
brain size with an exponent around −1/6, and this value is highly homogeneous across
many structures of gray matter [10]. The origin of these cerebral exponents has never
been explained, although it is interesting why brain metabolism scales different than
metabolism of other systems.
The brain, similar to other organs, uses capillaries for delivery of metabolic nutri-
ents (oxygen, glucose, etc.) to its cells [11]. Moreover, numerical density of cerebral
capillaries is strongly correlated with brain hemodynamics and metabolism [12, 13].
However, the cerebral microvascular network differs from other non-cerebral networks
in two important ways. First, in the brain there exists a unique physical border, called
the brain-blood barrier, which severely restricts influx of undesired molecules and ions
3
to the brain tissue. Second, cerebral capillaries exhibit a large degree of physical plastic-
ity, manifested in easy adaptation to abnormal physiological conditions. For instance,
during ischemia (insufficient amount of oxygen in the brain) capillaries can substantially
modify their diameter to increase blood flow and hence oxygen influx [14, 15, 16]. These
two factors, i.e. structural differences and plasticity of microvessels, can in principle
modify brain metabolism in such a way to yield different scaling rules in comparison
to e.g. lungs or muscles. Another, related factor that may account for the uncommon
brain metabolic scaling is the fact that brain is one of the most energy expensive organs
in the body [10, 17]. This is usually attributed to the neurons with their extended axons
and dendrites, which utilize relatively large amounts of glucose and ATP for synaptic
communication [18, 19].
The main purpose of this paper is to determine scaling laws for blood flow and
geometry of capillaries in the brain of mammals. Are they different from those found
or predicted for cardiovascular and respiratory systems? If so, do these differences
account for brain metabolic allometry? How the scalings of blood flow and capillary
dimensions relate to the scalings of neural characteristics, such as neural density and
axon (or dendrite) length? This study might have implications for expanding of our
understanding of mammalian brain evolution, in particular the relationship between
brain wiring, metabolism, and its underlying microvasculature [10, 20, 21]. The results
can also be relevant for research involving the microvascular basis of brain functional
imaging studies, which use relationships between blood flow and metabolism to decipher
regional neural activities [22, 23].
4
Results
The data for brain circulatory system were collected from different sources (see
Materials and Methods). They cover several mammals spanning 3-4 orders of magnitude
in brain volume, from mouse to human.
1. Empirical scaling data.
Cerebral blood flow CBF in different parts of mammalian gray matter decreases
systematically with gray matter volume, both in the cortical and subcortical regions
(Fig. 1).
In the cerebral cortex, the scaling exponent for regional CBF varies from
−0.13 for the visual cortex (Fig. 1A), −0.15 for the parietal cortex (Fig. 1B), −0.17 for
the frontal cortex (Fig. 1C), to −0.19 for the temporal cortex (Fig. 1D). The average
cortical exponent is −0.16 ± 0.02. In the subcortical regions, the CBF scaling exponent
is −0.14 for hippocampus (Fig. 2A), −0.17 for thalamus (Fig. 2B), and −0.18 for
cerebellum (Fig. 2C). The average subcortical exponent is identical with the cortical
one, i.e., −0.16 ± 0.02, and both of them are close to −1/6. It is interesting to note that
almost all of the cortical areas (except temporal cortex) have scaling exponents whose
95% confidence intervals do not include a quarter power exponent −1/4.
The microvessel system delivering energy to the brain consists of capillaries. The
capillary diameter increases very weakly but significantly with brain size, with an ex-
ponent of 0.08 (Fig. 3A). On the contrary, the volume-density of capillary length
decreases with brain size raised to a power of −0.16 (Fig. 3B). Thus, the cerebral
capillary network becomes sparser as brain size increases. Despite this, the fraction
of gray matter volume taken by capillaries is approximately independent of brain size
5
(Fig. 3C). Another vascular characteristic, the arterial partial oxygen pressure, is also
roughly invariant with respect to brain volume (Fig. 3D).
A degree of neurovascular coupling can be characterized by geometric relationships
between densities of capillaries and neurons. Scaling of the density of neuron number in
the cortical gray matter is not uniform across mammals [24, 25, 26]. In fact, the scaling
exponent depends to some extent on mammalian order and the animal sample used [26].
For the sample of mammals used in this study, it is found that cortical neuron density
decreases with cortical gray matter volume with an exponent of −0.13 (Fig. 4A). This
exponent is close to the exponent for the scaling of capillary length density, which is
−0.16 (Fig. 3B). Consistent with that, the ratio of cortical capillary length density to
neuron density across mammals is approximately constant and independent of brain
size (Fig. 4B). Typically, there is about 10 µm of capillaries per cortical neuron. The
scaling dependence between the two densities yields an exponent close to unity (Fig.
4C), which shows a proportionality relation between them.
Cerebral blood flow CBF scales with brain volume the same way as does capillary
length density (Figs. 1,2,3B), and thus, CBF should also be related to neural density.
Indeed, in the cerebral cortex the ratio of the average CBF to cortical neural density
is independent of brain scale (Fig. 5). This means that the average amount of cortical
blood flow per neuron is invariant among mammals, and about (1.45±0.4)10−8 mL/min.
Taken together, the findings in Figs. 4 and 5 suggest a tight global correlation between
neurons and their energy supporting microvascular network.
6
2. Theoretical scaling rules for cerebral capillaries.
Below I derive theoretical predictions for the allometry of brain capillary character-
istics, such as: capillary length and radius, capillary number, blood velocity, and time
taken by blood to travel through a capillary. I also find relationships connecting cere-
bral metabolic rate and blood flow with neuron density. The following assumptions are
made in the analysis: (i) Oxygen consumption rate in gray matter CMRO2 scales with
cortical gray matter volume V as V −1/6, in accordance with Ref.
[10]; (ii) Capillary
volume fraction, fc = πNcLcR2
c/V , is invariant with respect to V , which follows from
the empirical results in Fig. 3C. The symbol Nc denotes total capillary number in the
gray matter, Lc is the length of a single capillary segment, and Rc is its radius; (iii)
Driving blood pressure ∆pc through capillaries is independent of brain size, which is
consistent with a known fact that arterial blood pressure (both systolic and diastolic) of
resting mammals is independent of body size [27, 28, 29]; (iv) Partial oxygen pressure
pO2 in capillaries is also invariant, which is consistent with the empirical data in Fig.
3D on the invariance of arterial oxygen pressure; (v) Cerebral blood flow CBF is pro-
portional to oxygen consumption rate CMRO2, due to adaptation of capillary diameters
to oxygen demand.
The cerebral metabolic rate of oxygen consumption CMRO2, according to the modi-
fied Krogh model [11, 14], is proportional to the product of oxygen flux through capillary
wall and the tissue-capillary gradient of oxygen pressure ∆pO2, i.e.
CMRO2 ∼ D(cid:18)NcLc
V (cid:19) ∆pO2,
(1)
7
where D is the oxygen diffusion constant in the brain. The dependence of CMRO2
on capillary radius in this model has mainly a logarithmic character, and hence it is
neglected as weak. Since oxygen pressure in the brain tissue is very low [30], the pressure
gradient ∆pO2 is essentially equal to the capillary oxygen pressure pO2. Consequently,
the formula for CMRO2 simplifies to CMRO2 ∼ ρcpO2, where ρc is the density of capillary
length ρc = NcLc/V .
From the assumptions (i) and (iv) we obtain that capillary length density ρc ∼ V −1/6.
Additionally, from (ii) we have R2
c ∼ fc/ρc ∼ V 0/V −1/6 ∼ V 1/6, implying that capillary
radius (or diameter) Rc scales as V 1/12. Consequently capillary diameter does not
increase much with brain magnitude. As an example, a predicted capillary diameter for
elephant with its cortical gray matter volume 1379 cm3 [31] is 7.2 µm, which does not
differ much from those of rat (4.1 µm [15, 32]) or human (6.4 µm [33, 34]), who have
corresponding volumes 3450 and 2.4 times smaller.
The blood flow Qc through a capillary is governed by a modified Poiseuille's law in
which blood viscosity depends on capillary radius [35]:
Qc =
π∆pcR4
c
8ηef (Rc)Lc
,
(2)
where ∆pc is the axial driving blood pressure along a capillary of length Lc, and ηef (Rc)
is the capillary radius dependent effective blood viscosity. The latter dependence has a
nonmonotonic character, i.e. for small diameters the viscosity ηef (Rc) initially decreases
8
with increasing Rc, reaching a minimum at diameters about 5 − 7 µm. For 2Rc > 10
µm the blood viscosity ηef slowly increases with Rc approaching its bulk value for
diameters ∼ 500 µm. This phenomenon is known as the Fahraeus-Lindqvist effect [36].
In general, blood viscosity in narrow microvessels depends on microvessel thickness
because red blood cells tend to deform and place near the center of capillary leaving
a cell-free layer near the wall [35, 37]. These two regions have significantly different
viscosities, with the cell-free layer having essentially plasma viscosity ηp, which is much
smaller than the bulk (or center region) viscosity ηc. The formula relating the effective
blood viscosity ηef with capillary radius and both viscosities ηp and ηc is given by [35]:
ηef (Rc) =
ηp
1 − (1 − µ)(1 − w/Rc)4 ,
(3)
where µ = ηp/ηc ≪ 1, and w is the thickness of cell-free layer.
For capillary radiuses relevant for the brain, i.e. 1.5 µm < Rc < 3.5 µm (see Suppl.
Table S2), the ratio w/Rc increases with increasing Rc, which causes a decline in the
effective blood viscosity down to its minimal value at Rc = 3 − 3.5 µm (Table 1). Using
the data in Table 1 taken from [35], we can approximate the denominator in Eq. (3)
for this range of radiuses by a simple, explicit function of Rc. The best fit is achieved
with a logarithmic function, i.e. 1 − (1 − µ)(1 − w/Rc)4 ≈ 0.85[ln(Rc/Ro)]2/3, where
Ro = 1.2 µm (Table 1). As a result, the effective blood viscosity takes a simple form:
ηef (Rc) = 1.18ηp (ln(Rc/Ro))−2/3 .
(4)
9
Cerebral blood flow CBF in the brain gray matter is defined as CBF = Q/V , where
Q = NcQc is the total capillary blood flow through all Nc capillaries. Thus CBF is
given by
or
CBF ∼
∆pc
ηef (Rc)
ρcR4
c
L2
c
,
CBF ∼
∆pcρcR4
c
ηpL2
c
(cid:20)ln(cid:18) Rc
Ro(cid:19)(cid:21)
2/3
.
(5)
(6)
We can rewrite the logarithm present in Eq. (6), in an equivalent form, as a power
function (Rc/Ro)γ with a variable exponent γ given by (see Appendix S1 in the Supp.
Infor.):
so that CBF becomes
γ =
2
3
ln(ln(Rc/Ro))
ln(Rc/Ro)
,
CBF ∼
∆pc
ηp
ρcR4+γ
c
L2
c
.
(7)
(8)
The exponent γ in this equation can be viewed as a correction due to non-constant
10
blood viscosity (Fahraeus-Lindqvist effect [36]). The dependence of γ on the capillary
diameter is shown in Table 2. Because in general γ < 0, its presence in Eq.
(8)
reduces the power of Rc. However, this effect is weak for medium and large brains as
γ ≪ 1. Even for a small rat brain the relative influence of γ is rather weak, since
γ/4 ≈ 0.19. In contrast, for very small brains, such as mouse, the effect caused by γ
is strong (Table 2), which reflects a sharp increase in the effective blood viscosity for
the smallest capillaries [35, 37].
Now we are in a position to derive scaling rules for the capillary length segment
Lc, capillary blood velocity uc, and the number of capillaries Nc. From Eq. (8), using
the assumptions (i), (iii), and (v), we obtain L2
c ∼ ρcR4+γ
c
/CMRO2, which implies that
Lc ∼ R2+γ/2
c
(viscosity of blood plasma is presumably independent of brain scale [38]).
Consequently, Lc ∼ V 1/6+γ/24, i.e. capillary length should weakly increase with brain
size. Although there are no reliable data on Lc, we can compare our prediction with the
measured intercapillary distances, which generally should be positively correlated with
Lc. Indeed, the mean intercapillary distance in gray matter increases with increasing
brain volume, and is 17 − 24 µm in rat [39], 24 µm in cat [40], and 58 µm in human
[33].
Average velocity uc of blood flow in brain capillaries is given by uc = Qc/(πR2
c).
Using the expressions for Qc and ηef , we get uc ∼ (∆pcR2+γ
c
)/Lc. Since Lc ∼ R2+γ/2
c
above, and using the assumption (iii), we obtain uc ∼ Rγ/2
c ∼ V γ/24. Thus, capillary
blood velocity is almost independent of brain size for medium and large brains, as then
γ → 0 (Table 2). For very small brains, instead, there might be a weak dependence. A
11
related quantity, the blood transit time τc through a capillary, defined as τc = Lc/uc,
scales as τc ∼ V 1/6, regardless of the brain magnitude. This indicates that τc and
CBF are inversely related across different species, τc ∼ CBF−1, because of their scaling
properties.
We can find the scaling relation for the total number of capillaries Nc from the
volume-density of capillary length ρc. We obtain Nc = ρcV /Lc ∼ V −1/6V /V 1/6+γ/24 ∼
V 2/3−γ/24, i.e.
the exponent for Nc is close to 2/3 for not too small brains. As an
example, the number of capillary segments in the human cortical gray matter should
be 123 times greater than that in the rat (cortical volumes of both hemispheres in rat
and human are 0.42 cm3 [24] and 572.0 cm3 [41], respectively).
As was shown above, CMRO2 must be proportional to the volume density of capillary
length ρc (Eq. 1). On the other hand, the empirical results in Fig. 4 indicate that ρc
is roughly proportional to neuron density ρn. Thus, we have approximately CMRO2
∼ ρn across different mammals. This implies that oxygen metabolic energy per neuron
in the gray matter should be approximately independent of brain size. Exactly the
same conclusion was reached before in a study by Herculano-Houzel [26], based on
independent data analysis. Moreover, since cortical CMRO2 and CBF scale the same
way against brain size, we also have CBF ∼ ρn, which is confirmed by the results in
Fig. 5. In other words, both cerebral metabolic rate and blood flow per neuron are
scale invariant.
12
Discussion
1. General discussion.
The summary of the scaling results is presented in Table 3. Some of these allomet-
ric relations are directly derived from the experimental data (CBF, Rc, ρc, fc, ρc/ρn,
CBF/ρn), and others are theoretically deduced (Nc, Lc, uc, τc). The interesting result
is that cerebral blood flow CBF in gray matter scales with cortical gray matter volume
raised to a power of −0.16. The similar exponent governs the allometry of cortical
metabolic rate CMR [10], which indicates that brain metabolism and blood flow are
roughly linearly proportional across different mammals. This conclusion is compatible
with several published studies that have shown the proportionality of CMR and CBF
on a level of a single animal (rat, human) across different brain regions [12, 42].
The coupling between CMR and CBF manifests itself also in their relation to the
number of neurons.
In this respect, the present study extends the recent result of
Herculano-Houzel [26] about the constancy of metabolic energy per neuron in the brains
of mammals, by showing that also cerebral blood flow and capillary length per neuron
are essentially conserved across species. There are approximately 10 µm of capillar-
ies and 1.45 · 10−8 mL/min of blood flow per cortical neuron (Figs. 4 and 5; Supp.
Tables S2 and S3). This finding suggests that not only brain metabolism but also its
hemodynamics and microvascularization are evolutionarily constrained by the number
of neurons. This mutual coupling might be a result of optimization in the design of
cerebral energy expenditure and blood circulation.
It should be underlined that both CBF and CMR scale with brain volume with the
13
exponent about −1/6, which is significantly different from the exponent −1/4 relating
whole body resting specific metabolism with body volume [1, 2, 3]. Instead, the cerebral
exponent −1/6 is closer to an exponent −0.12 ± 0.02 characterizing maximal body
specific metabolic rate and specific cardiac output in strenuous exercise [43, 44].
In
this sense, the brain metabolism and its hemodynamics resemble more the metabolism
and circulation of exercised muscles than other resting organs, which is in line with the
empirical evidence that brain is an energy expensive organ [10, 17, 18]. This may also
suggest that there exists a common plan for the design of microcirculatory system in
different parts of the mammalian body that uses the same optimization principles [45].
The results of this study show that as brain increases in size its capillary network
becomes less dense, i.e. the densities of both capillary number and length decrease,
respectively as Nc/V ∼ V −1/3−γ/24 and ρc ∼ V −1/6 (Table 3). Contrary to that, the
capillary dimensions increase weakly with brain volume, their radius as Rc ∼ V 1/12 and
their length segment as Lc ∼ V 1/6+γ/24, which are sufficient to make the fraction fc
of capillary volume in the gray matter to be scale invariant (Table 3). The correction
γ/24 appearing in the scaling exponents for Nc/V and Lc reflects the fact that blood
viscosity depends on capillary radius (Fahraeus-Lindqvist effect [36]). This correction
is however small for sufficiently large brains, generally for brains larger or equal to that
of rat, for which typical values of γ/24 are in the range from −0.032 to −0.0004 (Table
2). On the contrary, for brains of mouse size or smaller, this correction is substantial,
about −0.15, which implies that for very small brains Lc is essentially constant.
Despite the changes in the geometry of microvessels, the velocity of capillary blood
14
uc is almost scale invariant for not too small brains (exponent γ/24 ≈ 0; Table 3). This
prediction agrees with direct measurements of velocity in the brains of mouse, rat, and
cat, which does not seem to change much, i.e.
it is in the range 1.5 − 2.2 mm/sec
[40, 46]. Consequently the transit time τc through a capillary increases with brain size
as τc ∼ V 1/6, i.e. the scaling exponent is again 1/6. Another variable that seems to be
independent of brain scale is partial oxygen pressure in cerebral capillaries (Table 3),
which is consistent with the empirical findings in Fig. 3D on the invariance of oxygen
pressure in arteries, as the two circulatory systems are mutually interconnected.
2. Capillary scaling in cerebral and non-cerebral tissue.
The above scaling results for the brain can be compared with available analogous
scaling rules for pulmonary, cardiovascular, and muscle systems. For these systems, it
was proposed (no direct measurements) that partial oxygen pressure in capillaries should
decline weakly with whole body volume (or organ volume as lung and heart volumes,
Vlung, Vheart, scale isometrically with body volume [2]) with an exponent around −1/12,
to account for the whole body specific metabolic exponent −1/4 [47, 48]. In the resting
pulmonary system, the capillary radius as well as the density of capillary length scale the
same way as they do in the brain, i.e., with the exponents 1/12 and −1/6, respectively,
against system's volume [49]. Also, the capillary blood velocity in cerebral and non-
cerebral tissues scale similarly, at least for not too small volumes, i.e. both are scale
invariant [2, 3] (Table 3). However, the number of capillaries and capillary length seem
to scale slightly different in the resting lungs, i.e. Nc ∼ V 5/8
lung and Lc ∼ V 5/24
lung
[47],
although the difference can be very mild. For the resting heart, it was predicted (again,
15
no direct measurements) that Nc ∼ V 3/4
heart, and blood transit time through a capillary
τ ∼ V 1/4
heart [48], i.e. the exponents are multiples of a quarter power and are slightly larger
than those for the brain (Table 3). Interestingly, for muscles and lungs in mammals
exercising at their aerobic maxima, the blood transit time scales against body mass
with an exponent close to 1/6 [50], which is the same as in the brain (Table 3). This
again suggests that brain metabolism is similar to the metabolism of other maximally
exercised organs. Overall, the small differences in the capillary characteristics among
cerebral and non-cerebral resting tissues might account for the observed differences in
the allometries of brain metabolism and whole body resting metabolism. In particular,
the prevailing exponent 1/6 found in this study for brain capillaries, instead of 1/4,
seems to be a direct cause for the distinctive brain metabolic scaling.
3. Brain microvascular network vs. neural network.
The interesting question from an evolutionary perspective is how the allometric
scalings for brain capillary dimensions relate to the allometry of neural characteristics.
The neural density ρn (number of cortical neurons Nn per cortical gray matter volume
V ) scales with cortical volume with a similar exponent as does the density of capillary
length ρc (Fig. 4A). Thus, as a coarse-grained global description we have approximately
ρn ∼ ρc (Fig. 4B,C), or Nn ∼ NcLc. The latter relation means that the total number
of neurons is roughly proportional to the total length of capillaries, or equivalently,
that capillary length per cortical neuron is conserved across different mammals. This
cross-species conclusion is also in agreement with the experimental data for a single
species. In particular, for mouse cerebral cortex it was found that densities of neural
16
number and microvessel length are correlated globally across cortical areas (but not
locally within a single column) [51]. Moreover, since axons and dendrites occupy a
constant fraction of cortical gray matter volume (roughly 1/3 each; [52, 53]), we have
Nnld2 ∼ V , where l and d are respectively axon (or dendrite) length per neuron and
diameter. Furthermore, because the average axon diameter d (unmyelinated) in the
cortical gray matter is approximately invariant against the change of brain scale [52, 54],
we obtain the following chain of proportionalities: l ∼ ρ−1
c , where
n ∼ ρ−1
c ∼ V 1/6 ∼ Lα
the exponent α = 1/(1 + γ/4). For medium and large brains, α ≈ 1, implying a nearly
proportional dependence of axonal and dendritic lengths on capillary segment length.
For very small brains (roughly below the volume of rat brain), α can be substantially
greater than 1, suggesting a non-linear dependence between capillary and neural sizes.
Given that the main exchange of oxygen between blood and brain takes place in
the capillaries, these results suggest that metabolic needs of larger brains with greater
but numerically sparser neurons must be matched by appropriately longer yet sparser
capillaries. This finding reflects a rough, global relationship, which might or might not
be related to the fact that during development neural and microvessel wirings share
mutual mechanisms [20, 55]. At the cortical microscale, however, things could be more
complicated, and a neuro-vascular correlation might be weaker, as both systems are
highly plastic even in the adult brain (e.g.
[56]). Regardless of its nature and precise
dependence, the neuro-vascular coupling might be important for optimization of neural
wiring [53, 57, 58]. In fact, neural connectivity in the cerebral cortex is very low, and
it decreases with brain size [58, 59], similar to the density of capillary length (Fig. 3B,
17
Table 3). To make the neural connectivity denser, it would require longer axons and
consequently longer capillaries. That may in turn increase excessively brain volume
and its energy consumption, i.e.
the costs of brain maintenance. As a result, the
metabolic cost of having more neural connections and synapses for storing memories
might outweigh its functional benefit.
The brain metabolism is obviously strictly related to neural activities. In general,
higher neural firing rates imply more cerebral energy consumed [18, 19]. It was esti-
mated, based on a theoretical formula relating CMR with firing rate, that the latter
should decline with brain size with an exponent around −0.15 [19]. This implies that
neurons in larger brain are on average less active than neurons in smaller brains. Such
sparse neural representations may be advantageous in terms of saving the metabolic en-
ergy [18, 60, 61]. At the same time, what may be related, neural activity is distributed
in such a way that both the average energy per neuron and the average blood flow per
neuron are approximately invariant with respect to brain size (Fig. 5; Table 3, [26]).
Additionally, average firing rate should be inversely proportional to the average blood
transit time τc through a capillary, because both of them scale reversely with brain size
(Table 3). Thus, it appears that global timing in neural activities should be correlated
with the timing of cerebral blood flow. These general considerations suggest that apart
from structural neuro-vascular coupling there is probably also a significant dynamic
coupling. This conclusion is qualitatively compatible with experimental observations
in which enhanced neural activity is invariably accompanied by increase in local blood
flow [62].
18
4. Relationship to brain functional imaging.
The interdependencies between brain metabolism, blood flow, and capillary param-
eters can have practical meaning. Currently existing techniques for non-invasive visu-
alization of brain function, such as PET or fMRI, are associated with measurements of
blood flow CBF and oxygen consumption CMRO2. It turns out that during stimulation
of a specific brain region, CBF increases often, but not always, far more than CMRO2
[63]. However, both of them increase only by a small fraction in relation to the back-
ground activity, even for massive stimulation [62, 63]. This phenomenon was initially
interpreted as an uncoupling between blood perfusion and oxidative metabolism [64].
Later, it was shown that this asymmetry between CBF and CMRO2 can be explained
in terms of mechanistic limitations on oxygen delivery to brain tissue through blood
flow [65]. We can provide a related, but simpler explanation of these observations that
involves physical limitations on the relative changes in capillary oxygen pressure and
radius.
During brain stimulation, both CBF and CMRO2 change by δCBF and δCMRO2,
which are according to Eqs. (1) and (8) related to modifications in capillary radius
(from Rc to Rc + δRc), and changes in partial oxygen pressure (pO2 7→ pO2 + δpO2). The
density of perfused capillary length ρc remains constant for normal neurophysiological
conditions. Accordingly, a small fraction of blood flow change is
δCBF
CBF
≈ (4 + γ)
δRc
Rc
(9)
19
and similarly, a small fractional change in the oxygen metabolic rate is:
δCMRO2
CMRO2
≈
δpO2
pO2
.
(10)
In general, oxygen pressure increases with increasing capillary radius, in response to
increase in blood flow CBF. This relationship can have a complicated character. We
simply assume that pO2 ∼ Ra
c , where the unknown exponent a (a > 0) contains all the
non-linear effects, however complicated they are. Thus, a small fractional change in
oxygen pressure can be written as δpO2/pO2 ≈ aδRc/Rc. As a result, we obtain
δCMRO2
CMRO2
≈
a
(4 + γ)
δCBF
CBF
.
(11)
If partial oxygen pressure pO2 depends on capillary radius linearly or sublinearly, i.e.,
if a ≤ 1, then the fractional increase in oxygen metabolism is significantly smaller
than a corresponding increase in cerebral blood flow. This case corresponds to the
experimental reports showing that this ratio is ≪ 1, for example, in the visual cortex
(∼ 0.1) [66] and in the sensory cortex (∼ 0.2 − 0.4) [64, 67]. If, in turn, pO2 depends
on Rc superlinearly, i.e.
if a > 1, then the coefficient a/(4 + γ) in Eq. (10) can be of
the order of unity. Such cases have been also reported experimentally during cognitive
activities [42] or anesthesia [68, 69].
20
Materials and Methods
The ethics statement does not apply to this study. CBF data were collected from
different sources: for mouse [70], rat [71], rabbit [72], cynomolgus monkey [73], rhesus
monkey [74], pig [75], and human [76]. Cerebral capillary characteristics were obtained
from several sources: for mouse [14, 51], rat [15, 77, 32], cat [40, 78], dog [79], rhesus
monkey [80], and human [33, 34]. Data for calculating neuron densities were taken
from [24, 25, 41, 52, 81, 82]. Cortical volume data (for 2 hemispheres) are taken from
[52, 41, 81]. Their values are: mouse 0.12 cm3, rat 0.42 cm3, rabbit 4.0 cm3, cat 14.0
cm3, cynomolgus monkey 21.0 cm3, dog 35.0 cm3, rhesus monkey 42.9 cm3, pig 45.0 cm3,
human 571.8 cm3. All the numerical data are provided in the Supporting Information
(Tables S1, S2, and S3).
21
Supporting Information
Appendix S1
Table S1
Regional cerebral blood flow CBF in mammals.
Table S2
Cerebral capillary and neural characteristics in mammals.
Table S3
Arterial partial oxygen pressure and average cortical CBF per neuron.
22
References
[1] Kleiber M (1947) Body size and metabolic rate. Physiol. Rev. 27: 511-
541.
[2] Schmidt-Nielsen K (1984) Scaling: why is animal size so important?
Cambridge: Cambridge Univ. Press.
[3] Calder WA (1984) Size, function, and life history. Cambridge, MA: Har-
vard Univ. Press.
[4] Dodds PS, Rothman DH, Weitz JS (2001) Re-examination of the '3/4-
law' of metabolism. J. Theor. Biol. 209: 9-27.
[5] Enquist BJ, West GB, Brown JH (2000) Quarter-power allometric scal-
ing in vascular plants: functional basis and ecological consequences. In
Scaling in biology, ed. Brown JH, West GB, pp. 113-128. Oxford: Oxford
Univ. Press.
[6] West GB, Brown JH, Enquist BJ (1997) A general model for the origin
of allometric scaling laws in biology. Science 276: 122-126.
[7] Banavar JR, Damuth J, Maritan A, Rinaldo A (2002) Supply-demand
balance and metabolic scaling. Proc. Natl. Acad. Sci. USA 99: 10506-
10509.
23
[8] Darveau C, Suarez R, Andrews R, Hochachka P (2002) Allometric cas-
cade as a unifying principle of body mass effects on metabolism. Nature
417: 166-170.
[9] Savage VM, Deeds EJ, Fontana W (2008) Sizing up allometric scaling
theory. PLoS Comput. Biol. 4(9): e1000171.
[10] Karbowski J (2007) Global and regional brain metabolic scaling and its
functional consequences. BMC Biology 5: 18.
[11] Krogh A (1929) The anatomy and physiology of capillaries, 2nd ed. New
Haven, CT: Yale Univ. Press.
[12] Klein B, Kuschinsky W, Schrock H, Vetterlein F (1986) Interdependency
of local capillary density, blood flow, and metabolism in rat brain. Am.
J. Physiol. 251: H1333-H1340.
[13] Borowsky IW, Collins RC (1989) Metabolic anatomy of brain: a com-
parison of regional capillary density, glucose metabolism, and enzyme
activities. J. Comp. Neurol. 288: 401-413.
[14] Boero JA, Ascher J, Arregui A, Rovainen C, Woolsey TA (1999) In-
creased brain capillaries in chronic hypoxia. J. Appl. Physiol. 86: 1211-
1219.
[15] Hauck EF, Apostel S, Hoffmann JF, Heimann A, Kempski O (2004) Cap-
illary flow and diameter changes during reperfusion after global cerebral
24
ischemia studied by intravital video microscopy. J. Cereb. Blood Flow
Metab. 24: 383-391.
[16] Ito H, Kanno I, Ibaraki M, Hatazawa J, Miura S (2003) Changes in hu-
man cerebral blood flow and cerebral blood volume during hypercapnia
and hypocapnia measured by positron emission tomography. J. Cereb.
Blood Flow Metab. 23: 665-670.
[17] Aiello LC, Wheeler P (1995) The expensive-tissue hypothesis: The brain
and the digestive system in human and primate evolution. Curr. Anthro-
pology 36: 199-221.
[18] Attwell D, Laughlin SB (2001) An energy budget for signaling in the
gray matter of the brain. J. Cereb. Blood Flow Metabol. 21: 1133-1145.
[19] Karbowski J (2009) Thermodynamic constraints on neural dimensions,
firing rates, brain temperature and size. J. Comput. Neurosci. 27: 415-
436.
[20] Carmeliet P, Tessier-Lavigne M (2005) Common mechanisms of nerve
and blood vessel wiring. Nature 436: 193-200.
[21] Attwell D, Gibb A (2005) Neuroenergetics and the kinetic design of
excitatory synapses. Nat. Rev. Neurosci. 6: 841-849.
[22] Heeger DJ, Ress D (2002) What does fMRI tell us about neuronal ac-
tivity? Nat. Rev. Neurosci. 3: 142-151.
25
[23] Logothetis NK, Wandell BA (2004) Interpreting the BOLD signal. Annu.
Rev. Physiol. 66: 735-769.
[24] Herculano-Houzel S, Mota B, Lent R (2006) Cellular scaling rules for
rodent brains. Proc. Natl. Acad. Sci. USA 103: 12138-12143.
[25] Herculano-Houzel S, Collins CE, Wong P, Kaas JH (2007) Cellular scal-
ing rules for primate brains. Proc. Natl. Acad. Sci. USA 104: 3562-3567.
[26] Herculano-Houzel S (2011) Scaling of brain metabolism with a fixed
energy budget per neuron: Implications for neuronal activity, plasticity,
and evolution. PLoS ONE 6(3): e17514.
[27] Woodbury RA, Hamilton WF (1937) Blood pressure studies in small
animals. Am. J. Physiol. 119: 663-674.
[28] Gregg DE, Eckstein RW, Fineberg MH (1937) Pressure pulses and blood
pressure values in unanesthetized dogs. Am. J. Physiol. 118: 399-410.
[29] Li JK-J (2000) Scaling and invariants in Cardiovascular biology. In Scal-
ing in biology, ed. Brown JH, West GB, pp. 113-128. Oxford: Oxford
Univ. Press.
[30] Lenigert-Follert E, Lubbers DW (1976) Behavior of microflow and local
PO2 of the brain cortex during and after electrical stimulation. Pflugers
Arch. 366: 39-44.
26
[31] Hakeem AY, Hof PR, Sherwood CC, Switzer RC, Rasmussen LEF, All-
man JM (2005) Brain of the african elephant (Loxodonta africana):
Neuroanatomy from magnetic resonance images. Anat. Rec. A 287 A:
1117-1127.
[32] Michaloudi H, Batzios C, Grivas I, Chiotelli M, Papadopoulos GC (2006)
Developmental changes in the vascular network of the rat visual areas
17, 18, and 18a. Brain Res. 1103: 1-12.
[33] Meier-Ruge W, Hunziker O, Schulz U, Tobler HJ, Schweizer A (1980)
Stereological changes in the capillary network and nerve cells of the aging
human brain. Mechanisms of Ageing and Development 14: 233-243.
[34] Lauwers F, Cassot F, Lauwers-Cances V, Puwanarajah P, Duvernoy
H (2008) Morphometry of the human cerebral cortex microcirculation:
General characteristics and space-related profiles. NeuroImage 39: 936-
948.
[35] Sugihara-Seki M, Fu BM (2005) Blood flow and permeability in mi-
crovessels. Fluid Dynamics Research 37: 82-132.
[36] Fahraeus R, Lindqvist T (1931) The viscosity of the blood in narrow
capillary tubes. Am. J. Physiol. 96: 562-568.
[37] Pries AR, Secomb TW, Gaehtgens P (1996) Biophysical aspects of blood
flow in the microvasculature. Cardiovasc. Res. 32: 654-667.
27
[38] Amin TM, Sirs JA (1985) The blood rheology of man and various animal
species. Q. J. Exp. Physiol. 70: 37-49.
[39] Schlageter KE, Molnar P, Lapin GD, Groothuis DR (1999) Microvessel
organization and structure in experimental brain tumors: microvessel
populations with distinctive structural and functional properties. Mi-
crovascular Res. 58: 312-328.
[40] Pawlik G, Rackl A, Bing RS (1981) Quantitative capillary topography
and blood flow in the cerebral cortex of cat: an in vivo microscopic
study. Brain Res. 208: 35-58.
[41] Herculano-Houzel S, Mota B, Wong P, Kaas JH (2010) Connectivity-
driven white matter scaling and folding in primate cerebral cortex. Proc.
Natl. Acad. Sci. USA 107: 19008-19013.
[42] Roland PE, Eriksson L, Stone-Elander S, Widen L (1987) Does mental
activity change the oxidative metabolism of the brain? J. Neurosci. 7:
2373-2389.
[43] Bishop CM (1999) The maximum oxygen consumption and aerobic
scope of birds and mammals: getting to the heart of the matter. Proc.
R. Soc. Lond. B 266: 2275-2281.
[44] Weibel ER, Hoppeler H (2005) Exercise-induced maximal metabolic rate
scales with muscle aerobic capacity. J. Exp. Biol. 208: 1635-1644.
28
[45] Weibel ER, Taylor CR, Hoppeler H (1991) The concept of symmorpho-
sis: a testable hypothesis of structure-function relationship. Proc. Natl.
Acad. Sci. USA 88: 10357-10361.
[46] Unekawa M, Tomita M, Tomita Y, Toriumi H, Miyaki K, Suzuki N
(2010) RBC velocities in single capillaries of mouse and rat brains are
the same, despite 10-fold difference in body size. Brain Res. 1320: 69-73.
[47] Dawson TH (2003) Scaling laws for capillary vessels of mammals at rest
and in exercise. Proc. R. Soc. Lond. B 270: 755-763.
[48] West GB, Brown JH, Enquist BJ (2000) The origin of universal scaling
laws in biology. In Scaling in biology, ed. Brown JH, West GB, pp. 87-
112. Oxford: Oxford Univ. Press.
[49] Dawson TH (2008) Modeling the vascular system and its capillary net-
works. In Vascular Hemodynamics: Bioengineering and Clinical Per-
spectives. Edited by Yim PJ. New York: Wiley.
[50] Kayar SR, Hoppeler H, Jones JH, Longworth K, Armstrong RB, et al
(1994) Capillary blood transit time in muscles in relation to body size
and aerobic capacity. J. Exp. Biol. 194: 69-81.
[51] Tsai PS, Kaufhold JP, Blinder P, Friedman B, Drew PJ, et al (2009)
Correlations of neuronal and microvascular densities in murine cortex
29
revealed by direct counting and colocalization of nuclei and vessels. J.
Neurosci. 29: 14553-14570.
[52] Braitenberg V, Schuz A (1998) Cortex: Statistics and Geometry of Neu-
ronal Connectivity. Berlin: Springer.
[53] Chklovskii DB, Schikorski T, Stevens CF (2002) Wiring optimization in
cortical circuits. Neuron 43: 341-347.
[54] Olivares R, Montiel J, Aboitiz F (2001) Species differences and simi-
larities in the fine structure of the mammalian corpus callosum. Brain
Behav. Evol. 57: 98-105.
[55] Stubbs D, DeProto J, Nie K, Englund C, Mahmud I, et al (2009) Neu-
rovascular congruence during cerebral cortical development. Cereb. Cor-
tex 19: 32-41.
[56] Chklovskii DB, Mel BW, Svoboda K (2004) Cortical rewiring and infor-
mation storage. Nature 431: 782-788.
[57] Mitchison, G (1992) Neuronal branching patterns and the economy of
cortical wiring. Proc. R. Soc. Lond. B 245: 151-158.
[58] Karbowski J (2001) Optimal wiring principle and plateaus in the degree
of separation for cortical neurons. Phys. Rev. Lett. 86: 3674-3677.
30
[59] Karbowski J (2003) How does connectivity between cortical areas de-
pend on brain size? Implications for efficient computation. J. Comput.
Neurosci. 15: 347-356.
[60] Levy WB, Baxter RA (1996) Energy efficient neural codes. Neural Com-
put. 8: 531-543.
[61] Laughlin SB, de Ruyter van Steveninck RR, Anderson JC (1998) The
metabolic cost of neural information. Nature Neurosci. 1: 36-41.
[62] Moore CI, Cao R (2008) The hemo-neural hypothesis: on the role of
blood flow in information processing. J. Neurophysiol. 99: 2035-2047.
[63] Raichle ME, Mintun MA (2006) Brain work and brain imaging. Annu.
Rev. Neurosci. 29: 449-476.
[64] Fox PT, Raichle ME (1986) Focal physiological uncoupling of cerebral
blood flow and oxidative metabolism during somatosensory stimulation
in human subjects. Proc. Natl. Acad. Sci. USA 83: 1140-1144.
[65] Buxton RB, Frank LR (1997) A model of the coupling between cerebral
blood flow and oxygen metabolism during neural stimulation. J. Cereb.
Blood Flow Metab. 17: 64-72.
[66] Fox PT, Raichle ME, Mintun MA, Dence C (1988) Nonoxidative glucose
consumption during focal physiologic neural activity. Science 241: 462-
464.
31
[67] Seitz RJ, Roland PE (1992) Vibratory stimulation increases and de-
creases the regional cerebral blood flow and oxidative metabolism: a
positron emission tomography (PET) study. Acta Neurol. Scand. 86:
60-67.
[68] Nilsson B, Siesjo BK (1975) The effect of phenobarbitone anaesthesia on
blood flow and oxygen consumption in the rat brain. Acta Anaesthesiol.
Scand. Suppl. 57: 18-24.
[69] Smith AL, Wollman H (1972) Cerebral blood flow and metabolism.
Anesthesiology 36: 378-400.
[70] Frietsch T, Maurer MH, Vogel J, Gassmann M, Kuschinsky W, et
al (2007) Reduced cerebral blood flow but elevated cerebral glucose
metabolic rate in erythropoietin overexpressing transgenic mice with
excessive erythrocytosis. J. Cereb. Blood Flow Metab. 27: 469-476.
[71] Frietsch T, Krafft P, Piepgras A, Lenz C, Kuschinsky W, et al (2000)
Relationship between local cerebral blood flow and metabolism during
mild and moderate hypothermia in rats. Anesthesiology 92: 754-763.
[72] Tuor UI (1991) Local cerebral blood flow in the newborn rabbit: an
autoradiographic study of changes during development. Pediatric Res.
29: 517-523.
32
[73] Orlandi C, Crane PD, Platts SH, Walovitch RC (1990) Regional cere-
bral blood flow and distribution of [99mTc]Ethyl Cysteinate dimer in
nonhuman primates. Stroke 21: 1059-1063.
[74] Noda A, Ohba H, Kakiuchi T, Futatsubashi M, Tsukada H, et al (2002)
Age-related changes in cerebral blood flow and glucose metabolism in
conscious rhesus monkeys. Brain Res. 936: 76-81.
[75] Delp MD, Armstrong RB, Godfrey DA, Laughlin MH, Ross CD, et al
(2001) Exercise increases blood flow to locomotor, vestibular, cardiores-
piratory and visual regions of the brain in miniature swine. J. Physiol.
533.3 849-859.
[76] Bentourkia M, Bol A, Ivanoiu A, Labar D, Sibomana M, et al (2000)
Comparison of regional cerebral blood flow and glucose metabolism in
the normal brain: effect of aging. J. Neurol. Sci. 181: 19-28.
[77] Bar TH (1980) The vascular system of the cerebral cortex. Adv. Anat.
Embryol. Cell Biol. 59: 71-84.
[78] Tieman SB, Moller S, Tieman DG, White JT (2004) The blood supply
of the cat's visual cortex and its postnatal development. Brain Res. 998:
100-112.
33
[79] Luciano MG, Skarupa DJ, Booth AM, Wood AS, Brant CL, et al (2001)
Cerebrovascular adaptation in chronic hydrocephalus. J. Cereb. Blood
Flow Metab. 21: 285-294.
[80] Weber B, Keller AL, Reichold J, Logothetis NK (2008) The microvascu-
lar system of the striate and extrastriate visual cortex of the macaque.
Cereb. Cortex 18: 2318-2330.
[81] Mayhew TM, Mwamengele GLM, Dantzer V (1996) Stereological and
allometric studies on mammalian cerebral cortex with implications for
medical brain imaging. J. Anat. 189: 177-184.
[82] Haug H (1987) Brain sizes, surfaces, and neuronal sizes of the cortex
cerebri: A stereological investigation of Man and his variability and a
comparison with some mammals (primates, whales, marsupials, insecti-
vores, and one elephant). Am. J. Anatomy 180: 126-142.
[83] Hatazawa J et al (1995) Regional cerebral blood flow, blood volume,
oxygen extraction fraction, and oxygen utilization rate in normal volun-
teers measured by the autoradiographic technique and the single breath
inhalation method. Ann. Nucl. Med. 9: 15-21.
[84] Heistad DD et al (1976) Effect of stimulation of carotid chemoreceptors
on total and regional cerebral blood flow. Cir. Res. 38: 20-25.
34
Figure Captions
Fig. 1
Scaling of cerebral blood flow CBF in the cortical gray matter.
(A) Visual cortex:
y = −0.127x + 1.98 (R2 = 0.93, p < 0.001). 95% confidence interval for the slope CI=(-
0.168,-0.086). (B) Parietal cortex: y = −0.150x + 2.00 (R2 = 0.89, p = 0.005), slope
CI=(-0.222,-0.078). (C) Frontal cortex: y = −0.170x + 2.00 (R2 = 0.89, p = 0.002),
slope CI=(-0.239,-0.100).
(D) Temporal cortex: y = −0.191x + 2.09 (R2 = 0.89,
p = 0.005), slope CI=(-0.286,-0.096).
Fig. 2
Scaling of cerebral blood flow CBF in the subcortical gray matter. (A) Hippocampus:
y = −0.135x + 1.89 (R2 = 0.90, p = 0.049), slope CI=(-0.271,0.000). (B) Thalamus:
y = −0.167x + 2.02 (R2 = 0.83, p = 0.011), slope CI=(-0.272,-0.062). (C) Cerebellum:
y = −0.177x + 2.03 (R2 = 0.88, p = 0.002), slope CI=(-0.252,-0.102).
Fig. 3
Scaling of brain capillary characteristics against brain size.
(A) Capillary diameter
scales against cortical gray matter volume with the exponent 0.075 (y = 0.075x + 0.60,
R2 = 0.87, p = 0.007), exponent CI=(0.034,0.117). (B) Volume density of capillary
length ρc scales with the exponent −0.16 (y = −0.162x + 2.86, R2 = 0.79, p = 0.044),
exponent CI=(-0.316,-0.008).
(C) Fraction of capillary volume fc in gray matter is
essentially independent of brain size (y = 0.029x−1.83, R2 = 0.07, p = 0.662), the same
as (D) the arterial partial oxygen pressure (y = 0.012x + 1.96, R2 = 0.09, p = 0.472).
35
Fig. 4
Neuron density versus capillary length density in the cerebral cortex. (A) Across our
sample of mammals, the cortical neuron number density ρn scales against cortical vol-
ume with the exponent −0.13 (y = −0.128x + 1.86, R2 = 0.87, p = 0.022), exponent
CI=(-0.221,-0.036). (B) The ratio of the density of capillary length ρc to the density
of neurons ρn in the cortex does not correlate with brain size (y = −0.034x + 0.99,
R2 = 0.09, p = 0.617), exponent CI=(-0.228,0.161).
(C) The log-log dependence
of the capillary length density ρc on neuron density ρn gives the exponent of 1.05
(y = 1.051x + 0.88, R2 = 0.63, p = 0.109).
Fig. 5
Invariance of cerebral blood flow per cortical neuron across mammals. The ratio of
CBF to neuron density ρn in the cerebral cortex does not correlate significantly with
brain size (log-log plot yields y = −0.033x + 0.183, R2 = 0.39, p = 0.261). The value
of CBF for each species is the arithmetic mean of regional CBF across cerebral cortex.
36
Table 1: Parameters affecting the effective blood viscosity.
Rc [µm] w [µm] w/Rc
0.05
1.5
0.15
2.0
0.24
2.5
3.0
0.30
0.07
0.30
0.60
0.90
1 − (1 − µ)(1 − w/Rc)4
0.27
0.54
0.71
0.79
0.85(ln(Rc/1.2))2/3
0.31
0.54
0.69
0.80
Data for w and Rc were collected from [35]. The value of µ was taken as 1/8.
The last column represents values of the fitting function to the function in
the fourth column.
Table 2: Exponent γ as a function of capillary diameter.
dog monkey human
Species
4.5
2Rc [µm]
γ
-0.49
mouse
3.1
-3.51
6.4
-0.01
rat
4.1
-0.77
cat
5.1
-0.25
5.6
-0.13
Data for Rc were taken from Suppl. Inform. Table S2 (references therein).
Table S1: Regional cerebral blood flow CBF in mammals.
Species
Mouse (a)
Rat (b)
Rabbit (c)
Cynomolgus
monkey (d)
Rhesus monkey (e)
Pig (f )
Human (g)
visual
frontal
cortex cortex
165.0
124.0
124.0
116.0
69.6
67.2
temporal parietal hippocam.
cortex
157.0
195.0
--
cortex
166.0
113.0
63.6
112.8
77.8
--
55.4
59.2
66.5
43.0
48.7
45.0
63.3
41.2
70.6
53.2
45.5
44.6
56.8
--
62.8
43.4
--
53.9
41.0
--
thalamus
cerebell.
195.5
116.0
60.6
56.7
--
48.5
47.5
195.0
123.0
62.4
57.2
47.8
64.0
41.5
CBF data are given in mL/(100g*min).
References: (a) [70]; (b) [71]; (c) [72]; (d) [73]; (e) [74]; (f ) [75]; (g) [76].
37
Table 3: Summary of scalings for brain capillaries and hemodynamics against
cortical gray matter volume V .
Parameter
Capillary radius, Rc
Capillary length density, ρc
Capillary volume fraction, fc
Total capillary number, Nc
Capillary segment length, Lc
Capillary blood velocity, uc
Capillary transit time, τc
Capillary oxygen pressure, pO2
Capillary length per neuron, ρc/ρn
Cerebral blood flow, CBF
Blood flow per neuron, CBF/ρn
Oxygen consumption rate, CMRO2 CMRO2 ∼ V −1/6
Oxygen use per neuron, CMRO2/ρn CMRO2/ρn ∼ V 0
Scaling rule
Rc ∼ V 1/12
ρc ∼ V −1/6
fc ∼ V 0
Nc ∼ V 2/3−γ/24
Lc ∼ V 1/6+γ/24
uc ∼ V γ/24
τc ∼ V 1/6
pO2 ∼ V 0
ρc/ρn ∼ V 0
CBF ∼ V −1/6
CBF/ρn ∼ V 0
38
Table S2: Cerebral capillary and neural characteristics in mammals.
ρc/ρn
Species
[µm]
7.59
11.29
15.60
0.015 (b)
0.014 (i)
0.021 (d)
fc
2Rc
[µm]
3.1 (a,b)
4.1 (c)
5.1 (d)
4.5 (e) --
Mouse
Rat
Cat
Dog
Rhesus monkey 5.6 (f )
Human
6.4 (g,h)
0.009 (f )
0.023 (g)
ρc
[mm/mm3]
880.0 (a,b)
806.0 (i)
780.0 (j)
--
329.9 (f )
219.0 (g)
ρn
[106/cm3]
116.0 (k)
71.4 (l)
50.0 (m)
12.6 (m) --
37.4 (n)
38.5 (n)
8.82
5.69
The ratio ρc/ρn is the average capillary length per neuron.
References: (a) [14]; (b) [51]; (c) [32]; (d) [40]; (e) [79]; (f ) [80]; (g) [33]; (h) [34]; (i) [77]; (j)
[78]; (k) [52]; (l) [24]; (m) [82]; (n) [41].
39
Table S3: Arterial partial oxygen pressure and average cortical CBF.
arterial pO2 [mm Hg] CBFcortex CBFcortex/ρn
93.0 (a)
93.0 (b)
75.0 (c)
157.4 (a)
131.8 (b)
70.2 (c)
Species
Mouse
Rat
Rabbit
Cynomolgus
102.4 (d)
monkey
107.0 (e)
Dog
Rhesus monkey 92.4 (f )
104.0 (g)
Pig
94.0 (h,i)
Human
1.36
1.84
--
57.9 (d)
--
52.5 (f )
59.5 (g)
42.6 (j)
1.51
--
1.40
--
1.11
The ratio CBFcortex/ρn is the average cortical blood flow per neuron (10−8 mL/min).
References: (a) [70]; (b) [71]; (c) [72]; (d) [73]; (e) [84]; (f ) [74]; (g) [75]; (h) [16]; (i) [83]; (j)
[76].
40
Appendix S1
In this Appendix we want to show a formal step connecting Eq. (6) and Eq. (8) in
the main text. We want to find the exponent γ, which satisfies the following equation:
(ln(Rc/Ro))2/3 = (Rc/Ro)γ.
Taking the natural logarithm of both sides leads to (2/3) ln(ln(Rc/Ro)) = γ ln(Rc/Ro),
which yields Eq. (7) in the main text.
41
y = − 0.127*x + 1.97
i
]
)
n
m
*
g
0
0
1
(
/
L
m
[
)
x
e
t
r
o
c
l
a
t
e
i
r
a
p
F
B
C
(
0
1
g
o
L
y = − 0.15*x + 2
B
2.4
2.3
2.2
2.1
2
1.9
1.8
1.7
1.6
1.5
A
2.4
2.3
2.2
2.1
2
1.9
1.8
1.7
1.6
1.5
1.4
−2
−1
0
i
]
)
n
m
*
g
0
0
1
(
/
L
m
[
)
x
e
t
r
o
c
l
i
a
u
s
v
F
B
C
(
g
o
L
0
1
C
2.4
2.3
2.2
2.1
2
1.9
1.8
1.7
1.6
1.5
1.4
−2
−1
0
i
]
)
n
m
*
g
0
0
1
(
/
L
m
[
)
x
e
t
r
o
c
l
a
t
n
o
r
f
F
B
C
(
g
o
L
0
1
1
2
Log
10
(cortical volume) [cm3]
3
4
1.4
−2
−1
0
1
2
3
4
Log
10
(cortical volume) [cm3]
y = − 0.17*x + 2
i
]
)
n
m
*
g
0
0
1
(
/
L
m
[
)
x
e
t
r
o
c
l
a
r
o
p
m
e
t
F
B
C
(
0
1
g
o
L
y = − 0.19*x + 2.1
D
2.4
2.3
2.2
2.1
2
1.9
1.8
1.7
1.6
1.5
1
2
Log
10
(cortical volume) [cm3]
3
4
1.4
−2
−1
0
1
2
3
4
Log
10
(cortical volume) [cm3]
Figure 1
A
2.3
2.2
2.1
2
1.9
1.8
1.7
1.6
1.5
1.4
−2
−1
i
]
)
n
m
*
g
0
0
1
(
/
L
m
[
)
s
u
p
m
a
c
o
p
p
h
F
B
C
i
(
g
o
L
0
1
y = − 0.14*x + 1.9
i
]
)
n
m
*
g
0
0
1
(
/
L
m
[
)
s
u
m
a
a
h
t
l
F
B
C
(
0
1
g
o
L
y = − 0.17*x + 2
B
2.4
2.2
2
1.8
1.6
0
1
Log
10
(cortical volume) [cm3]
C
2
3
1.4
−2
−1
0
1
2
3
4
Log
10
(cortical volume) [cm3]
y = − 0.18*x + 2
i
]
)
n
m
*
g
0
0
1
(
/
L
m
[
)
m
u
l
l
e
b
e
r
e
c
F
B
C
(
g
o
L
0
1
2.4
2.2
2
1.8
1.6
1.4
−2
−1
0
Figure 2
1
2
3
4
Log
10
(cortical volume) [cm3]
]
m
µ
[
)
r
e
t
e
m
a
d
y
r
a
i
l
l
i
p
a
c
(
0
1
g
o
L
A
1
0.9
0.8
0.7
0.6
0.5
0.4
−2
C
B
3.2
3
2.8
2.6
2.4
2.2
−2
D
−1
y = 0.075*x + 0.6
]
3
m
m
m
m
/
[
)
ρ
(
c
0
1
g
o
L
y = − 0.162*x + 2.86
−1
0
1
2
Log
10
(cortical volume) [cm3]
1
0
(cortical volume) [cm3]
2
Log
10
3
4
3
4
y = 0.029*x − 1.8
−1
−1.5
−2
−2.5
l
)
e
m
u
o
v
y
r
a
l
l
i
p
a
c
f
o
n
o
i
t
c
a
r
f
(
g
o
L
0
1
−3
−2
−1
1
0
(cortical volume) [cm3]
2
Log
10
3
4
3
4
]
g
H
m
m
[
)
e
r
u
s
s
e
r
p
2
O
l
a
i
r
e
t
r
a
(
0
1
g
o
L
y = 0.012*x + 2
3
2.8
2.6
2.4
2.2
2
1.8
1.6
1.4
1.2
1
−2
−1
0
1
2
Log
10
(cortical volume) [cm3]
Figure 3
A
2.2
2
1.8
1.6
1.4
1.2
−2
−1
0
]
3
m
c
/
6
0
1
[
)
y
t
i
s
n
e
d
n
o
r
u
e
n
(
0
1
g
o
L
y = − 0.13*x + 1.9
]
m
µ
[
)
ρ
/
ρ
(
n
c
0
1
g
o
L
B
3
2.5
2
1.5
1
0.5
0
−0.5
−1
−2
y = − 0.034*x + 0.99
−1
0
1
2
3
4
Log
10
(cortical volume) [cm3]
1
2
3
4
Log
10
(cortical volume) [cm3]
y = 1.05*x + 0.875
C
3.2
3
2.8
2.6
2.4
2.2
]
3
m
m
m
m
/
[
)
c
ρ
(
0
1
g
o
L
2
1.4
1.5
1.6
1.7
1.9
1.8
2
(ρ
) [106/cm3]
n
10
Log
2.1
2.2
Figure 4
y = − 0.033*x + 0.18
1.5
1
0.5
0
−0.5
)
n
ρ
/
F
B
C
(
0
1
g
o
L
−1
−2
−1
0
Log
10
(cortical volume) [cm3]
1
2
3
4
Figure 5
|
1312.1104 | 1 | 1312 | 2013-12-04T10:57:48 | New insights in gill/buccal rhythm spiking activity and CO2 sensitivity in pre- and post-metamorphic tadpoles (Pelophylax ridibundus) | [
"q-bio.NC",
"physics.bio-ph"
] | Central CO2chemosensitivity is crucial for all air-breathing vertebrates and raises the question of itsrole in ventilatory rhythmogenesis. In this study, neurograms of ventilatory motor outputs recorded infacial nerve of premetamorphic and postmetamorphic tadpole isolated brainstems, under normo- andhypercapnia, are investigated using Continuous Wavelet Transform spectral analysis for buccal activityand computation of number and amplitude of spikes during buccal and lung activities. Buccal burstsexhibit fast oscillations (20-30 Hz) that are prominent in premetamorphic tadpoles: they result from thepresence in periodic time windows of high amplitude spikes. Hypercapnia systematically decreases thefrequency of buccal rhythm in both pre- and postmetamorphic tadpoles, by a lengthening of the interburstduration. In postmetamorphic tadpoles, hypercapnia reduces buccal burst amplitude and unmasks smallfast oscillations. Our results suggest a common effect of the hypercapnia on the buccal part of the CentralPattern Generator in all tadpoles and a possible effect at the level of the motoneuron recruitment inpostmetamorphic tadpoles. | q-bio.NC | q-bio | 1
New insights in gill/buccal rhythm spiking activity and CO2
sensitivity in pre- and postmetamorphic tadpoles
(Pelophylax ridibundus)
Brigitte QUENETa, Christian STRAUS b, c,
Marie-Noëlle FIAMMAb, Isabelle RIVALS a,
Thomas SIMILOWSKI b, d, Ginette HORCHOLLE-BOSSAVIT a
a. ESPCI-ParisTech, Equipe de Statistique Appliquée, F-75005 Paris, France
b. UPMC Univ Paris 06, ER 10 UPMC, F-75013 Paris, France
c. Assistance Publique - Hôpitaux de Paris, Groupe Hospitalier Pitié-Salpêtrière,
Service Central d'Explorations Fonctionnelles Respiratoires, F-75013 Paris, France
d. Assistance Publique - Hôpitaux de Paris, Groupe Hospitalier Pitié-Salpêtrière,
Service de Pneumologie et Réanimation Médicale, F-75013 Paris, France
Corresponding author at: ESPCI-ParisTech, Equipe de Statistique Appliquée, F-75005 Paris,
France. Tel.: +33 140794461; Fax: +33 1 40794669 E-mail address [email protected] (B
Quenet).
ABSTRACT
Central CO2 chemosensitivity is crucial for all air-breathing vertebrates and raises the
question of its role in ventilatory rhythmogenesis. In this study, neurograms of ventilatory
motor outputs recorded in facial nerve of premetamorphic and postmetamorphic tadpole
isolated brainstems, under normo- and hypercapnia, are investigated using Continuous
Wavelet Transform spectral analysis for buccal activity and computation of number and
amplitude of spikes during buccal and lung activities. Buccal bursts exhibit fast oscillations
(20-30Hz) that are prominent in premetamorphic tadpoles: they result from the presence in
periodic time windows of high amplitude spikes. Hypercapnia systematically decreases the
frequency of buccal rhythm in both pre- and postmetamorphic tadpoles, by a lengthening of
the interburst duration. In postmetamorphic tadpoles, hypercapnia reduces buccal burst
amplitude and unmasks small fast oscillations. Our results suggest a common effect of the
hypercapnia on the buccal part of the Central Pattern Generator in all tadpoles and a possible
effect at the level of the motoneuron recruitment in postmetamorphic tadpoles.
KEYWORDS
Amphibian buccal ventilation, central CO2 chemosensitivity, fast oscillations, CWT spectrum,
spiking activity.
2
1. Introduction
in amphibians, characterized by a
the breathing control
Previous studies of
developmental transition from water breathing to air breathing, have addressed the issue of the
emergence of central CO2 chemosensitivity. Many investigations use the fact that cranial nerve
outputs of the isolated brainstem preparations from tadpoles and frogs match the corresponding
respiratory muscle sequences (Gdovin et al., 1998; Kogo et al., 1994a, 1994b; Sakakibara,
1984a, 1984b). The nerve activity exhibits regular low amplitude bursts associated with gill
ventilation and episodic high amplitude bursts corresponding to lung ventilation (Galante et al.,
1996).
Several studies have shown that hypercapnia increases systematically the occurrence of lung
episodes in postmetamorphic tadpoles (Kinkead and Milsom, 1994; Torgerson et al., 1997;
Taylor et al., 2003a, 2003b; Wilson et al., 2002; Straus et al., 2011). But regarding the buccal
rhythm, experimental results describe heterogeneous effects of hypercapnia and metamorphosis
on the gill/buccal frequency in pre- and postmetamorphic tadpoles, leading to contradictory
conclusions. In Lithobates catesbeianus (Rana catesbaiana) of the same premetamorphic
stages exposed to hypercapnia, Torgerson et al. (1997) reported an increase in the frequency
and the amplitude of the gill rhythm while Taylor et al. (2003a) observed a tendency for
frequency to decrease and no change in duty cycle. In postmetamorphic animals of the same
stage exposed to hypercapnia, Torgerson et al. (1997) did not observe any response; Taylor et
al. (2003b) observed a tendency for frequency to decrease. These conflicting conclusions could
be explained by physiological differences between species or/and by methodological
differences. Several studies reported effect of metamorphosis at normocapnia in Lithobates
catesbeianus (Burggren and Doyle, 1986; Togerson et al., 1998; Gdovin et al., 1999; Taylor et
al., 2003b; Hedrick et al., 2005; Fournier and Kinkead, 2006; Chen and Hedrick, 2008). These
studies gave conflicting results about the gill/buccal activity: for instance, Taylor et al. (2003b)
and Fournier and Kinkead (2006) reported no change of the buccal frequency, while Chen and
Hedrick (2008) observed a decrease of about 50%.
Developmental observations in tadpoles (Torgerson et al., 2001a, 2001b; Taylor et al., 2003b)
have suggested that the neural substrates for central chemoreception exist in the vicinity of the
structures responsible for both gill/buccal and
lung rhythm generation already
in
premetamorphic tadpoles. Here, we study how CO2 sensitivity affects premetamorphic and
postmetamorphic gill/buccal and lung rhythm in Pelophylax ridibundus. We analyze
neurograms recorded from isolated brainstems in Pelophylax ridibundus whose chaotic nature
has been evidenced by Straus et al. (2011). On the part of the filtered neurograms
corresponding to buccal activity only, we perform time-frequency transform with complex
Morlet wavelets (CWT) that is adapted to non-stationary signals (Akay, 2005; Mor and Lev-
Tov, 2007; Vialatte et al., 2007; Romcy-Pereira et al., 2008). With adapted signal-processing
tools, we evidenced, for the first time in amphibians, the presence of fast oscillations in the 20-
30 Hz frequency band in the buccal bursts, probably resulting from phase locked motoneuron
activities, which are reminiscent of some of the fast oscillations described in several species of
mammals (Funk and Parkis, 2002; Marchenko and Rogers, 2006, 2007). Regarding the effects
of hypercapnia (acidosis at pH=7.4) on the gill/buccal activity, our results show that there is no
change of the fast oscillations in the premetamorphic tadpole, but significant changes in
postmetamorphic ones. In addition, hypercapnia induces significant lengthening of the buccal
interburst period at both stages.
3
2. Materials and methods
2.1. Data collection
Using the brainstem preparation of the tadpole, Pelophylax ridibundus (Rana
esculenta - European Edible Frog) described in detail in Straus et al. (2011), the present study
investigates a subset of the signals used in this paper. The experimental protocol fulfilled
appropriate legal and ethical conditions (French Ministry for Agriculture and Animal Care
Committee - Ile-de-France, Paris, Comité 3). Briefly, the developmental stage of the animals
is determined according to Taylor and Köllros (1946). Tadpoles are chosen in order to assign
them either to the premetamorphic (stages X-XIII) or to the postmetamorphic (stage XXIV-
XXV) group of development. They are anesthetized and decerebrated. The brainstem is
dissected and transferred to a recording chamber. This chamber is superfused at room
temperature (~21°C) with mock cerebrospinal fluid (CSF) equilibrated with a gas mixture of
O2 and CO2: (in mM) NaCl, 104; KCl, 4; MgCl2, 1.4; D-glucose, 10; NaHCO3, 25; CaCl2 2.4.
The pH of the CSF is set to 7.8 or 7.4 by adjusting the fractional concentration of CO2. The
electrical activity – namely the neurogram - of the root of the facial nerve (VII) is recorded
using glass suction electrodes. The raw signal is then amplified (10 000 times) and filtered
(100 Hz–1 kHz) using a high gain, differential AC amplifier (model 1700, A-M Systems Inc.,
Evrett, WA), digitized at 2000 Hz (PowerLab, ADInstruments) and stored as text files.
Recordings last 20-35 min for each pH value. The first 10 minutes of each recording is
discarded. Then, 5 minutes of signal with the best identifiable bursts of buccal and lung
ventilation are selected for analysis. Indeed, the tadpole isolated brainstem preparation
exhibits neural ventilatory activity but also periods of apnea and, sometimes, also bursts of
undetermined kind, like, for example, the so-called ‘complex discharges’ (Straus et al., 2000).
We retain the part of the neurograms where the preparation is not apneic at that time the type
of the bursts is clear, either buccal or lung, with a minimum number of discharges of an
undeterminable kind.
2.2. Neurogram processing
The analysis is performed on neurograms from 5 premetamorphic (T, 1-5) and 5
postmetamorphic (F, 1-5) preparations, in normo- (pH=7.8) and hypercapnia (acidosis at
pH=7.4). The spectral analysis based on complex Morlet wavelets (CWT) leads to
information in the frequency domain taking temporal short-time variations into account.
After segmentation of the signal into cycles (buccal and lung), we identified the lung bursts,
and computed cross-correlation functions in order to insure the best temporal alignment of
these segmented cycles of each type, buccal or lung, separately. Finally, we come back to
the neurograms, which can be analyzed using the previous segmentation and temporal
alignment of the cycles, in order to measure the position and amplitude of each spike in
each cycle.
2.3. Neurogram filtering and spike detection
Each neurogram is squared, zero-phase filtered (Gustafsson, 1996) using a moving
average rectangular window of width W =100 ms. After taking the square root of the result,
the lung bursts are eliminated (see next paragraph). On the resulting S100 signal, a detection
of local minima is performed in order to segment this buccal signal into cycles, preserving
both the buccal burst and the inactive period in each cycle (Fig. 1A, black line). The same
filter with a window of width W=10 ms is applied to the buccal part of the neurogram, Sneur
(Fig. 1A, gray line) resulting in a new signal S10 (Fig. 1A, red line).
4
A detection of maxima, i.e. spikes, is performed independently on the positive and the
negative parts, in absolute value, on Sneur (Fig. 1B, red and blue dots), in order to store the
time position and the amplitude of each of them. This detection allows the computation of a
local spike density (per time unit).
2.4. Lung burst detection and analysis
The detection of local minima performed on the squared and filtered neurogram
(with W=100 ms) is completed by a detection of each local maximum between two
successive minima defining each burst. In a given neurogram, a burst is labeled ‘lung burst’
according to two criteria: its amplitude is higher than twice the mean amplitude of all the
recorded burst, and its spike density is higher than twice the mean spike density of all the
recorded bursts.
2.5. Cycle time alignment by cross-correlation function
The aim of this signal-processing step is to get a time alignment of the buccal bursts
able to exhibit phase locking of neuronal activity inside the bursts of a given neurogram.
Therefore, we define in S10 a “reference cycle” and position all the other cycles according
to the maximal value of the cross-correlation function (sliding dot product) between each
cycle and the reference (Lynn and Fuerst, 2000). At a first step, the amplitude vectors
representing the S10 buccal cycles, after segmentation, are padded with zeros when
necessary, in order to get an identical length for all these vectors. In a second step, a
symmetric matrix of correlations is computed, whose generic term of indices i and j is the
maximal value of the cross-correlation function between the amplitude vectors of cycle i
and cycle j. The “reference cycle” is then the cycle whose mean correlation with all the
other cycles is the highest: it is therefore considered as the most representative cycle. At a
third step, the time lag, which corresponds to the maximum of the cross-correlation
function between a cycle and the reference cycle, defines the time position of this cycle
relatively to the reference (Fig. 2A): when the cycles are positioned according to their time
lag and ranked according to their correlation with the reference, the bursts are aligned, and
each cycle can be visualized in a map by a horizontal band where its amplitude is coded in
gray scale (Fig. 2A). A same time alignment procedure is applied to the lung burst.
2.6. Amplitude profiles from the “spike matrices”
The aim of this procedure is to get canonical profiles of the burst and interburst part
representative of each neurogram buccal cycles. Using the relative time positions defined
previously, we build two matrices of the cycle positive and negative spikes respectively
(Fig 1B). For each matrix, the generic term of indices i and j is the amplitude of a spike,
when present, at sampling time j of cycle i. In maps visualizing these matrices, each row
corresponds to a cycle, each column corresponds to a time bin of the cycles, and a dot in a
row and a column corresponds to a spike in that time bin for this cycle (Fig 2B). An
amplitude profile is computed from these two matrices: the mean values, at each time bin,
of the positive and negative amplitudes over the cycles, after zero-phase filtering with a
window of 5 ms, constitute the two parts of the profile.
In order to define a canonical cycle for each neurogram, characterized by its amplitude
profile, with its positive and negative parts, whose duration corresponds to the median
value of the cycle durations, we use the most representative part of the amplitude profiles,
i.e. where the number of zeros related to the padding is the smallest: this criterion defines
the beginning and the end of the canonical cycle (Fig. 2A and 2B). In such a canonical
cycle, we want to define a central active zone, or duty cycle, corresponding to the buccal
burst, surrounded by zones of minor activity. The beginning of the burst is the first time
5
when the positive amplitude profile is higher than its mean value; symmetrically, the end of
the burst is the last time when the positive amplitude profile is higher that its mean value.
As the positive and the negative amplitude profiles are very similar in absolute value, only
the positive part is visualized when comparing the canonical cycles at the two values of the
pH; moreover, in graphical representations of a canonical activity, the positive part of the
canonical cycle is repeated twice, in order to get a complete view of the interburst part.
Amplitude profiles of lung burst are computed using the same procedure.
2.7. Number and amplitude profiles according to spike amplitude
The aim of this procedure is to evaluate the contributions of the spikes to the time structure of
the bursts according to their amplitude. The alternation of high and weak spike activity
characterizes the cyclic buccal part of the neurogram: we define the high activity window
(buccal burst duration or ‘duty cycle’) as the time interval of the buccal canonical cycle where
its profile amplitude is higher than the mean amplitude of this cycle (between the vertical
green lines in Fig. 2B). Then, two histograms of spike amplitude are computed in the buccal
cycles -aligned according to the previously described cross-correlation procedure (see § 2.5)-
corresponding to the spikes inside the active window (Pin) and to the spikes outside (Pout) this
window respectively. Populations Pin and Pout exhibit lognormal-like shaped distributions (not
shown) of the spike amplitudes. The threshold amplitude value Ath where the number of
spikes of Pin becomes higher that the number of spikes of Pout, defines the minimal amplitude
of a spike to have a higher probability to belong to Pin than to Pout. The positive spike
population is then divided into five subpopulations. The first one corresponds to the spikes
whose amplitude is smaller than Ath, the second, third and fourth subpopulations contain
spikes whose amplitude belongs to ]Ath, 3*Ath], ]3*Ath, 5*Ath] and ]5*Ath, 7*Ath] respectively,
and the fifth subpopulation contains the highest remaining spikes, higher than 7*At (Fig. 1B).
For each subpopulation k (k=1, 2, 3, 4 or 5), two new matrices are computed: a matrix of
the number of spikes over ten consecutive time bins Mk and a matrix of the mean amplitude
of these spikes Pk. In order to visualize the time distribution of the spiking activity for each
subpopulation, Mk is graphically represented by a map of dots, a dot corresponding to the
presence of one spike at least (Fig. 2B). The time successive mean values of the number of
spikes over the cycles define a number profile for each subpopulation, idem for the time
successive mean values of the amplitude, which define the amplitude profile for each
subpopulation.
2.8. Continuous Wavelet Transform maps and spectra
Given the non-stationarity of the buccal rhythm, a time-frequency analysis of S10 signal is
performed using a Continuous Wavelet Transform (CWT) (Akay, 2005; Mor & Lev-Tov;
2007; Vialatte et al., 2007; Romcy-Pereira et al., 2008) of S10 subsampled at 200 Hz (after
applying an anti-aliasing low pass filter eliminating all the frequencies above 100 Hz),
looking for relevant frequencies between 1 and 100 Hz, with a complex Morlet wavelet
basis function including five oscillations (Mallat, 2000). The coefficients of each time-
frequency amplitude map are computed for 100 values logarithmically distributed between
1 and 100 Hz, and at a time resolution corresponding to the sampling of S10. This time-
frequency map displays the temporal evolution of the S10 dominant frequencies in gray
scale (Fig. 3). All the CWT maps got from our signals exhibit a relative stability of the
dominant frequencies, which validates the representation of each S10 by a CWT spectrum
computed as the mean value, at each frequency, of the corresponding coefficients of the
amplitude map over time activity (Fig. 3). In order to compare these spectra, each of them
is normalized.
6
2.9. Statistical analysis
Statistical tests have been applied to the three following descriptors: the mean buccal cycle
duration, the relative energy in the band 1-10 Hz and the relative energy in the band 20-30
Hz, measured on the CWT spectrum. In order to test both the effects of pH and stage on
these descriptors, a two-way ANOVA is applied when the condition of homoscedasticity is
fulfilled, which is tested with Bartlett’s test. In this case, we also have to take into account
the fact that the same animal is tested for the two values of the pH, by including a random
individual effect. When homoscedasticity is rejected, pairwise statistical tests are
performed. The effect of the pH at a given stage is tested with a paired t-test, the effect of
the stage at a given pH is tested with a standard t-test if the variance of the two groups are
homogeneous (F-test), with a Wilcoxon rank sum test if not. Statistical significance was
assumed if p < 0.05. Signal processing algorithms and statistical computations have been
implemented with Matlab (The Mathworks, version 8.0.0.783 (R2012b)).
3. Results
3.1. CWT maps and spectra of buccal activity
Frequency information from S10 signal is visualized in both CWT maps and spectra, as
illustrated for T2 and F4 at both values of the pH (Fig. 3, lower part). Three seconds of the
corresponding S10 signals are illustrated in the upper part of this figure. In time domain, the
main effects induced by metamorphosis and by acidosis on the buccal rhythm are already
visible, i.e. a shortening of the buccal period after metamorphosis, a lengthening of the
period with acidosis in both premetamorphic and postmetamorphic tadpoles, and
oscillations superimposed on the buccal bursts at both pH in premetamorphic tadpoles,
much more discrete in postmetamorphic tadpoles.
For T2, the maps reveal a dark gray strip, discontinuous, at low frequency, 1.5 Hz,
corresponding to the fundamental buccal frequency at pH=7.8, whereas the equivalent strip
is darker, continuous, at 1.2 Hz at pH=7.4, which indicates that hypercapnia induces a
buccal activity at lower frequency and at higher rhythm regularity. A wider strip, at high
frequency centered at 21 Hz, is present with same characteristics at both values of the pH,
corresponding to fast oscillations exhibited on the bursts in S10 signal. For F4, the maps
reveal a dark, continuous, low frequency strip (3.7 Hz) at pH=7.8, while the corresponding
strip at pH=7.4 is lighter, at 2.5 Hz. Here, hypercapnia induces a lower and more variable
frequency of the buccal activity. A second strip is present at both pH values, at about twice
the corresponding low frequencies.
All these pieces of frequency information are also given in the spectra, which are illustrated
for all the signals in Fig. 4. For each premetamorphic tadpole at pH=7.8 and 7.4, the
spectra reveal two or three local maxima: the first one occurs systematically in the low
frequency range [1.1, 1.6] Hz, and the last one occurs in the high frequency range [20.6,
27.2] Hz. The frequency of the first peak, or buccal low frequency, is slightly lower in
hypercapnia than in normocapnia and the amplitude of this peak is higher, but for T5. The
high frequency parts of the spectra are very similar at both pH.
In the postmetamorphic tadpole, three maxima can be observed in the spectra. In
normocapnia, a first peak occurs at low frequency in the range [2.9, 3.7] Hz, systematically
followed by the peak at twice this fundamental frequency, i.e. [5.9 7.4] Hz. In two cases
(F1, F5) there is a smooth local maximum around 22 Hz. In hypercapnia, the first peak
occurs at a lower frequency, in the range [1.5, 2.5] Hz. When present, the second peak is
very small in amplitude and a smooth local maximum at high frequency around 22 Hz can
7
be seen in all cases. In contrast with the premetamorphic tadpoles spectra, the amplitude of
the first peak is systematically smaller in hypercapnia than in normocapnia, while there is
an increase in the high frequency band.
The statistical tests performed on the CWT spectral energies in the low frequency (1-10 Hz)
and in the high frequency band (20-30 Hz) give the following results. For the effect of
metamorphosis, the relative energy in the low frequency band is significantly higher for the
postmetamorphic tadpoles at both pH: at pH=7.8, +0.33 (p=0.008) and at pH=7.4, +0.09
(p=0.029), while in the high frequency band it is significantly lower: at pH=7.8, -0.12
(p=0.008) and at pH=7.4, -0.04 (p=0.001).
For the effect of the acidosis in premetamorphic tadpoles, there is no significant effect on
the relative spectral energy neither in the low frequency band nor in the high frequency
band. In postmetamorphic tadpoles, the decrease in the low frequency band is significant:
-0.23 (p=0.023), as well as the increase in the high frequency band: +0.07 (p=0.012).
3.2. Amplitude profiles of the buccal canonical cycles
In order to visualize in the temporal domain the dynamical modifications induced in the
buccal rhythm by hypercapnia, we superimpose the canonical activities at both pH. A
graphical comparison is thus obtained by positioning the positive part of the amplitude
profile of the canonical activity (i.e. a canonical cycle repeated twice) at pH=7.4 in such a
way that the correlation of its first cycle with the first cycle of the canonical activity at
pH=7.8 is maximal.
These amplitude profiles show
that oscillations are systematically present
in
premetamorphic tadpoles in normo and hypercapnia. For all premetamorphic tadpoles,
oscillations occupy the major part of the cycle (Fig. 5, left). These oscillations are rather
regular, with an interval of about 35-50 ms between two successive peaks. Under
hypercapnia, the cycle duration is systematically increased by about 0.14 s, i.e. by around
20%. This duration increase clearly affects the interburst part of the cycle. The burst
amplitude and duration seem very weakly modified.
For postmetamorphic tadpoles, the bursts are shorter and exhibit an asymmetry, with a
sudden interruption of activity which ends the burst, with very few and small oscillations, if
any (Fig. 5, right). Under hypercapnia, the cycle duration is systematically increased by
about 0.18 s, i.e. by around 60%. This duration increase clearly affects the interburst part of
the cycle also. The burst amplitude is markedly reduced with visible intra-burst oscillations.
The N-way ANOVA performed on the mean value of the buccal period (Table 1) indicates
that there is no individual effect, but a significant stage effect with a decrease of 335 ms
(p=0) and a significant pH effect with an increase of 155 ms (p=0.0004), and no significant
cross-effect between pH and stage, which suggests a similar lengthening of the period with
the pH for both pre- and postmetamorphic tadpoles.
3.3. Spike contributions to the buccal activity according to their amplitude
For T2, the matrices Mk containing the number of spikes in ten successive time bins (i.e. 5
ms) for each of the first four subpopulations (k=1,2,3 and 4) are visualized as maps of gray
dots for pH=7.8 and red dots for pH=7.4 (Fig. 6). By definition, the first subpopulation
contains the smallest ‘spikes’, with both neuron firings and noise, whose density is higher
outside the burst than inside (the burst limits are represented by the green lines in Fig.6):
this outside density is greater in hypercapnia than in normocapnia, while the inside burst
density is quite the same at both pH values. These features characterize both the number
and amplitude profiles, even if this high spike density brings a small contribution to the
amplitude profile (Fig. 6). For the second subpopulation, there is a small increase in spike
8
density and amplitude during the burst with slightly smaller values at pH=7.4, particularly
during the prolonged interburst part of the cycle. For the third population, there is an
increase in spike density and amplitude during the burst, carrying out its plateau part, with
a same activity profile of the burst at both pH values. The corresponding spike density has
a maximum at 0.31 spikes/5ms = 62 spikes/s. For the fourth subpopulation containing the
highest spikes, at both pH values we observe the quasi-absence of spikes outside the burst
and an alternation of bands of high and low density of these spikes, resulting to the
oscillations of the number and amplitude profiles. In conclusion, in T2 and all
premetamorphic tadpoles, in addition to the lengthening of the inactive period, the only
visible effect of hypercapnia is the reduction of the second subpopulation spikes during the
prolonged interburst.
For F2, the matrices Mk (k=1,2,3 and 4) are visualized for pH=7.8 and pH=7.4 in Fig. 7. At
pH=7.8, the first subpopulation exhibits a band of very high and constant density at the end
of the burst that cannot be of physiological origin: it means that this first peaks/spikes
subpopulation is clearly dominated by instrumental noise. For the second population, the
spikes are concentrated inside the burst with same densities inside and outside the burst at
both pH values, but the duration of the burst is longer at pH=7.4. The spike density of the
third subpopulation exhibits an increase during the burst followed by a quasi-absence of
spikes after the burst. The spike density inside the burst is smaller in hypercapnia than in
normocapnia. These features are accentuated in the fourth subpopulation, where some
oscillations are present. For all neurograms from the postmetamorphic tadpoles, as in F2, in
addition to the lengthening interburst duration, the main effect of hypercapnia is the strong
reduction of the third and fourth subpopulation spikes during the burst.
3.4. Analysis of the lung activity
Figure 8 illustrates the intensity of the neuronal activity for the ten neurograms of the
postmetamorphic tadpoles represented by the whole corresponding S100 signals, with both
buccal and lung components. This figure points out the higher amplitude of the buccal
bursts at pH = 7.8 than at pH=7.4. The irregular occurrence of lung bursts is visible in the
distribution of the vertical bars indicating these lung bursts. The number of lung bursts is
higher at pH=7.4 than at pH=7.8 (but F1, whose number of lung bursts is already high at
pH=7.8). The left part of figure 9 illustrates two typical lung bursts observed in the
neurograms of the postmetamorphic tadpole F2 at pH=7.8 and pH=7.4 respectively: in this
tadpole, all the lung bursts at pH=7.8 are preceded by a small amplitude burst, which is
totally absent at pH=7.4. The corresponding mean profiles confirm the presence of these
two lung burst types; in addition they show that the mean amplitude of the lung bursts is
smaller at pH=7.4 than at pH=7.8. Regarding the contribution to the different spike
populations according to their amplitude, the main observation is that the lung burst
amplitude is related to the spikes of population V that appears to be specific of the lung
activity and seems to be more present in normocapnia than in hypercapnia in F2. Of note,
there is no visible temporal structure in the lung bursts as observed in the buccal bursts.
4. DISCUSSION
In the present paper, our analysis of the neurograms was done in order to capture some
dynamic characteristics of buccal and lung rhythm generator at short-time scale and to
evaluate the effects of hypercapnia on these characteristics. In the European frog, Pelophylax
ridibundus the major findings of the study are the following:
1. In normocapnia, the frequency of the buccal rhythm in premetamorphic tadpoles is in
the range of [75, 96] min-1 and in postmetamorphic tadpoles, in the range [151, 220]
min-1, which is higher than all reported values in the American bullfrog (Table 1).
9
2. Fast oscillations in the frequency band 20-30 Hz are present in all tadpoles, revealed
by both the spectral analysis and by the buccal amplitude profiles (Fig. 4 and Fig. 5).
Such oscillations are visible in other amphibian species under the form of small
indentations in the standard filtered neurograms, but they have not been studied.
3. The analysis of spike time positions and amplitudes reveals their respective
contributions to the burst structure: the small spikes are uniformly distributed in the
burst, while the high ones are present in specific time windows separated by about 35-
50 ms at the origin of the fast oscillations (Fig. 6 and Fig. 7).
4. Metamorphosis induces a significant shortening of the cycle duration of about 335 ms
(Table 1)
5. Metamorphosis induces a significant increase of the relative energy in the low
frequency band (1-10 Hz) and a significant decrease of relative energy in the high
frequency band (20-30 Hz) (Fig. 4).
6. Hypercapnia induces a significant lengthening of about 155 ms of the interburst
duration in branchial/buccal rhythm in both premetamorphic and postmetamorphic
tadpoles (Table 1 and Fig. 5).
7. In the buccal activity of postmetamorphic tadpoles, hypercapnia induces a significant
decrease of the relative energy in the low frequency band (1-10 Hz) and a significant
increase of relative energy in the high frequency band (20-30 Hz) (Fig. 4).
8. In the buccal activity of postmetamorphic tadpoles, hypercapnia reduces noticeably in
the number of high amplitude spikes, which is related to the low frequency energy
decrease (Fig. 7).
9. Lung bursts are present in all postmetamorphic tadpoles, and, as in other species, are
more frequent in hypercapnia than in normocapnia (Fig. 8).
10. The lung bursts are characterized by the presence of spikes of high amplitude
(population V)
From our analysis of the neurograms with their buccal and lung activities, it seems that the
lung activity is “phase locked” with the buccal cycles in some preparations. In such
preparation, the phase locking may be different depending on the pH (for example see type I
or type II lung bursts of figure 9). This dynamic relationship suggests reliable dynamical
interconnections between the “lung network” and the “buccal network” of the CPG.
Our observations performed on in vitro brainstem preparations of tadpoles of the European
frog are similar to those of Taylor et al (2003a and b) with the same type of preparation in
Lithobates catesbeianus: hypercapnia is associated with a frequency decrease in buccal
oscillations in pre- and post-metamorphic tadpoles through a lengthening of the interburst
interval. In the case of premetamorphic tadpoles, the lengthening of the interburst duration
induced by hypercapnia leads to a prolonged inactivity of nerve VII motoneurons, which can
be attributed to a modification of the buccal rhythm genesis, since there is no lung activity.
This hypercapnia-related increase of the interburst in the postmetamorphic tadpoles induces a
higher effect on the buccal frequency, since this frequency in normocapnia is much higher in
postmetamorphic tadpoles (about 3 Hz) than in premetamorphic tadpoles (about 1.25 Hz). In
postmetamorphic tadpoles, this frequency effect is associated to a significant amplitude
reduction of the buccal bursts. These two effects that express a decrease in the buccal activity
could be the result of a global anesthetic effect of CO2; however, such a global effect would
also decrease the lung activity, in contrast with the observed increase of the lung frequency.
Both amplitude and frequency reduction of the buccal oscillation are also observed in vivo in
adult Lithobates catesbeianus (Kinkead and Milsom, 1994). In air-breathing postmetamorphic
tadpoles, several hypotheses have been proposed for the physiological function of the buccal
rhythm, which is no longer directly involved in gas exchange. The buccal rhythm ventilates
the oropharynx alone: it has been speculated that is optimizes the efficiency of lung breaths
10
by keeping the buccal cavity ventilated with fresh air (Gans et al., 1969). This point of view
has been challenged by the attribution of a likely olfactory function to the buccal activity
(Foxon, 1964; Vitalis and Shelton, 1990; Jorgensen, 2000).
The hypercapnia frequency effect seems to be a buccal Central Pattern Generator (CPG)
chemosensitive property already present before metamorphosis and retained after
metamorphosis despite the loss of a direct ventilatory function of the buccal activity. Our
results evidence that the amplitude decrease of the buccal burst induced by hypercapnia in
postmetamorphic tadpoles is due to a smaller number of high amplitude spikes. This reduced
motoneuronal activity could result either from a chemosensitivity of a motoneuron population
whose threshold may be increased at pH=7.4 or from a less efficient excitatory input from the
buccal CPG. In both cases, the motor units that are not recruited in hypercapnia during the
buccal activity are likely more efficiently recruited during lung episodes, which occur at a
higher frequency than in normocapnia.
The CWT spectra of the buccal activity (Fig. 4) reveal the peak at 20-30 Hz (Fig. 4)
corresponding to the oscillations observed in S10. These oscillations are mainly visible in the
premetamorphic tadpole buccal amplitude profiles (Fig. 5). When focusing on the
contribution of the spike sub-populations defined by their amplitude, the spike matrices reveal
a quasi-periodic density variation of the fourth population spikes (Fig. 6, top and left), with a
period of about 40 ms repeated about 10 times during a buccal burst. This density variation of
the highest spikes explains the oscillatory feature of the curves representing their mean
number and mean amplitude, and finally the complete amplitude profile (Fig. 5). It is quite
unlikely that this temporally organized motoneuron recruitment process is due to intrinsic
motoneuron properties, thus we suppose that it is a consequence of the driving of the
motoneurons by the buccal CPG, which may send oscillating excitatory inputs to the
motoneuronal population. If we suppose that, in premetamorphic tadpoles, the motoneuronal
population is quite homogeneous, i.e. weakly differentiated, all these motoneurons are driven
in the same way by the oscillating excitatory input and exhibit a higher firing probability at
the maxima of the oscillation: the presence of high spikes in the corresponding short time
windows may thus result from the summation of individual action potentials occurring in
coincidence. In postmetamorphic tadpoles, the peak in the CWT spectra at twice the
fundamental frequency [5.9 7.4] Hz is prominent in normocapnia (Fig. 4): it accounts for the
specific burst shape, higher and narrower than in the premetamorphic tadpoles (Fig. 3 and Fig.
5), that we relate to the recruitment of a more differentiated motoneuron population. Indeed,
the asymmetry of the burst shape exhibited in postmetamorphic tadpole amplitude profiles
suggests a progressive recruitment of motoneurons related to size effects of a differentiated
population, according to the Henneman size principle (Henneman et al., 1965), where the
largest neurons, with the highest threshold are recruitment only at the end of the generator
buccal cycle. This progressive recruitment is followed by a sudden stop: the fast and complete
de-recruitment of the motoneurons at the end of the burst may be due either to an active
inhibition process or to the end of the input burst. In normocapnia, such a recruitment process
of the motoneurons may also explain the quasi-absence of fast oscillations at pH=7.8, while
present at pH=7.4, although these oscillations are likely structuring the buccal generator
activity in both cases. In this hypothetical frame, at pH=7.4, the threshold of CO2
chemosensitive motoneurons may be increased in such a way that the generator activity input
is not sufficient to recruit them. However, some small motoneurons, with a low threshold, still
respond to this oscillating input (Fig. 4).
Fast oscillations have been described in the respiratory neural control system of several
mammalian species (for a review see Funk and Parkis, 2002). They are present in inspiratory-
related muscles, nerves, and neurons and include both Medium- (MFO) and High-Frequency
Oscillations (HFO). The first studies on phrenic and hypoglossal nerve discharges in cats
11
(Richardson and Mitchell, 1982; Cohen et al., 1987; Christakos et al., 1988,1991) revealed
frequency peaks in the two bands: 20–50 Hz (MFO) and 60–110 Hz (HFO). Similarly, in
adult rats (Marchenko and Rogers, 2006), the time-frequency spectra of phrenic and
hypoglossal nerve also exhibited two peaks: one for MFO (37–110 Hz) and one for HFO
(156–230 Hz) during eupnea. In mouse (O'Neal et al, .2005), spectral analyses of diaphragm
EMG bursts revealed peaks associated with the MFO (20 – 46 Hz) and HFO. In kitten
brainstems (Kato et al., 1996), spectra calculated for the inspiratory discharges of facial
respiratory nerves showed a stable oscillation in the MFO range only (27-32 Hz) and in
newborn rat (Kocsis et al, 1999), this band is also present at (22.8-43.0 Hz). As in kitten, in
juvenile rat there is a dominant band in the MFO at 20–55 Hz (Marchenko and Rogers, 2007).
Thus, the fast oscillations appear to be dependent on developmental stage, with MFO always
present in young animals.
The fast oscillations at 20-30 Hz that we observe in Pelophylax ridibundus are reminiscent of
MFO in mammals: indeed, Marchenko and Rogers (2007), for instance, attribute the genesis
of the MFO band to a central generator that could be of ancestral origin. The presence of fast
oscillations in the ventilatory central generator appears as a characteristic that has been
retained through the phylogeny, and it would be interesting to look for them in other anuran
species, particularly in the well-studied Lithobates catesbeianus. This temporal structure of
the buccal cycle, in addition to its fundamental low frequency (about 1 Hz) rhythm, suggests
the presence of high frequency pacemaker-like activities of the buccal generator, of cell
and/or network origin as observed in several types of locomotion, respiration and mastication
CPGs (Harris-Warrick, 2010). We suppose that the activity of the buccal network neurons is
the consequence of convergence of two pacemaker inputs with two slightly different
frequencies on these neurons. The interference between two such inputs may generate both a
slow beating frequency (Lefler Y. et al, 2013) (the buccal fundamental rhythm) and a high
frequency related to the mean of the pacemaker frequencies (intraburst oscillations), with or
without filter effects (Rose G.J. and Fortune E.S., 1999).
In a modeling approach of the buccal and lung CPG, Horcholle-Bossavit and Quenet have
proposed (Horcholle-Bossavit and Quenet, 2009) a synchronizing function of the lung part of
the CPG in the motoneuron recruitment process, which could account for the high amplitude
spikes specific of the lung bursts. The model also implements CO2 sensitivity at the level of
the lung part of the CPG to explain the increased lung episodic response to hypercapnia. The
results of the present work involve new mechanisms which must be included in the model of
the CPG, introducing interactions between pacemakers, in order to implement the buccal CO2
chemosensitivy effects, the fast oscillations structuring the buccal cycle, and the phase-
locking of the lung bursts with the buccal activity.
5. CONCLUSION
Our results about the buccal ventilatory rhythm control in Pelophylax ridibundus evidence the
presence of fast oscillations inside each buccal burst due to coincident firings of motoneurons
in quasi-periodic time windows. The effects of hypercapnia on the frequency components, i.e.
the low frequency (buccal rhythm) and the high frequency (fast oscillations), suggest that
chemosensitivity affects both buccal generator elements and some facial motoneurons
properties.
ACKNOWLEGMENTS
We thank Quentin Michard for technical assistance in the time-frequency analysis and Didier
Cassereau and Yacine Oussar for methodological advices.
12
REFERENCES
1. Akay, M., 2005. Hypoxia silences the neural activities in the early phase of the
phrenic neurogram of eupnea in the piglet. J. Neuroeng. Rehabil. 30, 32-41.
2. Burggren, W.W., Doyle, M., 1986. Ontogeny of regulation of gill and lung ventilation
in the bullfrog, Rana catesbeiana. Respir. Physiol. 66, 279–291
3. Chen, A.K., Hedrick, M.S., 2008. Role of glutamate and substance P in the
amphibian respiratory network during development. Resp. Physiol. Neurobiol. 162,
24–31.
4. Christakos, C.N., Cohen, M.I., Barnhardt, R., Shaw, C.F. 1991. Fast rhythms in
phrenic motoneuron and nerve discharges. J Neurophysiol. 66, 674-87.
5. Christakos, C.N., Cohen, M.I., See, W.R., Barnhardt, R. 1988. Fast rhythms in the
discharges of medullary inspiratory neurons. Brain Res., 463, 362-367.
6. Cohen, M.I., See, W.R., Christakos, C.N., Sica, A.L., 1987. High-frequency and
medium-frequency components of different inspiratory nerve and their modification
by various inputs. Brain Res. 417, 148-152.
7. Fournier, S. and Kinkead, R., 2008. Noradrenergic modulation of respiratory motor
output during tadpole development: role of adrenoceptors. J. Exp. Biol. 209, 3685-
3694.
8. Foxon, G.E.H., 1964. Physiology of the Amphibia, in: Moore, J.A. (Eds.) Blood and
respiration. Academic Press, New York, USA, pp. 151–209.
9. Funk, G.D. and Parkis, M.A., 2002. High frequency oscillations in respiratory
networks: functionally significant or phenomenological? Resp. Physiol. Neurobiol.
131, 101-20.
10. Galante, R.J., Kubin, L., Fishman, A.P., Pack, A.I., 1996. Role of chloride-mediated
inhibition in respiratory rhythmogenesis in an in vitro brainstem of tadpole, Rana
catesbeiana. J. Physiol. 492, 545-558.
11. Gans, C., De Jongh, H.J., Farber, J., 1969. Bullfrog (Rana catesbeiana) ventilation:
how does the frog breathe? Science 163, 1223-1225.
12. Gdovin, M.J., Torgerson, C.S., Remmers, J.E., 1998. Neurorespiratory pattern of gill
and lung ventilation in the decerebrate spontaneously breathing tadpole. Respir.
Physiol. 113, 135-146,
13. Gdovin, M.J., Torgerson, C.S., Remmers, J.E., 1999. The fictively breathing tadpole
brainstem preparation as a model for the development of respiratory pattern generation
and central chemoreception. Comparative Biochemistry and Physiology Part A:
Molecular & Integrative Physiology 124, 275–286.
14. Gustafsson, F., 1996. Determining the initial states in forward–backward filtering.
IEEE Trans. Signal Process. 44, 988–99.
15. Harris-Warrick R.M., 2010. General Principles of Rhythmogenesis in Central Pattern
Networks. Prog Brain Res . 187, 213–222.
16. Hedrick M.S., Chen A.K., Jessop K. L., 2005. Nitric oxide changes its role as a
modulator of respiratory motor activity during development in the bullfrog (Rana
catesbeiana) Comparative Biochemistry and Physiology Part A: Molecular &
Integrative Physiology, 142 (2), 231-240.
17. Henneman, E., Somjen, G. & Carpenter, D. O., 1965. Functional significance of cell
size in spinal motoneurons. J. Neurophysiol. 28, 560-580.
18. Horcholle-Bossavit, G., Quenet, B., 2009. Neural model of frog ventilatory
rhythmogenesis. Biosystems 97, 35-43.
19. Jorgensen, C.B., 2000. Amphibian respiration and olfaction and their relationships:
from Robert Townson (1794) to the present. Biol. Rev. Camb. Philos. Soc. 75, 297-
345.
13
20. Kato, F., Morin-Surun, M.P. and Denavit-Saubie, M., 1996. Coherent inspiratory
oscillation of cranial nerve discharges in perfused neonatal cat brainstem in vitro. J.
Physiol., 497, 539-549
21. Kinkead, R. and Milsom, W.K., 1994. Chemoreceptors and control of episodic
breathing in the bullfrog (Rana catesbeiana). Respir. Physiol. 95, 81–98.
22. Kocsis, B., Gyimesi-Pelczer, K., Vertes, R.P., 1999. Medium-frequency oscillations
dominate the inspiratory nerve discharge of anesthetized newborn rats. Brain Res. 818,
180–183.
23. Kogo, N., Perry, S.F., Remmers, J.E. 1994a. Neural organization of the ventilatory
activity in the frog, Rana catesbeiana I. J. Neurobiol. 25, 1067-1079.
24. Kogo, N., Remmers, J.E. 1994b. Neural organization of the ventilatory activity in the
frog, Rana catesbeina II. J. Neurobiol. 25, 1080-1094.
25. Lefler Y.,Torben-Nielsen B., Yarom Y., 2013. Oscillatory activity, phase differences,
and phase resetting in the inferior olivary nucleus. Front Syst Neurosci, 7:22-.
26. Lynn, A., Fuerst, W., 2000. Digital Signal Processing, 2nd ed. John Wiley and Sons,
Chichester.
27. Mallat, S., 2000. Une exploration des signaux en ondelettes. Edition de l’Ecole
Polytechnique, Paris.
28. Marchenko, V., Rogers, R.F., 2006. Selective loss of high-frequency oscillations in
phrenic and hypoglossal activity in the decerebrate rat during gasping. Am. J. Physiol.
Regul. Integr. Comp. Physiol. 291,1414-1429.
29. Marchenko, V., Rogers, R.F., 2007. Temperature and state dependence of dynamic
phrenic oscillations in the decerebrate juvenile rat. Am. J. Physiol. Regul. Integr.
Comp Physiol. 293, R2323-35.
30. Mor, Y. and Lev-Tov, A., 2007. Analysis of Rhythmic Patterns Produced by Spinal
Neural Networks. J. Neurophysiol. 98, 2807–2817.
31. O’Neal, I.M.H., Spiegel, E.T., Chon, K.H. and Solomon I.C., 2005. Time-Frequency
Representation of Inspiratory Motor Output in Anesthetized C57BL/6 Mice In Vivo. J.
Neurophysiol. 93, 1762-1775.
32. Richardson, C.A., Mitchell, R.A. 1982. Power spectral analysis of inspiratory nerve
activity in the decerebrate cat. Brain Res., 233, 317-336.
33. Romcy-Pereira, R.N., de Araujo, D.B., Leite, J.P., Garcia-Cairasco, N.A., 2008. Semi-
automated algorithm for studying neuronal oscillatory patterns: A wavelet-based time
frequency and coherence analysis. J. Neurosci. Methods 167, 384–392.
34. Rose G.J. and Fortune E.S., 1999. Mechanisms for generating temporal filters in the
electrosensory system. J Exp Biol. 202: 1281-9.
35. Sakakibara, Y., 1984a. The pattern of respiratory nerve activity in the bullfrog. Jpn J.
Physiol. 34, 269-282.
36. Sakakibara, Y., 1984b. Trigeminal nerve activity and buccal pressure as an index of
total inspiratory activity in the bullfrog. Jpn J. Physiol. 34, 827-838.
37. Straus, C., Samara, Z., Fiamma, M.N., Bautin, N., Ranohavimparany, A., Le Coz, P.,
Golmard, J.L., Darré, P., Zelter, M., Poon, C.S., Similowski, T., 2011. Effects of
maturation and acidosis on the chaos-like complexity of the neural respiratory output
in the isolated brainstem of the tadpole, Rana esculenta. Am. J. Physiol. Regul. Integr.
Comp. Physiol. 300, R1163-1174.
38. Straus C., Wilson R.J.A., Remmers J.E., 2000. Developmental disinhibition: turning
off inhibition turns on breathing in vertebrates. J Neurobiol 45: 75–83.
39. Taylor, A.C., Köllros, J., 1946. Stages in the normal development of Rana pipiens
larvae. Anat. Rec. 94, 7–24.
40. Taylor, B.E., Harris, M.B., Coates, E.L., Gdovin, M.J., Leiter, J.C., 2003a. Central
14
CO2 chemoreception in developing bullfrogs: anomalous response to acetazolamide.
J. Appl. Physiol. 94, 1204-1212.
41. Taylor, B.E., Harris, M.B., Leiter, J.C., Gdovin, M.J., 2003b. Ontogeny of central
CO2 chemoreception: chemosensitivity in the ventral medulla of developing bullfrogs.
Am. J. Physiol. Regul. Integr. Comp. Physiol. 285, R1461–1472.
42. Torgerson, C.S., Gdovin, M.J., Remmers, J.E., 2001a. Sites of respiratory
rhythmogenesis during development in the tadpole. Am. J. Physiol. Regul. Integr.
Comp. Physiol. 280, R913-920.
43. Torgerson, C.S., Gdovin, M.J., Brandt, R., Remmers, J.E., 2001b Location of central
respiratory chemoreceptors in the developing tadpole. Am. J. Physiol. Regul. Integr.
Comp. Physiol. 280: R921-928.
44. Torgerson, C.S., Gdovin, M.J., Remmers, J.E., 1997. The ontogeny of central
chemoreception during fictive gill and lung ventilation of an in vitro brainstem
preparation of Rana catesbeiana. J. Exp. Biol. 299, 2063–2072.
45. Torgerson, C.S., Gdovin, M.J., Remmers, J.E., 1998. Fictive gill and lung ventilation
in the pre- and postmetamorphic tadpole brain stem. J Neurophysiol. 80(4), 2015-22.
46. Vialatte, F.B., Martin, C., Dubois ,R., Quenet, B., Gervais, R., Dreyfus, G. 2007. A
machine learning approach to the analysis of time-frequency maps, and its application
to neural dynamics. Neural Netw. 20, 194-209.
47. Vitalis, T.Z. and Shelton, G.,1990. Breathing in Rana pipiens: the mechanism of
ventilation. J. Exp. Biol. 154, 537–556.
48. Wilson, R.J., Vasilakos, K., Harris, M.B., Straus, C., Remmers, J.E., 2002. Evidence
that ventilatory rhythmogenesis in the frog involves two distinct neuronal oscillators. J
Physiol 540, 557-570.
FIGURE LEGENDS
Figure 1: Neurogram analyses
A) By filtering neurogram Sneur (gray line), two signals are extracted: S100 (black line) and
S10 (red line). The neurogram and the signals are illustrated for a premetamorphic tadpole
neurogram portion during 1.3 seconds that includes two successive buccal cycles,
separated at the level of the local minima of S100 (blue diamonds). Thus, each neurogram
is segmented into successive cycles (red dashed vertical lines).
B) Identification of the spikes by location of the local maxima on the positive (red stars) and
negative (blue stars) parts of the neurogram (gray line, same portion as in A). The green
lines define five spike subpopulations (I to V) according to their amplitude
Figure 2: Cycle alignment and burst limit definition
A) The S10 cycle map is built by positioning and stacking cycles using cross-correlation with
a reference cycle (red line). In this map, the cycle amplitude values are coded in gray
scale. The blue line represents the proportion of zero padding in the map, between 0 and 1.
B) Using the previous positioning of the cycles, spike maps (positive and negative) are built
from the spike identification (Fig. 1). The spike amplitude values are represented in gray
scale. Positive and negative mean profiles are computed from the respective spike
matrices (gray lines); canonical positive and negative cycle amplitude profiles are defined
by filtering the previous mean profiles and by cutting these profiles at their most
significant portion to define a mean cycle, i.e. where the zero-padding proportion is
minimal (between blue lines). The buccal burst is defined as the most active zone of the
cycle (between green lines).
15
Figure 3: CWT Time-frequency maps and spectra from S10 signal
Upper part: S10 signal during 3 seconds for a premetamorphic tadpole (T2) and a
postmetamorphic tadpole (F4) in normocapnia (black line) and hypercapnia (red line) with
oscillations on the buccal bursts, particularly visible for T2.
Lower part: On the CWT time-frequency maps (shown here for T2 and F4 at pH=7.8 and
pH=7.4) of S10 signals with a logarithmic distribution of frequencies in the range [1-100] Hz,
normalized spectra in normocapnia (black lines) and hypercapnia (red lines) are superimposed.
A maximum of the T2 profiles appears around 20-30 Hz corresponding to the oscillation
frequency observed on S10 signal.
Figure 4: Effects of hypercapnia on the CWT spectra of premetamorphic and
postmetamorphic tadpoles
Normalized CWT spectra for premetamorphic tadpoles (T1-T5) and postmetamorphic
tadpoles (F1-F5) at pH=7.8 (black lines) and pH=7.4 (red lines). The values of the local
maximal frequencies are indicated near the corresponding maxima of the profiles.
Figure 5: Effects of hypercapnia on the amplitude profiles of the canonical buccal cycles
Amplitude profiles of two canonical buccal cycles computed from the respective positive
spike matrices on premetamorphic tadpoles (T1-T5) and postmetamorphic tadpoles (F1-F5) at
pH=7.8 (black lines) and pH =7.4 (red lines). The profiles exhibit around 20-30 Hz
oscillations as observed in the corresponding S10 signals. The red horizontal segment indicates
the interburst period lengthening induced by hypercapnia in both premetamorphic and
postmetamorphic tadpoles.
Figure 6: Contributions of the spike subpopulations on the buccal burst shape for T2
The left column shows the spike maps at pH=7.8 (black dots) and pH=7.4 (red dots) with the
corresponding mean number of spikes for each of the first four subpopulations (I to IV) into
bins of 5 ms during a canonical cycle (black line for normocapnia and red line for
hypercapnia).
The green lines delimit the canonical buccal burst (see Fig. 2). In the right column each
amplitude profile is computed from its corresponding spike subpopulation, for pH=7.8 (black
line) and pH=7.4 (red line).
Figure 7: Contributions of the spike subpopulations on the buccal burst shape for F2
Left and right columns exhibit respectively the number of spikes of each of the first four
subpopulations (I to IV) and the corresponding amplitude profiles as in Fig. 6. Note that the
number of buccal cycles is lower in hypercapnia than in normocapnia, for the same recording
time, due to the marked decrease of the buccal frequency and the higher number of lung
episodes which are eliminated from the analysis time.
Figure 8: Overview of hypercapnia effect on the buccal and lung activities in
neurograms of the postmetamorphic tadpoles (F1-F5).
For each postmetamorphic tadpole, the nerve activity recorded during five minutes and
filtered to get whole S100 signals, i.e. with both buccal and lung activities, are represented
according to a grey level scale. The time position of the lung bursts are indicated by vertical
bars, black at pH=7.8 and red at pH=7.4. The grey levels indicate that the buccal activity is
higher at pH=7.8 than at pH=7.4. The number of lung bursts is greater at pH=7.4 than at
pH=7.8, but for F1.
16
Figure 9: Example of lung bursts
Left: typical lung bursts observed in F2 at pH=7.8 (black) and pH=7.4 (red).
In F2, all the bursts at pH=7.8 are preceded by a period of small amplitude activity, while
such a pre-burst activity is absent at pH=7.4.
Right: mean amplitude of the lung burst and the pre-burst activity.
Figure 10: Contributions of spike subpopulations II to V on the lung burst shape for F2
Left and right columns exhibit respectively the number of spikes of each of the last four
subpopulations (II to V) and the corresponding amplitude profiles as in Fig. 6 and 7. Note that
the number of lung bursts is higher in hypercapnia than in normocapnia, for the same
recording time.
FIGURE 1
500
a.u
A
B
250
0
(cid:239)250
500
250
0
(cid:239)250
(cid:239)500
FIGURE 2
N
17
V
IV
III
II
I
0.25
0.5
0.75
time[s]
1
1.25
1
0
300
250
200
150
100
50
600
500
400
300
200
100
A
B
0.25
0.5
0.75
1
0.25
time [s]
0.5
0.75
1
0.25
0.5
FIGURE 3
18
0.05
T3
26
27.2
0.08
F3
1.3
0.04
T4
1.3
2.4
26
23.6
0.08
F4
3.7
2.5
7.4
5.1 20.6
1.6
1.2
2.3
T5
27.2
26
3.5
F5
7.1
2.2
20.6
22.6
4.4
6.4
26
1.5
1
10
100
1
10
100
1
10 100 Hz
FIGURE 4
a.u.
1.3
0.06
1.2
0.06
T1
0.06
T2
1.1
1.3
1.4
2.5
2.9
22.6
26
0.06
0.1
F1
2
2.9
2
1.5
20.6
3.7
3.1
21.5
0.1
F2
2.4
2.2
3.2
5.9
21.5
4.2
21.5
1
10
100
1
6.1
23.6
20.6
100
4.2
10
FIGURE 5
a.u
150
0
300
0
150
0
150
0
100
0
0.5
19
F1
0.12 s
F2
0.18 s
F3
0.32
F4
0.13 s
F5
0.16 s
0.75
0.25
time [s]
200
a.u
0.08 s
T1
0.22 s
T2
0.1 s
T3
T4
T5
0.14 s
0.15 s
1
0
300
0
400
0
400
0
600
0
1.5
time [s]
FIGURE 6
Mean number of spikes per cycle
Mean number of spikes per cycle
0.8
Amplitude a.u.
Amplitude a.u.
0.4
0.8
0.4
0.8
0.4
1.6
0.8
IV
III
II
I
100
50
0
100
50
0
100
50
0
100
50
0
0.25
0.5
time [s]
0
0
0.75
0.25
0.5
time [s]
0.75
FIGURE 7
Mean number of spikes per cycle
Mean number of spikes per cycle
Amplitude a.u.
Amplitude a.u.
20
0.4
0.2
0.4
0.2
0.4
0.2
1.6
0.8
IV
IV
IIIIII
IIII
I
I
100
50
0
100
50
0
100
50
0
100
50
0.3
0
0
0.4
0.1
0.2
time [s]
0.3
0.4
0
0.1
0.2
time [s]
FIGURE 8
21
FIGURE 9
Amplitude a.u.
1000
0
(cid:239)1000
1000
0
Amplitude a.u.
(cid:239)1000
0
0.25
time [s]
0.5
900 Amplitude a.u.
600
300
0
900 Amplitude a.u.
600
300
0
0
0.25
time [s]
FIGURE 10
Mean number of spikes per burst
Mean number of spikes per burst
Amplitude a.u.
Amplitude a.u.
0.5
1
0.5
1
0.5
1
0.5
1
0.5
0
V
V
IVIV
IIIIII
II
II
600
400
200
0
600
400
200
0
600
400
200
0
600
400
200
0.25
time [s]
0.5
0
0
0.25
time [s]
0.5
|
1805.09001 | 6 | 1805 | 2019-04-24T03:07:16 | One-to-one Mapping between Stimulus and Neural State: Memory and Classification | [
"q-bio.NC",
"cs.LG",
"stat.ML"
] | Synaptic strength can be seen as probability to propagate impulse, and according to synaptic plasticity, function could exist from propagation activity to synaptic strength. If the function satisfies constraints such as continuity and monotonicity, neural network under external stimulus will always go to fixed point, and there could be one-to-one mapping between external stimulus and synaptic strength at fixed point. In other words, neural network "memorizes" external stimulus in its synapses. A biological classifier is proposed to utilize this mapping. | q-bio.NC | q-bio | One-to-one mapping between stimulus and neural state:
Memory and classification
Sizhong Lan1
1China Mobile Research Institute, Beijing, 100053, Chinaa)
(Dated: 25 April 2019)
Synaptic strength can be seen as probability to propagate impulse, and according to synaptic plasticity, function could
exist from propagation activity to synaptic strength. If the function satisfies constraints such as continuity and mono-
tonicity, the neural network under external stimulus will always go to fixed point, and there could be one-to-one map-
ping between the external stimulus and the synaptic strength at fixed point. In other words, neural network "memorizes"
external stimulus in its synapses. A biological classifier is proposed to utilize this mapping.
9
1
0
2
r
p
A
4
2
]
.
C
N
o
i
b
-
q
[
6
v
1
0
0
9
0
.
5
0
8
1
:
v
i
X
r
a
I.
INTRODUCTION
Known experiment results show that synaptic connection
strengthens or weakens over time in response to increases or
decreases in impulse propagation1. It is also postulated that
"neurons that fire together wire together"2,3. This biochem-
ical mechanism, called synaptic plasticity4,5, is believed to
play a critical role in the memory formation6 -- 9, although it
is still argued if synapse is the sole locus of learning and
memory10,11. Meanwhile, a synapse propagates impulses
stochastically12 -- 14, which means that synaptic strength could
be measured with the probability of propagating an impulse
successfully. With this probabilistic treatment we find out
that, in the plasticity process a synapse's strength would be
inevitably attracted towards the same fixed point regardless of
its initial strength, and for a neural network there could ex-
ist a one-to-one mapping between the external stimulus from
environment and the synapses' strength at fixed point. This
one-to-one mapping serves the very purpose of ideal mem-
ory: to develop different stable neural state for different stim-
ulus from the environment, and develop the same stable neural
state for the same stimulus no matter what state is initialized
with. It follows that the synapses alone could sufficiently give
rise to persistent memory: they could the sole locus of learn-
ing and memory.
The remainder of paper goes as follows. Section II iden-
tifies the constraints under which synaptic plasticity of one
synaptic connection leads to its fixed state and one-to-one
stimulus-state mapping (memory). Section III extends the
concepts of fixed state and one-to-one mapping for the neu-
ral network consisting of many synaptic connections. Section
IV proposes a simple neural classifier utilizing this memory to
classify handwritten digit images.
II. SYNAPTIC CONNECTION AND ITS FIXED POINT
Let us start with one synaptic connection as shown in FIG 1.
In nature, synapses are known to be plastic, low-precision and
unreliable15. This stochasticity allows us to assume synap-
tic strength s to be the probability (reliability) of propagat-
a)Electronic mail: [email protected]
FIG. 1. A synaptic connection with strength s is directed from neuron
1 to neuron 2. The stimulus from environment or upstream neurons
stimulates neuron 1 to fire action potential with probability x. The
synaptic connection propagates nerve impulse (action potential) to
neuron 2. As a result, neuron 2 is stimulated to fire with probability
y. That is, neuron 1 and 2 fire simultaneously ("fire together") with
probability y.
ing a nerve impulse through, instead of being weight (usu-
ally unbounded real number) as in Artificial Neural Network16
(ANN). Easily we have y=xs where x,s,y∈[0,1]. Now we
treat synaptic plasticity, i.e.
the relation between synaptic
strength s and simultaneous firing probability y, as a function
s∗ = λ (y).
(1)
Here s∗∈[0,1] represents the target value that a connection's
strength will be strengthened or weakened to if the connec-
tion is under constant simultaneous firing probability y (while
s in y=xs represents current strength). By y=xs and Eq. (1),
we have s∗=λ (xs) stating that, under constant stimulus prob-
ability x, the connection initialized with strength s will evolve
towards s∗.
Function λ of Eq.
(1) truly links "firing together" and
"wiring together". For comparison, Hebbian learning rule2
treats synaptic plasticity, in the context of ANN, as a function
∆w=ηxy to learn connections' weight from the training pat-
terns; the function translates "firing together" into "neuron's
input and output both being positive or negative". Different
from ANN, our model actually makes no assumption of neu-
ron being computational unit, and aims to show that with λ
stimulus could sufficiently and precisely control the enduring
fixed state of synaptic connection. The following reasoning
will hinge on this "target strength function" λ , and we will
put constrains on this uncharted function to see how they af-
fect the dynamics of connection strength and most importantly
how stimulus is one-to-one mapped to the strength at fixated
state.
2
FIG. 3. The simulation results of Algorithm 1 for four typical λ func-
tions. For each λ , eleven trails parameteried with incremental initial
strength s0 are run for 105 iterations; all trails share the same stim-
ulus probability x=0.8. In each λ 's subfigure, the left chart depicts
λ (y) as black line, its horizontally scaled λ (xs)=λ (0.8s) in red line
and fixed points as blue dots; the right chart shows the strength tra-
jectories starting from incremental s0. (a) λ (y)=0.9y+0.05. There
exists one single fixed point for λ (0.8s). All strength trajectories
converge to this fixed point. (b) λ (y)=0.5sin(4πy)+0.5. There are
three fixed points for λ (0.8s), two of which are stable ones for the
(c) λ (y)=−y+1. All
trajectories to converge to with oscillation.
trajectories converge to one single fixed point. (d) λ (y) is discontin-
uous at y=0.5. There is no fixed point since there is no s∈[0,1] such
that λ (0.8s)=s, and consequently the trajectories don't converge. In-
stead, they fluctuate within "fixed interval", which as we will see is a
useable compromise of fixed point.
strengthened or weakened. Here the crucial Brouwer's theo-
rem is a fixed-point theorem in topology, which states that, for
any continuous function f (t) mapping a compact convex set
(e.g. interval [0,1] in our case; could be multi-dimensional) to
itself, there is always a point t+ such that f (t+)=t+. More-
over, as illustrated in FIG 2, given any initial value the strength
is always attracted towards fixed point. Therefore, a gentle
constraint of continuity on λ function could preferably drive
synaptic connection to the fixed state.
To verify connection strength's tendency towards fixed
points, we design Algorithm 1 to simulate our connection
model. In this simulation18, recent simultaneous firings are
recorded and the rate is supposed to approximate the simulta-
neous firing probability y; the connection updates its strength
FIG. 2. Two examples of λ (xs) are depicted as red bold lines, and
their fixed points as blue dots. (a) Given any initial s1<λ (xs1), there
must exist a fixed point s+∈(s1,1]; strength s tends to increase from
s1 as long as target strength λ (xs)>s. Given any initial s2>λ (xs2),
there must exist a fixed point s+∈[0,s2); strength s tends to de-
crease from s2 as long as target strength λ (xs)<s. Controlled by
these two tendencies, s will reach and stay at fixed point s+ such
that s+=λ (xs+). (b) There are three fixed points s+
2 ,s+
3 . Start-
ing from any initial s1∈(s+
1 . Starting
from any initial s2∈(s+
3 ,s+
2 . Then strength
tends to leave unstable fixed point s+
1 or s+
2 . Note that
if countable fixed points exist for λ (xs), one of them must be stable.
1 ,s+
2 ), strength increases to s+
3 for stable s+
1 ,s+
3 ), strength decreases to s+
Algorithm 1 connection strength's tendency to fixed points
Input: stimulus probability x, initial synaptic strength s0, target
strength function λ , strength step ∆s, and interations I.
neuron 1 and 2 fire together: recorder[p]←1.
endif
if recorder has been traversed once (i≥104) then
preset current pointed entry of recorder: recorder[p]←0.
pick random r1 and r2 from uniform distribution Uni f (0,1).
if x>r1 and s>r2 then
Output: trajectory of strength s.
1: initialize fire-together recorder (104-entries array): recorder⇐0.
2: initialize fire-together recorder pointer: p←0.
3: initialize current strength: s←s0.
4: for i=0 to I do
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21: end for
end if
forward recorder pointer: p←(p+1) mod 104.
set y with the proportion of 1-entries in recorder.
set target strength: s∗←λ (y).
if s∗>s then
step-increase current strength: s←min(s+∆s,1).
step-decrease current strength: s←max(0,s−∆s).
end if
if s∗<s then
end if
Here is our first constraint: λ is continuous on y. This
constraint is neurobiologically justifiable regarding synaptic
plasticity, since sufficiently small change in impulse probabil-
ity would most probably result in arbitrarily small change in
synaptic strength. In that case, given any x, λ (xs) is a continu-
ous function on s from unit interval [0,1] to unit interval [0,1],
and according to Brouwer's fixed-point theorem17 there must
exist a fixed point s+∈[0,1] such that s+=λ (xs+): connection
strength at s+ will evolve to s+ and hence fixate, no longer
by a small step ∆s=10−4 each iteration to the direction of tar-
get strength. As shown in FIG 3, we run the simulation for
four typical target strength functions, and the strength trajec-
tories resulted show that the constraint of continuity ensures
the tendency towards fixed points given any initial strength.
Our goal is to establish a one-to-one mapping between the
stimulus and the connection strength at fixed point. Specif-
ically, we could (1) given any stimulus x∈[0,1], identify the
fixed point s+ of connection strength without ambiguity; (2)
given any strength s+∈[0,1] at fixed point, identify stimulus
x without ambiguity. Among the four target strength func-
tions in FIG 3, λ (y)=0.9y+0.05 and λ (y)=−y+1 can lead
to one-to-one stimulus-strength mapping. Given any stimulus
x, a synaptic connection equipped with one of these functions
will have one single fixed point of strength regardless of its
initial strength, such that the relation between stimulus and
fixed point strength can be treated as a function s+=θ (x). In
FIG 4, simulation shows that θ could be strictly monotonic
and hence one-to-one mapping from x to s+, such that θ (x)
has one-to-one inverse function θ−1(s+). By contrast, FIG 5
shows that λ (y)=0.5sin(4πy)+0.5 cannot ensure the unique-
ness of fixed point and thus there is no such one-to-one θ (x);
FIG 6 shows that there is no θ either for the discontinuous λ
function in FIG 3(d).
3
5.
simulation
results
The
of Algorithm 1
FIG.
for
λ (y)=0.5sin(4πy)+0.5 in FIG 3(b) to identify the relation be-
(a) Given x=0.3 there is one single fixed point
tween x and s+.
(b) As with x=0.8 in FIG 3(b),
regardless of initial strength s0.
given x=0.9 there are two stable fixed points. Higher s0 converges to
higher fixed point; lower s0 converges to lower one; No convergence
(c) Trails with incremental x
to the middle unstable fixed point.
are run and the averaged s+ values are depicted as in FIG 4. From
x=0.65 and upwards, there are two possible stable fixed points to
converge to depending on what value initial strength is randomized
to, which means that there exists no θ function from x to s+.
FIG. 4. The simulation results of Algorithm 1 to reveal the relation of
stimulus probability x and fixed point strength s+. For each λ , sim-
ulation is parameterized with incremental x (rather than x=0.8 as in
FiG 3) and randomized s0. Ten trails are run for each incremental x,
and the ten converged s values are averaged to be the s+ value corre-
sponding to its input x. The red line represents the averaged s+ values
from simulation, while the blue line represents the true s+=θ (x). (a)
Simulation is parameterized with λ (y)=0.9y+0.05 and the results
match θ (x)=0.05/(1−0.9x) which is monotonically increasing. (b)
Simulation is parameterized with λ (y)=−y+1 and the results match
θ (x)=1/(1+x) which is monotonically decreasing.
In fact, we can pinpoint more constraints on λ as the con-
ditions for function θ to be one-to-one mapping.
In addi-
tion to constraint of continuity, let λ(y) be strictly mono-
tonic on [0, 1] and hence one-to-one; let λ(0)(cid:54)=0 to rule out
fixed point s+=0. In that case, λ has inverse function λ−1(s)
which is strictly monotonic between λ (0) and λ (1), and given
any fixed point strength s+ between we can identify stimu-
lus x=λ−1(s+)/s+. That is, function θ−1(s+) = λ−1(s+)/s+
exists. Let λ−1(s)/s be strictly monotonic between λ(0)
and λ(1). Then given any stimulus x∈[0,1] there is one sin-
gle fixed point s+ such that x=λ−1(s+)/s+. That is, function
s+=θ (x) exists. Both of λ (y)=0.9y+0.05 and λ (y)=−y+1
FIG. 6. The simulation results to find the θ function with respect to
the discontinuous λ in FIG 3(d). When x(cid:38)0.6, strength can evolve to
any point within a "fixed interval" each time simulation is finished.
The absence of fixed point doesn't allow the existence of θ.
obey all those constraints and their one-to-one θ functions
can be verified by the simulation results in FIG 4, whereas
λ (y)=0.5sin(4πy)+0.5 is not even strictly monotonic. How-
ever, neither λ (y)=0.9y+0.05 nor λ (y)=−y+1 is ideal for
our purpose. Guided by these constraints, we choose λ func-
tion carefully such that its derived θ (x) function is monoton-
ically increasing and range of which spans nearly the entire
[0,1] interval, as shown in FIG 7. Of all the λ constraints,
continuity and strong monotonicity are reasonable require-
ments of consistency on the neurobiological process of synap-
tic plasticity, whereas λ (0)(cid:54)=0 and strong monotonicity of
λ−1(s)/s are rather specific and peculiar claims. Admittedly,
those λ constraints need to be supported by experimental evi-
dences.
Now we have the one-to-one (continuous and strictly mono-
tonic) functions λ , λ−1, θ and θ−1, and in those functions ini-
tial strength s0 is irrelevant. Given s+ we can identify x and y
without ambiguity, and vice versa. Our interpretation of these
mappings is, the synaptic connection at fixed point precisely
"memorizes" the information of what (stimulus) it senses and
4
Stimulus and strength uniquely determine impulses prop-
agation within neural network, so there exists a mapping
Ψ:(X,S)→Y . Presumably, the mapping Ψ is continuous on
(1), there exists a mapping Λ:Y→S∗ such that
S. By Eq.
s∗
i j=λi j(yi j) for each yi j in Y and its counterpart s∗
i j in S∗.
Here S∗∈[0,1]c is c-dimensional vector of connections' tar-
get strength, and mapping Λ could be visualized as a vec-
tor of target strength functions such that entry Λi j is λi j.
Then with mapping Ψ and Λ we have a composite mapping
Λ◦Ψ:(X,S)→S∗.
If each λi j function is continuous on its
yi j, mapping Λ◦Ψ must be continuous on S and according to
Brouwer's fixed-point theorem given X there must exist one
fixed point S+∈[0,1]c such that Λ◦Ψ(X,S+)=S+. And under
constant stimulus X, neural network will go to fixed point S+
as each connection i(cid:32) j goes to its fixed point s+
i j. Our simu-
lation verifies that tendency as shown in FIG 9. In this simula-
tion, impulses traverse the neural network stochastically such
that each neuron is fired at most once per iteration; synaptic
connections update their strength as in Algorithm 1.
√
FIG. 7. (a) Our choices of λ functions are λL(y)=0.99
y+0.01 in
1+e−4.4(y+0.01) −1 in green. Here λT is a segment of
blue and λT (y)=
shifted and scaled Sigmoid function. They both obey the λ con-
straints as discussed previously. (b) Simulation results show that, λL
leads to linear-like θL in blue such that θL(x)≈x, and λT leads to
threshold-like θT in green.
2
how it responses (with impulse propagation).
III. NEURAL NETWORK AND ITS FIXED POINT
FIG. 8. A neural network (of one or multiple agents) consists of n≥2
neurons and c≥1 directed synaptic connections. An example of n=8
and c=7 is depicted. Each neuron receives stimulus from the envi-
ronment with probability and propagates out nerve impulses through-
out the synaptic connections, e.g., triggered by stimulus neuron 1
propagates impulses stochastically down along the directed paths
1(cid:32)7(cid:32)8(cid:32)2 and 1(cid:32)7(cid:32)8(cid:32)5(cid:32)3. Cyclic path (e.g. 3(cid:32)8(cid:32)5(cid:32)3)
is allowed and yet loop (e.g. 3(cid:32)3) isn't. Each neuron could have
either outbound or inbound connections, or neither, or both.
Now let us turn to the neural network shown in FIG 8. A
neural network could be treated as an "aggregate connection"
as it turns out. We shall see that, the definitions and reasoning
for neural network align well with neural connection in last
section.
As with synaptic connection, we can describe a neural net-
work by defining (1) the external stimulus as an n-dimensional
vector X∈[0,1]n in which each xi is the probability of neuron
i receiving stimulus; (2) the strength of all connections as a c-
dimensional vector S∈[0,1]c in which each si j is the strength
of connection from neuron i to neuron j (denoted as i(cid:32) j); (3)
the simultaneous firing probabilities over all connections as a
c-dimensional vector Y∈[0,1]c in which each yi j is the simul-
taneous firing probability over i(cid:32) j. In fact, one single neural
connection is a special case of neural network with c=1 and
n=2.
FIG. 9. Simulation results of neural network's tendency for the four
typical λ functions as in FIG 3. The neural network has n=8 and
c=19. The following observations hold true for any external stimu-
lus and connections configuration: (a) If all connections are equipped
with λ (y)=0.9y+0.05, the whole neural network has one single fixed
point and the trajectories of mean of all connections' strength con-
verge to one point. (b) λ (y)=0.5sin(4πy)+0.5. Because each con-
nection has two stable fixed points, there are 219 stable fixed points
for the whole neural network and 20 possible convergence points of
strength mean. (c) λ (y)=−y+1. There is one single fixed point for
the neural network. The trajectories converge to one point. (d) Dis-
continuous λ . The neural network has no fixed point as each synaptic
connection has no fixed point. The trajectories don't converge.
Generally the number of stable fixed points for a neural net-
work is ∏c fi j where each fi j is the number of stable fixed
points of i(cid:32) j. As in FIG 9(b), ∏c fi j can be enormous when
each fi j ≥ 2. As with synaptic connection, our goal is to estab-
lish one-to-one mapping between stimulus X and fixed point
S+ for neural network and meanwhile keep initial strength S0
out of picture. λ 's continuity alone cannot ensure the unique-
ness of fixed point, such that S0 can determine which fixed
point to go for. Now with all the λ constraints, we have:
(1) Λ is a one-to-one mapping and thus has inverse mapping
Λ−1:S∗→Y ; (2) there exists a mapping Θ:X→S+, because un-
der stimulus X the neural network will go to the same unique
fixed point S+ no matter what initial strength S0 to begin with;
(3) if Θ is a one-to-one mapping, Θ has inverse mapping
Θ−1:S+→X. With mapping Λ, Λ−1, Θ and Θ−1 being one-
to-one, given S+ we can identify X and Y without ambiguity,
and vice versa. Therefore, the same interpretation with respect
to synaptic connection could apply here: the neural network
at fixed point precisely "memorizes" the information about the
stimulus on many neurons and the impulse propagation across
many connections.
Nevertheless, even all of λ constraints are not sufficient to
secure one-to-one Θ:X→S+ for a neural network, as opposed
to the neural connection. Here is a case. For Θ to be one-
to-one, all neurons must have outbound connection. Other-
wise, e.g., for a neural network with three neurons (say 0,
1 and 2) and two connections (say 0(cid:32)1 and 1(cid:32)2), stimu-
lus X1=(1,1,0) and X2=(1,1,1) will result in the same fixed
point because stimulus on neuron 3, no matter what it is, af-
fects no connection. Or equivalently, for Θ to be one-to-one,
the definition of X should consider only the neurons with out-
bound connections such that X's dimension dim(X)≤n.
In
the perspective of information theory19, many-to-one Θ in-
troduces equivocation to the neural network at fixed point,
as if information loss occurred due to noisy channel.
If
dim(X)>dim(S)=c, mapping Θ conducts "dimension reduc-
tion" on stimulus X, and information loss is bound to occur.
Here is a trivial case regarding stimulus dependence. Con-
sider a neural network with 0(cid:32)2, 1(cid:32)2 and 2(cid:32)3, and stim-
ulus X=(x0,x1). When the neural network is at fixed point,
12x0 Pr(10) where Pr(10) is the proba-
x2=x0s+
bility of neuron 1 being stimulated conditional on neuron 0
being stimulated. Pr(10)(cid:54)=x1 if stimulus on neuron 1 and
2 are not independent. Pr(10) affects s+
23 and hence S+, or
in other words the neural network at fixed point gains the
hidden information of Pr(10). However, if Pr(10) varies,
given mere X there will be uncertainty about S+ such that
mapping Θ doesn't exist unless stimulus X is "augmented"
to X=(x0,x1,Pr(10)).
12−s+
02+x1s+
02s+
5
its particular training stimulus, and then a testing stimulus is
tested on all g neural networks independently to see which
gets the most connections propagated. For simplicity we as-
sume testing itself doesn't jeopardize the fixed points of neural
networks. And most importantly we assume that for each neu-
ral network given any stimulus there is one single fixed point
such that mapping Θ:X→S+ exists.
Consider a neural network in the classifier to be trained
by X to fixed point S+ and then tested by X.
In other
words, neural network memorizing X as S+ is tested by
X. Because impulses propagate across the neural network
stochastically, the count of synaptic connections propagated
in one test should be random variable. Let it be Z XX.
Then for the neural network in FIG 8 Z XX =∑c zi j where
each r.v. zi j∼Bernoulli(xis+
i j ), i.e., synaptic connection i(cid:32) j
is propagated with probability xis+
in the test such that
i j
i j. Easily zi j's expected value is E[zi j]=xis+
Pr(zi j=1)=xis+
i j,
and its variance is Var(zi j)=xis+
i j ). By central limit
theorem, Z's distribution could tend towards Gaussian-like
(bell curve) as c increases, even if all zi j are not independent
and identically distributed. We have
i j (1−xis+
E[Z XX ] =
c
∑
i(cid:32) j
E[zi j] =
c
∑
i(cid:32) j
xis+
i j .
(2)
And when c is large,
Var(Z XX ) ≈ c
∑
i(cid:32) j
Var(zi j) =
c
∑
i(cid:32) j
i j (1− xis+
i j ).
xis+
(3)
For any i(cid:32) j, in the training stage because S+=Θ( X) we
i j =θi j( X), and in the testing stage xi is uniquely deter-
have s+
mined by S+ and X such that xi is a function of X and X.
IV. AN APPLICATION FOR CLASSIFICATION
Ideally, a neural network with memory of stimulus X --
formally, mapping Θ casts memory of stimulus X as fixed
point S+ -- should response to stimulus X more "intensely"
than the neural network with different memory responses to
X. Memory would manifest itself as impulses propagation
throughout ensemble of neurons2,20 -- 22. Thus, it is natural to
differentiate response by counting the neurons fired or synap-
tic connections propagated by impulses. Given the reasoning
that synapse could be the sole locus of memory10,11, we adopt
the count of synaptic connections propagated as a macro-
scopic measure of how intensely memory responses to stim-
ulus or stimulus "recalls" memory. And accordingly we pro-
pose a classifier consisting of g neural networks, which clas-
sifies stimulus into one of g classes by the decision criteria
of which neural network gets the most synaptic connections
propagated. Reminiscent of supervised learning23, each neu-
ral network of our classifier is trained to its fixed point by
FIG. 10. The depicted neural network is basically the general one
in FIG 8 except that, to mimic real-life nervous system, an array of
sensor neurons are specialized for receiving stimulus from no other
neurons but the environment. There are 64 sensor neurons to accom-
modate 8×8-pixel image, and the rest are a cluster of 50 neurons.
Each sensor neuron has 6 outbound connections towards cluster, and
each cluster neuron has 5 outbound connections towards within clus-
ter. Connections are randomly put between neurons before training.
We experiment with this classifier to classify handwritten
digit images24. Ten identical neural networks (hence g=10)
of FIG 10, each designated for a digit from 0 to 9, are trained
to their fixed points by their training images in FIG 11 as
stimulus, and then testing images, also as stimulus, are clas-
sified into the digit whose designated neural network gets the
FIG. 11. A digit image has 8×8=64 pixels, and pixel grayscale is
normalized to the value between 0 and 1 (by dividing 16) as stimulus
probability. The upper row shows samples of digit images, and the
lower row shows the better written "average images", each of which
is actually pixel-wise average of a set of images of a digit. Each neu-
ral network is trained in each iteration by the same "average image",
or equivalently in each iteration by image randomly drawn from the
set of images.
biggest Z value. We run many tests to evaluate classification
accuracy, and collect Z values to approximate r.v. Z's dis-
tribution. With all synaptic connections equipped with λL in
FIG 7, the classifier has accuracy ∼44%, and ∼51% with λT .
Note that, equipped with λL or λT , the neural network of FIG
10 will have one-to-one ΘL or ΘT according to last section.
FIG 12 and FIG 13 show that, in positive testing (e.g. digit-6
image is tested in neural network trained by digit-6 images),
Z's expected value (sample mean) could be considerably big-
ger than that in negative testing (e.g. digit-6 image is tested
in neural network trained by digit-1 images), so as to discrim-
inate digit-6 images from the others. Given the same testing
image classification target can be different test by test since
the ten Z outcomes are randomized. To improve classifica-
tion accuracy, we shall distance the distribution of positive
testing Z as far as possible from those of negative testing Z.
We present another two special neural networks in FIG 14 to
demonstrate how our classifier utilizes memory to classify im-
ages and how to improve its accuracy in the neurobiological
way.
FIG. 12. The histogram (in probability density form) of Z. To collect
Z values, a digit-6 image is tested many times on each of the ten
trained neural networks. All connections are equipped with λT . Z66
of positive testing is in red, and the other nine Zk6 of negative testing,
where k=0,1,2,3,4,5,7,8,9, are in gray. Z's sample mean for each
digit is depicted as vertical dotted line.
When the classifier adopts ten neural networks of FIG 14(a)
and equips all connections with λL in FIG 7, classification ac-
curacy is ∼31% and Z's distribution for testing digit-6 im-
ages is shown in FIG 15(a). We already know that λL makes
θL(x)≈x. Then for one test we have
E[Z XX ] =
64
∑xiθi( xi) =
64
∑xiθL( xi) ≈ 64
∑xi xi = X(cid:124)X.
(4)
6
FIG. 13. The histograms of Z for all ten digits. For each digit, ran-
domly drawn testing image, instead of the same one, is used in each
test. From digit 0 to 9, classification accuracy is approximately 70%,
41%, 56%, 42%, 53%, 33%, 77%, 51%, 57% and 32%. Generally,
better Z-distribution separation of positive and nagative testing re-
sults in higher classification accuracy.
(2) and Eq.
i =θi( xi). By Eq.
FIG. 14. These two neural networks inherit the sensor-cluster struc-
ture of FIG 10. (a) Each sensor neuron connects to one single clus-
ter neuron such that each pixel stimulus xi only affects one single
connection. Then s+
(3), we have
E[Z XX ]=∑64 xiθi( xi) and Var(Z XX )≈∑64 xiθi( xi)[1 − xiθi( xi)]. (b)
Each sensor neuron connects to its own dedicated cluster of many
neurons and synaptic connections, and the clusters are of different
sizes. In that case, in a test each xi causes ωi (instead of just one)
synaptic connections to be propagated with probability xis+
i or none
with probability 1−xis+
i . When each ωi is a nonrandom variable,
we have E[Z XX ]=∑64 xiθi( xi)ωi and Var(Z XX )≈∑64 xiθi( xi)[1 −
xiθi( xi)]ω2
i .
Here X(cid:124)X is the dot product of training vector X∈[0,1]64
and testing vector X∈[0,1]64. Generally, the dot product of
two vectors, a scalar value, is essentially a measure of simi-
larity between the vectors. The bigger E[Z XX ] is, the more in-
tensely neural network with memory of training X responses
to testing X, and the more similar X and X are to each other.
Therefore, Eq. (4) simply links otherwise unrelated neural re-
sponse intensity and stimulus similarity. By comparing ten
E[Z XX ] values, we can tell which X is the most similar to X
and hence which digit is classification target. Only, Z XX value
from test actually deviates around the true E[Z XX ] randomly,
which makes it a useable and yet unreliable classification cri-
teria.
When the classifier equips all connections with threshold-
like λT in FIG 7, classification accuracy raises to ∼44%. By
comparing FIG 15(b) with FIG 15(a), the distance between
Z66's distribution and the other nine Zk6,k(cid:54)=6's distribution is
bigger with threshold-like λT than with linear-like λL. This
accuracy improvement can be explained conveniently with a
true threshold function (or step function)
(cid:40)
θstep(x) =
0, 0 ≤ x < xstep
xstep ≤ x ≤ 1
1,
.
Of the sum terms in ∑64 xi xi of Eq. (4), θstep basically di-
minishes small xi∈[0,xstep) to 0 and enhances big xi∈[xstep,1]
to 1, such that most probably E[Z66] would increase by hav-
ing xi=1 in the sum terms with big xi while the other nine
E[Zk6,k(cid:54)=6] would decrease by having xi=0 in the sum terms
with big xi, so as to preferably increase E[Z66]−E[Zk6,k(cid:54)=6].
And likewise Var(Z) would most probably decrease. As a re-
sult, θstep increases the distance between the distribution of
Z66 and Zk6,k(cid:54)=6 and thus better separates them.
FIG. 15. The Z histogram of testing digit-6 images for three dif-
ferent classifier settings. For each classifier setting, values of Z66
and Zk6,k(cid:54)=6 are transformed to z-scores (i.e. the number of standard
deviations from the mean a data point is) with respect to the distri-
bution of all Zk6,k(cid:54)=6's values combined. The distance between the
distribution of Z66 and Zk6,k(cid:54)=6 is approximately evaluated by E[Z66]
and standard deviation σ (Z66). (a) Classifier with neural networks
of FIG 14(a) and λL. E[Z66]≈0.93 and σ (Z66)≈0.98. (b) Classi-
fier with neural networks of FIG 14(a) and λT . E[Z66]≈1.33 and
σ (Z66)≈0.9. (c) Classifier with neural networks of FIG 14(b), λL
and ωi( xi)=100 xi
3. E[Z66]≈1.39 and σ (Z66)≈0.87.
FIG 14(b) provides another type of neural network to im-
prove classification accuracy without adopting threshold-like
λ function for all synaptic connections. Let the linear-like
3 simply for example.
λL be equipped back and take ωi=100 xi
With this setting our classifier has accuracy ∼47%. Here we
7
TABLE I. Classification accuracy on different classifier settings. In
each setting, neural network can be the one illustrated in FIG 10,
FIG 14(a) or 14(b), and target strength function can be one of those
in FIG 3 and FIG 7. The accuracy listed is the average of many out-
comes taken from the same trained classifier, and thus could fluctuate
slightly from one training to another.
λ or θ functions
λL(y)=0.99
λT (y)=
θstep
λ (y)=0.9y+0.05
λ (y)=0.5sin(4πy)+0.5
λ (y)=−y+1
Discontinuous λ
FIG 14(b)
47%
51%
60% b
19%
2%
1%d
28%
FIG 14(a)
31%
44%
48%a
16%
6%
5%
20%
FIG 10
44%
51%
-
14%
5%c
4%
23%
√
y+0.01
1+e−4.4(y+0.01) −1
2
a xstep is set to 0.6.
b xstep is set to 0.2.
c Accuracy under 10% is actually worse than wild guessing.
d If classification criteria is changed to "which neural network gets the
fewest synaptic connections propagated", the accuracy will be ∼40%.
4) where 100 xi
0.1) (here 4√
have E[Z XX ]≈∑64 xi(100 xi
transforms xi∈[0, 4√
and transforms xi∈[ 4√
4, like θstep, actually
0.1≈0.56) to within [0,10)
0.1,1] to across [10,100] -- again, the
strong training pixel-stimulus are greatly weighted while the
weak ones are relatively suppressed. As shown in FIG 15(c)
the distance between the distribution of Z66 and Zk6,k(cid:54)=6 is in-
creased compared to FIG 15(a). Here our neurobiological in-
3 is, the training stimulus af-
terpretation regarding ωi=100 xi
fects not only synaptic strength, but also the growth of neuron
cluster in the replication of neuron cells and in the formation
of new synaptic connections. Again this claim needs to be
supported by evidences.
TABLE I summarizes the performance of our classifier with
different types of neural networks and target strength func-
tions. The four typical λ functions in FIG 3 are also evalu-
ated to demonstrate how these somewhat "pathological" target
strength functions affect classification.
By Eq.
(4), the classification of handwritten digit im-
ages could be simplified to a task of restricted linear
classification23: given ten classes each with its discriminative
(cid:124)
i X where X,X∈[0,1]64, image X is classi-
function δi(X)= X
fied to the class with the largest δi value. Our neural classi-
fier simply takes over the computation of vectors' dot product
(cid:124)
X
i X and adds randomness to the ten results. To parameterize
the ten δi with their Xi, the "supervisors" could train the neural
networks in classifier with the images they deem best -- "av-
erage images" in our case or digits learning cards in teachers'
case. Our neural classifier is rather unreliable and primitive
compared to ANN which is also capable of linear classifica-
tion. On one hand, given the same image ANN always outputs
the same prediction result. On the other hand, ANN is not
only a classifier but also more importantly a "learner", which
learns from all kinds of handwritten digits to find the opti-
mal Xi for the ten δi; ANN with optimal Xi is more tolerant
with poor handwriting, and thus has less misclassification and
better prediction accuracy. Only, ANN's learning optimal Xi,
an optimization process of many iterations, requires massive
computational power to carry out, which is unlikely to be pro-
vided by the real-life nervous system -- there is no evidence
that an individual neuron can even conduct basic arithmetic
operations. Despite of its weakness, our neural classifier has
merit in its biological nature: it reduces the computation of
vectors' dot product to simple counting of synaptic connec-
tions propagated; its training and testing could be purely neu-
robiological development and activities where no arithmetic
operation is involved; its classification criteria, i.e. "deciding"
or "feeling" which neural (sub)network has the most connec-
tions propagated, could be an intrinsic capability of intelligent
agents. This classifier might project new insights on the neural
reality, hopefully.
V. CONCLUSION
This paper proposes a mathematical theory to explain how
memory forms and works. It all begins with synaptic plas-
ticity. We find out that, synaptic plasticity is more than im-
pulses affecting synapses; it actually plays as a force that can
drive neural network eventually to a long-lasting state. We
also find out that, under certain conditions there would be a
one-to-one mapping between the neural state and the external
stimulus that neural network is exposed to. With the map-
ping, given stimulus we know exactly what neural state will
be; given neural state we know precisely what stimulus has
been. The mapping is essentially a link between past event and
neural present; between the short-lived and the enduring. In
that sense, the mapping itself is memory, or the mapping casts
memory in neural network. Next, we study how memory af-
fects neural network's response to stimulus. We find out that,
the neural network with memory of stimulus can response to
similar stimulus more intensely than to the stimulus of less
similarity, if response intensity is evaluated by the number of
synaptic connections propagated by impulses. That is to say,
a neural network with memory is able to classify stimulus. To
verify this ability, we experiment with the classifier consist-
ing of ten neural networks, and they turn out to have consider-
able accuracy in classifying the handwritten digit images. The
classifier proves that neurons could collectively provide fully
biological computation for classification.
Our reasoning takes root in the mathematical treatment of
synaptic plasticity as target strength function λ from impulse
frequency to synaptic strength. We put hypothetical con-
straints on this λ function to ensure that the ideal one-to-one
mapping exists. Although these constraints are necessary to
keep our theory mathematically sound, they raise concerns.
Firstly, they could be overly restrictive. Take continuity con-
straint for example. Even the discontinuous function of FIG
3(d), whose nonexistent θ function would map certain stimu-
lus to any point within a "fixed interval" instead of a specific
fixed point as shown in FIG 6, can be a useable λ in our clas-
sifier according to TABLE I. In this case, fixed point per se
doesn't have to exist, and mere tendency to seek out for it
8
could serve the purpose. Secondly, as discussed in Section II
those λ constraints have yet to be supported by neurobiolog-
ical evidences. Above all, the evidence that reveals true λ is
vital to clarify the uncertainty.
VI. ACKNOWLEDGEMENT
We thank the anonymous reviewers for their comments that
improved the manuscript.
1J. Hughes. Post-tetanic potentiation. Physiol Rev., 38(1):91 -- 113, 1958.
2D. Hebb. The organization of behavior: a neuropsychological theory. John
Wiley & Sons, 1949.
3S. Lowel and W. Singer. Selection of intrinsic horizontal connections in the
visual cortex by correlated neuronal activity. Science, 255(5041):209 -- 212,
1992.
4G. Berlucchi and H. Buchtel. Neuronal plasticity: historical roots and evo-
lution of meaning. Experimental Brain Research, 192(3):307 -- 19, 2009.
5A. Citri and R. Malenka. Synaptic plasticity: multiple forms, functions,
and mechanisms. Neuropsychopharmacology, 33:18 -- 41, 2008.
6S. Martin, P. Grimwood, and R. Morris. Synaptic plasticity and memory:
an evaluation of the hypothesis. Annu Rev Neurosci., 23(649-711), 2000.
7E. Takeuchi, A. Duszkiewicz, and R. Morris. The synaptic plasticity and
memory hypothesis: encoding, storage and persistence. Philos Trans R Soc
Lond B Biol Sci., 369(1633), 2014.
8S. Nabavi, R. Fox, C. Proulx, J. Lin, and R. Tsien R. Malinow. Engineering
a memory with ltd and ltp. Nature, 511(348 -- 352), 2014.
9Y. Yang, D. Liu, W. Huang, J. Deng, Y. Sun, Y. Zuo, and M. Poo. Selective
synaptic remodeling of amygdalocortical connections associated with fear
memory. Nat. Neurosci., 19(1348 -- 1355), 2016.
10P. Trettenbrein. The demise of the synapse as the locus of memory: a
looming paradigm shift? Front. Syst. Neurosci., 10, 2016.
11J. Langille and R. Brown. The synaptic theory of memory: a historical
survey and reconciliation of recent opposition. Front. Syst. Neurosci., 12,
2018.
12C. Laing and G. Lord. Stochastic methods in neuroscience. Oxford Univer-
sity Press, 2010.
13G. Deco, E. Rolls, and R. Romo. Stochastic dynamics as a principle of
brain function. Progress in neurobiology, 88(1):1 -- 16, 2009.
14T. Branco and T. Staras. The probability of neurotransmitter release:
variability and feedback control at single synapses. Nat Rev Neurosci.,
10(5):373 -- 83, 2009.
15C. Baldassi, F. Gerace, H. Kappen, C. Lucibello, L. Saglietti, E. Tartaglione,
and R. Zecchina. Role of synaptic stochasticity in training low-precision
neural networks. Phys. Rev. Lett., 120(26), 2018.
16J. Schmidhuber. Deep learning in neural networks: an overview. Neural
Networks, 61:85 -- 117, 2015.
17I. Istratescu. Fixed point theory: an introduction. Springer, 1981.
18Source code can be found at https://github.com/lansiz/neuron.
19C. Shannon. A mathematical theory of communication. The Bell System
Technical Journal, 27(3):379 -- 423, 1948.
20G. Buzsaki. Neural syntax: cell assemblies, synapsembles and readers.
Neuron, 63(362 -- 385), 2010.
21C. Butler, Y. Wilson, J. Gunnersen, and M. Murphy. Tracking the fear
memory engram: discrete populations of neurons within amygdala, hy-
pothalamus and lateral septum are specifically activated by auditory fear
conditioning. Learn. Mem., 22(370 -- 384), 2015.
22C. Butler, Y. Wilson, J. Oyrer, T. Karle, S. Petrou, J. Gunnersen, M. Mur-
phy, and C. Reid. Neurons specifically activated by fear learning in lateral
amygdala display increased synaptic strength. eNeuro, 2018.
23T. Hastie, R. Tibshirani, and J. Friedman. The elements of statistical learn-
ing. Springer, 2009.
24The dataset of 1797 handwritten digit images can be obtained with Python
code "from sklearn import datasets; datasets.load_digits()".
|
1608.03619 | 1 | 1608 | 2016-08-11T21:09:40 | Modular Segregation of Structural Brain Networks Supports the Development of Executive Function in Youth | [
"q-bio.NC"
] | The human brain is organized into large-scale functional modules that have been shown to evolve in childhood and adolescence. However, it remains unknown whether structural brain networks are similarly refined during development, potentially allowing for improvements in executive function. In a sample of 882 participants (ages 8-22) who underwent diffusion imaging as part of the Philadelphia Neurodevelopmental Cohort, we demonstrate that structural network modules become more segregated with age, with weaker connections between modules and stronger connections within modules. Evolving modular topology facilitated network integration, driven by age-related strengthening of hub edges that were present both within and between modules. Critically, both modular segregation and network integration were associated with enhanced executive performance, and mediated the improvement of executive functioning with age. Together, results delineate a process of structural network maturation that supports executive function in youth. | q-bio.NC | q-bio |
Modular Segregation of Structural Brain Networks Supports the Development of Executive Function in Youth Graham L. Baum,a Rastko Ciric,a David R. Roalf,a Richard F. Betzel,c Tyler M. Moore,a Russell T. Shinohara,b Ari E. Kahn,c Megan Quarmley,a Philip A. Cook,d Mark A. Elliott,d Kosha Ruparel,a Raquel E. Gur,a Ruben C. Gur,a Danielle S. Bassett,c,e* & Theodore D. Satterthwaite a* Departments of Psychiatry,a Biostatistics and Epidemiology,b Bioengineering,c Radiology,d and Electrical and Systems Engineeringe of the University of Pennsylvania, Philadelphia PA 19104 USA *=Co-senior authors Address correspondence to: Theodore D. Satterthwaite, M.D., M.A. 10th floor, Gates Building Hospital of the University of Pennsylvania 34th and Spruce Street Philadelphia, PA 19104
SUMMARY: The human brain is organized into large-scale functional modules that have been shown to evolve in childhood and adolescence. However, it remains unknown whether structural brain networks are similarly refined during development, potentially allowing for improvements in executive function. In a sample of 882 participants (ages 8-22) who underwent diffusion imaging as part of the Philadelphia Neurodevelopmental Cohort, we demonstrate that structural network modules become more segregated with age, with weaker connections between modules and stronger connections within modules. Evolving modular topology facilitated network integration, driven by age-related strengthening of hub edges that were present both within and between modules. Critically, both modular segregation and network integration were associated with enhanced executive performance, and mediated the improvement of executive functioning with age. Together, results delineate a process of structural network maturation that supports executive function in youth. KEYWORDS: development, network, connectome, adolescence, executive, module ! 1 INTRODUCTION Modularity is a fundamental feature of complex systems, including social groups, cyber-physical systems, and diverse biological networks (Newman, 2006). A network module is a group of densely interconnected nodes, which often are the basis for specialized subunits of information processing (Sporns and Betzel, 2016). Functional neuroimaging studies have demonstrated that the human brain has a well-defined modular organization, as reflected in the presence of large-scale functional networks (Biswal et al., 1995; Damoiseaux et al., 2006; Power et al., 2011; Yeo et al., 2011). While the exact number and spatial distribution of functional network modules varies somewhat by analytic approach, a remarkable convergence exists across independent datasets and laboratories (Damoiseaux et al., 2006; Power et al., 2011; Yeo et al., 2011). Commonly described modules include somatomotor (Biswal et al., 1995), visual (Corbetta et al., 1998), default mode (Raichle et al., 2001; Buckner et al., 2008), and fronto-parietal systems (Dosenbach et al., 2007; Vincent et al., 2008). While brain network modules emerge very early in life (Fair et al., 2008; van den Heuvel et al., 2015; Thomason et al., 2013), a growing body of work has shown that these functional modules are refined during youth. During childhood and adolescence, functional modules become more distinct: connectivity within modules increases while connectivity between modules declines (Power et al., 2010; Fair et al., 2007; Fair et al., 2008; Fair et al., 2009; Dosenbach et al., 2010; Gu et al., 2015; Satterthwaite et al., 2013b; Supekar et al., 2009; Anderson et al., 2011). Such development allows for functional specialization, reducing interference among systems (Fornito et al., 2012) and facilitating cognitive performance (Hampson et al., 2010). Modularity is particularly relevant for executive function, which relies on co-activation of executive regions and reciprocal suppression of non-executive regions such as the default mode network (Anticevic et al., 2010; Barber et al., 2013; Persson et al., 2007; Satterthwaite et al., 2013a). Thus, available data suggests that development of network modularity may serve as a substrate for the evolution of executive capability during youth. Despite convergent evidence for the developmental emergence of functional network modularity, there is relatively scant data regarding the maturation of underlying structural brain networks that support this functional architecture (Hermundstad et al., 2014). Prior work demonstrates substantial correspondence between functional and structural measures of brain connectivity (Goñi et al., 2014; Honey et al., 2009; Mišić et al., 2016), although structural connections tend to be a subset of densely connected, polysynaptic functional networks (Betzel et al., 2014; Hermundstad et al., 2013). Structural networks in adults are highly modular (Bassett et al., 2010; Bassett et al., 2011), but it remains unknown if this topology evolves substantially during youth. Correspondence between functional and structural data intuitively suggests that, like functional networks, structural networks should become increasingly segregated during development. However, prior studies using relatively small samples report conflicting results, including declining modularity (Chen et al., 2013), increasing modularity (Chen and Deem, 2015; Huang et al., 2015), or no change with age (Hagmann et al., 2010b; Lim et al., 2015). Larger sample sizes may be necessary for resolving the variability of findings reported in previous studies. ! 2 Beyond this mixed data regarding normative developmental trends, the impact of structural network development on cognitive performance remains poorly described. Cognitive capability improves substantially during youth, with executive function undergoing a protracted phase of development throughout adolescence and young adulthood (Gur et al., 2012; Luna et al., 2004). Describing how structural brain networks evolve to support executive function is necessary to understand the basis for many sources of adolescent morbidity and mortality, which are prominently associated with failures of executive function (Shamosh et al., 2008; Casey et al., 2008; Casey et al., 2011). Finally, such data are a prerequisite for studies of neuropsychiatric disorders, which are increasingly understood as disorders of brain development (Insel, 2010; Rapoport et al., 2012; Casey et al., 2014), are marked by executive dysfunction (Shanmugan et al., 2016), and are linked to the disruption of evolving network topology (Alexander-Bloch et al., 2012; Di Martino et al., 2014; Castellanos and Proal, 2012; Voineskos et al., 2010). Here we sought to define the normative development of structural network modules, and delineate the impact of modular maturation on executive functioning. We tested the hypothesis that modules within structural brain networks become more segregated with age, as seen in functional brain networks. Further, we predicted that segregated structural modules would support enhanced executive functioning. To address these hypotheses, we capitalized upon a large sample of 882 youths who completed diffusion imaging as part of the Philadelphia Neurodevelopmental Cohort (PNC), a community-based study of brain development that includes rich neuroimaging and cognitive data (Satterthwaite et al., 2014a; Calkins et al., 2015; Satterthwaite et al., 2015). As described below, results provide novel evidence that structural brain networks undergo a process of modular segregation analogous to prior accounts of functional network development. Critically, these data reveal that the refinement of structural network modules mediate the development of executive function. RESULTS We investigated the evolution of structural brain networks in a sample of 882 youth aged 8-22 who completed neuroimaging as part of the PNC (Figure 1A). As expected, executive function improved markedly with age (Figure 1B). Structural brain networks were constructed using nodes defined based on a parcellation of each subject's structural image into 234 anatomically defined regions; structural connectivity between these nodes was estimated using deterministic tractography (Figure 2). Each network node was assigned a priori to one of the functional network modules defined by Yeo et al. (2011). Although these module partitions were defined in an independent dataset, using a different imaging modality, the modularity quality of the functional partition imposed on subject-level structural connectivity matrices (QYeo) was highly significant (p<1×10-10). Furthermore, data-driven analysis of structural networks using community detection procedures produced network modules that showed significant similarity to the a priori functional modules (p<1×10-10; Figure S1). Segregation of structural network modules increases with age We first sought to understand whether structural network modules became more segregated with age. To do this, we calculated the average participation coefficient for each subject's network. The participation coefficient quantifies the relative balance of a brain region's ! 3 between-module versus within-module connectivity (Guimerà and Nunes Amaral, 2005). Greater modular segregation is therefore indicated by lower participation coefficient values, with reduced between-module and elevated within-module connectivity. We examined the development of modular segregation using a generalized additive model with penalized splines, which allows for statistically rigorous modeling of both linear and non-linear effects while minimizing over-fitting (Wood, 2004; Wood, 2011). Based on emerging evidence that diffusion-weighted imaging measures are systematically biased by motion artifact (Yendiki et al., 2013), we included in-scanner motion as a covariate in all models in addition to participant sex. To ensure that results reflected changes in network topology, rather than global differences in network connectivity, total network strength was also included as a covariate in all analyses (Li et al., 2012). The participation coefficient declined significantly with age (Figure 3A; p<1×10-10), indicating enhanced modular segregation. Developmental increases in modular segregation were differentially distributed across modules (Figure 3B), with the most robust declines observed in the somatomotor and default mode modules. To further understand which regions were driving these effects, we examined the participation coefficient of individual network nodes. As expected, the nodal participation coefficient declined in many regions (Figure 3C), with many of the most significant reductions occurring in regions within the default mode system. Two exceptions to this overall trend were observed, with increasing participation coefficient in the right rostral frontal gyrus and frontal operculum. Next, we investigated the degree to which developmental effects on modular segregation were driven by changes in within-module connectivity, between-module connectivity, or both. We found that both effects were significant: within-module connectivity increased with age (Figure 4A; p<1×10-10), whereas between-module connectivity declined (Figure 4B; p<1×10-10). Moreover, modular segregation was reflected in individual network edges (Figure 4C), with permutation-based analysis revealing that a higher proportion of connections that strengthened with age were located within a module (Figure 4D; p<0.001). Results are robust to methodological approach Given this strong evidence for developmental modular segregation, we next pursued a set of analyses to determine if our results were dependent upon specific methodological choices. We evaluated multiple parameters, including alternative measures of network segregation, higher-resolution node systems, and different edge measures. First, we examined a complementary measure of modular segregation: the modularity quality index of the Yeo et al. partition (QYeo) applied to individual subject-level structural connectivity matrices. For this metric, greater modular segregation is denoted by a greater QYeo value. As expected, this analysis revealed that modular segregation increased with age (Figure 5A; p=1.06×10-9). Second, while our primary results used the functional partition defined by Yeo et al. (2011), we additionally evaluated modular segregation using a data-driven partition of the structural connectome (see Figures S1 and S2). As seen in Figure 5B, age-related decline in the mean participation coefficient remained evident (p<1×10-10). Third, we directly calculated the modularity quality index of a data-driven partition of each subject's structural connectivity matrix (Qsubj). We observed that this statistic – which provides an individualized measure of modularity that is not dependent on a group-level partition – also increased significantly with age (Figure 5C; p=0.0007), indicating greater modular segregation. Fourth, we investigated the impact of using a more fine-grained ! 4 network parcellation (n=463 nodes). This did not impact the observed results, with age-related declines in the participation coefficient remaining highly significant (Figure 5D; p<1×10-10). Fifth, we evaluated developmental effects when using different measures of structural connectivity instead of mean fractional anisotropy. Results using raw streamline count (Figure 5E; p=6.52×10-7) or volume-normalized streamline density (Figure 5F; p=4.56×10-8) remained highly similar. Lastly, in order to rule out the possibility that observed results were driven by potentially confounding variables, we included additional model covariates such as total brain volume, handedness, race, and maternal education; results were unchanged (p<1×10-10). Conversely, results were consistent when covariates such as total network strength, sex, and motion were removed from the model (p<1×10-10). Modular segregation contributes to global network efficiency Having established that network modules become more segregated with age, and that this finding was not dependent on specific analytic choices, we evaluated the impact of evolving network modularity on measures of network integration. Global network efficiency (Eglob) provides a measure of network integration by quantifying information flow across a network as the shortest path between pairs of nodes (Bassett et al., 2009). In many networks, modularity and global efficiency are inversely related, as network segregation by module partitions extends the path length. However, in some cases it is possible for networks to become both more modular and more efficient; this unusual situation occurs when connectivity within modules is efficiently organized and hub edges form strong links between otherwise segregated modules (Sporns and Betzel, 2016). To determine which scenario characterized human neurodevelopment, we first examined the relationship between global efficiency and age while controlling for the covariates described previously. Replicating previous reports (Chen et al., 2013; Hagmann et al., 2010b), we found that global efficiency increases with age (Figure 6A; p<1×10-10). Next, we calculated the correlation between modular segregation (mean participation coefficient) and global efficiency, while co-varying for age to control for shared developmental trends. Mean participation coefficient was negatively associated with global efficiency (Figure 6B; p<1×10-10), suggesting that the development of network modules does not result in fragmentation, but rather is associated with global network integration. Age effects are concentrated in hub edges that promote network integration To better understand this highly specialized association between network modularity and efficiency, we evaluated the edge betweenness centrality for each network connection. Edge betweenness identifies hub connections by providing a measure of how much a given network edge lies upon the shortest path of communication through a network, and thus contributes to global efficiency. Here we defined hub edges as those connections within the top quartile of edge betweenness across all network edges. Critically, edges that strengthened with age were enriched for hub edges (p<0.001; see Figure 6C). Both within- (p<0.001) and between-module (p<0.001) edges that strengthened with age had higher betweenness than expected by chance (Figure 6D; see Supplemental Information). Furthermore, the average strength of all within-module (Figure 6E; p<1×10-10) and between-module (Figure 6F; p<1×10-10) edges that strengthen with age was associated with global efficiency, suggesting that developmental effects are concentrated within connections that facilitate network integration. The striking combination of increasing modular segregation and enhanced global efficiency demonstrates that structural brain networks ! 5 become more modular and more integrated in development. These dual processes are driven by selective strengthening of network hub edges, which are present within network modules and also provide critical links between increasingly segregated modules. Modular segregation mediates development of executive function in youth Next, we evaluated the cognitive implications of modular segregation by examining associations with individual differences in executive function. Mean whole-brain participation coefficient was associated with improved executive performance (p=0.018). At the level of individual modules, we found that segregation of the frontoparietal control system was uniquely associated with executive ability (Figure 7A; p=0.005), suggesting a network-specific substrate for executive function. As a final step, we examined whether age-related changes in executive function and modularity were related. Mediation analyses revealed that this was indeed the case (Figure 7B; p=0.006), suggesting that the development of segregated structural brain modules mediates the age-related improvement in executive function. These mediating effects were specifically driven by the frontoparietal module (p=0.012). Similarly, global efficiency was associated with executive functioning (p=0.037), and also mediated executive development (p=0.002). To evaluate the specificity of these results, we examined associations with other domains of cognition, such as social cognition and memory performance. While no association with memory was found, modular segregation was also significantly associated with social cognition (p=0.022), which was driven by segregation of the default mode module (p=0.012). This effect mediated improvements in social cognition with age (p=0.008). Together, these results demonstrate that developmental segregation of specific structural network modules may support the development of disparate cognitive domains. DISCUSSION Capitalizing on a large sample of youth imaged as part of the PNC, we demonstrated that modules within human structural brain networks become increasingly segregated with age. This result was robust to specific methodological choices, and driven by a combination of enhanced within-module connectivity and declining between-module connectivity. Age related changes were concentrated within specific hub edges, allowing for networks to simultaneously become more modular and more integrated with age. Critically, segregation of network modules mediated the development of executive function during adolescence. Segregation of structural network modules parallels development of functional networks The delineation of robust, reproducible large-scale functional networks has had a tremendous impact on human neuroscience research (Power et al., 2011; Yeo et al., 2011). As a result, functional network modules have evolved to become the dominant framework by which human imaging data is interpreted. The conceptualization of the brain as a modular entity has had a particularly pronounced effect on theories of development, where convergent results have shown that functional network modules are present early in life (Thomason et al., 2013; van den ! 6 Heuvel et al., 2015), and continue to develop during youth (Power et al., 2010; Fair et al., 2007; Fair et al., 2008; Fair et al., 2009; Dosenbach et al., 2010; Gu et al., 2015; Satterthwaite et al., 2013b; Supekar et al., 2009; Anderson et al., 2011). In contrast, smaller studies of structural brain networks have produced heterogeneous results regarding the development of structural network modules that have not aligned well with functional imaging data (Chen et al., 2013; Chen and Deem, 2015; Hagmann et al., 2010b; Huang et al., 2015; Lim et al., 2015). When considered in light of prior studies that have reported substantial correspondence between brain structure and function (Goñi et al., 2014; Honey et al., 2009; Mišić et al., 2016), the disparity between developmental accounts of structural and functional network modules has been difficult to reconcile. Leveraging a large sample imaged as part of the PNC, we were able to resolve this discrepancy by demonstrating that structural network modules develop in a similar manner as functional brain networks, and become increasingly segregated with age. Modular segregation was present at every scale evaluated, including the whole network, individual network modules, and specific network nodes. Importantly, results were robust to a variety of analytic choices regarding network nodes, edges, and modules; such methodological replication is critical as parameter choices may sometimes impact inference (Hagmann et al., 2010a). For example, while we employed a commonly used set of functional network modules which were defined a priori, analysis of data-driven structural modules provided highly convergent results. Follow-up analyses revealed further parallels with functional imaging studies, and demonstrated that the process of modular segregation is driven by a combination of enhanced connectivity within a module as well as diminished connectivity between modules. Notably, the network modules that demonstrated the greatest developmental segregation were the somatomotor and default mode modules. Prior work has shown that both the default mode and somatomotor systems are highly segregated systems, with a low participation coefficient (Power et al., 2011). The relative isolation of these specific systems in the network may reflect a high degree of processing specialization (Power et al., 2013). The present results thus suggest that the differential evolution of structural network modules is similarly driven by each module's network role. Modular networks become more integrated through strengthening of hub edges In many networks, modular segregation is associated with reduced network integration, as measured by global efficiency. We found that this was not the case in development, and that increasing modularity was in fact associated with enhanced network integration. This robust association was the result of targeted strengthening of specific hub edges. These hub edges were present within but also between modules, allowing for integration across increasingly segregated partitions. These results are congruent with prior studies that have demonstrated that connections between network hubs strengthen preferentially with age (Baker et al., 2015), and that network efficiency increases during development (Chen et al., 2013; Hagmann et al., 2010b). The present data emphasize that increasing modular segregation does not result in isolation of functional sub-systems, but is associated with global network integration through strengthening of hub edges that facilitate both intra- and inter-module connectivity. ! 7 Structural network maturation supports the development of executive function in youth Having defined a normative process of modular segregation, we evaluated the cognitive impact of this developmental effect. While controlling for age, we found that greater modular segregation of structural networks was associated with better executive performance. Critically, modular segregation mediated the observed improvement of executive performance with age, and was driven by segregation of the frontoparietal module. Associations between module segregation and cognition were domain-specific: segregation of the default mode mediated age-related improvements in social cognition, which is reliant on regions within that network (Buckner and Carroll, 2007; Schacter et al., 2007). The process of structural network segregation may allow for functional specialization, and reduce competitive interference between brain systems (Fornito et al., 2012). Such a process is suggested by convergent data from task-based fMRI, which has shown that individual differences in performance are related to selective recruitment of executive regions and suppression of activity elsewhere (Anticevic et al., 2010; Barber et al., 2013; Persson et al., 2007; Satterthwaite et al., 2013a). Furthermore, building on prior work that reported an association between intelligence and the global efficiency of structural (Li et al., 2009) and functional networks (van den Heuvel et al., 2009) in relatively small adult samples, we found that global efficiency also mediated developmental improvements in executive function. Taken together, the current data suggest that structural brain networks re-configure with age, becoming both more modular and more integrated. This specific topology may allow for both functional specialization within modules as well as coordination across modules, which is necessary for effective implementation of dynamic executive processes (Hutchison and Morton, 2015; Braun et al., 2015). Limitations Notwithstanding the strengths of this study, several limitations should be noted. First and foremost, this is a cross-sectional dataset, which has inherent limitations for studies of development (Kraemer et al., 2000). The mediating role that network maturation plays in the development of executive function could be further interrogated using longitudinal data. These limitations offer clear directions for additional research. Ongoing follow-up of this cohort will yield informative data, as will other large-scale studies of brain development, including the IMAGEN consortium (Schumann et al., 2010), the NKI-Rockland sample (Nooner et al., 2012), and the forthcoming Adolescent Brain and Cognitive Development Study. Finally, it should be noted that diffusion-based tractography remains limited in its ability to fully resolve the complex architecture of the structural connectome (Jbabdi et al., 2015). Conclusions In this report, we demonstrated that structural brain modules become more segregated with age. Strengthening of specific within- and between-module hub edges allowed for a simultaneous process of network integration that evolves in concert with modular segregation. Finally, both modular segregation and global network integration mediated the development of executive function in youth. These data resolve an ongoing debate in the field regarding the normative development of structural brain networks, and delineate an important new mechanism for the development of executive functioning in youth. These findings may be relevant for understanding how individual differences in brain development associate with risk-taking behaviors, which are linked to failures of executive function, and are a major source of morbidity ! 8 and mortality in adolescence (Shamosh et al., 2008; Casey et al., 2008; Casey et al., 2011). Furthermore, as both abnormalities within developing networks and executive system dysfunction (Shanmugan et al., 2016) are a common feature of diverse types of psychopathology (Alexander-Bloch et al., 2012; Di Martino et al., 2014; Castellanos and Proal, 2012; Voineskos et al., 2010), structural network development may evolve to become an important imaging biomarker of risk and resilience during the critical period of adolescence. ! 9 EXPERIMENTAL PROCEDURES Participants This study included 882 subjects between 8 and 22 years of age (mean age=15.06, SD=3.15; 389 males, 493 females) who were imaged as part of the PNC (Satterthwaite et al., 2014a; Satterthwaite et al., 2015). Participants were excluded from analyses due to gross structural brain abnormalities (Gur et al., 2013), a history of inpatient psychiatric treatment, current use of psychotropic medications, and medical disorders that could impact brain function. Participants were only included if both structural (Vandekar et al., 2015) and diffusion images (Roalf et al., 2016) passed rigorous quality assurance procedures. All participants completed the Penn computerized neurocognitive battery, which included 14 tests (Gur et al., 2002; Gur et al., 2012). Cognitive performance was summarized by a recent factor analysis (Moore et al., 2014) of both speed and accuracy data, which delineated three factors corresponding to the efficiency of executive function, episodic memory, and social cognition. For further details see Supplemental Information. Image acquisition & processing All high-resolution structural and 64-direction diffusion images were collected on the same 3T Siemens TIM Trio scanner using the same sequences (Satterthwaite et al., 2014a). The T1 image was processed using FreeSurfer version 5.3 (Fischl, 2012), and parcellated into 234 cortical and subcortical regions (Cammoun et al., 2012). The diffusion images were distortion corrected, skull stripped, and motion and eddy current corrected with FSL's `eddy` tool (Jenkinson et al., 2012). Native-space T1 parcels were dilated by 4mm to extend regions into white matter, and co-registered to the first non-weighted (b=0) volume using a boundary-based rigid body transformation (Greve and Fischl, 2009). The diffusion tensor was estimated in DSI Studio and whole-brain deterministic fiber tracking was implemented with 1,000,000 streamlines per subject after removing all streamlines with length less than 10mm (Yeh et al., 2013). Edge weights in the adjacency matrix were defined by mean fractional anisotropy along streamlines connecting each pair of nodes (Mišić et al., 2016; Bohlken et al., 2016; Baker et al., 2015). See Figure 2 and Supplemental Information for more detail. Measurement of modular segregation Nodes were assigned to modules according to their overlap with seven functional networks defined a priori (Yeo et al., 2011); subcortical nodes were assigned to their own additional module. In addition to such a priori functional assignment, network modules were defined directly from the structural connectivity data using a generalized version of the Louvain community detection algorithm (Bassett et al., 2013; Blondel et al., 2008; Mucha et al., 2010). The similarity of data-driven partitions of the structural data were compared to the functional partition using the z-score of the Rand coefficient (Traud et al., 2011). Modular segregation was quantified using the participation coefficient, which measures the balance of between- versus within-module connectivity for each brain region (Guimerà and Nunes Amaral, 2005; Rubinov and Sporns, 2010). A higher participation coefficient at a given node indicates more between-module and less within-module connectivity. Each node's participation coefficient was averaged over modules and the whole brain in order to evaluate modular segregation at multiple scales. To ! 10 understand whether changes in within- or between-module connectivity were driving observed effects, we also calculated these measures separately for the whole brain and each network module (Gu et al., 2015). Also see Supplemental Information. Statistical analyses of age-related changes in modular segregation Linear and nonlinear effects of age were flexibly modeled with penalized thin-plate splines using generalized additive models (GAMs; Wood, 2004; Wood, 2011), which provide statistically rigorous analysis of non-linear effects while minimizing over-fitting. As in previous work (Satterthwaite et al., 2014b; Vandekar et al., 2015), nonlinearity was penalized within the GAMs using restricted maximum likelihood. Developmental models of each network measure were assessed by modeling age with a spline term, while including participant sex and in-scanner motion as covariates. To ensure that changes in global connectivity strength did not drive analyses of network topology, total network strength was also included as a covariate in all analyses (Li et al., 2012). Throughout, multiple comparisons were controlled using the False Discovery Rate (q<0.05; Genovese et al., 2002). Permutation testing was used to assess whether the a priori functional network partition fit subject-level structural connectivity matrices and whether within-module connections were enriched for age effects (see Supplemental Information). Methodological replications To verify that observed age-related increases in modular segregation were not simply due to specific processing choices, we repeated analyses of global network segregation using a variety of other parameters. Analyses were repeated using a subject-specific measure of modularity quality (Q) for both a priori functional networks and a data-driven structural partition. Using the functional network partition, we also examined age-related changes in modular segregation in a higher resolution node system (n=463 instead of n=234), and using two alternative edge definitions (streamline count and normalized streamline density). Finally, we evaluated the effect of additional model covariates (race, maternal education, handedness, and total brain volume). For additional details, see Supplemental Information. Relationship between modular segregation and global network efficiency Global efficiency was calculated for each participant's structural network (Latora and Marchiori, 2001; Rubinov and Sporns, 2010) and age-related effects were examined as above. The relationship between global efficiency and modular segregation was examined while controlling for age (and other covariates as above). To further describe how age-related changes in within- and between-module connectivity might drive measures of global efficiency, we calculated the edge betweenness centrality (Brandes, 2001; Rubinov and Sporns, 2010). Edge betweenness quantifies the degree to which each edge participates in the shortest paths within a network, and thus contributes to global efficiency. Normalized edge betweenness was calculated for each edge, split by type (within- versus between-module) and age effect (strengthens with age versus no change). Hub edges were defined as connections within the top quartile of edge betweenness across all network edges. Permutation tests were used to evaluate whether connections that strengthened with age were enriched for hub edges, and had higher edge betweenness than expected by chance (see Supplemental Information). ! 11 Associations with executive function To determine if network segregation was related to executive performance, a GAM was used which incorporated the factor score for executive efficiency as well as model covariates (spline age, sex, in-scanner motion, total network strength). Specificity analyses additionally evaluated relationships with other cognitive domains including episodic memory and social cognition. Global efficiency was also assessed for relationships with cognition. Linear mediation analyses investigated whether age-related improvements in cognition were mediated by modular segregation or integration; significance was assessed using bootstrapping procedures (Preacher and Hayes, 2008). SUPPLEMENTAL INFORMATION Supplemental Information includes Figure S1, Figure S2, and Supplemental Experimental Procedures. ACKNOWLDGEMENTS We thank the acquisition and recruitment team, including Karthik Prabhakaran and Jeff Valdez. Thanks to Chad Jackson for data management and systems support and Monica Calkins for phenotyping expertise. Supported by grants from the National Institute of Mental Health: R01MH107703 (TDS), R21MH106799 (DSB & TDS). DSB acknowledges support from the John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Army Research Laboratory and the Army Research Office through contract numbers W911NF-10-2-0022 and W911NF-14-1-0679, the National Institute of Mental Health (R01-DC-009209-11), the National Institute of Child Health and Human Development (R01HD086888-01), the Office of Naval Research, and the National Science Foundation (BCS-1441502, BCS-1430087, and PHY-1554488). The PNC was funded through NIMH RC2 grants MH089983 and MH089924 (REG). Additional support was provided by R01MH107235 (RCG), K23MH098130 (TDS), K01MH102609 (DRR), P50MH096891 (REG), R01NS085211 (RTS), and the Dowshen Program for Neuroscience. Support for developing statistical analyses (RTS & TDS) was provided by a seed grant by the Center for Biomedical Computing and Image Analysis (CBICA) at Penn. Data deposition: The data reported in this paper have been deposited in the database of Genotypes and Phenotypes (dbGaP), www.ncbi.nlm.nih.gov/gap (accession no. phs000607.v1.p1). ! 12 REFERENCES Alexander-Bloch, A., Lambiotte, R., Roberts, B., Giedd, J., Gogtay, N., and Bullmore, E. (2012). The discovery of population differences in network community structure: new methods and applications to brain functional networks in schizophrenia. Neuroimage 59, 3889-3900. Anderson, J.S., Ferguson, M.A., Lopez-Larson, M., and Yurgelun-Todd, D. (2011). Connectivity gradients between the default mode and attention control networks. Brain Connect 1, 147-157. Anticevic, A., Van Snellenberg, J.X., Cohen, R.E., Repovs, G., Dowd, E.C., and Barch, D.M. (2010). Amygdala Recruitment in Schizophrenia in Response to Aversive Emotional Material: A Meta-analysis of Neuroimaging Studies. Schizophr Bull Baker, S.T.E., Lubman, D.I., Yücel, M., Allen, N.B., Whittle, S., Fulcher, B.D., Zalesky, A., and Fornito, A. (2015). Developmental Changes in Brain Network Hub Connectivity in Late Adolescence. J Neurosci 35, 9078-9087. Barber, A.D., Caffo, B.S., Pekar, J.J., and Mostofsky, S.H. (2013). Developmental changes in within- and between-network connectivity between late childhood and adulthood. Neuropsychologia 51, 156-167. Bassett, D.S., Brown, J.A., Deshpande, V., Carlson, J.M., and Grafton, S.T. (2011). Conserved and variable architecture of human white matter connectivity. Neuroimage 54, 1262-1279. Bassett, D.S., Bullmore, E.T., Meyer-Lindenberg, A., Apud, J.A., Weinberger, D.R., and Coppola, R. (2009). Cognitive fitness of cost-efficient brain functional networks. Proc Natl Acad Sci U S A 106, 11747-11752. Bassett, D.S., Greenfield, D.L., Meyer-Lindenberg, A., Weinberger, D.R., Moore, S.W., and Bullmore, E.T. (2010). Efficient physical embedding of topologically complex information processing networks in brains and computer circuits. PLoS Comput Biol 6, e1000748. Bassett, D.S., Porter, M.A., Wymbs, N.F., Grafton, S.T., Carlson, J.M., and Mucha, P.J. (2013). Robust detection of dynamic community structure in networks. Chaos 23, 013142. Betzel, R.F., Byrge, L., He, Y., Goñi, J., Zuo, X.-N., and Sporns, O. (2014). Changes in structural and functional connectivity among resting-state networks across the human lifespan. Neuroimage 102 Pt 2, 345-357. Biswal, B., Yetkin, F.Z., Haughton, V.M., and Hyde, J.S. (1995). Functional connectivity in the motor cortex of resting human brain using echo-planar MRI. Magn Reson Med 34, 537-541. Blondel, V.D., Guillaume, J.-L., Lambiotte, R., and Lefebvre, E. (2008). Fast unfolding of communities in large networks. Journal of statistical mechanics: theory and experiment 2008, P10008. Bohlken, M.M., Brouwer, R.M., Mandl, R.C.W., Van den Heuvel, M.P., Hedman, A.M., De Hert, M., Cahn, W., Kahn, R.S., and Hulshoff Pol, H.E. (2016). Structural Brain Connectivity as a Genetic Marker for Schizophrenia. JAMA Psychiatry 73, 11-19. Braun, U., Schäfer, A., Walter, H., Erk, S., Romanczuk-Seiferth, N., Haddad, L., Schweiger, J.I., Grimm, O., Heinz, A., and Tost, H. (2015). Dynamic reconfiguration of frontal brain ! 13 networks during executive cognition in humans. Proceedings of the National Academy of Sciences 112, 11678-11683. Buckner, R.L., and Carroll, D.C. (2007). Self-projection and the brain. Trends Cogn Sci 11, 49-57. Buckner, R.L., Andrews-Hanna, J.R., and Schacter, D.L. (2008). The brain's default network: anatomy, function, and relevance to disease. Ann N Y Acad Sci 1124, 1-38. Calkins, M.E., Merikangas, K.R., Moore, T.M., Burstein, M., Behr, M.A., Satterthwaite, T.D., Ruparel, K., Wolf, D.H., Roalf, D.R., Mentch, F.D., Qiu, H., Chiavacci, R., Connolly, J.J., Sleiman, P.M.A., Gur, R.C., Hakonarson, H., and Gur, R.E. (2015). The Philadelphia Neurodevelopmental Cohort: constructing a deep phenotyping collaborative. J Child Psychol Psychiatry Cammoun, L., Gigandet, X., Meskaldji, D., Thiran, J.P., Sporns, O., Do, K.Q., Maeder, P., Meuli, R., and Hagmann, P. (2012). Mapping the human connectome at multiple scales with diffusion spectrum MRI. J Neurosci Methods 203, 386-397. Casey, B.J., Hare, T.A., and Galván, A. (2011). Risky and impulsive components of adolescent decision making. Decision Making, Affect, and Learning: Attention and Performance XXIII 23, 425. Casey, B.J., Jones, R.M., and Hare, T.A. (2008). The Adolescent Brain. Ann N Y Acad Sci 1124, 111-126. Casey, B.J., Oliveri, M.E., and Insel, T. (2014). A Neurodevelopmental Perspective on the Research Domain Criteria (RDoC) Framework. Biol Psychiatry 76, 350-353. Castellanos, F.X., and Proal, E. (2012). Large-scale brain systems in ADHD: beyond the prefrontal–striatal model. Trends Cogn Sci 16, 17-26. Chen, M., and Deem, M.W. (2015). Development of modularity in the neural activity of children's brains. Phys Biol 12, 016009. Chen, Z., Liu, M., Gross, D.W., and Beaulieu, C. (2013). Graph theoretical analysis of developmental patterns of the white matter network. Front Hum Neurosci 7, 716. Corbetta, M., Akbudak, E., Conturo, T.E., Snyder, A.Z., Ollinger, J.M., Drury, H.A., Linenweber, M.R., Petersen, S.E., Raichle, M.E., and Van Essen, D.C. (1998). A common network of functional areas for attention and eye movements. Neuron 21, 761-773. Damoiseaux, J.S., Rombouts, S.A.R.B., Barkhof, F., Scheltens, P., Stam, C.J., Smith, S.M., and Beckmann, C.F. (2006). Consistent resting-state networks across healthy subjects. Proc Natl Acad Sci U S A 103, 13848-13853. Di Martino, A., Fair, D.A., Kelly, C., Satterthwaite, T.D., Castellanos, F.X., Thomason, M.E., Craddock, R.C., Luna, B., Leventhal, B.L., Zuo, X.N., and Milham, M.P. (2014). Unraveling the Miswired Connectome: A Developmental Perspective. Neuron 83, 1335-1353. Dosenbach, N.U., Fair, D.A., Miezin, F.M., Cohen, A.L., Wenger, K.K., Dosenbach, R.A., Fox, M.D., Snyder, A.Z., Vincent, J.L., Raichle, M.E., Schlaggar, B.L., and Petersen, S.E. (2007). Distinct brain networks for adaptive and stable task control in humans. Proc Natl Acad Sci U S A 104, 11073-11078. Dosenbach, N.U.F., Nardos, B., Cohen, A.L., Fair, D.A., Power, J.D., Church, J.A., Nelson, S.M., Wig, G.S., Vogel, A.C., Lessov-Schlaggar, C.N., Barnes, K.A., Dubis, J.W., Feczko, E., Coalson, R.S., Pruett, J.R., Barch, D.M., Petersen, S.E., and Schlaggar, B.L. (2010). Prediction of individual brain maturity using fMRI. Science 329, 1358-1361. ! 14 Fair, D.A., Cohen, A.L., Dosenbach, N.U.F., Church, J.A., Miezin, F.M., Barch, D.M., Raichle, M.E., Petersen, S.E., and Schlaggar, B.L. (2008). The maturing architecture of the brain's default network. Proceedings of the National Academy of Sciences 105, 4028. Fair, D.A., Cohen, A.L., Power, J.D., Dosenbach, N.U.F., Church, J.A., Miezin, F.M., Schlaggar, B.L., and Petersen, S.E. (2009). Functional brain networks develop from a "local to distributed" organization. PLoS Comput Biol 5, e1000381. Fair, D.A., Dosenbach, N.U.F., Church, J.A., Cohen, A.L., Brahmbhatt, S., Miezin, F.M., Barch, D.M., Raichle, M.E., Petersen, S.E., and Schlaggar, B.L. (2007). Development of distinct control networks through segregation and integration. Proc Natl Acad Sci U S A 104, 13507-13512. Fischl, B. (2012). FreeSurfer. Neuroimage 62, 774-781. Fornito, A., Harrison, B.J., Zalesky, A., and Simons, J.S. (2012). Competitive and cooperative dynamics of large-scale brain functional networks supporting recollection. Proc Natl Acad Sci U S A 109, 12788-12793. Genovese, C.R., Lazar, N.A., and Nichols, T. (2002). Thresholding of statistical maps in functional neuroimaging using the false discovery rate. Neuroimage 15, 870-878. Goñi, J., van den Heuvel, M.P., Avena-Koenigsberger, A., de Mendizabal, N.V., Betzel, R.F., Griffa, A., Hagmann, P., Corominas-Murtra, B., Thiran, J.-P., and Sporns, O. (2014). Resting-brain functional connectivity predicted by analytic measures of network communication. Proceedings of the National Academy of Sciences 111, 833-838. Greve, D.N., and Fischl, B. (2009). Accurate and robust brain image alignment using boundary-based registration. Neuroimage 48, 63-72. Gu, S., Satterthwaite, T.D., Medaglia, J.D., Yang, M., Gur, R.E., Gur, R.C., and Bassett, D.S. (2015). Emergence of system roles in normative neurodevelopment. Proc Natl Acad Sci U S A 112, 13681-13686. Guimerà, R., and Nunes Amaral, L.A. (2005). Functional cartography of complex metabolic networks. Nature 433, 895-900. Gur, R.C., Richard, J., Calkins, M.E., Chiavacci, R., Hansen, J.A., Bilker, W.B., Loughead, J., Connolly, J.J., Qiu, H., Mentch, F.D., Abou-Sleiman, P.M., Hakonarson, H., and Gur, R.E. (2012). Age group and sex differences in performance on a computerized neurocognitive battery in children age 8-21. Neuropsychology 26, 251-265. Gur, R.C., Sara, R., Hagendoorn, M., Marom, O., Hughett, P., Macy, L., Turner, T., Bajcsy, R., Posner, A., and Gur, R.E. (2002). A method for obtaining 3-dimensional facial expressions and its standardization for use in neurocognitive studies. J Neurosci Methods 115, 137-143. Gur, R.E., Kaltman, D., Melhem, E.R., Ruparel, K., Prabhakaran, K., Riley, M., Yodh, E., Hakonarson, H., Satterthwaite, T., and Gur, R.C. (2013). Incidental findings in youths volunteering for brain MRI research. AJNR Am J Neuroradiol 34, 2021-2025. Hagmann, P., Cammoun, L., Gigandet, X., Gerhard, S., Ellen Grant, P., Wedeen, V., Meuli, R., Thiran, J.-P., Honey, C.J., and Sporns, O. (2010a). MR connectomics: Principles and challenges. J Neurosci Methods Hagmann, P., Sporns, O., Madan, N., Cammoun, L., Pienaar, R., Wedeen, V.J., Meuli, R., Thiran, J.-P., and Grant, P.E. (2010b). White matter maturation reshapes structural connectivity in the late developing human brain. Proc Natl Acad Sci U S A 107, 19067-19072. ! 15 Hampson, M., Driesen, N., Roth, J.K., Gore, J.C., and Constable, R.T. (2010). Functional connectivity between task-positive and task-negative brain areas and its relation to working memory performance. Magn Reson Imaging Hermundstad, A.M., Bassett, D.S., Brown, K.S., Aminoff, E.M., Clewett, D., Freeman, S., Frithsen, A., Johnson, A., Tipper, C.M., and Miller, M.B. (2013). Structural foundations of resting-state and task-based functional connectivity in the human brain. Proceedings of the National Academy of Sciences 110, 6169-6174. Hermundstad, A.M., Brown, K.S., Bassett, D.S., Aminoff, E.M., Frithsen, A., Johnson, A., Tipper, C.M., Miller, M.B., Grafton, S.T., and Carlson, J.M. (2014). Structurally-constrained relationships between cognitive states in the human brain. PLoS Comput Biol 10, e1003591. Honey, C.J., Sporns, O., Cammoun, L., Gigandet, X., Thiran, J.P., Meuli, R., and Hagmann, P. (2009). Predicting human resting-state functional connectivity from structural connectivity. Proc Natl Acad Sci U S A 106, 2035-2040. Huang, H., Shu, N., Mishra, V., Jeon, T., Chalak, L., Wang, Z.J., Rollins, N., Gong, G., Cheng, H., Peng, Y., Dong, Q., and He, Y. (2015). Development of human brain structural networks through infancy and childhood. Cereb Cortex 25, 1389-1404. Hutchison, R.M., and Morton, J.B. (2015). It's a matter of time: Reframing the development of cognitive control as a modification of the brain's temporal dynamics. Dev Cogn Neurosci Insel, T.R. (2010). Rethinking schizophrenia. Nature 468, 187-193. Jenkinson, M., Beckmann, C.F., Behrens, T.E.J., Woolrich, M.W., and Smith, S.M. (2012). FSL. Neuroimage 62, 782-790. Jbabdi, S., Sotiropoulos, S.N., Haber, S.N., Van Essen, D.C., and Behrens, T.E. (2015). Measuring macroscopic brain connections in vivo. Nat Neurosci 18, 1546-1555. Kraemer, H.C., Yesavage, J.A., Taylor, J.L., and Kupfer, D. (2000). How can we learn about developmental processes from cross-sectional studies, or can we? Am J Psychiatry 157, 163-171. Latora, V., and Marchiori, M. (2001). Efficient behavior of small-world networks. Phys Rev Lett 87, 198701. Li, L., Rilling, J.K., Preuss, T.M., Glasser, M.F., and Hu, X. (2012). The effects of connection reconstruction method on the interregional connectivity of brain networks via diffusion tractography. Hum Brain Mapp 33, 1894-1913. Li, Y., Liu, Y., Li, J., Qin, W., Li, K., Yu, C., and Jiang, T. (2009). Brain anatomical network and intelligence. PLoS Comput Biol 5, e1000395. Lim, S., Han, C.E., Uhlhaas, P.J., and Kaiser, M. (2015). Preferential detachment during human brain development: age- and sex-specific structural connectivity in diffusion tensor imaging (DTI) data. Cereb Cortex 25, 1477-1489. Luna, B., Garver, K.E., Urban, T.A., Lazar, N.A., and Sweeney, J.A. (2004). Maturation of cognitive processes from late childhood to adulthood. Child Dev 75, 1357-1372. Mišić, B., Betzel, R.F., de Reus, M.A., van den Heuvel, M.P., Berman, M.G., McIntosh, A.R., and Sporns, O. (2016). Network-Level Structure-Function Relationships in Human Neocortex. Cereb Cortex 26, 3285-3296. Moore, T.M., Reise, S.P., Gur, R.E., Hakonarson, H., and Gur, R.C. (2014). Psychometric Properties of the Penn Computerized Neurocognitive Battery. Neuropsychology ! 16 Mucha, P.J., Richardson, T., Macon, K., Porter, M.A., and Onnela, J.-P. (2010). Community structure in time-dependent, multiscale, and multiplex networks. Science 328, 876-878. Newman, M.E.J. (2006). Modularity and community structure in networks. Proc Natl Acad Sci U S A 103, 8577-8582. Nooner, K.B., Colcombe, S.J., Tobe, R.H., Mennes, M., Benedict, M.M., Moreno, A.L., Panek, L.J., Brown, S., Zavitz, S.T., Li, Q., Sikka, S., Gutman, D., Bangaru, S., Schlachter, R.T., Kamiel, S.M., Anwar, A.R., Hinz, C.M., Kaplan, M.S., Rachlin, A.B., Adelsberg, S., Cheung, B., Khanuja, R., Yan, C., Craddock, C.C., Calhoun, V., Courtney, W., King, M., Wood, D., Cox, C.L., Kelly, A.M.C., Di Martino, A., Petkova, E., Reiss, P.T., Duan, N., Thomsen, D., Biswal, B., Coffey, B., Hoptman, M.J., Javitt, D.C., Pomara, N., Sidtis, J.J., Koplewicz, H.S., Castellanos, F.X., Leventhal, B.L., and Milham, M.P. (2012). The NKI-Rockland Sample: A Model for Accelerating the Pace of Discovery Science in Psychiatry. Front Neurosci 6, 152. Persson, J., Lustig, C., Nelson, J.K., and Reuter-Lorenz, P.A. (2007). Age differences in deactivation: a link to cognitive control? J Cogn Neurosci 19, 1021-1032. Power, J.D., Cohen, A.L., Nelson, S.M., Wig, G.S., Barnes, K.A., Church, J.A., Vogel, A.C., Laumann, T.O., Miezin, F.M., Schlaggar, B.L., and Petersen, S.E. (2011). Functional network organization of the human brain. Neuron 72, 665-678. Power, J.D., Fair, D.A., Schlaggar, B.L., and Petersen, S.E. (2010). The development of human functional brain networks. Neuron 67, 735-748. Power, J.D., Schlaggar, B.L., Lessov-Schlaggar, C.N., and Petersen, S.E. (2013). Evidence for hubs in human functional brain networks. Neuron 79, 798-813. Preacher, K.J., and Hayes, A.F. (2008). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behav Res Methods 40, 879-891. Raichle, M.E., MacLeod, A.M., Snyder, A.Z., Powers, W.J., Gusnard, D.A., and Shulman, G.L. (2001). A default mode of brain function. Proc Natl Acad Sci U S A 98, 676-682. Rapoport, J.L., Giedd, J.N., and Gogtay, N. (2012). Neurodevelopmental model of schizophrenia: update 2012. Mol Psychiatry Roalf, D.R., Quarmley, M., Elliott, M.A., Satterthwaite, T.D., Vandekar, S.N., Ruparel, K., Gennatas, E.D., Calkins, M.E., Moore, T.M., Hopson, R., Prabhakaran, K., Jackson, C.T., Verma, R., Hakonarson, H., Gur, R.C., and Gur, R.E. (2016). The impact of quality assurance assessment on diffusion tensor imaging outcomes in a large-scale population-based cohort. Neuroimage 125, 903-919. Rubinov, M., and Sporns, O. (2010). Complex network measures of brain connectivity: uses and interpretations. Neuroimage 52, 1059-1069. Satterthwaite, T.D., Connolly, J.J., Ruparel, K., Calkins, M.E., Jackson, C., Elliott, M.A., Roalf, D.R., Ryan Hopsona, K.P., Behr, M., Qiu, H., Mentch, F.D., Chiavacci, R., Sleiman, P.M.A., Gur, R.C., Hakonarson, H., and Gur, R.E. (2015). The Philadelphia Neurodevelopmental Cohort: A publicly available resource for the study of normal and abnormal brain development in youth. Neuroimage Satterthwaite, T.D., Elliott, M.A., Ruparel, K., Loughead, J., Prabhakaran, K., Calkins, M.E., Hopson, R., Jackson, C., Keefe, J., Riley, M., Mentch, F.D., Sleiman, P., Verma, R., Davatzikos, C., Hakonarson, H., Gur, R.C., and Gur, R.E. (2014a). Neuroimaging of the Philadelphia neurodevelopmental cohort. Neuroimage 86, 544-553. ! 17 Satterthwaite, T.D., Shinohara, R.T., Wolf, D.H., Hopson, R.D., Elliott, M.A., Vandekar, S.N., Ruparel, K., Calkins, M.E., Roalf, D.R., Gennatas, E.D., Jackson, C., Erus, G., Prabhakaran, K., Davatzikos, C., Detre, J.A., Hakonarson, H., Gur, R.C., and Gur, R.E. (2014b). Impact of puberty on the evolution of cerebral perfusion during adolescence. Proc Natl Acad Sci U S A Satterthwaite, T.D., Wolf, D.H., Erus, G., Ruparel, K., Elliott, M.A., Gennatas, E.D., Hopson, R., Jackson, C., Prabhakaran, K., Bilker, W.B., Calkins, M.E., Loughead, J., Smith, A., Roalf, D.R., Hakonarson, H., Verma, R., Davatzikos, C., Gur, R.C., and Gur, R.E. (2013a). Functional Maturation of the Executive System during Adolescence. J Neurosci 33, 16249-16261. Satterthwaite, T.D., Wolf, D.H., Ruparel, K., Erus, G., Elliott, M.A., Eickhoff, S.B., Gennatas, E.D., Jackson, C., Prabhakaran, K., Smith, A., Hakonarson, H., Verma, R., Davatzikos, C., Gur, R.E., and Gur, R.C. (2013b). Heterogeneous impact of motion on fundamental patterns of developmental changes in functional connectivity during youth. Neuroimage 83, 45 - 57. Schacter, D.L., Addis, D.R., and Buckner, R.L. (2007). Remembering the past to imagine the future: the prospective brain. Nat Rev Neurosci 8, 657-661. Schumann, G., Loth, E., Banaschewski, T., Barbot, A., Barker, G., Büchel, C., Conrod, P.J., Dalley, J.W., Flor, H., Gallinat, J., Garavan, H., Heinz, A., Itterman, B., Lathrop, M., Mallik, C., Mann, K., Martinot, J.-L., Paus, T., Poline, J.-B., Robbins, T.W., Rietschel, M., Reed, L., Smolka, M., Spanagel, R., Speiser, C., Stephens, D.N., Ströhle, A., Struve, M., and IMAGEN consortium (2010). The IMAGEN study: reinforcement-related behaviour in normal brain function and psychopathology. Mol Psychiatry 15, 1128-1139. Shamosh, N.A., Deyoung, C.G., Green, A.E., Reis, D.L., Johnson, M.R., Conway, A.R.A., Engle, R.W., Braver, T.S., and Gray, J.R. (2008). Individual differences in delay discounting: relation to intelligence, working memory, and anterior prefrontal cortex. Psychol Sci 19, 904-911. Shanmugan, S., Wolf, D.H., Calkins, M.E., Moore, T.M., Ruparel, K., Hopson, R.D., Vandekar, S.N., Roalf, D.R., Elliott, M.A., Jackson, C., Gennatas, E.D., Leibenluft, E., Pine, D.S., Shinohara, R.T., Hakonarson, H., Gur, R.C., Gur, R.E., and Satterthwaite, T.D. (2016). Common and Dissociable Mechanisms of Executive System Dysfunction Across Psychiatric Disorders in Youth. Am J Psychiatry , appiajp201515060725. Sporns, O., and Betzel, R.F. (2016). Modular Brain Networks. Annu Rev Psychol 67, 613-640. Supekar, K., Musen, M., and Menon, V. (2009). Development of large-scale functional brain networks in children. PLoS Biol 7, e1000157. Thomason, M.E., Dassanayake, M.T., Shen, S., Katkuri, Y., Alexis, M., Anderson, A.L., Yeo, L., Mody, S., Hernandez-Andrade, E., Hassan, S.S., Studholme, C., Jeong, J.-W., and Romero, R. (2013). Cross-hemispheric functional connectivity in the human fetal brain. Sci Transl Med 5, 173ra24. Traud, A.L., Kelsic, E.D., Mucha, P.J., and Porter, M.A. (2011). Comparing community structure to characteristics in online collegiate social networks. SIAM review 53, 526-543. Vandekar, S.N., Shinohara, R.T., Raznahan, A., Roalf, D.R., Ross, M., DeLeo, N., Ruparel, K., Verma, R., Wolf, D.H., Gur, R.C., Gur, R.E., and Satterthwaite, T.D. (2015). Topologically dissociable patterns of development of the human cerebral cortex. J Neurosci 35, 599-609. ! 18 van den Heuvel, M.P., Kersbergen, K.J., de Reus, M.A., Keunen, K., Kahn, R.S., Groenendaal, F., de Vries, L.S., and Benders, M.J.N.L. (2015). The Neonatal Connectome During Preterm Brain Development. Cereb Cortex 25, 3000-3013. van den Heuvel, M.P., Stam, C.J., Kahn, R.S., and Hulshoff Pol, H.E. (2009). Efficiency of functional brain networks and intellectual performance. J Neurosci 29, 7619-7624. Vincent, J.L., Kahn, I., Snyder, A.Z., Raichle, M.E., and Buckner, R.L. (2008). Evidence for a frontoparietal control system revealed by intrinsic functional connectivity. J Neurophysiol 100, 3328-3342. Voineskos, A.N., Lobaugh, N.J., Bouix, S., Rajji, T.K., Miranda, D., Kennedy, J.L., Mulsant, B.H., Pollock, B.G., and Shenton, M.E. (2010). Diffusion tensor tractography findings in schizophrenia across the adult lifespan. Brain 133, 1494-1504. Wood, S.N. (2004). Stable and efficient multiple smoothing parameter estimation for generalized additive models. Journal of the American Statistical Association 99 Wood, S.N. (2011). Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 73, 3-36. Yeh, F.-C., Verstynen, T.D., Wang, Y., Fernández-Miranda, J.C., and Tseng, W.-Y.I. (2013). Deterministic diffusion fiber tracking improved by quantitative anisotropy. PLoS One 8, e80713. Yendiki, A., Koldewyn, K., Kakunoori, S., Kanwisher, N., and Fischl, B. (2013). Spurious group differences due to head motion in a diffusion MRI study. Neuroimage 88C, 79-90. Yeo, B.T.T., Krienen, F.M., Sepulcre, J., Sabuncu, M.R., Lashkari, D., Hollinshead, M., Roffman, J.L., Smoller, J.W., Zöllei, L., Polimeni, J.R., Fischl, B., Liu, H., and Buckner, R.L. (2011). The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol 106, 1125-1165. ! 19 Figure 1. Executive functioning improves with age. A. Age distribution of 882 youth completing diffusion imaging as part of the PNC. B. Executive performance on a neurocognitive battery improves with age. Blue line represents the best fit from a general additive model; shaded area indicates 95% confidence interval. Model includes participant sex as a covariate. ! 20 Figure 2. Connectome construction. For each subject, the T1 image was processed using FreeSurfer and parcellated into 234 network nodes on an individualized basis. Deterministic whole-brain fiber tracking was used to create a symmetric adjacency matrix (234×234), where the edge weight was defined as the mean fractional anisotropy (FA) along the connecting streamlines. Network nodes were each assigned to one of the seven large-scale functional modules defined by Yeo et al. (2011); subcortical nodes were assigned to an eighth modules. VIS=visual, SOM=somatomotor, DOR=dorsal attention, VEN=ventral attention, LIM=limbic, FPC=frontoparietal control, DMN= default mode network, SUB=subcortical. ! 21 Figure 3. Structural brain network modules become increasingly segregated with age. A. Modular segregation was quantified as the mean participation coefficient across all network nodes, with lower values indicating more segregation. Participation coefficient values declined significantly with age. B. Modular segregation is differentially distributed across functional systems. Age-related modular segregation is most robust in the somatomotor and default mode systems, but also present in other networks. C. Age-related changes in participation coefficient provide convergent results for individual nodes, and demonstrate widespread declines with age. The strongest age-related reductions of the participation coefficient were seen in default mode regions such as the posterior cingulate. Two exceptions to this overall trend were the right rostral frontal gyrus and frontal operculum, where participation coefficient increased with age. Blue line represents the best fit from a general additive model; shaded area indicates 95% confidence interval. All analyses control for sex, in-scanner motion, and network strength. Color palette represents z-transformed p values from a general additive model. Images are thresholded to control for multiple comparisons using the False Discovery Rate (Q<0.05). *indicates p<0.001. ! 22 Figure 4. Modular segregation is driven by a combination of both enhanced within-module connectivity and reduced between-module connectivity. A. Average strength of within-module connectivity increases with age. B. Between-module connectivity decreases across development. C. Convergent effects are seen at the level of individual graph edges (image thresholded using Bonferroni corrected p<0.05 for clarity). D. A higher percentage of within-module connections (red) strengthen with age than expected by chance. All analyses include sex, in-scanner motion, and network strength as model covariates. * indicates p<0.001. ! 23 Figure 5. Results are robust to methodological choices. Regardless of specific processing decisions, an increase in modular segregation with age was observed. A. Convergent findings result when using an index of the modularity quality for the Yeo partition, where higher Q indicates more segregated modules. B. When using a group-level structural partition, modular segregation (mean participation coefficient) decreases with age. C. Modularity quality of subject-level connectivity matrices also increases with age. D. Results remain unaffected when a higher-dimensional parcellation is used (n=463 nodes), E. when streamline count is used instead of FA as an edge weight, and F. when normalized streamline density is used as the edge weight. Lower participation coefficient indicates more segregated modules. Blue line represents the best fit from a general additive model; shaded area indicates 95% confidence interval. All models include sex, in-scanner motion, and total network strength as covariates. ! 24 Figure 6. Modularity is associated with network integration, and is driven by developmental strengthening of specific hub edges. A. Replicating prior work, global network efficiency increases with age. Model includes sex, in-scanner motion, and total network strength as covariates. B. While controlling for age, lower mean participation coefficient is associated with greater network efficiency, indicating a positive association between modular segregation and network integration. C. Connections that strengthen with age are enriched for hub edges (47%). Hub edges are defined as connections in the top quartile of edge betweenness centrality, which quantifies how often a given edge lies on the shortest path between nodes and thus facilitates global efficiency. Image thresholded using Bonferroni corrected p<0.05 for clarity. D. Both within- and between-module connections that strengthen with age have higher edge betweenness centrality than expected by chance. The average weight of within- (E) and between-module edges (F) that strengthen with age are positively associated with global efficiency. Blue line represents the best fit from a general additive model, shaded area indicates 95% confidence interval; * indicates p<0.001. Error bar represents standard error of the mean. ! 25 Figure 7. Segregation of structural modules supports the development of executive function in youth. A. While controlling for age, greater modular segregation in the frontoparietal control network is associated with better executive performance. B. Segregation of structural modules mediates the improvement of executive function with age. Mediation results shown as unstandardized regression coefficients. Significance of indirect effect (c'=0.007) was assessed using bootstrapped confidence intervals [0.002-0.012]. All models also include sex, in-scanner motion, and total network strength as covariates. * indicates p<0.01. ! 26 Figure 8. Modular evolution of structural brain networks across youth. From childhood through adulthood, structural brain networks become increasingly modular. The targeted strengthening of specific hub edges facilitates specialized information processing within distinct modules, and simultaneously enhances integration across modules. Hub edges are indicated by thick connections. ! 27 SUPPLEMENTAL INFORMATION! 28 Supplemental Figures
Supplemental Figure 1: Number of modules identified in group-level structural par-
titions. To examine alternative data-driven modular partitions of structural brain networks,
we varied γ over the interval [0,4] in increments of 0.05. The number of modules identified
in group-level consensus partitions increases as a function of γ. The similarity between
structural partitions and a priori functional partitions also increases with γ and the number
of identified structural modules. ∗ indicates alternative structural partitions identified at
plateaus for the number of modules. Bars are colored by the z-score of the Rand coeffi-
cient, which quantifies the similarity between structural partitions and the a priori func-
tional partition used throughout the main text. The 9-module structural partition identified
at γ=2.5 (marked by blue box) is used to examine age-related effects on modular segre-
gation in Figure 5. The z-score of the Rand coefficient is equal to 17.6 (p < 1× 10−10)
for this structural partition, suggesting a significant similarity with the functional partition
beyond chance.
30
051015201234GammaRastko_numComms_n882102030Zrand_n882_consensus_vs_Yeo7systemNumber of Modules123405101520102030ZRand*** = 1.5 = 3.1 = 2.5 Supplemental Figure 2: Data-driven structural network modules become more seg-
regated across youth. Here we demonstrate that regardless of the group-level consensus
partition used to define modules, modular segregation increases with age, as demonstrated
by a significant decrease in the mean participation coefficient. This developmental pattern
is replicated using a 5-module partition (A, γ=1.5), a 9-module partition (B, γ=2.5), and an
11-module partition (C, γ=3.1). The 9-module partition pictured in B is used to calculate
modular segregation in Figure 5. Blue line represents the best fit from a general additive
model; shaded area indicates 95% confidence interval. Models include participant sex,
in-scanner head motion, and total network strength as covariates.
31
8101214161820220.460.480.500.520.540.560.58age_in_yrsf(age_in_yrs)0.460.480.500.520.540.560.58810121416182022p < 1 × 10-10Age (years)Mean Participation CoefficientModuleModuleModule8101214161820220.670.680.690.700.710.720.730.74age_in_yrsf(age_in_yrs)810121416182022Age (years)0.670.680.690.700.710.720.730.74p < 1 × 10-10Mean Participation Coefficient8101214161820220.600.620.640.660.68age_in_yrsf(age_in_yrs)0.60Mean Participation Coefficient0.620.640.660.68810121416182022Age (years)p < 1 × 10-10 = 1.5 = 2.5 = 3.1234511234567891234567891011 Supplemental Experimental Procedures
Subjects
Diffusion tensor imaging (DTI) datasets were acquired as part of the Philadelphia Neu-
rodevelopmental Cohort (PNC), a large community-based study of brain development.
1601 subjects completed the cross-sectional neuroimaging protocol (Satterthwaite et al.,
2014). Datasets from 244 individuals were considered unusable due to incomplete ac-
quisition or incidental findings. The remaining 1357 participants underwent a rigorous
manual and automated quality assurance protocol for DTI datasets (Roalf et al., 2016),
which flagged 157 subjects for poor data quality (e.g., low temporal signal-to-noise ratio).
Of the remaining 1210 participants, 93 were flagged by automated quality assurance for
low quality or incomplete FreeSurfer reconstruction of T1-weighted images. Of the re-
maining 1117 participants, 235 subjects were excluded for meeting any of the following
criteria: gross radiological abnormalities, history of medical problems that might affect
brain function, history of inpatient psychiatric hospitalization, use of psychotropic medi-
cation at the time of data acquisition, missing data, and/or high levels of in-scanner head
motion (mean relative displacement between non-weighted volumes > 2mm), which has
been shown to impact measures derived from diffusion-weighted imaging (Roalf et al.,
2016; Yendiki et al., 2013). These exclusions produced a final sample consisting of 882
youths (mean age=15.06, SD=3.15; 389 males, 493 females).
Cognitive Assessment
The Penn computerized neurocognitive battery (Penn CNB) was administered to all
participants. The CNB consists of 14 tests adapted from tasks applied in functional neu-
roimaging to evaluate a broad range of cognitive domains (Gur et al., 2002; Gur et al.,
2012). These domains include executive control (abstraction and flexibility, attention,
working memory), episodic memory (verbal, facial, spatial), complex cognition (verbal
reasoning, nonverbal reasoning, spatial processing), social cognition (emotion identifica-
tion, emotion intensity differentiation, age differentiation) and sensorimotor and motor
speed. Accuracy and speed for each test were z-transformed. Cognitive performance was
summarized by a recent factor analysis (Moore et al., 2014) of both speed and accuracy
data, which delineated three factors corresponding to the efficiency of executive function,
episodic memory, and social cognition.
Data Acquisition
All MRI scans were acquired on the same 3T Siemens Tim Trio whole-body scanner
and 32-channel head coil at the Hospital of the University of Pennsylvania. DTI scans
were acquired using a twice- refocused spin-echo (TRSE) single-shot echo-planar imaging
32
(EPI) sequence (TR = 8100ms, TE = 82ms, FOV = 240mm2 /240mm2 ; Matrix = RL:
128/AP:128/Slices:70, in-plane resolution (x and y) 1.875 mm2; slice thickness = 2mm,
gap = 0; flip angle = 90◦/180◦/180◦, volumes = 71, GRAPPA factor = 3, bandwidth =
2170 Hz/pixel, PE direction = AP). This sequence used a four-lobed diffusion encoding
gradient scheme combined with a 90-180-180 spin-echo sequence designed to minimize
eddy-current artifacts . DTI data were acquired in two consecutive series consisting of 32
diffusion encoding gradient schemes. The complete sequence consisted of 64 diffusion-
weighted directions with b=1000s/mm2 and 7 interspersed scans where b=0 s/mm2. The
duration of DTI scans was approximately 11 minutes. The imaging volume was prescribed
in axial orientation covering the entire cerebrum with the topmost slice just superior to the
apex of the brain (Satterthwaite et al. 2014a).
In addition to the DTI scan, a map of
the main magnetic field (i.e., B0) was derived from a double-echo, gradient-recalled echo
(GRE) sequence, allowing us to estimate field distortions in each dataset.
Data Preprocessing
Two consecutive 32-direction acquisitions were merged into a single 64-direction time-
series. The skull was removed for each subject by registering a binary mask of a standard
fractional anisotropy (FA) map (FMRIB58 FA) to each subject's DTI image using a rigid-
body transformation (Smith et al., 2002). Eddy currents and subject motion were esti-
mated and corrected using FSL's eddy tool (Andersson and Sotiropoulos 2016). Diffusion
gradient vectors were then rotated to adjust for subject motion estimated by eddy. After
the field map was estimated, distortion correction was applied to DTI data using FSL's
FUGUE (Jenkinson et al., 2012). Lastly, DTI data was imported into DSI Studio software
and the diffusion tensor was estimated at each voxel.
DTI Tractography
Whole-brain fiber tracking was implemented for each subject in DSI Studio using a
modified fiber assessment by continuous tracking (FACT) algorithm with Euler interpola-
tion, initiating 1,000,000 streamlines after removing all streamlines with length less than
10mm or greater than 400mm (Yeh et al., 2013). Fiber tracking was performed with an
angular threshold of 45◦, a step size of 0.9375mm, and a fractional anisotropy (FA) thresh-
old determined empirically by Otzu's method, which optimizes the contrast between fore-
ground and background (Yeh et. al., 2013). Diffusivity measures (e.g., FA, mean diffusiv-
ity, radial diffusivity, axial diffusivity) were calculated along the path of each reconstructed
streamline. For each subject, tractography served as the basis for constructing structural
brain networks.
33
Network Construction
Following T1 reconstruction in FreeSurfer (version 5.3), cortical and subcortical gray
matter was parcellated according to the Lausanne atlas (Cammoun et al., 2012), which
includes whole-brain parcellations at multiple spatial scales (83, 129, 234, 463, and 1015
regions). Parcellations were defined in native space and co-registered to the first b = 0
volume of each subject's diffusion image using a rigid-body transform. To extend gray
matter region labels beyond the gray-white boundary, the atlas labels were dilated by 4mm
(Gu et al., 2015). Dilation involved filling non-labelled voxels with the statistical mode of
neighboring labels. 234 dilated brain regions defined the nodes for each subject's structural
brain network, which was represented as a weighted adjacency matrix A. Edges were
defined where at least one streamline connected a pair of nodes end-to-end. Edge weights
were primarily defined by the average FA along streamlines connecting any pair of nodes
(Misic et al., 2016; Bohlken et al., 2016). See Figure 2.
Functional Module Assignment
For the 234- and 463-region parcellations, we calculated a purity index for each Lau-
sanne label and corresponding voxels in the standard 7-system template image provided
by Yeo et al. (2011). This measure quantifies the maximum overlap of cortical Lausanne
labels and functional systems defined by Yeo et al. (2011). Each cortical Lausanne label
was assigned to a functional system by calculating the non-zero mode of all voxels in each
brain region. Subcortical regions were assigned to an eighth, subcortical module. The
primary modular partition defined for 234-node networks is shown in Figure 2. To de-
termine whether the functionally-defined network partition significantly fit the structural
connectivity data beyond chance, we quantified the modularity quality index (formally
defined below) of the functional partition imposed on structural brain networks. Briefly,
the modularity quality of a network partition quantifies how well that partition maximizes
the strength of within-module connections relative to a specified null model. Higher Q
values indicate that modules are highly segregated within a network, with strong within-
module connectivity and relatively weak between-module connectivity. We performed a
permutation test to examine the significance of the modularity quality of the functional
partition (QYeo) imposed on structural connectivity matrices. First, we permuted the as-
signment of N nodes to functional modules 1000 times, preserving the number of nodes
originally assigned to each module. We then calculated the modularity quality Qperm of
randomly-defined network partitions imposed on each subject's connectivity matrix, build-
ing a null distribution for Qperm. We used the calculated mean (µQperm) and standard de-
viation (σQperm) of the null distribution to derive a z-score based on the observed QYeo for
each subject (z-score = (QYeo−µQperm )
). Finally, we calculated the mean z-score across all
subjects to assess the significance of QYeo.
σQperm
34
Measures of Modular Segregation
We calculated the participation coefficient to quantify the relative balance of between-
module versus within-module connectivity for each brain region. Intuitively, this measure
describes the degree to which a brain region integrates information across distinct modules,
or the degree to which a brain region shows provincial connectivity among regions in its
own module. We define the participation coefficient Pi of node i as
(cid:16)ki(m)
(cid:17)2
ki
Pi = 1− ∑
m∈M
,
(1)
where m is a module in a set of modules M, and ki(m) is the weight of structural connec-
tions between node i and all nodes in module m (Guimera and Amaral 2005; Rubinov and
Sporns 2010). Moreover, Pi close to 1 indicates that a brain region is highly integrated
with regions in other modules, while a Pi close to 0 indicates that a brain region is highly
segregated, with strong connectivity among other regions in its own module. To quantify
the segregation of specific modules, we average Pi across all brain regions assigned to the
same module. To quantify global network segregation, we average Pi across all nodes in
the network.
Alternative Measures of Modular Segregation
To ensure that our results were not dependent on specific network metrics, we cal-
culated alternative measures of modular segregation. First, we calculated the average
strength of all within-module connections (a measure of structural coherence), and the
average strength of all between-module connections (a measure of structural integration)
in the network (Gu et al., 2015). These metrics provide additional insights into the segre-
gation of information processing within distinct modules, and the degree to which modules
are integrated across the network (see Figure 4). Alternatively, we calculated the subject-
specific modularity quality (Q) of group-level functional and structural network partitions.
As discussed above, this measure provides an index of how well a network can be decom-
posed into a hard partition where nodes within the same module demonstrate particularly
strong connectivity beyond chance. We also calculated Qsub j for subject-specific consen-
sus partitions (see detailed procedure below), which was not dependent on a group-level
partition. We calculated the modularity Q of a network partition S based on the following
modularity quality function:
(cid:104)
(cid:105)
Q(S) =
1
2m ∑
i j
Ai j − γPi j
δ (gi,g j),
(2)
where m is the total weight of A, P represents the expected strength of connections accord-
ing to a specified null model (Newman, 2004), γ is a structural resolution parameter that
35
determines the size of modules, and δ (gi,g j) is equal to unity when brain regions i and j
are assigned to same community gi, and is zero otherwise.
Community Detection in Structural Brain Networks
Primary analyses relied on an a priori functional partition to define network modules.
We additionally estimated network modules directly from the structural connectivity data
using community detection procedures. Communities were defined by maximizing the
modularity quality function using a generalization of the Louvain heuristic (Blondel et
al., 2008; Mucha et al., 2010). Because the Louvain algorithm is degenerate (Good et
al., 2010; Sporns and Betzel 2016), it is essential to perform modularity maximization
multiple times in order to identify a stable consensus partition that accurately reflects the
solutions offered by each optimization. Accordingly, we applied a locally greedy Louvain-
like modularity-optimization procedure (Blondel et al., 2008) 100 times for each subject
in order to define an "agreement" matrix A(cid:48) where A(cid:48)
i j was equal to the probability that
nodes i and j were assigned to the same community over the 100 iterations. If A(cid:48) was de-
terministic (edge weights were binary), then the algorithm had converged and the resultant
partition was defined as the consensus. Otherwise, we performed 100 iterations of modu-
larity optimization on A(cid:48) in order to generate a new agreement matrix A(cid:48)(cid:48). This procedure
was repeated until convergence (Lancichinetti and Fortunato, 2014). When performing
modularity optimization on an agreement matrix (e.g., A(cid:48) or A(cid:48)(cid:48)), we defined an alternative
null model P(cid:48) by permuting community assignments across nodes (Bassett et al., 2013).
Once a consensus partition was identified for each subject, we computed a group-
level consensus across the full PNC cohort (n=882). To do this, we used a Louvain-like
procedure to detect communities in a group-level agreement matrix A(cid:48)
group. Edge weights
in A(cid:48)
group were equal to the proportion of times that each pair of nodes was assigned to
the same community across subject-level consensus partitions. As above, 100 iterations
of modularity optimization were performed on A(cid:48)
group became
binary, indicating that the algorithm had converged on a group-level consensus partition.
Both subject-level and group-level consensus partitions were computed over a wide range
of γ ([0,4], in increments of 0.05) to explore variations in community structure. We plotted
the number of group-level consensus modules as a function of γ, and found several plateaus
indicating partition stability (Fenn et al., 2009; see Figure S1). In order to directly compare
the organization of data-driven, modularity-based partitions and the a priori functional
partition, we quantified the partition similarity using the z-score of the Rand coefficient
(Traud et al., 2011). For two partitions X and Y , we calculated the Rand z-score in terms
of the total number of node pairs in the network M, the number of pairs MX assigned to
the same module in partition X, the number of pairs MY that are in the same module in
partition Y , and the number of pairs of nodes wXY that are assigned to the same module
group until the resulting A(cid:48)(cid:48)
36
both in partition X and in partition Y . The z-score of the Rand coefficient is defined by:
zXY =
1
σwXY
wXY − MXMY
M ,
(3)
where σwXY is the standard deviation of wXY . The mean partition similarity is determined
by the mean value of zXY over all possible partition pairs for X (cid:54)= Y . Moreover, zXY denotes
the similarity of partitions X and Y beyond chance. Figure S1 shows the similarity between
all group-level structural partitions and the primary functional partition used in this study.
Measures of Network Integration
For each subject's structural brain network A, the topological length or distance of each
edge Ai j was computed as the reciprocal of the edge weight ( 1
). The path length between
Ai j
any pair of nodes is defined as the sum of the edge lengths along the shortest path connect-
ing them (Rubinov and Sporns, 2010). Global efficiency provides a theoretical prediction
of how easily information can flow across a network via the shortest path between all pairs
of nodes, and is defined by
(cid:18)
(cid:19)−1
Eglob(G) =
1
n ∑
i∈N
∑ j∈N, j(cid:54)=i
di j
n− 1
,
(4)
where n is the number of nodes, and di j is the shortest path length between node i and
node j.
To examine the possible role of specific edges as integrative hub connections within
the network, we calculated the weighted edge betweenness centrality (EBC) for each edge.
Edge betweenness identifies important hub connections by providing a measure of how
much a given connection participates in the shortest paths of communication through a
network, and thus contributes to global efficiency (Brandes, 2001).
EBC = ∑
hk
ρi j
hk
ρhk
,
(5)
where ρi j
hk denotes the number of shortest paths between nodes h and k that include
edge i j, and ρhk denotes the total number of shortest paths between h and k. After cal-
culating EBC individually for each weighted network Ai j (n=882), we normalized each
subjects(cid:48) EBC values by their maximum observed EBC, resulting in a bounded measure
[0,1] (Gong et al., 2009). We calculated the mean normalized EBC for each network edge
across subjects, and defined hub edges as those connections within the top quartile of
37
normalized edge betweenness across all network edges. Following group-level analysis,
which identified a subset of edges that significantly strengthened with age, we performed a
permutation-based test to assess whether connections that significantly strengthened with
age were enriched for hub edges (see below).
Group-level analyses
Prior work has demonstrated that brain development is not a linear process (Paus et al.,
1999, Shaw et al., 2006). Accordingly, group-level analyses of structural brain network
metrics were flexibly modeled using penalized splines within a General Additive Model
(GAM) implemented in the R package "mgcv" (https://cran.r-project.org/web/packages/mgcv/index.html;
Wood 2004; Wood 2011). Such an approach allows for detection of nonlinearities in the
relationship between age and measures of modular segregation without defining a set of
functions a priori (such as polynomials). Importantly, the GAM estimates nonlinearities
using restricted maximum likelihood (REML), and determines a penalty with increasing
nonlinearity in order to avoid overfitting the data. Due to this penalty, the GAM only
models nonlinearities when they explain additional variance in the data above and beyond
linear effects.
First, we used penalized splines to estimate nonlinear developmental patterns of mod-
ular segregation. Within this model we included covariates for sex, head motion, and total
network strength. Accordingly, the final model equations for estimating age effects on
modular segregation (mean participation coefficient) were as follows:
Modular segregation = spline(age) + sex + motion + total network strength
An identical model was used when estimating age effects on the participation coefficient
of individual brain regions. Similarly, we applied this model across all network edges in
order to assess linear and nonlinear age effects on the strength of individual connections.
For all analyses, multiple comparisons were controlled using the False Discovery Rate
(q<0.05).
Permutation Testing
We performed permutation-based tests across network edges in order to assess (i)
whether the edges that significantly strengthened with age were localized to within-module
connections beyond chance, (ii) whether edges that significantly strengthen with age were
enriched for hub edges, and (iii) whether these ages had elevated edge betweenness cen-
trality beyond chance.
First, we permuted a binary edge label specifying whether each edge connects nodes
within or between modules 1000 times. Then for permuted samples of within- and between-
module edges, we counted the number of edges that were shown to significantly strengthen
38
with age in group-level analysis. We then rank-ordered the number of edges shown to sig-
nificantly strengthen with age for permuted within-module edge samples, and determined
where the observed number of within-module edges that strengthen with age falls relative
to this null distribution.
Second, we evaluated whether edges that significantly strengthen with age were en-
riched for hub edges. We permuted a binary edge label defining hub or non-hub edges
1000 times. For each permuted sample, we counted the number of edges that significantly
strengthened with age in group-level analysis. Then, we rank-ordered the number of per-
muted hub edges shown to significantly strengthen with age, and compared these values
with the observed number of hub edges that strengthened with age.
Third, we evaluated whether edges that significantly strengthen with age had higher
edge betweenness centrality than anticipated by chance. We permuted normalized edge
betweenness centrality values 1000 times. For each permuted sample, we calculated the
mean EBC of within-module edges and between-module edges that significantly strength-
ened with age. We rank-ordered the mean EBC of permuted within-module and between-
module edges that strengthened with age, and compared these values with the observed
means for within- and between-module edges separately (Figure 6D).
Methodological Replications
To verify that observed age-related increases in modular segregation were not simply
due to specific network construction choices, we repeated developmental inferences on
modular segregation using a variety of other parameters. First, we examined age effects
on modular segregation (mean participation coefficient) using a data-driven structural par-
tition identified at the group level (see Figure S2B., Figure 5B, and detailed procedure
above). Alternatively, we also calculated the modularity quality index for each subject(cid:48)s
optimal partition at γ=2.5 (QSub j), where a higher QSub j indicates greater modular sege-
gration (Figure 5C). Next, we examined modular segregation (mean participation coeffi-
cient) using the a priori functional partition assigned to a higher-resolution parcellation
of the brain (463 nodes instead of 234; see Figure 5D). We also measured modular seg-
regation of the functional partition using structural networks with alternative edge weight
definitions. While primary analyses focused on FA-weighted structural networks, we also
measured modular segregation in streamline-weighted networks (see Figure 5E), where
edge weights were equal to the number of streamlines connecting a pair of nodes (Bassett
et al., 2011), and additionally, where edge weights were defined by streamline density:
the number of connecting streamlines divided by the total regional volume of each node
pair (Baker et al., 2015; see Figure 5F). In addition to examining age-related patterns of
modular segregation using alternative network measures and parameters, we also repeated
analyses including the following additional covariates in the GAM described above: race,
39
maternal education, handedness, and total brain volume.
Relationship Between Modular Segregation and Global Network Efficiency
First, we examined age-related effects on global efficiency using the same GAM as
above:
Global efficiency = spline(age) + sex + motion + total network strength
(see Figure 6A). The relationship between global efficiency and modular segregation was
assessed within a GAM while controlling for age in addition to other covariates described
above (Figure 6B). Moreover, the model equation was as follows:
Modular segregation = Global efficiency + spline(age) + sex + motion + total network
strength
To assess whether global efficiency was related to the weight of specific network connec-
tions that strengthened with age, we estimated the following GAMs:
Global efficiency = Average strength of within-module edges + spline(age) + sex + motion
+ total network strength
Global efficiency = Average strength of between-module edges + spline(age) + sex + mo-
tion + total network strength
(see Figure 6E and Figure 6F).
Associations with Executive Function
To examine the association between modular segregation and executive efficiency, we
included a spline age term in the model to account for the variance associated with linear
and nonlinear age-related changes in executive ability. The final model equation was as
follows:
Modular segregation = spline(age) + executive efficiency + sex + motion + total network
strength
Using the same GAM, we also evaluated the association between the segregation of indi-
vidual modules (e.g., frontoparietal) and three cognitive efficiency factor scores: executive
function, memory, and social cognition (see Figure 7A). We note that 2 participants of the
full 882 sample had incomplete cognitive datasets: subsequent analyses examining associ-
ations between executive function and modular segregation focused on the remaining 880
participants. Visualization of GAM model fits were created using the "visreg" package
in R (https://cran.r-project.org/web/packages/visreg/). In Figure 3A, Figure 5, and Figure
40
6B, one outlying datapoint was beyond the axis range, and was excluded for visualiza-
tion purposes only: group-level analyses and reported results include data points for all
subjects.
Mediation analyses
Linear mediation analyses investigated whether age-related improvement in execu-
tive function was mediated by modular segregation and/or global efficiency (Preacher and
Hayes, 2008). First, we regressed out the effects of nuisance covariates (sex, head motion,
and total network strength) on the independent (X), dependent (Y), and mediating (M)
variables. The residuals were then used in our mediation analysis. The significance of the
indirect effect was evaluated using bootstrapped confidence intervals within the R package
"lavaan" (https://cran.r-project.org/web/packages/lavaan/). Specifically, we examined the
total effect of age on executive performance (c path; Figure 7B), the relationship between
age and modular segregation (a path), the relationship between modular segregation and
executive function (b path), and the direct effect of age on executive efficiency after in-
cluding modular segregation as a mediator in the model (c(cid:48) path). The significance of the
indirect effect of age on executive function through the proposed mediator (modular seg-
regation) was tested using bootstrapping procedures, which minimize assumptions about
the sampling distribution (Preacher and Hayes, 2008). This approach involves calculating
unstandardized indirect effects for each of 10,000 bootstrapped samples and calculating
the 95% confidence interval. This procedure was repeated to assess (i) whether the seg-
regation of the frontoparietal module mediated developmental improvements in executive
function, (ii) whether the segregation of the default mode module mediated developmental
improvements in social cognition, and (iii) whether age-related increases in global effi-
ciency mediated improvements in executive function.
Data Visualization
Network partitions and regional results (Figure 2, Figure 3C, and Figure S2) were vi-
sualized on the cortical white matter surface using FreeSurfer visualization tools in MAT-
LAB. While age effects on the participation coefficient for subcortical brain regions are
not visualized in Figure 3C, these regions were included in all analyses. Brain network
visualizations in Figure 4 and Figure 6 were generated using BrainNet Viewer (Xia et al.
2013).
41
Supplemental References
1. Andersson, J.L. and Sotiropoulos, S.N. (2016). An integrated approach to correction
for off-resonance effects and subject movement in diffusion MR imaging. Neuroim-
age, 125,1063-1078.
2. Bassett, D.S., Porter, M.A., Wymbs, N.F., Grafton, S.T., Carlson, J.M. and Mucha,
P.J. (2013). Robust detection of dynamic community structure in networks. Chaos,
23, 013142.
3. Fenn, D.J. Porter, M.A., McDonald, M., Williams, S., Johnson, N.F., and Jones, N.S.
(2009). Dynamic communities in multichannel data: An application to the foreign
exchange market during the 2007-2008 credit crisis. Chaos 19, 033119.
4. Giedd, J.N., Blumenthal, J., Jeffries, N.O., Castellanos, F.X., Liu, H., Zijdenbos, A.,
Paus, T., Evans, A.C., Rapoport, J.L. (1999). Brain development during childhood
and adolescence: a longitudinal MRI study. Nature Neuroscience 2(10), 861-3.
5. Gong, G., He, Y., Concha, L., Lebel, C., Gross, D.W., Evans, A.C. and Beaulieu, C.
(2009). Mapping anatomical connectivity patterns of human cerebral cortex using
in vivo diffusion tensor imaging tractography. Cerebral Cortex, 19, 524-536.
6. Good, B. H., Montjoye, Y., and Clauset, A. (2010). Performance of modularity
maximization in practical contexts. Physical Review E 81, 046106.
7. Gu, S., Pasqualetti, F., Cieslak, M., Telesford, Q.K., Alfred, B.Y., Kahn, A.E.,
Medaglia, J.D., Vettel, J.M., Miller, M.B., Grafton, S.T. and Bassett, D.S. (2015).
Controllability of structural brain networks. Nature Communications, 6.
8. Lancichinetti, A. and Fortunato, S., 2012. Consensus clustering in complex net-
works. Scientific reports, 2.
9. Newman, M.E.J., 2004. Analysis of weighted networks. Physical Review E, 70,
056131.
10. Paus, T., Zijdenbos, A., Worsley, K., Collins, D.L., Blumenthal, J., Giedd, J.N.,
Rapoport, J.L. and Evans, A.C. (1999). Structural maturation of neural pathways in
children and adolescents: in vivo study. Science, 283, 1908-1911.
11. Shaw, P., Greenstein, D., Lerch, J., Clasen, L., Lenroot, R., Gogtay, N.E.E.A.,
Evans, A., Rapoport, J. and Giedd, J. (2006). Intellectual ability and cortical de-
velopment in children and adolescents. Nature, 440, 676-679.
12. Xia, M., Wang, J. and He, Y. (2013). BrainNet Viewer: a network visualization tool
for human brain connectomics. PLoS One, 8, e68910.
42
|
1904.10396 | 1 | 1904 | 2019-04-18T14:58:52 | Is coding a relevant metaphor for building AI? A commentary on "Is coding a relevant metaphor for the brain?", by Romain Brette | [
"q-bio.NC",
"cs.AI",
"cs.LG"
] | Brette contends that the neural coding metaphor is an invalid basis for theories of what the brain does. Here, we argue that it is an insufficient guide for building an artificial intelligence that learns to accomplish short- and long-term goals in a complex, changing environment. | q-bio.NC | q-bio |
Is coding a relevant metaphor for building AI?
A commentary on "Is coding a relevant metaphor for the brain?", by Romain Brette
Adam Santoro∗, Felix Hill∗, David G. T. Barrett,
David Raposo, Matthew Botvinick, & Timothy Lillicrap
{adamsantoro,felixhill,barrettdavid,draposo,botvinick,countzero}@google.com
Deepmind, London UK
Abstract
Brette contends that the neural coding metaphor is an invalid basis for theories of
what the brain does (Brette, 2019). Here, we argue that it is an insufficient guide for
building an artificial intelligence (AI) that learns to accomplish short- and long-term
goals in a complex, changing environment.
The goal of neuroscience is to explain how the brain enables intelligent behaviour, while
the goal of agent-based AI is to build agents that behave intelligently. Neuroscience, Brette
attests, has suffered from an exaggerated (and technically inaccurate) concern for the codes
transmitted by particular parts of the brain. In AI, on the other hand, some of the most
notable recent progress has been made not by deeply considering neural coding and its
implications, but by focusing on higher-level principles from optimization, learning, and
control.
Thanks to deep artificial networks trained via backpropagation, we now have artificial
learning systems capable of impressive exhibitions of specific human-like skills, such as object
recognition and language translation (e.g., He et al, 2016; Vaswani et al, 2017). In artificial,
rather than biological, neural networks, we can more tractably characterize the relationship
between a models neural codes, behaviour, and its external world. AI researchers have full
access to their models' input data distribution, can visualise weights and activations in any
part of the network and even make causal interventions on them, and can quickly implement
new models informed by any coding hypotheses they may have.
Nevertheless, in-depth analysis of a models internal representations is of increasingly rare
concern for getting these models to work. Consider AlphaGo, which is one of the more
∗Equal Contribution
1
compelling recent breakthroughs in AI (Silver et al, 2016). Researchers on this project
precisely defined the models goals, the dynamics underlying the models interactions with
its environment, how the model plans its actions, and how the model learns. Each of these
components contributes to the models success, and yet none of them fundamentally depend
on considerations from neural coding.
This is not to say that we cannot usefully apply representational analyses to such agents
post-hoc, regardless of whether the representations satisfy Brettes criteria for neural codes
(Barack et al, 2019). Indeed, since the earliest days of connectionism researchers have been
interested in the neural codes that emerge when a clearly-specified learning algorithm is
applied to a well-understood model trained to execute a particular task. A more recent and
important collaboration between AI and neuroscience revealed insight into the conditions
under which well-known codes can emerge: grid-cells can increasingly be understood as the
product of particular optimization processes (Banino et al, 2018; Cueva et al, 2018). A
key feature of these examples, however, is the central descriptive role given to the learning
algorithms, architectures, and optimization objectives; neural coding was incidental, and
in many cases the codes were not fundamental, privileged primitives on top of which the
models were built (Marblestone et al, 2016).
If the broader aim of agent-based AI (Russell et al, 2016) is to produce a system that
accomplishes short- and long-term goals in a complex, changing environment, then there
may be a more pernicious problem to the neural coding framework than it simply being
out of vogue in modern AI. How internal responses arrive from given stimuli -- a goal that is
implicit in the neural coding metaphor -- may be logically insufficient for producing intelligent
behaviour. In outlining the reasons why, we recall arguments that any system -- artificial or
biological -- needs to exert control over its environment to achieve intelligent behaviour .
First, the observations with which an agent may compute do not exist as a prespecified
dataset, independently of the agent's actions in the world. Rather, it is precisely the decisions
that the agent takes in that world that determine the sensory data from which it learns.
Second, [w]ithout an ongoing participation and perception of the world there is no meaning
for an agent (Brooks, 1991). An agent participating in an external world that responds to
its decisions learns useful, reliable, and meaningful interactions (Cisek, 1999). It is these
meaningful interactions that ground the agents representations and allow them to be used
for understanding and reasoning about its world. Therefore, insofar as neural coding is
understood as a framework to help understand a systems internal stimulus-response patterns,
it is a logically insufficient framework for designing AI because of its failure to engage with
the agent-environment causal loop.
Given these considerations, what, then, can we say about neural codings role in describing
the brain? In neuroscience, we ultimately care about understanding how the brain enables
intelligent behaviour. It is often argued that such an understanding cannot come from
analyzing low-level, mechanistic details such as neural codes, because [a] description of neural
activity and connections is not synonymous with knowing what they are doing to cause
behavior. (Krakauer, 2017). For this level of understanding we need high-level computational
and algorithmic theories that embrace agent-environment interactions. The history of AI tells
us that the most useful principles, and the richest theoretical insights, emerged from studying
2
control, optimization, and learning processes rather than the particularities of representations
or codes (Sutton, 2019). A focus on inferring such processes using our increasing quantities
of neural data, rather than characterizing neural codes for their own sake, may also be the
most productive way of making progress on understanding intelligent behaviour in humans
and animals.
Acknowledgments
Thanks to Francis Song and Drew Jaegle for helpful comments.
References
Banino, A., Barry, C., Uria, B., Blundell, C., Lillicrap, T., Mirowski, P., ... & Wayne, G.
(2018). Vector-based navigation using grid-like representations in artificial agents. Nature,
557(7705), 429.
Barack, D. and Jaegle, A. (2019) Codes, Functions, and Causes: A Critique of Brettes
Conceptual Analysis of Coding
Brooks, R. A. (1991). Intelligence without reason.
Cisek, P. (1999). Beyond the computer metaphor: Behaviour as interaction. Journal of
Consciousness Studies, 6(11-12), 125-142.
Cueva, C. J., & Wei, X. X. (2018). Emergence of grid-like representations by training
recurrent neural networks to perform spatial localization. arXiv preprint arXiv:1803.07770.
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition.
In Proceedings of the IEEE conference on computer vision and pattern recognition (pp.
770-778).
Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., & Poeppel, D. (2017).
Neuroscience needs behavior: correcting a reductionist bias. Neuron, 93(3), 480-490.
Marblestone, A. H., Wayne, G., & Kording, K. P. (2016). Toward an integration of deep
learning and neuroscience. Frontiers in computational neuroscience, 10, 94.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., ... &
Dieleman, S. (2016). Mastering the game of Go with deep neural networks and tree search.
nature, 529(7587), 484.
Russell, S. J., & Norvig, P. (2016). Artificial intelligence: a modern approach. Malaysia;
Pearson Education Limited.
Sutton, R. (2019). The Bitter Lesson. Available at http://incompleteideas.net/IncIdeas/
3
BitterLesson.html
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin,
I. (2017). Attention is all you need. In Advances in neural information processing systems
(pp. 5998-6008).
4
|
1509.01972 | 1 | 1509 | 2015-09-07T10:21:42 | The Brain Uses Reliability of Stimulus Information when Making Perceptual Decisions | [
"q-bio.NC"
] | In simple perceptual decisions the brain has to identify a stimulus based on noisy sensory samples from the stimulus. Basic statistical considerations state that the reliability of the stimulus information, i.e., the amount of noise in the samples, should be taken into account when the decision is made. However, for perceptual decision making experiments it has been questioned whether the brain indeed uses the reliability for making decisions when confronted with unpredictable changes in stimulus reliability. We here show that even the basic drift diffusion model, which has frequently been used to explain experimental findings in perceptual decision making, implicitly relies on estimates of stimulus reliability. We then show that only those variants of the drift diffusion model which allow stimulus-specific reliabilities are consistent with neurophysiological findings. Our analysis suggests that the brain estimates the reliability of the stimulus on a short time scale of at most a few hundred milliseconds. | q-bio.NC | q-bio |
The Brain Uses Reliability of Stimulus Information
when Making Perceptual Decisions
Sebastian Bitzer1
[email protected]
Stefan J. Kiebel1
[email protected]
1Department of Psychology, Technische Universitat Dresden, 01062 Dresden, Germany
Abstract
In simple perceptual decisions the brain has to identify a stimulus based on noisy
sensory samples from the stimulus. Basic statistical considerations state that the
reliability of the stimulus information, i.e., the amount of noise in the samples,
should be taken into account when the decision is made. However, for perceptual
decision making experiments it has been questioned whether the brain indeed uses
the reliability for making decisions when confronted with unpredictable changes
in stimulus reliability. We here show that even the basic drift diffusion model,
which has frequently been used to explain experimental findings in perceptual
decision making, implicitly relies on estimates of stimulus reliability. We then
show that only those variants of the drift diffusion model which allow stimulus-
specific reliabilities are consistent with neurophysiological findings. Our analysis
suggests that the brain estimates the reliability of the stimulus on a short time scale
of at most a few hundred milliseconds.
1
Introduction
In perceptual decision making participants have to identify a noisy stimulus. In typical experiments,
only two possibilities are considered [1]. The amount of noise on the stimulus is usually varied to
manipulate task difficulty. With higher noise, participants' decisions are slower and less accurate.
Early psychology research established that biased random walk models explain the response distri-
butions (choice and reaction time) of perceptual decision making experiments [2]. These models
describe decision making as an accumulation of noisy evidence until a bound is reached and cor-
respond, in discrete time, to sequential analysis [3] as developed in statistics [4]. More recently,
electrophysiological experiments provided additional support for such bounded accumulation mod-
els, see [1] for a review.
There appears to be a general consensus that the brain implements the mechanisms required for
bounded accumulation, although different models were proposed for how exactly this accumulation
is employed by the brain [5, 6, 1, 7, 8]. An important assumption of all these models is that the
brain provides the input to the accumulation, the so-called evidence, but the most established models
actually do not define how this evidence is computed by the brain [3, 5, 9, 1]. In this contribution, we
will show that addressing this question offers a new perspective on how exactly perceptual decision
making may be performed by the brain.
Probabilistic models provide a precise definition of evidence: Evidence is the likelihood of a de-
cision alternative under a noisy measurement where the likelihood is defined through a generative
model of the measurements under the hypothesis that the considered decision alternative is true. In
particular, this generative model implements assumptions about the expected distribution of mea-
surements. Therefore, the likelihood of a measurement is large when measurements are assumed,
1
by the decision maker, to be reliable and small otherwise. For modelling perceptual decision making
experiments, the evidence input, which is assumed to be pre-computed by the brain, should simi-
larly depend on the reliability of measurements as estimated by the brain. However, this has been
disputed before, e.g. [10]. The argument is that typical experimental setups make the reliability of
each trial unpredictable for the participant. Therefore, it was argued, the brain can have no correct
estimate of the reliability. This issue has been addressed in a neurally inspired, probabilistic model
based on probabilistic population codes (PPCs) [7]. The authors have shown that PPCs can imple-
ment perceptual decision making without having to explicitly represent reliability in the decision
process. This remarkable result has been obtained by making the comprehensible assumption that
reliability has a multiplicative effect on the tuning curves of the neurons in the PPCs. Reliability,
therefore, was implicitly represented in the tuning curves of model neurons which leaves the open
question why tuning curves exhibit a multiplicative effect of reliability.
In this paper we will consider this question from a more conceptual perspective under the interpre-
tation that the multiplicative effect on tuning curves results from adaptation of internal estimates
of measurement reliability in the brain. We show that even a simple, widely used bounded accu-
mulation model, the drift diffusion model, is based on some estimate of measurement reliability.
Using this result, we will analyse the results of a perceptual decision making experiment [11] and
will show that the recorded behaviour together with neurophysiological findings strongly favours
the hypothesis that the brain weights evidence using a current estimate of measurement reliability,
even when reliability changes unpredictably across trials.
This paper is organised as follows: We first introduce the notions of measurement, evidence and
likelihood in the context of the experimentally well-established random dot motion (RDM) stimulus.
We define these quantities formally by resorting to a simple probabilistic model which has been
shown to be equivalent to the drift diffusion model [12, 13]. This, in turn, allows us to formulate
three competing variants of the drift diffusion model that either do not use trial-dependent reliability
(variant CONST), or do use trial-dependent reliability of measurements during decision making
(variants DDM and DEPC, see below for definitions). Finally, using data of [11], we show that
only variants DDM and DEPC, which use trial-dependent reliability, are consistent with previous
findings about perceptual decision making in the brain.
2 Measurement, evidence and likelihood in the random dot motion stimulus
The widely used random dot motion (RDM) stimulus consists of a set of randomly located dots
shown within an invisible circle on a screen [14]. From one video frame to the next some of the
dots move into one direction which is fixed within a trial of an experiment, i.e., a subset of the dots
moves coherently in one direction. All other dots are randomly replaced within the circle. Although
there are many variants of how exactly to present the dots [15], the main idea is that the coherently
moving dots indicate a motion direction which participants have to decide upon. By varying the
proportion of dots which move coherently, also called the 'coherence' of the stimulus, the difficulty
of the task can be varied effectively.
We will now consider what kind of evidence the brain can in principle extract from the RDM stim-
ulus in a short time window, for example, from one video frame to the next, within a trial. For
simplicity we call this time window 'time point' from here on, the idea being that evidence is ac-
cumulated over different time points, as postulated by bounded accumulation models in perceptual
decision making [3, 1].
At a single time point, the brain can measure motion directions from the dots in the RDM display. By
construction, a proportion of measurable motion directions will be into one specific direction, but,
through the random relocation of other dots, the RDM display will also contain motion in random
directions. Therefore, the brain observes a distribution of motion directions at each time point. This
distribution can be considered a 'measurement' of the RDM stimulus made by the brain. Due to the
randomness of each time frame, this distribution varies across time points and the variation in the
distribution reduces for increasing coherences. We have illustrated this using rose histograms in Fig.
1 for three different coherence levels.
To compute the evidence for the decision whether the RDM stimulus contains predominantly motion
to one of the two considered directions, e.g., left and right, the brain must check how strongly these
2
Figure 1: Illustration of possible motion direction distributions that the brain can measure from
an RDM stimulus. Rows are different time points, columns are different coherences. The true,
underlying motion direction was 'left', i.e., 180◦. For low coherence (e.g., 3.2%) the measured
distribution is very variable across time points and may indicate the presence of many different
motion directions at any given time point. As coherence increases (from 9% to 25.6%), the true,
underlying motion direction will increasingly dominate measured motion directions simultaneously
leading to decreased variation of the measured distribution across time points.
directions are represented in the measured distribution, e.g., by estimating the proportion of motion
towards left and right. We call these proportions evidence for left, eleft, and evidence for right,
eright. As the measured distribution over motion directions may vary strongly across time points, the
computed evidences for each single time point may be unreliable. Probabilistic approaches weight
evidence by its reliability such that unreliable evidence is not over-interpreted. The question is: Does
the brain perform this reliability-based computation as well? More formally, for a given coherence,
c, does the brain weight evidence by an estimate of reliability that depends on c: l = e · r(c)1 and
which we call 'likelihood', or does it ignore changing reliabilities and use a weighting unrelated to
coherence: e(cid:48) = e · ¯r?
3 Bounded accumulation models
Bounded accumulation models postulate that decisions are made based on a decision variable. In
particular, this decision variable is driven towards the correct alternative and is perturbed by noise.
A decision is made, when the decision variable reaches a specific value. In the drift diffusion model,
these three components are represented by drift, diffusion and bound [3]. We will now relate the
typical drift diffusion formalism to our notions of measurement, evidence and likelihood by linking
the drift diffusion model to probabilistic formulations.
In the drift diffusion model, the decision variable evolves according to a simple Wiener process with
drift. In discrete time the change in the decision variable y can be written as
δy = yt − yt−δt = vδt +
√
(1)
where v is the drift, t ∼ N (0, 1) is Gaussian noise and s controls the amount of diffusion. This
equation bears an interesting link to how the brain may compute the evidence. For example, it has
been stated in the context of an experiment with RDM stimuli with two decision alternatives that
the change in y, often called 'momentary evidence', "is thought to be a difference in firing rates of
direction selective neurons with opposite direction preferences." [11, Supp. Fig. 6] Formally:
δtst
δy = ρleft,t − ρright,t
(2)
1For convenience, we use imprecise denominations here. As will become clear below, l is in our case a
Gaussian log-likelihood, hence, the linear weighting of evidence by reliability.
3
0°45°90°135°180°225°270°315°246810120°45°90°135°180°225°270°315°24681012140°45°90°135°180°225°270°315°510152025300°45°90°135°180°225°270°315°1234567890°45°90°135°180°225°270°315°2468100°45°90°135°180°225°270°315°5101520253035 3.2%time point 1 9.0%25.6%time point 2 √
l = − log(
2πδtσ) − 1
2
(xt ± 1)2
δtσ2
.
(5)
where ρleft,t is the firing rate of the population selective to motion towards left at time point t.
Because the firing rates ρ depend on the considered decision alternative, they represent a form of
evidence extracted from the stimulus measurement instead of the stimulus measurement itself (see
our definitions in the previous section). It is unclear, however, whether the firing rates ρ just represent
the evidence (ρ = e(cid:48)) or whether they represent the likelihood, ρ = l, i.e., the evidence weighted by
coherence-dependent reliability.
To clarify the relation between firing rates ρ, evidence e and likelihood l we consider probabilistic
models of perceptual decision making. Several variants have been suggested and related to other
forms of decision making [6, 16, 9, 7, 12, 17, 18]. For its simplicity, which is sufficient for our
argument, we here consider the model presented in [13] for which a direct transformation from
probabilistic model to the drift diffusion model has already been shown. This model defines two
Gaussian generative models of measurements which are derived from the stimulus:
p(xtleft) = N (−1, δtσ2)
(3)
where σ represents the variability of measurements expected by the brain. Similarly, it is assumed
that the measurements xt are sampled from a Gaussian with variance σ2 which captures variance
both from the stimulus and due to other noise sources in the brain:
p(xtright) = N (1, δtσ2)
(4)
Evidence for a decision is computed in this model by calculating the likelihood of a measurement
xt under the hypothesised generative models. To be precise we consider the log-likelihood which is
xt ∼ N (±1, δtσ2).
There are three important points: 1) The first term on the right hand side means that l increases
independently of the stimulus xt for decreasing σ. This contribution cancels when the difference
between the likelihoods for left and right is computed. 2) The likelihood is large for a measurement
xt, when xt is close to the values hypothesised for the decision alternatives, i.e., −1 and 1. 3) The
contribution of the stimulus is weighted by the assumed reliability r = σ−2.
This model of the RDM stimulus is simple but captures the most important properties of the stim-
ulus. In particular, a high coherence RDM stimulus has a large proportion of motion in the correct
direction with very low variability of measurements whereas a low coherence RDM stimulus tends
to have lower proportions of motion in the correct direction, with high variability (cf. Fig. 1). The
Gaussian model captures these properties by adjusting the noise variance such that a high coherence
corresponds to low noise and low coherence to high noise: Under high noise the values xt will vary
strongly and tend to be rather distant from −1 and 1, whereas for low noise the values xt will be close
to −1 or 1 with low variability. Hence, as expected, the model produces large evidences/likelihoods
for low noise and small evidences/likelihoods for high noise.
This intuitive relation between stimulus and probabilistic model is the basis for us to proceed to
show that the reliability of the stimulus r, connected to the coherence level c, appears at a prominent
position in the drift diffusion model. Crucially, the drift diffusion model can be derived as the sum
of log-likelihood ratios across time [3, 9, 12, 13]. In particular, a discrete time drift diffusion process
can be derived by subtracting the likelihoods of Eq. (5):
(xt + 1)2 − (xt − 1)2
2rxt
δt
.
δy = lright − lleft =
=
2δtσ2
(6)
Consequently, the change in y is Gaussian: δy ∼ N (2r/δt, 4r2σ2/δt). This replicates the model
described in [11, Supp. Fig. 6] where the parameterisation of the model, however, more di-
rectly followed that of the Gaussian distribution and did not explicitly take time into account:
δy ∼ N (Kc, S2), where K and S are free parameters and c is coherence of the RDM stimulus.
By analogy to the probabilistic model, we, therefore, see that the model in [11] implicitly assumes
that reliability r depends on coherence c.
More generally, the parameters of the drift diffusion model of Eq. (1) and that of the probabilistic
model can be expressed as functions of each other [13]:
v = ± 2
δt2 σ2 = ±r
2
δt2
4
(7)
s =
2σ
δtσ2 = r
2σ
δt
.
(8)
These equations state that both drift v and diffusion s depend on the assumed reliability r of the
measurements x. Does the brain use and necessarily compute this reliability which depends on
coherence? In the following section we answer this question by comparing how well three variants
of the drift diffusion model, that implement different assumptions about r, conform to experimental
findings.
4 Use of reliability in perceptual decision making: experimental evidence
We first show that different assumptions about the reliability r translate to variants of the drift dif-
fusion model. We then fit all variants to behavioural data (performances and mean reaction times)
of an experiment for which neurophysiological data has also been reported [11] and demonstrate
that only those variants which allow reliability to depend on coherence level lead to accumulation
mechanisms which are consistent with the neurophysiological findings.
4.1 Drift diffusion model variants
For the drift diffusion model of Eq. (1) the accuracy A and mean decision time T predicted by the
model can be determined analytically [9]:
A = 1 −
1 + exp( 2vb
s2 )
1
(cid:18) vb
(cid:19)
s2
T =
b
v
tanh
(9)
(10)
where b is the bound. These equations highlight an important caveat of the drift diffusion model:
Only two of the three parameters can be determined uniquely from behavioural data. For fitting
the model one of the parameters needs to be fixed. In most cases, the diffusion s is set to c = 0.1
arbitrarily [9], or is fit with a constant value across stimulus strengths [11]. We call this standard
variant of the drift diffusion model the DDM.
If s is constant across stimulus strengths, the other two parameters of the model must explain dif-
ferences in behaviour, between stimulus strengths, by taking on values that depend on stimulus
strength. Indeed, it has been found that primarily drift v explains such differences, see also be-
low. Eq. (7) states that drift depends on estimated reliability r. So, if drift varies across stimulus
strengths, this strongly suggests that r must vary across stimulus strengths, i.e., that r must depend
on coherence: r(c). However, the drift diffusion formalism allows for two other obvious variants
of parameterisation. One in which the bound b is constant across stimulus strengths, b = ¯b, and,
conversely, one in which drift v is constant across stimulus strengths, v = ¯v ∝ ¯r (Eq. 7). We call
these variants DEPC and CONST, respectively, for their property to weight evidence by reliability
that either depends on coherence, r(c), or not, ¯r.
4.2 Experimental data
In the following we will analyse the data presented in [11]. This data set has two major advantages
for our purposes: 1) Reported accuracies and mean reaction times (Fig. 1d,f) are averages based on
15,937 trials in total. Therefore, noise in this data set is minimal (cf. small error bars in Fig. 1d,f)
such that any potential effects of overfitting on found parameter values will be small, especially in
relation to the effect induced by different stimulus strengths. 2) The behavioural data is accompanied
by recordings of neurons which have been implicated in the decision making process. We can,
therefore, compare the accumulation mechanisms resulting from the fit to behaviour with the actual
neurophysiological recordings. Furthermore, the structure of the experiments was such that the
stimulus in subsequent trials had random strength, i.e., the brain could not have estimated stimulus
strength of a trial before the trial started.
In the experiment of [11], that we consider here, two monkeys performed a two-alternative forced
choice task based on the RDM stimulus. Data for eight different coherences were reported. To avoid
5
ceiling effects, which prevent the unique identification of parameter values in the drift diffusion
model, we exclude those coherences which lead to an accuracy of 0.5 (random choices) or to an
accuracy of 1 (perfect choices). The behavioural data of the remaining six coherence levels are
presented in Table 1.
Table 1: Behavioural data of [11] used in our analysis. RT = reaction time.
coherence (%):
accuracy (fraction):
mean RT (ms):
3.2
0.63
613
6.4
0.76
590
9
0.79
580
12
0.89
535
25.6
0.99
440
The analysis of [11] revealed a nondecision time, i.e., a component of the reaction time that is
unrelated to the decision process (cf. [3]) of ca. 200ms. Using this estimate, we determined the
mean decision time T by subtracting 200ms from the mean reaction times shown in Table 1.
The main findings for the neural recordings, which replicated previous findings [19, 1], were that i)
firing rates at the end of decisions were similar and, particularly, showed no significant relation to
coherence [11, Fig. 5] whereas ii) the buildup rate of neural firing within a trial had an approximately
linear relation to coherence [11, Fig. 4].
4.3 Fits of drift diffusion model variants to behaviour
We can easily fit the model variants (DDM, DEPC and CONST) to accuracy A and mean decision
time T using Eqs. (9) and (10). In accordance with previous approaches we selected values for the
respective redundant parameters. Since the redundant parameter value, or its inverse, simply scales
the fitted parameter values (cf. Eqs. 9 and 10), the exact value is irrelevant and we fix, in each model
variant, the redundant parameter to 1.
Figure 2: Fitting results: values of the free parameters, that replicate the accuracy and mean RT
recorded in the experiment (Table 1), in relation to coherence. The remaining, non-free parameter
was fixed to 1 for each variant. Left: the DDM variant with free parameters drift v (green) and
bound b (purple). Middle: the DEPC variant with free parameters v and diffusion s (orange). Right:
the CONST variant with free parameters s and b.
Fig. 2 shows the inferred parameter values. In congruence with previous findings, the DDM variant
explained variation in behaviour due to an increasing coherence mostly with an increasing drift v
(green in Fig. 2). Specifically, drift and coherence appear to have a straightforward, linear relation.
The same finding holds for the DEPC variant. In contrast to the DDM variant, however, which also
exhibited a slight increase in the bound b (purple in Fig. 2) with increasing coherence, the DEPC
variant explained the corresponding differences in behaviour by decreasing diffusion s (orange in
Fig. 2). As the drift v was fixed in CONST, this variant explained coherence-dependent behaviour
with large and almost identical changes in both diffusion s and bound b such that large parameter
values occurred for small coherences and the relation between parameters and coherence appeared
to be quadratic.
We further investigated the properties of the model variants with the fitted parameter values. The top
row of Fig. 3 shows example drift diffusion trajectories (y in Eq. (1)) simulated at a resolution of
1ms for two coherences. Following [11], we interpret y as the decision variables represented by the
6
051015202530coherence (%)0510152025bDDM051015202530coherence (%)0.000.010.020.030.040.05sDEPC051015202530coherence (%)01020304050607080sCONST0.000.020.040.060.080.10v0.0000.0010.0020.0030.004v02004006008001000120014001600b Figure 3: Drift-diffusion properties of fitted model variants. Top row: 15 example trajectories of y
for different model variants with fitted parameters for 6.4% (blue) and 25.6% (yellow) coherence.
Trajectories end when they reach the bound for the first time which corresponds to the decision
time in that simulated trial. Notice that the same random samples of were used across variants
and coherences. Bottom row: Trajectories of y averaged over trials in which the first alternative (top
bound) was chosen for the three model variants. Format of the plots follows that of [8, Supp. Fig. 4]:
Left panels show the buildup of y from the start of decision making for the 5 different coherences.
Right panels show the averaged drift diffusion trajectories when aligned to the time that a decision
was made.
firing rates of neurons in monkey area LIP. These plots exemplify that the DDM and DEPC variants
lead to qualitatively very similar predictions of neural responses whereas the trajectories produced
by the CONST variant stand out, because the neural responses to large coherences are predicted to
be smaller than those to small coherences.
We have summarised predicted neural responses to all coherences in the bottom row of Fig. 3 where
we show averages of y across 5000 trials either aligned to the start of decision making (left pan-
els) or aligned to the decision time (right panels). These plots illustrate that the DDM and DEPC
variants replicate the main neurophysiological findings of [11]: Neural responses at the end of the
decision were similar and independent of coherence. For the DEPC variant this was built into the
model, because the bound was fixed. For the DDM variant the bound shows a small dependence
on coherence, but the neural responses aligned to decision time were still very similar across coher-
ences. The DDM and DEPC variants, further, replicate the finding that the buildup of neural firing
depends approximately linear on coherence (normalised mean square error of a corresponding linear
model was 0.04 and 0.03, respectively). In contrast, the CONST variant exhibited an inverse rela-
tion between coherence and buildup of predicted neural response, i.e., buildup was larger for small
coherences. Furthermore, neural responses at decision time strongly depended on coherence. There-
fore, the CONST variant, as the only variant which does not use coherence-dependent reliability, is
also the only variant which is clearly inconsistent with the neurophysiological findings.
5 Discussion
We have investigated whether the brain uses online estimates of stimulus reliability when making
simple perceptual decisions. From a probabilistic perspective fundamental considerations suggest
that using accurate estimates of stimulus reliability lead to better decisions, but in the field of percep-
tual decision making it has been questioned that the brain estimates stimulus reliability on the very
short time scale of a few hundred milliseconds. By using a probabilistic formulation of the most
7
0100200300400500600700800time from start (ms)201001020yDDM 6.425.60100200300time from start (ms)05101520mean of y3.26.49.012.025.6-300-200-100DTtime from end (ms)0100200300400500600700800time from start (ms)1.00.50.00.51.0DEPC0100200300time from start (ms)0.00.20.40.60.81.0-300-200-100DTtime from end (ms)0100200300400500600700800time from start (ms)6004002000200400600CONST0100200300time from start (ms)0200400600800100012001400-300-200-100DTtime from end (ms) widely accepted model we were able to show that only those variants of the model which assume
online reliability estimation are consistent with reported experimental findings.
Our argument is based on a strict distinction between measurements, evidence and likelihood which
may be briefly summarised as follows: Measurements are raw stimulus features that do not relate to
the decision, evidence is a transformation of measurements into a decision relevant space reflecting
the decision alternatives and likelihood is evidence scaled by a current estimate of measurement
reliabilities. It is easy to overlook this distinction at the level of bounded accumulation models,
such as the drift diffusion model, because these models assume a pre-computed form of evidence as
input. However, this evidence has to be computed by the brain, as we have demonstrated based on
the example of the RDM stimulus and using behavioural data.
We chose one particular, simple probabilistic model, because this model has a direct equivalence
with the drift diffusion model which was used to explain the data of [11] before. Other models may
have not allowed conclusions about reliability estimates in the brain. In particular, [13] introduced
an alternative model that also leads to equivalence with the drift diffusion model, but explains dif-
ferences in behaviour by different mean measurements and their representations in the generative
model. Instead of varying reliability across coherences, this model would vary the difference of
means in the second summand of Eq. (5) directly without leading to any difference on the drift
diffusion trajectories represented by y of Eq. (1) when compared to those of the probabilistic model
chosen here. The interpretation of the alternative model of [13], however, is far removed from basic
assumptions about the RDM stimulus: Whereas the alternative model assumes that the reliability of
the stimulus is fixed across coherences, the noise in the RDM stimulus clearly depends on coherence.
We, therefore, discarded the alternative model here.
As a slight caveat, the neurophysiological findings, on which we based our conclusion, could have
been the result of a search for neurons that exhibit the properties of the conventional drift diffusion
model (the DDM variant). We cannot exclude this possibility completely, but given the wide range
and persistence of consistent evidence for the standard bounded accumulation theory of decision
making [1, 20] we find it rather unlikely that the results in [19] and [11] were purely found by
chance. Even if our conclusion about the rapid estimation of reliability by the brain does not en-
dure, our formal contribution holds: We clarified that the drift diffusion model in its most common
variant (DDM) is consistent with, and even implicitly relies on, coherence-dependent estimates of
measurement reliability.
In the experiment of [11] coherences of the RDM stimulus were chosen randomly for each trial.
Consequently, participants could not predict the reliability of the RDM stimulus for the upcoming
trial, i.e., the participants' brains could not have had a good estimate of stimulus reliability at the
start of a trial. Yet, our analysis strongly suggests that coherence-dependent reliabilities were used
during decision making. The brain, therefore, must had adapted reliability within trials even on the
short timescale of a few hundred milliseconds. On the level of analysis dictated by the drift diffusion
model we cannot observe this adaptation. It only manifests itself as a change in mean drift that is
assumed to be constant within a trial. First models of simultaneous decision making and reliability
estimation have been suggested [21], but clearly more work in this direction is needed to elucidate
the underlying mechanism used by the brain.
References
[1] Joshua I Gold and Michael N Shadlen. The neural basis of decision making. Annu Rev Neu-
rosci, 30:535 -- 574, 2007.
[2] I. D. John. A statistical decision theory of simple reaction time. Australian Journal of Psy-
chology, 19(1):27 -- 34, 1967.
[3] R. Duncan Luce. Response Times: Their Role in Inferring Elementary Mental Organization.
Number 8 in Oxford Psychology Series. Oxford University Press, 1986.
[4] Abraham Wald. Sequential Analysis. Wiley, New York, 1947.
[5] Xiao-Jing Wang. Probabilistic decision making by slow reverberation in cortical circuits. Neu-
ron, 36(5):955 -- 968, Dec 2002.
[6] Rajesh P N Rao. Bayesian computation in recurrent neural circuits. Neural Comput, 16(1):1 --
38, Jan 2004.
8
[7] Jeffrey M Beck, Wei Ji Ma, Roozbeh Kiani, Tim Hanks, Anne K Churchland, Jamie Roitman,
Michael N Shadlen, Peter E Latham, and Alexandre Pouget. Probabilistic population codes for
Bayesian decision making. Neuron, 60(6):1142 -- 1152, December 2008.
[8] Anne K Churchland, R. Kiani, R. Chaudhuri, Xiao-Jing Wang, Alexandre Pouget, and M. N.
Shadlen. Variance as a signature of neural computations during decision making. Neuron,
69(4):818 -- 831, Feb 2011.
[9] Rafal Bogacz, Eric Brown, Jeff Moehlis, Philip Holmes, and Jonathan D. Cohen. The physics
of optimal decision making: a formal analysis of models of performance in two-alternative
forced-choice tasks. Psychol Rev, 113(4):700 -- 765, October 2006.
[10] Michael N Shadlen, Roozbeh Kiani, Timothy D Hanks, and Anne K Churchland. Neurobi-
ology of decision making: An intentional framework. In Christoph Engel and Wolf Singer,
editors, Better Than Conscious? Decision Making, the Humand Mind, and Implications For
Institutions. MIT Press, 2008.
[11] Anne K Churchland, Roozbeh Kiani, and Michael N Shadlen. Decision-making with multiple
alternatives. Nat Neurosci, 11(6):693 -- 702, Jun 2008.
[12] Peter Dayan and Nathaniel D Daw. Decision theory, reinforcement learning, and the brain.
Cogn Affect Behav Neurosci, 8(4):429 -- 453, Dec 2008.
[13] Sebastian Bitzer, Hame Park, Felix Blankenburg, and Stefan J Kiebel. Perceptual decision
making: Drift-diffusion model is equivalent to a bayesian model. Frontiers in Human Neuro-
science, 8(102), 2014.
[14] W. T. Newsome and E. B. Par´e. A selective impairment of motion perception following lesions
of the middle temporal visual area MT. J Neurosci, 8(6):2201 -- 2211, June 1988.
[15] Praveen K. Pilly and Aaron R. Seitz. What a difference a parameter makes: a psychophysical
comparison of random dot motion algorithms. Vision Res, 49(13):1599 -- 1612, Jun 2009.
[16] Angela J. Yu and Peter Dayan.
Inference, attention, and decision in a Bayesian neural ar-
chitecture. In Lawrence K. Saul, Yair Weiss, and L´eon Bottou, editors, Advances in Neural
Information Processing Systems 17, pages 1577 -- 1584. MIT Press, Cambridge, MA, 2005.
[17] Alec Solway and Matthew M. Botvinick. Goal-directed decision making as probabilistic in-
ference: a computational framework and potential neural correlates. Psychol Rev, 119(1):120 --
154, January 2012.
[18] Yanping Huang, Abram Friesen, Timothy Hanks, Mike Shadlen, and Rajesh Rao. How prior
probability influences decision making: A unifying probabilistic model. In P. Bartlett, F.C.N.
Pereira, C.J.C. Burges, L. Bottou, and K.Q. Weinberger, editors, Advances in Neural Informa-
tion Processing Systems 25, pages 1277 -- 1285. 2012.
[19] Jamie D Roitman and Michael N Shadlen. Response of neurons in the lateral intraparietal area
during a combined visual discrimination reaction time task. J Neurosci, 22(21):9475 -- 9489,
Nov 2002.
[20] Timothy D. Hanks, Charles D. Kopec, Bingni W. Brunton, Chunyu A. Duan, Jeffrey C. Erlich,
and Carlos D. Brody. Distinct relationships of parietal and prefrontal cortices to evidence
accumulation. Nature, Jan 2015.
[21] Sophie Den`eve. Making decisions with unknown sensory reliability. Front Neurosci, 6:75,
2012.
9
|
1909.02931 | 1 | 1909 | 2019-09-06T14:31:54 | Integrated Information in the Spiking-Bursting Stochastic Model | [
"q-bio.NC"
] | This study presents a comprehensive analytic description in terms of the empirical "whole minus sum" version of Integrated Information in comparison to the "decoder based" version for the "spiking-bursting" discrete-time, discrete-state stochastic model, which was recently introduced to describe a specific type of dynamics in a neuron-astrocyte network. The "whole minus sum" information may change sign, and an interpretation of this transition in terms of "net synergy" is available in the literature. This motivates our particular interest to the sign of the "whole minus sum" information in our analytical consideration. The behavior of the "whole minus sum" and "decoder based" information measures are found to bear a lot of similarity, showing their mutual asymptotic convergence as time-uncorrelated activity is increased, with the sign transition of the "whole minus sum" information associated to a rapid growth in the "decoder based" information. The study aims at creating a theoretical base for using the spiking-bursting model as a well understood reference point for applying Integrated Information concepts to systems exhibiting similar bursting behavior (in particular, to neuron-astrocyte networks). The model can also be of interest as a new discrete-state test bench for different formulations of Integrated Information. | q-bio.NC | q-bio |
Integrated Information in the Spiking-Bursting Stochastic Model
Oleg Kanakov,1 Susanna Gordleeva,1 and Alexey Zaikin1, 2, 3, ∗
1Lobachevsky State University of Nizhny Novgorod, Nizhny Novgorod, Russia
2Institute for Women's Health and Department of Mathematics,
3Department of Pediatrics, Faculty of Pediatrics, Sechenov University, Moscow, Russia
University College London, London, United Kingdom
This study presents a comprehensive analytic description in terms of the empirical "whole mi-
nus sum" version of Integrated Information in comparison to the "decoder based" version for the
"spiking-bursting" discrete-time, discrete-state stochastic model, which was recently introduced to
describe a specific type of dynamics in a neuron-astrocyte network. The "whole minus sum" in-
formation may change sign, and an interpretation of this transition in terms of "net synergy" is
available in the literature. This motivates our particular interest to the sign of the "whole minus
sum" information in our analytical consideration. The behavior of the "whole minus sum" and
"decoder based" information measures are found to bear a lot of similarity, showing their mutual
asymptotic convergence as time-uncorrelated activity is increased, with the sign transition of the
"whole minus sum" information associated to a rapid growth in the "decoder based" information.
The study aims at creating a theoretical base for using the spiking-bursting model as a well un-
derstood reference point for applying Integrated Information concepts to systems exhibiting similar
bursting behavior (in particular, to neuron-astrocyte networks). The model can also be of interest
as a new discrete-state test bench for different formulations of Integrated Information.
I.
INTRODUCTION
Integrated information (II) [1 -- 4] is a measure of inter-
nal information exchange in complex systems which was
initially proposed to quantify consciousness [5]. This ini-
tial aim still remaining a matter of research and debate
[6 -- 9], the II concept itself is by now a widely acknowl-
edged tool in the field of complex dynamics analysis [10 --
12]. The general concept gave rise to specific "empirical"
formalizations of II [13 -- 16] aimed at computability from
empirical probability distributions based on real data.
For a systematic taxonomy of II measures see [17], and
a comparative study of empirical II measures in applica-
tion to Gaussian autoregressive network models has been
recently done in [18].
A recent study [19] addressed the role of astrocytic reg-
ulation of neurotransmission [20] in generating positive
II by small networks of brain cells -- neurons and astro-
cytes. Empirical "whole minus sum" II as defined in [13]
was calculated in [19] from the time series produced by a
biologically realistic model of neuro-astrocytic networks.
A simplified, analytically tractable stochastic "spiking-
bursting" model (in complement to the realistic one) was
designed to describe a specific type of activity in neuro-
astrocytic networks which manifests itself as a sequence
of intermittent system-wide excitations of rapid pulse
trains ("bursts") on the background of random "spik-
ing" activity in the network. The spiking-bursting model
is a discrete-time, discrete-state stochastic process which
mimics the main features of this behavior. The model
was successfully used in [19] to produce semi-analytical
estimates of II in good agreement with direct compu-
∗ E-mail: [email protected]
tation of II from time series of the biologically realistic
network model.
The present study aims at creating a theoretical base
for using the spiking-bursting model of [19] as a well un-
derstood reference point for applying Integrated Infor-
mation concepts to systems exhibiting similar bursting
behavior (in particular, to other neuron-astrocyte net-
works). We also aim at extending the knowledge of com-
parative features of different empirical II measures, which
are currently available mainly in application to Gaussian
autoregressive models [17, 18], by applying two such mea-
sures [13, 16] to our discrete-state model.
In Sections II, III we specify the definitons of the used
II measures and the model. Specific properties of the
model which lead to a redundance in its parameter set are
addressed in Section IV. In Section V we provide with an
analytical treatment for the empirical "whole minus sum"
[13] version of II in application to our model. This choice
among other empirical II measures is inherited from the
preceding study [19] and is in part due to its easy analyt-
ical tractability, and also due to its ability to change sign,
which naturally identifies a transition point in the param-
eter space. This property may be considered a violation
of the natural non-negativeness requirement for II [16];
on the other hand, the sign of the "whole minus sum" in-
formation has been given interpretation in terms of "net
synergy" [21] as a degree of redundancy in the evolution
of a system [18].
In this sense this transition may be
viewed as a useful marker in its own right in the toolset
of measures for complex dynamics. This motivates our
particular focus on identifying the sign transition of the
"whole minus sum" information in the parameter space
of the model. We also identify a scaling of II with a
parameter determining time correlations of the bursting
(astrocytic) subsystem when these correlations are weak.
In Section VI we compare the outcome of the "whole
minus sum" II measure [13] to the "decoder based" mea-
sure Φ∗, which was specifically designed in [16] to sat-
isfy the non-negativeness property. We compute Φ∗ di-
rectly by definition from known probability distributions
of the model. Despite their inherent difference consisting
in changing or not changing sign, the two compared mea-
sures are shown to bear similarities in their dependence
upon model parameters, including the same scaling with
the time correlation parameter.
II. DEFINITION OF II MEASURES IN USE
The empirical "whole minus sum" version of II is for-
mulated according to [13] as follows. Consider a station-
ary stochastic process ξ(t) (binary vector process), whose
instantaneous state is described by N binary digits (bits),
each identified with a node of the network (neuron). The
full set of N nodes ("system") can be split into two non-
overlapping non-empty subsets ("subsystems") A and B,
such splitting further referred to as bipartition AB. De-
note by x = ξ(t) and y = ξ(t+τ ) two states of the process
separated by a specified time interval τ 6= 0. States of
the subsystems are denoted as xA, xB, yA, yB.
Mutual information between x and y is defined as
where
Ixy = Hx + Hy − Hxy,
Hx = −Xx
p(x) log2 p(x)
(1)
(2)
is entropy (base 2 logarithm gives result in bits), summa-
tion is hereinafter assumed to be taken over the whole
range of the index variable (here x), Hy = Hx due to
assumed stationarity.
Next, a bipartition AB is considered, and "effective
information" as a function of the particular bipartition
is defined as
Φeff(AB) = Ixy − IxA,yA − IxB ,yB .
(3)
II is then defined as effective information calculated
for a specific bipartition ABMIB ("minimum information
bipartition") which minimizes specifically normalized ef-
fective information:
Φ = Φeff(ABMIB),
ABMIB = argminAB(cid:20)
Φeff(AB)
min{H(xA), H(xB)}(cid:21) .
(4a)
(4b)
Note that this definition prohibits positive II, whenever
Φeff turns out to be zero or negative for at least one
bipartition AB.
We compare the result of the "whole minus sum" effec-
tive information (3) to the "decoder based" information
measure Φ∗, which is modified from its original formula-
tion of [16] by setting the logarithms base to 2 for con-
sistency:
Φ∗(AB) = Ixy − I ∗
xy(AB),
(5a)
where
2
I ∗
xy(AB) = max
β "−Xy
+Xxy
p(y) log2Xx
p(x)qAB(yx)β
p(xy) log2 qAB(yx)β# ,
(5b)
qAB(yx) = p(yAxA)p(yBxB) =
p(xAyA)p(xByB)
p(xA)p(xB)
.
(5c)
III. SPIKING-BURSTING STOCHASTIC
MODEL
We consider a stochastic model, which produces a
binary vector valued, discrete-time stochastic process.
In keeping with [19], this "spiking-bursting" model is
defined as a combination M = {V, S} of a time-
correlated dichotomous component V which turns on and
off system-wide bursting (that mimics global bursting of
a neuronal network, when each neuron produces a train
of pulses at a high rate [19]), and a time-uncorrelated
component S describing spontaneous (spiking) activity
(corresponding to a background random activity in a
neural network characterized by relatively sparse ran-
dom appearance of neuronal pulses -- spikes [19]) oc-
curring in the absence of a burst. The model mimics
the spiking-bursting type of activity which occurs in a
neuro-astrocytic network, where the neural subsystem
normally exhibits time-uncorrelated patterns of spiking
activity, and all neurons are under the common influ-
ence of the astrocytic subsystem, which is modeled by
the dichotomous component V and sporadically induces
simultaneous bursting in all neurons. A similar network
architecture with a "master node" spreading its influence
on subordinated nodes was considered for example in [1]
(Figure 4b therein).
The model is defined as follows. At each instance of
(discrete) time the state of the dichotomous component
can be either "bursting" with probability pb, or "sponta-
neous" (or "spiking") with probability ps = 1−pb. While
in the bursting mode, the instantaneous state of the re-
sulting process x = ξ(t) is given by all ones: x = 11..1
(further abbreviated as x = 1). In case of spiking, the
state x is a (time-uncorrelated) random variate described
by a discrete probability distribution sx (where an occur-
rence of '1' in any bit is referred to as a "spike"), so that
the resulting one-time state probabilities read
p(x 6= 1) = pssx,
p(x = 1) = p1,
p1 = pss1 + pb,
(6a)
(6b)
where s1 is the probability of spontaneous occurrence of
x = 1 (hereafter referred to as a system-wide simultane-
ous [22] spike) in the absence of a burst.
To describe two-time joint probabilities for x = ξ(t)
and y = ξ(t + τ ), consider a joint state xy which is a con-
catenation of bits in x and y. The spontaneous activity
is assumed to be uncorrelated in time, which leads to the
factorization
sxy = sxsy.
(7)
The time correlations of the dichotomous component [23]
are described by a 2 × 2 matrix
pq∈{s,b},r∈{s,b} =(cid:18)pss psb
pbs pbb(cid:19)
(8)
whose components are joint probabilities to observe the
respective spiking (index "s") and/or bursting (index
"b") states in x and y. The probabilities obey psb = pbs
(due to stationarity), pb = pbb + psb, ps = pss + psb,
thereby allowing to express all one- and two-time prob-
abilities describing the dichotomous component in terms
of two independent quantities, which for example can be
a pair {ps, pss}, then
psb = pbs = ps − pss,
pbb = 1 − (pss + 2psb),
(9a)
(9b)
or {pb, ρ} as in [19], where ρ is correlation coefficient
defined by
psb = pspb(1 − ρ).
(10)
In Section IV we justify the use of another effective pa-
rameter ǫ (13) instead of ρ to determine time correlations
in the dichotomous component.
The two-time joint probabilities for the resulting pro-
cess are then expressed as
p(x 6= 1, y 6= 1) = psssxsy,
p(x 6= 1, y = 1) = πsx,
p(x = 1, y = 1) = p11,
π = psss1 + psb,
p(x = 1, y 6= 1) = πsy,
p11 = psss2
1 + 2psbs1 + pbb.
(11a)
(11b)
(11c)
(11d)
Note that the above notations can be applied to any
subsystem instead of the whole system (with the same
dichotomous component, as it is system-wide anyway).
3
hand and bursts on the other, because the latter are as-
sumed to be correlated in time, unlike the former. That
said, the "labeling" of bursts versus system-wide spikes
exists in the model (by the state of the dichotomous com-
ponent), but not in the realizations. Proceeding from the
realizations, it must be possible to relabel a certain frac-
tion of system-wide spikes into bursts (more precisely,
into a time-uncorrelated portion thereof). Such relabel-
ing would change both components of the model {V, S}
(dichotomous and spiking processes), in particular dilut-
ing the time correlations of bursts, without changing the
actual realizations of the resultant process. This implies
the existence of a transformation of model parameters
which keeps realizations (i.e. the stochastic process as
such) invariant. The derivation of this transformation
is presented in Appendix A and leads to the following
scaling
sx = αs′
x,
1 − s1 = α(1 − s′
1),
ps′ = αps,
ps′s′ = α2pss,
(12a)
(12b)
(12c)
(12d)
where α is a positive scaling parameter, and all other
probabilities are updated according to Eq. (9).
The mentioned invariance in particular implies that
any characteristic of the process must be invariant to
the scaling (12a-d). This suggests a natural choice of a
scaling-invariant effective parameter ǫ defined by
pss = p2
s(1 + ǫ)
(13)
to determine time correlations in the dichotomous com-
ponent. In conjunction with a second independent pa-
rameter of the dichotomous process, for which a straight-
forward choice is ps, and with full one-time probability
table for spontaneous activity sx, these constitute a nat-
ural full set of model parameters {sx, ps, ǫ}.
The two-time probability table (8) can be expressed in
terms of ps and ǫ by substituting Eq. (13) into Eq. (9):
pq∈{s,b},r∈{s,b} =(cid:18) p2
s + ǫp2
pspb − ǫp2
s
s (cid:19) .
s pspb − ǫp2
p2
b + ǫp2
s
(14)
The requirement of non-negativeness of probabilities im-
poses simultaneous constraints
IV. MODEL PARAMETERS SCALING
and
ǫ ≥ −1
(15a)
The spiking-bursting stochastic model as described in
Section III is redundant in the following sense. In terms
of the model definition, there are two distinct states of the
model which equally lead to observing the same one-time
state of the resultant process with 1's in all bits: firstly
-- a burst, and secondly -- a system-wide simultaneous
spike in the absence of a burst, which are indistinguish-
able by one-time observations. Two-time observations
reveal a difference between system-wide spikes on one
ps ≤ ps max =( 1
1+ǫ(cid:16)1 −pǫ(cid:17) if − 1 ≤ ǫ < 0,
if
1
1+ǫ
ǫ ≥ 0,
or, equivalently,
− ǫ2
max ≤ ǫ ≤ ǫmax =
pb
ps
.
(15b)
(16)
Comparing the off-diagonal term psb in (14) to the def-
inition of correlation coefficient ρ in (10), we get
ǫ = ρ
pb
ps
= ρ ǫmax,
(17)
thus the sign of ǫ has the same meaning as that of ρ.
Hereinafter we limit ourselves to non-negative correla-
tions ǫ ≥ 0.
V. ANALYSIS OF THE EMPIRICAL "WHOLE
MINUS SUM" MEASURE FOR THE
SPIKING-BURSTING PROCESS
In this Section we analyze the behavior of the "whole
minus sum" empirical II [13] defined by Eqs. (3), (4) for
the spiking-bursting model in dependence of the model
parameters, particularly focusing on its transition from
negative to positive values.
Mutual information Ixy for two time instances x and
y of the spiking-bursting process is expressed by insert-
ing all one- and two-time probabilities of the process ac-
cording to (6), (11) into the definition (1), (2). The full
derivation is given in Appendix B and leads to an expres-
sion which was used in [19]
Ixy = 2(1 − s1){ps} + 2{p1} − (1 − s1)2{pss}
− 2(1 − s1){π} − {p11},
where we denote for compactness
{q} = −q log2 q.
(18)
(19)
We exclude from further consideration the following
degenerate cases which automatically give Ixy = 0 by
definition (1):
s1 = 1,
or ps = 0,
or
ps = 1,
or ρ = ǫ = 0,
(20)
where the former two correspond to a deterministic "al-
ways 1" state for which all entropies in (1) are zero, and
the latter two produce no predictability, which implies
Hxy = Hx + Hy.
The particular case s1 = 0 in (18) reduces to
Ixys1=0 = 2(cid:0){ps} + {pb}(cid:1) −(cid:0){pss} + 2{psb} + {pbb}(cid:1)
= I0(ps, ǫ),
(21)
which coincides with mutual information for the dichoto-
mous component taken alone and can be seen as a func-
tion denoted in (21) as I0(·,·) of just two independent
parameters of the dichotomous component, for which we
chose ps and ǫ as described in Section IV. Typical plots
of I0(ps, ǫ) versus ps at fixed ǫ are shown with blue solid
lines in Fig. 1.
The general case (18) can be recovered back from (21)
1 = 0
by virtue of the scaling (12a-d), by assuming s′
4
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1
FIG. 1. Blue solid lines -- plots of I0(ps, ǫ) versus ps varied
from 0 to ps max as per (15), at ǫ = 0.01, 0.1, 0.2, 0.5, 1 (from
right to left). Red dashed lines -- approximation (35). Red
dots -- upper bounds of approximation applicability range
(36c).
in (12b) and substituting the corresponding scaled value
ps′ = (1−s1)ps as per (12c) in place of the first argument
of function I0(ps′ , ǫ) defined in (21), while parameter ǫ re-
mains invariant to the scaling. This produces a simplified
expression
Ixy = I0(cid:0)(1 − s1)ps, ǫ(cid:1),
(22)
which is still exactly equivalent to (18). We emphasize
that hereinafter expressions containing I0(·,·) like (22),
(23), (30b) etc. imply that all probabilities in (21) must
be expressed in terms of ps and ǫ, and ps in turn be
accordingly substituted by the actual first argument of
I0(·,·), e.g. by (1− s1)ps in (22). The same applies when
the approximate expression for I0(·) (35) is used.
Given a bipartition AB (see Section II), this result
is applicable as well to any subsystem A (B), with s1
replaced by sA (sB) which denote the probability of a
subsystem-wide simultaneous spike xA = 1 (xB = 1)
in the absence of a burst, and with same parameters of
the dichotomous component (here ps, ǫ). Then effective
information (3) is expressed as
Φeff = I0(cid:0)(1−s1)ps, ǫ(cid:1)−I0(cid:0)(1−sA)ps, ǫ(cid:1)−I0(cid:0)(1−sB)ps, ǫ(cid:1).
(23)
Hereafter in this section we assume the independence
of spontaneous activity across the system, which implies
sAsB = s1,
Φeff = f (sA),
then (23) turns into
where
f (s) = I0(cid:0)(1 − s1)ps, ǫ(cid:1) − I0(cid:0)(1 − s)ps, ǫ(cid:1)
− I0(cid:0)(1 − s1/s)ps, ǫ(cid:1).
(24)
(25a)
(25b)
Note that the function I0(·,·) in (21) is defined only when
the first argument is in the range (0, 1), thus the defini-
tion domain of f (s) in (25b) is
s1 < s < 1.
(26)
According to (4), the necessary and sufficient condition
for the "whole minus sum" empirical II be positive is the
requirement that Φeff be positive for any bipartition AB.
Due to (25), this requirement can be written in the form
min
s∈{sA}
f (s) > 0,
(27)
where {sA} is the set of sA values for all possible bipar-
titions AB (if A is any non-empty subsystem, then sA is
defined as the probability of spontaneous occurrence of
1's in all bits in A in the same instance of the discrete
time).
Expanding the set of s in (27) to the whole defini-
tion domain of f (s) (26) leads to a sufficient (generally,
stronger) condition for positive II
min
s1<s<1
f (s) > 0.
(28)
Note [24] that f (s) by definition (25b) satisfies f (s =
s1) = f (s = 1) = 0, f ′(s = s1) > 0 and (due to the
invariance to mutual renaming of subsystems A and B)
f (s1/s) = f (s). The latter symmetry implies that the
quantity of extrema on (s1, 1) must be odd, one of them
always being at s = √s1.
If the latter is the only ex-
tremum, then it is a positive maximum, and (28) is thus
fulfilled automatically. In case of three extrema, f (√s1)
is a minimum, which can change sign. In both these cases
the condition (28) is equivalent to the requirement
f (√s1) > 0,
which can be rewritten as
g(s1) > 0,
where
(29)
(30a)
g(s1) = f (√s1) = I0(cid:0)(1 − s1)ps, ǫ(cid:1) − 2I0(cid:0)(1 − √s1)ps, ǫ(cid:1).
(30b)
The equivalence of (29) to (28) would be broken in case
of 5 or more extrema. As suggested by numerical evi-
dence [25], this exception never holds, although we did
not prove this rigorously. Based on the reasoning above,
in the following we assume the equivalence of (29) (and
(30)) to (28).
A typical scenario of transformations of f (s) with the
change of s1 is shown in Fig. 2. Here the extremum
f (√s1) (shown with a dot) transforms with the decrease
of s1 from a positive maximum into a minimum, which in
turn decreases from positive through zero into negative
values.
Note that by construction, the function g(s1) defined
in (30b) expresses effective information Φeff from (3) for
5
0.006
0.004
0.002
0
−0.002
−0.004
−0.006
0
0.2
0.4
0.6
0.8
1
FIG. 2. Plots of f (s) on s1 < s < 1 with s1 = 0.03, 0.025,
0.0182919, 0.01 (from top to bottom) at ps = 0.7, ǫ = 0.1.
For each value of s1, the extremum (√s1, f (√s1)) is indicated
with a dot.
a specific bipartition characterized by sA = sB = √s1. If
such "symmetric" bipartition exists, then the value √s1
belongs to the set {sA} in (27), which implies that (29)
(same as (30)) is equivalent not only to (28), but also to
the necessary and sufficient condition (27). Otherwise,
(28) (equivalently, (29) or (30)), formally being only suf-
ficient, still may produce a good estimate of the neces-
sary and sufficient condition in cases when {sA} contains
values which are close to √s1 (corresponding to nearly
symmetric partitions, if such exist).
Except for the degenerate cases (20), g(s1) is negative
at s1 = 0
g(s1 = 0) = −I0(ps, ǫ) < 0
and tends to +0 [26] with s1 → 1 − 0 as soon as
(31)
(32)
lim
s1→1−0
I0(cid:0)(1 − s1)ps, ǫ(cid:1)
2I0(cid:0)(1 − √s1)ps, ǫ(cid:1) = 2,
hence g(s1) changes sign at least once on s1 ∈ (0, 1). Ac-
cording to numerical evidence [27], we assume that g(s1)
changes sign exactly once on (0, 1) without providing a
rigorous proof for the latter statement (note however that
for the asymptotic case (38) this statement is rigorous).
In line with the above, the solution to (30a) has the form
smin
1
(ps, ǫ) < s1 < 1,
(33)
1
where smin
(ps, ǫ) is the unique root of g(s1) on (0, 1).
Several plots of smin
(ps, ǫ) versus ps at ǫ fixed and versus
ǫ at ps fixed, which are obtained by numerically solving
for the zero of g(s1), are shown in Fig. 3 with blue solid
lines.
1
Further insight into the dependence of mutual infor-
mation Ixy (and, consequently, of Φeff and II) upon pa-
rameters can be obtained by inserting the expressions
6
0.15
0.1
0.05
0
0
0.2
0.4
0.6
0.8
1
(a)
(b)
0.05
0.04
0.03
0.02
0.01
0
0
0.2
0.4
0.6
0.8
1
FIG. 3. (a) Blue solid lines -- plots of smin
(ps, ǫ) versus ps varied from 0 to ps max as per (15), at ǫ = 0.1, 0.5, 1 (from right to
left). Red dashed line -- plot of the asymptotic formula (38). (b) Blue solid lines -- plots of smin
(ps, ǫ) versus ǫ varied from
0 to ǫmax as per (16), at ps = 0.5, 0.6, 0.7 (from top to bottom). Vertical position of red dashed lines is the result of (38),
horizontal span denotes the estimated applicability range (36b).
1
1
for the two-time probabilities (14) into the definition of
I0(ps, ǫ) in (21) and expanding it in powers of ǫ (weak
time correlation limit), which yields
I0(ps, ǫ) =
1
2 log 2(cid:18) ps
1 − ps(cid:19)2
· ǫ2 + O(ǫ3).
(34)
Estimating the residual term (see details in Appendix C)
indicates that the approximation by the leading term
I0(ps, ǫ) ≈
is valid when
ǫ2
2 log 2(cid:18) ps
1 − ps(cid:19)2
ǫ ≪ 1,
ps(cid:19)2
ǫ ≪(cid:18) pb
= ǫ2
max .
(35)
(36a)
(36b)
Solving (36b) for ps rewrites it in the form of an upper
bound [28] for ps
ps <
.
(36c)
1
1 +pǫ
Note how inequalities (36b), (36c) compare to the formal
upper bounds ǫmax in (16) and ps max in (15) which arise
from the definition of ǫ (13) due to the requirement of
positive probabilities.
which do not depend on ǫ at all, this scaling of Φeff does
not change the minimum information bipartition, finally
implying that II also scales as ǫ2. That said, as factor ǫ2
does not affect the sign of Φeff, the lower bound smin
in
(33) exists and is determined only by ps in this limit.
1
Substituting the approximation (35) for I0(·) into the
definition of g(s1) in (30b) after simplifications reduces
the equation g(s1) = 0 to the following [29]:
ps(√2 − 1) s1 − √s1 + (1 − ps)(√2 − 1) = 0 ,
(37)
whose solution in terms of s1 on 0 < s1 < 1 equals
smin
, according to the reasoning behind Eq. (33). Solving
1
(37) as a quadratic equation in terms of √s1 produces a
unique root on (0, 1), which yields
smin
1
(ps)ǫ→0 =
1 −q1 − 4ps(1 − ps)(√2 − 1)2
2ps(√2 − 1)
2
.
(38)
Result of (38) is plotted in Fig. 3 with red dashed lines:
in panel (a) as a function of ps, and in panel (b) as hori-
zontal lines whose vertical position is the result of (38),
and horizontal span denotes the estimated applicability
range (36b) (note that condition (36a) also applies, and
becomes stronger than (36b) when ps < 1/2).
Approximation (35) is plotted in Fig. 1 with red dashed
lines along with corresponding upper bounds of approx-
imation applicability range (36c) denoted by red dots
(note that large ǫ violates (36a) anyway, thus in this case
(36c) has no effect). Mutual information (35) scales with
ǫ within range (36) as ǫ2 and vanishes with ǫ → 0. The
same holds for effective information (23). Since the nor-
malizing denominator in (4b) contains one-time entropies
VI. COMPARISON OF INTEGRATED
INFORMATION MEASURES
In this Section we compare the outcome of two versions
of empirical Integrated Information measures available
in the literature, one being the "all-minus-sum" effec-
tive information (3) from [13] which is used elsewhere in
this study, and the other "decoder based" information as
introduced in [16] and expressed by Eqs. (5a-c). We cal-
culate both measures by their respective definitions using
the one- and two-time probabilities from Eqs. (6a,b) and
(11a-d) for the spiking-bursting model with N = 6 bits,
assuming no spatial correlations among bits in spiking
activity, with same spike probability P in each bit. In
this case
sx = P m(x)(1 − P )N −m(x), P = s
1
N
1 ,
(39)
where m(x) is the number of ones in the binary word x.
We consider only a symmetric bipartition with subsys-
tems A and B consisting of N/2 = 3 bits each. Due
to the assumed equal spike probabilities in all bits and
in the absence of spatial correlations of spiking, this im-
plies complete equivalence between the subsystems. In
particular, in the notations of Sec. V we get
sA = sB = √s1.
s1 = sAsB,
(40)
This choice of the bipartition is firstly due to the fact
that the sign of effective information for this bipartition
determines the sign [30] of the resultant "whole minus
sum" II. This has been established in Sec. V (see reason-
ing behind Eqs. (27) -- (30) and further on); moreover, this
effective information has been denoted in Eq. (30) as a
function
Φeff(AB) = g(s1),
(41)
which has been analyzed in Sec. V.
Moreover, the choice of the symmetric bipartition is
consistent with available comparative studies of II mea-
sures [18], where it was substantiated by the conceptual
requirement that highly asymmetric partitions should be
excluded [2], and by the lack of a generally accepted spec-
ification of minimum information bipartition; for further
discussion, see [18].
We have studied the dependence of the mentioned ef-
fective information measures upon spiking activity, which
is controlled by s1, at different fixed values of the parame-
ters ps and ǫ specifying the bursting component. Typical
dependence of both measures upon s1, taken at ps = 0.6
with several values of ǫ, is shown in Fig. 4, panel (a).
The behavior of the "whole minus sum" effective infor-
mation Φeff (41) (blue lines in Fig. 4) is found to agree
with the analytical findings of Sec. V:
1
• Φeff transitions from negative values to positive at
a certain threshold value of s1 = smin
, which is
well approximated by the formula (38) when ǫ is
small, as required by (36a,b); the result of Eq. (38)
is indicated in each panel of Fig. 4 by an additional
vertical grid line labeled smin
on the abscissae axis,
cf. Fig. 3;
• Φeff reaches a maximum on the interval smin
1 and tends to zero (from above) at s1 → 1;
• Φeff scales with ǫ as ǫ2, when (36a,b) hold.
1 < s1 <
1
7
To verify the scaling observation, we plot the scaled
values of both information measures Φeff/ǫ2, Φ∗/ǫ2 in
the panels (b) -- (d) of Fig. 4 for several fixed values of ps
and ǫ. Expectedly, the scaling fails at ps = 0.7, ǫ = 0.4
in panel (d), as (36b) is not fulfilled in this case.
Furthermore, the "decoder based" information Φ∗
(plotted with red lines in Fig. 4) behaves mostly the same
way, apart from being always non-negative (which was
one of key motivations for introducing this measure in
[16]). At the same time, the sign transition point smin
of the "whole minus sum" information associates with a
rapid growth of the "decoder based" information. When
s1 is increased towards 1, the two measures converge. Re-
markably, the scaling as ǫ2 is found to be shared by both
effective information measures.
1
VII. DISCUSSION
In general, the spiking-bursting model is completely
specified by the combination of a full single-time proba-
bility table sx (consisting of 2N probabilities of all pos-
sible outcomes, where N is the number of bits) for the
time-uncorrelated spontaneous activity, along with two
independent parameters (e.g. ps and ǫ) for the dichoto-
mous component. This combination is, however, redun-
dant in that it admits a one-parameter scaling (12) which
leaves the resultant stochastic process invariant.
Condition (30) was derived assuming that spiking ac-
tivity in individual bits (i.e. nodes, or neurons) consti-
tuting the system is independent among the bits, which
implies that the probability table sx is fully determined
by N spike probabilities for individual nodes. The con-
dition is formulated in terms of ps, ǫ and a single pa-
rameter s1 (system-wide spike probability) for the spon-
taneous activity, thus ignoring the "internal structure"
of the system, i.e. the spike probabilities for individual
nodes. This condition provides that the "whole minus
sum" effective information is positive for any bipartition,
regardless of the mentioned internal structure. More-
over, in the limit (36) of weak correlations in time, the
inequality (30a) can be explicitly solved in terms of s1,
producing the solution (33), (38).
In this way, the inequality (33) together with the
asymptotic estimate (38) supplemented by its applicabil-
ity range (36) specifies the region in the parameter space
of the system, where the "whole minus sum" II is posi-
tive regardless of the internal system structure (sufficient
condition). The internal structure (though still without
spike correlations across the system) is taken into account
by the necessary and sufficient condition (27) for positive
II.
The mentioned conditions were derived under the as-
sumption of absent correlation between spontaneous ac-
tivity in individual bits (24). If correlation exists and is
positive, then s1 > sAsB, or sB < s1/sA. Then com-
paring the expressions for Φeff (23) (general case) to (25)
(space-uncorrelated case), and taking into account that
(a)
0.1
0.08
0.06
0.04
0.02
0
0
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
(b)
0.1
0.2
0.3
0.4
8
0.2
0.15
0.1
0.05
0
0
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.1
0.2
0.3
0.4
(c)
(d)
FIG. 4. Comparison of two versions of empirical effective information for the symmetric bipartition -- "whole-minus-sum"
measure (3) from [13] (blue lines) and "decoder based" information (5) from [16] (red lines) versus spiking activity parameter
s1 at various fixed values of the bursting component parameters ps (indicated on top of the panels) and ǫ (indicated in the
legends). Panel (a) -- unnormalized values, panels (b) -- (d) -- normalized by ǫ2. Threshold smin
calculated according to (38)
is shown in each panel with an additional vertical grid line.
1
I0(ps) is an increasing function, we find Φeff < f (sA),
cf. (25a). This implies that any necessary condition for
positive II remains as such. Likewise, in the case of neg-
ative correlations we get Φeff > f (sA), implying that a
sufficient condition remains as such.
We found that II scales as ǫ2 for ǫ small (namely, within
(36)) when other parameters (i.e. ps and spiking prob-
ability table sx) are fixed. For the "whole minus sum"
information, this is an analytical result. Note that the
reasoning behind this result does not rely upon the as-
sumption of spatial uncorrelation of spiking activity (be-
tween bits) and thus applies to arbitrary spiking-bursting
systems. According to a numerical calculation, this scal-
ing is applicable to the "decoder based" information as
well.
Remarkably, II can not exceed the time delayed mutual
information for the system as a whole, which in case of
the spiking-bursting model in its present formulation is
no greater than 1 bit.
The present study substantiates, refines and quantifies
1
1
qualitative observations in regard to II in the spiking-
bursting model which were initially made in [19]. The
existence of lower bounds in spiking activity (character-
ized by s1) which was noticed in [19] is now expressed in
the form of an explicit inequality (33) with the estimate
(38) for the bound smin
. The observation of [19] that
typically smin
is mostly determined by burst probability
and weakly depends upon time correlations of bursts also
becomes supported by the quantitative result (33), (38).
The model provides a basis for possible modifications
in order to apply Integrated Information concepts to sys-
tems exhibiting similar, but more complicated behavior
(in particular, to neuron-astrocyte networks). Such mod-
ifications might incorporate non-trivial spatial patterns
in bursting, and causal interactions within and between
the spiking and bursting subsystems.
The model can also be of interest as a new discrete-
state test bench for different formalizations of Integrated
Information, while available comparative studies of II
measures mainly focus on Gaussian autoregressive mod-
els [17, 18].
ACKNOWLEDGMENTS
This research was funded by the Ministry of Education
and Science of the Russian Federation within Agreement
No. 075-15-2019-871.
Appendix A: Derivation of parameters scaling of the
spiking-bursting model
In order to formalize the reasoning in Section IV, we
introduce an auxiliary 3-state process W with set of one-
time states {s′, d, b}, where s′ and b are always inter-
preted as spiking and bursting states in terms of Sec-
tion III, and d is another state, which is assumed to
produce all bits equal 1 like in a burst, but in a time-
uncorrelated manner (which is formalized by Eq. (A4)
below) like in a system-wide spike. When W is properly
defined (by specifying all necessary probabilities, see be-
low) and supplemented with a time-uncorrelated process
S as a source of spontaneous activity for the state s′,
these together constitute a completely defined stochastic
model {W, S}.
This 3-state based model may be mapped on equivalent
(in terms of resultant realizations) 2-state based models
as in Section III in an ambiguous way, because the state d
may be equally interpreted either as a system-wide spike,
or as a time-uncorrelated burst, thus producing two dif-
ferent dichotomous processes (which we denote as V and
V ′) for the equivalent spiking-bursting models. The rela-
tionship between the states of W , V and V ′ is illustrated
by the following diagram.
V = s
b
(A1)
As soon as d-states of W are interpreted in V as (spik-
ing) s-states, the spontaneous activity process S accom-
panying V has to be supplemented with system-wide
spikes whenever W = d, in addition to the spontaneous
activity process S′ for V ′. In order to maintain the ab-
sence of time correlations in spontaneous activity (which
is essential for the analysis in Section V), we assume time-
uncorrelated choice between W = s′ and W = d when
V = s (which manifests below in Eq. (A4)). Then the
difference between the spontaneous components S and S′
comes down to a difference in the corresponding one-time
probability tables sx and s′
x.
In the following, we proceed from the dichotomous pro-
cess V defined as in Section III, then define a consistent
3-state process W , and further obtain another dichoto-
mous process V ′ for an equivalent model. Finally, we
establish the relation between the corresponding proba-
bility tables of spontaneous activity sx and s′
x.
W = s′ d b
V ′ = s′
z}{
{z}
b′
9
The first dichotomous process V has states denoted by
{s, b} and is related to W according to the rule V = s
when W = s′ or W = d, and V = b whenever W = b (see
diagram (A1)). Assume fixed conditional probabilities
p(W = s′ V = s) = α,
p(W = d V = s) = β = 1 − α,
which implies one-time probabilities for W as
(A2a)
(A2b)
ps′ = αps,
pd = βps.
(A3)
The mentioned requirement of
time-uncorrelated
choice between W = s′ and W = d when V = s is ex-
pressed by factorized two-time conditional probabilities
p(W = s′s′ V = ss) = α2,
p(W = s′d V = ss) = αβ = p(W = ds′ V = ss),
p(W = dd V = ss) = β2.
(A4a)
(A4b)
(A4c)
Given the two-time probability table for V (8) along with
the conditional probabilities (A2), (A4), we arrive at a
two-time probability table for W
pq∈{s′,d,b},r∈{s′,d,b} =
s′
d
b
s′ α2pss αβpss αpsb
d αβpss β2pss βpsb
pbb
b αpbs βpbs
.
!
(A5)
Note that (A5) is consistent both with (A3), which is ob-
tained by summation along the rows of (A5), and with
(8), which is obtained by summation within the line-
separated cell groups in (A5):
pss ≡ ps′s′ + ps′d + pds′ + pdd
psb ≡ ps′b + pdb
pbs ≡ pbs′ + pbd
pbb ≡ pbb.
(A6a)
(A6b)
(A6c)
(A6d)
Consider the other dichotomous process V ′ with states
{s′, b′} obtained from W according to the rule V ′ = b′
when W = d or W = b, and V ′ = s′ whenever W = s′
(see diagram (A1)). The two-time probability table for
V ′ is obtained by another partitioning of the table (A5)
s′
d
b
pqr =
s′ α2pss αβpss αpsb
d αβpss β2pss βpsb
b αpbs
pbb
βpbs
!
(A7)
with subsequent summation of cells within groups, which
yields
ps′s′ = α2pss,
ps′b′ = α(βpss + psb) = pb′s′,
pb′b′ = β2pss + 2βpsb + pbb.
(A8a)
(A8b)
(A8c)
The corresponding one-time probabilities for V ′ read
ps′ = αps,
pb′ = βps + pb.
(A9a)
(A9b)
In order to establish the relation between the one-time
probability tables of spontaneous activity sx and s′
x, we
equate the resultant one-time probabilities of observing
a given state x as per (6) for the two equivalent models
{V, S} and {V ′, S′}
p(x 6= 1) = pssx = ps′ s′
x,
p(x = 1) = pss1 + pb = ps′ s′
1 + pb′ .
(A10a)
(A10b)
Taking into account (A9), we finally get
sx = αs′
x,
1 − s1 = α(1 − s′
1).
(A11a)
(A11b)
Equations (A8), (A9) and (A11) fully describe the
transformation of the spiking-bursting model which keeps
the resultant stochastic process invariant by the construc-
tion of the transform. Taking into account that the di-
chotomous process is fully described by just two inde-
pendent quantities, e.g. ps and pss, all other probabili-
ties being expressed in terms of these due to normaliza-
tion and stationarity, the full invariant transformation is
uniquely identified by a combination of (A11a,b), (A8a)
and (A9a), which together constitute the scaling (12).
Note that parameter α within its initial meaning (A2)
may take on values in the range 0 < α ≤ 1 (case α = 1
producing the identical transform). That said, in terms
10
of the scaling (12a-d), all values α > 0 are equally
possible, so that mutually inverse values α = α1 and
α = α2 = 1/α1 produce mutually inverse transforms.
Appendix B: Expressing mutual information for the
spiking-bursting process
One-time entropy Hx for the spiking-bursting process
is expressed by (2) with probabilities p(x) taken from (6):
Hx =Xx
{p(x)} =Xx
{pssx} + {p1} − {pss1},
(B1)
where the additional terms besides the sum over x ac-
count for the specific expression (6b) for p(x = 1). Using
the relation
{ab} ≡ a{b} + {a}b,
(B2)
which is derived directly from (19), and collecting similar
terms, we arrive at
Hx = psHs − ps{s1} + (1 − s1){ps} + {p1},
(B3)
where Hs is the entropy of the spiking component taken
alone
Hs =Xx
{sx}.
(B4)
Two-time entropy is expressed similarly, by substitut-
ing probabilities p(xy) from (11) into the definition of
entropy and taking into account the special cases with
x = 1 and/or y = 1:
Hxy =Xxy
{p(xy)} =Xxy
{psssxsy} −Xx
−Xy
{psssxs1} +Xx
{psss1sy} +Xy
{πsx}
{πsy} + {psss2
1} − 2{πs1} + {p11}.
(B5)
Further, applying (B2) and using the notation (B4), we
find
sxsy
= pss · 2Hs + {pss},
{sxsy} + {pss}Xxy
Xxy
{psssxsy} = pssXxy
where we used the reasoning thatPxy{sxsy} is the two-
time entropy of the spiking component taken alone, which
is (due to the postulated absence of time correlations
in it) twice the one-time entropy Hs (this of course can
equally be found by direct calculation). Similarly, we get
(B6a)
also
and exactly the same expression for Py{psss1sy}, and
Xy
{sx} + {π}Xx
{πsx} = πXx
{πsy} =Xx
(B6c)
sx
= πHs + {π}.
Substituting (B6a-c) into (B5), using (B2) where ap-
plicable, and collecting similar terms with the relation
pss + π − psss1 ≡ ps
taken into account, we arrive at
Xx
{psssxs1} = psss1Xx
{sx} + {psss1}Xx
= psss1Hs + {psss1}
sx
(B6b)
Hxy = 2psHs + (1 − s1)2{pss} − 2ps{s1}
+ 2(1 − s1){π} + {p11}.
(B7)
(B8)
Finally, the expression (18) for mutual information is
obtained by inserting (B3) and (B8) into the definition
(1), with stationarity Hy = Hx taken into account.
and in the case ξ < 0 to
Appendix C: Expanding I0 in powers of ǫ
where
Taylor series expansion for a function f (x) up to the
quadratic term reads
1
3
Φ(cid:18)ξ
x0(cid:19) ≪ 1,
Φ(ζ) =
ζ
(1 − ζ)2 .
11
(C8a)
(C8b)
f (x0 + ξ) = f (x0) + f ′(x0)ξ + f ′′(x0)
ξ2
2
+ R(ξ).
(C1)
Replacing Φ(·) in (C8a) by its linearization Φ(ζ) ≈ ζ
for small ζ, we reduce both (C7) and (C8a) to a single
condition
The remainder term R(ξ) can be represented in the La-
grange's form as
ξ ≪ 3x0.
(C9)
We use these considerations to expand in powers of ǫ
the function I0(ps, ǫ) defined in (21) with pss, psb, pbb
substituted by their expressions in terms of ǫ according
to (14). We note that the braces notation {·} defined in
(19) is expressed via the function f (x) from (C4) as
(C3a)
Expanding this way the subexpressions of (21)
R(ξ) = f ′′′(c)
ξ3
6
,
(C2)
where c is an unknown real quantity between x0 and
x0 + ξ.
The function f (x) can be approximated by omit-
ting R(ξ) in (C1) if R(ξ) is negligible compared to the
quadratic term, for which it is sufficient that
for any c between x0 and x0 + ξ, namely for
f ′′(x0)
ξ2
2 (cid:12)(cid:12)(cid:12)(cid:12)
ξ3
f ′′′(c)
(cid:12)(cid:12)(cid:12)(cid:12)
6 (cid:12)(cid:12)(cid:12)(cid:12) ≪(cid:12)(cid:12)(cid:12)(cid:12)
c ∈((x0, x0 + ξ),
(x0 − ξ, x0),
if ξ > 0,
if ξ < 0.
(C3b)
Consider the specific case
f (x) = −x log x,
x > 0,
(C4)
for which we get
,
1
x
f ′′′(x) =
f ′′(x) = −
1
f ′(x) = − log x − 1,
x2 .
(C5)
As long as f ′′′(x) is a falling function for any x > 0,
fulfilling (C3a) at the left boundary of (C3b) (at c = x0
if ξ > 0, and at c = x0 − ξ if ξ < 0) makes sure (C3a)
is fulfilled in the whole interval (C3b). Precisely, the
requirement is
1
x2
0
(cid:12)(cid:12)(cid:12)(cid:12)
1
(x0 − ξ)2
(cid:12)(cid:12)(cid:12)(cid:12)
1
x0
1
x0
ξ3
ξ3
6 (cid:12)(cid:12)(cid:12)(cid:12) ≪(cid:12)(cid:12)(cid:12)(cid:12)
6 (cid:12)(cid:12)(cid:12)(cid:12) ≪(cid:12)(cid:12)(cid:12)(cid:12)
ξ2
ξ2
2 (cid:12)(cid:12)(cid:12)(cid:12)
2 (cid:12)(cid:12)(cid:12)(cid:12)
which in the case ξ > 0 reduces to
,
,
if
ξ > 0,
(C6a)
if
ξ < 0,
(C6b)
ξ
3x0 ≪ 1,
(C7)
{q} =
f (q)
log 2
.
s + ǫp2
{pss} = {p2
s},
{psb} = {pspb − ǫp2
s},
b + ǫp2
{pbb} = {p2
s},
(C10)
(C11a)
(C11b)
(C11c)
we find by immediate calculation that the zero-order and
linear in ǫ terms vanish, and the quadratic term yields
(35). The condition (C9) has to be applied to all three
subexpressions (C11a-c). Omitting the insignificant fac-
tor 3 in (C9), we obtain the applicability conditions
ǫp2
ǫp2
ǫp2
s ≪ p2
s,
s ≪ pspb,
s ≪ p2
b,
which is equivalent to
ǫ ≪ 1,
pb
ǫ ≪
ps
ǫ ≪ ǫ2
max,
= ǫmax,
(C12a)
(C12b)
(C12c)
(C13a)
(C13b)
(C13c)
where the notation ǫmax from (16) is used. We note that
when ǫmax < 1, the condition (C13c) is the strongest
among (C13a-c); when ǫmax > 1, the condition (C13a)
is the strongest. Therefore, in both cases (C13b) can be
dropped, thus producing (36).
[1] G. Tononi, BMC Neurosci. 5, 42 (2004).
[2] D. Balduzzi and G. Tononi, PLoS Comput. Biol. 4,
[17] M. Tegmark, PLoS Comput. Biol. 12, e1005123 (2016).
[18] P. Mediano, A. Seth, and A. Barrett, Entropy 21, 17
e1000091 (2008).
(2019).
[3] G. Tononi, Archives italiennes de biologie 150, 293
[19] O. Kanakov, S. Gordleeva, A. Ermolaeva, S. Jalan, and
12
(2012).
[4] M. Oizumi, L. Albantakis, and G. Tononi, PLoS Com-
put. Biol. 10, e1003588 (2014).
[5] G. Tononi, Biol. Bull. 215, 216 (2008).
[6] A. Peressini, J. Conscious. Stud. 20, 180 (2013).
[7] N. Tsuchiya, S. Taguchi, and H. Saigo, Neurosci. Res.
107, 1 (2016).
[8] G. Tononi, M. Boly, M. Massimini, and C. Koch, Nat.
Rev. Neurosci. 17, 450 (2016).
[9] R. Norman and A. Tamulis, J. Comput. Theor. Nanosci.
14, 2255 (2017).
[10] D. Engel and T. W. Malone, PLoS ONE 13, e0205335
(2018).
[11] P. A. M. Mediano, J. C. Farah,
arXiv
han,
arXiv:1606.08313 [q-bio.NC].
e-prints
,
arXiv:1606.08313
and M. Shana-
(2016),
[12] D. Toker and F. T. Sommer, PLoS Comput. Biol. 15,
e1006807 (2019).
[13] A. B. Barrett and A. K. Seth, PLoS Comput. Biol. 7,
e1001052 (2011).
[14] V. Griffith, arXiv e-prints , arXiv:1401.0978 (2014),
arXiv:1401.0978 [cs.IT].
[15] M. Oizumi, N. Tsuchiya, and S.-i. Amari, Proc. Natl.
Acad. Sci. USA 113, 14817 (2016).
[16] M. Oizumi, S.-i. Amari, T. Yanagawa, N. Fujii,
and
N. Tsuchiya, PLoS Comput. Biol. 12, e1004654 (2016).
A. Zaikin, Phys. Rev. E 99, 012418 (2019).
[20] A. Araque, G. Carmignoto, P. G. Haydon, S. H. Oliet,
R. Robitaille, and A. Volterra, Neuron 81, 728 (2014).
[21] A. B. Barrett, Phys. Rev. E 91, 052802 (2015).
[22] In a real network "simultaneous" implies occuring within
the same time discretization interval [19].
[23] In a neural network these correlations are conditioned by
burst duration [19]; e.g., if this (in general, random) du-
ration mostly exceeds τ , then the correlation is positive.
[24] All mentioned properties and subsequent reasoning can
be observed in Fig. 2, which shows a few sample plots of
f (s).
[25] The equivalence of (28) to (30) was confirmed up to ma-
chine precision for each combination of ps ∈ [0.01, 0.99]
and ρ ∈ [0.01, 1] (both with step 0.01).
[26] Notations −0 and +0 denote the left ang right one-sided
limits.
[27] This statement was confirmed up to machine precision
for each combination of ps ∈ [0.01, 0.99] and ρ ∈ [0.01, 1]
(both with step 0.01).
[28] The use of '≪' sign is not appropriate in (36c), because
this inequality does not imply a small ratio between its
left-hand and right-hand parts.
[29] See comment below Eq. (22).
[30] Although the actual value of II is determined by the min-
imal information bipartition which may be different.
|
1502.07907 | 1 | 1502 | 2015-02-27T14:22:14 | Closed Form Jitter Analysis of Neuronal Spike Trains | [
"q-bio.NC"
] | Interval jitter and spike resampling methods are used to analyze the time scale on which temporal correlations occur. They allow the computation of jitter corrected cross correlograms and the performance of an associated statistically robust hypothesis test to decide whether observed correlations at a given time scale are significant. Currently used Monte Carlo methods approximate the probability distribution of coincidences. They require generating $N_{\rm MC}$ simulated spike trains of length $T$ and calculating their correlation with another spike train up to lag $\tau_{\max}$. This is computationally costly $O(N_{\rm MC} \times T \times \tau_{\max})$ and it introduces errors in estimating the $p$ value. Instead, we propose to compute the distribution in closed form, with a complexity of $O(C_{\max} \log(C_{\max}) \tau_{\max})$, where $C_{\max}$ is the maximum possible number of coincidences. All results are then exact rather than approximate, and as a consequence, the $p$-values obtained are the theoretically best possible for the available data and test statistic. In addition, simulations with realistic parameters predict a speed increase over Monte Carlo methods of two orders of magnitude for hypothesis testing, and four orders of magnitude for computing the full jitter-corrected cross correlogram. | q-bio.NC | q-bio | Closed Form Jitter Analysis
of Neuronal Spike Trains
Daniel M. Jeck1,2 and Ernst Niebur1,3
1Zanvyl Krieger Mind/Brain Institute, Johns Hopkins
University, Baltimore, MD
2Department of Biomedical Engineering, Johns Hopkins
3Solomon Snyder Department of Neuroscience, Johns Hopkins
University, Baltimore, MD
University, Baltimore, MD
July 16, 2018
5
1
0
2
b
e
F
7
2
]
.
C
N
o
i
b
-
q
[
1
v
7
0
9
7
0
.
2
0
5
1
:
v
i
X
r
a
1
Closed Form Jitter Analysis
Running Title
Closed Form Jitter Analysis
Address correspondence to
Daniel Jeck
The Johns Hopkins University
Krieger Mind/Brain Institute
3400 N. Charles Street
Baltimore, MD 21218
(email: [email protected])
Keywords
Jitter, Spike timing, Synchrony, Cross Correlation
Abstract
Interval jitter and spike resampling methods are used to analyze
the time scale on which temporal correlations occur. They allow
the computation of jitter corrected cross correlograms and the per-
formance of an associated statistically robust hypothesis test to de-
cide whether observed correlations at a given time scale are signifi-
cant. Currently used Monte Carlo methods approximate the proba-
bility distribution of coincidences. They require generating NMC sim-
ulated spike trains of length T and calculating their correlation with
another spike train up to lag τmax. This is computationally costly
O(NMC × T × τmax) and it introduces errors in estimating the p value.
Instead, we propose to compute the distribution in closed form, with
a complexity of O(Cmax log(Cmax)τmax), where Cmax is the maximum
possible number of coincidences. All results are then exact rather
than approximate, and as a consequence, the p-values obtained are
the theoretically best possible for the available data and test statis-
tic. In addition, simulations with realistic parameters predict a speed
increase over Monte Carlo methods of two orders of magnitude for
hypothesis testing, and four orders of magnitude for computing the
full jitter-corrected cross correlogram.
2
Closed Form Jitter Analysis
1
Introduction
Though there are many means by which neurons communicate, using both
chemical and electrical mechanisms, most attention has been paid to series
of action potentials (spikes).
It is known that in some cases the detailed
time structure of these spike trains is used for information transmission
while in others, only the overall number of spikes in some interval seems
to be important but not their position in the interval. Examples of the first
type are various kinds of "temporal coding" schemes proposed for different
neural system and for different functional roles, e.g.
refs Abeles (1991);
Singer and Gray (1995); Riehle et al. (1997); Niebur et al. (1993); Softky
(1995); Steinmetz et al. (2000), while the latter is the well-known rate-code
mechanism, e.g. refs. Adrian and Zotterman (1926); Shadlen and Newsome
(1998). To distinguish between these two possibilities, it is necessary to find
whether reproducible correlations at the relevant time scale are present in
neuronal data. One common way to approach this problem is to use auto-
or cross-correlation functions as test statistics. Then one can (a) search for
non-trivial structure in the function, like deviations from uniformity, and
(b) detect whether there are differences between these functions in different
experimental (e.g. behavioral) conditions.
The situation is complicated by the influence of rate variations on the raw
correlations. To increase the signal-to-noise ratio, correlations are typically
computed as averages over many trials. Changes in the behavioral state
of an animal, e.g. due to onset of sensory stimuli or motor responses that
occur always at the same time during a trial, typically result in changes in
neural firing rates which are common to many neurons. While these are
genuine correlations, they are unrelated to the neuronal coding question.
Different techniques have been developed to remove them, e.g. subtraction
of a "shuffle predictor" Perkel et al. (1967), the average of cross correlations
between spike trains from permuted trials1. While this correction removes
correlations that are time-locked to trial onset, it was later pointed out that
peaks in the correlation function that may be taken as indicative of correlated
firing (e.g. at zero lag) can also be caused by slow rate covariations Brody
(1998, 1999). After finding a significant peak in the cross correlation function,
this ambiguity can be addressed by analyzing the time scale at which the
1A "shift predictor" is very similar but the correlation function is computed from trials
that immediately follow each other, rather than from randomly selected trials.
3
Closed Form Jitter Analysis
measured correlation arises.
It was pointed out more recently Amarasingham et al. (2012) that the
null hypothesis of spike trains being independent in earlier work Perkel et al.
(1967) is useless if their mean rates co-vary. Rate co-variation of means the
spike trains are not independent which leads to its immediate rejection of the
null without providing any further insight. Amarasingham and his colleagues
instead proposed a more detailed null hypothesis, namely that within an in-
terval of width ∆, the exact location of spikes does not matter. Then, under
the null hypothesis, simulated spike trains can be generated by modifying
the spike times of an original measured spike train within a range of ∆. The
cross correlations obtained from these modified ("jittered") spike trains are
then compared to those obtained from the original. If significant differences
are found, the null hypothesis is rejected and it is likely that non-random
correlations at time scales ≤ ∆ are present in the data. Additionally, this
method gives rise to the computation of jitter-corrected cross correlograms,
which have been used to compare changes in synchrony across experimen-
tal conditions (Hirabayashi et al., 2013a,b; Smith et al., 2013). Because the
method relies on repeated simulation of spike trains, it will be referred to as
the Monte Carlo jitter method for the purposes of this paper.
While the Monte Carlo jitter method (described fully in Section 2.1) is
useful and easily generalized to complex statistical tests and hypotheses,
its practical utility is limited by the inherent trade-off between accuracy and
computation time in all Monte Carlo methods. As we will show in Section 3.2,
the computation time may be prohibitively long, and even at this cost, it will
only be a numerical approximation of the true solution. In the case where
the test statistic is the cross-correllation value at a single lag, the p value
can be computed exactly, as was shown by Harrison (2013). In the present
study, we therefore explore the benefits of computing in closed form the
distribution which is only approximated by the Monte Carlo simulations.
Accordingly we refer to this method, described in Section 2.2, as the closed
form jitter method. In addition to computing the p value for rejecting the
null hypothesis exactly we show that the computation of the jitter-corrected
cross correlogram follows readily from that derivation. The computational
performance of the closed form jitter and Monte Carlo jitter methods are
compared theoretically (as computational complexity) in Section 3.1 and
practically (as computational time) in Section 3.2.
4
Closed Form Jitter Analysis
Length of binned spike trains
Two spike trains
Correlation Lag
Correlation of X and Y
Jitter interval width
Index over Monte Carlo simulated signals
Index over jitter intervals
Number of spikes in X in interval j
ith Monte Carlo simulated signal
Number of Monte Carlo simulated signals
Correlation of X MC
Number of cases where C MC
p Value for correlation at lag τ
T
X, Y
τ
C(τ )
∆
i
j
N(X, j)
X MC
NMC
C MC
i
Rτ
pτ
JCCG(τ ) Jitter-corrected Cross Correlogram
Pτ (C MC) Monte Carlo estimate of the distribution of correlations at lag τ
C int
Pτ (C(τ )) True distribution of correlations at lag τ
τmax
Nmax
Cmax
Maximum τ value processed
Maximum value of N(X, j) or N(Y, j)
Maximum possible number of coincidences
Number of coincidences in the jth interval
(τ ) > C(τ )
i
and Y
i
i
j
Table 1: Glossary. Variables are listed in the order in which they are
introduced.
5
Closed Form Jitter Analysis
2 Materials and Methods
2.1 The Monte Carlo Jitter Method
Utilizing the Monte Carlo jitter method (Amarasingham et al., 2012), it is
possible to determine whether correlations arise from fine temporal structure
or larger scale variations, sometimes referred to as rate covariations. This
determination is made by comparing a test statistic (in this case cross cor-
relations) of an original pair of spike trains against those computed from a
set of jittered spike trains as described below. The jitter method, like cross
correlation, operates on binned spike trains which we take as binary signals
with values 0 and 1 and integer arguments 0 to T − 1, where T is the number
of bins in the spike train. The binary assumption implies that the bin size is
small enough (typically 1ms or so) such that two spikes cannot be recorded
in a single time bin. A sufficiently small bin size can always be chosen since
there are limits on the minimal inter-spike interval time due to the refractory
period of the neurons in question.
Let X(t) and Y (t) be two such binned spike trains. The processing then
consists of the following steps:
1. Compute the cross correlation C(τ ) between the original X and Y ,
C(τ ) =Xt
X(t − τ )Y (t)
where the sum runs from 0 to T − 1 and X is assumed to be 0 if its
argument is outside that range.
2. Subdivide one of the signals, say X, into intervals of width ∆.
3. Count the number of spikes in each interval of X. In interval j this is,
N(X, j) =
(j+1)∆−1
Xk=j∆
X(k)
(1)
4. For X generate NMC Monte Carlo simulated signals {X MC
}, in which
the spike counts for each interval are the same as in the corresponding
interval in X, such that N(X MC
, j) = N(X, j) for all i, j. However,
now spike times within the interval are all equally likely. Spike times
should be sampled without replacement to ensure that the spike count
stays constant without putting multiple spikes in a single bin.
i
i
6
Closed Form Jitter Analysis
5. Compute the cross correlation C MC
i
(τ ) for lag time τ between each
and the second spike train Y to get an estimate of the distribution
X MC
Pτ (C MC) of cross correlation values for each time lag τ .
i
C MC
i
(τ ) =Xt
X MC
i
(t − τ )Y (t)
6. Let Rτ be the number of simulations where C MC
i
(τ ) ≥ C(τ ). Then the
p value for a given lag τ is computed as
pτ =
Rτ + 1
NMC + 1
7. If desired, a jitter-corrected cross correlogram (JCCG), defined using
the expectation operator E[·], can be computed as
JCCG(τ ) = C(τ ) − E[CMC(τ )] ≈ C(τ ) −
1
NMC Xi
CMC
i
(τ )
(2)
where the approximation approaches equality with NM C → ∞.
Jittering the spikes within an interval of size ∆ destroys all correlations
at time scales within this interval. The cross correlations computed from the
jittered spike trains therefore are not correlated on time scales ∆ or smaller,
and Pτ (C MC) is the distribution of correlations at time lag τ obtained under
the null hypothesis that correlations at time scales ≤ ∆ are indistinguishable
from random correlations. If the measured cross correlation C(τ ) is signif-
icantly outside this distribution, we have to reject the null hypothesis and
we conclude that nonrandom correlations at lag τ with time scales ≤ ∆ are
found in the observed spike trains. If, on the other hand, the observed cor-
relation is consistent with what is seen in the distribution of jittered spike
trains, then we cannot reject the null hypothesis. This means we cannot
exclude that the observed synchrony is caused by correlations on time scales
outside the jittered range, in other words that the observed correlation at lag
time τ is caused by rate variations on time scales greater than ∆.
In practice, X(t) and Y (t) do not have to be gathered from a continuous
block of time. In the case of multiple trials of the same experimental con-
dition, it may be useful to concatenate the recorded spike trains (possibly
7
Closed Form Jitter Analysis
after removing sections of them, like those recorded during stimulus onsets).
In doing so, a period of no spiking of width τmax (the largest correlation lag
of interest) should be added between the trials so that correlations between
trials don't affect the outcome of the jitter procedure.
As mentioned in Section 1, the practical utility of the Monte Carlo method
is limited by the trade-off between accuracy and computation time inherent
in all Monte Carlo algorithms. Furthermore, in practice a single set of Monte
Carlo simulations is often generated for many hypothesis tests (i.e. tests at
multiple lags), introducing potential dependencies between the different tests
when they should be treated independently 2. In order to avoid both of these
issues, the probability distribution Pτ (C MC) can be computed exactly and
independently for each time lag as described in the following.
2.2 Closed Form Computation
2.2.1 Probability Distribution For One Interval
First, let us consider a single interval consisting of ∆ time bins. For example,
if spike times have been binned to 2 ms, for an interval of width 20 ms we
have ∆ = 10. Since time has been discretized, it is still possible to discuss
this unitless value as a length of time, a time scale, or a interval width for
a given bin size. As before, we assume that the sequence is binary, so each
bin has either zero or one spike. This is true even when the spike times
are jittered because spike times are sampled without replacement. For this
single interval, the probability of a given number of coincidences occurring is
determined by three values: ∆, N(X, j) the number of spikes in interval j of
spike train X, and N(Y, j) the number of spikes in interval j of spike train
Y.
As a first step, we count the number of perfect coincidences, in which one
spike occurs in both X and Y within the same time bin, meaning τ = 0.
Using the standard notation of(cid:0)a
b), we find that there are (cid:0) ∆
each of them falls into one of these empty bins is (cid:0)∆−N (Y,j)
b(cid:1) for the combinatorial operation (a choose
N (X,j)(cid:1) ways to distribute N(X, j) spikes in ∆
N (X,j) (cid:1). These are all
available bins. The number of empty (spike-less) bins in spike train Y is
[∆ − N(Y, j)]. The number of ways to distribute N(X, j) spikes such that
2The procedure of generating one set of spike trains for multiple lags is appropriate
inappropriate only if each lag is being tested independently. If the test statistic is the sum
of C(τ ) over a range of lag values, a single set of simulated spike trains is appropriate.f
8
Closed Form Jitter Analysis
possible cases in which a coincidence is avoided. The probability that zero
coincidences occur in the j-th interval is therefore
P (C int
j = 0∆, N(X, j), N(Y, j)) = (cid:0)∆−N (Y,j)
N (X,j) (cid:1)
(cid:0) ∆
N (X,j)(cid:1)
where C int
j
is the number of coincidences in this interval.
(3)
We can generalize equation 3 to a non-zero number c of coincidences
by breaking the numerator up into the number of ways that c spikes can
coincide with the spikes in Y , and N(X, j) − c spikes coincide with the gaps
(or non-spikes) in Y . We can thus compute a probability distribution for
each interval j,
P (C int
j = c∆, N(X, j), N(Y, j)) = (cid:0)∆−N (Y,j)
N (X,j)−c(cid:1)(cid:0)N (Y,j)
(cid:1)
(cid:0) ∆
N (X,j)(cid:1)
c
(4)
where we follow the customary convention of setting the value of a "choose"
operation to zero if either of its arguments is negative, or if its upper argu-
ment is less than the lower. If this happens in the numerator of eq. 4, the
probability on the left hand side becomes zero. Of course, the denominator
is always positive since N(X, j) ≤ ∆. This is a hypergeometric distribution.
Equation 4 is easily generalized to nonzero values of τ by applying the
analysis leading to it to a shifted version of Y . For the computation of
N(Y, j), this implies adding τ to the summation limits in eq. 1. As with
other correlation algorithms, the boundaries of finite spike trains (beginning
and end) result in fewer intervals to analyze as τ gets further away from
zero. Thus generalizing eq. 4 to non-zero τ , we denote the resulting number
j (τ ) and the associated probability distributions as P int
of coincidences as C int
.
τ
2.2.2 Jitter-Corrected Cross Correlation
Once we have the analytical probability distribution for the correlations, we
can obtain all relevant quantities to characterize the pairwise correlations be-
tween two spike trains. It is straightforward to compute the commonly used
jitter-corrected cross correlogram (e.g., Hirabayashi et al., 2013a,b; Smith et al.,
2013) which shows the correlation function after all correlations on time scales
longer than ∆ have been removed.
It is defined in analogy to equation 2
where the expectation value of the stochastic solution, E[C MC(τ )], was used.
9
Closed Form Jitter Analysis
We can replace this approximation by the exact solution E[C int(τ )]. Fur-
thermore, by the null hypothesis each interval is conditionally independent
based on the spike counts. Therefore, the JCCG can be computed without
approximation by
JCCG(τ ) = C(τ ) − E"Xj
Cint
j (τ )# = C(τ ) −Xj
E(cid:2)Cint
j (τ )(cid:3)
(5)
which, as should be remembered, is computed for a specific jitter interval
width ∆. The expectation on the right can either be calculated for each
window as N(x, j) × N(y, j)/∆.
The jitter-corrected cross correlogram is used, for instance, when the
scientific question of interest is whether there are significant changes in syn-
chrony between conditions, rather than a test of the presence or absence of
synchrony.
It is then used as part of a bootstrap statistical test in which
the observed pairwise correlation is compared with the distribution obtained
from eq. 5.
2.2.3 Probability Distribution For Spike Train
τ
One can also obtain the probability distribution for the entire signal Pτ (C(τ ))
as the convolution of the individual probability distributions for all intervals,
P int
. This is identical to computing Pτ (C MC) from Section 2.1 with an infinite
number of Monte Carlo simulations for each value of τ . One can then evaluate
how likely it is that the observed cross correlation C(τ ) is explained by this
probability distribution. The likelihood p that this is the case is obtained as
the integral of the probability density function exceeding C(τ ), as in
pτ =
∞
Xc=C(τ )
Pτ (c)
(6)
3 Results
3.1 Computational Complexity
In many situations, the statistical distributions underlying the phenomena
under study are complicated or unknown and performing Monte Carlo simu-
lations are the only way to make progress, even though it may be costly and
10
Closed Form Jitter Analysis
it introduces additional randomness in the processing. In the case considered
here (binary spike trains, null hypothesis of uniform spike time distribution
in fixed interval, cross correlation test statistic), the distribution Pτ (C) can
be computed directly, using the closed form jitter method described above,
without the need for repeated simulations. This section will compare the
computational complexity of using the Monte Carlo jitter method against
the direct computation using the closed form jitter method.
3.1.1 Monte Carlo Method
In the Monte Carlo algorithm, the data generation step requires a permuta-
tion of ∆ data points for each interval and simulation. Since a single per-
mutation operation has a computational complexity O(∆), and ∆ times the
number of intervals is the length of the signal T , generating the set of Monte
Carlo simulations {X MC
} is O(NMC ×T ). The complexity of cross correlation
or convolution of two signals with lengths T is O(T × log T ), assuming an
FFT-based method (Cooley and Tukey, 1965) is used. So computing the full
Monte Carlo probability distribution for all values of τ is O(NMC×T ×log T ).
In many cases, not all values of τ are needed. If the correlation is computed
only for the subset of delay values from 0 to τmax, the complexity for the
Monte Carlo jitter method is
i
O(NMC × T × τmax)
Computing the jitter-corrected cross correlogram by this method only
requires one additional sum, with complexity O(NMC × τmax) so the total
complexity remains unchanged.
3.1.2 Closed Form Probability Distribution
To compute the exact probability distribution with the closed form jitter
method, note that the values of the distribution can be precomputed based
on the maximum values of N(X, j) and N(Y, j) over all j; call this maxi-
mum Nmax. Also, n choose k operations can be as fast as O(min(k, n − k))
(Manolopoulos, 2002). Therefore, a three dimensional table of all possible
values of P (C int
N(X, j), N(Y, j)) can be precomputed and then looked up
for each interval. Generating this table requires up to Nmax different values
of C, Nmax values of N(X, j), and Nmax values of N(Y, j). Computing each
value requires three n choose k operations, which are on the order of O(Nmax),
j
11
Closed Form Jitter Analysis
4). While the ex-
so the total computation of the probability table is O(Nmax
ponent is high, the expression does not have any dependence on the length
of the signal and, furthermore, Nmax ≤ ∆ is a small number in essentially
all cases of interest. In practice, for analyzing neurophysiological data it is
rare that a time resolution finer than 1 ms is needed, or controlling for cross
correlations at time scales larger than approximately 100 ms (i.e. ∆ ≈ 100).
The full lookup table is therefore maximally a 100 × 100 × 100 matrix, which
requires negligible resources to compute and store.
To compute the combined probability distribution Pτ (C) over all inter-
vals, all interval probability distributions Pτ (C int
j ) must be convolved, and
the computational complexity of the problem is dominated by these convo-
lution operations. As will be shown, we can improve performance by taking
advantage of the structure of the problem at hand, since many of the convo-
lution operations are identical. As a result, the convolutions can be grouped
together based on N(X, j) and N(Y, j) and quickly combined so that O(T )
2) convolutions. This can be done by the
convolutions will turn into O(Nmax
following procedure:
1. Take the Fast Fourier Transform (FFT) of Pτ (C int
j
for each encountered value of N(X, j) and N(Y, j).
N(X, j), N(Y, j))
2. Raise each complex frequency spectrum value to a power equal to the
number of times that the (N(X, j), N(Y, j)) pair appears.
3. Multiply these frequency spectra.
4. Take the inverse FFT of the result to get the final probability distribu-
tion Pτ (C) and compute pτ as in equation 6.
5. Repeat steps 2 through 4 for each value of τ to be tested.
The FFT operations in step 1 must be zero-padded up to the maximum
number of coincident spikes Cmax to account for the highest possible number
of synchronous spikes in the combined probability distribution. Therefore the
FFT operation in step 1 is O(Cmax × log(Cmax)). In step 2, exponentiation is
2) exponents to be taken, repeated τmax
O(1). However there are O(T × Nmax
2) multiplications, again repeated
times in step 5. Step 3 requires O(T ×Nmax
τmax times. In step 4, the length of the spectral signal (to be inverted by FFT)
is Cmax, so the operation is O(Cmax × log(Cmax)) repeated τmax times. When
12
Closed Form Jitter Analysis
combining these steps, note that Cmax is proportional to T . However Cmax
will be used when relevant because it captures the frequency dependence of
the computation time. Therefore the total computational complexity is
O(Cmax × log(Cmax) × τmax)
.
Note that because the zero-frequency component of a probability distribu-
tion is always exactly unity, the inverse FFT computation will have accuracy
limited by the precision of the numerical system.
In practice this implies
that p values less than 10−13 will not be estimated accurately.
3.1.3 Jitter-Corrected Cross Correlation
Both the complexity analysis and the actual computation of the jitter-corrected
cross correlogram is much simpler than that of the probability distribution.
We generate a lookup table of possible E(cid:2)C int(cid:3) values and, from equation 5,
the jitter-corrected correlogram can be computed at a speed of
O(T × τmax)
.
3.2 Computational Execution Time
For practical applications, consumption of resources is an important limita-
tion for any computational method. For the size of problems encountered in
typical neurophysiological experiments, the only limiting resource is execu-
tion time. To compare the performance of the Monte Carlo jitter and the
closed form method, the two algorithms were run side by side in the MAT-
LAB environment (Mathworks, Natick MA). Synthetic spike trains were gen-
erated that varied in both frequency of spiking (5 to 500 Hz) and length (1
to 91 seconds). For each (time, frequency) condition, 50 spike trains were
generated, binned to 1 ms, and the average processing time was computed.
Processing was performed with τmax = 100ms and ∆ = 20. All computations
were performed on an Intel i7 920 processor with 12 GB of RAM running
Linux Ubuntu 12.04.
For the Monte-Carlo method, NMC was set to 1000. This selection of
NMC is unrealistically low for two reasons. First, it can at best result in a
13
Closed Form Jitter Analysis
Bonferroni corrected p value of 0.201 due to the 201 p values being tested
in the range of −τmax to τmax. As the execution for NMC = 1000 already
takes 5.7 days to run, increasing NMC is impractical. Second, only a single
set of Monte-Carlo trials were generated for all lag values computed, induc-
ing potential correlations between the p values. These correlations should
decrease as more trials are generated. Therefore results are extrapolated to
NMC = 20, 000 (resulting in a minimum p ≈ 0.01) under the assumption
that the processing for 20 times as many simulations would take 20 times as
long. Though the bonferroni correction used here is conservative, it is less
conservative (by an order of magnitude) than simulating a whole new set of
spike trains for each p value as would be required to entirely eliminate any
correlations between the p values.
For the closed form jitter method, all lookup tables were computed de
novo for each spike train. This is a conservative approach (favoring the
Monte Carlo technique) since performance of the closed form jitter method
could be improved by computing the tables only once and using them for all
spike trains. This is certainly advised in a "production environment."
The results of this simulation, shown in Figure 1, illustrate a number of
features about the speed of the two algorithms. Plotted is the performance
gain, defined as the ratio of the computation time between the Monte Carlo
jitter method and the closed form jitter method. The first observation is
that the closed form method is substantially faster than the Monte Carlo
method in all cases considered. Second, while the performance gain depends
only weakly on spike train length, it does decrease with increasing firing
rate. This is because the computation time of the closed form jitter method
increases with firing rate. In practice, however, it is rare to observe firing
at sustained frequencies exceeding 100 Hz in physiological recordings. In the
physiological range, the closed form jitter algorithm is faster by a factor of
approximately 180 to 7200.
Harrison (2013) uses importance sampling to accelerate the Monte Carlo
hypothesis testing process which requires drawing fewer samples.
In that
work the number of samples needed, even for a low Bonferroni corrected p
value, is reduced to 100. However, each sample is reported to take 18 times
as long to generate and process as before, effectively resulting in a simulation
about 11 times faster than the Monte Carlo simulation with NMC = 20, 000.
Therefore, under physiological conditions the closed form computation has an
expected speed-up of 16 to 650 times compared to the importance sampling
method. It should be noted that in cases where even lower p values are needed
14
Closed Form Jitter Analysis
because of multiple hypothesis constraints, importance sampling will provide
larger gains in estimating very small p values. In these cases, increasing the
p value requirements has no effect on the computation speed of the closed
form method, so the closed form method can be expected to be faster in all
cases.
Another improvement mentioned in Section 2.2.2 is the ability of the
closed form jitter method to compute the jitter corrected correlogram very
rapidly, without computing the null hypothesis distribution of correlation
values. To show the magnitude of the improvement, the simulation was
repeated with only the mean of the null hypothesis distribution computed
under the closed form jitter method since this is all that is needed for the
corrected correlation function, eq. 5. We also restricted firing frequencies to
the range 5 -- 200 Hz. Figure 2 shows the ratio of the time it takes to compute
equation 2 vs. equation 5. In these cases, the closed form jitter calculation
is substantially faster (480x -- 13,000x), with increasing benefits for increasing
spike train lengths. As discussed previously, the spike train length is typically
not that of individual trials but of the concatenation of many trials.
4 Discussion
The importance, or absence of it, of precise timing of neural spikes has been
discussed for the last half-century. Several techniques have been developed
to characterize neuronal responses at fine time scales and it is clear that
statistical methods have to be developed with much care to avoid wrong
conclusions (e.g. Gawne and Richmond, 1993; Roy et al., 2000). One impor-
tant difficulty is that firing rates can co-vary in the neurons under study. It
is well-known that such co-variations are observable in quantities like pair-
wise cross correlation functions but they are typically considered as irrelevant
from the point of view of neuronal coding or of determining the connectivity
in the underlying circuitry. For instance, the onset of a stimulus will typi-
cally generate a temporary increase in firing rates in sensory cortex but the
resulting increase in cross correlation is usually not considered of importance
for neural coding (for an exception see Chase and Young, 2007, who showed
that spike timing relative to onset-related population activity is informative).
One common way to subtract such stimulus-locked effects is by subtracting
a "shuffle predictor" (Perkel et al., 1967), obtained by computing cross cor-
relations between spike trains from permuted trials. It has been pointed out
15
Closed Form Jitter Analysis
repeatedly (see references in the Introduction) that this does not eliminate
spurious correlations, including close to τ = 0 (synchrony).
Brody (1998, 1999) and Amarasingham et al. (2012) proved that adopt-
ing the null hypothesis of independent neurons can not solve the problem.
Observation of such correlations is, indeed, evidence against the null hypoth-
esis of independence between the two observed spike trains. Rejection of this
null hypothesis can, however, occur either because spikes in the two spike
trains are correlated "one-by-one" (synchrony), or because of slow firing rate
covariations common to both spike trains. The fact that this null hypothesis
can be rejected does not tell us why it is rejected. If the question is whether
synchrony exists at less than a given time scale (only), this is the wrong null
hypothesis. Instead, the time scale needs to be specified explicitly. The null
hypothesis chosen by Amarasingham et al. (2012) is that changes of spike
times within a time interval of size ∆ have no effect on the computed statis-
tic, in this case the correlation function.
It is this null hypothesis that is
tested by computer simulation in the Amarasingham et al. (2012) study and
analytically in this report.
A key element of the methods discussed here is that the jitter intervals
are defined without reference to the original spike trains. This ensures that
if the null hypothesis is true, there is no way to distinguish the original
spike trains from the Monte Carlo simulated spike trains. This characteristic
(called exchangeability) ensures that the obtained p values are from a well
formulated hypothesis test. If, on the other hand, the resampling method
was changed so that each spike was jittered about it's original spike time,
then even under the null hypothesis the original spike train would stand out
from the rest because all of its spikes would be at the center of the jitter
intervals. Therefore the resulting test would not be a proper statistical test
and should be avoided (Amarasingham et al., 2012).
We have discussed two ways one can choose to characterize the correla-
tions between two spike trains. One is a strict hypothesis testing approach.
A null hypothesis is formulated, namely that the observed correlations are
indistinguishable from correlations between spike trains whose spikes have
been distributed randomly within intervals of length ∆, without changing
the number of spikes in each interval. By comparing the observed correla-
tion with those in the distribution generated under the null hypothesis, it is
then decided for a given α whether the null hypothesis can be rejected.
The alternative is to compute the time-resolved correlation function and
"correct" for the correlations as observed under the null hypothesis, by sub-
16
Closed Form Jitter Analysis
tracting the expectation value of the latter. This is the more commonly cho-
sen approach, perhaps because the time-resolved correlation function is both
intuitive and familiar. The distribution of JCCG values can be compared
between experimental conditions (indicating a change in 'excess synchrony')
using a bootstrap test to test for significance. Also, its shape (e.g. the lo-
cation of peaks) may provide insight that goes beyond the yes-no answer
whether the null hypothesis can be rejected or not.
In the Amarasingham et al. (2012) study, the Monte Carlo procedure is
further developed to account for more potential causes of fine timing effects
besides synchrony such as ramping spike rates within an interval or inter-
spike interval distribution effects. These methods are straightforward and
statistically well-defined. Like any Monte Carlo method, however, they only
generate an approximation to the underlying distribution whose quality de-
pends on the number of surrogate spike trains. In practice what is more prob-
lematic is that the method can be computationally very costly. For instance,
as discussed in section 3.2, our example problem using the simplest of the
null hypotheses discussed (50 spike trains of a few seconds long each, mean
rates between 1 and 100 Hz, maximal time lag of 100 ms, α = 0.01 with Bon-
ferroni correction applied) would have required a simulation several months
long on a reasonably fast machine. We therefore only simulated NM C = 1000
trials and extrapolated to the execution time needed for NM C = 20, 000 but
even that abbreviated Monte Carlo run took nearly six days. Some progress
can be made by using much faster machines or many machines (the problem
parallelizes easily) but execution time is clearly a problem.
In contrast, the closed form jitter methods this report focuses on are ex-
act, rather than approximate. More important for practical applications may
be that they are extremely efficient, with a speed-up of at least two orders of
magnitude for the hypothesis testing approach, and four orders of magnitude
for the full correlation functions. Even over importance sampling methods
(Harrison, 2013), they have been shown to provide a substantial increase in
speed. For the hypothesis testing examples used in our study (whose scope is
quite comparable to that of typical neurophysiological experiments, assum-
ing a proper Bonferroni correction is applied), computation time is reduced
from more than 100 days under the original Monte Carlo method to about
one night. Computational time required for the full correlation function is
reduced from over 100 days to a few minutes. An increase in performance on
this scale is more than merely a quantitative improvement. For instance, it is
essentially impossible to explore variations in the analyses (like the influence
17
Closed Form Jitter Analysis
of the jitter time scale ∆) if each computational run takes a few months, but
it is easy to do if it takes minutes.
So far we were only concerned with correlations between two spike trains.
Modern recording techniques are already increasing the number of simulta-
neously recorded spike trains to tens or hundreds. Unfortunately, the closed-
form jitter method is limited in the ability to analyze large ensembles. This
is because the correlation functions of some pairs in an ensemble will restrict
the possible correlation values of other pairs. For example, if there are three
neurons X, Y , and Z, and the pairs XY and Y Z have perfect correlation,
then the pair XZ must also have perfect correlation. A Monte-Carlo jitter
analysis that jitters an entire ensemble of neurons and then performs a hy-
pothesis test on the ensemble can be performed relatively simply, but no such
closed-form method exists yet. In order to avoid the constraints of the type
described above, only N − 1 pairs of neurons can be analyzed with closed
form methods when N neurons are recorded.
Additionally, the nature of the exact solutions provides an opportunity
for further exact analysis. Having a closed form solution allows questions
about the effects of spike sorting errors, the value of ∆, or the structure of
JCCG(τ ) to be addressed rigorously and more precisely than is possible with
any numerical method.
In conclusion, we study a statistical framework for quantifying correla-
tions between spike trains at given time scales. It can be applied both for
hypothesis testing and for correcting observed correlation functions for cor-
relations at these time scales. Results are exact, and both computational
complexity and computational time for realistic examples are several orders
of magnitude lower than related approaches based on Monte Carlo simula-
tions.
Matlab code will be made available by the authors upon request.
Acknowledgments
This work was supported by the Office of Naval Research under MURI grant
N000141010278 and by NIH under R01EY016281-02. We acknowledge dis-
cussions with Drs. Rudiger von der Heydt and Anne Martin who also gave
us access to unpublished data.
18
Closed Form Jitter Analysis
References
M. Abeles. Corticonics -- Neural circuits of the cerebral cortex. Cambridge
University Press, 1991.
E. D. Adrian and Y. Zotterman. The impulses produced by sensory nerve
endings. Part 2. The response of a single end organ. J. Physiol., 61:151 --
171, 1926.
A. Amarasingham, M. T. Harrison, N. G. Hatsopoulos, and S. Geman. Con-
ditional modeling and the jitter method of spike resampling. Journal of
Neurophysiology, 107(2):517 -- 531, 2012. PMC3349623.
C. D. Brody. Slow covariations in neuronal resting potentials can lead to arte-
factually fast cross-correlations in their spike trains. J. Neurophysiology,
80(6):3345 -- 51, December 1998.
C. D. Brody. Correlations without synchrony. Neural Comput, 11(7):1527 --
1535, Oct 1999.
S. M Chase and E. D Young. First-spike latency information in single neurons
increases when referenced to population onset. Proceedings of the National
Academy of Sciences (USA), 104(12):5175 -- 5180, 2007.
J. W. Cooley and J. W. Tukey. An algorithm for the machine calculation
of complex Fourier series. Mathematics of computation, 19(90):297 -- 301,
1965.
T. J. Gawne and B. J. Richmond. How independent are the messages carried
by adjacent inferior temporal cortical neurons? J. Neurosci., 13(7):2758 --
2771, 1993.
M. T. Harrison. Accelerated spike resampling for accurate multiple testing
controls. Neural Computation, 25(2):418 -- 449, 2013.
T. Hirabayashi, D. Takeuchi, K. Tamura, and Y. Miyashita. Functional mi-
crocircuit recruited during retrieval of object association memory in mon-
key perirhinal cortex. Neuron, 77(1):192 -- 203, January 2013a.
T. Hirabayashi, D. Takeuchi, K. Tamura, and Y. Miyashita. Microcircuits
for hierarchical elaboration of object coding across primate temporal areas.
Science, 341(6142):191 -- 5, July 2013b.
19
Closed Form Jitter Analysis
Y. Manolopoulos. Binomial coefficient computation: recursion or iteration?
ACM SIGCSE Bulletin, 34(4):65 -- 67, 2002.
E. Niebur, C. Koch, and C. Rosin. An oscillation-based model for the neural
basis of attention. Vision Research, 33:2789 -- 2802, 1993.
D. H. Perkel, G. L. Gerstein, and G. P. Moore. Neuronal spike trains and
stochastic point processes. II: Simultaneous spike trains. Biophys. J., 7:
419 -- 440, 1967.
A. Riehle, S. Grun, M. Diesmann, and A. Aertsen. Spike synchronization and
rate modulation differentially involved in motor cortical function. Science,
278:1950 -- 1953, December 1997.
A. Roy, P. N. Steinmetz, and E. Niebur. Rate limitations of unitary event
analysis. Neural Computation, 12(9):2063 -- 2082, 2000.
M. N. Shadlen and W. T. Newsome. The variable discharge of cortical neu-
rons: implications for connectivity, computation, and information coding.
J. Neurosci., 18:3870 -- 3896, 1998.
W. Singer and C. M. Gray. Visual feature integration and the temporal
correlation hypothesis. Annu. Rev. Neurosci., 18:555 -- 586, 1995.
M. A. Smith, X. Jia, A. Zandvakili, and A. Kohn. Laminar dependence of
neuronal correlations in visual cortex. Journal of Neurophysiology, 109(4):
940 -- 7, March 2013.
W. Softky. Simple codes versus efficient codes. Current Opinion in Neurobi-
ology, 5:239 -- 247, 1995.
P. N. Steinmetz, A. Roy, P. Fitzgerald, S. S. Hsiao, K. O. Johnson, and
E. Niebur. Attention modulates synchronized neuronal firing in primate
somatosensory cortex. Nature, 404:187 -- 190, 2000.
20
Closed Form Jitter Analysis
P value Computation
5 Hz
10 Hz
20 Hz
40 Hz
50 Hz
100 Hz
200 Hz
400 Hz
500 Hz
20
40
60
80
100
Spike Train Length [s]
4
10
3
10
i
n
a
G
e
c
n
a
m
r
o
f
r
e
P
2
10
1
10
0
Figure 1: Performance gain in implementing the closed form jitter method
for p value computations. Gain is defined as the ratio in computation time
between the Monte Carlo Jitter method and the closed form jitter method.
Processing parameters used are τmax = 100ms, ∆ = 20, and NMC = 20, 000
(see text for details).
21
Closed Form Jitter Analysis
JCCG Computation
4
10
5 Hz
50 Hz
100 Hz
200 Hz
i
n
a
G
e
c
n
a
m
r
o
f
r
e
P
3
10
0
20
40
60
80
100
Spike Train Length [s]
Figure 2: Performance gain in implementing the closed form jitter method
for Jitter Corrected Correlogram computations. Gain is defined as the ratio
in computation time between the Monte Carlo Jitter method and the closed
form jitter method. Processing parameters used are τmax = 100ms, ∆ = 20,
and NMC = 20, 000 (see text for details). The lowered performance gain at
signal length of 1 second is due to the overhead of computing the probability
table de novo for each spike train.
22
|
1302.3943 | 1 | 1302 | 2013-02-16T08:10:04 | Interplay between Network Topology and Dynamics in Neural Systems | [
"q-bio.NC",
"cond-mat.dis-nn",
"nlin.AO",
"physics.bio-ph"
] | This thesis is a compendium of research which brings together ideas from the fields of Complex Networks and Computational Neuroscience to address two questions regarding neural systems:
1) How the activity of neurons, via synaptic changes, can shape the topology of the network they form part of, and
2) How the resulting network structure, in its turn, might condition aspects of brain behaviour.
Although the emphasis is on neural networks, several theoretical findings which are relevant for complex networks in general are presented -- such as a method for studying network evolution as a stochastic process, or a theory that allows for ensembles of correlated networks, and sets of dynamical elements thereon, to be treated mathematically and computationally in a model-independent manner. Some of the results are used to explain experimental data -- certain properties of brain tissue, the spontaneous emergence of correlations in all kinds of networks... -- and predictions regarding statistical aspects of the central nervous system are made. The mechanism of Cluster Reverberation is proposed to account for the near-instant storage of novel information the brain is capable of. | q-bio.NC | q-bio |
Department of Electromagnetism and Physics of Matter &
Institute Carlos I for Theoretical and Computational Physics,
University of Granada, Spain.
Interplay between Network
Topology and Dynamics in
Neural Systems
Samuel Johnson
Ph.D. Thesis
Advisors: Joaqu´ın J. Torres Agudo
& Joaqu´ın Marro Borau
Granada, April 2011
Dedicated to my family
Acknowledgements
Aunque hay muchas m´as personas a las que estoy agradecido, por tantas cosas,
de las que puedo enumerar en un espacio razonable, tratar´e de poner el umbral
en alg´un sitio y mencionar expl´ıcitamente s´olo a aquellas que me han ayudado
de alguna manera directa con esta tesis. En primer lugar quiero hacer constar
mi sincero agradecimiento a mis directores: Joaqu´ın Torres no s´olo es quien
me introdujo al caos, los fractales, el SOC, las redes, la simulaci´on... sino el
que me mostr´o que la neurociencia es algo tan concreto y estudiable como la
f´ısica estad´ıstica; Joaqu´ın Marro, por su parte, al compartir conmigo su genial
perspectiva sobre la ciencia y la complejidad, me ha abierto los o jos a toda una
manera de ver el mundo. Miguel ´Angel Munoz ha sido tambi´en una influencia
cientf´ıfica important´ısima, con su maravillosa combinaci´on de talento y buen
humor. Agradezco tambi´en a tod@s l@s dem´as companer@s con quienes he
traba jado y convivido estos anos, tanto por lo que he aprendido de ellos como
por los muchos buenos ratos: Pablo Hurtado, Pedro Garrido, Paco de los
Santos, Antonio Lacomba, Elvira Romera, Ram´on Contreras, Paco Ramos,
Jes´us Cort´es, Juan Antonio Bonachela, Luca Donetti, Jos´e Manuel Mart´ın,
Omar al-Hammal, Jes´us del Pozo, Carlos Espigares, Clara Guglieri, Pablo
Sartori, Marina Manrique, Jordi Hidalgo, Virginia Dom´ınguez, Jorge Mej´ıas,
Sebastiano de Franciscis, Alejandro Pinto, Leticia Rubio, Jordi Garces, Luca
Sabino, Simone Pigolotti, Lu´ıs Seoane, Daniele Vilone y Miguel Ib´anez. Como
bien sab´eis, sois companer@s mucho m´as que de traba jo, si´endolo tambi´en,
seg´un el caso, de piso, de grupo musical, de decrepitud, de juegos diversos,
del alma... Por supuesto, hay otras personas con esta multiplicidad de roles a
quienes sin embargo no voy a tratar de nombrar aqu´ı: por favor, no pens´eis
que es por falta de reconocimiento, sino por ahorrar papel y tinta, y porque
de todas maneras seguramente no vay´ais a leer esto.
I’m grateful to Mike Ramsey, Mohammed Boudjada and Helmut Rucker for
treating me so well in Graz as well as introducing me to the world of research.
Doy gracias a Marcelo del Castillo por invitarme a la UNAM y por ser, junto
v
vi
con su familia y amig@s, tan buen anfitri´on en M´exico; a Ezequiel Albano,
Gabriel Baglietto, Bel´en Moglia, Nara Guisoni y Luis Diambra por tratarme
tan bien en La Plata; and to Nick Jones and Sumeet Agarwal for making me
so welcome in Oxford. I am also grateful to all the people who have helped
in one way or another with my research, be it reading manuscripts, providing
data, suggesting ideas, or simply with stimulating conversations. I’m proba-
bly leaving many deserving people out when I mention Dante Chialvo, ´Alex
Arenas, Ginestra Bianconi, Yamir Moreno, Jennifer Dunne, Alberto Pascual,
V´ıctor Egu´ıluz, Sasha Goltsev, Gorka Zamora, Lars Rudolf, Sabine Hilfiker,
Tiago Peixoto, Ole Paulsen and Peter Latham. Tambi´en estoy agradecido a
Juan Soler, Alberto Prieto, Juan Calvo, Pilar Guerrero, Irene Mendoza, Es-
trella Ryan, Luna ´Alvarez, Javier (Chancly) Pascual, Nikolina Dimitrov, Felisa
Torralba y Caroline de Cannart por su diversa ayuda. Three members of my
family who have been particularly influential on my scientific interests and
helpful in various ways are my uncle Dave Jones, my grandfather Tony Jones,
and my aunt Sue Ziebland.
C´ecile Poirier, qui as ´et´e avec moi pendant la plus part de ces derni`eres
quatre ans, tu m’as fait surmonter tant d’obstacles et m’as aid´e tellement que
je n’ai vraiment pas de mots pour te remercier assez. And the mental freedom
and emotional basis I need to undertake this or any other pro ject is grounded
in the unconditional support and love from my parents, Jenni and Alan, and
my sisters, Jazz and Abby. Thanks so much to all of you.
Finally, thanks are also due to the government of the United States for its
perseverance in attempting to crush the news organization Wikileaks, whose
continual publication of enlightening documents so often kept me from work-
ing on this thesis; and to myself, for at all times resisting the temptation of
perfectionism when writing this up – as the reader will no doubt observe.
“To say that a man is made up of certain chemical elements is a satisfactory
description only for those who intend to use him as a fertilizer.”
Hermann Joseph Muller
“Je n’avais pas besoin de cette hypoth`ese-l`a.”
Pierre-Simon Laplace
“Research is what I’m doing when I don’t know what I’m doing.”
Werner von Braun
Abstract
This thesis is a compendium of research which brings together ideas from the
fields of Complex Networks and Computational Neuroscience to address two
questions regarding neural systems:
1) How the activity of neurons, via synaptic changes, can shape the topology
of the network they form part of, and
2) How the resulting network structure, in its turn, might condition aspects of
brain behaviour.
Although the emphasis is on neural networks, several theoretical findings which
are relevant for complex networks in general are presented – such as a method
for studying network evolution as a stochastic process, or a theory that allows
for ensembles of correlated networks, and sets of dynamical elements thereon,
to be treated mathematically and computationally in a model-independent
manner. Some of the results are used to explain experimental data – cer-
tain properties of brain tissue, the spontaneous emergence of correlations in
all kinds of networks... – and predictions regarding statistical aspects of the
central nervous system are made. The mechanism of Cluster Reverberation is
proposed to account for the near-instant storage of novel information the brain
is capable of.
ix
Preamble: The Ant, the
Grasshopper and Complexity
Once upon a time, in a charming and peaceful little valley, a grasshopper
sat under the shade of a sunflower, idly strumming up a tune, when a young
worker ant came into view. The grasshopper watched as she trundled her way
laboriously up an incline under the weight of a large piece of leaf. When she
was close enough, he hailed her:
‘Ahoy there, friend. I hope I won’t seem tactless if I point out what a sin-
gularly cumbersome bit of leaf you have there. Would you not rather put it
down for a while and join me for a quick jam session? You could bang along
on some twigs or something.’
‘Thank you for the offer, but I must continue on my way,’ replied the ant,
glancing up in slight surprise at being thus addressed.
‘Oh, what a pity,’ the grasshopper rejoined. ‘And where, if I may be so bold as
to inquire, would you be taking your rather unappetising ration of cellulose?’
‘Well, I can’t say I really know... I just follow this trail of pheromones I’ve
come across. I’m sure it’s for some noble purpose though.’
‘Ah, that must be reassuring. And I suppose when you get to wherever it is
you don’t know you’re going you intend to eat your bit of leaf...’
‘Oh no, I can’t digest something like this – who do you take me for?’
‘You can’t? Well, how strange...’
‘What’s strange?’
‘However did an animal evolve which, instead of engaging in biologically rea-
sonable (not to mention enjoyable) activities, such as playing music to attract
sexual partners, prefers to lug useless bits of leaf about? How on earth can
that serve to spread your genes?’
‘I’m not interested in music or sex, whatever those are. I just follow simple
rules, like all my identical sisters. You could say we’re automata.’
‘Thanks, I was going to but wasn’t sure whether you’d be offended. Well, let
xi
xii
me wish you an agreeable day of toil, you frigid little automaton.’
With that, the grasshopper gave a big leap into the air, slightly exasper-
ated by the folly so often displayed by his fellow insects. Looking down, he
spotted a few more ants, all carrying leaves in the same direction as the one
he had just met.
Intrigued, he fluttered slightly higher (since grasshoppers
can, actually, fly, if not all that well). He realised the ants were all heading
for a nest some way off. In fact, there were many ant trails leading to various
sources of food. It dawned on the grasshopper that although the individual
ants were just boring little morons idiotically following rules, the nest as a
whole was managing to find the closest leaves, bring them back along optimal
routes, and feed them to its plantations of fungi. The colony was behaving
like an intelligent organism, in some respects not so different from he himself,
who functioned thanks to the cells of his body – each with the same genome,
like the ants – cooperating through the obedience to relatively simple rules.
This thought impressed the grasshopper very much, driving him to flutter
even higher so as to see things in greater perspective. From there he considered
the apparently fragile web of trophic, parasitical and symbiotic interactions
linking all the living beings in the valley – a network which nonetheless must
have evolved a particularly robust structure not to shatter at the first environ-
mental fluctuation. He became so enthralled by the idea of such complexity
on one scale emerging from simplicity on another that he didn’t even pay any
attention to an attractive young grasshopperess making her wanton way just
below him. Instead, he couldn’t help fearing that a butterfly he noticed gently
flapping his wings would probably set off a hurricane somewhere. As he flew
ever higher, he began to see snowflakes glide by, overwhelmingly intricate and
beautiful patterns self-organised out of the simplest little water molecules. Fi-
nally he was so high that he began to reflect on how the very stellar systems,
galaxies, clusters, superclusters, filaments of galaxies... – of which his whole
world was but an infinitesimal component – also interacted with each other
via the simple rules of gravity and pressure to form ob jects marvellous beyond
conception.
What he didn’t notice until it was too late, as he left behind the cosy
protection of the atmosphere, was how ultraviolet sunlight and ionising cos-
mic rays were steadily burning his wings each to a crisp. Beginning to fall, he
xiii
only hoped he would have time to consider the several morals to his tragic tale.
After a while spent plummeting to his doom he realised that, the freefall
terminal velocity and life expectancy of a grasshopper being what they respec-
tively were, he would most likely die peacefully of old age somewhere along
his way down – never again contemplating his Edenic valley except, like some
prophetic locust, from afar.
Contents
Acknowledgements
Abstract
Preamble: The Ant, the Grasshopper and Complexity
List of figures
List of tables
1 Resumen en espanol
v
ix
xi
xxv
xxvii
1
7
2 Where we are and where we’d like to go
2.1 From bridges to brains . . . . . . . . . . . . . . . . . . . . . . .
7
2.2 Neural networks in neuroscience . . . . . . . . . . . . . . . . . . 13
2.3 A declaration of intent . . . . . . . . . . . . . . . . . . . . . . . 18
3 Evolving networks and the development of neural systems
21
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1
3.2 Basic considerations
. . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Synaptic pruning . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Phase transitions . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.6 The C. Elegans neural network . . . . . . . . . . . . . . . . . . 34
3.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
39
4 Bringing on the Edge of Chaos with heterogeneity
4.1 Exciting cooperation . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 The Fast-Noise model . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 Edge of Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.4 Network performance . . . . . . . . . . . . . . . . . . . . . . . . 47
xiv
Contents
xv
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5 Correlated networks and natural disassortativity
51
5.1 Assortativity of networks . . . . . . . . . . . . . . . . . . . . . . 51
5.2 The entropy of network ensembles . . . . . . . . . . . . . . . . . 53
5.3 Entropic origin of disassortativity . . . . . . . . . . . . . . . . . 55
5.4 To sum up... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6 Enhancing robustness to noise via assortativity
61
6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.2 Preliminary considerations . . . . . . . . . . . . . . . . . . . . . 64
6.2.1 Model neurons on networks
. . . . . . . . . . . . . . . . 64
6.2.2 Network ensembles . . . . . . . . . . . . . . . . . . . . . 65
6.2.3 Correlated networks
. . . . . . . . . . . . . . . . . . . . 66
6.3 Analysis and results . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.3.1 Mean field . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.3.2 Generating correlated networks
. . . . . . . . . . . . . . 71
6.3.3 Assortativity and dynamics
. . . . . . . . . . . . . . . . 72
6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7 Cluster Reverberation: A mechanism for robust short-term
memory without synaptic learning
79
7.1 Slow but sure, or fast and fleeting? . . . . . . . . . . . . . . . . 79
7.2 The simplest neurons on modular networks . . . . . . . . . . . . 83
7.3 Cluster Reverberation . . . . . . . . . . . . . . . . . . . . . . . 84
7.4 Energy and topology . . . . . . . . . . . . . . . . . . . . . . . . 88
7.5 Forgetting avalanches . . . . . . . . . . . . . . . . . . . . . . . . 88
7.6 Clustered networks . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.7 Yes, but does it happen in the brain? . . . . . . . . . . . . . . . 93
8 Concluding remarks
9 Conclusiones en espanol
94
98
A Nonlinear preferential rewiring in fixed-size networks as a dif-
fusion process
102
B Effective modularity of highly clustered networks
110
xvi
Contents
C Nestedness of networks
112
C.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
C.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
C.3 The effect of the degree distribution . . . . . . . . . . . . . . . . 115
C.4 Nestedness and assortativity . . . . . . . . . . . . . . . . . . . . 116
C.5 Bipartite networks
. . . . . . . . . . . . . . . . . . . . . . . . . 118
C.6 Overlapping networks . . . . . . . . . . . . . . . . . . . . . . . . 120
C.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
D Publications derived from the thesis
D.1 Journals and book chapters (the most relevant ones marked with
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
an asterisk)
D.2 Abstracts
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
124
References
129
List of Figures
2.1 The problem of the Seven Bridges of Konigsberg can be reduced
to a graph in which nodes and edges represent land masses and
. . . . . . . . . . . . . . . . . . . . . . . .
bridges, respectively.
8
2.2 Drawing of the cells of the chick cerebellum, from “Estructura
de los centros nerviosos de las aves”, Madrid, 1905. Notice how
the neurons make up a complex network of synaptic interactions. 14
2.3
In the Hopfield neural-network model, the interaction strengths
(representing synaptic weights) store information in the form
of particular patterns, or configurations of firing neurons, which
become attractors of the dynamics. Whatever the initial state of
the system, it will always evolve towards one of these patterns,
thus allowing for the storage and retrieval of information. The
mechanism, known as associative memory, is thought to be at
the basis of memory in the brain. In this case, the network is
“remembering” an illustration by Jean-Baptiste Oudry for Jean
. . . . . . . . . . 16
de la Fontaine’s fable La Cigale et la Fourmi.
xviii
List of Figures
xix
3.1 Synaptic densities in layers 1 (red squares) and 2 (black circles)
of the human auditory cortex against time from conception.
Data from Huttenlocher and Dabholkar (1997), obtained by di-
rectly counting synapses in tissues from autopsies. Lines follow
best fits to Eq. (3.7), where the parameters were: for layer 1,
τp = 5041 days; and for layer 2, τp = 3898 days (for ρf we have
used the last data pints: 30.7 and 40.8 synapses/µm3 , for layers
1 and 2 respectively). Data pertaining to the first year and to
days 4700, 5000 7300, shown with smaller symbols, where omit-
ted from the fit. Assuming the existence of transient growth
factors, we can include the data points for the first year in the
fit by using Eq. (3.8). This is done in the inset (where time
is displayed logarithmically). The best fits were:
for layer 1,
τg = 151.0 and τp = 5221; and for layer 2, τg = 191.1 and
τp = 4184, all in days (we have approximated t0 to the time of
the first data points, 192 days).
. . . . . . . . . . . . . . . . . . 27
3.2 Evolution of the degree distributions of networks beginning as
regular random graphs with κ(0) = 20 in the critical (top) and
supercritical (bottom) regimes. Local probabilities are σ(k) =
k/(hkiN ) in both cases, and π(k) = 2σ(k) − N −1 and π(k) =
k3/2 /(hk3/2iN ) for the critical and supercritical ones, respec-
tively. Global probabilities as in Eq. (3.6), with n = 10 and
κmax = 20. Symbols in the main panels correspond to p(k , t) at
different times as obtained from MC simulations. Lines result
from numerical integration of Eq.
Insets show typical
(3.1).
time series of κ and m. Light blue lines are from MC simula-
tions and red lines are theoretical, given by Eq. (3.7) and Eq.
. . . . . . . . . . . . . . . . . . . 28
(3.1), respectively. N = 1000.
xx
List of Figures
3.3 Phase transitions in mst for π(k) ∼ kα and σ(k) ∼ k , and u(κ)
and d(κ) as in Eq. (3.6). N = 1000 (blue squares), 1500 (red
triangles) and 2000 (black circles); κ(0) = κmax = 2n = N/50.
Corresponding lines are from numerical integration of Eq. (3.1).
The bottom left inset shows values of the highest eigenvalue of
the Laplacian matrix (red squares) and of Q = λN /λ2 (black
circles), a measure of unsynchronizability; N = 1000. The top
right inset shows transitions for the same parameters in the final
values of Pearson’s correlation coefficient r (see Section 3.5),
both for only one edge allowed per pair of nodes (red squares)
and without this restriction (black circles).
. . . . . . . . . . . 29
3.4 Degree distribution (binned) of the C. Elegans neural network
(circles) (White et al., 1986) and that obtained with MC simu-
lations (line) in the stationary state (t = 105 steps) for an equiv-
alent network in which edges are removed randomly (β = 1) at
the critical point (α = 1.35). N = 307, κst = 14.0, averages
over 100 runs. Global probabilities as in Eq. (3.6). The slope
is for k−5/2 . Top right inset: mean-neighbour-degree function
knn (k) as measured in the same empirical network (circles) and
as given by the same simulations (line) as in the main panel.
The slope is for k−1/2 . Bottom left inset: mst of equivalent net-
work for a range of α, both from simulations (circles) and as
obtained with Eq. (3.1). (See also Table 3.1.)
. . . . . . . . . . 35
4.1 The temperature dependence of the difference between the val-
ues for the fatigue at which the ferromagnetic–periodic transi-
tion occurs, as obtained analytically for T = 0 (Φ0 ) and from
MC simulations at finite T (Φc ). The critical temperature is cal-
culated as Tc = hk2i (hki N )−1 for each topology. Data are for
bimodal distributions with varying ∆ and for scale–free topolo-
gies with varying γ , as indicated. Here, hki = 20, N = 1600 and
α = 2. Standard deviations, represented as bars in this graph,
were shown to drop with N −1/2 (not depicted).
. . . . . . . . . 45
List of Figures
xxi
4.2 The critical fatigue values Φ0 (solid lines) and Φc from MC av-
erages over 10 networks (symbols) with T = 2/N , hki = 20,
N = 1600, α = 2. The dots below the lines correspond to
changes of sign of the Lyapunov exponent as given by the iter-
ated map, which qualitatively agree with the other results. This
. . . . . . 46
is for bimodal and scale–free topologies, as indicated.
4.3 Network “performance” (see the main text) against ∆ for bi-
modal topologies (above) and against γ for scale–free topologies
(below). Φ = 0.8 for the first case and Φ = 1 in the second.
Averages over 20 network realizations with stimulation every 50
MC steps for 2000 MC steps, δ = 5 and M = 4; other param-
eters as in Fig. 6.5. Inset shows sections of typical time series
of mν for ∆ = 10 (above) and γ = 4 (below); the corresponding
stimulus for pattern ν is shown underneath.
. . . . . . . . . . . 48
5.1 Evolution of the Internet at the AS level. Empty (blue) squares
and circles: entropy per node of randomized networks in the
fully random and in the configuration ensembles, as obtained
by Bianconi (hence the “B” superscription) (Bianconi, 2008,
2009; Anand and Bianconi, 2009). Filled (red) triangles and
diamonds: Shannon entropy for an ER network and a scale-free
one with γ = 2.3, respectively.
. . . . . . . . . . . . . . . . . . 55
5.2 Shannon entropy of correlated scale-free networks against pa-
rameter β (left panel) and against Pearson’s coefficient r (right
panel), for various values of γ (increasing from bottom to top).
hki = 10, N = 104 .
. . . . . . . . . . . . . . . . . . . . . . . . . 56
5.3 Lines from top to bottom: r at which the entropy is maxi-
mized, r∗ , against γ for random scale-free networks with mean
degrees hki = 1
2 , 1, 2 and 4 times k0 = 5.981, and N = N0 =
10697 nodes (k0 and N0 correspond to the values for the Inter-
net at the AS level in 2001 (Park and Newman, 2003), which
had r = r0 = −0.189). Symbols are the values obtained in
(Park and Newman, 2003) as those expected solely due to the
one-edge-per-pair restriction (with k0 , N0 and γ = 2.1, 2.3 and
2.5). Inset: r∗ against N for networks with fixed hki/N (same
values as the main panel) and γ = 2.5; the arrow indicates
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
N = N0 .
xxii
List of Figures
5.4 Level of assortativity that maximizes the entropy, r∗ , for vari-
ous real-world, scale-free networks, as predicted theoretically by
Eq. (5.1) (circles) and as directly measured (horizontal lines),
against exponent γ .
. . . . . . . . . . . . . . . . . . . . . . . . 60
6.1 Mean-nearest-neighbour functions knn (k) for scale-free networks
with β = −0.5 (disassortative), 0.0 (neutral), and 0.5 assor-
tative, generated according to the algorithm described in Sec.
6.3.2. Inset: degree distribution (the same in all three cases).
Other parameters are γ = 2.5, hki = 12.5, N = 104 .
. . . . . . . 66
6.2 Stable stationary value of the weighted overlap µ1 against tem-
perature T for scale-free networks with correlations according
to knn ∼ kβ , for β = −0.5 (disassortative), 0.0 (neutral), and
0.5 (assortative). Symbols from MC simulations, with errorbars
representing standard deviations, and lines from Eqs.
(6.6).
Other network parameters as in Fig. 6.1. Inset: µ1 against T
for the assortative case (β = 0.5) and different system sizes:
N = 104 , 3 · 104 and 5 · 104 .
. . . . . . . . . . . . . . . . . . . . 72
6.3 Stable stationary values of order parameters µ0 , µ1 and µβ+1
against temperature T , for assortative networks according to
β = 0.5. Symbols from MC simulations, with errorbars repre-
senting standard deviations, and lines from Eqs. (6.6). Other
parameters as in Fig. 6.1.
. . . . . . . . . . . . . . . . . . . . . 73
6.4 Difference between the stationary values µ1 and µ0 for networks
with β = −0.5 (disassortative), 0.0 (neutral) and 0.5 (assorta-
tive), against temperature. Symbols from MC simulations, with
errorbars representing standard deviations, and lines from Eqs.
(6.6). Line shows the expected level of fluctuations due to noise,
∼ N − 1
2 . Other parameters as in Fig. 6.1.
. . . . . . . . . . . . 73
6.5 Phase diagrams for scale-free networks with γ = 2.5, 3, and 3.5.
Lines show the critical temperature Tc marking the second-order
transition from a memory (ferromagnetic) phase to a memory-
less (paramagnetic) one, against the assortativity β , as given by
Eq. (6.7). Other parameters as in Fig. 6.1.
. . . . . . . . . . . 75
6.6 Parameter space β − γ partitioned into the regions in which
b(β , γ ) has the same functional form – where b is the scaling
exponent of the critical temperature: Tc ∼ N b . Exponents ob-
. . . . . . . . . 75
tained by taking the large N limit in Eq. (6.7).
List of Figures
xxiii
6.7 Examples of how Tc scales with N for networks belonging to
regions I, II, III and IV of Fig. 6.6 (β = −0.8, −0.35, 0.0 and
0.9, respectively). Symbols from MC simulations, with errorbars
representing standard deviations, and slopes from Eq. (6.7). All
parameters – except for β and N – are as in Fig. 6.1.
. . . . . 76
6.8 Global order parameter ζ for assortative (β = 0.5), neutral
(β = 0.0) and disassortative (β = −0.5) networks with P = 3
(left panel) and P = 10 (right panel) stored patterns. Symbols
from MC simulations, with errorbars representing standard de-
viations. All parameters are as in Fig. 6.1.
. . . . . . . . . . . 77
7.1 Diagram of a modular network composed of four five-neuron
clusters. The four circles enclosed by the dashed line represent
the stimulus: each is connected to a particular module, which
adopts the input state (red or blue) and retains it after the
stimulus has disappeared via Cluster Reverberation. . . . . . . . 85
7.2 Performance η against λ for networks of the sort described in the
main text with M = 160 modules of n = 10 neurons, hki = 9;
patterns are shown with intensities δ = 8.5, 9 and 10, and T =
0.02 (lines – splines – are drawn as a guide to the eye). Inset:
typical time series of mstim (i.e., the overlap with whichever
pattern was last shown) for λ = 0.0, 0.25, and 0.5, and δ =
hki = 9.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.3 Configurational energy of a network composed of M = 20 mod-
ules of n = 10 neurons each, according to Eq. (7.1), for various
values of the rewiring probability λ. The minima correspond to
situations such that all neurons within any given module have
the same sign.
. . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.4 Left panel: distribution of escape times τ , as defined in the
main text, for λ = 0.25 and T = 2. Slope is for β = 1.35. Other
parameters as in Fig. 7.2. Symbols from MC simulations and
line given by Eqs. (7.2) and (7.3). Right panel: exponent β of
the quasi-power-law distribution p(τ ) as given by Eq. (7.3) for
temperatures T = 1 (red line), T = 2 (green line) and T = 3
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
(blue line).
xxiv
List of Figures
7.5 Proportion of outgoing edges, λ, from boxes of linear size l
against exponent γ for scale-free networks embedded on 2D lat-
tices. Lines from Eq. (7.4) and symbols from simulations with
hki = 4 and N = 1600.
. . . . . . . . . . . . . . . . . . . . . . 92
7.6 Performance η against exponent γ for scale-free networks, em-
bedded on a 2D lattice, with patterns of M = 16 modules of
n = 100 neurons each, hki = 4 and N = 1600; patterns are
shown with intensities δ = 3.5, 4, 5 and 10, and T = 0.01 (lines
– splines – are drawn as a guide to the eye). Inset: typical time
series for γ = 2, 3, and 4, with δ = 5.
. . . . . . . . . . . . . . . 92
A.1 Degree distribution p(k , t) at four different stages of evolution:
t = 102 [(yellow) squares], 103 [(blue) circles], 104 [(red) trian-
gles)] and 105 MCS [(black) diamonds]. From top to bottom
panels, subcritical (α = 0.5), critical (α = 1) and supercritical
(α = 1.5) rewiring exponents. Symbols from MC simulations
and corresponding solid lines from numerical integration of Eq.
. . . . . . . . 106
(A.1). β = 1, hki = 10 and N = 1000 in all cases.
A.2 Adjusted variance σ 2/hki2 of the degree distribution after 2×105
MCS against α, as obtained from MC simulations, for system
sizes N = 800 [(yellow) squares], 1200 [(blue) circles], 1600
[(red) triangles] and 2000 [(black) diamonds]. Top left inset
shows final degree distributions for α = 0.5 [light gray (blue)], 1
[dark gray (red)] and 1.5 (black), with N = 1000. Bottom right
inset shows typical time series of σ 2/hki2 for the same three
values of α and N = 1200. In all cases, β = 1 and hki = 10. . . . 107
C.1 Maximally packed matrix representing a network of plants and
islands off Perth (Abbott and Black, 1980) (because the net-
work is bipartite, the adjacency matrix is composed of four
blocks: two identical to this matrix, the other two composed
of zeros). Data, image and line obtained from NESTEDNESS
CALCULATOR, which returns a “temperature” of T = 0.69o
for this particular network.
. . . . . . . . . . . . . . . . . . . . 113
C.2 Nestedness against assortativity (as measured by Pearson’s cor-
relation coefficient) for scale-free networks as given by Eq. (C.13).
. . . . . . . . . . . . . . . . . . . . . . . . 119
hki = 10, N = 1000.
List of Figures
xxv
C.3 Nestedness against assortativity (as measured by Pearson’s cor-
relation coefficient) for data on a variety of networks. Blue
squares are food webs (Table C.4) and red circles are networks
of all other types (Table C.4).
. . . . . . . . . . . . . . . . . . . 119
List of Tables
3.1 Values of small-world parameters C and l, and Pearson’s cor-
relation coefficient r , as measured in the neural network of the
worm C. Elegans (White et al., 1986), and as obtained from sim-
ulations in the stationary state (t = 105 steps) for an equivalent
network at the critical point when edges are removed randomly
– i.e., for α = 1.35 and β = 1. N = 307, κst = 14.0; averages
over 100 runs and global probabilities as in Eq.
(3.6). The-
oretical estimates correspond to Eqs. (3.12), (3.14) and (3.11)
applied to the networks generated by the same simulations. The
last column lists the respective configuration model values: C
and l are obtained theoretically as in (Newman, 2003c), while
r , from MC simulations as in (Maslov et al., 2004), is the value
expected due to the absence of multiple edges. (See also Fig. 3.4.) 36
C.1 Food webs appearing in Fig. C.3 (listed from least to most
assortative) : r is the assortativity and ν the nestedness. The
origins of all data cited in Ref. (Dunne et al., 2004), and kindly
provided to us by Jennifer Dunne.
. . . . . . . . . . . . . . . . 120
C.2 Empirical networks appearing in Fig. C.3 (listed from least to
most assortative) : r is the assortativity and ν the nestedness.
All data available on the personal Web pages of ´Alex Arenas,
. . . . . . . . . . . . . . . . 121
Mark Newman and Duncan Watts.
xxvii
Chapter 1
Resumen en espanol
Paradigma de sistema complejo y el peor comprendido de nuestros ´organos, el
cerebro es, esencialmente, una inmensa red de c´elulas que se comunican entre
s´ı mediante senales electro-qu´ımicas. Este traba jo recoge y desarrolla ideas
del joven campo de las Redes Complejas para tratar de mejorar nuestro en-
tendimiento acerca del comportamiento colectivo complejo que puede emerger
en las redes de neuronas a partir de din´amicas individuales relativamente sen-
cillas.
El Cap´ıtulo 2 es una breve introducci´on a las Redes Complejas y a la
Neurociencia Computacional. Se describe, entre otras cosas, el modelo de
Hopfield de red neuronal atractora, en que cada nodo representa una neurona
y las sinapsis son representadas por los enlaces. Este sistema puede almacenar
informaci´on, en forma de patrones o configuraciones concretas de neuronas
activas e inactivas, en los pesos sin´apticos; es decir, en la intensidad con la
que la actividad de una neurona influye sobre sus vecinas. Si, para representar
un patr´on dado, dos neuronas vecinas han de adoptar el mismo estado (activo
o inactivo), se refuerza la interacci´on entre ambas, mientras que se debilita
en caso contrario. Repitiendo esta operaci´on para cada pareja conectada de
neuronas y para cada patr´on, estos patrones se convierten en los estados que
minimizan la energ´ıa total (atractores de la din´amica), y el sistema evolu-
ciona siempre hacia el patr´on que m´as se parezca a su estado inicial. Este
mecanismo, llamado de memoria asociativa, es la responsable del almacena je
y la recuperaci´on de informaci´on tanto en modelos m´as realistas de medios
neuronales, como en muchos aparatos artificiales que desempenan tareas tales
como la identificaci´on y la clasificaci´on de im´agenes. Adem´as, hoy en d´ıa
existen evidencias experimentales suficientes para asegurar que algo parecido
ocurre en el cerebro: mediante los procesos bioqu´ımicos de potenciaci´on de
1
2
Chapter 1. Resumen en espanol
largo plazo (LTP, por sus siglas en ingl´es) y depresi´on de largo plazo (LTD),
las sinapsis modifican gradualmente sus conductancias durante el aprendiza je.
El Cap´ıtulo 3 aborda el problema de c´omo puede desarrollarse una red con
el tipo de estructuras que se observa en el cerebro. Para ello se formaliza como
un proceso estoc´astico una red que evoluciona mediante cambios probabil´ısticos
que dependen de cualquier manera de informaci´on local y global de los grados
(n´umeros de vecinos) de los nodos, tal como se hace en la Ref. (Johnson et al.,
2010a). Se considera que estas suposiciones son relevantes para el caso del
cerebro ya que la arborizaci´on y la atrofia sin´apticas dependen de la actividad
el´ectrica de la neurona en cuesti´on, que a su vez puede estar relacionada con
el n´umero de vecinos que tenga, y con la densidad sin´aptica media en la red.
Se demuestra c´omo esta situaci´on viene descrita por una ecuaci´on de Fokker-
Planck, y se aplica a dos conjuntos de datos reales neurofisiol´ogicos: por una
parte, la curvas de poda sin´aptica (fuerte reducci´on de la densidad sin´aptica
que sufre el c´ortex durante la infancia) de autopsias humanas pueden explicarse
con unas suposiciones m´ınimas; por otra, varias magnitudes estad´ısticas de la
red del an´elido C. Elegans (distribuci´on de grados, perfil de correlaciones, clus-
tering o agrupamiento y camino m´ınimo medio) emergen con cierta precisi´on
y de manera natural justo en la transici´on de fase que presenta el modelo.
Esto da fuerza a la idea de que el sistema nervioso optimiza su rendimiento
coloc´andose cerca de un punto cr´ıtico. Un caso parecido, en que los enlaces
de la red, en vez de desaparecer o aparecer, son redirigidos estoc´asticamente,
presentado en la Ref. (Johnson et al., 2009b), se describe en el Ap´endice A.
El resto de la tesis se centra en los efectos que pueden tener sobre el com-
portamiento colectivo de sistemas de neuronas las caracter´ısticas topol´ogicas
descritas en el Cap´ıtulo 3. Se sabe que la heterogeneidad de la distribuci´on
de grados de la red suele tener una influencia significativa en la din´amica de
elementos conectados mediante sus enlaces. En el caso de redes neuronales de
Hopfield, Torres et al. (Torres et al., 2004) demostraron que, en redes libres
de escala (que son altamente heterog´eneas), el rendimiento aumenta con la
heterogeneidad. El Cap´ıtulo 4 examina el mismo efecto en una red neuronal
que incluye otro ingrediente biol´ogico: la depresi´on sin´aptica, gracias a la cual
se observa una transici´on entre una fase de memoria est´atica a otra en que el
sistema salta ca´oticamente entre los patrones guardados. Resulta que cerca
de este punto cr´ıtico (el famoso Borde del Caos) la red es capaz de realizar
una tarea din´amica necesaria para los seres vivos: reconocer, de entre varios
patrones que tenga almacenados, uno dado que se le “ensene” y retenerlo in-
3
definidamente despu´es. Como demostramos en la Ref. (Johnson et al., 2008),
la heterogeneidad de la distribuci´on de grados de la red acerca el punto cr´ıtico
a una regi´on del espacio de par´ametros en que apenas hay depresi´on sin´aptica.
Teniendo en cuenta que esta depresi´on empeora la capacidad de memoria del
sistema, se concluye que una red altamente heterog´enea es la ´optima para re-
alizar este tipo de tarea din´amica. Las redes funcionales medidas en el c´ortex
humano durante tareas del estilo adopta la red libre de escala m´as heterog´ena
posible, por lo que cabe la hip´otesis de que el cerebro est´e maximizando as´ı su
rendimiento.
Otra propiedad altamente estudiada de las redes complejas es la existencia
de correlaciones entre los grados de nodos vecinos. Cuando dichas correla-
ciones son positivas (nodos muy conectados se suelen conectar con otros que
tambi´en tienen muchos vecinos, y los que tienen pocos con otros parecidos) se
dice que la red es asortativa; mientras que es disasortativa si las correlaciones
son negativas (los que tienen muchas conexiones se conectan, preponderante-
mente, con los que tienen pocas). Curiosamente, se hab´ıa observado que por
lo general las redes sociales (por ejemplo, redes de colaboraciones profesionales
o de contactos sexuales) son asortativas, mientras que pr´acticamente todas las
dem´as (gen´eticas, tr´oficas, proteicas, de transportes, de palabras, Internet, la
Web...) son significativamente disasortativas. Aunque se hab´ıa estudiado los
efectos de estas correlaciones en varios sistemas, las t´ecnicas matem´aticas y
computacionales para ello padec´ıan de inconvenientes que restaban general-
idad a los resultados. Para solventar esto, en el Cap´ıtulo 5 se describe un
nuevo m´etodo para particionar el espacio de las fases de redes en regiones de
correlaciones iguales, una t´ecnica que permite tanto an´alisis te´orico como com-
putacional de este tipo de sistemas. Utilizando este m´etodo junto con ideas
de Teor´ıa de la Informaci´on se demuestra tambi´en el resultado principal de la
Ref. (Johnson et al., 2010b): que la disasortatividad es el estado “natural”
(en cuanto a situaci´on de equlibrio) de las redes heterog´eneas, lo cual explica
la preponderancia en la realidad de este tipo de configuraciones. La prefer-
encia de los humanos por agregarse en funci´on de propiedades similares ser´ıa
la explicaci´on de que las redes sociales se encuentren fuera del equilibrio, en
regiones asortativas del espacio de fases.
En el Cap´ıtulo 6 se aplica el m´etodo del Cap´ıtulo 5 al caso de una red neu-
ronal de Hopfield que no s´olo presenta heterogeneidad, sino tambi´en correla-
ciones nodo-nodo. Se encuentra, como ya fue descrito en la Ref. (de Franciscis et al.,
2011), que estos sistemas pueden aumentar de manera notable su robustez
4
Chapter 1. Resumen en espanol
frente a ruido gracias a las correlaciones positivas. De nuevo, esto parece
enca jar, al menos cualitativamente, con resultados experimentales que han
encontrado redes funcionales en el c´ortex humano altamente asortativas.
Hemos dicho que las redes neuronales pueden aprender gracias a una apropi-
ada modificaci´on de los pesos sin´apticos mediante LTP y LTD, lo que explica
la memoria de largo plazo. Pero estos procesos bioqu´ımicos ocurren en un
tiempo caracter´ıstico de al menos minutos. Los modelos de memoria de corto
plazo, que ocurren en escalas de tiempo menores, suelen dar por hecho que
la informaci´on que se utiliza est´a de antemano almacenada en el cerebro, y
que el sujeto realizando la tarea s´olo ha de activar y mantener de alguna
manera la configuraci´on correcta (como en el Cap´ıtulo 4). Pero es f´acil darse
cuenta de que esto no puede ser el caso para cualquier tarea: basta mirar
algo totalmente nuevo por un instante, cerrar los o jos, y pensar en lo que
se ha visto. Los ´unicos modelos de memoria de corto plazo existentes que
no requieren aprendiza je sin´aptico se basan en que cada neurona mantenga
de alguna manera la informaci´on que le corresponde (t´ıpicamente gracias a
una serie de procesos sub-celulares). Pero al no emerger la memoria como
propiedad colectiva del sistema, sino como suma de memorias individuales,
estos modelos padecen de una gran falta de robustez frente a ruido. Y, lejos
de presentar un comportamiento individual fiable, las neuronas se caracterizan
justamente por ser c´elulas de una alta variabilidad, con tendencia a disparar
de manera m´as o menos aleatoria. En el Cap´ıtulo 7 se propone un mecanismo,
llamado Cluster Reverberation (CR), o Reverberaci´on de Grupo, gracias al cual
incluso sistemas como redes con unidades simples, binarias, como en el modelo
de Hopfield pueden almacenar informaci´on instant´aneamente sin necesidad de
aprendiza je sin´aptico, y de una manera que puede ser todo lo robusto frente
a ruido como se quiera (Johnson et al., 2011). Para ello el sistema aprovecha
la existencia de estados metastables (situaciones que minimizan la energ´ıa del
sistema localmente, sin corresponder al m´ınimo global) y como consecuencia
aparecen transitorios en la dinmica de la actividad neuronal cuyas propiedades
son consecuencia inmediata de las caracter´ısticas de la topolog´ıa subyacente y
que es del tipo de las descritas anteriormente en el Cap´ıtulo 3 y en experimen-
tos, esto es, el grado de agrupamiento o la modularidad de la red. B´asicamente,
grupitos densamente interconectados de neuronas pueden mantener un estado
conjunto de alta o ba ja actividad, en promedio. Considerando cada grupito
como un elemento funcional elemental, en vez de cada neurona, se consigue la
aparici´on de las propiedades requeridas. Es m´as, algunas otras caracter´ısticas
5
de la memoria de corto plazo emergen de manera natural de este mecanismo.
En particular, se demuestra que la informaci´on se pierde gradualmente con el
tiempo seg´un una ley aproximadamente potencial, como ha sido descrito en
experimentos psicof´ısicos.
En conclusi´on, las principales aportaciones originales de esta Tesis son:
• M´etodos anal´ıticos y computacionales para estudiar redes evolutivas (Johnson et al.,
2009b, 2010a) y redes con correlaciones nodo-nodo (Johnson et al., 2010b;
de Franciscis et al., 2011).
• Una respuesta a la pregunta de por qu´e la mayor´ıa de las redes reales
son disasortativas (Johnson et al., 2010b).
• Propiedades topol´ogicas que pueden optimizar el rendimiento de redes
neuronales (Johnson et al., 2008; de Franciscis et al., 2011).
• Un mecanismo que pudiera estar detr´as de la memoria de corto plazo
(Johnson et al., 2011).
Chapter 2
Where we are and where we’d
like to go
2.1 From bridges to brains
Strolling through the streets of Konigsberg, a young Immanuel Kant may have
wondered whether, as some hoped, a path could be found that would take him
once and only once over each of the city’s celebrated seven bridges and back to
where he started. In 1736 Leonard Euler pointed out that for this or any other
problem of the kind all that mattered was which land masses were connected
to each other, and by how many bridges. In other words, the situation could
be captured by a graph, as in Fig. 2.1, in which each land mass is represented
by a node (also called vertex) and each bridge by a link (or edge). He showed
that in the case of Konigsberg no such walk could be found, since an “Eulerian
cycle” in a connected graph exists if and only if the degrees of all nodes are
even numbers – the degree of a node being its number of edges (Euler, 1736).
And thus was Graph Theory born.
For over two centuries, the graphs people were interested in were precisely
defined ob jects, usually sufficiently small to be drawn on a piece of paper. But
in the late nineteen fifties, mathematicians began to study random graphs –
i.e., defined only by some random generation process – perhaps with a view to
better dealing with ever-growing communications networks (Bollob´as, 2001).
E. N. Gilbert considered a situation in which there are n nodes and each pair
is connected by an edge with probability q (Gilbert, 1959). For different values
of these parameters, he was able to obtain the likelihood of the graph being
connected (that is, of there being a path joining any two nodes). A similar
model was proposed by Paul Erdos and Alfr´ed R´enyi: each of all the possi-
7
8
Chapter 2. Where we are and where we’d like to go
Figure 2.1: The problem of the Seven Bridges of Konigsberg can be reduced
to a graph in which nodes and edges represent land masses and bridges, re-
spectively.
ble graphs with n nodes and m edges had an equal chance of being “picked”
(Erdos and R´enyi, 1959). In fact, a given graph will be generated with equal
probability in either scenario, so the descriptions are equivalent, and usually
known as the Erdos-R´enyi (ER) model. It is simple to see that if one were to
average over many graphs generated by either of these processes, the degrees
would follow a binomial distribution – tending, for large n, to a Poisson dis-
tribution. That is, p(k) is symmetrically centred around its mean value and
drops off exponentially – where ki is the degree of node i. An interesting phe-
nomenon that can be observed using the ER model is that of percolation. If we
measure the size Φ of the largest connected component (that is, of the highest
number of nodes in the graph forming a connected subgraph) we obtain at
different values of the probability q (or, equivalently, of m = 1
2 qn2), we see
that there is a critical value, qc = 1/n, above which Φ/n does not vanish for
high n – that is, there will usually exist a connected subgraph of a size compa-
rable to that of the whole system. This passing from one situation (or phase)
to another, each characterised by some qualitatively different characteristic, is
known in physics as a phase transition. In this case, it is a second-order tran-
sition, since the control parameter Φ varies continuously (and not abruptly, as
in first-order transitions), and has innumerable applications. For instance, the
nodes might be people susceptible to some disease, trees which may be set on
fire, or oil bubbles in a porous medium. The epidemic will spread, the forest
will burn, or the oil will be extractable if the density of edges – contagious
contact, fire-conducting proximity, or links between pores – is over the critical
2.1 From bridges to brains
9
value.
In his 1929 short story Chains (L´ancszemek, in the original Hungarian)
Frigyes Karinthy suggested that the number of people in a chain of acquain-
tances grows exponentially with size, and thus that very few steps are needed
to join anyone with any other person. This Small World idea was taken up in
1967 by Stanley Milgram, who performed a series of experiments that, while
somewhat less controversial than his well-known obedience-to-authority explo-
rations, have nonetheless been widely discussed (Milgram, 1967). He and his
colleagues sent various letters to random people with the request to attempt
to send them on to a particular individual many thousands of miles away, plus
the constraint that this had to be done via people with whom the sender was
on first-name terms. Many of the letters reached their destination, after having
been sent on by a surprisingly small number of intervening people. This was
later popularised as the Six Degrees of Separation – the famous idea that any
two people are linked by a path of only six acquaintances. Within the con-
nected component of an ER random graph any two nodes are also joined by a
short path – of the order of the logarithm of the number of nodes. However,
this is less surprising, since these networks are not clustered; that is, they do
not have the property typical of social networks whereby “the friends of my
friends are (also likely to be) my friends.” In 1998, Duncan Watts and Steven
Strogatz put forward a network model which took this feature into account.
They considered a ring of n nodes, each connected to their k nearest neigh-
bours (they set k = 4). Each edge was then broken from one of its nodes and
rewired to some other random node with a probability p. Thus, p = 0 leaves
the ring intact, while p = 1 changes it into an ER random graph. Two magni-
tudes were measured for different values of p: the mean-minimum-path length,
l, and the clustering coefficient, C . The first is simply an average over all pairs
of nodes of the minimum-paths connecting them. The latter can be seen as
the probability that two neighbours of a given node are directly connected to
each other. For p = 0, the clustering is high (C = 1
2 ) and independent of
the network size, while the mean minimum path scales with n (l ≃ n/8). At
the other extreme, p = 1 yields a vanishingly small clustering (C = k/n) but
short paths (l ≃ ln n/ ln k). The most interesting case is found at intermediate
values of p. As p grows from zero, l falls very rapidly to a value close to the
random case, but C does not present this drop until a much higher value is
reached. Watts and Strogatz called this intermediate zone the Small-World
region, since everyone is highly-interconnected while much of the local struc-
10
Chapter 2. Where we are and where we’d like to go
ture is conserved. They suggested that this is actually a property of many
real networks (as has since turned out to be the case (Newman, 2003c)), most
especially of social networks – in which C is often several orders of magnitude
greater than if the graph were random, while l is not much larger than in such
a case. As the authors point out, it is essential to take this feature into account
for the study of, say, epidemics.
Another feature of networks which is quite ubiquitous in the real world is
that degree distributions are highly heterogeneous; in fact, they often follow
power-laws, p(k) ∼ k−γ , with γ a positive constant typically between 2 and 3.
Such networks are nowadays referred to as scale free. In the nineteen fifties,
Herbet Simon showed that these distributions come about when “the rich get
richer” (Simon, 1955). Applying this idea to the case of scientific citations,
Derek de Solla Price proposed the first known model of a scale-free network,
in which nodes represent papers and edges are citations (de S. Price, 1965).
Each node has an in-degree (the number of papers citing it) and an out-degree
(papers it cites). That is, the network is directed, since edges have a direc-
tion. Assuming that the probability a paper has of being cited by a new one
is proportional to the number already citing it (its in-degree), the network is
built up through the gradual addition of nodes, the neighbours of these be-
ing chosen according to their existing in-degrees. Price showed analytically
that such a mechanism leads to an in-degree distribution p(k) ∼ k−(2+1/m) ,
with m a parameter of the model equivalent to the mean degree1 . He called
this mechanism cumulative advantage. Somewhat ironically – considering that
Price, with a PhD in history from Cambridge, is best known as the father
of scientometrics – this work was mostly ignored by the scientific community.
The model was rediscovered in 1999 by Albert-L´aszl´o Barab´asi and Reka Al-
bert, with the difference that they considered the network to be undirected
(Barab´asi and Albert, 1999). They coined the term preferential attachment
for the rich-get-richer mechanism, which is now generally assumed to be be-
hind the formation of most scale-free networks (although other mechanisms
exist (Caldarelli et al., 2002; Krapivsky et al., 2000; Newman, 2005)). Among
many interesting consequences of such degree heterogeneity, Mark Newman
showed that the clustering and mean-minimum-path length are respectively
higher and lower than in homogeneous networks, making all scale-free net-
works to some extent small worlds (Newman, 2003b). It also has important
1Note that in a directed network, the mean in-degree and mean out-degree coincide.
2.1 From bridges to brains
11
consequences for dynamical processes taking place among elements on the net-
work, such as the synchronisation of coupled oscillators (Barahona and Pecora,
2002).
As mentioned above, networks can be made up of separate components
such that no path exists between nodes in different subgraphs. This is an
extreme case of community structure. However, what is usually more inter-
esting is the fact that communities may exist such that there is a higher
density of edges within them than without, even if the network is connected
(Girvan and Newman, 2002). These communities are also at times called mod-
ules or clusters (although this can create some confusion with the related but
distinct idea of clustering referred to above). Given a network, one can make
a partition – that is, divide the nodes up into groups – and calculate what
proportion of the edges fall within these, compared with the random expecta-
tion. This measure is called the modularity of this partition, and sometimes
one speaks of the modularity of a graph referring to that of the partition
for which this value is maximum. Determining the community structure of
empirical networks can often provide useful insights into aspects of the sys-
tems. For instance, the communities may correspond to functional groups
in a metabolic network, or groups of people who share some trait. How-
ever, there are many problems related to making this kind of measurement.
For one thing, there are so many possible partitions that even an ER ran-
dom graph can have a fairly high modularity due simply to statistical fluctua-
tions (Reichardt and Bornholdt, 2006). Then there is the fact that community
structure can exist on may different levels – that is, the groups considered
can be of any size – so one must usually consider a hierarchy of modules
(Arenas et al., 2006). Furthermore, finding an optimal partition is an NP-
Complete problem (Garey and Johnson, 1979), which makes comparing the
modularity of each possible partition intractable in all but very small net-
works. For these and other reasons, in recent years much work has gone into
finding efficient algorithms to determine the community structure of networks,
albeit approximately (Girvan and Newman, 2002; Donetti and Munoz, 2004;
Blondel et al., 2008) – as well as into comparing the results offered by each
approach (L. Danon et al., 2005).
Finally, another feature of networks worth mentioning arises when the
nodes of a network are endowed with some property and this is reflected by
the layout of the edges: the situation is called a mixing pattern (Newman,
2002, 2003a). For instance, people might tend to choose sexual partners who
12
Chapter 2. Where we are and where we’d like to go
share certain characteristics, such as mother-tongue or self-defined race.
In
these cases the network is assortative, since nodes of a kind assort, or group
together. However, if the property in question were, say, gender, then the same
graphs would be disassortative if most of the relations were heterosexual. In
these cases the property can be considered discrete, but it can be continuous
– for instance, people might assort according to age. An interesting case is
when the property in question is the degree of each node, since it is then an
entirely topological issue. The extent to which the degrees of neighbouring
nodes are correlated – as given, for example, by Pearson’s correlation coeffi-
cient (Newman, 2002) – is then a measure of the assortativity of the graph,
being positive for assortative networks and negative for disassortative ones.
It turns out that there is a striking universality in the nature of the degree-
degree correlations displayed by real-world networks, whether natural or artifi-
cial: social networks, like the ones just described, tend to be assortative, while
almost all other kinds of network are disassortative (Pastor-Satorras et al.,
2001; Newman, 2003c). Often specific functional constraints can be found to
justify correlations of one or other kind, but in Chapter 5 of this thesis the
purely topological explanation put forward in Ref. (Johnson et al., 2010b) is
described. In any case, the degree-degree correlations of a network can play a
significant role in the behaviour of processes taking place thereon. For exam-
ple, assortative networks have lower percolation thresholds and are more robust
to targeted attack (Newman, 2003a), while disassortative ones make for more
stable ecosystems and seem to be more synchronizable (Brede and Sinha).
All the aspects of networks mentioned in this brief overview, as well as many
others, have been shown to be relevant for a wide range of complex systems
(Albert and Barab´asi, 2002; Newman, 2003c; Boccaletti et al., 2006). Among
these is the brain, a paradigm of complexity as well as the least understood
of our organs. However, research focusing on how the collective behaviour
of neural systems, as observed in mathematical models, is influenced by the
topology of the underlying network is relatively scarce. This is perhaps due in
part to the attention that other biological properties of the nervous system have
tended to draw from the Computational Neuroscience community. Thanks to
the flurry of activity that the field of Complex Networks has been enjoying
over the last decade, this is a particularly good moment to undertake a more
systematic analysis of how dynamics and topology are related in this kind of
systems.
2.2 Neural networks in neuroscience
13
2.2 Neural networks in neuroscience
Ever since the publication of Santiago Ram´on y Ca jal’s drawings of neurons –
in his words, those “mysterious butterflies of the soul” – it has been clear that
the nervous system is composed of a large number of such cells connected to
one another to form a network (y Ca jal, 1995). Long axons, ending in termi-
nals which form synapses to the dendrites which branch out from neighbouring
neurons, transmit action potentials (APs) – changes in the cellular membrane
voltage – and enable neurons somehow to cooperate and give rise the aston-
ishing emergent phenomenon called thought. One of these APs is formed each
time the membrane electric potential of a neuron surpasses a threshold value,
leading to the opening of a great many voltage-dependent ionic gates between
the cell and the extra-cellular medium. In turn, the membrane potential of a
given neuron is constantly affected by action potentials arriving from neigh-
bouring neurons, and thus an extremely complex web of cellular signalling is
achieved.
The first model neuron was proposed by McCulloch and Pitts (1943). This
was simply an element that would return a Heaviside step function of the sum
of its inputs. Sets of such “artificial neurons” could be used to implement
any logical gate. Shortly after this, another important suggestion was made,
this time by the psychologist Donald Hebb. Attempting to relate Pavlovian
conditioning experiments with cellular plasticity, he conjectured, in 1949, the
existence of some biological mechanism that would lead to neurons which re-
peatedly fired (i.e., let off action potentials) together becoming more strongly
coupled (Hebb, 1949). The initiation and propagation of action potentials
in individual neurons was first modelled mathematically by Alan L. Hodgkin
and Andrew Huxley in 1952 by means of a set of nonlinear ordinary differential
equations which took into account the various ion fluxes (Hodgkin and Huxley,
1952).
However, the concept of a neural network (as understood in theoretical
and computational neuroscience) was partly inspired by mathematical models
of spin systems. The first of these was the Ising model, put forward in 1920
by Wilhelm Lenz and studied by Ernst Ising with a view with a view to un-
derstanding phase transitions and magnets (Onsager, 1944; Brush, 1967). It
was known that the spin of electrons conferred a magnetic moment to indi-
vidual atoms, but it wasn’t clear how exactly a very many such spins could
self-organise into a large body producing a net magnetic field. By considering
an infinite set of entities (spins) with possible values plus or minus one (up
14
Chapter 2. Where we are and where we’d like to go
Figure 2.2: Drawing of the cells of the chick cerebellum, from “Estructura de
los centros nerviosos de las aves”, Madrid, 1905. Notice how the neurons make
up a complex network of synaptic interactions.
or down, say) which, when placed at the nodes of a lattice, interact in such a
way that energy is lowest when neighbours are aligned, and a temperature pa-
rameter to govern the extent of random fluctuations, it was eventually shown
that, below a certain critical temperature (in two or more dimensions), sym-
metry is spontaneously broken and most of the spins end up aligned (Baxter,
1982). This ferromagnetic solution comes about and is then maintained be-
cause it has a lower energy than any other configuration of spins. Subsequent
models, in particular that of Sherrington and Kirkpatrick (1975), incorporated
inhomogeneities in the coupling strengths such that there was no longer a con-
figuration which simultaneously minimized all interaction energies, leading to
frustrated states (spin-glasses).
These ideas were put together, by Amari (1972) and then by Hopfield
(1982), in the first neural network models to exhibit the mechanism known as
associative memory. Each model neuron was placed at the node of a network,
originally assumed to be fully connected (all nodes connected to all the rest),
and followed a dynamics which can be seen either as that of Ising spins or of
McCulloch-Pitts neurons. However, a noise parameter usually referred to as
“temperature” by analogy with spin systems could be included to allow for
non-deterministic behaviour. By setting the interaction strengths (synaptic
weights) not randomly, as in the Sherrington-Kirkpatrick model, but according
2.2 Neural networks in neuroscience
15
to the Hebb rule referred to above, information could be stored and retrieved
by the system. More specifically, a set of particular patterns, or configurations
of positive and negative elements (firing and non-firing neurons), are recorded
in the following way: for each pattern, one looks at each pair of neurons and
adds a quantity to the weight of the synapse joining them if the pattern in
question requires them to be in the same state, and subtracts it when they
should be opposite. In this way, the minimum energy configurations correspond
to the stored patterns, which therefore become attractors of the dynamics: if
the temperature is not too heigh to destroy all order, the system will evolve
towards whichever of these patterns most resembles the initial configuration
it is placed in. Figure 2.3 illustrates how this mechanism works for a system
such that the firing and non-firing neurons represent black and white pixels of
a bitmap image.
Thanks to associative memory, if we were to store, say, a set of photos of
various people and then “show” the network a different picture of one of the
same sub jects, it would be able to retrieve the correct identity. Not only is this
mechanism used nowadays in technology capable of performing tasks such as
pattern discrimination and classification, but it is widely considered to underlie
our own capacity for learning and recalling information. There is evidence
from neuronal readouts that this is so (Amit, 1995), and not long ago, in vivo
experiments finally established that learning is indeed related to the processes
of long term potentiation (LTP) and depression (LTD) – by which synapses
between neurons that fire nearly simultaneously gradually increase or decrease
their conductance depending on the interval of time elapsed (Gruart et al.,
2006; Roo et al., 2008).
The neural network models studied nowadays generally include more real-
istic dynamics both for the neurons and for the synapses, taking into account
a variety of cellular and subcellular processes (Amit, 1989; Torres and Varona,
2010). For example, the fact that the conductance of synapses in reality de-
pends on their workload has been found to enable a network to switch from
one pattern to another – either spontaneously or as a reaction to sensory stim-
uli – providing a means for the performance of dynamic tasks (Cortes et al.,
2006; Holcman and Tsodyks, 2006); this result also seems to agree well with
physiological data (Korn and Faure, 2003). In fact, there is evidence that the
brain somehow maintains itself close to a boundary – called, in physics, a
critical point – between an ordered and a chaotic regime (Egu´ıluz et al., 2005;
Chialvo, 2004; Chialvo et al., 2008; Bonachela et al., 2010; Torres and Varona,
16
Chapter 2. Where we are and where we’d like to go
Figure 2.3: In the Hopfield neural-network model, the interaction strengths
(representing synaptic weights) store information in the form of particular
patterns, or configurations of firing neurons, which become attractors of the
dynamics. Whatever the initial state of the system, it will always evolve to-
wards one of these patterns, thus allowing for the storage and retrieval of
information. The mechanism, known as associative memory, is thought to be
at the basis of memory in the brain. In this case, the network is “remember-
ing” an illustration by Jean-Baptiste Oudry for Jean de la Fontaine’s fable La
Cigale et la Fourmi.
2010). This would be in line with research that shows how certain useful prop-
erties – such as the computational capacity of some neural-network models
(Bertschinger and Natschlager, 2004), or the dynamic range of sensitivity to
stimuli in sensory systems (Kinouchi and Copelli, 2006) – are optimised at this
“edge of chaos (Chialvo, 2006)”.
That these models should actually reflect, albeit in an enormously simpli-
fied way, what actually goes on in our brains tends to fit in quite well with
intuitive expectations – to the extent that so-called connectionist models seem
to be gradually becoming the accepted paradigm in relevant areas of psychol-
ogy and philosophy (Marcus and G.F., 2001; Frank, 1997). For instance, from
2.2 Neural networks in neuroscience
17
this point of view the way in which the recollection of a particular detail often
evokes, almost instantly, a whole landscape of sensations and emotions makes
sense, since these concepts will have been stored in some way as the same pat-
tern. Also, the fact that new memories are recorded in synapses which were
already being used to store previous information would seem to explain why
memories tend to fade slowly with time, yet can still be recalled, at least to
some extent, when a particular thought in some sense overlaps with (reminds
one of ) one of them. When this happens, the old memory springs to mind and,
if held there for long enough, can be refreshed via long term potentiation and
depression – although interaction with other patterns or with current stimuli
may well modify the refreshed information. Similarly, previous information
influences the storing of new memories, leading to the well known fact that we
tend to “see” things the way we expect them to be.
It seems, then, that the basic mechanisms behind the ability of our brains
to remember things, at least when the information is stored slowly enough for
the biochemical processes of LTP and LTD to be at work (long-term memory),
are now understood. Not only are the implications of such knowledge far-
reaching in themselves. The way in which it was developed is also particularly
notworthy. More or less sketchy ideas from areas as diverse as behavioural
psychology, neurophysiology and theoretical physics were brought together in
order to come up with a minimal mathematical model capable of manifesting
the sought-after phenomenon of information retrieval as a consequence of the
known properties of a great many simple elements. This kind of research can
at first seem more like a mathematical game than anything to do with nitty-
gritty reality. But the fact the basic mechanism of associative memory has
since borne up to decades of experimenting and theoretical probing reveals
how insightful it can actually be.
It is likely that other features of brain
function – short-term memory, information processing or emotional tagging,
to name but the first few that spring to mind – will eventually be thrown under
a similar light. In fact, we can expect the nature of even such an elusive and
intimate phenomenon as consciousness some day to become clear. After all,
the explanations behind other emergent properties of matter which in their day
seemed almost mystical, such as temperature or life, are now well understood.
18
Chapter 2. Where we are and where we’d like to go
2.3 A declaration of intent
As Zora Neale Hurston put it, “Research is formalized curiosity. It is poking
and prying with a purpose.” But there are many possible purposes, and even
more different ways of poking and prying. The motivation behind the work
presented here is to understand how the phenomena we observe in certain
systems on a macroscopic scale can come about from interactions of their
many relatively simple constituent elements. In the case of neural systems, it
seems reasonable to assume that these basic elements are neurons, and that
it is thanks to the cooperation of a great many of these cells that such organs
are able to think and feel. The human brain – with about 100 billion neurons
connected by 100 trillion synapses – being the most complex system we know
of, an enormous degree of simplification will be required for our description to
be of any use to this purpose. (In fact, if we could somehow simulate a brain
in all detail, the result would be just as unfathomable as the original ob ject,
however exciting the activity may prove for other reasons.) The physiology of
the neuron is nowadays quite well understood. However, just as the properties
of atoms or transistors that are key to understanding phase transitions or the
workings of a microchip are, respectively, magnetic interactions and voltage-
dependent gating, we must try to ascertain exactly which neuronal features are
necessary for the macroscopic behaviour we are interested in to occur. One
way to do this is to start by considering only the most basic characteristics
and explore what non-trivial phenomena emerge from these, allowing us then
to add new ingredients one at a time to pinpoint the relevant ones. In this
line, we consider large sets of Hopfield’s simple binary model neurons to study
how network properties are related to collective behaviour.
This work is laid out as follows. Chapter 3 deals with development. The
appearance and disappearance of edges in a network (growth and death of
synapses, in the case of the brain) is formalised as a stochastic process and
studied in a general setting (Johnson et al., 2009b, 2010a). It turns out that
many of the topological features observed in experiments are well modelled in
this way – which to some extent justifies, a posteriori, our initial assumptions.
The following chapters describe particular phenomena that emerge as a di-
rect consequence of some of those topological features: degree heterogeneity in
conjunction with synaptic depression improves the performance of dynamical
tasks (Johnson et al., 2008) (Chapter 4); assortativity serves to enhance a neu-
ral network’s robustness to noise (de Franciscis et al., 2011) (Chapter 6); and
clustering or modularity can lead to metastable states with certain properties
2.3 A declaration of intent
19
essential for some short-term memory abilities (properties hitherto lacking in
previous models) (Johnson et al., 2011) (Chapter 7). Thanks to the extreme
simplicity of the basic elements we are considering, we are able not only to
simulate but also to understand mathematically how exactly the interesting
phenomena emerge. This makes it possible to predict, to some extent, which
extra ingredients wil l not invalidate the results if they are taken into account
explicitly.
Some of the work has a more general scope than the study of neural net-
works. In particular, the equations obtained in Chapter 3 can be applied to any
network that evolves under the influence of probabilistic addition and deletion
of edges. And the method put forward in Chapter 5 for the study of correlated
networks can be used not just for analysing particular models, as we go on
to do in Chapter 6, but to solve many other problems – such as that of the
ubiquity of disassortative networks in nature and technology (Johnson et al.,
2010b), or how the property of nestedness typical of ecosystems is related to
other topological characteristics (c.f. Appendix C).
To sum up, the aim of the thesis is to shed light on how cellular dy-
namics can lead to the complex network structures of neural systems,
and, in its turn, in what ways this topology can influence, optimise
and determine the collective behaviour of such systems.
The main contributions made are:
• An analytical method to study the evolution of networks governed by a
combination of local and global stochastic rules.
• A mathematical and computational technique for the study of correlated
networks in a model-independent way.
• Possible biological justifications for two non-trivial features of the topol-
ogy of the human cortex: heterogeneity of the degree distribution and
high assortativity.
• An answer to the long-standing question of why most networks are dis-
assortative.
• Cluster Reverberation: the first mechanism proposed which would allow
neural systems to store information instantaneously in a robust manner.
Chapter 3
Evolving networks and the
development of neural systems
The highly heterogeneous degree distributions of most empirical networks is
assumed in many cases to arise from some form of cummulative advantage,
or preferential attachment. However, the origin of various other topological
features is often not clear and attributed to specific functional requirements.
We show how it is possible to analyse a very general scenario in which nodes
gain or lose edges according to arbitrary functions of local and/or global degree
information. Applying our method to two rather different examples of brain
development – synaptic pruning in humans and the neural network of the worm
C. Elegans – we find that simple biologically motivated assumptions lead to
very good agreement with experimental data. In particular, many nontrivial
topological features of the worm’s brain arise naturally at a critical point.
3.1 Introduction
The conceptual simplicity of a network – a set of nodes, some pairs of which
connected by edges – often suffices to capture the essence of cooperation in
complex systems. Ever since Barab´asi and Albert presented their evolving
network model (Barab´asi and Albert, 1999), in which linear preferential at-
tachment leads asymptotically to a scale-free degree distribution (the degree,
k , of a node being its number of neighbouring nodes), there have been many
variations or refinements to the original scenario (Albert and Barab´asi, 2000;
G. Bianconi and Barab´asi, 2001; Krapivsky et al., 2000; Bianconi and Barab´asi,
2001; Park et al., 2005; Ree, 2007) (see also the review by Boccaletti et al.
(2006)). In Ref. (Johnson et al., 2009b), we show how topological phase tran-
21
Chapter 3. Evolving networks and the development of neural
systems
22
sitions and scale-free solutions can emerge in the case of nonlinear rewiring
in fixed-size networks, and this work is summarized in Appendix A. In Ref.
(Johnson et al., 2010a) we extend our scope to more general and realistic situa-
tions, considering the evolution of networks making only minimal assumptions
about the attachment/detachment rules. In fact, all we assume is that these
probabilities factorize into two parts: a local term that depends on node de-
gree, and a global term, which is a function of the mean degree of the network.
This is the work described in this chapter.
Our motivation can be found in the mechanisms behind many real-world
networks, but we focus, for the sake of illustration, on the development of bio-
logical neural networks, where nodes represent neurons and edges play the part
of synaptic interaction (Amit, 1989; Sporns et al., 2004; Torres and Varona,
2010). Experimental neuroscience has shown that enhanced electric activity
induces synaptic growth and dendritic arborization (Lee et al., 1980; Frank,
1997; Klintsova and Greenough, 1999; Roo et al., 2008). Since the activity
of a neuron depends on the net current received from its neighbours, which
tends to be higher the more neighbours it has, we can see node degree as a
proxy for this activity – accounting for the local term alluded to above. On
the other hand, synaptic growth and death also depend on concentrations of
various molecules, which can diffuse through large areas of tissue and there-
fore cannot in general be considered local. A feature of brain development
in many animals is synaptic pruning – the large reduction in synaptic den-
sity undergone throughout infancy. Chechik et al. (1999, in press) have shown
that via an elimination of less needed synapses, this can reduce the energy
consumed by the brain (which in a human at rest can account for a quarter
of total energy used) while maintaining near optimal memory performance.
Going on this, we will take the mean degree of the network – or mean synaptic
density – to reflect total energy consumption, hence the global terms in our
attachment/detachment rules (Johnson et al., 2009a).
An alternative approach would be to consider some kind of model neu-
rons explicitly and couple the probabilities of synaptic growth and death to
neuronal dynamic variables, such as local and global fields.
In an Amari-
Hopfield network, for example, the expected value of the field (total incoming
current) at node i is proportional to i’s degree (Torres et al., 2004), the to-
tal current (energy consumption) in the network therefore being proportional
to the mean degree; qualitatively, these observations are likely to hold also
in more realistic situations (Magistretti, 2009), although relations need not
3.2 Basic considerations
23
be linear. Co-evolving networks of this sort are currently attracting atten-
tion, with dynamics such as Prisoner’s Dilemma (Poncela et al., 2008), Voter
Model (Vazquez et al., 2008) or Random Walkers (Antiqueira et al., 2009).
Although we consider this line of work particularly interesting, for generality
and analytical tractability we opt here to use only topological information for
the attachment/detachment rules, although our results can be applied to any
situation in which the dynamical states of the elements at the nodes can be
functionally related to degrees1 .
Following a brief general analysis, we show how appropriate choices of func-
tions induce the system to evolve towards heterogeneous (sometimes scale-free)
networks while undergoing synaptic pruning in quantitative agreement with
experiments. At the same time, degree-degree correlations emerge naturally,
thus making the resulting networks disassortative – as tends to be the case for
most biological networks – and leading to realistic small-world parameters.
3.2 Basic considerations
Consider a simple undirected network with N nodes defined by the adjacency
matrix a, the element aij representing the existence or otherwise of an edge
between nodes i and j . Each node can be characterized by its degree, ki =
Pj aij . Initially, the degrees follow some distribution p(k , t = 0) with mean
κ(t). We wish to study the evolution of networks in which nodes can gain or
lose edges according to stochastic rules which only take into account local and
global information on degrees. So as to implement this in the most general
way, we will assume that at every time step, each node has a probability of
gain
gaining a new edge, P
, to a random node; and a probability of losing a
i
gain
randomly chosen edge, P lose
. We assume these factorize as P
= u(κ)π(ki)
i
i
and P lose
= d(κ)σ(ki), where u, d, π and σ can be arbitrary functions, but
i
impose nothing else other than normalization.
For each edge that is withdrawn from the network, two nodes decrease in
degree:
i, chosen according to σ(ki ), and j , a random neighbour of i’s; so
there is an added effective probability of loss kj /(κN ). Similarly, for each edge
placed in the network, not only l chosen according to π(kl ) increases its degree;
1For instance, the stationary distribution of walkers used for edge dynamics by
Antiqueira et al. (2009) is actually obtained purely from topological information, although
it can only be written in terms of local degrees for undirected networks.
Chapter 3. Evolving networks and the development of neural
systems
24
a random node m will also gain, with the consequent effective probability
N −1 (though see2 ). Let us introduce the notation π(k) ≡ π(k) + N −1 and
σ(k) ≡ σ(k) + k/(κN ). Network evolution can now be seen as a one step
process (van Kampen, 1992) with transition rates u(κ) π(k) and d(κ) σ(k). The
expected value for the increment in a given p(k , t) at each time step (which we
equate with a temporal derivative) defines a master equation for the degree
distribution (Johnson et al., 2009b):
dp(k , t)
dt
= u (κ) π(k − 1)p(k − 1) + d (κ) σ(k + 1)p(k + 1)
(3.1)
(3.2)
∂pst (k)
∂k
− [u (κ) π(k) + d (κ) σ(k)] p(k , t).
Assuming now that p(k , t) evolves towards a stationary distribution, pst (k),
then this must necessarily satisfy detailed balance since it is a one step process
(van Kampen, 1992); i.e., the flux of probability from k to k + 1 must equal the
flux from k + 1 to k , for all k (Marro and Dickman). This condition (sufficient
for Eq. (3.1) to be zero) can be written as
σ(k + 1) − 1(cid:21) pst(k),
= (cid:20) u(κst)
π(k)
d(κst)
where we have substituted a difference for a partial derivative and κst ≡
Pk kpst(k). Setting π and σ so as to be normalized to one (i.e., Pk p(k)π(k) =
Pk p(k)σ(k) = 1, ∀t), which is equivalent to saying that at each time step ex-
actly u(κ) nodes are chosen to gain edges and d(κ) to lose them, then in the
stationary state we will have u(κst) = d(κst) since the total number of edges
will be conserved. From Eq. (3.2) we can see that pst (k) will have an extremum
at some value ke if it satisfies π(ke ) = σ(ke + 1). ke will be a maximum (mini-
mum) if the numerator in Eq. (3.2) is smaller (greater) than the denominator
for k > ke , and viceversa for k < ke . Assuming, for example, that there is
one and only one such ke , then a maximum implies a relatively homogeneous
distribution, while a minimum means pst(k) will be split in two, and therefore
highly heterogeneous. More intuitively, if for nodes with large enough k there
is a higher probability of gaining edges than of losing them, the degrees of
these nodes will grow indefinitely, leading to heterogeneity. If, on the other
hand, highly connected nodes always lose more edges than they gain, we will
2We are ignoring the small corrections that arise because j 6= i and l 6= m, which in any
case would disappear if self-connections were allowed.
3.3 Synaptic pruning
25
obtain quite homogeneous networks. From this reasoning we can see that a
particularly interesting case (which turns out to be critical) is that in which
π(k) and σ(k) are such that
π(k) = σ(k) ≡ v (k),
∀k .
According to Eq. (3.2), Condition (3.3) means that for large k , ∂pst (k)/∂k →
0, and pst(k) flattens out – as for example a power-law does.
The standard Fokker-Planck approximation for the one step process defined
by Eq. (3.1) is (van Kampen, 1992):
(3.3)
∂p(k , t)
∂ t
=
1
2
∂ 2
∂k2 {[d(κ) σ(k) + u(κ) π(k)] p(k , t)}
(3.4)
∂
∂k {[d(κ) σ(k) − u(κ) π(k)] p(k , t)} .
For transition rates which meet Condition (3.3), Eq. (3.4) can be written as:
+
∂p(k , t)
∂ t
=
1
2
[d (κ) + u (κ)]
∂ 2
∂k2 [v (k)p(k , t)]
(3.5)
∂
∂k
[v (k)p(k , t)] .
+ [d (κ) − u (κ)]
Ignoring boundary conditions, the stationary solution must satisfy, on the one
hand, v (k)pst(k) = Ak + B , so that the diffusion is stationary, and, on the
other, u(κst) = d(κst), to cancel out the drift. For this situation to be reach-
able from any initial condition, u(κ) and d(κ) must be monotonous functions,
decreasing and increasing respectively.
3.3 Synaptic pruning
As a simple example, we will first consider global probabilities which have the
linear forms:
and
d[κ(t)] =
,
u[κ(t)] =
κmax (cid:19)
N (cid:18)1 −
n
κ(t)
where n is the expected value of the number of additions and deletions of edges
per time step, and κmax is the maximum value the mean degree can have. This
choice describes a situation in which the higher the density of synapses, the less
likely new ones are to sprout and the more likely existing ones are to atrophy
– a situation that might arise, for instance, in the presence of a finite quantity
κ(t)
κmax
(3.6)
n
N
Chapter 3. Evolving networks and the development of neural
systems
26
h
(3.7)
of nutrients. Again taking into account that π and σ are normalized to one,
gain
− P lose
we find that the expected increment in κ(t) is
summing over P
i
i
N (cid:20)1 − 2
κmax (cid:21)
∆κ(t)
κ(t)
n
∆t i = 2{u[κ(t)] − d[κ(t)]} = 2
(independently of the local probabilities). Therefore, the mean degree will
increase or decrease exponentially with time, from κ(0) to 1
2 κmax . Assuming
that the initial condition is, say, κ(0) = κmax , and expressing the solution in
terms of the mean synaptic density – i.e., ρ(t) ≡ κ(t)N/(2V ), with V the total
volume considered – we have
ρ(t) = ρf (cid:16)1 + e−t/τp (cid:17) ,
where we have defined ρf ≡ ρ(t → ∞) and the time constant for pruning is
τp = ρfN/n. This equation was fitted in Fig. 3.1 to experimental data on
layers 1 and 2 of the human auditory cortex3 obtained during autopsies by
Huttenlocher and Dabholkar (1997).
It seems reasonable to assume that the initial overgrowth of synapses is due
to the transient existence of some kind of growth factors. If we account for
these by including a nonlinear, time-dependent term g (t) ≡ a exp(−t/τg ) in
the probability of growth, i.e., u[κ(t), t] = (n/N )[1 − κ(t)/κmax + g (t)], leaving
d[κ(t)] as before, we find that ρ(t) becomes
− t−t0
ρ(t) = ρf (cid:20)1 + e−t/τp − (cid:16)1 + e−t0 /τp (cid:17) e
τg (cid:21) ,
where t0 is the time at which synapses begin to form (t = 0 corresponds to
the moment of conception) and τg is the time constant related to growth. The
inset in Fig. 3.1 shows the best fit to the auditory cortex data. Since the
contour conditions ρf and (for Eq.
(3.8)) t0 are simply taken as the value
of the last data point and the time of the first one, in each case, the time
constants τp and τg are the only parameters needed for the fit.
(3.8)
3.4 Phase transitions
The drift-like evolution of the mean degree we have just illustrated with the
example of synaptic pruning is independent of the local probabilities π(k)
3Data points for three particular days (smaller symbols) are omitted from the fit, since
we believe these must be from sub jects with inherently lower synaptic density.
3.4 Phase transitions
27
Cortical layer 1
Cortical layer 2
80
80
80
80
)
)
)
)
t
t
t
t
(
(
(
(
ρ
ρ
ρ
ρ
0
0
0
0
100
10000
1000
100
100
100
1000
1000
1000
10000
10000
10000
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
3
3
3
3
3
3
3
3
m
m
m
m
m
m
m
m
µ
µ
µ
µ
µ
µ
µ
µ
/
/
/
/
/
/
/
/
s
s
s
s
s
s
s
s
e
e
e
e
e
e
e
e
s
s
s
s
s
s
s
s
p
p
p
p
p
p
p
p
a
a
a
a
a
a
a
a
n
n
n
n
n
n
n
n
y
y
y
y
y
y
y
y
S
S
S
S
S
S
S
S
:
:
:
:
:
:
:
:
)
)
)
)
)
)
)
)
t
t
t
t
t
t
t
t
(
(
(
(
(
(
(
(
ρ
ρ
ρ
ρ
ρ
ρ
ρ
ρ
80
80
80
80
80
80
80
80
70
70
70
70
70
70
70
70
60
60
60
60
60
60
60
60
50
50
50
50
50
50
50
50
40
40
40
40
40
40
40
40
30
30
30
30
30
30
30
30
20
20
20
20
20
20
20
20
0
0
0
0
0
0
0
0
5000
5000
5000
5000
5000
5000
5000
5000
10000
15000
10000
10000
10000
10000
10000
10000
10000
15000
15000
15000
15000
15000
15000
15000
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
Conceptual age (days)
20000
20000
20000
20000
20000
20000
20000
20000
25000
25000
25000
25000
25000
25000
25000
25000
Figure 3.1:
Synaptic densities in layers 1 (red squares) and 2 (black cir-
cles) of the human auditory cortex against time from conception. Data from
Huttenlocher and Dabholkar (1997), obtained by directly counting synapses in
tissues from autopsies. Lines follow best fits to Eq. (3.7), where the param-
eters were: for layer 1, τp = 5041 days; and for layer 2, τp = 3898 days (for
ρf we have used the last data pints: 30.7 and 40.8 synapses/µm3 , for layers 1
and 2 respectively). Data pertaining to the first year and to days 4700, 5000
7300, shown with smaller symbols, where omitted from the fit. Assuming the
existence of transient growth factors, we can include the data points for the
first year in the fit by using Eq. (3.8). This is done in the inset (where time
is displayed logarithmically). The best fits were: for layer 1, τg = 151.0 and
τp = 5221; and for layer 2, τg = 191.1 and τp = 4184, all in days (we have
approximated t0 to the time of the first data points, 192 days).
and σ(k). The effect of these is rather in the diffusive behaviour which can
lead, as mentioned, either to homogeneous or to heterogeneous final states.
A useful bounded order parameter to characterize these phases is therefore
m ≡ exp (−σ 2/κ2 ) , where σ 2 = hk2i − κ2 is the variance of the degree dis-
tribution (h·i ≡ N −1 Pi (·) represents an average over nodes). We will use
mst ≡ limt→∞ m(t) to distinguish between the different phases, since mst = 1
for a regular network and mst → 0 for one following a highly heterogeneous
distribution. Although there are particular choices of probabilities which lead
to Eq. (3.5), these are not the only critical cases, since the transition from
Chapter 3. Evolving networks and the development of neural
systems
28
100
100
100
100
100
100
100
Critical
Critical
Critical
Critical
Critical
Critical
Critical
)
)
)
)
)
)
)
t
t
t
t
t
t
t
,
,
,
,
,
,
,
k
k
k
k
k
k
k
(
(
(
(
(
(
(
p
p
p
p
p
p
p
10-2
10-2
10-2
10-2
10-2
10-2
10-2
10-4
10-4
10-4
10-4
10-4
10-4
10-4
)
)
t
t
(
(
κ
κ
20
20
10
10
0
0
1
1
k-2
k-2
k-2
k-2
k-2
k-2
k-2
)
)
)
)
)
)
)
)
)
t
t
t
t
t
t
t
t
t
(
(
(
(
(
(
(
(
(
m
m
m
m
m
m
m
m
m
2.5 103
2.5 103
0
0
5 104
5 104
0
0
Time (MCS)
Time (MCS)
t=102
t=103
t=106
100
100
100
100
100
100
100
Supercritical
Supercritical
Supercritical
Supercritical
Supercritical
Supercritical
Supercritical
10-2
10-2
10-2
10-2
10-2
10-2
10-2
10-4
10-4
10-4
10-4
10-4
10-4
10-4
)
)
t
t
(
(
κ
κ
20
20
10
10
0
0
1
1
)
)
t
t
(
(
m
m
2.5 103
2.5 103
0
0
5 104
5 104
0
0
Time (MCS)
Time (MCS)
k-4
k-4
k-4
k-4
k-4
k-4
k-4
k-4
k-4
k-4
k-4
t=102
t=103
t=105
100
100
100
100
100
100
100
101
101
101
101
101
101
101
102
102
102
102
102
102
102
k
k
k
k
k
k
k
103
103
103
103
103
103
103
100
100
100
100
100
100
100
101
101
101
101
101
101
101
102
102
102
102
102
102
102
k
k
k
k
k
k
k
103
103
103
103
103
103
103
Figure 3.2: Evolution of the degree distributions of networks beginning as
regular random graphs with κ(0) = 20 in the critical (top) and supercritical
(bottom) regimes. Local probabilities are σ(k) = k/(hkiN ) in both cases, and
π(k) = 2σ(k)−N −1 and π(k) = k3/2/(hk3/2 iN ) for the critical and supercritical
ones, respectively. Global probabilities as in Eq.
(3.6), with n = 10 and
κmax = 20. Symbols in the main panels correspond to p(k , t) at different times
as obtained from MC simulations. Lines result from numerical integration of
Eq. (3.1). Insets show typical time series of κ and m. Light blue lines are
from MC simulations and red lines are theoretical, given by Eq. (3.7) and Eq.
(3.1), respectively. N = 1000.
homogeneous to heterogeneous stationary states can come about also with
functions which never meet Condition (3.3). Rather, this is a classic topolog-
ical phase transition, the nature of which depends on the choice of functions
(Park and Newman, 2004; Burda et al., 2004; Der´enyi et al., 2004) .
Evolution of the degree distribution is shown in Fig. 3.2 for critical and
supercritical choices for the probabilities, as given by MC simulations (starting
from regular random graphs) and contrasted with theory.
(The subcritical
regime is not shown since the stationary state has a distribution similar to
the ones at t = 103 MCS in the other regimes.) The disparity between the
theory and the simulations for the final distributions is due to the build up of
certain correlations not taken into account in our analysis. This is because the
existence of some very highly connected nodes reduces the probability of there
being very low degree nodes. In particular, if there are, say, x nodes connected
to the rest of the network, then a natural cutoff, kmin = x, emerges. Note
that this occurs only when we restrict ourselves to simple networks, i.e., with
only one edge allowed for each pair of nodes. This topological phase transition
is shown in Fig. 3.3, where mst is plotted against parameter α for global
3.4 Phase transitions
29
0
0
0
0
0
0
r
r
r
r
r
r
-1
-1
-1
-1
-1
-1
0.5
0.5
0.5
0.5
0.5
0.5
1.5
1.5
1.5
1.5
1.5
1.5
2
2
2
2
2
2
1
1
1
1
1
1
α
α
α
α
α
α
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
t
t
t
t
t
t
t
t
t
s
s
s
s
s
s
s
s
s
m
m
m
m
m
m
m
m
m
N=1000
N=1500
N=2000
103
103
Ν
Ν
λ
λ
Q
Q
λ
λ
N
N
0
0
0.5
0.5
0.5
0.5
1.5
1.5
1.5
1.5
2
2
2
2
1
1
1
1
α
α
α
α
5 102
5 102
Q
Q
0
0
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
1
1
1
1
1
1
1
1
1
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
2
2
2
2
2
2
2
2
2
α
α
α
α
α
α
α
α
α
Figure 3.3: Phase transitions in mst for π(k) ∼ kα and σ(k) ∼ k , and u(κ)
and d(κ) as in Eq. (3.6). N = 1000 (blue squares), 1500 (red triangles) and
2000 (black circles); κ(0) = κmax = 2n = N/50. Corresponding lines are from
numerical integration of Eq. (3.1). The bottom left inset shows values of the
highest eigenvalue of the Laplacian matrix (red squares) and of Q = λN /λ2
(black circles), a measure of unsynchronizability; N = 1000. The top right
inset shows transitions for the same parameters in the final values of Pearson’s
correlation coefficient r (see Section 3.5), both for only one edge allowed per
pair of nodes (red squares) and without this restriction (black circles).
probabilities as in Eq. (3.6) and local ones π(k) ∼ kα and σ(k) ∼ k . This
situation corresponds to one in which edges are eliminated randomly while
nodes have a power-law probability of sprouting new ones (note that power-
laws are good descriptions of a variety of monotonous response functions, yet
only require one parameter). Although, to our knowledge, there are not yet
enough empirical data to ascertain what degree distribution the structural
topology of the human brain follows, it is worth noting that its functional
topology, at the level of brain areas, has been found to be scale-free with an
exponent very close to 2 (Egu´ıluz et al., 2005).
In general, most other measures can be expected to undergo a transition
along with its variance. For instance, highly heterogeneous networks (such
as scale-free ones) exhibit the small-world property, characterized by a high
clustering coefficient, C ≫ hki/N , and a low mean minimum path, l ∼ ln(N )
(Watts and Strogatz, 1998). A particularly interesting topological feature of
Chapter 3. Evolving networks and the development of neural
systems
30
a network is its synchronizability – i.e., given a set of oscillators placed at
the nodes and coupled via the edges, how wide a range of coupling strengths
will result in them all becoming synchronized. Barahona and Pecora showed
analytically that, for linear oscillators, a network is more synchronizable the
lower the relation Q = λN /λ2 – where λN and λ2 are the highest and lowest
non-zero eigenvalues of the Laplacian matrix ( Λij ≡ δij ki − aij ), respectively
(Barahona and Pecora, 2002). The bottom left inset in Fig. 3.3 displays the
values of Q and λN obtained for the different stationary states. There is a
peak in Q at the critical point.
It has been argued that this tendency of
heterogeneous topologies to be particularly unsynchronizable poses a paradox
given the wide prevalence of scale-free networks in nature, a problem that has
been deftly got around by considering appropriate weighting schemes for the
edges (Motter et al., 2005; Chavez et al., 2005) (see also4 , and the review by
Arenas et al. (2008a)). However, there is no generic reason why high synchro-
nizability should always be desirable. In fact, it has recently been shown that
heterogeneity can improve the dynamical performance of model neural net-
works precisely because the fixed points are easily destabilised (Johnson et al.,
2008) (as well as conferring robustness to thermal fluctuations and improving
storage capacity (Torres et al., 2004)). This makes intuitive sense, since, pre-
sumably, one would not usually want all the neurons in one’s brain to be doing
exactly the same thing. Therefore, this point of maximum unsynchronizability
at the phase transition may be a particularly advantageous one.
On the whole, we find that three classes of network – homogeneous, scale-
free (at the critical point) and ones composed of starlike structures, with a
great many small-degree nodes connected to a few hubs – can emerge for any
kind of attachment/detachment rules.
It follows that a network sub ject to
some sort of optimising mechanism, such as Natural Selection for the case of
living systems, could thus evolve towards whichever topology best suits its
requirements by tuning these microscopic actions.
3.5 Correlations
Most real networks have been found to exhibit degree-degree correlations, also
known as mixing by degree (Pastor-Satorras et al., 2001; Newman, 2003c).
4Using pacemaker nodes, scale-free networks have also been shown to emerge via rules
which maximize synchrony (Sendina-Nadal et al., 2008).
3.5 Correlations
31
They can thus be classified as assortative, when the degree of a typical node
is positively correlated with that of its neighbours, or disassortative, when
the correlation is negative. This property has important implications for net-
work characteristics such as connectedness and robustness (Newman, 2002,
2003a). A useful measure of this phenomenon is Pearson’s correlation coef-
ficient applied to the edges (Newman, 2003c,a; Boccaletti et al., 2006): r =
([klk ′
l ] − [kl ]2)/([k2
l ] − [kl ]2), where kl and k ′
l are the degrees of each of the two
nodes pertaining to edge l, and [·] ≡ (hkiN )−1 Pl (·) represents an average
over edges; r ≤ 1. Writing Pl (·) = Pij aij (·), r can be expressed in terms of
averages over nodes:
r = hkihk2knn(k)i − hk2i2
(3.9)
,
hkihk3i − hk2i2
where knn(k) is the mean nearest-neighbour-degree function; i.e., if knn,i ≡
k−1
i Pj aij kj is the mean degree of the neighbours of node i, knn (k) is its average
over all nodes such that ki = k . Whereas most social networks are assortative
(r > 0) – due, probably, to mechanisms such as homophily (Newman, 2003c)
– almost all other networks, whether biological, technological or information-
related, seem to be generically disassortative. The top right inset in Fig. 3.3
displays the stationary value of r obtained in the same networks as in the main
panel and lower inset. It turns out that the heterogeneous regime is disassorta-
tive, the more so the larger α. (Note that a completely homogeneous network
cannot have degree-degree correlations, since all degrees are the same.) It is
known that the restriction of having at most one edge per pair of nodes induces
disassortativity (Park and Newman, 2003; Maslov et al., 2004). However, in
our case this is not the sole origin of the correlations, as can also be seen in
the same inset of Fig. 3.3, where we have plotted r for networks in which we
have lifted the restriction and allowed any number of edges per pair of nodes.
In fact, when multiple edges are allowed, the correlations are slightly stronger.
As we shall discuss in Chapter 5, there is a general entropic reason for hetero-
geneous networks, in their equilibrium state (i.e., in the absence of correlating
mechanisms), to become disassortative (Johnson et al., 2010b). But neither is
this here the case, since the networks generated are driven from equilibrium
by the mechanisms of preferential attachment and detachment.
To understand how these specific correlations come about, consider a pair
of nodes (i, j ), which, at a given moment, can either be occupied by an edge or
unoccupied. We will call the expected times of permanence for occupied and
unoccupied states τ o
ij and τ u
ij , respectively. After sufficient evolution time (so
Chapter 3. Evolving networks and the development of neural
systems
32
that occupancy becomes independent of the initial state5 ), the expected value
of the corresponding element of the adjacency matrix, E (aij ) ≡ ǫij , will be
τ o
ij
ij + τ u
τ o
ij
ǫij =
.
If p+
ij (p−
ij ) is the probability that (i, j ) will become occupied (unoccupied) given
ij and τ u
that it is unoccupied (occupied), then τ o
ij ∼ 1/p+
ij ∼ 1/p−
ij , yielding
ij !−1
ǫij = 1 +
p−
ij
p+
Taking into account the probability that each node has of gaining or losing an
edge, we obtain6 : p+
ij = u(hki)N −1 [π(ki ) + π(kj )] and p−
ij = d(hki)[σ(ki)/ki +
ij ≫ p+
σ(kj )/kj ]. Then, assuming that the network is sparse enough that p−
ij
(since the number of edges is much smaller than the number of pairs), and
particularising for power-law local probabilities π(k) ∼ kα and σ(k) ∼ kβ , the
expected occupancy of the pair is
.
=
p+
ij
p−
ij
ǫij ≃
hkαiN kα
j ! .
i + kα
hkβ i
u(hki)
j
kβ−1
i + kβ−1
d(hki)
Considering the stationary state, when u(hki) = d(hki), and for the case of
random deletion of edges, β = 1 (so that the only nonlinearity is due to α),
the previous expression reduces to
ǫij ≃ hki
2hkαiN (cid:0)kα
i + kα
j (cid:1) .
(Note that this matrix is not consistent term by term, since Pj ǫij
6= ki ,
although it is globally consistent: Pij ǫij = hkiN .) The nearest-neighbour-
degree function is now
1
ǫij kj = hki
ki Xj
i + hkα+1 ik−1
(hkikα−1
knn (ki ) =
i )
2hkαi
(a decreasing function for any α), with the result that Pearson’s coefficient
becomes
hkα i (cid:18) hki3hkα+1 i − hk2i2 hkαi
(cid:19) .
1
hkihk3i − hk2i2
5Note that this will always happen eventually since the process is ergodic.
6Again, we are ignoring corrections due to the fact that i is necessarily different from j .
(3.10)
r =
(3.11)
3.5 Correlations
33
More generally, one can understand the emergence of these correlations in
the following way. For the network to become heterogeneous, we must have
π(k) + N −1 ≥ σ(k) + k/(hkiN ) for large enough k , so that highly connected
nodes do not lose more edges than they can acquire (see Section 3.2). This
implies that π(k) must be increasing and approximately linear or superlinear.
The expected value of the degree of a node i, chosen according to π(ki ), is then
E (ki ) = N −1 Pk π(k)k & hk2i/hki, while that of its new, randomly chosen
neighbour, j , is only E (kj ) = hki. This induces disassortative correlations
which can never be compensated by the breaking of edges between nodes whose
expected degree values are N −1 Pk σ(k)k and hk2i/hki if σ(k) is an increasing
function. It thus ensues that a scenario such as the one analysed in this paper
will never lead to assortative networks except for some cases in which σ(k)
is a decreasing function – meaning that less connected nodes should be more
likely to lose edges. Assortativity could, however, arise if there were some bias
also on the node chosen to be i’s neighbour, e.g. on the postsynaptic neuron
– which is precisely what happens in most social networks, where individuals
do not generally choose their friends, partners, etc. randomly. Although there
seem to be other reasons for the ubiquity of disassortative networks in nature
(Johnson et al., 2010b), it is possible that the generality of the scenario studied
here may also play a part.
We can use the expected value matrix ǫ to estimate other magnitudes.
For example, the clustering coefficient, as defined by Watts and Strogatz
(Watts and Strogatz, 1998), is an average over nodes of Ci , with Ci the pro-
portion of i’s neighbours which are connected to each other; so its expected
value is E (Ci ) = ǫj l conditioned to j and l being neighbours of i’s. This means
that, on average, we can make the approximation that
kj = kl = hknn i = hki
[hkihkα−1i + hkα+1 ihk−1i].
2hkαi
Substituting this value in Eq. (3.10), and taking into account that one edge
of j ’s and one of l’s are taken up by i, we have
C ≃ hki
(hknn i − 1)α .
hkαiN
For a rough estimate of the mean minimum path (the minimum path between
two nodes being the smallest number of edges one has to follow to get from one
to the other), we can proceed as Albert and Barab´asi (2002). For a given node,
let us define the number of nearest neighbours, z1 , next-nearest neighbours,
z2 , and in general mth neighbours, zm . Using the relation zm = z1 (z2/z1)m−1 ,
(3.12)
Chapter 3. Evolving networks and the development of neural
systems
34
and assuming that the network is connected and can be obtained in l steps,
this yields
1 +
zm = N .
(3.13)
l
X1
On average, z1 = hki and z2 = hki[(1 − C )hknni − 1] (since for each second
nearest neighbour, one edge goes to the reference node and a proportion C to
mutual neighbours). Now, if N ≫ z1 and z2 ≫ z1 , Eq. (3.13) leads to
ln(N/hki)
ln[(1 − C )hknni − 1]
l ≃ 1 +
(3.14)
.
3.6 The C. Elegans neural network
There exists a biological neural network which has been entirely mapped
(although not, to the best of our knowledge, at different stages of develop-
ment) – that of the much-investigated worm C. Elegans (White et al., 1986;
Watts and Strogatz, 1998). With a view to testing whether such a network
could arise via simple stochastic rules of the kind we are here considering, we
ran simulations for the same number of nodes, N = 307, and (stationary) mean
degree, hki = 14.0 (in the simple, undirected representation of the network).
Using the global probabilities given by Eq. (3.6) and local ones π(k) ∼ kα
and σ(k) ∼ k (as in Fig. 3.3), we obtain a surprising result. Precisely at the
critical point, α = αc ≃ 1.35, there are some remarkable similarities between
the biological network and the ones produced by the model.
Figure 3.4 displays the degree distributions, both for the empirical net-
work and for the average (stationary) simulated network corresponding to the
critical point, while the top inset shows the mean-nearest-neighbour degree
function knn(k) for the same networks. Both p(k) and knn(k) of the simulated
networks can be seen to be very similar to those measured in the biological one.
Furthermore, as is displayed in Table 3.1, the clustering coefficient obtained in
simulation is almost the same as the empirical one. The mean minimum path is
similar though slightly smaller in simulation, probably due to the worm’s brain
having modules related to functions (Arenas et al., 2008b). Finally, Pearson’s
coefficient is also in fairly good agreement, although the simulated networks
are actually a bit more disassortative. It should, however, be stressed that the
simulation results are for averages over 100 runs, while the biological system is
equivalent to a single run; given the small number of neurons, statistical fluc-
tuations can be fairly large, so one should refrain from attributing too much
3.6 The C. Elegans neural network
35
0.1
0.1
0.1
0.01
0.01
0.01
0.001
0.001
0.001
)
)
)
k
k
k
(
(
(
p
p
p
1
1
m
m
0
0
1e-04
1e-04
1e-04
1
1
1
102
102
102
102
n
n
n
n
n
n
n
n
k
k
k
k
101
101
101
101
100
100
100
100
k-0.5
k-0.5
k-0.5
k-0.5
102
102
102
102
k
k
k
k
k-2.5
k-2.5
k-2.5
α
α
c=1.35
c=1.35
0.5 1 1.5 2
0.5 1 1.5 2
α
α
10
10
10
100
100
100
1000
1000
1000
k
k
k
Figure 3.4: Degree distribution (binned) of the C. Elegans neural network
(circles) (White et al., 1986) and that obtained with MC simulations (line) in
the stationary state (t = 105 steps) for an equivalent network in which edges
are removed randomly (β = 1) at the critical point (α = 1.35). N = 307, κst =
14.0, averages over 100 runs. Global probabilities as in Eq. (3.6). The slope is
for k−5/2 . Top right inset: mean-neighbour-degree function knn (k) as measured
in the same empirical network (circles) and as given by the same simulations
(line) as in the main panel. The slope is for k−1/2 . Bottom left inset: mst
of equivalent network for a range of α, both from simulations (circles) and as
obtained with Eq. (3.1). (See also Table 3.1.)
importance to the precise values obtained – at least until we can average over
100 worms. Table 3.1 also shows the values of C , l and r both as estimated
form the theory laid out in Section 3.5, and for the equivalent network in the
configuration model (Newman, 2003c) – generally taken as the null model for
heterogeneous networks, where the probability of an edge existing between
nodes i and j is kikj /(hkiN ). It is clear that whereas the configuration-model
predictions deviate substantially from the magnitudes measured in the C. El-
egans neural network, the growth process we are here considering accounts for
them quite well.
It is interesting that it should be at the critical point that a
structural topology so similar to the empirical one emerges, since it seems that
the brain’s functional topology may also be related to a critical point (Chialvo,
2004; Chialvo et al., 2008).
Chapter 3. Evolving networks and the development of neural
systems
36
Experiment Simulation Theory Config.
C 0.28
2.46
l
r
0.28
2.19
0.23
0.15
1.86
1.96
-0.163
-0.207
-0.305
-0.101
Table 3.1: Values of small-world parameters C and l, and Pearson’s correla-
tion coefficient r , as measured in the neural network of the worm C. Elegans
(White et al., 1986), and as obtained from simulations in the stationary state
(t = 105 steps) for an equivalent network at the critical point when edges are
removed randomly – i.e., for α = 1.35 and β = 1. N = 307, κst = 14.0;
averages over 100 runs and global probabilities as in Eq. (3.6). Theoretical
estimates correspond to Eqs. (3.12), (3.14) and (3.11) applied to the networks
generated by the same simulations. The last column lists the respective con-
figuration model values: C and l are obtained theoretically as in (Newman,
2003c), while r , from MC simulations as in (Maslov et al., 2004), is the value
expected due to the absence of multiple edges. (See also Fig. 3.4.)
3.7 Discussion
With this work we have attempted, on the one hand, to extend our under-
standing of evolving networks so that any choice of transition probabilities
dependent on local and/or global degrees can be treated analytically, thereby
obtaining some model-independent results; and on the other, to illustrate how
such a framework can be applied to realistic biological scenarios. For the latter,
we have used two examples relating to two rather different nervous systems:
i) synaptic pruning in humans, for which the use of nonlinear global prob-
abilities reproduces the initial increase and subsequent depletion in synaptic
density in good accord with experiments – to the extent that nonmonotonic
data points spanning a lifetime can be very well fitted with only two parame-
ters; and
ii) the structure of the C. Elegans neural network, for which it turns out that
by only considering the numbers of nodes and edges, and imposing random
deletion of edges and power-law probability of growth, the critical point leads
3.7 Discussion
37
to networks exhibiting many of the worm’s nontrivial features – such as the
degree distribution, small-world parameters, and even level of disassortativity.
These examples indicate that it is not far-fetched to contemplate how
many structural features of the brain or other networks – and not just the
degree distributions – could arise by simple stochastic rules like the ones con-
sidered; although, undoubtedly, other ingredients such as natural modularity
(Arenas et al., 2008b), a metric (Kaiser and Hilgetag, 2004) or functional re-
quirements (Sporns et al., 2004) can also be expected to play a role in many
instances. We hope, therefore, that the framework laid out here – in which
for simplicity we have assumed the network to be undirected and to have a
fixed size, although generalizations are straightforward – may prove useful for
interpreting data from a variety of fields. It would be particularly interesting
to try to locate and quantify the biological mechanisms assumed to be behind
this kind of network dynamics.
Chapter 4
Bringing on the Edge of Chaos
with heterogeneity
The collective behaviour of systems of coupled excitable elements, such as
neurons, has been shown to depend significantly on the heterogeneity of the
degree distribution of the underlying network of interactions. For instance,
broad – in particular, scale-free – distributions have been found to improve
static memory performance in neural-network models. Here we look at the
influence of degree heterogeneity in a neural network which, due to the effect
of synaptic depression (a kind of fatigue of the interaction strengths), exhibits
chaotic behaviour. Not only can the existence of a chaotic phase be related
to neurophysiological experiments;
it allows the system to perform a class
of dynamic pattern-recognition tasks. We find first of all that, as has been
described in a few other systems, optimal performance is achieved close to the
phase transition – i.e., at the so-called Edge of Chaos. Furthermore, we obtain
a functional relationship between the level of synaptic depression required to
bring on chaos and the heterogeneity of the degree distribution. This result
points to a clear advantage of low-exponent scale-free networks, and suggests
an explanation for their apparent ubiquity in certain biological systems.
4.1 Exciting cooperation
Excitable systems allow for the regeneration of waves propagating through
them, and may thus respond vigorously to weak stimulus. The brain and other
parts of the nervous system are well–studied paradigms, and forest fires with
constant ignition of trees and autocatalytic reactions in surfaces, for instance,
also share some of the basics (Bak et al., 1990; Meron, 1992; Lindner et al.,
39
40 Chapter 4. Bringing on the Edge of Chaos with heterogeneity
2004; Izhikevich , 2007; Arenas et al., 2008a). The fact that signals are not
gradually damped by friction in these cases is a consequence of cooperation
among many elements in a nonequilibrium setting. These systems can be
seen as large networks of nodes that are “excitable”. This admits various
realizations, but typically means that each element has a threshold and a
refractory time between consecutive responses – a behaviour that impedes
thermal equilibrium.
Some brain tasks can be simulated with mathematical neural networks. As
described in Chapter 2, these consist of neurons – often modelled as variables
which are as simple as possible while still able to display the essence of the
cooperative behaviour of interest1 – connected by edges representing synapses
(Amari, 1972; Hopfield, 1982; Amit, 1989; Torres and Varona, 2010). If the
edges are weighted according to some prescription – such as the Hebb rule
(Hebb, 1949) – which saves information from a set of given patterns of activ-
ity (particular configurations of active and inactive neurons), these patterns
become attractors of the phase-space dynamics. Therefore, the system is then
able to retrieve the stored patterns; this mechanism is known as associative
memory. Actual neural systems do much more than just recalling a memory
and staying there, however. That is, one should expect dynamic instabilities
or some other destabilizing mechanism. This expectation is reinforced by re-
cent experiments suggesting that synapses undergo rapid changes with time
which may both determine brain tasks (Abbott et al., 1997; Tsodyks et al.,
1998; Hilfiker et al., 1999; Pantic et al., 2002) and induce irregular and per-
haps chaotic activity (Barrie et al., 1996; Korn and Faure, 2003).
One may argue that the observed rapid changes (which have been found
to cause “synaptic depression” and/or “facilitation” on the time scale of mil-
liseconds (Tsodyks et al., 1998; Pantic et al., 2002) – i.e., much faster than
the plasticity processes whereby synapses store patterns (Malenka and Nicoll,
1999)) may simply correspond to the characteristic behaviour of single ex-
citable elements. Furthermore, a fully-connected network which describes co-
operation among such excitable elements has recently been shown to exhibit
both attractors and chaotic instabilities (Marro et al., 2008). The work de-
scribed here, first reported by Johnson et al. (2008), extends and generalizes
this study to conclude on the influence of the excitable network topology on
1Several studies have already shown that binary neurons can capture the essence of
cooperation in many more complex settings. See, for instance, (Pantic et al., 2002) in the
case of integrate and fire neuron models of pyramidal cells.
4.2 The Fast-Noise model
41
dynamic behaviour. We show, in particular, an interesting correlation between
certain wiring topology and optimal functionality.
4.2 The Fast-Noise model
Consider N binary nodes (si = ±1) and the adjacency matrix, aij = 1, 0, which
indicates the existence or not of an edge between nodes i, j = 1, 2, ..., N . Let
there be a set of M patterns, ξ ν
i = ±1, ν = 1, ...M (which we generate here at
random), and assume that they are “stored” by giving each edge a base weight
ωij = N −1 Pν ξ ν
i ξ ν
j . Actual weights are dynamic, however, namely, ωij = ωij xj
where xj is a stochastic variable. Assuming the limit in which this varies in a
time scale infinitely smaller than the one for node dynamics, we can consider
a stationary distribution such as P (xj S ) = qδ (xj − Ξj ) + (1 − q)δ (xj − 1),
S = {sj } , for instance. This amounts to assuming that, at each time step,
every connection has a probability q of altering its weight by a factor Ξj which
is a function (to be determined) of the local field at j, defined as the net
current arriving to j from other nodes. This choice differs essentially from the
one used by Marro et al. (2008), where q depends on the global degree of order
and Ξj is a constant independent of j.
Assume independence of the noise at different edges, and that the transition
rate for the stochastic changes is
= Yj /aij =1 R dxj P (xj S )Ψ(uij )
¯c (S → S i )
R dxj P (xj S i)Ψ(−uij )
¯c (S i → S )
where uij ≡ sisj xj ωij T −1 , Ψ(u) = exp (cid:0)− 1
2 u(cid:1) to have proper contour condi-
tions, T is a “temperature” or stochasticity parameter, and S i stands for S
after the change si → −si . (This formalism and its interpretation is described
in detail by Marro and Dickman.) We define the effective local fields heff
i =
i si /T (cid:1) , where ϕ±
i (S, T , q) via Qj ϕ−
ij /ϕ+
heff
ij = exp (cid:0)−heff
ij ≡ q exp (±Ξj vij ) +
(1 − q) exp (±vij ), with vij = 1
2 aij uij . Effective weights ω eff
ij then follow from
heff
i = Pj ω eff
ij sj aij . To obtain an analytical expression, we linearize around
ωij = 0 (a good approximation when M ≪ N ), which yields
ω eff
ij = [1 + q (Ξj − 1)] ωij .
In order to fix Ξj here, we first introduce the overlap vector −→m = (m1 , ...mM ),
with mν ≡ N −1 Pi ξ ν
i si , which measures the correlation between the current
configuration and each of the stored patterns, and the local one −→mj of com-
ponents mν
j ≡ hki−1 Pl ξ ν
l sl aj l , where hki is the mean node connectivity, i.e.,
,
42 Chapter 4. Bringing on the Edge of Chaos with heterogeneity
the average of ki = Pj aij . We then assume, for any q 6= 0, that the relevant
factor is Ξj = 1 + ζ (hν
j )(Φ − 1)/q , with
χα
1 + M/N Xν
ζ (hν
hν
j α ,
j ) =
where χ ≡ N/hki and α > 0 is a parameter. This comes from the fact that
the field at node j can be written as a sum of components from each pattern,
namely, hj = PM
ν hν
j , where
j N −1 Xi
j = ξ ν
hν
Our choice for Ξj , which amounts to assuming that the “fatigue” at a given
edge increases with the field at the preceding node j (and allows to recover the
fully–connected limit described by Marro et al. (2008) if α = 2), finally leads
to
j mν
i si = χ−1ξ ν
aij ξ ν
j .
ij = [1 + (Φ − 1)ζj (−→mj )] ωij .
ω eff
Varying Φ one sets the nature of the weights. That is, 0 < Φ < 1 corresponds
to resistance (depression ) due to heavy local work, while the edge facilitates
– i.e., tends to increase the effect of the signal under the same situation –
for Φ > 1. (The action of the edge is reversed for negative Φ.) We performed
Monte Carlo simulations using standard parallel updating with the effective
rates ¯c (S → S i ) computed using the latter effective weights.
4.3 Edge of Chaos
It is possible to solve the single pattern case (M = 1) under a mean-field as-
sumption, which is a good approximation for large enough connectivity. That
is, we may substitute the matrix aij by its mean value over network realizations
to obtain analytical results that are independent of the underlying disorder.
Imagine that each node hosts ki half–edges according to a distribution p(k),
the total number of half–edges in the network being hkiN . Choose a node i
at random and randomly join one of its half–edges to an available free half–
edge. The probability that this half–edge ends at node j is kj / (hki N ) . Once
all the nodes have been linked up, the expected value (as a quenched average
over network realizations) for the number of edges joining nodes i and j is2
2Assuming one edge at most between any two nodes, aij = 0, 1, the value will be slightly
smaller, but it is easy to prove that this is also a good approximation if the network has a
structural cut-off: ki < phkiN , ∀i.
4.3 Edge of Chaos
43
E (aij ) = kikj / (hki N ). This expression, which can be seen as a definition
of the so-called configuration model for complex networks (Newman, 2003c),
is valid for random networks with a given degree sequence (or, in practise, a
given degree distribution) that have zero degree-degree correlations between
neighbours (Johnson et al., 2010b). Using the notation ηi ≡ ξisi , we have
mj = χhηi aij ii = χ
N Pi ηi aij . Because node activity is not statistically in-
dependent of connectivity (Torres et al., 2004), we must define a new set of
overlap parameters, analogous to m and mj . That is, µn ≡ hkn
i ηi ii/hkni and
the local versions µj
n ≡ χhkn
i ηi aij ii/hkni. After using aij = E (aij ), one ob-
tains the relation µi
n = hkn+1ikiµn+1/(hknihki2). Inserting this expression into
the definition of µn , and substituting hsi i = tanh[T −1hef f
(S )] (for large N),
i
standard mean-field analysis yields
MT ,Φ =
µn (t + 1) =
1
hkn i hkn tanh MT ,Φ (k , t)ik ,
where the last quantity is defined as
T N (cid:20)µ1(t) + (Φ − 1) hkα+1 i
hkiα+1 µ1(t)α µα+1(t)(cid:21) .
k
This is a two-dimensional map which is valid for any random topology of
distribution p(k). Note that the macroscopic magnitude of interest is µ0 =
m ≡ −→m .
A main consequence of this is the existence of a critical temperature, Tc , un-
der very general conditions. More specifically, as T is decreased, the overlap m
describes a second–order phase transition from a disordered or, say, “paramag-
netic” phase to an ordered (“ferromagnetic”) phase which exhibits associative
memory. The mean-field temperature at which this transition occurs is
Tc = hk2i
hki N
On the other hand, the map reduces to
µn (t + 1) = sign (cid:26)µn (t) (cid:20)1 + (Φ − 1) hkα+1i
hkiα+1 (cid:21)(cid:27)
for T = 0. This implies the existence at Φ = Φ0 , where
Φ0 = 1 − hkiα+1
hkα+1i
of a transition as Φ is decreased from the ferromagnetic phase to a new phase
in which periodic hopping between the attractor and its negative occurs. This
.
,
44 Chapter 4. Bringing on the Edge of Chaos with heterogeneity
is confirmed by the Monte Carlo simulations for M > 1; that is, the hopping is
also among different attractors for finite T . The simulations also indicate that
this transition washes out at low enough finite temperature. Instead, Monte
Carlo evolutions show that, for a certain range of Φ values, the system activity
then exhibits chaotic behaviour.
The transition from ferromagnetic to chaotic states is a main concern here-
after. Our interest in this regime follows from several recent observations
concerning the relevance of chaotic activity in a network. In particular, it has
been shown that chaos might be responsible for certain states of attention dur-
ing brain activity (Torres et al., 2008, 2009), and that some network properties
such as the computational capacity (Bertschinger and Natschlager, 2004) and
the dynamic range of sensitivity to stimuli (de Assis and Copelli, 2008) may
become optimal at the Edge of Chaos in a variety of settings.
We next note that the critical values Tc and Φ0 only depend on the mo-
ments of the generic distribution p(k), and that the ratio hkai/hkia , a > 1,
is a convenient way of characterizing heterogeneity. We studied in detail
two particular types of connectivity distributions with easily tunable hetero-
geneity; that is, networks with hkiN/2 edges randomly distributed with p (k)
such that the heterogeneity depends on a single parameter. Our first case
2 δ (k − k1) + 1
is the bimodal distribution, p(k) = 1
2 δ (k − k2) with parameter
∆ = (k2 − k1)/2 = hki − k1 = k2 − hki. Our second case is the scale–free
distribution, p(k) ∼ k−γ , which does not have any characteristic size but k is
1
γ−1 , N − 1) for finite N . Notice
confined to the limits, k0 and km ≤ min(k0N
that the network in this case gets more homogeneous as γ is increased3 , and
that this kind of distribution seems to be most relevant in nature (Newman,
2003c; Boccaletti et al., 2006). In particular, it seems important to mention
that the functional topology of the human brain, as defined by correlated ac-
tivity between small clusters of neurons, has been shown to correspond to this
case with exponent γ ≃ 2 (Egu´ıluz et al., 2005). (It has not yet been possible
to ascertain the brain’s structural topology experimentally, but there is some
evidence that function reflects structure at least to some extent (Zhou et al.,
2006b). Furthermore, it has been suggested, based on indirect methods, that
the structural connectivity of cat and macaque brains, at the level of brain
areas, may indeed be scale free (Kaiser et al., 2007) – and in any case dis-
3The distribution is truncated and therefore not strictly scale free for γ < 2. However,
nature shows examples for which γ is slightly larger than 1, so we consider the whole range
here.
4.3 Edge of Chaos
45
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
∆=4
∆=4
∆=4
∆=4
∆=8
∆=8
∆=8
∆=12
∆=12
γ=1.5
γ=1.5
γ=1.5
γ=1.5
γ=2.5
γ=2.5
γ=2.5
γ=3.5
γ=3.5
c
c
c
c
c
c
c
c
Φ
Φ
Φ
Φ
Φ
Φ
Φ
Φ
-
-
-
-
-
-
-
-
0
0
0
0
0
0
0
0
Φ
Φ
Φ
Φ
Φ
Φ
Φ
Φ
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
T/Tc
T/Tc
T/Tc
T/Tc
T/Tc
T/Tc
T/Tc
T/Tc
Figure 4.1: The temperature dependence of the difference between the values
for the fatigue at which the ferromagnetic–periodic transition occurs, as ob-
tained analytically for T = 0 (Φ0 ) and from MC simulations at finite T (Φc ).
The critical temperature is calculated as Tc = hk2 i (hki N )−1 for each topology.
Data are for bimodal distributions with varying ∆ and for scale–free topologies
with varying γ , as indicated. Here, hki = 20, N = 1600 and α = 2. Standard
deviations, represented as bars in this graph, were shown to drop with N −1/2
(not depicted).
plays significantly higher heterogeneity than that of, say, Erdos–R´enyi random
graphs.)
We obtained the critical value of the fatigue, Φc (T ) , from Monte Carlo
simulations at finite temperature T . These indicate that chaos never occurs for
T & 0.35Tc . On the other hand, a detailed comparison of the value Φc with Φ0
– as obtained analytically for T = 0 – indicates that Φc ≃ Φ0 .
Figure 4.1 illustrates the “error” Φ0 − Φc (T ) for different topologies. This
shows that the approximation Φc ≃ Φ0 is quite good at low T for any of the
cases examined. Therefore, assuming the critical values for the main param-
eters, Tc and Φ0 , as given by our map, we conclude that the more heteroge-
neous the distribution of connectivities of a network is, the lower the amount
of fatigue, and the higher the critical temperature, needed to destabilize the
dynamics. As an example of this interesting behaviour, consider a network
with hki = ln(N ), and dynamics according to α = 2. If the distribution were
46 Chapter 4. Bringing on the Edge of Chaos with heterogeneity
0.5
0.5
0.5
Stable
Stable
Stable
Φ
Φ
Φ
Unstable
Unstable
Unstable
0
0
0
1
1
1
1
1
1
0
0
0
4
4
4
8
8
8
∆
∆
∆
12
12
12
16
16
16
Stable
Stable
Stable
Stable
Stable
Stable
Φ
Φ
Φ
Φ
Φ
Φ
0.5
0.5
0.5
0.5
0.5
0.5
Unstable
Unstable
Unstable
Unstable
Unstable
Unstable
0
0
0
0
0
0
1
1
1
1
1
1
2
2
2
2
2
2
3
3
3
3
3
3
4
4
4
4
4
4
5
5
5
5
5
5
6
6
6
6
6
6
γ
γ
γ
γ
γ
γ
Figure 4.2:
The critical fatigue values Φ0 (solid lines) and Φc from MC
averages over 10 networks (symbols) with T = 2/N , hki = 20, N = 1600,
α = 2. The dots below the lines correspond to changes of sign of the Lyapunov
exponent as given by the iterated map, which qualitatively agree with the other
results. This is for bimodal and scale–free topologies, as indicated.
regular, the critical values would be Tc = ln(N )/N (which goes to zero in the
thermodynamic limit) and Φ0 = 0. However, a scale–free topology with the
same number of edges and γ = 2 would yield Tc = 1 and Φ0 = 1 − 2(ln N )3/N 2
(which goes to 1 as N → ∞).
Figure 6.5 illustrates, for two topologies, the phase diagram of the ferromagnetic–
chaotic transition. Most remarkable is the plateau observed in the Edge-of-
4.4 Network performance
47
Chaos or transition curve for scale–free topologies around γ ≃ 2, for which
very little fatigue, namely, Φ . 1 which corresponds to slight depression, is
required to achieve chaos. The limit γ → ∞ corresponds to hki–regular graphs
(equivalent to ∆ = 0). If γ is reduced, km increases and k0 decreases. The
network is truncated when km = N . It follows that a value of γ exits at which
k0 cannot be smaller, so that km must drop to preserve hki. This explains the
fall in Φc as γ → 1.
Assuming that the “ferromagnetic phase” here corresponds to a synchronous
state, our results are in qualitative agreement with the ones obtained recently
for coupled oscillators (Nishikawa et al., 2003; Zhou et al., 2006a). As a matter
of fact, the range of coupling strengths which allow for stability of synchronous
states in these systems has been shown to depend on the spectral gap of the
Laplacian matrix (Barahona and Pecora, 2002), implying that the more het-
erogeneous a topology is, the more easily activity can become unstable.
It
should be emphasized, however, that the dynamics we are considering here
does not come within the scope of the formalism used to derive these results,
since activity at node i depends on the local field at node j .
4.4 Network performance
As a further illustration of our findings, we monitored the performance as a
function of topology during a simulation of pattern recognition. That is, we
“showed” the system a pattern, say ν chosen at random from the set of M
previously stored, every certain number of time steps. This was performed in
practice by changing the field at each node for one time step, namely, hi →
hi + δξ ν , where δ measures the intensity of the input signal.
Ideally, the
network should remain in this configuration until it is newly stimulated. The
performance may thus be estimated from a temporal average of the overlap
between the current state and the input pattern, hmν itime . This is observed
to simply increase monotonically with ∆ for the bimodal case. The scale–free
case, however, as illustrated in Fig. 4.3, shows how the task is better performed
the closer to the Edge of Chaos the network is. This is because the system is
then easily destabilized by the stimulus while being able to retrieve a pattern
with accuracy. Figure 4.3 also shows that the best performance for the scale–
free topology when Φ = 1, i.e., in the absence of any fatigue, definitely occurs
around γ = 2.
48 Chapter 4. Bringing on the Edge of Chaos with heterogeneity
1
1
Φ=0.8
Φ=0.8
t
t
>
>
ν
ν
m
m
<
<
0.5
0.5
0
0
1
1
5
5
Φ=1.0
Φ=1.0
t
t
>
>
ν
ν
m
m
<
<
0.5
0.5
1
1
2
2
1
1
1
1
0
0
0
0
1
1
1
1
0
0
0
0
∆
∆
1
1
1
1
0
0
0
0
1
1
1
1
0
0
0
0
3
3
γ
γ
Time
Time
Time
Time
10
10
15
15
Time
Time
Time
Time
4
4
5
5
Figure 4.3: Network “performance” (see the main text) against ∆ for bimodal
topologies (above) and against γ for scale–free topologies (below). Φ = 0.8 for
the first case and Φ = 1 in the second. Averages over 20 network realizations
with stimulation every 50 MC steps for 2000 MC steps, δ = 5 and M = 4;
other parameters as in Fig. 6.5. Inset shows sections of typical time series
of mν for ∆ = 10 (above) and γ = 4 (below); the corresponding stimulus for
pattern ν is shown underneath.
4.5 Discussion
The model network we have studied is one of the simplest relevant situations
one may conceive. In particular, as emphasized above, we are greatly simply-
fieng the elements at the nodes (neurons) as binary variables. However, our
4.5 Discussion
49
assumption of dynamic connections which depend on the local fields in such
a simple scenario happens to show that a close relation may exist between
topological heterogeneity and function, thus suggesting this may indeed be a
relevant property for a realistic network efficiently to perform certain high level
tasks. In a similar way to networks shown previously to be useful for pattern
recognition and family identification (Cortes et al., 2005), our system retrieves
memory patterns with accuracy in spite of noise, and yet it is easily destabi-
lized so as to change state in response to an input signal – without requiring
excessive fatigue for the purpose. There is a relation between the amount Φ
of fatigue and the value of γ for which performance is maximized. One may
argue that the plateau of “good” behaviour shown around γ ≃ 2 for scale–free
networks with Φ . 1 (Fig. 6.5) is a possible justification for the supposed
tendency of certain systems in nature to evolve towards this topology. It may
also prove useful for implementing some artificial networks.
Chapter 5
Correlated networks and
natural disassortativity
An intriguing feature of complex networks is the ubiquity of strong negative
degree-degree correlations between neighbouring nodes – the only exceptions
being social systems, which tend to be assortative instead of disassortative.
With the double purpose of addressing this mystery and uncovering the effects
of correlations on network behaviour, we put forward a method which allows
for the model-independent study of ensembles of correlated networks. We go
on to show, by means of an information theory approach, that the expected
value of correlations for a network at equilibrium (i.e., in the absence of spe-
cific correlating mechanisms) is not, as had been supposed, uncorrelated, but
rahter disassortative. It turns out that the correlations of some networks are in
excellent agreement with our predictions, while others, with known correlating
or anticorrelating mechanisms, indeed appear to have been driven from their
equilibrium points as expected. Therefore, our approach not only provides a
parsimonious topological answer to a long-standing question, but also a neu-
tral model against which to contrast experimental data to determine whether
mechanisms must be sought to account for observed correlations. We go on
to use our method, in Chapter 6, to study the influence of assortativity on
neural-network dynamics.
5.1 Assortativity of networks
Complex networks, whether natural or artificial, have non-trivial topologies
which are usually studied by analysing a variety of measures, such as the degree
distribution, clustering, average paths, modularity, etc. (Albert and Barab´asi,
51
52
Chapter 5. Correlated networks and natural disassortativity
2002; Dorogovtsev and Mendes, 2003; Pastor-Satorras and Vespignani, 2004;
Newman, 2003c; Boccaletti et al., 2006) The mechanisms which lead to a par-
ticular structure and their relation to functional constraints are often not clear
and constitute the sub ject of much debate (Newman, 2003c; Boccaletti et al.,
2006). When nodes are endowed with some additional “property,” a feature
known as mixing or assortativity can arise, whereby edges are not placed be-
tween nodes completely at random, but depending in some way on the property
in question. If similar (dissimilar) nodes tend to wire together, the network is
said to be assortative (disassortative ) (Newman, 2002, 2003a).
An interesting situation is when the property taken into account is the
degree of each node – i.e., the number of neighbouring nodes connected to it.
It turns out that a high proportion of empirical networks – whether biological,
technological, information-related or linguistic – are disassortatively arranged
(high-degree nodes, or hubs, are preferentially linked to low-degree neighbours,
and viceversa) while social networks are usually assortative. Such degree-
degree correlations have important consequences for network characteristics
such as connectedness and robustness (Newman, 2002, 2003a).
However, while assortativity in social networks can be explained taking into
account homophily (Newman, 2002, 2003a) or modularity (Newman and Park,
2003), the widespread prevalence and extent of disassortative mixing in most
other networks remains somewhat mysterious. Maslov et al.
found that the
restriction of having at most one edge per pair of nodes induces some dis-
assortative correlations in heterogeneous networks (Maslov et al., 2004), and
Park and Newman showed how this analogue of the Pauli exclusion principle
leads to the edges following Fermi statistics (Park and Newman, 2003) (see
also (Capocci and Colaiori, 2006)). However, this restriction is not sufficient
to fully account for empirical data. In general, when one attempts to consider
computationally all the networks with the same distribution as a given empiri-
cal one, the mean assortativity is not necessarily zero (Holme and Zhao, 2007).
But since some “randomization” mechanisms induce positive correlations and
others negative ones (Farkas et al., 2004; Johnson et al., 2010a), it is not clear
how the phase space can be properly sampled numerically.
In this chapter we develop a method for the study of correlated networks
which is model-independent, and describe the main result of Ref. (Johnson et al.,
2010b) – namely, that there is a general reason, consistent with empirical
data, for the “natural” mixing of most networks to be disassortative. Us-
ing an information-theory approach we find that the configuration which can
5.2 The entropy of network ensembles
53
be expected to come about in the absence of specific additional constraints
turns out not to be, in general, uncorrelated. In fact, for highly heterogeneous
degree distributions such as those of the ubiquitous scale-free networks, we
show that the expected value of the mixing is usually disassortative: there are
simply more possible disassortative configurations than assortative ones. This
result provides a simple topological answer to a long-standing question. Let us
caution that this does not imply that all scale-free networks are disassortative,
but only that, in the absence of further information on the mechanisms behind
their evolution, this is the neutral expectation.
5.2 The entropy of network ensembles
The topology of a network is entirely described by its adjacency matrix a; the
element aij represents the number of edges linking node i to node j (for undi-
rected networks, a is symmetric). Among all the possible microscopically dis-
tinguishable configurations a set of L edges can adopt when distributed among
N nodes, it is often convenient to consider the set of configurations which have
certain features in common – typically some macroscopic magnitude, like the
degree distribution. Such a set of configurations defines an ensemble.
In a
seminal series of papers Bianconi has determined the partition functions of
various ensembles of random networks and derived their statistical-mechanics
entropy (Bianconi, 2008, 2009; Anand and Bianconi, 2009). This allows the
author to estimate the probability that a random network with certain con-
straints has of belonging to a particular ensemble, and thus assess the relative
importance of different magnitudes and help discern the mechanisms respon-
sible for a given real-world network. For instance, she shows that scale-free
networks arise naturally when the total entropy is restricted to a small finite
value. Here we take a similar approach: we obtain the Shannon information
entropy encoded in the distribution of edges. As we shall see, both methods
yield the same results (Jaynes, 1957; Anand and Bianconi, 2009), but for our
purposes the Shannon entropy is more tractable.
The Shannon entropy associated with a probability distribution pm is
s = − Xm
where the sum extends over all possible outcomes m. For a given pair of
nodes (i, j ), pm can be considered to represent the probability of there being
m edges between i and j . For simplicity, we shall focus here on networks such
pm ln(pm),
54
Chapter 5. Correlated networks and natural disassortativity
that aij can only take values 0 or 1, although the method is applicable to any
number of edges allowed. In this case, we have only two terms: p1 = ǫij and
p0 = 1 − ǫij , where ǫij ≡ E (aij ) is the expected value of the element aij given
that the network belongs to the ensemble of interest. The entropy associated
with pair (i, j ) is then
(5.1)
sij = − [ǫij ln(ǫij ) + (1 − ǫij ) ln(1 − ǫij )] ,
while the total entropy of the network is S = PN
ij sij :
N
Xij
S = −
[ǫij ln(ǫij ) + (1 − ǫij ) ln(1 − ǫij )] .
Since we have not imposed symmetry of the adjacency matrix, this expression
is in general valid for directed networks. For undirected networks, however,
the sum is only over i ≤ j , with the consequent reduction in entropy.
For the sake of illustration, we shall estimate the entropy of the Internet
at the autonomous system (AS) level and compare it with the values obtained
in (Bianconi, 2008, 2009; Anand and Bianconi, 2009) assuming the network
belongs to two different ensembles: the fully random graph, or Erdos-R´enyi
(ER) ensemble, and the configuration ensemble with a scale-free degree distri-
bution (p(k) ∼ k−γ ) (Newman, 2003c) and structural cutoff, ki < phkiN , ∀i
(Bianconi, 2008, 2009; Anand and Bianconi, 2009) (hki is the mean degree).
In this example, we assume the network to be sparse enough to expand the
term ln(1 − ǫij ) in Eq. (5.1) and keep only linear terms. This reduces Eq. (5.1)
to
N
Xij
In the ER ensemble, each of N nodes has an equal probability of receiving
each of 1
2 hkiN undirected edges. So, writing ǫER
ij = hki/N , we have
1
2 hkiN [ln (hki/N ) − 1] .
SER = −
The configuration ensemble, which imposes a given degree sequence (k1 , ...kN ),
is defined via the expected value of the adjacency matrix (Newman, 2003c;
Johnson et al., 2008):
ǫij [ln(ǫij ) − 1] + O(ǫ2
ij ).
Ssparse ≃ −
This value leads to
ǫc
ij = kikj /(hkiN ).
Sc = hkiN [ln(hkiN ) + 1] − 2N hk ln ki,
5.3 Entropic origin of disassortativity
55
SB
ER
SB
c
SER
Sc(γ=2.3)
20
20
20
20
20
20
20
20
15
15
15
15
15
15
15
15
10
10
10
10
10
10
10
10
e
e
e
e
e
e
e
e
d
d
d
d
d
d
d
d
o
o
o
o
o
o
o
o
n
n
n
n
n
n
n
n
r
r
r
r
r
r
r
r
e
e
e
e
e
e
e
e
p
p
p
p
p
p
p
p
y
y
y
y
y
y
y
y
p
p
p
p
p
p
p
p
o
o
o
o
o
o
o
o
r
r
r
r
r
r
r
r
t
t
t
t
t
t
t
t
n
n
n
n
n
n
n
n
E
E
E
E
E
E
E
E
11/97
11/97
11/97
11/97
11/97
11/97
11/97
11/97
10/98
10/99
10/99
10/99
10/99
10/99
10/99
10/99
10/99
10/98
10/98
10/98
10/98
10/98
10/98
10/98
Date (month/year)
Date (month/year)
Date (month/year)
Date (month/year)
Date (month/year)
Date (month/year)
Date (month/year)
Date (month/year)
10/00
10/00
10/00
10/00
10/00
10/00
10/00
10/00
3/01
3/01
3/01
3/01
3/01
3/01
3/01
3/01
Figure 5.1: Evolution of the Internet at the AS level. Empty (blue) squares
and circles: entropy per node of randomized networks in the fully random and
in the configuration ensembles, as obtained by Bianconi (hence the “B” su-
perscription) (Bianconi, 2008, 2009; Anand and Bianconi, 2009). Filled (red)
triangles and diamonds: Shannon entropy for an ER network and a scale-free
one with γ = 2.3, respectively.
where h·i ≡ N −1 Pi (·) stands for an average over nodes.
Fig. 5.1 displays the entropy per node obtained in (Bianconi, 2008, 2009;
Anand and Bianconi, 2009) for the first two levels of approximation (ensem-
bles) to the Internet at the AS level, first taking into account only the numbers
of nodes N and edges L = 1
2 hkiN , and then also the degree sequence. Along-
side these, we plot the Shannon entropy both for an ER random network,
(which coincides exactly with Bianconi’s expression), and for a scale-free net-
work with γ = 2.3 (the slight disparity arising from this exponent’s changing
a little with time).
5.3 Entropic origin of disassortativity
We shall now go on to analyse the effect of degree-degree correlations on the
entropy. In the configuration ensemble, the expected value of the mean degree
56
Chapter 5. Correlated networks and natural disassortativity
Figure 5.2: Shannon entropy of correlated scale-free networks against param-
eter β (left panel) and against Pearson’s coefficient r (right panel), for various
values of γ (increasing from bottom to top). hki = 10, N = 104 .
of the neighbours of a given node is
,
ij kj = hk2i
i Xj
ǫc
knn,i = k−1
hki
which is independent of ki . However, as mentioned above, real networks often
display degree-degree correlations, with the result that knn,i = knn(ki ).
If
knn(k) increases (decreases) with k , the network is assortative (disassortative).
A measure of this phenomenon is Pearson’s coefficient applied to the edges
(Newman, 2003c, 2002, 2003a; Boccaletti et al., 2006):
r =
l ] − [kl ]2
[klk ′
l ] − [kl ]2 ,
[k2
where kl and k ′
l are the degrees of each of the two nodes belonging to edge l,
and [·] ≡ (hkiN )−1 Pl (·) is an average over edges. Writing Pl (·) = Pij aij (·),
r can be expressed as
r = hkihk2knn(k)i − hk2i2
(5.2)
.
hkihk3i − hk2 i2
The ensemble of all networks with a given degree sequence (k1 , ...kN ) contains a
subset for all members of which knn (k) is constant (the configuration ensemble),
but also subsets displaying other functions knn (k). We can identify each one
5.3 Entropic origin of disassortativity
57
kj ǫij = kiknn (ki ), ∀i, and
ǫij = ki , ∀i (for consistency).
of these subsets (regions of phase space) with an expected adjacency matrix ǫ
which simultaneously satisfies the following conditions:
i) Xj
ii) Xj
An ansatz which fulfils these requirements is any matrix of the form
N (cid:20) (kikj )ν
j + hkν i(cid:21) ,
kikj
f (ν )
+ Z dν
hkν i − kν
i − kν
hkiN
where ν ∈ R and the function f (ν ) is in general arbitrary, although depending
on the degree sequence it shall here be restricted to values which maintain
ǫij ∈ [0, 1], ∀i, j . This ansatz yields
knn(k) = hk2 i
+ Z dν f (ν )σν+1 (cid:20) kν−1
k (cid:21)
1
hkν i −
hki
(the first term being the result for the configuration ensemble), where σb+1 ≡
hk b+1i − hkihk bi. In practice, one could adjust Eq. (5.4) to fit any given func-
tion knn(k) and then wire up a network with the desired correlations: it suffices
to throw random numbers according to Eq. (5.3) with f (ν ) as obtained from
the fit to Eq. (5.4)1 . To prove the uniqueness of a matrix ǫ obtained in this
way (i.e., that it is the only one compatible with a given knn (k)) assume that
there exists another valid matrix ǫ′
6= ǫ. Writting ǫ′
ij − ǫij ≡ h(ki , kj ) = hij ,
then i) implies that Pj kj hij = 0, ∀i, while ii) means that Pj hij = 0, ∀i. It
follows that hij = 0, ∀i, j .
ǫij =
(5.3)
(5.4)
In many empirical networks, knn (k) has the form knn (k) = A + Bkβ , with
A, B > 0 (Boccaletti et al., 2006; Pastor-Satorras et al., 2001) – the mixing
being assortative (disassortative) if β is positive (negative). Such a case is
fitted by Eq. (5.4) if
f (ν ) = C (cid:20)δ (ν − β − 1)
σβ+2 − δ (ν − 1)(cid:21) ,
σ2
1Although, as with the configuration ensemble, it is not always possible to wire a network
according to a given ǫ.
58
Chapter 5. Correlated networks and natural disassortativity
r =
(5.6)
(5.5)
with C a positive constant, since this choice yields
knn (k) = hk2i
+ C σ2 (cid:20) kβ
hki (cid:21) .
1
hkβ+1i −
hki
After plugging Eq. (5.5) into Eq. (5.2), one obtains:
hkβ+1i (cid:18) hkihkβ+2i − hk2ihkβ+1i
(cid:19) .
C σ2
hkihk3i − hk2i2
Inserting Eq. (5.3) in Eq. (5.1), we can calculate the entropy of correlated
networks as a function of β and C – or, by using Eq. (5.6), as a function of r .
Particularizing for scale-free networks, then given hki, N and γ , there is always
a certain combination of parameters β and C which maximizes the entropy; we
shall call these β ∗ and C ∗ . For γ . 5/2 this point corresponds to C ∗ = 1. For
higher γ , the entropy can be slightly higher for larger C . However, for these
values of γ , the assortativity r of the point of maximum entropy obtained with
C = 1 differs very little from the one corresponding to β ∗ and C ∗ (data not
shown). Therefore, for the sake of clarity but with very little loss of accuracy,
in the following we shall generically set C = 1 and vary only β in our search
for the level of assortativity, r∗ , that maximizes the entropy given hki, N and
γ . Note that C = 1 corresponds to removing the linear term, proportional
to kikj , in Eq. (5.3), and leaving the leading non-linearity, (kikj )β+1 , as the
dominant one.
Fig. 5.2 displays the entropy curves for various scale-free networks, both
as functions of β and of r : depending on the value of γ , the point of max-
imum entropy can be either assortative or disassortative. This can be seen
more clearly in Fig. 5.3, where r∗ is plotted against γ for scale-free networks
with various mean degrees hki. The values obtained by Park and Newman
(Park and Newman, 2003) as those resulting from the one-edge-per-pair re-
striction are also shown for comparison: notice that whereas this effect alone
cannot account for the Internet’s correlations for any γ , entropy considerations
would suffice if γ ≃ 2.1. As shown in the inset, the results are robust in the
large system-size limit.
Since most networks observed in the real world are highly heterogeneous,
with exponents in the range γ ∈ (2, 3), it is to be expected that these should
display a certain disassortativity – the more so the lower γ and the higher hki.
In Fig. 5.4 we test this prediction on a sample of empirical, scale-free net-
works quoted in Newman’s review (Newman, 2003c) (p. 182). For each case,
we found the value of r that maximizes S according to Eq. (5.1), after insert-
ing Eq. (5.3) with the quoted values of hki, N and γ . In this way, we obtained
5.3 Entropic origin of disassortativity
59
<k>=1/2 k0
<k>=1/2 k0
<k>=1/2 k0
<k>=1/2 k0
<k>=1/2 k0
<k>=k0
<k>=k0
<k>=k0
<k>=k0
<k>=2 k0
<k>=2 k0
<k>=2 k0
<k>=4 k0
<k>=4 k0
r0=-0.189
r0=-0.189
r0=-0.189
r0=-0.189
r0=-0.189
*
*
*
*
*
*
r
r
r
r
r
r
0.1
0.1
0.1
0.1
0.1
0.1
0
0
0
0
0
0
-0.1
-0.1
-0.1
-0.1
-0.1
-0.1
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
-0.3
-0.3
-0.3
-0.3
-0.3
-0.3
2
2
2
2
2
2
0.1
0.1
0.1
0.1
*
*
*
*
r
r
r
r
0
0
0
0
-0.1
-0.1
-0.1
-0.1
103
103
103
103
3
3
3
3
3
3
γ
γ
γ
γ
γ
γ
γ=2.5
γ=2.5
γ=2.5
γ=2.5
3´•104
3´•104
3´•104
3´•104
N
N
N
N
4
4
4
4
4
4
Figure 5.3: Lines from top to bottom: r at which the entropy is maximized,
r∗ , against γ for random scale-free networks with mean degrees hki = 1
2 , 1, 2
and 4 times k0 = 5.981, and N = N0 = 10697 nodes (k0 and N0 correspond
to the values for the Internet at the AS level in 2001 (Park and Newman,
2003), which had r = r0 = −0.189). Symbols are the values obtained in
(Park and Newman, 2003) as those expected solely due to the one-edge-per-
pair restriction (with k0 , N0 and γ = 2.1, 2.3 and 2.5). Inset: r∗ against N for
networks with fixed hki/N (same values as the main panel) and γ = 2.5; the
arrow indicates N = N0 .
the expected assortativity for six networks, representing: a peer-to-peer (P2P)
network, metabolic reactions, the nd.edu domain, actor collaborations, protein
interactions, and the Internet (see (Newman, 2003c) and references therein).
For the metabolic, Web domain and protein networks, the values predicted are
in excel lent agreement with the measured ones; therefore, no specific anticor-
relating mechanisms need to be invoked to account for their disassortativity.
In the other three cases, however, the predictions are not accurate, so there
must be additional correlating mechanisms at work. Indeed, it is known that
small routers tend to connect to large ones (Pastor-Satorras et al., 2001), so
one would expect the Internet to be more disassortative than predicted, as is
the case2 – an effect that is less pronounced but still detectable in the more
2However, as Fig. 5.3 shows, if the Internet exponent were the γ = 2.2 ± 0.1 reported
in (Pastor-Satorras et al., 2001) rather than γ = 2.5, entropy would account more fully for
60
Chapter 5. Correlated networks and natural disassortativity
*
*
*
*
*
*
r
r
r
r
r
r
0.2
0.2
0.2
0.2
0.2
0.2
0
0
0
0
0
0
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
-0.4
-0.4
-0.4
-0.4
-0.4
-0.4
2
2
2
2
2
2
u
d
e
.
d
n
c
i
l
o
b
a
t
e
m
s
r
o
t
c
a
t
e
n
r
e
t
n
i
n
i
e
t
o
r
p
P
2
P
2.25
2.25
2.25
2.25
2.25
2.25
γ
γ
γ
γ
γ
γ
2.5
2.5
2.5
2.5
2.5
2.5
Figure 5.4: Level of assortativity that maximizes the entropy, r∗ , for various
real-world, scale-free networks, as predicted theoretically by Eq. (5.1) (circles)
and as directly measured (horizontal lines), against exponent γ .
egalitarian P2P network. Finally, as is typical of social networks, the actor
graph is significantly more assortative than predicted, probably due to the ho-
mophily mechanism whereby highly connected, big-name actors tend to work
together (Newman, 2002, 2003a).
5.4 To sum up...
We have shown how the ensemble of networks with a given degree sequence
can be partitioned into regions of equally correlated networks and found, using
an information-theory approach, that the largest (maximum entropy) region,
for the case of scale-free networks, usually displays a certain disassortativity.
Therefore, in the absence of knowledge regarding the specific evolutionary
forces at work, this should be considered the most likely state. Given the
accuracy with which our approach can predict the degree of assortativity of
certain empirical networks with no a priori information thereon, we suggest
this as a neutral model to decide whether or not particular experimental data
require specific mechanisms to account for observed degree-degree correlations.
these correlations.
Chapter 6
Enhancing robustness to noise
via assortativity
As we saw in Chapter 4, the performance of attractor neural networks depends
crucially on the heterogeneity of the underlying topology’s degree distribution.
We take this analysis a step further by examining the effect of degree-degree
correlations – or assortativity – on neural-network behaviour. In Chapter 5 we
described a method for studying correlated networks and dynamics thereon,
both analytically and computationally, which is independent of how the topol-
ogy may have evolved. We now make use of this to show how the robustness to
noise is greatly enhanced in assortative (positively correlated) neural networks,
especially if it is the hub neurons that store the information.
6.1 Background
For a dozen years or so now, the study of complex systems has been heavily
influenced by results from network science – which one might regard as the fu-
sion of graph theory with statistical physics (Newman, 2003c; Boccaletti et al.,
2006). Phenomena as diverse as epidemics (Watts and Strogatz, 1998), cel-
lular function (Suel et al., 2006), power-grid failures (Buldyrev et al., 2010)
or internet routing (Bogun´a et al., 2010), among many others (Arenas et al.,
2008a), depend crucially on the structure of the underlying network of in-
teractions. One of the earliest systems to have been described as a net-
work was the brain, which is made up of a great many neurons connected
to each other by synapses (y Ca jal, 1995; Amit, 1989; Abbott and Kepler,
1990; Torres and Varona, 2010). Mathematically, the first neural networks
combined the Ising model (Baxter, 1982) with the Hebb learning rule (Hebb,
61
62
Chapter 6. Enhancing robustness to noise via assortativity
1949) to reproduce, very successfully, the storage and retrieval of informa-
tion (Amari, 1972; Hopfield, 1982; Amit, 1995). Neurons were simplified
to binary variables (like Ising spins) representing firing or non-firing cells.
By considering the trivial fully-connected topology, exact solutions could be
reached, which at the time seemed more important than attempting to in-
troduce biological realism. Subsequent work has tended to focus on consider-
ing richer dynamics for the cells rather than on the way in which these are
interconnected (Vogels et al., 2005; Torres et al., 2007; Mejias et al., 2010).
However, the topology of the brain – whether at the level of neurons and
synapses, cortical areas or functional connections – is obviously far from triv-
ial (Amaral et al., 2000; Sporns et al., 2004; Egu´ıluz et al., 2005; Arenas et al.,
2008b; Bullmore and Sporns, 2009; Johnson et al., 2010a).
The number of neighbours a given node in a network has is called its degree,
and much attention is paid to degree distributions since they tend to be highly
heterogeneous for most real networks. In fact, they are often approximately
scale-free (i.e., described by power laws) (Newman, 2003c; Boccaletti et al.,
2006; Peretto, 1992; Barab´asi and Oltvai, 2004). By including this topologi-
cal feature in a Hopfield-like neural-network model, Torres et al. Torres et al.
(2004) found that degree heterogeneity increases the system’s performance at
high levels of noise, since the hubs (high degree nodes) are able to retain
information at levels well above the usual critical noise. To prove this ana-
lytically, the authors considered the configurational ensemble of networks (the
set of random networks with a given degree distribution but no degree-degree
correlations) and showed that Monte Carlo (MC) simulations were in good
agreement with mean-field analysis, despite the approximation inherent to the
latter technique when the network is not fully connected. A similar approach
can also be used to show how heterogeneity may be advantageous for the per-
formance of certain tasks in models with a richer dynamics (Johnson et al.,
2008). It is worth mentioning that this influence of the degree distribution on
dynamical behaviour is found in many other settings, such as the more general
situation of systems of coupled oscillators (Barahona and Pecora, 2002).
Another property of empirical networks that is quite ubiquitous is the exis-
tence of correlations between the degrees of nodes and those of their neighbours
(Pastor-Satorras et al., 2001; Newman, 2002, 2003a). If the average degree-
degree correlation is positive the network is said to be assortative, while it is
called disassortative if negatively correlated. Most heterogeneous networks are
disassortative (Newman, 2003c), which, as described in Chapter 5, seems to be
6.1 Background
63
because this is in some sense their equilibrium (maximum entropy) state given
the constraints imposed by the degree distribution (Johnson et al., 2010b).
However, there are probably often mechanisms at work which drive systems
from equilibrium by inducing different correlations, as appears to be the case
for most social networks, in which nodes (people) of a kind tend to group
together. This feature, known as assortativity or mixing by degree, is also
relevant for processes taking place on networks. For instance, assortative net-
works have lower percolation thresholds and are more robust to targeted attack
(Newman, 2003a), while disassortative ones make for more stable ecosystems
and are – at least according to the usual definition – more synchronizable
(Brede and Sinha).
The approach usually taken when studying correlated networks computa-
tionally is to generate a network from the configuration ensemble and then
introduce correlations (positive or negative) by some stochastic rewiring pro-
cess (Maslov et al., 2004). A drawback of this method, however, is that results
may well then depend on the details of this mechanism: there is no guarantee
that one is correctly sampling the phase space of networks with given corre-
lations. For analytical work, some kind of hidden variables from which the
correlations originate are often considered (Caldarelli et al., 2002; Soderberg,
2002; Bogun´a and Pastor-Satorras, 2003; Fronczak and Fronczak, 2006) – an
assumption which can also be used to generate correlated networks compu-
tationally (Bogun´a and Pastor-Satorras, 2003). This can be a very powerful
method for solving specific network models. However, it may not be appropri-
ate if one wishes to consider all possible networks with given degree-degree cor-
relations, independently of how these may have arisen. In this chapter, we get
round the problem by making use of the method put forward by Johnson et al.
(2010b) (and described in Chapter 5) whereby the ensemble of all networks
with given correlations can be considered theoretically without recurring to
hidden variables (de Franciscis et al., 2011). Furthermore, we show how this
approach can be used computationally to generate random networks that are
representative of the ensemble of interest (i.e., they are model-independent). In
this way, we study the effect of correlations on a simple neural network model
and find that assortativity increases performance in the face of noise – partic-
ularly if it is the hubs that are mainly responsible for storing information (and
it is worth mentioning that there is experimental evidence suggestive of a main
functional role played by hub neurons in the brain (Morgan and Soltesz, 2008;
Bonifazi et al., 2009)). The good agreement between the mean-field analysis
64
Chapter 6. Enhancing robustness to noise via assortativity
and our MC simulations bears witness both to the robustness of the results as
regards neural systems, and to the viability of using this method for studying
dynamics on correlated networks.
6.2 Preliminary considerations
6.2.1 Model neurons on networks
The attractor neural network model put forward by Hopfield (Hopfield, 1982)
consists of N binary neurons, each with an activity given by the dynamic vari-
able si = ±1. Every time step (MCS), each neuron is updated according to
the stochastic transition probability P (si → ±1) = 1
2 [1 ± tanh (hi /T )] (paral-
lel dynamics), where the field hi is the combined effect on i of all its neighbours,
hi = Pj wij sj , and T is a noise parameter we shall call temperature, but which
represents any kind of random fluctuations in the environment. This is the
same as the Ising model for magnetic systems, and the transition rule can be
derived from a simple interaction energy such that aligned variables s (spins)
contribute less energy than if they were to take opposite values. However,
this system can store P given configurations (memory patterns) ξ ν
i = ±1 by
having the interaction strengths (synaptic weights) set according to the Hebb
rule (Hebb, 1949): wij ∝ PP
ν=1 ξ ν
i ξ ν
j . In this way, each pattern becomes an
attractor of the dynamics, and the system will evolve towards whichever one is
closest to the initial state it is placed in. This mechanism is called associative
memory, and is nowadays used routinely for tasks such as image identifica-
tion. What is more, it has been established that something similar to the
Hebb rule is implemented in nature via the processes of long-term potenti-
ation and depression at the synapses (Malenka and Nicoll, 1999; Roo et al.,
2008; Rodr´ıguez-Moreno and Paulsen, 2008; Kwag and Paulsen, 2009), and
this phenomenon is indeed required for learning (Gruart et al., 2006).
To take into account the topology of the network, we shall consider the
weights to be of the form wij = ωij aij , where the element aij of the adjacency
matrix represents the number of directed edges (usually interpreted as synapses
in a neural network) from node j to node i, while ω stores the patterns, as
before:
P
1
Xν=1
hki
For the sake of coherence with previous work, we shall assume a to be sym-
metric (i.e., the network is undirected), so each node is characterized by a
ωij =
i ξ ν
ξ ν
j .
6.2 Preliminary considerations
65
single degree ki = Pj aij . However, all results are easily extended to directed
networks – in which nodes have both an in degree, k in
i = Pj aij , and an out
degree, kout
= Pj aj i – by bearing in mind it is only a neuron’s pre-synaptic
i
neighbours that influence its behaviour. The mean degree of the network is
hki, where the angles stand for an average over nodes1 : h·i ≡ N −1 Pi (·).
6.2.2 Network ensembles
When one wishes to consider a set of networks which are randomly wired while
respecting certain constraints – that is, an ensemble – it is usually useful to
define the expected value of the adjacency matrix2 , E (a) ≡ ǫ. The element
ǫij of this matrix is the mean value of aij obtained by averaging over the en-
semble. For instance, in the Erdos-R´enyi (ER) ensemble all elements (outside
the diagonal) take the value ǫER
ij = hki/N , which is the probability that a
given pair of nodes be connected by an edge. For studying networks with a
given degree sequence, (k1 , ...kN ), it is common to assume the configuration
ensemble, defined as
ǫconf
ij =
kikj
hkiN
This expression can usually be applied also when the constraint is a given de-
gree distribution, p(k), by integrating over p(ki ) and p(kj ) where appropriate.
One way of deriving ǫconf is to assume one has ki dangling half-edges at each
node i; we then randomly choose pairs of half-edges and join them together
until the network is wired up. Each time we do this, the probability that we
join i to j is kikj /(hkiN )2 , and we must perform the operation hkiN times.
Bianconi showed that this is also the solution for Barab´asi-Albert evolved net-
works (Bianconi, 2002). However, we should bear in mind that this result is
only strictly valid for networks constructed in certain particular ways, such
as in these examples. It is often implicitly assumed that were we to average
over all random networks with a given degree distribution, the mean adjacency
matrix obtained would be ǫconf . However, as we discussed in Chapter 5, this
is not in fact necessarily true (Johnson et al., 2010b).
1 In directed networks the mean in degree and the mean out degree necessarily coincide,
whatever the forms of the in and out distributions.
2As in statistical physics, one can consider the microcanonical ensemble, in which each
element (network) satisfies the constraints exactly, or the canonical ensemble, where the
constraints are satisfied on average (Bianconi, 2009). Throughout this work, we shall refer
to canonical ensembles.
66
Chapter 6. Enhancing robustness to noise via assortativity
102
>
k
<
-
n
n
k
101
100
10-2
10-4
)
k
(
P
β=-0.5
β=0
β=0.5
k-2.5
101
102
103
k0.5
101
k-0.5
102
k
Figure 6.1: Mean-nearest-neighbour functions knn (k) for scale-free networks
with β = −0.5 (disassortative), 0.0 (neutral), and 0.5 assortative, generated
according to the algorithm described in Sec. 6.3.2. Inset: degree distribution
(the same in all three cases). Other parameters are γ = 2.5, hki = 12.5,
N = 104 .
6.2.3 Correlated networks
In the configuration ensemble, the expected value of the mean degree of the
i Pj ǫconf
neighbours of a given node is knn,i = k−1
kj = hk2i/hki, which is in-
ij
dependent of ki . However, as mentioned above, real networks often display
degree-degree correlations, with the result that knn,i = knn(ki ). If knn (k) in-
creases with k , the network is said to be assortative – whereas it is disassorta-
tive if it decreases with k (see Fig. 6.1). This is from the more general nomen-
clature (borrowed form sociology) in which sets are assortative if elements of
a kind group together, or assort. In the case of degree-degree correlated net-
works, positive assortativity means that edges are more than randomly likely
to occur between nodes of a similar degree.
The ensemble of all networks with a given degree sequence (k1 , ...kN ) con-
tains a subset for all members of which knn (k) is constant (the configuration
ensemble), but also subsets displaying other functions knn (k). We can iden-
tify each one of these subsets (regions of phase space) with an expected ad-
jacency matrix ǫ which simultaneously satisfies the following conditions:
i)
Pj kj ǫij = kiknn (ki ), ∀i (by definition of knn(k)), and ii) Pj ǫij = ki , ∀i (for
consistency). As we showed in Chapter 5, the general solution to this problem
6.2 Preliminary considerations
67
is a matrix of the form
ǫij =
N (cid:20) (kikj )ν
j + hkν i(cid:21) ,
kikj
f (ν )
+ Z dν
i − kν
hkν i − kν
hkiN
where ν ∈ R and the function f (ν ) is determined by knn (k) (Johnson et al.,
2010b). (If the network were directed, then ki = k in
and kj = kout
in this
i
j
expression.) This yields
(6.1)
(6.2)
+ Z dν f (ν )σν+1 (cid:20) kν−1
knn (k) = hk2i
k (cid:21)
1
hkν i −
hki
(the first term being the result for the configuration ensemble), where σb+1 ≡
hk b+1i − hkihk bi. This means that ǫ is not just one possible way of obtaining
correlations according to knn (k); rather, there is a two-way mapping between
ǫ and knn (k): every network with this particular function knn (k) and no other
ones are contained in the ensemble defined by ǫ. Thanks to this, if we are
able to consider random networks drawn according to this matrix (whether
we do this analytically or computationally; see Section 6.3.2), we can be con-
fident that we are correctly taking account of the whole ensemble of interest.
In other words, whatever the reasons behind the existence of degree-degree
correlations in a given network, we can study the effects of these with only
information on p(k) and knn (k) by obtaining the associated matrix ǫ. This is
not to say, of course, that all topological properties are captured in this way: a
particular network may have other features – such as higher order correlations,
modularity, etc. – the consideration of which would require concentrating on
a sub-partition of those with the same p(k) and knn (k). But this is not our
purpose here.
In many empirical networks, knn (k) has the form knn (k) = A + Bkβ , with
A, B > 0 (Boccaletti et al., 2006; Pastor-Satorras et al., 2001) – the mixing
being assortative if β is positive, and disassortative when negative. Such a
case is fitted by Eq. (6.2) if
δ (ν − β − 1) − δ (ν − 1)(cid:21) ,
f (ν ) = C (cid:20) σ2
σβ+2
with C a positive constant, since this choice yields
+ C σ2 (cid:20) kβ
hkβ+1i −
knn(k) = hk2i
hki
hki (cid:21) .
1
(6.3)
(6.4)
68
Chapter 6. Enhancing robustness to noise via assortativity
In Chapter 5 we discussed how the most likely configurations for networks
with scale-free degree distributions (p(k) ∼ k−γ ) and correlations given by
Eq.
(6.4) are generally disassortative. We also showed that the maximum
entropy is usually obtained for values of C close to one. Here, we shall use
this result to justify concentrating on correlated networks with C = 1, so that
the only parameter we need to take into account is β . It is worth mentioning
that Pastor-Satorras et al. originally suggested using this exponent as a way
of quantifying correlations (Pastor-Satorras et al., 2001), since this seems to
be the most relevant magnitude. Because β does not depend directly on p(k)
(as r does), and can be defined for networks of any size (whereas r , in very
heterogeneous networks, always goes to zero for large N due to its normaliza-
tion (Dorogovtsev et al., 2005)), we shall henceforth use β as our assortativity
parameter.
So, after plugging Eq. (6.3) into Eq. (6.1), we find that the ensemble of
networks exhibiting correlations given by Eq. (6.4) (and C = 1) is defined by
the mean adjacency matrix
1
N
σ2
σβ+2
ǫij =
[ki + kj − hki]
N (cid:20) (kikj )β+1
j + hkβ+1i(cid:21) .
1
hkβ+1i − kβ+1
i − kβ+1
6.3 Analysis and results
+
(6.5)
6.3.1 Mean field
Let us consider the single-pattern case (P = 1, ξi = ξ 1
i ). Substituting the
adjacency matrix a for its expected value ǫ (as given by Eq. (6.5)) in the ex-
pression for the local field at i – which amounts to a mean-field approximation
– we have
hi =
1
hki
)(cid:21) µ0
ξi (cid:26)(cid:20)(ki − hki) +
σ2
(hkβ+1i − kβ+1
i
σβ+2
i − hkβ+1i)µβ+1(cid:27) ,
σ2
(kβ
+ hkiµ1 +
σβ+2
µα ≡ hkα
i ξisi i
hkαi
for α = 0, 1, β + 1. These order parameters measure the extent to which the
system is able to recall information in spite of noise (Johnson et al., 2008).
where we have defined
6.3 Analysis and results
69
For the first order we have µ0 = m ≡ hξisi i, the standard overlap measure
in neural networks (analogous to magnetization in magnetic systems), which
takes account of memory performance. However, µ1 , for instance, weighs the
sum with the degree of each node, with the result that it measures information
per synapse instead of per neuron. Although the overlap m is often assumed
to represent, in some sense, the mean firing rate of neurological experiments,
it is possible that µ1 is more closely related to the empirical measure, since the
total electric potential in an area of tissue is likely to depend on the number of
synapses transmitting action potentials. In any case, a comparison between the
two order parameters is a good way of assessing to what extent the performance
of neurons depends on their degree – larger-degree model neurons can in general
store information at higher temperatures than ones with smaller degree can
(Torres et al., 2004).
Substituting si for its expected value according to the transition probability,
si → tanh(hi/T ), we have, for any α,
i ξisi i = hkα
hkα
i ξi tanh(hi /T )i;
or, equivalently, the following 3-D map of closed coupled equations for the
macroscopic overlap observables µ0 , µ1 and µβ+1 – which describes, in this
mean-field approximation, the dynamics of the system:
µ0(t + 1) = Z p(k) tanh[F (t)/(hkiT )]dk
1
hki Z p(k)k tanh[F (t)/(hkiT )]dk
µ1(t + 1) =
1
hkβ+1i Z p(k)kβ+1 tanh[F (t)/(hkiT )]dk ,
µβ+1(t + 1) =
(6.6)
with
F (t) ≡ (kµ0(t) + hkiµ1(t) − hkiµ0(t))
+
σ2
σβ+2
[kβ+1 (µβ+1(t) − µ0(t))
+ hkβ+1i(µ0(t) − µβ+1(t))].
This can be easily computed for any degree distribution p(k). Note that taking
β = 0 (the uncorrelated case) the system collapses to the 2-D map obtained by
70
Chapter 6. Enhancing robustness to noise via assortativity
Torres et al. (2004), while it becomes the typical 1-D case for a homogeneous
p(k) – say a fully-connected network (Hopfield, 1982). It is in principle pos-
sible to do similar mean-field analysis for any number P of patterns, but the
map would then be 3P -dimensional, making the problem substantially more
complex.
At a critical temperature Tc , the system will undergo the characteristic
second order phase transition from a phase in which it exhibits memory (akin
to ferromagnetism) to one in which it does not (paramagnetism). To obtain
this critical temperature, we can expand the hyperbolic tangent in Eqs. (6.6)
around the trivial solution (µ0 , µ1 , µβ+1) ≃ (0, 0, 0) and, keeping only linear
terms, write
µ0 = µ1/Tc ,
µ1 =
µβ+1 =
1
hki2Tc (cid:2)hki2µ1 + σ2µβ+1(cid:3) ,
1
Tchkihkβ+1i hσβ+2µ0
σ2
σβ+2 (cid:0)hkβ+1 i2 − hk2(β+1) i(cid:1) µ0
σβ+2 (cid:0)hkβ+1i2 − hk2(β+1) i(cid:1) µβ+1(cid:21) .
σ2
+ hkihkβ+1iµ1 −
+
Defining
A ≡
B ≡
σ2
hki2 ,
σ2
σβ+2
hk2(β+1) i − hkβ+1 i2
hkihkβ+1i
,
σβ+2
hkihkβ+1i
Tc will be the solution to the third order polynomial equation:
D ≡
,
c − (B + 1)T 2
T 3
c + (B − A)Tc + A(B − D) = 0.
(6.7)
Note that for neutral (i.e., uncorrelated) networks, β = 0, and so A = B = D .
We then have Tc = hk2i/hki2 , as expected (Johnson et al., 2008).
6.3 Analysis and results
71
6.3.2 Generating correlated networks
Given a degree distribution p(k), the ensemble of networks compatible with
this constraint and with degree-degree correlations according to Eq.
(6.4)
(with some exponent β ) is defined by the mean adjacency matrix ǫ of Eq.
(6.5) – as described in Section 6.2.3 and by Johnson et al. (2010b). Therefore,
although there will generally be an enormous number of possible networks in
this volume of phase space, we can sample them correctly simply by generating
them according to ǫ. To do this, first we have to assign to each node a degree
drawn from p(k). If the elements of ǫ were probabilities, it would suffice then
to connect each pair of nodes (i, j ) with probability ǫij to generate a valid
network. Strictly speaking, ǫ is an expected value, which in certain cases can
be greater than one. To get round this, we write a probability matrix p = ǫ/a
with a some value such that all elements of p are smaller than one.
If we
then take random pairs of nodes (i, j ) and, with probability pij , place an edge
between them, repeating the operation until 1
2 hkiN edges have been placed,
the expected value of edges joining i and j will be ǫij . This method is like the
hidden variable technique (Bogun´a and Pastor-Satorras, 2003) in that edges
are placed with a predefined probability (which is why the resulting ensemble
is canonical). The difference lies in the fact that in the method here described
correlations only depend on the degrees of nodes.
We are interested here in neural networks, in which a given pair of nodes
can be joined by several synapses, so we shall not impose the restriction of
so-called simple networks of allowing only one edge at most per pair. We
shall, however, consider networks with a structural cutoff: ki < phkiN , ∀i
(Bianconi, 2008). This ensures that, at least for β ≤ 0, all elements of ǫ are
indeed smaller than one.
Because we can expect effects due to degree-degree correlations to be largest
when p(k) is very broad, and since most networks in nature and technol-
ogy seem to exhibit approximately power-law degree distributions (Newman,
2003c; Arenas et al., 2008a; Peretto, 1992; Barab´asi and Oltvai, 2004), we shall
here test our general theoretical results against simulations of scale-free net-
works: p(k) ∼ k−γ . This means that a network (or the region of phase space
to which it belongs) is characterized by the set of parameters {hki, N , γ , β }.
72
Chapter 6. Enhancing robustness to noise via assortativity
1
0.5
0
0
N=104
N=3x104
N=5x104
β=0.5
4
8
12
1
0.5
1
µ
β=-0.5
β=0
β=0.5
0
0
1
2
3
T
4
5
6
7
Figure 6.2: Stable stationary value of the weighted overlap µ1 against tem-
perature T for scale-free networks with correlations according to knn ∼ kβ , for
β = −0.5 (disassortative), 0.0 (neutral), and 0.5 (assortative). Symbols from
MC simulations, with errorbars representing standard deviations, and lines
from Eqs. (6.6). Other network parameters as in Fig. 6.1. Inset: µ1 against T
for the assortative case (β = 0.5) and different system sizes: N = 104 , 3 · 104
and 5 · 104 .
6.3.3 Assortativity and dynamics
In Fig. 6.2 we plot the stationary value of µ1 against the temperature T ,
as obtained from simulations and Eqs. (6.6), for disassortative, neutral and
assortative networks. The three curves are similar at low temperatures, but
as T increases their behaviour becomes quite different. The disassortative
network is the least robust to noise. However, the assortative one is capable
of retaining some information at temperatures considerably higher than the
critical value, Tc = hk2i/hki, of neutral networks. A comparison between µ1
and µ0 (see Fig. 6.3) shows that it is the high degree nodes that are mainly
responsible for this difference in performance. This can be seen more clearly in
Fig. 6.4, which displays the difference µ1 − µ0 against T for the same networks.
It seems that, because in an assortative network a sub-graph of hubs will
have more edges than in a disassortative one, it has a higher effective critical
temperature. Therefore, even when most of the nodes are acting randomly,
the set of nodes of sufficiently high degree nevertheless displays associative
memory.
6.3 Analysis and results
73
1
0.5
x
µ
µ
0µ
1µβ+1
β=0.5
0
0 1 2 3 4 5 6 7
T
Figure 6.3:
Stable stationary values of order parameters µ0 , µ1 and µβ+1
against temperature T , for assortative networks according to β = 0.5. Symbols
from MC simulations, with errorbars representing standard deviations, and
lines from Eqs. (6.6). Other parameters as in Fig. 6.1.
0.3
0
µ
0.15
-
1
µ
β=-0.5
β=0
β=0.5
2N-1/2
0
0
4
T
8
10
Figure 6.4: Difference between the stationary values µ1 and µ0 for networks
with β = −0.5 (disassortative), 0.0 (neutral) and 0.5 (assortative), against
temperature. Symbols from MC simulations, with errorbars representing stan-
dard deviations, and lines from Eqs. (6.6). Line shows the expected level of
fluctuations due to noise, ∼ N − 1
2 . Other parameters as in Fig. 6.1.
74
Chapter 6. Enhancing robustness to noise via assortativity
The phase diagram if Fig. 6.5 shows the critical temperature, Tc , as ob-
tained from Eq.
(6.7).
In addition to the effect reported by Torres et al.
(2004) whereby the Tc of scale-free networks grows with degree heterogeneity
(decreasing γ ), it also increases very significantly with positive degree-degree
correlations (increasing β ).
At large values of N , the critical temperature scales as Tc ∼ N b , with b ≥ 0
a constant. However, because the moments of k appearing in the coefficients of
Eq. (6.7) can have different asymptotic behaviour depending on the values of
γ and β , the scaling exponent b differs from one region to another in the space
of these parameters. These are the seven regions shown in Fig. 6.6, along with
the scaling behaviour exhibited by each one. This can be seen explicitely in
Fig. 6.7, where Tc , as obtained from MC simulations, is plotted against N for
cases in each of the regions with γ < 3. In each case, the scaling is as given by
Eq. (6.7) and shown in Fig. 6.6. For the four regions with γ < 3, from lowest
to highest assortativity we have scaling exponents which are dependent on:
only γ (region I), only β (region II), both γ and β (region III), and, perhaps
most interestingly, neither of the two (region IV) – with Tc scaling, in the latter
case, as √N . As for the more homogeneous γ > 3 part, regions V and VI have
a diverging critical temperature despite the fact that the second moment of
p(k) is finite, simply as a result of assortativity.
The case in which more than one pattern are stored (P > 1) can be explored
numerically. Assuming there are P uncorrelated patterns, we have an order
parameter µν
1 for each pattern ν . A global measure of the degree to which
there is memory can be captured by the parameter ζ , where
ζ 2 ≡
1
1 + P /N
(µν
1 )2 .
P
Xν=1
Notice that the normalization factor is due to the fact that if one pattern is
condensed – i.e., µ1 . 1 – the others have µν ∼ 1/√N , ν = 2, ..P , and so ζ ≃
1. Figure 6.8 shows how ζ decreases with T in variously correlated networks
for P = 3 (left panel) and P = 10 patterns (right panel). The behaviour is
not qualitatively different from that observed for the single-pattern case in the
main panel of Fig. 6.2, suggesting that the influence of assortativity we report
is robust as to the number of patterns stored, P .
6.3 Analysis and results
75
P
γ=2.5
8
6
γ=3
4
T
γ=3.5
2
F
0.5
0
1
-1
-0.5
0
β
Figure 6.5: Phase diagrams for scale-free networks with γ = 2.5, 3, and 3.5.
Lines show the critical temperature Tc marking the second-order transition
from a memory (ferromagnetic) phase to a memoryless (paramagnetic) one,
against the assortativity β , as given by Eq. (6.7). Other parameters as in Fig.
6.1.
β+2
β+3
2β+3
4
3.5
(0) finite Tc
∝
Tc
N(3+2β-γ)/2
(V)
β/2
∝ N
(VI)Tc
γ
3
∝ N(3-γ)/4
(I) Tc
2.5
2
-1
(II)
∝ N-β/2
Tc
-0.5
∝ N(3+β-γ)/2
(III)Tc
∝ N1/2
(IV)Tc
0.5
1
0
β
Figure 6.6: Parameter space β − γ partitioned into the regions in which b(β , γ )
has the same functional form – where b is the scaling exponent of the critical
temperature: Tc ∼ N b . Exponents obtained by taking the large N limit in Eq.
(6.7).
76
Chapter 6. Enhancing robustness to noise via assortativity
β=-0.8
β=-0.35
β=0
β=0.9
N(-γ+3)/4
N(-β)/2
N(-γ+β+3)/2
N1/2
10
c
T
1
102
103
104
105
N
Figure 6.7:
Examples of how Tc scales with N for networks belonging to
regions I, II, III and IV of Fig. 6.6 (β = −0.8, −0.35, 0.0 and 0.9, respec-
tively). Symbols from MC simulations, with errorbars representing standard
deviations, and slopes from Eq. (6.7). All parameters – except for β and N –
are as in Fig. 6.1.
6.4 Discussion
We have shown that assortative networks of simple model neurons are able to
exhibit associative memory in the presence of levels of noise such that uncor-
related (or disassortative) networks cannot. This may appear to be in con-
tradiction with a recent result obtained using spectral graph analysis – that
synchronizability of a set of coupled oscillators is highest for disassortative
networks (Brede and Sinha). A synchronous state of model oscillators and a
memory phase of model neurons are both sets of many simple dynamical el-
ements coupled via a network in such a way that a macroscopically coherent
situation is maintained (Barahona and Pecora, 2002). Obviously both systems
require the effective transmission of information among the elements. So why
are opposite results as regards the influence of topology reported for each sys-
tem? The answer is simple: whereas the definition of a synchronous state is
that every single element oscillate at the same frequency, it is precisely when
most elements are actually behaving randomly that the advantages to assor-
tativity we report become apparent. In fact, it can be seen in Fig. 6.2 that at
low temperatures disassortative networks perform the best, although the ef-
6.4 Discussion
77
1
2
/
1
)
2
)
0.5
1
ν
µ
(
ν
Σ
(
0
0
β=-0.5β=0
β=0.5
4
T
8
β=-0.5β=0
β=0.5
0
4
T
8
Figure 6.8: Global order parameter ζ for assortative (β = 0.5), neutral
(β = 0.0) and disassortative (β = −0.5) networks with P = 3 (left panel) and
P = 10 (right panel) stored patterns. Symbols from MC simulations, with
errorbars representing standard deviations. All parameters are as in Fig. 6.1.
fect is small. This is reminiscent of percolation: at high densities of edges the
giant component is larger in disassortative networks, but in assortative ones a
non-vanishing fraction of nodes remain interconnected even at densities below
the usual percolation threshold (Newman, 2002, 2003a). Because in the case
of targeted attacks it is this threshold which is taken as a measure of resilience,
we say that assortative networks perform the best. The relevance of partial
synchronization and the important role of hubs have already been noted for
systems of (weakly) coupled oscillators (G´omez-Gardenes et al., 2007; Pereira,
2010) – for which, however, assortativity has not been expected to be of con-
sequence (Pereira, 2010). In general, the optimal network for good conditions
(i.e., complete synchronization, high density of edges, low levels of noise) is not
necessarily the one which performs the best in bad conditions (partial synchro-
nization, low density of edges, high levels of noise). It seems that optimality
– whether in resilience or robustness – should thus be defined for particular
conditions.
We have used the technique suggested by Johnson et al. (2010b) to study
the effect of correlations on networks of model neurons, but many other sys-
78
Chapter 6. Enhancing robustness to noise via assortativity
tems of dynamical elements should be susceptible to a similar treatment. In
fact, Ising spins (Bianconi, 2002), Voter Model agents (Suchecki et al., 2005),
or Boolean nodes (Peixoto, 2010), for instance, are similar enough to binary
neurons that we should expect similar results for these models. If a moral can
be drawn, it is that persistence of partial synchrony, or coherence of a subset of
highly connected dynamical elements, can sometimes be as relevant (or more
so) as the possibility of every element behaving in the same way. In the case of
real brain cells, experiments suggest that hub neurons play key functional roles
(Morgan and Soltesz, 2008; Bonifazi et al., 2009). From this point of view,
there may be a selective pressure for brain networks to become assortative –
although, admittedly, this organ engages in such complex behaviour that there
must be many more functional constraints on its structure than just a high
robustness to noise. Nevertheless, it would be interesting to investigate this as-
pect of biological systems experimentally. For this, it should be borne in mind
that heterogeneous networks have a natural tendency to become disassortative,
so it is against the expected value of correlations discussed by Johnson et al.
(2010b) that empirical data should be contrasted in order to look for meaning-
ful deviations towards assortativity. Similarly, it may be necessary to take into
account the correlations that could emerge due to the spatial layout of neurons
(Kaiser et al., 2007; Johnson et al., 2011). In any case, it would be in areas of
the cortex specifically related to memory – such as the temporal (long-term
memory) (Miyashita, 1988; Sakai and Miyashita, 1991) or prefrontal (short-
term memory) (Camperi and Wang, 1998b; Compte et al., 2003) lobes – that
this effect might be relevant. A curious fact that would seem to support our hy-
pothesis is that whereas the vast ma jority of non-social networks are disassor-
tative (Newman, 2003c), one that appears actually to be strongly assortative
is the functional network of the human cortex (Egu´ıluz et al., 2005).
Chapter 7
Cluster Reverberation: A
mechanism for robust
short-term memory without
synaptic learning
Short-term memory cannot in general be explained the way long-term memory
can – as a gradual modification of synaptic conductances – since it takes place
too quickly. Theories based on some form of cellular bistability, however, do
not seem to be able to account for the fact that noisy neurons can collectively
store information in a robust manner. We show how a sufficiently clustered
network of simple model neurons can be instantly induced into metastable
states capable of retaining information for a short time. Cluster Reverbera-
tion, as we call it, could constitute a viable mechanism available to the brain
for robust short-term memory with no need of synaptic learning. Relevant phe-
nomena described by neurobiology and psychology, such as power-law statistics
of forgetting avalanches, emerge naturally from this mechanism.
7.1 Slow but sure, or fast and fleeting?
Of all brain phenomena, memory is probably one of the best understood (Amit,
1989; Abbott and Kepler, 1990; Torres and Varona, 2010). Consider a set of
many neurons, defined as elements with two possible states (firing or not firing,
one or zero) connected among each other in some way by synapses which carry
a proportion of the current let off by a firing neuron to its neighbours; the
probability that a given neuron has of firing at a certain time is then some
79
80
Chapter 7. Cluster Reverberation: A mechanism for robust
short-term memory without synaptic learning
function of the total current it has just received. Such a simplified model of
the brain is able to store and retrieve information, in the form of patterns of
activity (i.e., particular configurations of firing and non-firing neurons) when
the synaptic conductances, or weights, have been appropriately set according
to a learning rule (Hebb, 1949). Because each of the stored patterns becomes
an attractor of the dynamics, the system will evolve towards whichever of the
patterns most resembles the initial configuration. Artificial systems used for
tasks such as pattern recognition and classification, as well as more realistic
neural network models that take into account a variety of subcellular processes,
all tend to rely on this basic mechanism, known as Associative Memory (Amari,
1972; Hopfield, 1982).
Synaptic conductances in animal brains have indeed been found to be-
come strengthened or weakened during learning, via the biochemical processes
of long-term potentiation (LTP) and depression (LTD) (Malenka and Nicoll,
1999; Gruart et al., 2006; Roo et al., 2008; Rodr´ıguez-Moreno and Paulsen,
2008; Kwag and Paulsen, 2009). Further support for the hypothesis that such a
mechanism underlies long-term memory (LTM) comes from psychology, where
it is being found more and more that so-called connectionist models fit in well
with observed brain phenomena (Marcus and G.F., 2001; Frank, 1997). How-
ever, some memory processes take place on timescales of seconds or less and in
many instances cannot be accounted for by LTP and LTD (Durstewitz et al.,
2000), since these require at least minutes to be effected (Lee et al., 1980;
Klintsova and Greenough, 1999). For example, Sperling found that visual
stimuli are recalled in great detail for up to about one second after expo-
sure (iconic memory) (Sperling, 1960); similarly, acoustic information seems
to linger for three or four seconds (echoic memory) (Cowan, 1984). In fact,
it appears that the brain actually holds and continually updates a kind of
buffer in which sensory information regarding its surroundings is maintained
(sensory memory) (Baddeley, 1999). This is easily observed by simply closing
one’s eyes and recalling what was last seen, or thinking about a sound after it
has finished. Another instance is the capability referred to as working mem-
ory (Durstewitz et al., 2000; Baddeley and A.D., 2003):
just as a computer
requires RAM for its calculations despite having a hard drive for long term
storage, the brain must continually store and delete information to perform
almost any cognitive task. To some extent, working memory could consist in
somehow labelling or bringing forward previously stored concepts, like when
one is asked to remember a particular sequence of digits or familiar shapes.
7.1 Slow but sure, or fast and fleeting?
81
But we are also able to manipulate – if perhaps not quite so well – shapes and
symbols we have only just become acquainted with, too recently for them to
have been learned synaptically. We shall here use short-term memory (STM)
to describe the brain’s ability to store information on a timescale of seconds
or less1 .
Evidence that short-term memory is related to sensory information while
long-term memory is more conceptual can again be found in psychology. For
instance, a sequence of similar sounding letters is more difficult to retain for
a short time than one of phonetically distinct ones, while this has no bear-
ing on long-term memory, for which semantics seems to play the main role
(Conrad, 1964a,b); and the way many of us think about certain concepts,
such as chess, geometry or music, is apparently quite sensorial: we imagine
positions, surfaces or notes as they would look or sound. Most theories of
short-term memory – which almost always focus on working memory – make
use of some form of previously stored information (i.e., of synaptic learning)
and so can account for the labelling tasks referred to above but not for the
instant recall of novel information (Wang, 2001; Barak and Tsodyks, 2007;
Roudi and Latham; Mongillo et al., 2008; Mejias and Torres, 2009). Attempts
to deal with the latter have been made by proposing mechanisms of cel lular
bistability : neurons are assumed to retain the state they are placed in (such
as firing or not firing) for some period of time thereafter (Camperi and Wang,
1998a; Teramae and Fukai, 2005; Tarnow, 2008). Although there may indeed
be subcellular processes leading to a certain bistability, the main problem with
short-term memory depending exclusively on such a mechanism is that if each
neuron must act independently of the rest the patterns will not be robust to
random fluctuations (Durstewitz et al., 2000) – and the behaviour of individual
neurons is known to be quite noisy (Compte et al., 2003). It is worth pointing
out that one of the strengths of Associative Memory is that the behaviour
of a given neuron depends on many neighbours and not just on itself, which
means that robust global recall can emerge despite random fluctuations at an
1We should mention that sensory memory is usually considered distinct from STM – and
probably has a different origin – but we shall use “short-term memory” generically since
the mechanism we propose in this paper could be relevant for either or both phenomena.
On the other hand, the recent flurry of research in psychology and neuroscience on working
memory has lead to this term sometimes being used to mean short-term memory; strictly
speaking, however, working memory is generally considered to be an aspect of cognition
which operates on information stored in STM.
82
Chapter 7. Cluster Reverberation: A mechanism for robust
short-term memory without synaptic learning
individual level.
Something that, at least until recently, most neural network models have
failed to take into account is the structure of the network – its topology – it
often being assumed that synapses are placed among the neurons completely
at random, or even that all neurons are connected to all the rest (a mathe-
matically convenient but unrealistic situation). Although relatively little is yet
known about the architecture of the brain at the level of neurons and synapses,
experiments have shown that it is heterogeneous (some neurons have very many
more synapses than others), clustered (two neurons have a higher chance of be-
ing connected if they share neighbours than if not) and highly modular (there
are groups, or modules, with neurons forming synapses preferentially to those
in the same module) (Sporns et al., 2004; Johnson et al., 2010a). This chapter
describes the main result of Ref. (Johnson et al., 2011) – namely, that it suf-
fices to use a more realistic topology, in particular one which is modular and/or
clustered, for a randomly chosen pattern of activity the system is placed in to
be metastable. This means that novel information can be instantly stored and
retained for a short period of time in the absence of both synaptic learning and
cellular bistability. The only requisite is that the patterns be coarse grained
versions of the usual patterns – that is, whereas it is often assumed that each
neuron in some way represents one bit of information, we shall allocate a bit
to a small group or neurons2 (four or five can be enough).
If
The mechanism, which we call Cluster Reverberation, is very simple.
neurons in a group are more highly connected to each other than to the rest
of the network, either because they form a module or because the network
is significantly clustered, they will tend to retain the activity of the group:
when they are all initially firing, they each continue to receive many action
potentials and so go on firing, whereas if they start off silent, there is not
usually enough input current from the outside to set them off. The fact that
each neuron’s state depends on its neighbours conferres to the mechanism a
certain robustness in the face of random fluctuations. This robustness is par-
ticularly important for biological neurons, which as mentioned are quite noisy.
Furthermore, not only does the limited duration of short-term memory states
emerge naturally from this mechanism (even in the absence of interference
from new stimuli) but this natural forgetting follows power-law statistics, as
2This does not, of course, mean that memories are expected to be encoded as bitmaps.
Just as with individual neurons, positions or orientations, say, could be represented by the
activation of particular sets of clusters.
7.2 The simplest neurons on modular networks
83
in experimental settings (Wixted and Ebbesen, 1991, 1997; Sikstrom, 2002).
The process is reminiscent both of block attractors in ordinary neural net-
works (Dominguez et al., 2009) and of domains in magnetic materials (A. and R.,
1998), while Munoz et al. have recently highlighted a similarity with Griffiths
phases on networks (Munoz et al., 2010). It can also be interpreted as a multi-
scale phenomenon: the mesoscopic clusters take on the role usually played by
individual neurons, yet make use of network properties. Although the mecha-
nism could also work in conjunction with other ones, such as synaptic learning
or cellular bistability, we shall illustrate it by considering the simplest model
which has the necessary ingredients: a set of binary neurons linked by synapses
of uniform weight according to a topology whose modularity or clustering we
shall tune. As with Associative Memory, this mechanism of Cluster Reverber-
ation appears to be simple and robust enough not to be qualitatively affected
by the complex subcellular processes incorporated into more realistic neuron
models – such as integrate-and-fire or Hodgkin-Huxley neurons. However, such
refinements are probably needed to achieve graded persistent activity, since the
mean frequency of each cluster could then be set to a certain value.
7.2 The simplest neurons on modular networks
We consider a network of N model neurons, with activities si = ±1. The topol-
ogy is given by the adjacency matrix aij = {1, 0}, each element representing
the existence or absence of a synapse from neuron j to neuron i (a need not
be symmetric). In this kind of model, each edge usually has a synaptic weight
associated, ωij ∈ R; however, we shall here consider these to have all the same
value: ωij = ω ∀i, j . Neurons are updated in parallel (Little dynamics) at each
time step, according to the stochastic transition rule
T (cid:19) +
tanh (cid:18) hi
P (si → ±1) = ±
1
2
1
2
,
where the field of neuron i is defined as
aij sj
hi = ω
N
Xj
and T is a parameter we shall call temperature.
First of all, we shall consider the network defined by a to be made up
of M distinct modules. To achieve this, we can first construct M separate
random directed networks, each with n = N/M nodes and mean degree (mean
84
Chapter 7. Cluster Reverberation: A mechanism for robust
short-term memory without synaptic learning
number of neighbours) hki. Then we evaluate each edge and, with probability
λ, eliminate it, to be substituted for another edge between the original post-
synaptic neuron and a new pre-synaptic neuron chosen at random from among
any of those in other modules3 . Note that this protocol does not alter the
number of pre-synaptic neighbours of each node, k in
i = Pj aij (although the
number of post-synaptic neurons, kout
i = Pj aj i , can vary). The parameter λ
can be seen as a measure of modularity of the partition considered, since it
coincides with the expected value of the proportion of edges that link different
modules.
In particular, λ = 0 defines a network of disconnected modules,
while λ = 1 − M −1 yields a random network in which this partition has no
modularity. If λ ∈ (1 − M −1 , 1), the partition is less than randomly modular
– i.e., it is quasi-multipartite (or multipartite if λ = 1).
If the size of the modules is of the order of hki, the network will also be
highly clustered. Taking into account that the network is directed, let us
define the clustering coefficient Ci as the probability, given that there is a
synapse from neuron i to a neuron j and from another neuron l to i, that
there be a synapse from j to l:
that is, that there exist a feedback loop
i → j → l → i. Then, assuming M ≫ 1, the expected value of the clustering
coefficient C ≡ hCii is
C & hki − 1
n − 1
(1 − λ)3 .
7.3 Cluster Reverberation
A memory pattern, in the form of a given configuration of activities, {ξi =
±1}, can be stored in this system with no need of prior learning. Imagine a
pattern such that the activities of all n neurons found in any module are the
same, i.e., ξi = ξµ(i) , where the index µ(i) denotes the module that neuron i
belongs to. This can be thought of as a coarse graining of the standard idea of
memory patterns, in which each neuron represents one bit of information. In
our scheme, each module represents – and stores – one bit. The system can be
induced into this configuration via the application of an appropriate stimulus
(see Fig. 7.1): the field of each neuron will be altered for just one time step
according to
hi → hi + δξµ(i) ,
∀i,
3We do not allow self-edges (although these can occur in reality) since they can be
regarded as a form of cellular bistability.
7.3 Cluster Reverberation
85
Figure 7.1: Diagram of a modular network composed of four five-neuron clus-
ters. The four circles enclosed by the dashed line represent the stimulus: each
is connected to a particular module, which adopts the input state (red or blue)
and retains it after the stimulus has disappeared via Cluster Reverberation.
where the factor δ is the intensity of the stimulus. This mechanism for dy-
namically storing information will work for values of parameters such that the
system is sensitive to the stimulus, acquiring the desired configuration, yet also
able to retain it for some interval of time thereafter.
The two main attractors of the system are si = 1 ∀i and si = −1 ∀i. These
are the configurations of minimum energy (see the next section for a more
detailed discussion on energy). However, the energy is locally minimised for
any configuration in which si = dµ(i) ∀i with dµ = ±1; that is, configurations
such that each module comprises either all active or all inactive neurons. These
are the configurations that we shall use to store information. We define the
mean activity4 of each module,
mµ ≡
1
n
n
Xi∈µ
si ,
4The mean activity in a neural network model is usually taken to represent the mean
firing rate measured in experiments (Torres and Varona, 2010).
86
Chapter 7. Cluster Reverberation: A mechanism for robust
short-term memory without synaptic learning
which is a mesoscopic variable, as well as the global mean activity,
1
N
si =
1
M
m ≡
N
M
Xi
Xµ
(these magnitudes change with time, but, where possible, we shall avoid writ-
ing the time dependence explicitely for clarity). The extent to which the
network, at a given time, retains the pattern {ξi} with which it was stimulated
is measured with the overlap parameter
mµ
1
N
1
M
ξisi =
mstim ≡
M
N
Xµ
Xi
Ideally, the system should be capable of reacting immediately to a stimulus by
adopting the right configuration, yet also be able to retain it for long enough
to use the information once the stimulus has disappeared. A measure of per-
formance for such a task is therefore
ξµmµ .
1
τ
η ≡
mstim (t),
t0+τ
Xt=t0+1
where t0 is the time at which the stimulus is received and τ is the period of
time we are interested in (η ≤ 1) (Johnson et al., 2008).
If the intensity
of the stimulus, δ , is very large, then the system will always adopt the right
pattern perfectly and η will only depend on how well it can then retain it.
In this case, the best network will be one that is made up of unconnected
modules. However, since the stimulus in a real brain can be expected to arrive
via a relatively small number of axons, either from another part of the brain or
directly from sensory cells, it is more realistic to assume that δ is of a similar
order as the input a typical neuron receives from its neighbours, hhi ∼ ω hki.
Fig. 7.2 shows the mean performance obtained when the network is repeat-
edly stimulated with different randomly generated patterns. For low enough
values of the modularity λ and stimuli of intensity δ & ω hki, the system can
capture and successfully retain any pattern it is “shown” for some period of
time, even though this pattern was in no way previously learned. For less in-
tense stimuli (δ < ω hki), performance is nonmonotonic with modularity: there
exists an optimal value of λ at which the system is sensitive to stimuli yet still
able to retain new patterns quite well.
It is worth noting that performance can also break down due to thermal
fluctuations. The two main attractors of the system (si = 1 ∀i and si = −1 ∀i)
7.3 Cluster Reverberation
87
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
η
η
η
η
η
η
η
η
η
η
η
η
η
η
η
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
n
n
n
n
n
n
n
n
n
n
n
n
n
n
n
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
1
1
1
m
m
m
i
i
i
0.5
0.5
0.5
t
t
t
s
s
s
m
m
m
0
0
0
500
500
500
δ=8.5
δ=8.5
δ=8.5
δ=9
δ=9
δ=10
Time (MCS)
Time (MCS)
Time (MCS)
3500
3500
3500
λ=0.0
λ=0.0
λ=0.0
λ=0.25
λ=0.25
λ=0.5
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Figure 7.2: Performance η against λ for networks of the sort described in the
main text with M = 160 modules of n = 10 neurons, hki = 9; patterns are
shown with intensities δ = 8.5, 9 and 10, and T = 0.02 (lines – splines – are
drawn as a guide to the eye).
Inset: typical time series of mstim (i.e., the
overlap with whichever pattern was last shown) for λ = 0.0, 0.25, and 0.5, and
δ = hki = 9.
suffer the typical second order phase transition of the Hopfield model (Hopfield,
1982), from a memory phase (one in which m = 0 is not stable and stable
solutions m 6= 0 exist) to one with no memory (with m = 0 the only stable
solution), at the critical temperature (Johnson et al., 2008)
Tc = ω hk2
ini
hki
(Note that, in a directed network, hkini = hkouti ≡ hki, although the other
moments can in general be different.) The metastable states we are interested
in, though, have a critical temperature
.
T ′
c = (1 − λ)Tc
(assuming that the mean activity of the network is m ≃ 0). That is, the
temperature at which the modules are no longer able to retain their individual
activity is in general lower than that at which the the solution m = 0 for the
whole network becomes stable.
88
Chapter 7. Cluster Reverberation: A mechanism for robust
short-term memory without synaptic learning
7.4 Energy and topology
Each pair of nodes contributes a configurational energy eij = −ω 1
2 (aij +aj i )sisj ;
that is, if there is an edge from i to j and they have opposite activities, the
energy is increased in 1
2 ω , whereas it is decreased by the same amount if their
activities are the same. Given a configuration, we can obtain its associated
energy by summing over all pairs. We shall be interested in configurations with
x neurons that have s = +1 (and N − x with s = −1), chosen in such a way
that one module at most, say µ, has neurons in both states simultaneously.
Therefore, x = nρ + z , where ρ is the number of modules with all their neu-
rons in the positive state and z is the number of neurons with positive sign in
module µ. We can write m = (2x − 1)/N and mµ = (2z − 1)/n. The total con-
figurational energy of the system will be E = Pij eij = 1
2 ω (L↑↓ − hkiN ), where
L↑↓ is the number of edges linking nodes with opposite activities. By simply
counting over edges, we can obtain the expected value of L↑↓ (which amounts
to a mean-field approximation because we are substituting the number of edges
between two neurons for its expected value), yielding:
+
z(n − z)
E
= (1 − λ)
n − 1
ω hki
λn
1
N − n {ρ[n − z + n(M − ρ − 1)] + (M − ρ − 1)(z + nρ)} −
N .
2
Fig. 7.3 shows the mean-field configurational energy curves for various values
of the modularity on a small modular network. The local minima (metastable
states) are the configurations used to store patterns. It should be noted that
the mapping x → m is highly degenerate: there are C M
mM patterns with mean
activity m that all have the same energy.
(7.1)
7.5 Forgetting avalanches
In obtaining the energy we have assumed that the number of synapses rewired
from a given module is always ν = hkinλ. However, since each edge is evaluated
with probability λ, ν will in fact vary somewhat from one module to another,
being approximately Poisson distributed with mean hν i = hkinλ. The depth
of the energy well corresponding to a given module is then, neglecting all but
the first term in Eq. (7.1) and approximating n − 1 ≃ n,
1
4
ω (nhki − ν ).
∆E ≃
7.5 Forgetting avalanches
89
M=5
M=5
M=5
M=5
λ=0.2
λ=0.2
λ=0.2
λ=0.2
λ=0.4
λ=0.4
λ=0.4
λ=0.6
λ=0.6
λ=0.8
0
0
0
0
-0.2
-0.2
-0.2
-0.2
-0.4
-0.4
-0.4
-0.4
-0.6
-0.6
-0.6
-0.6
-0.8
-0.8
-0.8
-0.8
)
)
)
)
N
N
N
N
k
k
k
k
ω
ω
ω
ω
(
(
(
(
/
/
/
/
E
E
E
E
2
2
2
2
-1
-1
-1
-1
-1
-1
-1
-1
-0.5
-0.5
-0.5
-0.5
0
0
0
0
m
m
m
m
0.5
0.5
0.5
0.5
1
1
1
1
Figure 7.3: Configurational energy of a network composed of M = 20 modules
of n = 10 neurons each, according to Eq.
(7.1), for various values of the
rewiring probability λ. The minima correspond to situations such that all
neurons within any given module have the same sign.
where
(7.2)
(7.3)
τ −β (τ ) ,
β (τ ) = 1 +
The typical escape time τ from an energy well of depth ∆E at temperature
T is τ ∼ e∆E /T (Levine and R.D., 2005). Using Stirling’s approximation in the
Poissonian distribution of ν and expressing it in terms of τ , we find that the
escape times are distributed according to
ln τ (cid:19)− 3
P (τ ) ∼ (cid:18)1 −
4T
2
ωnhki
ωnhki "1 + ln λnhki
ωnhki ln τ !# .
4T
1 − 4T
Therefore, at low temperatures, P (τ ) will behave approximately like a power-
law. The left panel of Fig. 7.4 shows the distribution of time intervals between
events in which the overlap mµ of at least one module µ changes sign. The
power-law-like behaviour is apparent, and justifies talking about forgetting
avalanches – since there are cascades of many forgetting events interspersed
with long periods of metastability. This is very similar to the behaviour ob-
served in other nonequilibrium settings in which power-law statistics arise from
the convolution of exponentials (Hurtado et al., 2008; Munoz et al., 2010).
It is known from experimental psychology that forgetting in humans is in-
deed well described by power-laws (Wixted and Ebbesen, 1991, 1997; Sikstrom,
2002). The right panel of Fig. 7.4 shows the value of the exponent β (τ ) as
a function of τ . Although for low temperatures it is almost constant over
90
Chapter 7. Cluster Reverberation: A mechanism for robust
short-term memory without synaptic learning
)
)
τ
τ
(
(
p
p
100
100
10-1
10-1
10-2
10-2
10-3
10-3
10-4
10-4
101
101
)
)
)
τ
τ
τ
(
(
(
β
β
β
2
2
2
1.8
1.8
1.8
1.6
1.6
1.6
1.4
1.4
1.4
1.2
1.2
1.2
1
1
1
T=1
T=2
T=3
101
101
101
103
103
103
τ
τ
τ
105
105
105
107
107
107
102
102
103
103
τ
τ
104
104
105
105
Figure 7.4: Left panel: distribution of escape times τ , as defined in the main
text, for λ = 0.25 and T = 2. Slope is for β = 1.35. Other parameters as in
Fig. 7.2. Symbols from MC simulations and line given by Eqs. (7.2) and (7.3).
Right panel: exponent β of the quasi-power-law distribution p(τ ) as given by
Eq. (7.3) for temperatures T = 1 (red line), T = 2 (green line) and T = 3
(blue line).
many decades of τ – approximating a pure power-law – for any finite T there
will always be a τ such that the denominator in the logarithm of Eq. (7.3)
approaches zero and β diverges, signifying a truncation of the distribution.
7.6 Clustered networks
Although we have illustrated how the mechanism of Cluster Reverberation
works on a modular network, it is not actually necessary for the topology to
have this characteristic – only for the patterns to be in some way “coarse-
grained,” as described, and that each region of the network encoding one bit
have a small enough parameter λ, defined as the proportion of synapses to
other regions. For instance, for the famous Watts-Strogatz smal l-world model
(Watts and Strogatz, 1998) – a ring of N nodes, each initially connected to its
k nearest neighbours before a proportion p of the edges are randomly rewired
– we have λ ≃ p (which is not surprising considering the resemblance between
this model and the modular network used above). More precisely, the expected
modularity of a randomly imposed box of n neurons is
2 (cid:19) ,
n (cid:18) k
1
1 − p
4 −
λ = p −
n − 1
N − 1
p +
7.6 Clustered networks
91
(7.4)
the second term on the right accounting for the edges rewired to the same box,
and the third to the edges not rewired but sufficiently close to the border to
connect with a different box.
Perhaps a more realistic model of clustered network would be a random
network embedded in d-dimensional Euclidean space. For this we shall use the
scheme laid out by Rozenfeld et al. (Rozenfeld et al., 2002), which consists
simply in allocating each node to a site on a d-torus and then, given a par-
ticular degree sequence, placing edges to the nearest nodes possible – thereby
attempting to minimise total edge length5 . For a scale-free degree sequence
(i.e., a set {ki} drawn from a degree distribution p(k) ∼ k−γ ) according to
some exponent γ , then, as shown in B, such a network has a modularity
1
d(γ − 2) − 1 (cid:2)d(γ − 2)l−1 − l−d(γ−2) (cid:3) ,
λ ≃
where l is the linear size of the boxes considered.
Fig. 7.5 compares this expression with the value obtained numerically after
averaging over many network realizations, and shows that it is fairly good –
considering the approximations used for its derivation. It is interesting that
even in this scenario, where the boxes of neurons which are to receive the
same stimulus are chosen at random with no consideration for the underlying
topology, these boxes need not have very many neurons for λ to be quite low
(as long as the degree distribution is not too heterogeneous).
Carrying out the same repeated stimulation test as on the modular net-
works in Fig. 7.2, we find a similar behaviour for the scale-free embedded
networks. This is shown in Fig. 7.6, where for high enough intensity of stim-
uli δ and scale-free exponent γ , performance can, as in the modular case, be
η ≃ 1. We should point out that for good performance on these networks we
require more neurons for each bit of information than on modular networks
with the same λ (in Fig. 7.6 we use n = 100, as opposed to n = 10 in Fig.
7.2). However, that we should be able to obtain good results for such diverse
network topologies underlines that the mechanism of Cluster Reverberation
is robust and not dependent on some very specific architecture. In fact, we
have recently shown that similar metastable memory states can also occur on
networks which have random modularity and clustering, but a certain degree
of assortativity 6 (de Franciscis et al., 2011).
5The authors also consider a cutoff distance, but we shall take this to be infinite here.
6The assortativity of a network is here understood to mean the extent to which the
degrees of neighbouring nodes are correlated (Johnson et al., 2010b).
92
Chapter 7. Cluster Reverberation: A mechanism for robust
short-term memory without synaptic learning
λ
λ
λ
λ
λ
λ
λ
λ
λ
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0
0
0
0
0
0
0
0
0
l=2
l=4
l=8
2
2
2
2
2
2
2
2
2
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
3
3
3
3
3
3
3
3
3
3.5
3.5
3.5
3.5
3.5
3.5
3.5
3.5
3.5
γ
γ
γ
γ
γ
γ
γ
γ
γ
4
4
4
4
4
4
4
4
4
4.5
4.5
4.5
4.5
4.5
4.5
4.5
4.5
4.5
5
5
5
5
5
5
5
5
5
Figure 7.5: Proportion of outgoing edges, λ, from boxes of linear size l against
exponent γ for scale-free networks embedded on 2D lattices. Lines from Eq.
(7.4) and symbols from simulations with hki = 4 and N = 1600.
δ=3.5
δ=3.5
δ=3.5
δ=3.5
δ=4
δ=4
δ=4
δ=5
δ=5
δ=10
1
1
1
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
η
η
η
η
η
η
η
η
η
η
η
η
e
e
e
e
e
e
e
e
e
e
e
e
c
c
c
c
c
c
c
c
c
c
c
c
n
n
n
n
n
n
n
n
n
n
n
n
a
a
a
a
a
a
a
a
a
a
a
a
m
m
m
m
m
m
m
m
m
m
m
m
r
r
r
r
r
r
r
r
r
r
r
r
o
o
o
o
o
o
o
o
o
o
o
o
f
f
f
f
f
f
f
f
f
f
f
f
r
r
r
r
r
r
r
r
r
r
r
r
e
e
e
e
e
e
e
e
e
e
e
e
P
P
P
P
P
P
P
P
P
P
P
P
γ=2
γ=2
γ=2
γ=3
γ=3
γ=4
1
1
1
0.5
0.5
0.5
m
m
m
i
i
i
t
t
t
s
s
s
m
m
m
0
0
0
500
500
500
Time (MCS)
Time (MCS)
Time (MCS)
3000
3000
3000
0
0
0
0
0
0
0
0
0
0
0
0
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
4
4
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
5
5
5
5
5
5
5
5
5
5
5
5
6
6
6
6
6
6
6
6
6
6
6
6
Figure 7.6: Performance η against exponent γ for scale-free networks, embed-
ded on a 2D lattice, with patterns of M = 16 modules of n = 100 neurons
each, hki = 4 and N = 1600; patterns are shown with intensities δ = 3.5, 4, 5
and 10, and T = 0.01 (lines – splines – are drawn as a guide to the eye). Inset:
typical time series for γ = 2, 3, and 4, with δ = 5.
7.7 Yes, but does it happen in the brain?
93
7.7 Yes, but does it happen in the brain?
As we have shown, Cluster Reverberation is a mechanism available to neural
systems for robust short-term memory without synaptic learning. To the best
of our knowledge, this is the first mechanism proposed which has these charac-
teristics – essential for, say, sensory memory or certain working-memory tasks.
All that is needed is for the network topology to be highly clustered or modu-
lar, and for small groups of neurons to store one bit of information, as opposed
to the conventional view which assumes one bit per neuron. Considering the
enormous number of neurons in the brain, and the fact that real individual
neurons are probably too noisy to store information reliably, these hypotheses
do not seem farfetched. The mechanism is furthermore consistent both with
what is known about the topology of the brain, and with experiments which
have revealed power-law forgetting.
Since the purpose of this paper is only to describe the mechanism of Clus-
ter Reverberation, we have made use of the simplest possible model neurons –
namely, binary neurons with static, uniform synapses – for the sake of clarity
and generality. However, there is no reason to believe that the mechanism
would cease to function if more neuronal ingredients were to be incorporated.
In fact, cellular bistability, for instance, would increase performance, and the
two mechanisms could actually work in conjunction. Similarly, we have also
used binary patterns to store information. But it is to be expected that pat-
terns depending on any form of frequency coding, for instance, could also
be maintained with more sophisticated neurons – such that different modules
could be set to different mean frequencies.
Whether Cluster Reverberation would work for biological neural systems
could be put to the test by growing such modular networks in vitro, stimulating
appropriately, and observing the duration of the metastable states. In vivo
recordings of neural activity during short-term memory tasks, together with
a mapping of the underlying synaptic connections, could be used to ascertain
whether the brain does indeed make use of this mechanism – although for this
it must be borne in mind that the neurons forming a module need not find
themselves close together in metric space. We hope that experiments such as
these will be carried out and eventually reveal something more about the basis
of this puzzling emergent property of the brain’s known as thought.
Chapter 8
Concluding remarks
“As long as the brain is a mystery, the universe will remain a mystery,” claimed
Santiago Ram´on y Ca jal. Our very essence seems to reside somehow in the
workings of this organ, probably as a consequence of electro-chemical signalling
that goes on among its hundred billion or so constituent neurons. Will this
mystery ever be cleared up? We know of other ob jects that process informa-
tion in highly sophisticated ways – electronic computers. Faced with a sudden
blue screen, one may be forgiven for calling these devices incomprehensible and
capricious, even malevolent. But in fact most educated people understand, on
some level at least, what mechanisms and physical processes are behind the
complex behaviour displayed by computers, and do not consider the issue a
mystery. This is not to suggest that the analogy between brain and computer
should be taken any further than to illustrate how a great many elements, each
executing some fairly simple and obvious operation, can “cooperate” to yield
astonishingly complicated yet functional behaviour; and that one can grasp
how this occurs without having to know every detail. But we have not yet
reached this point as regards the brain. Much progress has been made con-
cerning aspects of physiology, while once unassailable mental disorders such
as phobias can now be easily cured by psychology. Yet as far as what mecha-
nisms relate these two levels of description goes, perhaps all we can safely say
for now is that synaptic plasticity is responsible for long-term memory. The
origins of even some well-defined and much studied cognitive abilities – such
as probabilistic reasoning or short-term memory – remain somewhat elusive,
while the nature of consciousness, say, is still truly a mystery. However, if
instead of developing computers ourselves we had been given them by an alien
species, we could still hope one day to unravel the mysteries of their magic. In
much the same manner, by searching for ways in which collections of neurons
94
95
might perform tasks such as we know them to be capable of, we will some day
understand not only how our stomachs digest and our hearts pump, but also
how our brains think.
I cannot pretend that the work described here takes us more than, at best,
a tiny step of the way along this path. The brain is, among other things,
a network, and networks are a kind of mathematical ob ject about which we
now know much more than just a few years ago. In fact, they are a central
element of what can arguably be called the most challenging frontier currently
facing human understanding about the world – the nature of complex systems.
So, from among the innumerable aspects likely to shape and determine the
way neurons cooperate, the research presented here focuses on the structure
of the underlying network. First of all it looks at how this structure can
develop. Chapter 3 addressed this by formalizing as a stochastic process a
situation governed by probabilistic events like synaptic growth and death. Such
simple individual behaviour was shown to be enough to explain many statistical
features of real neural systems. Furthermore, this Fokker-Planck description
relating microscopic, stochastic actions to a macroscopic evolution of properties
such as mean synaptic density, degree heterogeneity or assortativity may help
to gain insights into the biochemical processes taking place.
The rest of the thesis is mostly devoted to how aspects of a neural network’s
topology might influence or even determine its ability to carry out certain tasks
akin to those the brain undertakes. The fact that dynamical memory perfor-
mance ensuing from synaptic depression is favoured by a highly heterogeneous
degree distribution, laid out in Chapter 4, may help to explain why the brain
seems to display such a topology at several levels of description – perhaps
somehow maintaining itself close to a critical point. Similarly, the enhanced
robustness to noise found for positively correlated networks in Chapter 6 sug-
gests a functional advantage to a neural network being thus wired; a prediction
also in agreement with some experimental findings.
As far as unearthing the mechanisms underpinning how neurons can per-
form cognitive tasks goes, though, perhaps the most interesting idea proposed
is that of Cluster Reverberation, in Chapter 7, whereby thanks to modularity
and/or clustering a neural network is able to store information instantly, with-
out requiring biochemical changes in the synapses. Time will tell whether real
neural systems do indeed harness this mechanism to perform certain short-term
memory tasks.
A collateral but noteworthy aspect of this research is the potentiality for
96
Chapter 8. Concluding remarks
application elsewhere of some of the mathematical techniques developed. Most
of all, the method for studying correlated networks and dynamics thereon put
forward in Chapter 5 for use in Chapter 6 can be expected to find widespread
use. The answer to the question of why most networks are disassortative given
in Chapter 5, or the relation between degree-degree correlations and nestedness
described in Appendix C are examples of this.
Finally I must mention not just the answers I hope to have provided, or at
least hinted at, to some unsolved problems, but also the questions that have
been posed and challenges laid bare: Would a more detailed description of
brain development still be possible with Fokker-Planck equations? Are these
topological effects, found to be at work for the simplest neural models, indeed
so relevant for real neurons? Can Cluster Reverberation be performed in vitro?
The greatest function this thesis could perform would be to stimulate others
to look into these or related issues in more depth than here. But I also hope
it may serve to illustrate the sentiment, What matters how long the path to
the final unravelling of the mysteries is, as long as the going is fun?
Chapter 9
Conclusiones en espanol
“Mientras el cerebro sea un misterio, el universo continuar´a siendo un miste-
rio”, dijo una vez Santiago Ram´on y Ca jal. Parece que nuestra misma esencia
reside de alguna manera en el funcionamiento de este ´organo, probablemente
como consecuencia de las senales electro-qu´ımicas entre sus aproximadamente
cien mil millones de neuronas. ¿Se resolver´a alg´un d´ıa este misterio? Conoce-
mos otros ob jetos capaces de procesar informaci´on de manera altamente sofisti-
cada:
los ordenadores electr´onicos. Confrontados con un pantallazo azul, se
nos podr´ıa perdonar el tildar este tipo de aparatos de incomprensibles y capri-
chosos, por no decir mal´evolos. Pero en realidad la mayor parte de la gente
entiende, al menos en alg´un nivel, cu´ales son los mecanismos y procesos f´ısicos
que subyacen el comportamiento complejo del que hacen gala los ordenadores,
y no consideran que el tema sea un misterio. No es que la analog´ıa entre
cerebro y ordenador deba ser llevado m´as lejos que para ilustrar c´omo muchos
elementos, cada uno ejecutando alguna operaci´on relativamente simple y obvia,
pueden “cooperar” y mostrar un comportamiento colectivo asombrosamente
complicado, pero funcional; y que se puede comprender c´omo ocurre esto sin
necesidad de conocer hasta el ´ultimo detalle. A´un no hemos llegado a poder
responder a esta pregunta en lo que respecta al cerebro. Hemos ampliado
enormemente nuestro conocimiento de aspectos fisiol´ogicos, y trastornos men-
tales antano incurables, como las fobias, son f´acilmente tratadas hoy en d´ıa
por la psicolog´ıa. En cuanto a los mecanismos que relacionan estos dos niveles
de descripci´on, posiblemente lo ´unico que podamos decir a ciencia cierta es que
la plasticidad sin´aptica est´a detr´as de la memoria a largo plazo. Los or´ıgenes
incluso de algunas habilidades cognitivas bien definidas y extensamente estu-
diadas, como el razonamiento probabil´ıstico o la memoria a corto plazo, est´an
a´un por descifrar completamente; mientras que, por ejemplo, la naturaleza de
98
99
la consciencia es verdaderamente a´un un misterio. Sin embargo, si en lugar de
haber desarrollado los ordenadores nosotros mismos los hubi´esemos recibido de
una especie alien´ıgena, a´un as´ı podr´ıamos esperar alg´un d´ıa desenmaranar los
misterios de su magia. Del mismo modo, buscando maneras de que conjuntos
de neuronas puedan realizar el tipo de tareas de las que las sabemos capaces,
alg´un d´ıa entenderemos no s´olo c´omo nuestros est´omagos digieren y nuestros
corazones laten, sino tambi´en c´omo nuestros cerebros piensan.
Este traba jo, en el mejor de los casos, nos avanza un paso infinitesimal por
este camino. El cerebro es, entre otras muchas cosas, una red, y las redes son
ob jetos matem´aticos sobre los que sabemos hoy mucho m´as que hace tan s´olo
unos anos. De hecho, son un elemento fundamental para uno de los mayores re-
tos con los que se enfrenta actualmente el conocimiento humano: la naturaleza
de los sistemas complejos. As´ı que, de entre los innumerables aspectos suscepti-
bles de modificar y determinar c´omo las neuronas cooperan, esta investigaci´on
se centra en la estructura de la red subyacente. Primero analiza c´omo dicha
estructura puede desarrollarse. El Cap´ıtulo 3 enfoca esto formalizando medi-
ante la teor´ıa de los procesos estoc´asticos una situaci´on gobernada por eventos
probabil´ısticos tales como el crecimiento y la muerte sin´apticas. Se demuestra
que este tipo de comportamiento individual es suficiente para explicar muchas
propiedades estad´ısticas de las redes de cerebros reales. Por otra parte, este
marco te´orico puede ser reducido a una descripci´on en t´erminos de ecuaciones
de Fokker-Planck, que relacionan acciones microsc´opicas estoc´asticas con la
evoluci´on macrosc´pica de propiedades como la densidad sin´aptica media, la
heterogeneidad de la distribuci´on de grados o la asortatividad, que quiz´as nos
permita extraer informaci´on relevante acerca de los procesos bioqu´ımicos in-
volucrados.
La mayor parte del resto de la tesis trata de c´omo aspectos de la topolog´ıa
de una red neuronal pueden influenciar o incluso determinar su habilidad para
ejecutar ciertas tareas cognitivas como las que se describen en un cerebro o
medio neuronal real. Por ejemplo, el hecho de que, en cuanto a la memoria
din´amica que emerge gracias a la depresi´on sin´aptica, el rendimiento es mayor
para una distribuci´on de grados altamente heterog´enea, como demuestra el
Cap´ıtulo 4, podr´ıa ayudar a explicar por qu´e el cerebro parece mostrar una
topolog´ıa de este tipo en varios niveles de descripci´on, quiz´as incluso man-
teni´endo su actividad, de alguna manera todav´ıa no comprendida del todo,
cerca de un punto cr´ıtico. De igual modo, la mayor robustez durante los pro-
cesos cognitivos en presencia de ruido en el caso de redes con correlaciones
100
Chapter 9. Conclusiones en espanol
positivas como se ha descrito en el Cap´ıtulo 6 sugiere que existe una venta ja
funcional para una red neuronal en adoptar esta propiedad; una predicci´on
que tambi´en enca ja con algunos hallazgos experimentales.
En lo que se refiere a desentranar los mecanismos que permiten a las
neuronas realizar colectivamente tareas cognitivas, quiz´as la idea m´as intere-
sante aqu´ı propuesta es la de Cluster Reverberation (Reverberaci´on de Grupo),
en el Cap´ıtulo 7, seg´un la cual, gracias a la modularidad y/o el grado de
“agrupamiento”, una red neuronal es capaz de almacenar informaci´on in-
stant´aneamente, sin requerir para ello cambios bioqu´ımicos de potenciaci´on o
depresi´on a largo plazo en las sinapsis. El tiempo dir´a si el cerebro aprovecha
realmente este mecanismo para realizar ciertas tareas de memoria de corto
plazo.
Un aspecto colateral pero digno de menci´on de este traba jo es el de la
potencialidad de algunas de las t´ecnicas matem´aticas desarrolladas de ser apli-
cadas para otras situaciones de inter´es. Sobre todo, es de esperar que el m´etodo
para estudiar redes correlacionadas, y din´amicas sobre ellas, propuesto en el
Cap´ıtulo 5 y utilizado en el Cap´ıtulo 6, sea ´util para una amplia gama de prob-
lemas. La respuesta, en el Cap´ıtulo 5, a la pregunta de por qu´e la mayor´ıa de
las redes son disasortativas, o la relaci´on entre correlaciones entre los nodos y
el “anidamiento” descrita en el Ap´endice C son ejemplos de aplicaciones.
Finalmente, hay que mencionar no s´olo las respuestas que se han intentado
dar, o al menos sugerir, con esta tesis para algunos problemas sin resolver,
sino tambi´en las preguntas y los nuevos retos que han surgido: por ejemplo,
¿ser´ıa posible, tambi´en con ecuaciones de Fokker-Planck, una descripci´on m´as
detallada del desarrollo cerebral? ¿Son estos efectos topol´ogicos, descritos para
los modelos neuronales m´as sencillos, realmente tan relevantes para neuronas
de verdad? ¿Puede el mecanismo de Cluster Reverberation ocurrir in vitro? En
definitiva, la mayor funci´on que pudiera cumplir esta tesis ser´ıa la de estimular
a otra/os para que indaguen en estos y otros temas m´as profundamente que
aqu´ı. Pero quiz´as tambi´en sirva para ilustrar el siguiente sentimiento: ¿qu´e
m´as da cu´an largo sea el camino hacia el desenmaranamiento ´ultimo de los
misterios, siempre que el trayecto sea divertido?
Appendix A
Nonlinear preferential rewiring
in fixed-size networks as a
diffusion process
We present an evolving network model in which the total numbers of nodes
and edges are conserved, but in which edges are continuously rewired according
to nonlinear preferential detachment and reattachment. Assuming power-law
kernels with exponents α and β , the stationary states the degree distribu-
tions evolve towards exhibit a second order phase transition – from relatively
homogeneous to highly heterogeneous (with the emergence of starlike struc-
tures) at α = β . Temporal evolution of the distribution in this critical regime
is shown to follow a nonlinear diffusion equation, arriving at either pure or
mixed power-laws, of exponents −α and 1 − α.
Complex systems may often be described as a set of nodes with edges con-
necting some of them – the neighbours – (see, for instance, Refs.(Boccaletti et al.,
2006; Arenas et al., 2008a; Marro et al., 2008)). The number of edges a partic-
ular node has is called its degree, k . The study of such large networks is usually
made simpler by considering statistical properties, e.g., the degree distribution,
p(k) (probability of finding a node with a particular degree). It turns out that
a high proportion of real-world networks follow power-law degree distributions,
p(k) ∼ k−γ – referred to as scale-free due to their lack of a characteristic size.
Also, many of them have their edges placed among the nodes apparently in a
random way – i.e., there is no correlation between the degree of a node and
any other of its properties, such as the degrees of its neighbours. Barab´asi
and Albert (Barab´asi and Albert, 1999) applied the mechanism of preferential
attachment to an evolving network model and showed how this resulted in the
102
103
degree distributions becoming scale-free for long enough times. For this to
work, attachment had to be linear – i.e., the probability a node with degree k
has of receiving a new edge is π(k) ∼ k + q . This results in scale-free stationary
degree distributions with an exponent γ = 3 − q .
Preferential attachment seems to be behind the emergence of many real-
world, continuously growing networks. However, not all networks in which
some nodes at times gain (or loose) new edges have a continuously growing
number of nodes. For example, a given group of people may form an evolv-
ing social network (Kossinets and Watts, 2006) in which the edges represent
friendship. Preferential attachment may be relevant here – the more people
you know, the more likely it is that you will be introduced to someone new –
but probabilities are not expected to depend linearly on degree. For instance,
there may be saturations (highly connected people might become less acces-
sible), threshold effects (hermits may be prone to antisocial tendencies), and
other non-linearities. The brain may also be a relevant case. Once formed, the
number of neurons does not seem to continually augment, and yet its struc-
tural topology is dynamic (Klintsova and Greenough, 1999). Synaptic growth
and dendritic arborization have been shown to increase with electric stimula-
tion (Lee et al., 1980; Roo et al., 2008) – and, in general, the more connected
a neuron is, the more current it receives from the sum of its neighbours.
Barab´asi and Albert showed that both (linear) preferential attachment and
an ever-growing number of nodes are needed for scaling to emerge in their
model. In a fixed population, their mechanism would result in a fully-connected
network. However, this is not normally observed in real systems. Rather, just
as some new edges sprout, others disappear – less used synapses suffer at-
rophy, unstimulating friendships wither. Often, the numbers of both nodes
and edges remain roughly constant. The same authors did therefore extend
their model so as to include the effects of preferential rewiring (which could
be applied to fixed-size networks), although again probabilities depended lin-
early on node degree (Albert and Barab´asi, 2000). Another mechanism which
(roughly) maintains constant the numbers of nodes and edges is node fus-
ing (Thurner et al., 2007), once more according to linear probabilities. As to
nonlinear preferential attachment, the (growing) BA model was extended to
take power-law probabilities into account (Krapivsky et al., 2000), although
the solutions are only scale free for the linear case.
In this note we present an evolving network model with preferential rewiring
according to nonlinear (power-law) probabilities. The number of nodes and
Chapter A. Nonlinear preferential rewiring in fixed-size networks
104
as a diffusion process
edges is conserved but the topology evolves, arriving eventually at a macro-
scopically (nonequilibrium) stationary state – as described by global properties
such as the degree distribution. Depending on the exponents chosen for the
rewiring probabilities, the final state can be either fairly homogeneous, with
a typical size, or highly heterogeneous, with the emergence of starlike struc-
tures. In the critical case marking the transition between these two regimes,
the degree distribution is shown to follow a nonlinear diffusion equation. This
describes a tendency towards stationary states that are characterized either by
scale-free or by mixed scale-free distributions, depending on parameters.
Our model consists of a random network with N nodes of respective de-
gree ki , i = 1, 2, ..., N , and 1
2 N hki edges. Initially, the degrees have a given
distribution p(k , t = 0). At each time step, one node is chosen with a proba-
bility which is a function of its degree, ρ(ki ). One of its edges is then chosen
randomly and removed from it, to be reconnected to another node j chosen
according to a probability π(kj ). That is, an edge is broken and another one is
created, and the total number of edges, as well as the total number of nodes,
is conserved. The functions π(k) and ρ(k) are arbitrary, but we shall explic-
i and ρ(ki ) ∼ kβ
itly illustrate here π(ki ) ∼ kα
i that capture the essence of a
wide class of nonlinear monotonous response functions and are easy to handle
analytically.
The probabilities π and ρ a given node has, at each time step, of increasing
or decreasing its degree can be interpreted as transition probabilities between
states. The expected value of the increment in a given p(k , t) at each time
step, ∆p(k , t), may then be written as
∂p(k , t)
∂ t
(A.1)
= (k − 1)α ¯k−1
α p(k − 1, t)
+ (k + 1)β ¯k−1
β p(k + 1, t)
− (cid:0)kα ¯k−1
α + kβ ¯k−1
β (cid:1) p(k , t),
where ¯ka = ¯ka (t) = Pk kap(k , t). If it exists, any stationary solution must
satisfy the condition pst(k + 1) (k + 1)β ¯k st
α = pst (k) kα ¯k st
β which, for k ≫ 1,
implies that
= ¯k st
(k + 1)β − 1! pst(k).
kα
∂pst (k)
α
¯k st
∂k
β
1
Therefore, the distribution will have an extremum at ke = (cid:0)¯k st
β /¯k st
α (cid:1)
α−β (where
If α < β , this will be a maximum,
we have approximated ke ≃ ke + 1).
(A.2)
105
signalling the peak of the distribution. On the other hand, if α > β , ke will
correspond to a minimum. Therefore, most of the distribution will be broken
in two parts, one for k < ke and another for k > ke . The critical case for
α = β will correspond to a monotonously decreasing stationary distribution,
but such that limk→∞∂pst (k)/∂k = 0. In fact, Eq. (A.1) is for this situation
(α = β ) the discretized version of a nonlinear diffusion equation,
∂ 2
∂k2 [kαp(k , τ )],
after dynamically modifying the time scale according to τ = t/¯kα (t). Ignoring,
for the moment, border effects, the solutions of this equation are of the form
∂p(k , τ )
∂ τ
=
(A.3)
(A.4)
pst (k) ∼ Ak−α + Bk−α+1 ,
with A and B constants. If α > 2, then given A we can always find a B which
allows pst (k) to be normalized in the thermodynamic limit 1 . For example,
if the lower limit is k ≥ 1, then B = (α − 2) [1 − A/(α − 1)]. However, if
1 < α ≤ 2, then only A can remain non-zero, and pst (k) will be a pure power
law. For α ≤ 1, both constants must tend to zero as N → ∞.
In finite
networks, no node can have a degree larger than N − 1 or lower than 0. In
fact, one would usually wish to impose a minimum nonzero degree, e.g. k ≥ 1.
The temporal evolution of the degree distribution is illustrated in Fig. A.1.
This shows the result of integrating Eq. (A.1) for k ≥ 1, different times, β = 1,
and three different values of α, along with the respective values obtained from
Monte Carlo simulations.
The main result may be summarized as follows. For α < β , the network
will evolve to have a characteristic size, centred around hki. At the critical
case α = β , all sizes appear, according either to a pure or a composite power
law, as detailed above.
If we impose, say, k ≥ 1, then starlike structures will emerge, with a great
many nodes connected to just a few hubs 2 .
Figure A.2 illustrates the second order phase transition undergone by the
variance of the final (stationary) degree distribution, depending on the expo-
nent α, where β is set to unity. It should be mentioned that this particular
1Although all moments of k will diverge unless B = 0.
2There is a finite-size effect not taken into account by the theory – but relevant when
α > β – which provides a natural lower cutoff for pst (k): if there are, say, m nodes which
are connected to the whole network, then the minimum degree a node can have is m.
Chapter A. Nonlinear preferential rewiring in fixed-size networks
106
as a diffusion process
100
100
100
100
100
100
100
100
α=0.5
α=0.5
α=0.5
α=0.5
α=0.5
α=0.5
α=0.5
α=0.5
)
)
)
)
)
)
)
)
k
k
k
k
k
k
k
k
(
(
(
(
(
(
(
(
t
t
t
t
t
t
t
t
s
s
s
s
s
s
s
s
p
p
p
p
p
p
p
p
)
)
)
)
)
)
)
)
)
k
k
k
k
k
k
k
k
k
(
(
(
(
(
(
(
(
(
t
t
t
t
t
t
t
t
t
s
s
s
s
s
s
s
s
s
p
p
p
p
p
p
p
p
p
)
)
)
)
)
)
)
)
k
k
k
k
k
k
k
k
(
(
(
(
(
(
(
(
t
t
t
t
t
t
t
t
s
s
s
s
s
s
s
s
p
p
p
p
p
p
p
p
100
100
100
100
100
100
100
100
104
104
104
104
104
104
104
104
100
100
100
100
100
100
100
100
100
104
104
104
104
104
104
104
104
104
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
104
104
104
104
104
104
104
104
100
100
100
100
100
100
100
100
t=102
t=102
t=102
t=102
t=103
t=103
t=103
t=104
t=104
t=105
k
k
k
k
k
k
k
k
102
102
102
102
102
102
102
102
α=1.0
α=1.0
α=1.0
α=1.0
α=1.0
α=1.0
α=1.0
α=1.0
α=1.0
1/k
α=1.5
α=1.5
α=1.5
α=1.5
α=1.5
α=1.5
α=1.5
α=1.5
k
k
k
k
k
k
k
k
k
k
k
k
k
k
k
k
k
103
103
103
103
103
103
103
103
103
103
103
103
103
103
103
103
103
Figure A.1: Degree distribution p(k , t) at four different stages of evolution:
t = 102 [(yellow) squares], 103 [(blue) circles], 104 [(red) triangles)] and 105
MCS [(black) diamonds]. From top to bottom panels, subcritical (α = 0.5),
critical (α = 1) and supercritical (α = 1.5) rewiring exponents. Symbols from
MC simulations and corresponding solid lines from numerical integration of
Eq. (A.1). β = 1, hki = 10 and N = 1000 in all cases.
case, β = 1, corresponds to edges being chosen at random for disconnection,
since the probability of a random edge belonging to node i is proportional to
ki .
This topological phase transition is similar to the ones that have been
107
100
100
100
100
10
10
10
10
10
10
10
10
10
10
10
10
)
k
(
t
s
p
2
2
2
2
2
2
2
2
2
2
2
2
>
>
>
>
>
>
>
>
>
>
>
>
k
k
k
k
k
k
k
k
k
k
k
k
<
<
<
<
<
<
<
<
<
<
<
<
/
/
/
/
/
/
/
/
/
/
/
/
2
2
2
2
2
2
2
2
2
2
2
2
σ
σ
σ
σ
σ
σ
σ
σ
σ
σ
σ
σ
10-3
10-3
10-3
10-3
100
100
100
100
N=800
N=1200
N=1600
N=2000
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
α=0.5
α=0.5
α=0.5
α=0.5
α=1.0
α=1.0
α=1.0
α=1.5
α=1.5
β=1
β=1
β=1
β=1
β=1
β=1
β=1
β=1
β=1
β=1
β=1
β=1
103
103
103
103
k
10
10
10
1
1
1
2
2
2
>
>
>
k
k
k
<
<
<
/
/
/
2
2
2
σ
σ
σ
0.1
0.1
0.1
1
1
1
1
1
1
1
1
1
1
1
1
α
α
α
α
α
α
α
α
α
α
α
α
105
105
105
0
0
0
Time (in MCS)
Time (in MCS)
Time (in MCS)
2
2
2
2
2
2
2
2
2
2
2
2
Figure A.2: Adjusted variance σ 2/hki2 of the degree distribution after 2 × 105
MCS against α, as obtained from MC simulations, for system sizes N = 800
[(yellow) squares], 1200 [(blue) circles], 1600 [(red) triangles] and 2000 [(black)
diamonds]. Top left inset shows final degree distributions for α = 0.5 [light
gray (blue)], 1 [dark gray (red)] and 1.5 (black), with N = 1000. Bottom right
inset shows typical time series of σ 2/hki2 for the same three values of α and
N = 1200. In all cases, β = 1 and hki = 10.
described in equilibrium network ensembles defined via an energy function,
in the so-called synchronic approach to network analysis (Farkas et al., 2004;
Park and Newman, 2004; Burda et al., 2004; Der´enyi et al., 2004). However,
our (nonequilibrium) model does not come within the scope of this body of
work, since the rewiring rates cannot, in general, be derived from a potential.
Furthermore, we are here concerned with the time evolution rather than the
stationary states, making our approach diachronic.
Summing up, in spite of its simplicity, our model captures the essence of
many real-world networks which evolve while leaving the total numbers of
nodes and edges roughly constant. The grade of heterogeneity of the station-
ary distribution obtained is seen to depend crucially on the relation between
the exponents modelling the probabilities a node has of obtaining or loos-
ing a new edge. It is worth mentioning that the heterogeneity of the degree
distribution of a random network has been found to determine many rele-
vant behaviours and magnitudes such as its clustering coefficient and mean
minimum path (Newman, 2003c), critical values related to the dynamics of
Chapter A. Nonlinear preferential rewiring in fixed-size networks
108
as a diffusion process
excitable networks (Johnson et al., 2008), or the synchronizability for systems
of coupled oscillators (since this depends on the spectral gap of the Laplacian
matrix) (Barahona and Pecora, 2002).
The above shows how scale-free distributions, with a range of exponents,
may emerge for nonlinear rewiring, although only in the critical situation in
which the probabilities of gaining or loosing edges are the same. We believe
that this non-trivial relation between the microscopic rewiring actions (gov-
erned in our case by parameters α and β ) and the emergent macroscopic degree
distributions could shed light on a class of biological, social and communica-
tions networks.
Appendix B
Effective modularity of highly
clustered networks
The number of nodes within a radius r is n(r) = Adrd , with Ad a constant.
We shall therefore assume a node with degree k to have edges to all nodes
up to a distance r(k) = (k/Ad)1/d , and none beyond (note that this is not
necessarily always feasible in practice). To estimate λ, we shall first calcu-
late the probability that a randomly chosen edge have length x. The chance
that the edge belong to a node with degree k is π(k) ∼ kp(k) (where p(k) is
the degree distribution). The proportion of edges that have length x among
those belonging to a node with degree k is ν (xk) = dAdxd−1/k if Adxd < k ,
and 0 otherwise. Considering, for example, scale-free networks (as in Ref.
(Rozenfeld et al., 2002)), so that the degree distribution is p(k) ∼ k−γ in some
interval k ∈ [k0 , kmax ] (Barab´asi and Albert, 1999), and integrating over p(k),
we have the distribution of lengths,
P (x) = (C onst.) Z kmax
max(k0 ,Axd )
where we have assumed, for simplicity, that the network is sufficiently sparse
that max(k0 , Axd ) = Axd , ∀x ≥ 1, and where we have normalised for the
interval 1 ≤ x < ∞; strictly, x ≤ (kmax/A)1/d , but we shall also ignore this
effect. Next we need the probability that an edge of length x fall between two
compartments of linear size l. This depends on the geometry of the situation
as well as dimensionality; however, a first approximation which is independent
of such considerations is
π(k)ν (k x)dk = d(γ − 2)x−[d(γ−2)+1] ,
Pout (x) = min (cid:16)1,
110
x
l (cid:17) .
111
We can now estimate the modularity λ as
λ = Z ∞
1
d(γ − 2) − 1 (cid:2)d(γ − 2)l−1 − l−d(γ−2) (cid:3) .
1
Fig. 7.5 shows how λ depends on γ for d = 2 and various box sizes.
Pout (x)P (x)dx =
Appendix C
Nestedness of networks
The property of nestedness has for some time aroused a fair amount of interest
as regards ecological networks – especially since a high nestedness in mutu-
alistic systems has been shown to enhance biodiversity. However, because it
is usually estimated with software, no analytical work has been done relating
nestedness with other network characteristics, and consequently comparisons
of experimental data with null-models can only be done computationally. We
suggest a slightly refined version of the measure recently defined by Bastolla et
al. and go on to study the effect of the degree distribution and degree correla-
tions (assortativity). Our work provides a benchmark against which empirical
networks can be contrasted.
C.1 Introduction
The intense study that complex networks have undergone over the past decade
or so has shown how important topological features can be for properties of
complex systems, such as dynamical behaviour, spreading of information, re-
silience to attacks, etc. (Newman, 2003c; Boccaletti et al., 2006). A paradig-
matic case is that of ecosystems. The solution to May’s paradox (May, 1973)
– the fact that large ecosystems seem to be especially stable, when theory
predicts the contrary – is still not clear, but it is widely suspected that there
is some structural feature of ecological networks which as yet eludes us. One
aspect of such networks, which has been studied for some time by ecologists
and may be related to this problem, is called nestedness. Loosly speaking, a
network – say of species and islands, linked whenever the former inhabit the
latter – is said to be highly nested if the species which exist on scarcely popu-
lated islands tend always to be found also on those islands inhabited by many
112
C.1 Introduction
113
Figure C.1: Maximally packed matrix representing a network of plants and
islands off Perth (Abbott and Black, 1980) (because the network is bipartite,
the adjacency matrix is composed of four blocks: two identical to this ma-
trix, the other two composed of zeros). Data, image and line obtained from
NESTEDNESS CALCULATOR, which returns a “temperature” of T = 0.69o
for this particular network.
different species. This can be most easily seen by graphically representing a
matrix such that animals are columns and islands are files, with elements equal
to one whenever two nodes are linked and zero if not. If, after ordering each
kind of node by degree (number of neighbours), all the ones can be quite neatly
packed into one corner, the network is considered highly nested. This is done
in Fig. C.1 for a network of plants inhabiting islands off Perth. This rather
vague concept is usually measured with software for the purpose. For Fig.
C.1, we have used NESTEDNESS CALCULATOR, which estimates a curve
of equal density of ones and zeros, calculates how many ones and zeros are on
the “wrong” side and by how much, and returns a number between 0 and 100
called “temperature” by analogy with some system such as a subliming solid.
A low temperature indicates high nestedness. To determine how significantly
nested a given network is, the usual procedure is to generate equivalent ran-
dom networks computationally (with sone constraint such as the number of
edges or the degree of each node being conserved) and estimate how likely it
is that such a network be “colder” than that of the data.
Bastolla et al. (Bastolla et al., 2009) have recently shown how symbiotic
interactions can reduce the effective competition between two species, say of
insect, via common symbiotic hosts – such as plants they pollinate. These
authors define a measure to take into account the average number of shared
partners in these mutualistic networks, and call it “nestedness” because it
would seem to be related to the concept referred to above. They go on to show
114
Chapter C. Nestedness of networks
evidence of how the nestedness of empirical mutualistic networks is correlated
with the biodiversity of the corresponding ecosystems. This beneficial effect
“enemy” nodes can gain from sharing “friendly” partners is not confined to
ecosystems. It is expected also to play a role, for instance, in financial networks
or other economic systems (Sugihara and Ye, 2009). The principle is simple.
Say nodes A and B are in competition with each other. An increase in A will
be to B’s detriment, and viceversa; but if both A and B engage in a symbiotic
relationship with node C, then A’s thriving will stimulate C, which in turn will
be helpful to B. Thus, the effective competition between A and B is reduced,
and the whole system becomes more stable and capable of sustaining more
nodes (Dom´ınguez-Chibet´ın et al., 2011).
In Ref. (Johnson and Munoz) we take up this idea of shared neighbours
(though characterised with a slightly different measure, for reasons we shall
explain in Section C.2) and study analytically the effect of other topological
properties, such as the degree distribution and degree-degree correlations. This
allows us to contrast empirical data with null-models and thus test for statis-
tical significance with no need of computer randomisations. We also comment
on how mutual-neighbour structure could develop in systems of interdependent
networks (such as competition and symbiosis) so as to minimise the risk of a
“cascade of failures” (Buldyrev et al., 2010). Although we are not here con-
cerned specifically with neural systems, a description of this work is included
as an appendix since it serves as an example application of the method put
forward in Ref. (Johnson et al., 2010b) and presented in Chapter 5.
C.2 Definition
Consider a network with N nodes defined by the adjacency matrix a: the
element aij is equal to the number on links, or edges, from node j to node
i (typically considered to be either one or zero).
If a is symmetric, then
the network is undirected and each node i can be characterised by a degree
ki = Pj aij . (If it is directed, i has both an in degree, k in
i = Pj aij , and an out
degree, kout
= Pj aj i ; we shall focus here on undirected networks, although
i
most of the results could be easily extended to directed ones.).
Bastolla et al. (Bastolla et al., 2009) have shown that the effective compe-
tition between two species (say two species of insect) can be reduced if they
have common neighbours with which they are in symbiosis (for instance, if
they both pollinate the same plant). Therefore, in mutualistic networks (net-
C.3 The effect of the degree distribution
115
,
(C.1)
works of symbiotic interactions) it is beneficial to the species at two nodes i
and j for the number of shared symbiotic partners, nij = Pl ail alj = (a2 )ij , to
be high. Going on this, and assuming the network is undirected, the authors
suggest taking into account the following measure:
ηB = Pi<j nij
Pi<j min(ki , kj )
which they call nestedness because it would seem to be highly correlated with
the measures returned by nestedness software. Note that, although the authors
were considering only bipartite graphs, this characteristic is not imposed in the
above definition. In this work, we shall take up the idea of the importance of
nij , but use a slightly different measure of nestedness, for several reasons. One
is that ηB has a serious shortcoming. If we commute the sums1 in the numer-
ator of Eq. (C.1), we find that the result only depends on the heterogeneity of
the degree distribution: Pij nij = Pl Pi ail Pj alj = N hk2 i. Also, although
the maximum value nij can take is min(ki , kj ), this is not necessarily the best
normalisation factor, since the expected number of paths of length 2 connect-
ing nodes i and j depends on both ki and kj (as we show explicitely in Section
C.3). Furthermore, it can sometimes be convenient to have a local measure of
nestedness. For these reasons, we shall use
ηij ≡
nij
kikj
=
(a2)ij
kikj
,
(C.2)
which is defined for every pair of nodes (i, j ). This allows for the consideration
of a nestedness per node, ηi = N −1 Pj ηij , or of the global measure
1
N 2 Xij
ηij .
η =
C.3 The effect of the degree distribution
(C.3)
Most networks have quite broad degree distributions p(k), most notably the
fairly ubiquitous scale-free networks, for which they follow power-laws, p(k) ∼
k−γ . Since this heterogeneity tends to have an importante influence on any
network measure, it will be useful to take this effect into account analytically.
As is standard, the null-model we shall use to do this is the configurational
1 In an undirected network, Pi<j = 1
2 Pij ; we shall always sum over all i and j , since it
is easier to generalise to directed networks and often avoids writing factors 2.
116
Chapter C. Nestedness of networks
model (Newman, 2003c): the set of random networks wired according to the
constraints that a given degree sequence (k1 , ..., kN ) is respected, and also that
there be no degree-degree correlations. The expected value of an element of
the adjacency matrix for networks belonging to this ensemble is
.
aij ≡ ǫc
ij =
kikj
hkiN
We shall use a line, (·), to represent expected values given certain constraints,
and angles, h·i, for averages over nodes of a given network2 . For the case of
the adjacency matrix, we use the notation ǫc
ij = aij for clarity and coherence
with previous work. Plugging Eq. (C.4) into Eq. (C.2), we have the expected
value in the configuration ensemble,
(C.4)
(C.5)
ηij = hk2i
hki2N ≡ ηconf .
Since ηc is independent of i and j , it coincides with the expected value for the
global measure, η = ηconf – a fact that justifies the normalisation chosen in
Eq. (C.2). It is obvious from Eq. (C.5) that degree heterogeneity will have
an important effect on η . Therefore, if we are to capture aspects of network
structure other than those directly induced by the degree distribution, it will
in general be useful to consider the nestedness normalised to this expected
value,
η ≡
η
ηconf
= hki2
hk2iN Xij
Although η is unbounded, it has the advantage that it is equal to unity for
any uncorrelated random network, independently of its degree heterogeneity,
thereby making it possible to detect non-trivial structure in a given empirical
network without the need for computational randomisations.
(a2)ij
kikj
.
(C.6)
C.4 Nestedness and assortativity
In the configuration ensemble, the expected value of the mean degree of the
neighbours of a given node is knn,i = k−1
i Pj ǫc
ij kj = hk2i/hki, which is indepen-
dent of ki . However, real networks usually display degree-degree correlations,
If knn (k) increases (decreases) with k ,
with the result that knn,i = knn (ki ).
2 In this case, for instance, the network considered for hki is any of the members of the
ensemble, since they all have the same mean degree by definition.
C.4 Nestedness and assortativity
117
the network is assortative (disassortative). A measure of this phenomenon
is Pearson’s coefficient applied to the edges (Newman, 2003c, 2002, 2003a;
l ] − [kl ]2 ), where kl and k ′
l ] − [kl ]2 )/([k2
Boccaletti et al., 2006): r = ([klk ′
l are the
degrees of each of the two nodes belonging to edge l, and [·] ≡ (hkiN )−1 Pl (·)
is an average over edges. Writing Pl (·) = Pij aij (·), r can be expressed as
r = hkihk2knn(k)i − hk2i2
.
(C.7)
hkihk3i − hk2i2
The ensemble of all networks with a given degree sequence (k1 , ...kN ) con-
tains a subset for all members of which knn (k) is constant (the configuration
ensemble), but also subsets displaying other functions knn (k).
In Chapter 5 (Johnson et al., 2010b) we showed that there is a one-to-one
mapping between any mean-nearest-neighbour function knn(k) and its corre-
sponding mean-adjacency-matrix ǫ, which is as follows: writing knn (k) as
+ Z dν f (ν )σν+1 (cid:20) kν−1
knn(k) = hk2 i
k (cid:21)
1
hkν i −
hki
with σν+1 ≡ hkν+1i − hkihkν i (which can always be done), the corresponding
matrix ǫ takes the form
(C.8)
ǫij =
(C.9)
N (cid:20) (kikj )ν
j + hkν i(cid:21) .
kikj
f (ν )
+ Z dν
hkν i − kν
i − kν
hkiN
In many empirical networks, knn (k) has the form knn (k) = A + Bkβ , with
A, B > 0 (Boccaletti et al., 2006; Pastor-Satorras et al., 2001) – the mixing
being assortative (disassortative) if β is positive (negative). Such a case is
fitted by Eq. (C.8) if f (ν ) = C [δ (ν − β − 1)σ2/σβ+2 − δ (ν − 1)], with C a
positive constant, since this choice yields
+ C σ2 (cid:20) kβ
knn (k) = hk2i
hki (cid:21) .
1
hkβ+1i −
hki
After plugging Eq. (C.10) into Eq. (C.7), one obtains:
hkβ+1i (cid:18) hkihkβ+2i − hk2 ihkβ+1i
(cid:19) .
C σ2
hkihk3i − hk2i2
It turns out that the configurations most likely to arise naturally (those with
maximum entropy) usually have C ≃ 1 (Johnson et al., 2010b) (c.f. Chapter
5). Therefore, and for the sake of analytical tractability, we shall do as in
(C.11)
(C.10)
r =
118
Chapter C. Nestedness of networks
ǫij =
Chapter 6 and consider this particular case3 – that is, we shall use
σβ+2 (cid:20) (kiki )β+1
N (cid:26) σ2
j + hkβ+1 i(cid:21) + ki + kj − hki(cid:27) .
1
i − kβ+1
hkβ+1i − kβ+1
(C.12)
Substituting the adjacency matrix for this expression in the definition of η (Eq.
(C.6)), we obtain its expected value as a function of the remaining parameter
β :
ηβ =
hk2i "1 + (σ2 − α2
hkβ+1 i (cid:19)2# , (C.13)
hki2
β ρβ ) (cid:18)2 hkβ ihk−1i
β ρβ (cid:18) hkβ i
hkβ+1i − hk−1i2(cid:19) + α2
where αβ ≡ σ2/σβ+2 and ρβ ≡ hk2(β+1) i − hki2(β+1) . Note that η0 = 1.
Fig. C.2 shows the value of ηβ given by Eq.
(C.13) against the assor-
tativity r for various scale-free networks. Nestedness is seen to grow very
fast with increasing disassortativity (decreasing negative r), while in general
slightly assortative networks are less nested than neutral ones. However, highly
heterogeneous networks (γ → 2) show an increase in ηβ for large positive r .
Fig. C.3 shows a plot of nestedness against assortativity for the selection of
empirical networks listed in Table C.4. Although these networks are highly
disparate as regards size, density, degree distribution, etc., it is apparent from
the similarity to Fig. C.2 that the main contribution to ηβ comes indeed from
the assortativity.
C.5 Bipartite networks
Mutualistic networks are usually bipartite: two sets of nodes exist such that
all edges are between nodes in one set and those of another. The ones con-
sidered in Ref. (Bastolla et al., 2009), for instance, are composed of animals
and plants which interact in symbiotic relations of feeding-pollination; these
interactions only take place between animals and plants. Let us therefore con-
sider a bipartite network and call the sets Γ1 and Γ2 , with n1 and n2 nodes,
respectively (n1 + n2 = N ). Using the notation h·ii for averages over set Γi ,
the total number of edges is hki1n2 = hki2n1 = 1
2 hkiN . Assuming that the
network is defined by the configuration ensemble, though with the additional
3Note that C = 1 corresponds to removing the linear term, proportional to ki kj , in Eq.
(C.9), and leaving the leading non-linearity, (ki kj )β+1 , as the dominant one.
C.5 Bipartite networks
119
γ=2.00
γ=2.00
γ=2.00
γ=2.00
γ=2.00
γ=2.00
γ=2.25
γ=2.25
γ=2.25
γ=2.25
γ=2.25
γ=2.50
γ=2.50
γ=2.50
γ=2.50
γ=2.75
γ=2.75
γ=2.75
γ=3.00
γ=3.00
2.5
2.5
2.5
2.5
2.5
2.5
2
2
2
2
2
2
1.5
1.5
1.5
1.5
1.5
1.5
η
η
η
η
η
η
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
-0.1
-0.1
-0.1
-0.1
-0.1
-0.1
0
0
0
0
0
0
0.1
0.1
0.1
0.1
0.1
0.1
0.2
0.2
0.2
0.2
0.2
0.2
r
r
r
r
r
r
0.3
0.3
0.3
0.3
0.3
0.3
0.4
0.4
0.4
0.4
0.4
0.4
0.5
0.5
0.5
0.5
0.5
0.5
0.6
0.6
0.6
0.6
0.6
0.6
Figure C.2: Nestedness against assortativity (as measured by Pearson’s cor-
relation coefficient) for scale-free networks as given by Eq. (C.13). hki = 10,
N = 1000.
1.75
1.75
1.75
1.5
1.5
1.5
1.25
1.25
1.25
η
η
η
1
1
1
0.75
0.75
0.75
0.5
0.5
0.5
-0.3
-0.3
-0.3
-0.2
-0.2
-0.2
-0.1
-0.1
-0.1
0
0
0
0.2
0.2
0.2
0.3
0.3
0.3
0.4
0.4
0.4
0.5
0.5
0.5
0.1
0.1
0.1
r
r
r
Figure C.3: Nestedness against assortativity (as measured by Pearson’s cor-
relation coefficient) for data on a variety of networks. Blue squares are food
webs (Table C.4) and red circles are networks of all other types (Table C.4).
constraint of being bipartite, the probability of node l being connected to node
i is
ǫil = 2
kikl
hkiN
120
Chapter C. Nestedness of networks
Food web
r
ν
hki N
σ/hki
Little Rock lake
Ythan Estuary (w/p)
Stony Stream
Canton Creek
Skipwith Pond
El Verde
Caribbean Reef (small)
St. Martin Island
UK Grassland
Chesapeake Bay
NE US Shelf
Coachella Valley
St. Mark’s Estuary
-0.343
-0.249
-0.201
-0.196
-0.194
-0.183
-0.172
-0.165
-0.125
-0.123
-0.088
0.043
0.118
1.219
1.323
1.163
1.171
0.891
1.088
1.000
1.071
0.907
0.801
0.971
0.857
0.816
20.4
8.9
14.7
13.5
14.2
18.4
19.7
9.3
2.8
4.1
34.3
14.6
8.5
92
82
109
102
25
155
50
42
61
31
79
29
48
0.73
0.93
0.75
0.69
0.37
0.88
0.49
0.56
0.82
0.60
0.45
0.41
0.55
Table C.1:
Food webs appearing in Fig. C.3 (listed from least to most
assortative) : r is the assortativity and ν the nestedness. The origins of all
data cited in Ref. (Dunne et al., 2004), and kindly provided to us by Jennifer
Dunne.
ηbip =
if they belong to different sets, and zero if they are in the same one. Proceeding
as before, we find that the expected value of the nestedness for a bipartite
network is
N 2 " Xi,j∈Γ1
1
hki2n1 # =
kikl
1
klkj
kikj Xl∈Γ1
hki1n2
n1hk2 i2 + n2 hk2i1
(C.14)
hki1hki2(n1 + n2 )2 .
Interestingly, if n1 = n2 , the fact that the network is bipartite has no effect on
the nestedness: ηbip = ηconf .
1
kikj Xl∈Γ2
kikl
hki1n2
klkj
hki2n1
+ Xi,j∈Γ2
C.6 Overlapping networks
If the adjacency matrix a describes a mutualistic network, the benefit to its
being nested resides in a counteraction of the competition matrix c, which
C.6 Overlapping networks
121
Network
r
ν
hki N
σ/hki Ref.
Political blogs
Metabolic
Political books
Adjectives and nouns
Dolphins
Power grid
Neural
Jazz musicians
Email
American football
PGP
High-energy arXiv
Net-science arXiv
-0.221
-0.220
-0.138
-0.125
-0.063
0.003
0.005
0.020
0.078
0.133
0.239
0.294
0.462
1.496
1.688
0.996
1.057
0.922
0.834
0.907
0.924
0.923
0.904
0.867
0.533
0.443
22.4
9.0
8.4
7.6
5.1
2.7
5.9
27.6
9.6
10.6
4.6
3.8
3.45
1490
453
104
111
61
4940
306
198
1133
114
10680
8360
1588
1.62
1.87
0.65
0.89
0.58
0.67
0.81
0.63
0.97
0.08
1.77
1.14
1.00
(Adamic and Glance, 2005)
(Duch and Arenas, 2005)
(Krebs)
(Newman, 2006)
(Lusseau et al., 2003)
(Watts and Strogatz, 1998)
(Watts and Strogatz, 1998)
(P.Gleiser and Danon, 2003)
(Guimer`a et al., 2003)
(Girvan and Newman, 2002)
(Bogn´a et al., 2004)
(Newman, 2001)
(Newman, 2006)
Table C.2: Empirical networks appearing in Fig. C.3 (listed from least to most
assortative) : r is the assortativity and ν the nestedness. All data available on
the personal Web pages of ´Alex Arenas, Mark Newman and Duncan Watts.
takes into account the extent to which one species is detrimental to another
due to predation, sharing of resources, etc. From this point of view, it may be
interesting to study to what extent matrices c and a2 overlap (note that both
networks have the same nodes, but different edges). Presumably, if ecological
networks are assembled in such a way that effective competition is minimised,
this overlap should be higher than randomly expected. On the other hand, a
certain degree of overlap may also arise from the fact that species interacting
symbiotically with the same host are perhaps more than averagely likely to
be phylogenetically close and/or phenotypically similar, leading (as Darwin
noted) to a higher competition element.
In any case, a measure of this overlap is
1
hkicN Xij
where h·ic represents an average over the competion network; similarly, h·ia
will stand for an average over the mutualistic network. If the two networks
cij (a2 )ij ,
r ≡
(C.15)
122
Chapter C. Nestedness of networks
are mutually uncorrelated4 – i.e., if the existence of an edge in one provides no
information as to whether there is a corresponding one in the other – we can
write
cij
(C.16)
1
1
hkicN Xij
N 2 Xij
(a2)ij ≡ runc .
r ≃
Using Pij (a2 )ij = hk2 iaN , and assuming that c is normalised so that Pij cij =
hkicN , we have5
runc ≃ hk2ia
(C.17)
,
N
which only depends on the heterogeneity of the degree distribution of the
mutualistic network. Again, it may be useful to consider the overlap normalised
to this value,
r ≡
=
r
runc
cij (a2)ij .
1
hkic hk2ia Xij
This measure will equal unity when there is no statistical relation between
the competition matrix and the mutualistic one, but can be expected to be
greater if indeed such an overlap were contributing to a reduction in effective
competition.
(C.18)
It has recently been shown that interconnected networks are prone to dan-
gerous “cascades of failures” (Buldyrev et al., 2010). It seems that the northen
half of Italy was once left temporarily with no electric supply due to failures in
the power-grid closing down dependent internet servers, which in turn further
disrupted the grid, until many nodes of both networks were rendered dysfunc-
tional. If two inter-dependent networks were to coincide perfectly (r = 1), the
resilience of the system to node removal would be the same as that of just one
network; however, lower overlap leads to increased vulnerability to such cas-
cades of failures. Since the extinction of a species can result in its host species
also going extinct, such cascades of failures may be a threat to mutualistic
systems. In such a case, it would seem that a high overlap r , as defined here,
between the competition matrix and the mutualistic one would minimise this
possibility. It would be interesting to test this experimentally.
4Note that we are saying nothing of the internal correlations that each network may
display.
5The competition matrix will in general be weighted, as could be the mutualistic one; we
shall treat both as though they were not, but using weighted networks would only influence
results by a normalisation factor.
C.7 Discussion
C.7 Discussion
123
Whether or not the topological feature here described should be considered a
measure of nestedness as it is usually understood in ecology is not clear. What
is certain is that interactions between dynamical elements that are mediated
by third parties, or common neighbours, can be relevant in a wide variety of
settings. We have mentioned the paradigmatical case of ecosystems as well
as financial and communications networks. But other examples spring easily
to mind. For instance, two excitatory neighbouring neurons might have their
mutual effect dampened if they share inhibitory neighbours. Genetic networks
are riddled with motifs such that switches activate or inactivate each other
indirectly, via common neighbours. As we have shown, there are nontrivial
relationships between nestedness, as it is here defined, and other topological
features. If it turns out that this network property is indeed relevant for many
complex systems, then we hope the null models we have laid out and analysed
will prove useful in assessing its functional significance.
Appendix D
Publications derived from the
thesis
D.1 Journals and book chapters (the most rel-
evant ones marked with an asterisk)
1. * Cluster Reverberation: A mechanism for robust short-term memory
without synaptic learning, S. Johnson, J. Marro, and JJ. Torres, submit-
ted, arXiv:1007.3122
2. * Enhancing neural-network performance via assortativity, S. de Fran-
ciscis, S. Johnson, and J.J. Torres, Physical Review E 83, 036114 (2011)
3. Why are so many networks disassortative? S. Johnson, J.J. Torres, J.
Marro, and M.A. Munoz, AIP Conf. Proc. 1332, 249–50 (2011)
4. Shannon entropy and degree-degree correlations in complex networks, S.
Johnson, J.J. Torres, J. Marro, and M.A. Munoz, “Nonlinear Systems
and Wavelet Analysis”, Ed. R. L´opez-Ruiz, WSEAS Press, pp. 31–35
(2010)
5. * Entropic origin of disassortativity in complex networks, S. Johnson,
J.J. Torres, J. Marro, and M.A. Munoz, Physical Review Letters 104,
108702 (2010)
6. * Evolving networks and the development of neural systems, S. Johnson,
J. Marro, and J.J. Torres, Journal of Statistical Mechanics (2010) P03003
124
D.2 Abstracts
125
7. Excitable networks: Nonequilibrium criticality and optimum topology,
J.J. Torres, S. de Franciscis, S. Johnson, and J. Marro, International
Journal of Bifurcation and Chaos 20, 869–875 (2010)
8. Nonequilibrium behavior in neural networks: criticality and optimal per-
formance, J.J. Torres, S. Johnson, J.F. Mejias, S. de Franciscis, and J.
Marro, “Advances in Cognitive Neurodynamics (II)” Eds. R. Wang and
F. Gu, pp 597–603, Springer, 2011, ISBN: 978-90-481-9694-4, Proceed-
ings of Second International Conference on Cognitive Neurodynamics
(ICCN2009), Hangzhou 15-19 November 2009.
9. Development of neural network structure with biological mechanisms, S.
Johnson, J. Marro, J.F. Mejias, and J.J. Torres, Lecture Notes in Com-
puter Science 5517, 228–235 (2009)
10. Switching dynamics of neural systems in the presence of multiplicative
colored noise, J.F. Mejias, J.J. Torres, S. Johnson, and H.J. Kappen,
Lecture Notes in Computer Science 5517, 17–23 (2009)
11. * Nonlinear preferential rewiring in fixed-size networks as a diffusion
process, S. Johnson, J.J. Torres, and J. Marro, Physical Review E 79,
050104(R) (2009)
12. * Functional optimization in complex excitable networks, S. Johnson, J.J.
Torres, and J. Marro, EPL 83, 46006 (2008)
13. Excitable networks: Non-equilibrium criticality and optimum topology,
J.J. Torres, S. de Franciscis, S. Johnson, and J. Marro, “Modelling and
Computation on Complex Networks and Related Topics”, Eds. Criado,
Gonzalez-Vias, Mancini and Romance. Proceedings of the conference
”Net-Works 2008”, 185–192, ISBN:978-84-691-3819-9.
14. Topology-induced instabilities in neural nets with activity-dependent synapses,
S. Johnson, J. Marro, and J. J. Torres, “New Trends and Tools in Com-
plex Networks”, Eds. Criado, Pello and Romance. Proceedings of the
conference ”Net-Works 2007”, 59–71, ISBN:978-84-690-6890-8.
D.2 Abstracts
1. Network topology and dynamical task performance, S. Johnson, J. Marro,
and J.J. Torres, AIP Conf. Proc. 1091, 280 (2009)
126
Chapter D. Publications derived from the thesis
2. Constructive chaos in excitable networks with tuneable topologies, S. John-
son, J. Marro, and J.J. Torres, XV Congreso de Fsica Estadstica FisEs08,
104 (2008)
3. The effect of topology on neural networks with unstable memories, S.
Johnson, J. Marro, and J.J. Torres, AIP Conf. Proc. 887 261 (2006)
4. Relationship between the solar wind and the upper-frequency limit of
Saturn Kilometric Radiation, M.Y. Boudjada, P.H.M. Galopeau, H.O.
Rucker, A. Lecacheux, W.S. Kurth, D.A. Gurnett, U. Taubencshuss,
J.T. Steinberg, S. Johnson, and W. Vollerr, European Geosciences Union
(2006)
128
Chapter D. Publications derived from the thesis
References
H. A. and S. R. Magnetic Domains. Springer, Berlin, 1998.
I. Abbott and R. Black. Changes in species composition of floras on islets near
perth, western australia. Journal of Biogeography, 7:399–410, 1980.
L. Abbott and T. Kepler. From Hodgkin-Huxley to Hopfield, in Statistical
mechanics of neural networks. Springer-Verlag, Berlin, 1990.
L. Abbott, J. Varela, K. Sen, and S. Nelson. Synaptic depression and cortical
gain control. Science, 275:220, 1997.
L. Adamic and N. Glance. The political blogosphere and the 2004 US Election.
Proceedings of the WWW-2005 Workshop on the Weblogging Ecosystem,
2005.
R. Albert and A.-L. Barab´asi. Topology of evolving networks: Local events
and universality. Phys. Rev. Lett., 85(24):5234–5237, 2000.
R. Albert and A.-L. Barab´asi. Statistical mechanics of complex networks. Rev.
Mod. Phys., 74:47–97, 2002.
L. Amaral, A. Scala, M. Barth´el`emy, and H. Stanley. Classes of small-world
networks. Proc. Natl. Acad. Sci. USA, 97:11149–52, 2000.
S. Amari. Characteristics of random nets of analog neuron-like elements. IEEE
Trans. Syst. Man. Cybern., 2:643–657, 1972.
D. Amit. The Hebbian paradigm reintegraged: Local reverberations as internal
representations. Behavioral and Brain Sciences, 18:617–57, 1995.
D. J. Amit. Modeling Brain Function: The World of Attractor Neural Net-
works. Cambridge University Press, Cambridge, UK, 1989.
K. Anand and G. Bianconi. Entropy measures for networks: Toward an infor-
mation theory of complex topologies. Phys. Rev. E, 80:045102, 2009.
129
130
REFERENCES
L. Antiqueira, F. Rodrigues, and L. Costa. Letter: Modeling connectivity in
terms of network activity. Journal of Statistical Mechanics: Theory and
Experiment, 9, 2009.
A. Arenas, A. D´ıaz-Guilera, and C. P´erez-Vicente. Synchronization reveals
topological scales in complex networks. Phys. Rev. Lett., 96:114102, 2006.
A. Arenas, A. D´ıaz-Guilera, J. Kurths, Y. Y. Moreno, and C. Zhou. Synchro-
nization in complex networks. Phys. Rep., 469:93–153, 2008a.
A. Arenas, A. Fern´andez, and S. G. S. A complex network approach to the
determination of functional groups in the neural system of c. elegans. Lect.
Notes Comp. Sci., 5151:9–18, 2008b.
looking back and looking forward.
Baddeley and A.D. Working memory:
Nature Reviews Neuroscience, 4:829–39, 2003.
A. Baddeley. Essentials of Human Memory. Psychology Press, East Sussex,
UK, 1999.
P. Bak, K. Chen, and C. Tang. A forest-fire model and some thoughts on
turbulence. Phys. Lett. A, 147:297, 1990.
A. Barab´asi and Z. Oltvai. Network biology: understanding the cell’s func-
tional organization. Nature Reviews Genetics, 5:101–3, 2004.
A.-L. Barab´asi and R. Albert. Emergence of scaling in random networks.
Science, 286:509–512, 1999.
M. Barahona and L. Pecora. Synchronization in small-world systems. Phys.
Rev. Lett., 89:054101, 2002.
O. Barak and M. Tsodyks. Persistent activity in neural networks with dynamic
synapses. PLoS Comput. Biol., 3(2):e35, 2007.
J. Barrie, J. Freeman, and M. Lenhart. Spatiotemporal analysis of prepyri-
form, visual, auditory, and somesthetic surface EEGs in trained rabbits. J.
Neurophysiol., 76:520, 1996.
U. Bastolla, M. Fortuna, A. Pascual-Garc´ıa, A. Ferrera, B. Luque, and J. Bas-
compte. Nature, 458:1018–21, 2009.
R. Baxter. Exactly Solved Models in Statistical Mechanics. Academic Press,
London, 1982.
REFERENCES
131
N. Bertschinger and T. Natschlager. Real-time computation at the edge of
chaos in recurrent neural networks. Neural Comp., 16:1413, 2004.
G. Bianconi. Mean-field solution of the Ising model on a Barab´asi-Albert
network. Phys. Lett. A, 303:166–8, 2002.
G. Bianconi. The entropy of randomized network ensembles. EPL, 81:28005,
2008.
G. Bianconi. Entropy of network ensembles. Phys. Rev. E, 79(3):036114, 2009.
G. Bianconi and A.-L. Barab´asi. Competition and multiscaling in evolving
networks. EPL, 54:436–442, 2001.
V. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre. Fast unfolding of
community hierarchies in large networks. J. Stat. Mech., P10008, 2008.
S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D. Hwang. Complex
networks: Structure and dynamics. Physics Reports, 424:175–308, 2006.
M. Bogn´a, R. Pastor-Satorras, A. D´ıaz-Guilera, and A. Arenas. Models of
social networks based on social distance attachment. Phys. Rev. E, 70:
056122, 2004.
M. Bogun´a and R. Pastor-Satorras. Class of correlated random networks with
hidden variables. Phys. Rev. E, 68:036112, 2003.
M. Bogun´a, F. Papadopoulos, and D. Krioukov. Sustaining the internet with
hyperbolic mapping. Nature Communications, 1:62, 2010.
B. Bollob´as. Random Graphs. Cambridge University Press, Cambridge, 2001.
J. Bonachela, S. de Franciscis, J. Torres, and M. Munoz. Self-organization
without conservation: Are neuronal avalanches generically critical? J. Stat.
Mech., page P02015, 2010.
P. Bonifazi, M. Goldin, M. Picardo, I. Jorquera, A. Cattani, G. Bianconi,
A. Represa, Y. Ben-Ari, and R. Cossart. Gabaergic hub neurons orchestrate
synchrony in developing hippocampal networks. Science, 326:1419, 2009.
M. Brede and S. Sinha. Assortative mixing by degree makes a network more
unstable. arXiv:cond-mat/0507710.
S. Brush. History of the Lenz-Ising model. Rev. Mod. Phys., 39:883–93, 1967.
132
REFERENCES
S. Buldyrev, R. Parshani, G. Paul, H. Stanley, and S. Havlin. Catastrophic
cascade of failures in interdependent networks. Nature, 464:1025–8, 2010.
E. Bullmore and O. Sporns. Complex brain networks: graph theoretical anal-
ysis of structural and functional systems. Nature Reviews Neuroscience, 10:
186–98, 2009.
Z. Burda, J. Jurkiewicz, and A. Krzywicki. Statistical mechanics of random
graphs. Physica A: Statistical Mechanics and its Applications, 344:56–61,
2004.
G. Caldarelli, A. Capocci, P. D. L. Rios, and M. M. oz. Scale-free networks
from varying vertex intrinsic fitness. Phys. Rev. Lett., 89:258702, 2002.
M. Camperi and X.-J. Wang. A model of visuospatial working memory in
prefrontal cortex: recurrent network and cellular bistability. J. Comp. Neu-
rosci., 5:383–405, 1998a.
M. Camperi and X.-J. Wang. A model of visuospatial working memory in
prefrontal cortex: recurrent network and cellular bistability. J. Comp. Neu-
rosci., 5:383–405, 1998b.
A. Capocci and F. Colaiori. Mixing properties of growing networks and simp-
son’s paradox. Phys. Rev. E., 74:026122, 2006.
M. Chavez, D.-U. Hwang, A. Amann, H. Hentschel, and S. Boccaletti. Syn-
chronization is enhanced in weighted complex networks. Phys. Rev. Lett.,
94:218701, 2005.
G. Chechik, I. Meilijson, and E. Ruppin. Neuronal regulation: a biologically
plausible mechanism for efficient synaptic pruning in development. Neuro-
computing, 26:633, 1999.
G. Chechik, I. Meilijson, and E. Ruppin. Synaptic pruning in development: A
computational account. Neural Computation, in press.
D. Chialvo. Critical brain networks. Physica A: Statistical Mechanics and its
Applications, 340:756–765, 2004.
D. Chialvo. Psychophysics: Are our senses critical? Nature Physics, 2:301–2,
2006.
REFERENCES
133
D. Chialvo, P. Balenzuela, and D. Fraiman. The brain: What is critical about
it? In . E. P. L. M. Ricciardi, A. Buonocore, editor, Col lective Dynamics:
Topics on Competition and Cooperation in the Biosciences, volume 1028 of
American Institute of Physics Conference Series, pages 28–45, 2008.
A. Compte, C. Constantinidis, J. Tegner, S. Raghavachari, M. Chafee, and et
al. Temporally irregular mnemonic persistent activity in prefrontal neurons
of monkeys during a delayed response task. J. Neurophysiol., 90:3441–54,
2003.
R. Conrad. Acoustic confusion in immediate memory. B. J. Psychol., 55:
75–84, 1964a.
R. Conrad.
Information and acoustic confusion and memory span. B. J.
Psychol., 55:429–32, 1964b.
J. Cortes, P. Garrido, H. Kappen, J. Marro, C. Morilla, D. Navidad, and
J. Torres. Algorithms for identification and categorization. AIP Conf. Proc.,
779:178, 2005.
J. Cortes, J. Torres, J. Marro, P. Garrido, and H. Kappen. Effects of fast
presynaptic noise in attractor neural networks. Neural Comp., 18:614, 2006.
N. Cowan. On short and long auditory stores. Psychological Bul letin, 96:
341–70, 1984.
V. de Assis and M. Copelli. Dynamic range of hypercubic stochastic excitable
media. Phys. Rev. E, 77:011923, 2008.
S. de Franciscis, S. Johnson, and J. Torres. Enhancing neural-network perfor-
mance via assortativity. Phys. Rev. E, in press, 2011.
D. J. de S. Price. Networks of scientic papers. Science, 149:510–5, 1965.
I. Der´enyi, I. Farkas, G. Palla, and T. Vicsek. Topological phase transitions
of random networks. Physica A: Statistical Mechanics and its Applications,
334:583–590, 2004.
D. Dominguez, M. Gonz´alez, E. Serrano, and F. Rodr´ıguez. Structured infor-
mation in small-world neural networks. Phys. Rev. E, 79:021909, 2009.
V. Dom´ınguez-Chibet´ın, S. Johnson, and M. Munoz. Hidden interactions in
food-webs: Unearthing the missing links. in preparation, 2011.
134
REFERENCES
L. Donetti and M. A. Munoz. Detecting network communities: a new sys-
tematic and powerful algorithm. J. Stat. Mech.: Theor. Exp., page P10012,
2004.
S. Dorogovtsev and J. Mendes. Evolution of networks: From biological nets to
the Internet and WWW. Oxford Univ. Press, Oxford, 2003.
S. Dorogovtsev, A. Ferreira, A. Goltsev, and J. Mendes. Zero pearson coeffi-
cient for strongly correlated growing trees. Phys. Rev. E, 81:031135, 2005.
J. Duch and A. Arenas. Community identification using Extremal Optimiza-
tion. Phys. Rev. E, 72:027104, 2005.
J. Dunne, R. Williams, and N. Martinez. Network structure and robustness of
marine food webs. Mar. Ecol. Prog. Ser., 273:291–302, 2004.
D. Durstewitz, J. Seamans, and T. S. and. Neurocomputational models of
working memory. Nature Neuroscience, 3:1184–91, 2000.
V. Egu´ıluz, D. Chialvo, G. Cecchi, M. Baliki, and A. Apkarian. Scale-free
brain functional networks. Phys. Rev. Lett., 94:018102, 2005.
P. Erdos and A. R´enyi. On random graphs I. Publ. Math. Debrecen, 6:290–7,
1959.
L. Euler. Solutio problematis ad geometriam situs pertinentis. Acad. Sci. U.
Petrop., 8:128–40, 1736.
I. Farkas, I. Der´enyi, G. Palla, and T. Vicsek. Lect. Notes in Phys., 650:163,
2004.
E. Frank. Synapse elimination: For nerves it’s all or nothing. Science, 275:
324–325, 1997.
A. Fronczak and P. Fronczak. Networks with given two-point correlations:
hidden correlations from degree correlations. Phys. Rev. E, 74:026121, 2006.
G. G. Bianconi and A.-L. Barab´asi. Bose-Einstein condensation in complex
networks. Phys. Rev. Lett., 86:5632–5635, 2001.
M. Garey and D. Johnson. Computers and Intractability: A Guide to the
Theory of NP-Completeness. W.H. Freeman and Company, USA, 1979.
REFERENCES
135
E. Gilbert. Random graphs. Annals of Mathematical Statistics, 30:1141–4,
1959.
M. Girvan and M. Newman. Community structure in social and biological
networks. Proc. Natl. Acad. Sci. USA, 99:7821–6, 2002.
J. G´omez-Gardenes, Y. Moreno, and A. Arenas. Paths to synchronization on
complex networks. Phys. Rev. Lett., 98:034101, 2007.
A. Gruart, M. Munoz, and J. Delgado-Garc´ıa.
Involvement of the ca3-ca1
synapse in the acquisition of associative learning in behaving mice. J. Neu-
rosci., 26:1077–87, 2006.
R. Guimer`a, L. Danon, A. D´ıaz-Guilera, F. Giralt, and A. Arenas. Self-similar
community structure in a network of human interactions. Phys. Rev. E, 68:
065103(R), 2003.
D. Hebb. The Organization of Behavior. Wiley, New York, 1949.
S. Hilfiker, V. Pieribone, H.-T. Kao, A. Czernik, G. Augustine, and P. Green-
gard. Synapsins as regulators of neurotransmitter release. Phil. Trans. R.
Soc. London B, 354:269–79, 1999.
A. Hodgkin and A. Huxley. A quantitative description of membrane current
and its application to conduction and excitation in nerve. J. Physiol., 117:
500–44, 1952.
H. Holcman and M. Tsodyks. The emergence of Up and Down states in cortical
networks. PLoS Comput. Biol., 2:e23, 2006.
P. Holme and J. Zhao. Exploring the assortativity-clustering space of a net-
work’s degree sequence. Phys. Rev. E, 75:046111, 2007.
J. Hopfield. Neural networks and physical systems with emergent collective
computational abilities. Proc. Natl. Acad. Sci. USA, 79:2554–8, 1982.
P. Hurtado, J. Marro, and P. Garrido. Demagnetization via nucleation of the
nonequilibrium metastable phase in a model of disorder. J. Stat. Phys., 133:
29–58, 2008.
P. R. Huttenlocher and A. S. Dabholkar. Regional differences in the synap-
togensis in human cerebral cortex. The Journal of Comparative Neurology,
387:167–178, 1997.
136
REFERENCES
E. M. Izhikevich. Dynamical Systems in Neuroscience: The Geometry of Ex-
citability and Bursting. MIT Press, Cambridge MA, 2007.
E. Jaynes. Information theory and statistical mechanics. Phys. Rev., 106(4):
620–630, 1957.
S. Johnson and M. Munoz. Nestedness of networks: Null models and topolog-
ical effects. in preparation.
S. Johnson, J. Torres, and J. Marro. Functional optimization in complex
excitable networks. EPL, 83:46006, 2008.
S. Johnson, J. Marro, and J. Torres. Development of neural network structure
with biological mechanisms. Lecture Notes in Computer Science, 5517:228–
35, 2009a.
S. Johnson, J. Torres, and J. Marro. Nonlinear preferential rewiring in fixed-
size networks as a diffusion process. Phys. Rev. E, 79:050104, 2009b.
S. Johnson, J. Marro, and J. Torres. Evolving networks and the development
of neural systems. J. Stat. Mech., P03003, 2010a.
S. Johnson, J. Torres, J. Marro, and M. Munoz. Entropic origin of disassorta-
tivity in complex networks. Physical Review Letters, 104:108702, 2010b.
S. Johnson, J. Marro, and J. Torres. Cluster Reverberation: A mecha-
nism for robust short-term memory without synaptic learning. submitted.
arXiv:1007.3122, 2011.
M. Kaiser and C. Hilgetag. Spatial growth of real-world networks. Phys. Rev.
E, 69:036103, 2004.
M. Kaiser, R. Martin, P. Andras, and M. Young. Simulation of robustness
against lesions of cortical networks. Eur. J. Neurosci., 25:3185–92, 2007.
O. Kinouchi and M. Copelli. Optimal dynamical range of excitable networks
at criticality. Nature Physics, 2(5):348–351, 2006.
A. Klintsova and W. Greenough. Synaptic plasticity in cortical systems. Cur-
rent Opinion in Neurobiology, 9:203, 1999.
H. Korn and P. Faure. Is there chaos in the brain? II. Experimental evidence
and related models. C. R. Biol., 326:787, 2003.
REFERENCES
137
G. Kossinets and D. Watts. Science, 311:88–90, 2006.
P. L. Krapivsky, S. Redner, and F. Leyvraz. Connectivity of growing random
networks. Phys. Rev. Lett., 85:4629–4632, 2000.
V. Krebs. unpublished, http://www.orgnet.com/.
A. Kwag and O. Paulsen. The timing of external input controls the sign of
plasticity at local synapses. Nat Neurosci., 12:1219–21, 2009.
J. D. L. Danon, A. D´ıaz-Guilera, and A. Arenas. Comparing community
structure identification. J. Stat. Mech., P09008, 2005.
K. Lee, F. Schottler, M. Oliver, and G. Lynch. Brief bursts of high-frequency
stimulation produce two types of structural change in rat hippocampus. J.
Neurophysiol., 44:247, 1980.
Levine and R.D. Molecular Reaction Dynamics. Cambridge University Press,
Cambridge, 2005.
B. Lindner, J. Garc´ıa-Ojalvo, A. Neiman, and L. Schimansky-Geier. Effects of
noise in excitable systems. Phys. Rep., 392:321, 2004.
D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase, E. Slooten, and S. Daw-
son. The bottlenose dolphin community of doubtful sound features a large
proportion of long-lasting associations. Behavioral Ecology and Sociobiology,
54:396–405, 2003.
P. J. Magistretti. Neuroscience. low-cost travel in neurons. Science, 325:1349–
51, 2009.
R. Malenka and R. Nicoll. Long-term potentiation – a decade of progress?
Science, 285:1870–4, 1999.
Marcus and G.F. The Algebraic Mind: Integrating Connectionism and Cogni-
tive Science. MIT Press, Cambrigde and MA, 2001.
J. Marro and R. Dickman. Nonequilibrium Phase Transitions in Lattice Models.
Cambridge.
J. Marro, J. Torres, and J. Cortes. Complex behavior in a network with time-
dependent connections and silent nodes. J. Stat. Mech., page P02017, 2008.
138
REFERENCES
S. Maslov, K. Sneppen, and A. Zaliznyak. Pattern detection in complex net-
works: correlation profile of the internet. Physica A, 333:529–40, 2004.
R. May. Stability and Complexity in Model Ecosystems. Princeton University
Press, Princeton, 1973.
W. McCulloch and W. Pitts. A logical calculus of the ideas immanent in
nervous activity. Bul letin of Mathematical Biophysics, 7:115–33, 1943.
J. Mejias and J. Torres. Maximum memory capacity on neural networks with
short-term synaptic depression and facilitation. Neural Comput., 21:851–71,
2009.
J. Mejias, H. Kappen, and J. Torres.
Irregular dynamics in up and down
cortical states. PLoS ONE, 5 (11):e13651, 2010.
E. Meron. Pattern formation in excitable media. Phys. Rep., 218:1, 1992.
S. Milgram. The Small World problem. Psychology Today, 1:61–67, 1967.
Y. Miyashita. Neuronal correlate of visual associative long-term memory in
the primate temporal cortex. Nature, 335:817–20, 1988.
G. Mongillo, O. Barak, and M. Tsodyks. Synaptic theory of working memory.
Science, 319:1543–6, 2008.
R. Morgan and I. Soltesz. Nonrandom connectivity of the epileptic dentate
gyrus predicts a ma jor role for neuronal hubs in seizures. Proc. Nat. Acad.
Sci. USA, 105:6179, 2008.
A. Motter, C. Zhou, and J. Kurths. Network synchronization, diffusion, and
the paradox of heterogeneity. Phys. Rev. E, 71:016116, 2005.
M. Munoz, R. Juhasz, C. Castellano, and G. Odor. Griffiths phases on complex
networks. Phys. Rev. Lett., 105:128701, 2010.
M. Newman. The structure of scientific collaboration networks. Proc. Natl.
Acad. Sci. USA, 98:404–9, 2001.
M. Newman. Assortative mixing in networks. Phys. Rev. Lett., 89:208701,
2002.
M. Newman. Mixing patterns in networks. Phys. Rev. E, 67:026126, 2003a.
REFERENCES
139
M. Newman. Random graphs as models of networks, in Handbook of Graphs
and Networks, S. Bornholdt and H.G. Schuster (eds.). Wiley-VCH, Berlin,
2003b.
M. Newman. The structure and function of complex networks. SIAM Review,
45:167–256, 2003c.
M. Newman. Power laws, Pareto distributions and Zipf ’s law. Contemporary
Physics, 46:323, 2005.
M. Newman and J. Park. Why social networks are different from other types
of networks. Phys. Rev. E, 68:036122, 2003.
M. E. J. Newman. Finding community structure in networks using the eigen-
vectors of matrices. Preprint physics/0605087, 2006.
T. Nishikawa, A. Motter, Y.-C. Lai, and F. Hoppensteadt. Heterogeneity in
oscillator networks: Are smaller worlds easier to synchronize? Phys. Rev.
Lett., 91:014101, 2003.
L. Onsager. Crystal statistics. I. A two-dimensional model with an order-
disorder transition. Phys. Rev., (2) 65:117–49, 1944.
L. Pantic, J. Torres, H. Kappen, and S. Gielen. Associative memory with
dynamic synapses. Neural Comp., 14:2903, 2002.
J. Park and M. Newman. Origin of degree correlations in the internet and
other networks. Phys. Rev. E, 68:026112, 2003.
J. Park and M. E. J. Newman. Statistical mechanics of networks. Phys. Rev.
E, 70:066117, 2004.
K. Park, Y.-C. Lai, and N. Ye. Self-organized scale-free networks. Phys. Rev.
E, 72:026131, 2005.
R. Pastor-Satorras and A. Vespignani. Evolution and structure of the Internet:
A statistical physics approach. Cambridge Univ. Press, Cambridge, 2004.
R. Pastor-Satorras, A. V´azquez, and A. Vespignani. Dynamical and correlation
properties of the internet. Phys. Rev. Lett., 87:258701, 2001.
T. Peixoto. Redundancy and error resilience in boolean networks. Phys. Rev.
Lett., 104:048701, 2010.
140
REFERENCES
T. Pereira. Hub synchronization in scale-free networks. Phys. Rev. E, 82:
036201, 2010.
P. Peretto. An Introduction to the Modeling of Neural Networks. Cambridge
Univ. Press, Cambridge, 1992.
P.Gleiser and L. Danon. Community structure in Jazz. Adv. Complex Syst.,
6:565, 2003.
J. Poncela, J. Gmez-Gardees, L. Flor´ıa, A. S´anchez, and Y. Moreno. Complex
cooperative networks from evolutionary preferential attachment. PLoS ONE,
3(6):e2449, 2008.
S. Ree. Generation of scale-free networks using a simple preferential-rewiring
dynamics. Physica A: Statistical Mechanics and its Applications, 376:692–
698, 2007.
J. Reichardt and S. Bornholdt. Statistical mechanics of community detection.
Phys. Rev. E, 74:016110, 2006.
A. Rodr´ıguez-Moreno and O. Paulsen. Spike timing-dependent long-term de-
pression requires presynaptic nmda receptors. Nat Neurosci., 11:744–5, 2008.
M. D. Roo, P. Klauser, P. Mendez, L. Poglia, and D. Muller. Activity-
dependent PSD formation and stabilization of newly formed spines in hip-
pocampal slice cultures. Cerebral Cortex, 18:151–161, 2008.
Y. Roudi and P. Latham. A balanced memory network. PLoS Comput. Biol.,
3.
A. Rozenfeld, R. Cohen, D. ben Avraham, and S. Havlin. Scale-free networks
on lattices. Phys. Rev. Lett., 89:218701, 2002.
K. Sakai and Y. Miyashita. Neuronal organization for the long-term memory
of paired associates. Nature, 354:152–5, 1991.
I. Sendina-Nadal, J. Buld´u, I. Leyva, and S. Boccaletti. Phase locking induces
scale-free topologies in networks of coupled oscillators. PLoS ONE, 3:e2644,
2008.
D. Sherrington and S. Kirkpatrick. Solvable model of a spin-glass. Phys. Rev.
Lett., 35:1792–6, 1975.
REFERENCES
141
S. Sikstrom. Forgetting curves: implications for connectionist models. Cogni-
tive Psychology, 45:95–152, 2002.
H. A. Simon. On a class of skew distribution functions. Biometrika, 42:425–40,
1955.
B. Soderberg. General formalism for inhomogeneous random graphs. Phys.
Rev. E, 66:066121, 2002.
G. Sperling. The information available in brief visual persentation. Psycho-
logical Monographs: General and Applied, 74:1–30, 1960.
O. Sporns, D. R. Chialvo, M. Kaiser, and C. C. Hilgetag. Organization, de-
velopment and function of complex brain networks. Trends in Cognitive
Sciences, 8:418–425, 2004.
K. Suchecki, V. Egu´ıluz, and M. S. Miguel. Conservation laws for the voter
model in complex networks. EPL, 69:228–34, 2005.
G. Suel, J. Garcia-Ojalvo, L. Liberman, and M. Elowitz. An excitable gene
regulatory circuit induces transient cellular differentiation. Nature, 440:545–
50, 2006.
G. Sugihara and H. Ye. Complex systems: Cooperative network dynamics.
Nature, 458:979–80, 2009.
E. Tarnow. Short term memory may be the depletion of the readily releasable
pool of presynaptic neurotransmitter vesicles. Cognitive Neurodynamics, 3:
263–9, 2008.
J.-N. Teramae and T. Fukai. A cellular mechanism for graded persistent ac-
tivity in a model neuron and its implications for working memory. J. Comp.
Neurosci., 18:105–21, 2005.
S. Thurner, F. Kyriakopoulos, and C. Tsallis. Phys. Rev. E., 76:036111, 2007.
J. Torres and P. Varona. Modeling Biological Neural Networks, in Handbook
of Natural Computing, ed. G. Rozenberg et al. Springer Verlag, Berlin, Ger-
many, 2010.
J. Torres, M. Munoz, J. Marro, and P. L. Garrido. Influence of topology on
the performance of a neural network. Neurocomputing, 58-60:229–234, 2004.
142
REFERENCES
J. Torres, J. Cortes, J. Marro, and H. Kappen. Competition between synaptic
depression and facilitation in attractor neural networks. Neural Computa-
tion, 19:2739–55, 2007.
J. Torres, J. Marro, J. M. Cortes, and B. Wemmenhove. Instabilities in at-
tractor networks with fast synaptic fluctuations and partial updating of the
neurons activity. Neural Nets., 21:1272–77, 2008.
J. Torres, J. Marro, and S. de Franciscis. Chaos in heterogeneous networks
with temporally inert nodes. Int. J. Bifur. Chaos, 19:677–86, 2009.
M. Tsodyks, K. Pawelzik, and H. Markram. Neural networks with dynamic
synapses. Neural Comp., 10:821–35, 1998.
N. van Kampen. Stochastic Processes in Physics and Chemistry. Elsevier
Science, Amsterdam, 1992.
F. Vazquez, V. Egu´ıluz, and M. S. Miguel. Generic absorbing transition in
coevolution dynamics. Phys. Rev. Lett., 100:108702, 2008.
T. Vogels, K. Ra jan, and L. Abbott. Neural network dynamics. Annu. Rev.
Neurosci., 28:357–76, 2005.
X.-J. Wang. Synaptic reverbaration underlying mnemonic presistent activity.
TRENDS in Neurosci., 24:455–63, 2001.
D. J. Watts and S. H. Strogatz. Collective dynamics of ’small-world’ networks.
Nature, 393:440–442, 1998.
J. White, E. Southgate, J. N. Thomson, and S. Brenner. The structure of the
nervous system of the nematode c. elegans. Philosophical transactions Royal
Society London, 314:1–340, 1986.
J. Wixted and E. Ebbesen. On the form of forgetting. Psychological Science,
2:409–15, 1991.
J. Wixted and E. Ebbesen. Genuine power curves in forgetting: A quantitative
analysis of individual sub ject forgetting functions. Memory & Cognition, 25:
731–9, 1997.
S. R. y Ca jal. Histology of the Nervous System of Man and Vertebrates. Oxford
University Press, Oxford, 1995.
REFERENCES
143
C. Zhou, A. Motter, and J. Kurths. Universality in the synchronization of
weighted random networks. Phys. Rev. Lett., 96:034101, 2006a.
C. Zhou, L. Zemanov´a, G. Zamora, C. Hilgetag, and J. Kurths. Hierarchical
organization unveiled by functional connectivity in complex brain networks.
Phys. Rev. Lett., 97:238103, 2006b.
|
1806.06809 | 2 | 1806 | 2018-07-05T16:48:00 | Drifting perceptual patterns suggest prediction errors fusion rather than hypothesis selection: replicating the rubber-hand illusion on a robot | [
"q-bio.NC"
] | Humans can experience fake body parts as theirs just by simple visuo-tactile synchronous stimulation. This body-illusion is accompanied by a drift in the perception of the real limb towards the fake limb, suggesting an update of body estimation resulting from stimulation. This work compares body limb drifting patterns of human participants, in a rubber hand illusion experiment, with the end-effector estimation displacement of a multisensory robotic arm enabled with predictive processing perception. Results show similar drifting patterns in both human and robot experiments, and they also suggest that the perceptual drift is due to prediction error fusion, rather than hypothesis selection. We present body inference through prediction error minimization as one single process that unites predictive coding and causal inference and that it is responsible for the effects in perception when we are subjected to intermodal sensory perturbations. | q-bio.NC | q-bio | Drifting perceptual patterns suggest prediction
errors fusion rather than hypothesis selection:
replicating the rubber-hand illusion on a robot
Nina-Alisa Hinz†, Pablo Lanillos‡∗, Hermann Mueller†, Gordon Cheng‡
† Experimental Neuro-Cognitive Psychology, Ludwig-Maximilians-Universitat, Munchen, Germany
‡ Institute for Cognitive Systems, Technische Universitat Munchen, Munchen, Germany
∗Email: [email protected]
8
1
0
2
l
u
J
5
]
.
C
N
o
i
b
-
q
[
2
v
9
0
8
6
0
.
6
0
8
1
:
v
i
X
r
a
Abstract-Humans can experience fake body parts as theirs
just by simple visuo-tactile synchronous stimulation. This body-
illusion is accompanied by a drift in the perception of the
real limb towards the fake limb, suggesting an update of body
estimation resulting from stimulation. This work compares body
limb drifting patterns of human participants, in a rubber hand
illusion experiment, with the end-effector estimation displace-
ment of a multisensory robotic arm enabled with predictive
processing perception. Results show similar drifting patterns in
both human and robot experiments, and they also suggest that
the perceptual drift is due to prediction error fusion, rather
than hypothesis selection. We present body inference through
prediction error minimization as one single process that unites
predictive coding and causal inference and that it is responsible
for the effects in perception when we are subjected to intermodal
sensory perturbations.
Index Terms-Sensorimotor self, rubber-hand illusion, predic-
tive coding, robotics
I. INTRODUCTION
Distinguishing between our own body and that of others
is fundamental for our understanding of the self. By learning
the relationship between sensory and motor information and
integrating them into a common percept, we gradually develop
predictors about our body and its interaction with the world
[1]. This body learning is assumed to be one of the major
processes underlying embodiment. Body-ownership illusions,
like the rubber hand illusion [2], are the most widely used
methodology to reveal information about the underlying mech-
anisms, helping us to understand how the sensorimotor self
is computed. Empirical evidence has shown that embodiment
is flexible, adaptable, driven by bottom-up and top-down
modulations and sensitive to illusions.
We replicated the passive rubber hand illusion on a multi-
sensory robot and compared it with human participants, there-
fore gaining insight into the perceptive contribution to self-
computation. Enabling a robot with human-like self-perception
[3] is important for: i) improving the machine adaptability
has
by
been
supported
This work
SELFCEPTION project
(www.selfception.eu) European Union Horizon 2020 Programme (MSCA-
IF-2016) under grant agreement n. 741941 and the ENB Master Program in
Neuro-Cognitive Psychology at Ludwig-Maximilians Universitat. Video to
this paper: http://web.ics.ei.tum.de/~pablo/rubberICDL2018PL.mp4
Accepted for publication at 2018 IEEE International Conference on Devel-
opment and Learning and Epigenetic Robotics
and providing safe human-robot interaction, and ii) testing
computational models for embodiment and obtaining some
clues about the real mechanisms. Although some computa-
tional models have already been proposed for body-ownership,
agency and body-illusions, the majority of them are restricted
to the conceptual level or simplistic simulation [4]. Examining
real robot data and using body illusions as a benchmark
for testing the underlying mechanism enriches the evaluation
considerably. To the best of our knowledge, this is the first
study replicating the rubber hand illusion on an artificial agent
and comparing it to human data.
the robot
We showed that when inferring the robot body location
through prediction error minimization [5],
limb
drifting patterns are similar to those observed in human par-
ticipants. Human and robot data suggest that the perception of
the real hand and the rubber hand location drifts to a common
location between both hands. This supports the idea that,
instead of selecting one of two hypotheses (common cause for
stimulation vs. different causes) [6], visual and proprioceptive
information sources are merged generating an effect similar to
averaging both hypotheses [7].
The remainder of this paper is as follows: Sec. II describes
current rubber-hand illusion findings and its neural basis; Sec.
III defines the computational model designed for the robot;
Sec. IV describes the experimental setup for both human
participants and the robot; Sec. V presents the comparative
analysis of the drifting patterns; Finally, Sec. VI discusses
body estimation within the prediction error paradigm as the
potential cause of the perceptual displacement.
II. BACKGROUND
A. Rubber-hand illusion
Botvinick and Cohen [10] demonstrated that humans can
embody a rubber hand only by means of synchronous visuo-
tactile stimulation of the rubber hand and their hidden real
hand. This was measured using a questionnaire about the
illusion, but also by proprioceptive localization of the partici-
pant's real hand. After experiencing the illusion, the perception
of their own hand's position had drifted towards that of the
rubber hand. Since then, multiple studies have replicated the
illusion under different conditions (for a review, see [11]).
Collectively, these studies show that top-down expectations
found an increased proprioceptive drift for a larger distance,
the relative amount of drift (i.e. corrected for distance) did
not differ between the small and large distances. In [17], they
replicated this result, as long as the fake hand was near the
body. If the real hand was closer to body midline than the
fake hand, increasing distance between both hands decreased
the proprioceptive drift.
In the present study, we systematically varied the distance
between both hands. The real hand, however stayed in the
same position for all conditions and only the fake hand had a
varying distance from the real hand in anatomically plausible
positions. In [8], where the distance between both hands
was varied by displacing the fake hand in relation to the
real hand, the fake hand was also increasingly rotated with
increasing distance. Rotational differences, nevertheless, may
influence the illusion [20], probably confounding the results
of [8]. In the current study, we systematically examined the
influence of the distance between the rubber and the real hand
on proprioceptive and visual drift. This provided the basis
for validating the computational model proposed in Sec. III
and comparing the drift of the body estimation in different
distances between the robot and humans.
B. Body illusions in the brain
A seminal contribution to possible neuronal mechanisms
underlying the rubber hand illusion came from [21], who
discovered parietal neurons in the primate brain coding for
the position of the real arm and a plosturally plausible fake
monkey arm. Several fMRI studies looked into the neural
correlates for body-ownership illusions in humans (see [20] for
a review). Three areas were consistently found activated during
the rubber hand illusion: posterior parietal cortex (including
intra-parietal cortex and temporo-parietal junction), premotor
cortex and lateral cerebellum. The cerebellum is assumed
to compute the temporal relationship between visual and
tactile signals, thus playing a role in the integration of visual,
tactile and proprioceptive body-related signals [22], [23]. The
premotor and intra-parietal cortex are multisensory areas, also
integrating visual, tactile and proprioceptive signals present
during the rubber hand illusion [24]. In [20], they differentiated
the role of posterior parietal cortex, being responsible for the
recalibration of visual and tactile coordinate systems into a
common reference frame, and the role of premotor cortex,
being responsible for the integration of signals in this common,
hand-centered reference frame. Although it is known that these
areas participate in evoking the rubber hand illusion, little is
known about the underlying computations [25]. In [26], they
used dynamic causal modeling during the rubber hand illusion
to confirm to some extent that, during the illusion, visual in-
formation is weighted more than proprioceptive information -
which would be congruent with predictive coding models. Dur-
ing the illusion the intrinsic connectivity of the somatosensory
cortex was reduced, indicative of less somatosensory precision,
while the extrinsic connectivity between the lateral occipital
complex and premotor cortex was increased, indicative of
(a) Illusion depending on distance (b) Proprio. drift depending on time
Fig. 1. (a) Quality of the illusion as a function of the distance between hands:
the higher the better. (b) Intensity of the proprioceptive drift depending on
the visuo-tactile stimulation time. Adapted from [8] and [9] respectively.
about the physical appearance of a human hand, resulting
from abstract internal body models, and bottom-up sensory
information, especially spatiotemporal congruence of the stim-
ulation and distance between the hands, influence embodiment
of the fake hand [9]. In [12], they were even able to induce
body-ownership on a robotic arm. The common assumption
is that the spatial representations of both hands are merged,
as a result of minimizing the error between predicted sensory
outcomes from seeing the stimulation of the rubber hand and
the actual sensory outcome of feeling the stimulation of one's
own hand [13]. Recently, [7] undermined this by showing that
not only the perception of the real hand's position is drifted
towards the rubber hand (proprioceptive drift), but the one
of the rubber hand is drifted towards the real hand as well
(visual drift), i.e. to a common location between both hands.
In [6], they proposed a Bayesian Causal Inference Model for
this prediction error minimization, considering visual, tactile
and proprioceptive information, weighted according to their
precision. In combination with the prior probability of assum-
ing a common cause or different causes for the conflicting
multi-sensory information, the posterior probability of each
hypothesis is computed. A common cause, i.e. ownership of
the rubber hand, is assumed if the posterior probability exceeds
a certain threshold. This binarity of the illusion, however, is
at variance with findings of [9], demonstrating a continuous
proprioceptive drift of the stimulated hand. The proprioceptive
drift was shown to increase exponentially during the first
minute of stimulation and increasing further over the following
four minutes (Fig. 1(b)). Although the reported onset of
the illusion ranged from 11 seconds [14] to 23 seconds,
with 90 percent of subjects experiencing it within the first
minute of stimulation [15], the ongoing drifting suggests a
continuous, rather than a discrete mechanism, being involved
in embodiment.
The proposed computational models in the literature of
body-ownership illusions need further verification from exper-
imental data. Several studies showed reduced illusion scores
for larger, as compared to smaller, distances between the real
and the rubber hand [8], [16]–[19] (Fig. 1(a)), though only
few studies measured the proprioceptive drift in dependence
on the distance between the two hands [16], [17]. While [16]
more attention to visual input. Further functional evidence for
the proposed computations is needed.
III. COMPUTATIONAL MODEL
(cid:88)
i
x =
(si − gi(x))
(cid:124)
(cid:123)(cid:122)
σsi
error expected sensation
(cid:125)
i(x) − sx − µx
(cid:124) (cid:123)(cid:122) (cid:125)
g(cid:48)
σx
error prior
(3)
We assume that all sensations / features follow a Gaussian dis-
tribution with (linear or non-linear) mean gi(x) and variance
σsi. The forward models learned should be differentiable with
respect to the body configuration (g(cid:48)
By rewriting the prediction error as e = s − g(x) and
defining µx as the prior belief about the body configuration,
the dynamics of the body perception model are described
by (see Appendix for derivation and [27] for the detailed
algorithm):
i(x) = ∂gi(x)/∂x).
v(x) + evtg(cid:48)
vt(x)
x = −ex + ep + evg(cid:48)
ex = sx − µx − σxex
ep = sp − x − σpep
ev = sv − gv(x) − σvev
evt = svt − gvt(x) − σvtevt
µx = µx + λex
(4)
where λ is the learning ratio parameter that specifies how
fast the prior of the body configuration µx is adjusted to the
prediction error.
The visual forward function gv and its derivative are cal-
culated using Gaussian process estimation (see Sec. IV). The
visuo-tactile generative function is computed by means of a
hand-crafted likelihood, which uses the visual ov and tactile
st stimulation information (temporal hs and spatial ht), and
the expected position of the hand gv(x):
gt(x) = st · hs · ht = sta1e−b1
(cid:80)(gv(x)−ov)2 · a2e−b2δ2
(5)
where a1, b1, a2, b2 are parameters that shape the likelihood
of the spatial plausibility and have been tuned in accordance
with the data acquired from human participants; δ is the level
of synchrony of the events (e.g. timing difference between the
visual and the tactile event); and ov is the other agent end-
effector location in the visual field.
IV. EXPERIMENTAL SET-UP
Fig. 2. Rubber hand illusion modelled as a body estimation problem solved
using prediction error minimization. Visual features of the rubber hand are
incorporated when there is synchronous visuo-tactile stimulation, though it is
constrained by the prior belief and the expected location of the hand according
to the generative visual forward model and the estimated body configuration.
We formalized the rubber hand illusion as a body estimation
problem under the predictive processing framework. The core
idea behind this is that all features and sensory modalities
are contributing to refine body estimation through the min-
imization of the errors between sensations and predictions
[27]. During synchronous visuo-tactile stimulation, the most
plausible body configuration is perturbed due to the merging of
visual and proprioceptive information. This is coherent with
the drift of both the real hand and the rubber hand as the
participants are just pointing to the prediction of their hand
according to the current body configuration. Figure 2 shows
how sensory modalities or features are contributing to the
estimation of the participants limb.
We define x as the latent space variable that expresses
the body configuration. We model the problem as inferring
the most plausible body configuration x given the sensation
likelihood and the prior: P (xs) = p(sx)p(x). We further
define sp, sv, svt as the proprioceptive, visual and visuo-tactile
sensation respectively. Assuming independence of the different
sources of information we get:
P (xs) = p(spx)p(svx)p(svtx)p(x)
(1)
A. Participants selection
The perception or estimation of the body is then solved
by learning an approximation of the forward model for each
feature or modality s = g(x) and minimizing a lower bound
on the KL-divergence known as negative free energy F [5],
[28].
log P (xs) = −F =
log p(six) + log p(x)
(2)
(cid:88)
i
We obtain the estimated value of the latent variables through
gradient descent minimization x = ∂F
x :
20 volunteers (mean age: 25, 75 % female) took part in the
experiment. They received 8 euros per hour in compensation.
All participants were right-handed, had no disability of per-
ceiving touch on their right hand, did not wear nail polish and
did not have any special visual features (e.g. scars / tattoos)
on their right hand. They had no neurological or psychological
disorders, as indicated by self-report, and normal or corrected-
to-normal vision. None of them had experienced the rubber
hand illusion before. All participants gave informed consent
prior to the experiment.
(a) Human
(b) Robot
Fig. 3. Experimental setups used in the current study. (a) Rubber-hand illusion in different distance conditions and (b) artificial version.
B. Humans experiment details
We performed the rubber-hand illusion experiment, focusing
on the proprioceptive drift of the real and the visual drift of the
rubber hand, as a function of the distance between both hands.
The experiment, depicted in Fig. 3, comprised six conditions,
each with a different distance between the real hand and the
rubber hand. The participant's real right hand was placed in
a box, with the index finger 20cm away from the participants
body midline. The rubber hand was placed with its index
finger 15cm, 20cm, 25cm, 30cm, 35cm or 40cm away from the
participants real right hand (5cm, 0cm, -5cm, -10cm, -15cm or
-20cm away from the participants body midline respectively).
Participants sat in front of a wooden box, placing their hands
near the outer sides of the box. They wore a rain cape covering
their body and arms. In one of the arms, a rubber hand was
placed such that it seemed coherent with the body. With their
left hand they held a computer mouse. Each trial consisted
of four phases: localization of the real hand, localization of
the rubber hand, the stimulation phase and post-stimulation
localization of both hands.
1) First, we covered the box with a wooden top and
a blanket above it, so that no visual cues could be
used. Participants had to indicate where they currently
perceived the location of the index finger of their right
hand, pointing with the mouse on a horizontal
line
presented on the screen. The line did cover the whole
length of the box.
2) After fixating the rubber hand for one minute, we again
covered the box and the same task was repeated for the
rubber hand.
3) The box was remodeled, removing the cover and intro-
ducing a vertical board next to the participant's right
hand so that it was not visible to the participant (Fig.
3(a)). Then the experimenter began stimulating the rub-
ber and the real hand synchronously with two similar
paintbrushes, starting at the index finger, continuing to
the little finger and then starting at the index finger again,
with one brush stroke each about two seconds.
4) Subsequently, participants were again asked to indicate
where they perceived the index finger of the real or the
rubber hand, starting with the real or the rubber hand
in randomized fashion. The box was covered during the
localization.
5) At the end of each trial, participants were asked to
answer ten questions related to the illusion adapted from
[29], presented randomized on the screen, using a con-
tinuous scale from -100 (indicating strong disagreement)
to 100 (indicating strong agreement).
For the localization trials, a horizontal line was presented on
the screen opposite to the box, with the screen having the
same size as the box. The localization trials were repeated ten
times to account for high variance. The proprioceptive drift
and visual drift were calculated by subtracting the average of
the first localization phase from the second localization phase
for the real hand and the rubber hand separately. The illusion
index was calculated by subtracting the average response
to control statements S4-S10 from the average response to
illusion statements S1-S3 [30]. Between all phases participants
were blindfolded, so they did not observe the remodeling,
which might potentially have impeded the illusion.
C. Robot experiment details
We tested the model on the multisensory UR-5 arm of
robot TOMM [31], as depicted in Fig. 3(b). The proprioceptive
input data were three joint angles with added noise (shoulder1,
shoulder2 and elbow - Fig. 4(a)). The visual input was an rgb
camera mounted on the head of the robot, with 640×480 pixels
definition. The tactile input was generated by multimodal skin
cells distributed over the arm [32].
D. Learning g(x) from visual and proprioceptive data
In order to learn the sensory forward model, we applied
Gaussian process (GP) regression: gv(x) ∼ GP. We pro-
grammed random trajectories in the joint space that resembled
horizontal displacements of the arm. Figure 4(a) shows the
extracted data: noisy joint angles and visual location of the
end-effector, obtained by color segmentation. To learn the
visual forward model sv = gv(x), each sample was defined as
the input joint angles sensor values x = (x1, x2, x3) and the
output sv = (i, j) pixel coordinates. As an example, Figure
4(b) shows the learned visual forward model by GP regression
with 46 samples (red dots). It describes the horizontal mean
and variance (in pixels) with respect to two joints angles. The
GP learning and its partial derivative computation with respect
to x is described in the Appendix B.
(a) Data recorded example (joints angles + noise, end-effector visual,
end-effector cartesian) and schematic picture of the 3-DOF.
(b) Learnt gv(x) for visual horizontal location
(c) Tactile (left) and visual (right) event trajectories
Fig. 4. Collected data. (a) Joints angles, visual and ground truth information of
the end-effector. (b) Mean and the variance computed by the GP for the visual
horizontal location depending on two joints. (c) Touch patterns extracted from
tactile (117 forearm skin cells) and visual sources.
E. Extracting visuo-tactile data
We used proximity sensing information (infrared sensors)
from 117 different skin cells to discern when the arm was
being touched. The sensor outputed a filtered signal ∈ (0, 1).
From the other's hand visual trajectory and the skin proximity
activation, we computed the level of synchrony between the
two patterns (Fig. 4(c)). Timings for tactile stimuli st were
obtained by setting a threshold over the proximity value: prox
> 0.7 → activation. Timings for the other's trajectory events
were obtained through the velocity components. Detected
initial and end positions of the visual touching are depicted in
Fig. 4(c) (right, green circles).
V. RESULTS
We compared the drifting data extracted from the rubber-
hand illusion experiment in human participants and the robot.
In order to obtain the robot results, we fixed in advance the
model parameters for the learning and the body estimation
stages. gv(x) ∼ GP learning hyperparameters: signal variance
σn = exp(0.02) and kernel length scale l = exp(0.1). The
integration step was ∆t = 0.05 (20Hz) and the error variances
were σx ∈ R3 = [1, 1, 1], σp ∈ R3 = [1, 1, 1], σvt ∈ Z 2 =
[2, 2]. The adaptability rate of µx was λ = 1. The visual feature
from the real hand sv was not used in the rubber hand illusion
experiment as the arm was covered. Finally, the visuo-tactile
, where σt =
function (Eq. 5) parameters were: b1 = σt
d2
0.001 and dmax = 0.0016; b2 = 25; and a1 = a2 = 1.
The robot drift was computed by subtracting the estimated
end-effector position gv(x) and the ground truth location, and
x was dynamically updated minimizing the prediction error
using the proposed model.
A. Comparative analysis
max
(a) Drift
(b) Relative drift
Fig. 5. Human proprioceptive drift vs end-effector robot estimation drift
after the rubber hand illusion experiment. (a) Drift in cm. Positive values
express displacement towards the fake hand. (b) Relative drift depending on
the distance between fake and real arm.
Figure 5 shows the proprioceptive drift comparison. Fig.
5(a) shows similar drifting patterns in both the robot and the
human participants. A drift towards the fake hand emerges in
both cases when the distance is small and then vanishes with
longer distances. The prior information used for the tactile
likelihood function parameters is modulated when the effect
is taking place, as the error will start propagating when the
gradient of the function is noticeable. Furthermore, the relative
drift (Fig. 5(b)) also showed that, for close distances, the
amount of displacement is the same, and then it decreases
until vanishing. The robot was tested on even closer distances
than humans, since the human experimental setup was not
equipped for distances beneath 15cm. The large increase in
proprioceptive drift for 10cm distance between fake and real
hand is an interesting prediction for human data, that could be
tested in future work.
B. Human data analysis
Data exceeding a range of two standard deviations around
the mean was excluded from further analysis. T-tests were
used to test the proprioceptive drift, the visual drift and the
illusion score in each condition against zero. In the first three
conditions the proprioceptive drift was significantly different
from zero, while it was not in the other three conditions
(Table I). Employing Bonferroni-Correction only leaves a
trend towards significance in the 20cm distance condition.
However,
three conditions is still
the average of the first
significantly different from zero M : 13.72, SD : 14.84,
p < .001) while the average of the other conditions is not
(M 8.24 :, SD : 19.79, p > .05). The visual drift was
only significantly different from zero in the 30cm distance
condition (M : 14.01, SD : 29.72, p < .01). The illusion
score was significantly different from zero in all conditions
(all p < .05). Partial Pearson correlations between illusion
score and proprioceptive drift, illusion score and visual drift
and between proprioceptive drift and visual drift were not
significant (all p > .05).
top plot, also shows how smooth body configuration output
µx1:3 is (blue line). The robot inferred the most plausible
body joints angles given the sensory information, which in this
case yielded a horizontal drift on the estimated end-effector
location. A video of the evolution of the variables during
the artificial rubber-hand illusion experiment can be found at
http://web.ics.ei.tum.de/∼pablo/rubberICDL2018PL.mp4.
DESCRIPTIVE AND INFERENCE STATISTICS FROM PROPRIOCEPTIVE DRIFT
TABLE I
DATA IN EACH CONDITION.
Condition
15 cm
20 cm
25 cm
30 cm
35 cm
40 cm
std
mean
df
14.89 mm 27.08 mm 19
12.79 mm 18.99 mm 17
13.22 mm 23.51 mm 16
15.52 mm 16
4.39 mm
18.00 mm 18
7.84 mm
5.45 mm
31.34 mm 17
t-value
2.46
2.86
2.32
1.17
1.90
0.74
p-value
.024
.011
.034
.260
.074
.471
(a) End-effector drift depending on the sensing information (propio.,
visuo-tactile, and proprio+visuo-tactile)
(a) propriceptive drift
(b) illusion score
Fig. 6. Mean and standard deviation of (a) the visual and the proprioceptive
drift, and (b) the illusion score depending on the distance.
C. Robot model analysis
We analyzed the internal variables of the proposed model
during the visuo-tactile stimulation and the induced end-
effector estimation drift towards the fake arm. Figure 7(a)
shows the robot camera view with the final end-effector
estimation overlaid after 12 seconds. Depending on the dif-
ferent enabled modalities (proprioceptive, visuo-tactile and
proprioceptive+visuo-tactile), body estimation evolved differ-
entially, accordingly the prediction of the end-effector gv(x).
Fig. 7(b) shows the evolution of the body configuration in
term of joint angles and the corresponding prediction errors.
We did initialize the robot belief in a wrong body configuration
to further show the adaptability of the model. During the
first five seconds,
the system converged to the real body
configuration. Afterwards, when perturbing with synchronous
visuo-tactile stimulation, a bias appeared on the body joints.
This implies a drift of the robot end-effector towards the
location of the fake arm visual feature. Tactile perturbations
are shown as prediction error bumps (yellow line). Fig. 7(b),
(b) Evolution of body estimation and predictive errors
Fig. 7. Replicating the rubber-hand illusion on a robot. (a) End-effector
estimation depending on the modality used. (b top) Joints angles in radians:
real (black dotted line), estimated µx (blue dotted line) and current belief
x (red line). (b middle) Errors between expected and observed values. (b
bottom) g(x) values evolution during the experiment. In the first five seconds,
in which there is no tactile stimulation, the estimation is refined. Next, we
inject tactile stimulation while the experimenter is touching a fake arm. When
visuo-tactile stimulation becomes synchronous, a horizontal drift appears on
the end-effector estimation.
VI. DISCUSSION: BODY ESTIMATION AS AN EXPLANATION
FOR THE PERCEPTUAL DRIFT
It has been shown that during the rubber hand illusion, the
location of the real hand is perceived to be closer to the rubber
hand than before. Similarly, the location of the rubber hand is
mislocalized towards the real hand [7]. Our results from the
robot and humans support the former finding: in our predictive
coding scheme, the representation of both hands merged into
a common location between both hands, due to inferring
one's body's location from minimizing free energy. This body
estimation generated a drift of the perceived location of the
real hand towards the equilibrium location, which was visible
in the data from both the humans and the robot. In comparative
analysis, the patterns of the drift resembled each other, both in
terms of absolute and relative values. The three closest tested
distances showed a substantial proprioceptive drift. All other
distances showed a smaller drift, approaching zero. For these,
the distance between the fake and the real hand was probably
too large for the fake hand to be fully embodied, supporting
[6] simulation data exhibiting a reduced illusion probability
for distances over 30cm. Although previous and the present
research support predictive coding as a probable underlying
mechanism of the rubber hand illusion, other accounts can
not be ruled out by the present work.
Human illusion score data, however, did not mirror the
proprioceptive shift pattern found. For all distances, illusion
scores ranged between 20 and 35, which on our continuous
scale up to 100 resembles illusion scores previously found
from 1 on a discrete scale to 3, e.g. [8]. Given this, we
can assume that we were able to induce the illusion in
every condition. Illusion scores and the proprioceptive drift,
additionally, were not correlated. This supports the current
debate that body-ownership illusions and the drift are two
different, but related, processes [30]. The proprioceptive drift
is an unconscious process - in contrast to the illusion, which
is consciously accessible to the subjects. Hence, it might be
possible that the predictive coding formulation in its uncon-
scious form can explain drifting patterns, while it is not as
such sufficient to explain body-ownership illusions.
In contrast
to [7], we did not find a conclusive visual
drift in the human experiment. From participants' personal
communication, we know that many used visual landmarks to
estimate the position of the rubber hand, but of the real hand
also. Differences in this strategy for localization would not
only account for the high variance we observed in the drift
data, but also for the small magnitude of the proprioceptive
drift - as compared to the values reported in other studies
(e.g. [16]). Beyond that, the method we used for localization
is probably prone to high variance due to small mouse move-
ments. Although we tried to account for that by repeating
the localization ten times, more trials might be needed as
performed in [6]. Arguably, however, the lack of visual drift
in our study does not contradict the predictive coding scheme.
Actually, some of our participants communicated that they
experienced that
the representation of both hands merged
together. This is supported by the positive mean (14.16) of
the actual control statement S10 "It felt as if the rubber hand
and my own right hand lay closer to each other", which
was even higher than the mean response (-4.95) to the actual
illusion statement S3 "I felt as if the rubber hand were my
hand". Further investigations, accounting for the variance in
localization, are needed to support this conjecture.
The computational model presented here also generates
predictions about the temporal dynamics of the rubber hand
illusion. The constant accumulation of information resulting
in an also accumulated drift of the body estimation (see 7(a))
is comparable to findings from [9] (see 1), who also found an
accumulation of the drift over time in humans. To provide a
finer temporal comparison, the dynamics of the human illusion
should be further investigated.
VII. CONCLUSION
We implemented the rubber hand illusion experiment on a
multisensory robot. The perception of the real hand's position
drifted towards the rubber hand, following a similar pattern
in humans and the robot. We suggest that this proprioceptive
drift resulted from a merging body estimation between both
hands. This supports an account of the proprioceptive drift
underlying body-ownership illusions in terms of the predictive
coding scheme. Future work will address the mechanisms
behind awareness of body-ownership.
APPENDIX A
FREE ENERGY GRADIENT
p(xsp, sv) = p(spx)p(svx)p(svtx)p(x)
(6)
Applying logarithms we get the negative free energy formu-
lation:
F = ln p(spx) + ln p(svx) + ln p(svtx) + ln p(x)
(7)
Substituting the probability distributions by their functions
f (.; .), and under the Laplace approximation [33] and assum-
ing normally distributed noise, we can compute the negative
free energy as:
F = ln f (sp; gp(x), σp) + ln f (sp; gp(x), σp) + ln f (x; µx, σx)
= − (x − µx)2
− (sp − gp(x))2
2σx
+
− (sv − gv(x))2
(cid:2)− ln σx − ln σsp − ln σsv − ln σsvt
− (svt − gvt(x))2
(cid:3) + C.
2σvt
2σp
2σv
(8)
+
1
2
In order to find x in a gradient-descent scheme we minimize
Eq. 8 through the following differential equation:
x = − x − µx
σx
+
sp − gp(x)
+
σp
(cid:48)
g
p(x) +
sv − gv(x)
σv
(cid:48)
g
v(x) +
svt − gvt(x)
σvt
(cid:48)
vt(x)
g
(9)
In the case that x is equivalent to gp(x) like using the joint
angles values directly as the body configuration,
then the
proprioceptive error can be rewritten as: sp−x and the gradient
becomes 1. Generalizing for i sensors we finally have:
si − gi(x)
(cid:88)
∂gi(x)T
= − (x − µx)
(10)
+
σx
∂ x
σi
i
∂F
∂ x
APPENDIX B
GP REGRESSION
Given sensor samples s from the robot in several body
configurations x and the covariance function k(xi, xj), the
training is performed by computing the covariance matrix
K(X, X) on the collected data with noise σ2
n:
∀i, j ∈ x
nI + k(xi, xj)
kij = σ2
(11)
The prediction of the sensory outcome s given x is then
computed as [34]:
g(x) = k(x, X)K(X, X)−1s = k(x, X)α
(12)
where α = choleski(K)T\(choleski(K)\s).
Finally, in order to compute the gradient of the posterior
g(x)(cid:48) we differentiate the kernel [35], and obtain its prediction
analogously as Eq. 12:
g(x)(cid:48) =
∂k(x, X)
∂ x
K(X, X)−1s =
∂k(x, X)
∂ x
α
(13)
Using the squared exponential kernel with the Mahalanobis
distance covariance function, the derivative becomes:
g(x)(cid:48) = −Λ−1(x − X)T (k(x, X)T · α)
(14)
where Λ is a matrix where the diagonal is populated with the
length scale for each dimension (diag(1/l2)) and · is element-
wise multiplication.
REFERENCES
[1] P. Lanillos, E. Dean-Leon, and G. Cheng, "Enactive self: a study
of engineering perspectives to obtain the sensorimotor self through
enaction," in IEEE Int. Conf. on Developmental Learning and Epigenetic
Robotics, 2017.
[2] M. Botvinick and J. Cohen, "Rubber hands feeltouch that eyes see,"
Nature, vol. 391, no. 6669, p. 756, 1998.
[3] P. Lanillos, E. Dean-Leon, and G. Cheng, "Yielding self-perception in
robots through sensorimotor contingencies," IEEE Trans. on Cognitive
and Developmental Systems, no. 99, pp. 1–1, 2016.
[4] K. Kilteni, A. Maselli, K. P. Kording, and M. Slater, "Over my fake
body: body ownership illusions for studying the multisensory basis of
own-body perception," Frontiers in human neuroscience, vol. 9, p. 141,
2015.
[5] K. Friston, "A theory of cortical responses," Philosophical Transactions
of the Royal Society of London B: Biological Sciences, vol. 360, no.
1456, pp. 815–836, 2005.
[6] M. Samad, A. J. Chung, and L. Shams, "Perception of body ownership
is driven by bayesian sensory inference," PloS one, vol. 10, no. 2, p.
e0117178, 2015.
[7] R. Erro, A. Marotta, M. Tinazzi, E. Frera, and M. Fiorio, "Judging the
position of the artificial hand induces a visual drift towards the real one
during the rubber hand illusion," Scientific reports, vol. 8, no. 1, p. 2531,
2018.
[8] D. M. Lloyd, "Spatial limits on referred touch to an alien limb may
reflect boundaries of visuo-tactile peripersonal space surrounding the
hand," Brain and cognition, vol. 64, no. 1, pp. 104–109, 2007.
[9] M. Tsakiris and P. Haggard, "The rubber hand illusion revisited:
Visuotactile integration and self-attribution," Journal of experimental
psychology. Human perception and performance, vol. 31, no. 1, pp.
80–91, 2005.
[10] M. Botvinick and J. Cohen, "Rubber hands 'feel' touch that eyes see,"
Nature, vol. 391, no. 6669, p. 756, 1998.
[11] K. Kilteni, A. Maselli, K. P. Kording, and M. Slater, "Over my fake
body: Body ownership illusions for studying the multisensory basis of
own-body perception," Frontiers in human neuroscience, vol. 9, 2015.
[12] L. Aymerich-Franch, D. Petit, G. Ganesh, and A. Kheddar, "Non-
human looking robot arms induce illusion of embodiment," International
Journal of Social Robotics, vol. 9, no. 4, pp. 479–490, 2017.
[13] M. Tsakiris, "My body in the brain: A neurocognitive model of body-
ownership," Neuropsychologia, vol. 48, no. 3, pp. 703–712, 2010.
[14] H. H. Ehrsson, C. Spence, and R. E. Passingham, "That's my hand!
activity in premotor cortex reflects feeling of ownership of a limb,"
Science, vol. 305, no. 5685, pp. 875–877, 2004.
[15] A. Kalckert and H. H. Ehrsson, "The onset time of the ownership
sensation in the moving rubber hand illusion," Frontiers in psychology,
vol. 8, p. 344, 2017.
[16] R. Zopf, G. Savage, and M. A. Williams, "Crossmodal congruency
measures of lateral distance effects on the rubber hand illusion," Neu-
ropsychologia, vol. 48, no. 3, pp. 713–725, 2010.
[17] C. Preston, "The role of distance from the body and distance from
the real hand in ownership and disownership during the rubber hand
illusion," Acta psychologica, vol. 142, no. 2, pp. 177–183, 2013.
[18] S. C. Pritchard, R. Zopf, V. Polito, D. M. Kaplan, and M. A. Williams,
"Non-hierarchical influence of visual form, touch, and position cues
on embodiment, agency, and presence in virtual reality," Frontiers in
psychology, vol. 7, p. 1649, 2016.
[19] N. Ratcliffe and R. Newport, "The effect of visual, spatial and temporal
manipulations on embodiment and action," Frontiers in human neuro-
science, vol. 11, p. 227, 2017.
[20] T. R. Makin, N. P. Holmes, and H. H. Ehrsson, "On the other hand:
Dummy hands and peripersonal space," Behavioural brain research, vol.
191, no. 1, pp. 1–10, 2008.
[21] M. S. Graziano, D. F. Cooke, and C. S. Taylor, "Coding the location of
the arm by sight," Science, vol. 290, no. 5497, pp. 1782–1786, 2000.
[22] H. H. Ehrsson, N. P. Holmes, and R. E. Passingham, "Touching a
rubber hand: Feeling of body ownership is associated with activity in
multisensory brain areas," Journal of neuroscience, vol. 25, no. 45, pp.
10 564–10 573, 2005.
[23] A. Guterstam, G. Gentile, and H. H. Ehrsson, "The invisible hand
illusion: Multisensory integration leads to the embodiment of a discrete
volume of empty space," Journal of cognitive neuroscience, vol. 25,
no. 7, pp. 1078–1099, 2013.
[24] A. Guterstam, H. Zeberg, V. M. Ozc¸iftci, and H. H. Ehrsson, "The mag-
netic touch illusion: A perceptual correlate of visuo-tactile integration
in peripersonal space," Cognition, vol. 155, pp. 44–56, 2016.
[25] M. A. J. Apps and M. Tsakiris, "The free-energy self: A predictive
coding account of self-recognition," Neuroscience and biobehavioral
reviews, vol. 41, pp. 85–97, 2014.
[26] D. Zeller, K. J. Friston, and J. Classen, "Dynamic causal modeling of
touch-evoked potentials in the rubber hand illusion," NeuroImage, vol.
138, pp. 266–273, 2016.
[27] P. Lanillos and G. Cheng, "Adaptive robot body learning and estimation
through predictive coding," arXiv preprint arXiv:1805.03104, 2018.
[28] R. Bogacz, "A tutorial on the free-energy framework for modelling
perception and learning," Journal of mathematical psychology, 2015.
[29] M. P. M. Kammers, F. de Vignemont, L. Verhagen, and H. C. Dijkerman,
"The rubber hand illusion in action," Neuropsychologia, vol. 47, no. 1,
pp. 204–211, 2009.
[30] Z. Abdulkarim and H. H. Ehrsson, "No causal link between changes in
hand position sense and feeling of limb ownership in the rubber hand
illusion," Attention, perception & psychophysics, vol. 78, no. 2, pp. 707–
720, 2016.
[31] E. Dean-Leon, B. Pierce, F. Bergner, P. Mittendorfer, K. Ramirez-
Amaro, W. Burger, and G. Cheng, "Tomm: Tactile omnidirectional
mobile manipulator," in Robotics and Automation (ICRA), IEEE Int.
Conf. on, 2017, pp. 2441–2447.
[32] P. Mittendorfer and G. Cheng, "Humanoid multimodal tactile-sensing
modules," IEEE Trans. on robotics, vol. 27, no. 3, pp. 401–410, 2011.
[33] K. Friston, "Hierarchical models in the brain," PLoS computational
biology, vol. 4, no. 11, p. e1000211, 2008.
[34] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine
The MIT
Learning (Adaptive Computation and Machine Learning).
Press, 2005.
[35] A. McHutchon, "Differentiating gaussian processes," 2013.
|
1803.05840 | 1 | 1803 | 2018-03-15T16:25:47 | Effective Connectivity from Single Trial fMRI Data by Sampling Biologically Plausible Models | [
"q-bio.NC",
"physics.data-an"
] | The estimation of causal network architectures in the brain is fundamental for understanding cognitive information processes. However, access to the dynamic processes underlying cognition is limited to indirect measurements of the hidden neuronal activity, for instance through fMRI data. Thus, estimating the network structure of the underlying process is challenging. In this article, we embed an adaptive importance sampler called Adaptive Path Integral Smoother (APIS) into the Expectation-Maximization algorithm to obtain point estimates of causal connectivity. We demonstrate on synthetic data that this procedure finds not only the correct network structure but also the direction of effective connections from random initializations of the connectivity matrix. In addition--motivated by contradictory claims in the literature--we examine the effect of the neuronal timescale on the sensitivity of the BOLD signal to changes in the connectivity and on the maximum likelihood solutions of the connectivity. We conclude with two warnings: First, the connectivity estimates under the assumption of slow dynamics can be extremely biased if the data was generated by fast neuronal processes. Second, the faster the time scale, the less sensitive the BOLD signal is to changes in the incoming connections to a node. Hence, connectivity estimation using realistic neural dynamics timescale requires extremely high-quality data and seems infeasible in many practical data sets. | q-bio.NC | q-bio |
Effective Connectivity from Single Trial fMRI Data by Sampling
Biologically Plausible Models
H.C. Ruiz Euler, H.J. Kappen
August 19, 2020
Abstract
The estimation of causal network architectures in the brain is fundamental for understanding cognitive
information processes. However, access to the dynamic processes underlying cognition is limited to indirect
measurements of the hidden neuronal activity, for instance through fMRI data. Thus, estimating the network
structure of the underlying process is challenging. In this article, we embed an adaptive importance sampler
called Adaptive Path Integral Smoother (APIS) into the Expectation-Maximization algorithm to obtain point
estimates of causal connectivity. We demonstrate on synthetic data that this procedure finds not only the correct
network structure, but also the direction of effective connections from random initializations of the connectivity
matrix. In addition -- motivated by contradictory claims in the literature -- we examine the effect of the neuronal
time scale on the sensitivity of the BOLD signal to changes in the connectivity and on the maximum likelihood
solutions of the connectivity. We conclude with two warnings: First, the connectivity estimates under the
assumption of slow dynamics can be extremely biased if the data was generated by fast neuronal processes.
Second, the faster the time scale, the less sensitive the BOLD signal is to changes in the incoming connections
to a node. Hence, connectivity estimation using realistic neural dynamics time scale requires extremely high
quality data and seems infeasible in many practical data sets.
Introduction
1
In recent years, the field of neuroimaging has seen a rapidly increasing interest in effective connectivity estimations
from functional magnetic resonance imaging (fMRI) data. Although this data acquisition method is very powerful
to investigate human brain function by identifying brain regions that are active during perceptual or cognitive
tasks, fMRI time series are an indirect, delayed and blurred measurement of the actual signal of interest, the latent
neuronal activation of a specific region of interest (ROI). Thus, estimating the underlying connectivity is challenging.
Roughly speaking there are two approaches to this problem based on the distinction between functional and
effective connectivity. The first approach seeks to estimate the temporal correlations between separable ROIs [7].
Examples of these, which are known to correctly estimate the network structure, are partial correlation, regularized
inverse covariance and some Bayes net methods [28].
These symmetric measures, however, have no information about the directionality of the connection, i.e. the
influence that one node exerts over another. This "effective" connectivity is in general harder to estimate but
it is often of great interest. We can distinguish between two families of approaches; purely data-driven methods
that attempt to infer directionality directly from the time series using statistical measures and methods based on
dynamic models that seek to find a forward model to fit the data.
Purely data-driven methods fall generally into three classes. First, lag-based methods, e.g. Granger causality
[13, 12, 5], are variations of the well-known auto-regressive models. In this framework one considers the similarity
between a pair of time series, one of which is shifted in time. If the lagged time series helps predict the zero-lagged
time series, then a causal relation is inferred. A second family of methods is based on the concept of conditional
independence and follow the ideas from structural learning in Bayes networks [14, 22]. The third class considers
higher order statistics to infer causality. For instance, Patel's τ approach measures the asymmetry between the
conditional probabilities of functionally connected nodes. If the activation probability of a node 1 given node 2 is
larger than the activation probability of node 2 given node 1, this is interpreted as a directed connection 2 → 1
[20]. Although many of these methods are widely used, the estimation of directed connectivity proves difficult with
these methods [28].
Dynamic model methods frame the inference problem as a state-space model (SSM). The latent states are the
dynamic degrees of freedom of the process underlying the data. Hence, in general, estimating the parameters of the
1
model requires also the estimation of the hidden process, which is in most cases analytically infeasible and some
approximations are needed. Once the latent process estimation is addressed, one typically uses the Expectation-
Maximization (EM) algorithm in either its Maximum-Likelihood (ML) or variational Bayes versions [26, 25].
For fMRI data, model-based methods usually distinguish themselves in the approximations they make of the
dynamic system or the latent process estimation (E-step). For instance, in [25] the fMRI time series is modeled as
a linear convolution of the bi-linear latent system with the canonical hemodynamic response function (HRF) and
its time derivative [21]. This approximation allows using the Kalman smoother for the E-step.
On the contrary, dynamic causal modeling (DCM) uses a biologically plausible nonlinear model for the hidden
process. Here, the neuronal activity is given as a bi-linear system coupled to the Balloon model describing the
nonlinear relation between blood flow and volume [2, 18, 10]. In addition, the fMRI observations are modeled as
a nonlinear function of the hidden variables. Hence, to deal with the nonlinearities of the biological model, DCM
resorts to either a Volterra expansion of the dynamic system or the generalized filtering method [8, 6, 9].
In addition, DCM uses a Gaussian approximation of the posterior over the parameters and its mean and
covariance are estimated using EM update rules. For this, the adjacency matrix is specified beforehand, effectively
imposing a strong prior on the connectivity of the system. Using the log-evidence, the proposed models are scored
to find the best model amongst them [9]. Although DCM has been widely used, concerns about this approach have
spurred a discussion about its feasibility and validity due to the combinatorial explosion of the network structure and
the apparent inability of the scoring procedure to distinguish between generally accepted networks and randomly
generated ones [17, 1, 4].
In this article, we present an alternative method for connectivity estimation. Our approach is similar to DCM in
the sense that it uses the same model but differs in two important ways. First of all, in the E-step the full posterior
over the latent process is estimated using an optimal control approach that was first introduced in [23]. Second, in
the M-step the connectivity is optimized without prior assumptions as in [9, 8]. The Monte Carlo estimates involved
in these steps are the only approximations required and it is proven to work for nonlinear deconvolution of fMRI
time series [24].
We show using synthetic data that the connectivity estimates obtained are close to the ground truth and that
adding a small L1-regularization on the connections is beneficial to obtain sparse estimates that generalize better
on unseen time series. Furthermore, the proposed method obtains estimates from single event fMRI time series that
are robust against random initializations of the connectivity matrix.
In addition, we study the sensitivity of the BOLD signal depending on the neuronal time scale and the effect of
this parameter on the connectivity estimates. This analysis is motivated by the different claims about the value of
the neuronal time scale in the literature. For instance, [28] used a time constant resulting in a mean neural lag of
50 ms and it is argued that this value is towards the upper limit of observed lags in general. Contrary to this claim
is the assumption made in the DCM literature, where it is argued that the scale ought to be of order 1 s [8, 9].
We observe a significant effect of the neuronal time scale on the connectivity estimates and a remarkable lower
sensitivity of the BOLD signal to changes in the connectivity for fast neuronal dynamics. Hence if the underlying
neuronal processes are fast, the quality of the data must be significantly higher to obtain reliable, unbiased estimates
of the effective connectivity.
2 Method
2.1 Modeling fMRI Data
Similar to DCMs [8], we consider the fMRI forward model as a network of m regions z = (z1, z2, . . . , zm) that follow
stochastic dynamics given by
dzt = A(Czt + u(zt, t) + BIt)dt +
(1)
where dWt ∼ N (0, dt) is a m-dimensional Wiener process with variance1 dt = 0.01 and σz > 0. The parameter A
sets the time scale of the neuronal response, while the connectivity matrix C has diagonal with −1 and off-diagonal
elements Cij ≤ 1, i (cid:54)= j. This form assumes that the fastest scale is the within-node temporal decay. The term
√
A in the diffusion ensures that the stationary distribution remains invariant under changes in the time scale A.
Contrary to DCMs, the input strength has been rescaled such that the stationary point of the system with non-zero
constant input is independent of A. Hence, the inverse time scale A determines only how fast the neuronal system
follows the input.
AσzdWt
√
1In all our simulations we used this discretization step unless stated otherwise.
2
Parameter Value Parameter
A
τs
τf
τ0
1 Hz
0.8
1.54
2.44
1.02
σz
E0
k1
k2
k3
Value
10−3
0.4
7E0
2
2E0 − 0.2
Parameter Value Parameter Value
√
0
σz/
2
TBD
TBD
0.002
2.5
0.32
0.018
B
α−1
V0
µs,0
µz,0
σz,0
σs,0
σf,q,v,0
µf,q,v,0
0
1
σy
Table 1: Parameters for the neural dynamics (top row) and for the BOLD transformation (bottom rows).
Notice that we define the process in equation (1) to have an unknown function u(z, t). This function is the
importance sampling controller learned with APIS. In general, u(z, t) can be any parametrized function [?], but in
this paper it is chosen to have the simple form u(z, t) = az + bt with a a constant and bt a time varying function.
We assume that the external input It and its strength B are known.
Each node's activity zi is coupled to a nonlinear deterministic system modeling the hemodynamic transformation.
For each node i there are two Hemodynamic equations [10]
(cid:19)
dt
(cid:18)
zi − si
τs
− fi − 1
τf
dsi =
dfi = sidt
(2)
(3)
and two equations of the Balloon model2 [2],
dqi =
dvi =
1
τ0
1
τ0
The BOLD signal change is given by
(cid:18)
1 − (1 − E0)1/fi
fi
(fi − vα
i ) dt.
E0
(cid:19)
dt
− vα−1
i
qi
v
(cid:0)1 − q
(cid:2)k1 (1 − qt) + k2
yi(t) = B(qi, viθ) + σydWy
(cid:1) + k3 (1 − v)(cid:3), dWy ∼ N (0, 1) and θ denotes all parameters of the
where B(qt, vtθ) := V0
system. For simplicity, we assume the same hemodynamic transformation for all nodes since our focus is on
connectivity estimates. The parameters are found in [10]. A summary of all the values used in this article is given
in table 1, unless a different value is stated explicitly. For now, we restrict our attention to slow neuronal processes
with a neuronal lag of A = 1 Hz, which is around the typical values in the literature on effective connectivity, e.g.
[8, 9]. The input strength B is chosen such that the resulting amplitude of BOLD responses are around 2-4%.
In addition, we consider for the prior over the initial states in each node xi,0 = (zi,0, si,0, fi,0, qi,0, vi,0) a normal
distribution3 with mean µ0 = (µz,0, µs,0, µf,0, µq,0, µv,0) and a covariance given by a diagonal matrix with entries
(z,s,f,q,v),0 set to be the variance of the stationary distribution induced by the Ornstein-Uhlenbeck process in
σ2
√
equation (1) when It = u(z, t) = 0. Hence, the variance for p(z0) is set to σz,0 = σz/
2. Since all hemodynamic
variables are deterministic, their variance is estimated by forward sampling.
(4)
Now that the model has been described, we proceed with a more detailed explanation of the parameter inference
procedure given by the EM-algorithm.
2.2 EM-algorithm for Time Series
The Expectation-Maximization algorithm [3] is an iterative method to find maximum-likelihood or maximum a
posteriori (MAP) estimates. There are several applications to state space models (SSM), e.g.
[27]. In its general
form, its objective is to maximize a lower bound of the marginal likelihood p(y0:Tθ) =
dz[0:T ]p(z[0:T ], y0:Tθ) where
y0:T = {y(tk)k = 0, . . . , K} is the time series, θ the set of parameters to estimate and z[0:T ] = {ztt ∈ [0, T ]} are
´
2For clarity in the notation, we denote the exponent of the volume fraction 1/α simply as α.
3Notice that due to the small discretization step dt and noise levels used here, the log-transformation of the hemodynamic variables
was not required [29]. Nevertheless, it is straightforward to use this transformation in our procedure.
3
the hidden (continuous) processes underlying the observations. For any probability density q(z[0:T ]) we get from
Jensen's inequality
(cid:2)q(z[0:T ])p(z[0:T ], y0:Tθ)(cid:3) is the Kullback-Leibler (KL) divergence. Hence, the variational density
where DKL
q(z[0,T ]) must be optimized to tighten this inequality. It turns out that the optimal variational density q(z[0:T ]) for
a fixed value of the parameters θ = θ is precisely the posterior p(z[0:T ]y0:T , θ) [19]. Hence, given an initial value θ,
the objective is
(cid:2)q(z[0:T ])p(z[0:T ], y0:Tθ)(cid:3)
log [p(y0:Tθ)] ≥ −DKL
(cid:105)(cid:105)
(cid:104)−DKL
(cid:104)
θ∗ = argmax
θ
p(z[0:T ]y0:T , θ)p(z[0:T ], y0:Tθ)
(5)
This algorithm entails two steps in each iteration. First, the E-step is used to obtain expectations with respect
to the posterior over the latent process z[0,T ]. Then, the M-step is a gradient update to maximize the objective
function with respect to the parameters in question.
2.2.1 E-step: APIS
Given the above model, we can sample from the posterior using APIS. This adaptive importance sampling method
samples N trajectories, or particles, by forward integration of (1)-(3) initialized with u(z, t) = 0. Then, a total cost
Sξ is assigned to each hidden process ξ [11, 15, 16] via
(cid:104)
0
Sξ = −
T
dt log
g(y0:Tz(ξ)
[0,T ])
(cid:105) − 1
2
t )2 −
(u(ξ)
A
σ2
z
√
A
σz
u(ξ)
t dW (ξ)
t
, t) and g(y0:Tz(ξ)
t
t
denotes the noise realization of process ξ, u(ξ)
where dW (ξ)
[0,T ]) is the likelihood with
respect to the latent process z(ξ)
[0,T ] given by equation (4). We call the first term of Sξ the "state cost" of particle
ξ. The negative of the total cost exponentiated gives the unnormalized importance weight used in the Monte Carlo
estimates to learn the control function u(z, t). After u(z, t) is updated, it is used to sample again forward particles.
These steps are repeated until the effective sample size (ESS) converges or reaches a predefined value. For details
on APIS we refer the reader to [23].
t = u(z(ξ)
APIS not only increases the efficiency of sampling, but it also minimizes the mean cost S of the particle
ensemble. This ensures that the posterior neuronal and BOLD signals give the best possible fit, even if there is a
strong mismatch between the model and the ground truth [24]. In addition, the control appears in the computation
of the gradients of the KL divergence, giving reliable Monte Carlo estimates in the M-step.
For a 3 node network, we use N = 20000 particles over 500 iterations to ensure that the scheme has sufficient
time to bootstrap the E-step and converge to the MAP connections in the M-step. Depending on the initialization,
the estimations can take more or less iterations, but we find that this number allows for convergence in most cases.
A learning rate of ηE = 0.1 achieves a fast bootstrapping while maintaining the estimations of the control signals
stable. For the annealing threshold we use γE = 0.02 − 0.03, corresponding to 2-3% of effective samples used for
the estimations during bootstrapping.
2.2.2 M-Step: Gradient Ascent in the Connectivity
In the M-step, we maximize equation (5) with respect to the connectivity matrix C and assume that all other
parameters have correct values. Since C is assumed to be time independent, this optimization requires estimates
of zt at all time. Thus, we wish to maximize at the n-th iteration,
E(cid:110)
log [p(ztzt−dt, C)]y0:T , C (n)(cid:111)
T(cid:88)
t=dt
Qn(C) =
where the expectation is over the posterior latent process. Using equation (1) and the conditional independence of
the components zi
t given zt−dt, the above results in
T(cid:88)
m(cid:88)
t=dt
i=1
E
(cid:26) −1
2σ2
z dt
(cid:0)dzi
t − Fi(zt, C)dt(cid:1)2 y0:T , C (n)
(cid:27)
Qn(C) =
4
where F (zt, C) = A(Czt + BIt). With ∂ikF := ∂F/∂Cik = Azk
finally obtain
T(cid:88)
t=dt
E
(cid:26) A
(cid:16)
σ2
z
∂ikQn =
Aui
tdt +
AσzdW i
t
t and dzt − F (zt, C)dt = Autdt +
(cid:17)
(cid:27)
t y0:T , C (n)
zk
.
√
√
AσzdWt we
This gradient is used to update the connectivity at each iteration,
C (n+1)
ik
= C (n)
ik + ηM ∂ikQn
where ηM is the learning rate.
Note that both the E- and the M-step involve optimization. In the E-step we optimize the importance sampler
for given C (n). In the M-step, we optimize C (n). In practice, we find that it is beneficial to wait until the ESS in
the E-step is sufficiently high. This is ensured by imposing a threshold γM on the ESS above which the connectivity
is updated. Empirically, it is found that γM ∈ [2γE, 5γE] and a learning rate of ηM = 0.01 work well. The latter is
usually found such that after each update, the ESS takes only few iterations to reach γM again. Naturally, there is
a trade-off between large updates of the connectivity and the stability of the ESS. In addition, we use momentum
with a rate of κ = 0.9 to improve the gradient ascent procedure.
Remember that the within-node time scale is assumed to be the fastest time scale in the network and all nodes
have the same value. Hence, we fix the diagonal elements of C to −1. Although they are not optimized, the above
procedure can be readily extended to estimate these elements as well.
The initial random matrices for n = 0 are generated with the following procedure similar to [17]. First, on-edges
are sampled with probability p1 = 0.5. This gives a random graph that defines the adjacency matrix, i.e. the
directed connections among the ROIs. Then, the strength of these connections are randomly chosen from a uniform
distribution on a small interval [c, d] where c , d ≤ 1. Here, the bound on the strength are chosen to be c = −0.5
and d = 0.5. This ensures that the sampled matrix has negative eigenvalues with very high probability, which is
important to guarantee the stability of the latent process. The resulting matrix is accepted if all eigenvalues are
negative, otherwise, the procedure is repeated.
2.3 Studying the Effects of the Neuronal Time Scale on Effective Connectivity Esti-
mates
The use of slow dynamics for the neuronal processes underlying fMRI data is argued in [9], but it stands in contrast
to the case studied in [28] with a neural lag of approximately 50 ms, corresponding to A = 20 Hz. This is a
significant difference in the assumptions made on the generative process of the data, hence, it is important to
understand the consequences of these assumptions for effective connectivity estimates. For this, we study two
important effects that depend on the time scale of the neuronal activity, the sensitivity of the BOLD response to
changes in the connectivity and the maximum-likelihood solutions under different assumptions of the time scale.
As a measure of sensitivity, the difference in the mean and variance of the posterior BOLD between two significant
different models is computed. We call this difference the "sensitivity" of the mean sµ and variance sσ respectively.
Hence, if the change in the response is large when moving from one connectivity model to another, we say that the
system is highly sensible to changes in the connectivity. On the contrary, small changes in the BOLD response will
make the discrimination between connectivity models more difficult.
In addition, the inverse time scale A not only affects the sensitivity of the BOLD response, but also its overall
shape and delay. This could have significant effects on the connectivity estimates. Thus, we study the solutions
depending on the assumed neuronal time scale given that the underlying generative process has fast neuronal
activation, i.e. we consider a model mismatch between the ground truth inverse time scale AGT and the assumed
one for the reconstruction of the network A.
For simplicity, we consider the deterministic system because we find that the variance of the posterior contributes
marginally to the neg. log-likelihood (sσ (cid:28) sµ) and the posterior mean follows the same dynamics as the determin-
istic system for sufficiently small noise levels. Hence, since the only randomness is in the observations, we do not
use the EM approach described above. Instead, we use a brute force search on the neg. log-likelihood landscape.
This approach requires the discretization of the connectivity space and, for each grid point, the integration of the
full system on the time interval [0, T ].
In this analysis, consider for the ground truth a chain network I → 1 → 2 → 3 with connectivity matrix CGT,c
and input I. To lower the computational effort in finding the connections we restrict the problem by fixing all
connections but two, say (C31, C32). The brute force approach even in this restricted scenario is computationally
too expensive to study the grid search solutions for a large number of time series. Hence, the search on this plane
5
is constrained further to the line C31 = −gC32 + h on which, given the input, the fix point of the neuronal system
is invariant to changes in the connectivity. This constrain gives g = (CGT,c)23 and h = g · (CGT,c)32.
To see this, consider the fix point solution z∗ of equation (1) with u(z, t) = 0, σz = 0. For node 3, the dependence
of z∗
3 on the connections (C31, C32) is given by z∗
1 is the fix point of the activity in node
1. Keeping the input fixed sets the value of z∗
3 will be completely determined by the proportionality factor
h = gC32 + C31. By setting h = g · (CGT,c)32, we ensure that the fix point of the system remains constant at the
value of the ground truth while we vary the connections on this line, which we call the z∗-line. Hence, the only
relevant information to differentiate between connections is in the transients given by the values on this line. We will
denote estimates obtained by the brute force search on the z∗-line as grid search connections to distinguish them
from the EM approach described above. For a validation on this procedure we refer the reader to the appendix.
3 = [gC32 + C31]z∗
1, so z∗
1, where z∗
We use the grid search on two studies on the effect of the neuronal time scale on the grid search connections and
the measurement precision needed to obtain reliable connectivity estimates. In these studies, we perform a brute
force search on the z∗-line as follows. Given a value C32, we integrate the system to obtain the BOLD response,
which is used to compute the log-likelihood and find the grid search solution of 1000 noise realizations, first keeping
σy fixed and varying A, then keeping A fixed to the correct value AGT and varying σy.
Notice that the variance observed in this analysis is clearly an under estimate of the true variance of the grid
search solutions in the full connectivity space, because we restricted ourselves to a single line C31 = −gC32 + h in
this space. However, the results of this analysis capture already the consequences of fast neuronal dynamics for
connectivity estimates.
3 Results
3.1 Single Trial Estimates of Effective Connectivity from Slow Neuronal Activation
In the following analysis we consider two connectivity structures with different topologies with no bidirectional
connection, a chain network CGT,c and a triangle network CGT,∆. In both cases, we consider an external input to
node 1 given by a box-car function It,1 = Θ[t − ton]Θ[tof f − t] where (ton, tof f ) = (3.2, 4.70) seconds and Θ is the
Heaviside function. To generate the data from both examples CGT,c and CGT,∆, the systems are integrated forward
T = 16 seconds with the initial state set to the mean value µ0. In each case, the response signals of the nodes are
subsampled with a T R = 0.4 seconds, resulting in 3 time series each with 41 observations corrupted by Gaussian
noise (σy = 0.002). This noise level is chosen to be 5-8% of the BOLD change, roughly the same as in [9].
are assumed, we use an L1-regularization term Σi,jλL1 Cij in equation (5).
We illustrate how the EM-algorithm together with APIS obtains the correct directed edges. Since sparse networks
For the first example, we consider the chain network with connections
−1
CGT,c =
0
0
0.5 −1
0
0.75 −1
0
.
In figure 1, the connectivity estimation for the chain network C c
λL1=0.05 is shown over iterations. In this case, we
choose a regularization of λL1 = 0.05. Each graph in the panels corresponds to a different random initialization
of the connectivity matrix. At the beginning, the connections do not change because the sampler must learn
an appropriate controller to bootstrap the procedure. Once the controller is estimated with sufficient accuracy,
the ESS surpasses a threshold and the connectivity is updated via gradient ascent, hence the sudden jumps in the
connections. After a few jumps, the learning procedure stabilizes and the connections are updated at each iteration.
In most cases, it takes around 200-250 iterations to converge to the MAP solution.
used. The mean state cost of the sampled process E(cid:0)log(cid:2)g(y0:Tz[0,T ])(cid:3)(cid:1) decreased in all cases from around 1000
The ESS after convergence of most examples increased from less than 1% to higher than 90% of total particles
to 68, meaning that the model found fits the data well.
Notice that the randomly initialized matrices converged all to the same values close to the ground truth. The
31 is sufficiently weak and could be regarded as non existent since the other estimated connections
connection C c
32 can be understood from the L1-
dominate by an order of magnitude. The underestimation of the connection C c
penalty in the M-step to obtain sparse solutions. This regularization shrinks the connections by penalizing their
absolute value. In turn, this shrinkage is compensated by an increase in the other incoming connection to node 3 to
reduce the state cost, i.e. C c
31 becomes non-zero. This hints to a strong regularization. To see the effect of strong
regularization, we study the dependency of the solutions on the regularization strength λL1.
6
Figure 1: MAP-solutions for chain network: The figure shows the consistency of the estimation procedure against
random initializations of the connectivity matrix. Each line represents the estimation procedure over iterations of
a randomly initialized example. The regularization strength is set to λL1 = 0.05. All runs converged to the same
matrix C c
λL1=0.05. The estimated connections are not identical to the ground truth because of the finite data used
in reconstruction. However, the qualitative structure of the network, including which connections are present and
their directionality, is correctly reconstructed. On the lower right panel only an example of the upper off-diagonal
connections is shown. In all cases they are estimated to vanish with very high precision.
7
Figure 2: MAP-solutions for chain network vs L1-regularization strength: a small amount of regularization λL1 =
0.01 improves the estimation of the connectivity compared to the unregularized case. The mean validation error
of the BOLD signal over 20 new time series decreases compared to the unregularized case but increases rapidly
with stronger regularization (right). Taking the connectivity at the minimum of the validation error gives us a
connectivity with a small error relative to the ground truth connection (black dashed line, left panel)
For a fixed initialization of uncoupled nodes C c
0 = −13×3, the connections of the chain network are estimated
using different values λL1 ∈ [0, 0.1]. The analysis shows -- left panel of figure 2 -- that the ground truth connections
(black dashed line) are consistently revealed across a wide range of λL1-values. As expected, zero regularization
gives small non-zero contributions in edges that would not be allowed, but still the ground truth connections are
easy to differentiate. Hence, although the L1-regularization is not crucial in this example, we see that applying a
small regularization helps shrinking the value of non existent connections towards zero, which improves the estimate
of the network making it generalize better, as seen on the right panel of figure 2, where the mean validation error
is shown.
To estimate the validation error we generate 1000 new time series using the ground truth connection and
computed the mean negative log-likelihood or sate cost of the response obtained from each connectivity. This is a
measure of how well our estimates fit the new data. We observe a decrease for weak regularization λL1 = 0.01 vis-
ï¿oe-vis the unregularized case and a strong increase in the validation error for increasing λL1. From the minimum
of this profile, we conclude that a weak regularization is beneficial, in this case using λL1 = 0.01 we obtain a solution
pinpointing the ground truth to
Finally, we proceed with the estimation of the triangle network
to show that we obtain again the ground truth structure with a small regularization of λL1 = 0.01. In figure 3,
we observe a robust estimation against random initializations of the connectivity matrix and a clear distinction
between correct and incorrect edges and directions. In this case, all estimated connections are around
C c
λL1=0.01 =
0
0
0.45 −1
0
0.78 −1
0
CGT,∆ =
0
0
0.5 −1
0
0.5 −1
0.375
C ∆
λL1=0.01 =
0
0
0.48 −1
0.03
0.4 −1
0.41
.
.
−1
−1
−1
8
Figure 3: Triangle network with similar connection strengths as the chain network: The MAP-solution is again
robust against random initializations of the connectivity and we can clearly distinguish between existent and non-
existent edges and directions.
Interestingly, in this case the overestimated connection is C ∆
23 with the same order of magnitude as C ∆
31 in the
previous example. Again, we may vary λL1 to select its value by cross validation.
Identifying Networks with Fast Neuronal Activations
3.2
3.2.1 BOLD Sensitivity to Changes in the Connectivity Diminishes with Fast Neuronal Dynamics
The connectivity estimation above assumed that the temporal scale of the neuronal dynamics is of the order of a
second. Here, we study the sensitivity of the BOLD response to changes in the connectivity as a function of A. We
consider 3 different time scales A = 1, 5 and 50. The neuronal noise level σz = 0.01 is chosen such that the prior
process has a standard deviation of 0.7%. The learning rate of APIS has to be adapted for the different values of
A, roughly with the relation ηE = 0.1/A. All other parameters of APIS are kept fixed.
Given the data generated by the chain network CGT,c, the posterior statistics of the BOLD signal are estimated
in two extreme cases using APIS; one case defines the structure 3 ← 1 → 2 and the other is close to the ground truth
CGT,c. The connectivity matrices are chosen such that the only difference to the ground truth are the connections
into node 3,
.
−1
C1 =
0
0
0.5 −1
0
0 −1
0.375
; C2 =
−1
0
0
0.5 −1
0
0.63 −1
0.06
In figure 4, we show the sensitivity of the BOLD response in node 3 when changing the connections from C1 to
C2. We appreciate two characteristics of the posterior BOLD response.
First, on the right panel, it is apparent that the variance of the signal has little sensitivity to the connectivity,
with a change several orders of magnitude smaller than the change of the mean signal on the left. This suggests
9
Figure 4: Sensitivity of the BOLD posterior to changes in the connectivity. Notice that the change in the variance of
the posterior BOLD (right) is five orders of magnitude smaller than the change in the posterior mean (left). Hence,
faster neuronal activity implies less sensitivity of the BOLD signal to changes in the connectivity. The dashed black
line is the noise level σy = 0.002 of the data generated with C c
GT .
that -- for small neuronal noise -- the contribution of the variance to the identification of the connectivity is marginal
and we can focus on the mean signal, which is equal to the BOLD signal of a deterministic system.
Second, on the left panel, for slow dynamics the amplitude of the mean signal change is larger than the observation
noise level σy. However, already for A = 5 the sensitivity decreases below the observation noise and it becomes
an order of magnitude smaller for large A. Hence, the faster the neuronal dynamics, the less sensible the BOLD
response becomes to changes in the connectivity. The lack of sensitivity for faster time scales implies flat likelihood
functions that contain no information to differentiate connectivity structures. Thus, A should be sufficiently small
such that the sensitivity is higher than the observation noise. This seems to be a fundamental problem to reconstruct
connectivity.
One solution to this problem is more data, which effectively reduces the measurement noise. Another, option is
to reconstruct with small A and hope for the best. Thus, we study the grid search solutions obtained for different
inverse time scales A by the brute force search described before.
3.2.2 Connectivity Estimates Depend on the Assumed Neuronal Time Scale
We study now the connectivity estimates as a function of the inverse neuronal time scale A. For each value of A,
we find the connections from 1000 different time series with the ground truth given by the chain network CGT,c
with the same parameters as before but AGT = 35. Figure 5 shows the mean and standard deviation of these
estimates. Taking the amplitude of sµ in figure 4 as a rough upper limit for the noise level that allows connectivity
estimations, the analysis is performed with two noise levels, σy = 2 × 10−4 (on the top panels) and σy = 2 × 10−3
(on the bottom panels).
Upon examination of the top panels, we reinforce our expectation that with sufficient precision, the variance
of the grid search solutions is small. Interestingly, the inverse time scale A biases the connectivity estimates in a
significant way. Nevertheless, it is possible to obtain in principle the correct connectivity if the time scale is jointly
estimated because there is a clear minimum in the negative log-likelihood profile (red dot on the right panel).
On the contrary, with a noise level an order of magnitude higher, the identification of connections is problematic.
Although the bias of the estimations is the same as above, the variance of the grid search solutions increases much
faster with the inverse time scale because of the flat profiles of the neg. log-likelihood function. Thus, the grid search
solutions spread across a wide range of values, making it difficult to pinpoint the connectivity. As a consequence,
for A ≥ 15 Hz identifying the correct connections will not be possible, even after learning the inverse time scale.
This can be appreciated by the flat profile of the mean neg. log-likelihood at the grid search solutions.
Our analysis focuses on grid search solutions on the restricted connectivity space C31 = −gC32 + h for simplicity.
However, a similar analysis on the bidirectional connections results also in an extreme bias in the connectivity
10
Figure 5: Connections obtained via grid search vs inverse time scale A for σy = 0.0002 (top) and σy = 0.002
(bottom). Left: Mean (solid) and standard deviation (dashed) of grid search solutions on C31 = −gC32 + h over
different realizations of the observation noise. All connections are fixed besides the in-connections C32, C31 to node
3. The horizontal black lines are the values of the ground truth connections at C32 = 0.75 and C31 = 0. Right:
The averaged minimum of the neg. log-likelihood function for each inverse time scale. The ground truth is given
by the model described in section 2.1 but with AGT = 35. The red dot on the right panels represent the minimum
of the mean neg. log-likelihood profile. Due to the large variance and the flat neg. log-likelihood the minimum in
the bottom case was estimated around A = 40.
11
Figure 6: Standard deviation of grid search solutions for connection C32 depending on the observation noise. The
increase in precision needed to obtain reliable connectivity estimates is a direct result of the insensitivity of the
BOLD signal to changes in the connectivity. Notice the log-scaling of both axes.
estimates due to the wrong time scale. This can be appreciated by inspecting the neg. log-likelihood landscape of
a bidirectional connection with a mismatch in A (not shown). As before, if the data is generated by a fast process
but one assumes a slow process, the landscape is distorted and barely changes across different noise realizations and
values of σy. The distorted landscape features a minimum around (C32, C23) = (1.3,−1.3), although the data was
generated using (C32, C23) = (0.75, 0).
Finally, we study the variance of the connectivity estimates as a function of the observation noise σy. For this,
data is generated with AGT = 1, 5 and 35 and the brute force search is performed without model mismatch. As
expected from figure 5, the mean connectivity remains close to the ground truth values for all noise levels (not
shown).
Figure 6 shows the standard deviation of the grid search connections C32. We observe that the noise level
allowed for connectivity estimation depends on the inverse time scale of the neuronal dynamics. There are two
effects here. First and most intuitive, the variance of the grid search connections increases with the observation
noise. Notice the linear increase for large values in this log-log-graph. Second but more important, fast time scales
in the neuronal signal exacerbate this increase at least an order of magnitude due to the insensitivity of the BOLD
signal to changes in the connectivity. Notice how the dependency flattens by lowering the noise, first for A = 1,
then for A = 5 and finally for A = 35 in the extreme high precision regime σy ∝ 10−5, where the standard deviation
of the grid search solutions for A = 1 and A = 5 collapse. This shows that -- even in this idealized case -- to resolve
connections when the neuronal time scale is around 30 ms we require at most σy < 10−3.
4 Discussion
In this paper we present a method to obtain point estimates of effective connectivity from fMRI time series. This
method is based on the Expectation-Maximization algorithm with the E-step being performed by APIS on the
same biologically plausible model used in DCM. We show that APIS-EM obtains robust estimations independent
of random initializations of the connectivity matrix. In addition, we analyze the effect of the neuronal time scale on
estimates of the effective connectivity. This analysis shows that if the underlying neuronal dynamics are fast, the
connectivity estimates will be extremely biased if slow processes are assumed in the reconstruction. Hence, jointly
estimating the neuronal time scale and the connectivity is important. However, due to the high insensitivity of the
BOLD signal for faster neuronal dynamics, this becomes increasingly more demanding and requires higher quality
data.
We conclude that robust reconstruction of the connectivity between ROIs including the directionality requires
much higher precision in the data when assuming a realistic (fast) time scale for neuronal dynamics. In addition,
12
if a slow time scale is assumed (as in [8, 9]) the inferred connectivity can be completely wrong if the generative
neuronal process is fast.
The adaptive importance sampler APIS is an efficient tool to compute posterior estimates over the hidden
processes. Importantly, the sampling step is easily parallelizable such that one can lower the computational burden
involved in integrating the system of non-linear differential equations while having a large amount of samples to
ensure reliable estimates.
Furthermore, APIS and its generalization based on PICE [?], are flexible frameworks that allow an easy adap-
tation of the M-step to learn the hemodynamic parameters by including noise in the hemodynamic degrees of
freedom. Hence, instead of an m-dimensional controller, APIS would learn a 5m-dimensional controller. Although
the computational burden increases in this case, we find empirically that the fully controlled version behaves more
stable in terms of the effective sample size. Nevertheless, for simplicity we focused here on the reconstruction of the
neural network and assumed the hemodynamic system is known. Hence, to maintain a lower computational cost,
we chose the simplified, under-actuated setting.
Since the proposed scheme obtains robust estimates from randomly initialized connectivity matrices, it can be
considered as a gradient-based exploratory procedure able to change the initial connectivity to find better fitting
models. This addresses some of the concerns raised in [17]. Naturally, better initializations improve performance.
A natural choice is the symmetric connectivity obtained from methods known to infer the right undirected network,
for instance partial correlation or regularized inverse covariance.
Moreover, the addition of modulatory inputs or the consideration of more complex neuronal systems, e.g. firing
rate models with non-linear activation functions, is straightforward. Furthermore, extensions of the proposed
method to obtain Gaussian approximations of the posterior over all parameters similar to DCM can be worked out.
This makes the combination of APIS and EM a flexible and efficient alternative to DCMs.
Finally, a word of caution. The bias in the connectivity estimates assuming the wrong time scale has significant
consequences for the estimation of the hemodynamic parameters because the resulting BOLD signal is shifted
roughly a second with respect to the ground truth. If both sets of parameters are jointly estimated, this shift will bias
the hemodynamic parameters as well. On the contrary, fast neuronal dynamics make the delays being completely
determined by the hemodynamics. This might lower the "interference" between delays caused by the neuronal time
scale and the ones caused by the hemodynamics, possibly increasing identifiability of the hemodynamic system.
Hence, it is important to clarify the temporal scale of the underlying processes, specially because this assumption
has a major impact on the required quality of the data to obtain correct directed networks.
Acknowledgements
This work was supported by the European Commission through the FP7 Marie Curie Initial Training Network
289146, NETT: Neural Engineering Transformative Technologies.
Appendix: Validation of Grid Search on z∗-Line
We validate the grid search approach by studying the neg. log-likelihood landscape on (C31, C32) for two extreme
cases with and without the correct inverse time scale. We show graphically that when there is a mismatch in A,
the bias caused by the restriction on the z∗-line is small compared to the bias caused by the wrong time scale.
For the validation, data was generated using the chain network with AGT = 35 and two different noise levels
σy = 2 · 10−3 and σy = 2 · 10−4. Figure 7 left panel shows the neg.
log-likelihood on (C31, C32) for σy = 0.002
estimated using the correct time scale. Although the high noise level 'washes out' the z∗-line (magenta), there is
a clear direction along this line where the values are smallest. Perpendicular to this direction the value increases.
Thus, minima lie in this case on the line. Due to the high noise level, the neg. log-likelihood along this line is flat.
For lower noise levels, the valley becomes more and more narrow and the profile of the neg. log-likelihood on this
line shows a clear minimum.
On the right panel of figure 7, we consider more precise data generated with σy = 2 · 10−4 but compute the
neg.
log-likelihood with a model mismatch in the time scale (A = 1). This distorts the landscape such that the
low valued region around the z∗-line transforms into an elliptic shaped valley with the minimum (dot #2) far away
from the ground truth (dot #1). Notice how the major axis is parallel to the z∗-line and, although the minimum
does not lie exactly on the line, the error is small compare to the bias caused by the wrong time scale. Thus,
restricting the search to the z∗-line is a good approximation for the purpose of studying the grid search connections
as a function of A.
13
Figure 7: Validating the restriction of the search to the z∗-line (magenta line). Negative log-likelihood in log-scale
for two different cases of the parameters A and σy. The upper most data point gives the position of the ground
truth. Left: No mismatch in the time scale of the model (A = AGT = 35 Hz) and σy = 0.002. Notice the direction
of the valley along the z∗-line. Right: There is a mismatch between the ground truth time scale (AGT = 35 Hz)
and the model (A = 1 Hz). Although the noise is small σy = 2· 10−4, there is a wide valley in the form of an ellipse.
Notice how the major axis is parallel to the z∗-line. The minimum (lower data point) falls in this case outside
the line but the bias introduced by disregarding the perpendicular direction is small compared to the overall bias
caused by the wrong time scale.
References
[1] Michael Breakspear. Dynamic and stochastic models of neuroimaging data: A comment on lohmann et al.
Neuroimage, 75:270 -- 274, 2013.
[2] Richard B Buxton, Eric C Wong, and Lawrence R Frank. Dynamics of blood flow and oxygenation changes
during brain activation: the balloon model. Magnetic resonance in medicine, 39(6):855 -- 864, 1998.
[3] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the
em algorithm. Journal of the royal statistical society. Series B (methodological), pages 1 -- 38, 1977.
[4] Karl Friston, Jean Daunizeau, and Klaas Enno Stephan. Model selection and gobbledygook: Response to
lohmann et al. Neuroimage, 75:275 -- 278, 2013.
[5] Karl Friston, Rosalyn Moran, and Anil K Seth. Analysing connectivity with granger causality and dynamic
causal modelling. Current opinion in neurobiology, 23(2):172 -- 178, 2013.
[6] Karl Friston, Klaas Stephan, Baojuan Li, and Jean Daunizeau. Generalised filtering. Mathematical Problems
in Engineering, 2010, 2010.
[7] Karl J Friston. Functional and effective connectivity in neuroimaging: a synthesis. Human brain mapping,
2(1-2):56 -- 78, 1994.
[8] Karl J Friston, Lee Harrison, and Will Penny. Dynamic causal modelling. Neuroimage, 19(4):1273 -- 1302, 2003.
[9] Karl J Friston, Baojuan Li, Jean Daunizeau, and Klaas E Stephan. Network discovery with dcm. Neuroimage,
56(3):1202 -- 1221, 2011.
[10] Karl J Friston, Andrea Mechelli, Robert Turner, and Cathy J Price. Nonlinear responses in fmri: the balloon
model, volterra kernels, and other hemodynamics. NeuroImage, 12(4):466 -- 477, 2000.
[11] Igor Vladimirovich Girsanov. On transforming a certain class of stochastic processes by absolutely continuous
substitution of measures. Theory of Probability & Its Applications, 5(3):285 -- 301, 1960.
[12] Rainer Goebel, Alard Roebroeck, Dae-Shik Kim, and Elia Formisano. Investigating directed cortical interac-
tions in time-resolved fmri data using vector autoregressive modeling and granger causality mapping. Magnetic
resonance imaging, 21(10):1251 -- 1261, 2003.
14
[13] Clive WJ Granger. Investigating causal relations by econometric models and cross-spectral methods. Econo-
metrica: Journal of the Econometric Society, pages 424 -- 438, 1969.
[14] Pearl Judea. Causality: models, reasoning, and inference. Cambridge University Press. ISBN 0, 521(77362):8,
2000.
[15] Hilbert J Kappen. Linear theory for control of nonlinear stochastic systems. Physical review letters,
95(20):200201, 2005.
[16] Hilbert J Kappen, Vicenç Gómez, and Manfred Opper. Optimal control as a graphical model inference problem.
Machine learning, 87(2):159 -- 182, 2012.
[17] Gabriele Lohmann, Kerstin Erfurth, Karsten Müller, and Robert Turner. Critical comments on dynamic causal
modelling. Neuroimage, 59(3):2322 -- 2329, 2012.
[18] Joseph B Mandeville, John JA Marota, C Ayata, Greg Zaharchuk, Michael A Moskowitz, Bruce R Rosen, and
Robert M Weisskoff. Evidence of a cerebrovascular postarteriole windkessel with delayed compliance. Journal
of Cerebral Blood Flow & Metabolism, 19(6):679 -- 689, 1999.
[19] Radford M Neal and Geoffrey E Hinton. A view of the em algorithm that justifies incremental, sparse, and
other variants. In Learning in graphical models, pages 355 -- 368. Springer, 1998.
[20] Rajan S Patel, F DuBois Bowman, and James K Rilling. A bayesian approach to determining connectivity of
the human brain. Human brain mapping, 27(3):267 -- 276, 2006.
[21] Will Penny, Zoubin Ghahramani, and Karl Friston. Bilinear dynamical systems. Philosophical Transactions of
the Royal Society of London B: Biological Sciences, 360(1457):983 -- 993, 2005.
[22] Joseph D Ramsey, Stephen José Hanson, Catherine Hanson, Yaroslav O Halchenko, Russell A Poldrack, and
Clark Glymour. Six problems for causal inference from fmri. neuroimage, 49(2):1545 -- 1558, 2010.
[23] Hans-Christian Ruiz and Hilbert J Kappen. Particle smoothing for hidden diffusion processes: Adaptive path
integral smoother. IEEE Transactions on Signal Processing, 65(12):3191 -- 3203, 2017.
[24] H.-Ch. Ruiz Euler and H.J. Kappen. Nonlinear deconvolution by sampling a biophysically plausible hemody-
namic model. 2017.
[25] Srikanth Ryali, Kaustubh Supekar, Tianwen Chen, and Vinod Menon. Multivariate dynamical systems models
for estimating causal interactions in fmri. Neuroimage, 54(2):807 -- 823, 2011.
[26] Simo Särkkä. Bayesian filtering and smoothing, volume 3. Cambridge University Press, 2013.
[27] Thomas B Schön, Adrian Wills, and Brett Ninness. System identification of nonlinear state-space models.
Automatica, 47(1):39 -- 49, 2011.
[28] Stephen M Smith, Karla L Miller, Gholamreza Salimi-Khorshidi, Matthew Webster, Christian F Beckmann,
Thomas E Nichols, Joseph D Ramsey, and Mark W Woolrich. Network modelling methods for fmri. Neuroimage,
54(2):875 -- 891, 2011.
[29] Klaas Enno Stephan, Lars Kasper, Lee M Harrison, Jean Daunizeau, Hanneke EM den Ouden, Michael Break-
spear, and Karl J Friston. Nonlinear dynamic causal models for fmri. Neuroimage, 42(2):649 -- 662, 2008.
15
|
1606.08232 | 1 | 1606 | 2016-06-27T12:15:39 | Information integration from distributed threshold-based interactions | [
"q-bio.NC",
"cond-mat.stat-mech"
] | We consider a collection of distributed units that interact with one another through the sending of messages. Each message carries a positive ($+1$) or negative ($-1$) tag and causes the receiving unit to send out messages as a function of the tags it has received and a threshold. This simple model abstracts some of the essential characteristics of several systems used in the field of artificial intelligence, and also of biological systems epitomized by the brain. We study the integration of information inside a temporal window as the model's dynamics unfolds. We quantify information integration by the total correlation, relative to the window's duration ($w$), of a set of random variables valued as a function of message arrival. Total correlation refers to the rise of information gain above and beyond that which the units already achieve individually, being therefore related to consciousness studies in some models. We report on extensive computational experiments that explore the interrelations of the model's parameters (two probabilities and the threshold), highlighting relevant scenarios of message traffic and how they impact the behavior of total correlation as a function of $w$. We find that total correlation can occur at significant fractions of the maximum possible value and provide semi-analytical results on the message-traffic characteristics associated with values of $w$ for which it peaks. We then reinterpret the model's parameters in terms of the current best estimates of some quantities pertaining to cortical structure and dynamics. We find the resulting possibilities for best values of $w$ to be well aligned with the time frames within which percepts are thought to be processed and eventually rendered conscious. | q-bio.NC | q-bio |
Information Integration from Distributed
Threshold-Based Interactions
Valmir C. Barbosa∗
Programa de Engenharia de Sistemas e Computa¸cao, COPPE
Universidade Federal do Rio de Janeiro
Caixa Postal 68511
21941-972 Rio de Janeiro - RJ, Brazil
Abstract
We consider a collection of distributed units that interact with one
another through the sending of messages. Each message carries a pos-
itive (+1) or negative (−1) tag and causes the receiving unit to send
out messages as a function of the tags it has received and a threshold.
This simple model abstracts some of the essential characteristics of sev-
eral systems used in the field of artificial intelligence, and also of biological
systems epitomized by the brain. We study the integration of information
inside a temporal window as the model's dynamics unfolds. We quantify
information integration by the total correlation, relative to the window's
duration (w), of a set of random variables valued as a function of message
arrival. Total correlation refers to the rise of information gain above and
beyond that which the units already achieve individually, being therefore
related to consciousness studies in some models. We report on extensive
computational experiments that explore the interrelations of the model's
parameters (two probabilities and the threshold), highlighting relevant
scenarios of message traffic and how they impact the behavior of total
correlation as a function of w. We find that total correlation can occur
at significant fractions of the maximum possible value and provide semi-
analytical results on the message-traffic characteristics associated with
values of w for which it peaks. We then reinterpret the model's parame-
ters in terms of the current best estimates of some quantities pertaining
to cortical structure and dynamics. We find the resulting possibilities for
best values of w to be well aligned with the time frames within which
percepts are thought to be processed and eventually rendered conscious.
Keywords: Threshold-based system, Information integration, Total cor-
relation, Cortical dynamics, Percept, Consciousness.
∗[email protected].
1
1 Introduction
A threshold-based system is a collection of loosely coupled units, each charac-
terized by a state function that depends on how the various inputs to the unit
relate to a threshold parameter. The coupling in question refers to how the
units interrelate, which is by each unit communicating its state to some of the
other units whenever that state changes. Given a set of timing assumptions
and how they relate to the exchange of states among units as well as to state
updates, each individual unit processes inputs (states communicated to it by
other units) and produces a threshold-dependent output (its own new state,
which gets communicated to other units).
The quintessential threshold-based system is undoubtedly the brain, where
each neuron's state function determines whether an action potential is to be fired
down the neuron's axon. This depends on how the combined action potentials
the neuron perceives through the synapses connecting other neurons' axons
to its dendrites (its synaptic potentials) relate to its threshold potential [1].
The greatly simplified model of the natural neuron known as the McCulloch-
Pitts neuron [2], introduced over seventy years ago, retained this fundamental
property of being threshold-based, and so did generalizations thereof such as
generalized Petri nets [3] and threshold automata [4]. In fact, this holds for much
of the descent from the McCulloch-Pitts neuron, which has extended through
the present day in a succession of ever more influential dynamical systems.
Such descent includes the essentially deterministic Hopfield networks of the
1980s [5,6] and moves on through generalizations of those networks' Ising-model
type of energy function and the associated need for stochastic sampling. The
resulting networks include the so-called Boltzmann machines [7] and Bayesian
networks [8, 9], as well as the more general Markov (or Gibbs) Random Fields
[10 -- 12] and several of the probabilistic graphical models based on them [13]. A
measure of the eventual success of such networks can be gained by considering,
for example, the restricted form of Boltzmann machines [14, 15] used in the
construction of deep belief networks [16], as well as some of the other deep
networks that have led to landmark successes in the field of artificial intelligence
recently [17 -- 19].
Our interest in this paper is the study of how information gets integrated
as the dynamics of a threshold-based system is played out. The meaning we
attach to the term information integration is similar to the one we used previ-
ously in other contexts [20, 21]. Given a certain amount of time w and a set
of random variables, each related to the firing activity of each of the system's
units inside a temporal window of duration w, we quantify integrated informa-
tion as the amount of information the system generates as a whole (relative to
a global state of maximum entropy) beyond that which accounts for the aggre-
gated information the units generate individually (now relative to local states
of maximum entropy). This surplus of information is known as total correla-
tion [22] and is fundamentally dependent on how the units interact with one
another.
2
Our understanding of information integration, therefore, lies in between
those used by approaches that seek it in the synchronization of input/output
signals (see [23], for example) and those that share our view but would con-
sider not just the whole and the individual units but all partitions in between
as well [24]. By virtue of this, we remain aligned with the latter theory by ac-
knowledging that information gets integrated only when it is generated by the
whole in excess of the total its parts can generate individually. On the other
hand, by sticking with total correlation as an information-theoretic quantity
that requires only two partitions of the set of units to be considered (one that is
fully cohesive and another that is maximally disjointed), we ensure tractability
way beyond that of the all-partitions theory.
We conduct all our study on a simple model of threshold-based systems.
In this model, the units are placed inside a cube and exchange messages whose
delivery depends on their propagation speed and the Euclidean distance between
sender and receiver. Every message is tagged, and upon reaching its destination,
its tag is used to move a local accumulator either toward or away from the
threshold. Reaching the threshold makes the unit send out messages and the
accumulator is reset. There are three parameters in the model. Two of them are
probabilities (that a message is tagged so that the accumulator at the destination
gets decreased upon its arrival, and that a unit sends a message to each of the
other units), the other being the value of the threshold. Parameter values are
the same for all units.
This model is by no means offered as an accurate representation of any par-
ticular threshold-based system, but nevertheless summarizes some key aspects
of such systems through its relatively few parameters. In particular, it gives
rise to three possible expected global regimes of message traffic. One of them
is perfectly balanced, in the sense that on average as much traffic reaches the
units as leaves them. In this case, message traffic is sustained indefinitely. In
each of the other two regimes, by contrast, either more traffic reaches the units
than leaves them, or conversely. Message traffic dies out in the former of these
two (unless the units receive further external input) but grows indefinitely in
the latter one.
We find that information integration is guaranteed to occur at high levels
for some window durations whenever message traffic is sustained at the perfect-
balance level or grows. We also find that this happens nearly independently of
parameter variations. On the other hand, we also find that information inte-
gration is strongly dependent on the model's parameters, with significant levels
occurring only for some combinations, whenever message traffic is imbalanced
toward the side that prevents it from being sustained. Here we once again turn
to the brain, whose cortical activity is in many accounts characterized as tending
to be sparse [25, 26], as an emblematic example.
We proceed as follows. Our message-passing model is laid out in Section 2,
where its geometry and distributed algorithm are detailed and the question of
message imbalance is introduced. An account of our use of total correlation is
given in Section 3, followed by our methodology in Section 4. Results, discussion,
and conclusion follow, respectively in Sections 5, 6, and 7.
3
2 Model
Our system model comprises a structural component and an algorithmic one.
The two are described in what follows, along with some analysis of how they
interrelate.
2.1 Underlying geometry
For 1 ≤ D ≤ 3, our model is based on N simple processing units, henceforth
referred to as nodes, each one placed at a fixed position inside the D-dimensional
cube of side ℓ. The position of node i has coordinates X (1)
, so the
Euclidean distance between nodes i and j is
, . . . , X (D)
i
i
δij = q(X (1)
i − X (1)
j
)2 + · · · + (X (D)
i − X (D)
j
)2.
(1)
We assume that nodes can communicate with one another by sending messages
that propagate at the fixed speed σ on a straight line. The delay incurred by a
message sent between nodes i and j in either direction is therefore δij /σ.
Our computational experiments will all be such that the N nodes are placed
in the D-dimensional cube uniformly at random. In this case, and in the limit
of infinite N , the expected distance between two randomly chosen nodes i and
j is given by
ℓ = (cid:18) 1
∆(D)
ℓ(cid:19)2DZ ℓ
0
· · ·Z ℓ
0
δijdX (1)
i
· · · dX (D)
i
dX (1)
j
· · · dX (D)
j
,
(2)
where 1/ℓ is the probability density for each of the 2D variables. Letting X (d)
ℓx(d)
k
in this equation for k ∈ {i, j} and d ∈ {1, . . . , D} yields
k =
∆(D)
ℓ =
ℓ2D+1
ℓ2D Z 1
0
· · ·Z 1
0
δijdx(1)
i
· · · dx(D)
i dx(1)
j
· · · dx(D)
j
,
where δij now has x's in place of the X's. We then have
∆(D)
ℓ = ℓ∆(D)
1
,
(3)
(4)
with the expected distances in the unit cube for the numbers of dimensions of
interest being well known: ∆(1)
1 ≈ 0.6617
[28].
1 ≈ 0.5214 [27], and ∆(3)
1 = 1/3, ∆(2)
In addition to expected distances, the associated variances will also at one
point be useful. Analytical expressions for most of them seem to have remained
unknown thus far, but the underlying probability densities have been found to
be more concentrated around the expected values given above as D grows [29].
That is, variance is greatest for D = 1.
4
2.2 Network algorithmics
We view the N nodes as running an asynchronous message-passing algorithm
collectively. By asynchronous we mean that each node remains idle until a
message arrives. When this happens, the arriving message is processed by the
node, which may result in messages being sent out as well. Such a purely reactive
stance on the part of the nodes requires at least one node to send out at least
one message without having received one, for startup. We assume that this is
done by all nodes initially, after which they start behaving reactively.
We assume that each message carries a signed-unit tag (i.e., either +1 or −1)
with it. The specific tag to go with the message is chosen probabilistically by its
sender at send time, with −1 being chosen with probability p−. The processing
done by node i upon arrival of a message is the heart of the system's thresholding
nature and involves manipulating an accumulator Ai, initially equal to 0, to
which every tag received is added (unless Ai = 0 and the tag is −1, in which
case Ai remains unchanged). Whenever Ai reaches a preestablished integer
value τ > 0, node i sends out messages of its own and Ai is reset to 0. Thus,
the integer τ acts as a threshold governing the sending of messages by (the firing
of) node i. The values of p− and τ are the same for all nodes.
It follows from this simple rule that the value of Ai is perpetually confined to
the interval [0, τ ]. The expected number of messages that node i has to receive
in order for Ai to be increased all the way from 0 to τ is the node's expected
number of message arrivals between firings, henceforth denoted by µ = µ(p−, τ ).
The value of µ can be calculated easily once we recognize that µ is simply the
expected number of steps for the following discrete-time Markov chain to reach
state τ having started at state 0. The chain has states 0, . . . , τ and transition
probability prs, from state r to state s, given by
1 − p− if s = r + 1;
p−
1
0
if s + 1 = r < τ or r = s = 0;
if r = s = τ ;
otherwise.
prs =
It is easily solved and yields
µ =
p−
(1 − 2p−)2 (cid:20)(cid:18) p−
1 − p−(cid:19)τ
− 1(cid:21) +
τ
1 − 2p−
(5)
(6)
for p− 6= 0.5 [30, page 348].1
The sending of messages when a node fires is based on another parameter,
p→, which is the probability with which the node sends a message to each of
the other N − 1 nodes. It follows that the expected number of messages that
get sent out is (N − 1)p→. The value of p→ is the same for all nodes as well.
1For p− = 0.5 we have µ = τ (τ + 1) [30, page 349], but this holds for none of our
computational experiments.
5
2.3 Local imbalance and global message traffic
At node i, a balance exists between message output and message input when
the expected number of messages sent out at each firing is the same as the
expected number of messages received between two successive firings. That is,
message traffic is locally balanced when (N − 1)p→ = µ. It is locally imbalanced
otherwise, which can be quantified by the difference I, defined to be
I =
(N − 1)p→
µ
− 1.
(7)
Given I, clearly the instantaneous density of global message output at time
t, denoted by M (t), is expected to remain constant with t if I = 0, or to
decrease or increase exponentially with t depending on whether I < 0 or I > 0,
respectively. This behavior is described by
dM (t)
dt
=
I
T0
M (t),
(8)
where the T0 participating in the time constant T0/I is some fundamental
amount of time related to the system's geometry and kinetics.
In Section 5,
we provide empirical evidence that T0 is the expected delay undergone by a
message, given by
T0 =
∆(D)
ℓ
σ
.
Equation (8) is of immediate solution, yielding
where
M (t) = M0eIt/T0,
M0 = M (0) = N (N − 1)p→
(9)
(10)
(11)
is the expected number of messages that all nodes, collectively, send out initially.
Similarly, the cumulative global message output inside a temporal window
of duration w starting at time t is
Mt(w) = Z t+w
t
M (t)dt =
M0T0
I
(cid:16)eI(t+w)/T0 − eIt/T0(cid:17) .
(12)
In the case of locally balanced message traffic (I = 0), this expression is easily
seen to yield
Mt(w) = M0w,
(13)
therefore independent of t. Otherwise, Mt(w) either decreases or increases ex-
ponentially with t, depending respectively on whether I < 0 or I > 0.
6
2.4 Graph-theoretic interpretations
The model described so far in Section 2 can be regarded as a directed geometric
graph, that is, a graph whose nodes are positioned in some region of interest
(the D-dimensional cube of side ℓ) and whose edges are directed. It is moreover
a complete graph without self-loops, in the sense that an edge exists directed
from every node i to every node j 6= i.
Our use of the model in the sequel will require the nodes to be positioned at
random before each new run of the distributed algorithm of Section 2.2, so an
alternative interpretation that one might wish to consider views the model as a
variation of the traditional random geometric graph [31]. In this variation, an
edge exists directed from i to j 6= i with fixed probability p→, independently of
any other node pair. That is, aside from node positioning the graph underlying
our model is an Erdos-R´enyi random graph [32] as extended to the directed
case [33].
This interpretation is somewhat loose, though, because it requires that we
view each individual run of the algorithm as being equivalent to several runs on
independent instances of the underlying random graph, with multiple further
runs serving to validate any statistics that one may come up with at the end.
This is hard to justify, however, particularly when one considers the nonlinear-
ities characterizing the quantities we will average over all runs of the algorithm
(see Section 3). Even so, interpreting our model in terms of random graphs
remains tantalizing in some contexts. For example, it allows the parameter p−
to be regarded as the fraction of a node's in-neighbors from which messages
with negative tags are received. In the context of networks of the brain at the
neuronal level, for example, an abstract rendering of the fraction of neurons
that are inhibitory is obtained (see Section 6.1).
3 Total correlation
We use the total correlation of N random variables [22], each corresponding to
one of the N nodes, as a measure of information integration. Each of these
variables is relative to a temporal window of fixed duration w, the variable
corresponding to node i being denoted by X (w)
and taking up values from the
set {0, 1}. The intended semantics is that X (w)
= 1 if and only if node i
receives at least one message in a time interval of duration w. We also use the
shorthands X(w) and x to denote the sequence of variables (X (w)
, . . . , X (w)
N )
and the sequence of values (x1, . . . , xN ), respectively. Thus, X(w) = x stands
for the joint valuation X (w)
1 = x1, . . . , X (w)
N = xN .
1
i
i
Given the marginal Shannon entropy of each variable X (w)
,
i
H(X (w)
i
) = − Xx∈{0,1}
Pr(X (w)
i = x) log2 Pr(X (w)
i = x),
(14)
7
and the joint Shannon entropy
H(X(w)) = − Xx∈{0,1}N
Pr(X(w) = x) log2 Pr(X(w) = x),
(15)
the total correlation of the N variables given w is defined as2
C(X(w)) =
N
Xi=1
H(X (w)
i
) − H(X(w)).
(16)
To see the significance of this definition in our context, consider the flat
joint probability mass function, Pr(X(w) = x) = 2−N for all x ∈ {0, 1}N. This
mass function entails maximum uncertainty of the variables' values, hence the
maximum possible value of the joint entropy, H(X(w)) = N .
It also implies
flat marginals, Pr(X (w)
i = x) = 0.5 for all x ∈ {0, 1} and all i, and again the
maximum possible value of each marginal entropy, H(X (w)
) = 1. The difference
from the actual joint entropy H(X(w)) to its maximum reflects a reduction of
uncertainty, or an information gain,
i
G(X(w)) = H(X(w)) − H(X(w)) = N − H(X(w)),
the same holding for each of the marginals,
G(X (w)
i
) = H(X (w)
i
) − H(X (w)
i
) = 1 − H(X (w)
i
).
(17)
(18)
Thus, it is possible to rewrite the expression for C(X(w)) in such a way that
G(X(w)) = C(X(w)) +
N
Xi=1
G(X (w)
i
).
(19)
That is, the total correlation of all N variables is the information gain that
surpasses their combined individual gains. This surplus is zero if and only if the
variables are independent of one another, that is, precisely when Pr(X(w) = x) =
i = xi) for all x ∈ {0, 1}N , since in this case we have H(X(w)) =
). It is strictly positive otherwise, with a maximum possible value
i
i=1 Pr(X (w)
QN
i=1 H(X (w)
PN
of N − 1.
Achieving this maximum requires a joint probability mass function assigning
no mass to any but two members of {0, 1}N , say x and y, and moreover that
these two be equiprobable (i.e., Pr(X(w) = x) = Pr(X(w) = y) = 0.5) and
complementary to each other (i.e., xi + yi = 1 for every i). Referring back to
the intended meaning of the N random variables, total correlation is maximized
in those runs of the distributed algorithm of Section 2.2 for which a partition
(X1, X2) of the set {X (w)
N } exists with the following two properties.
, . . . , X (w)
1
2The N = 2 case of this formula coincides with that for mutual information, but one is to
note that in the general case the two formulas are completely different [34].
8
First, no matter which particular window of duration w we concentrate on, the
set of nodes that receive at least one message inside the window is either X1 or
X2. Second, the first property holds with X1 for exactly half such windows.
While these are exceedingly stringent conditions both spatially and tempo-
rally, perhaps implying that values of total correlation equal to or near N −1 are
practically unachievable, they serve to delineate those scenarios with a chance of
generating substantial amounts of total correlation. Specifically, such scenarios
will on average have a pattern of global message traffic, inside a window, that is
neither too sparse nor too dense. Furthermore, sustaining such an amount of to-
tal correlation as time elapses will also require traffic patterns that deviate only
negligibly from the ones yielding the average, possibly entailing some variability
on window sizes. Our methodology to track and validate values of w leading
to noteworthy total correlation is described next. It involves computational ex-
periments for a variety of values for w, and also gauging the cumulative global
message output Mt(w) that results from each experiment against a function of
the reference number of messages and reference delay embodied in M0 and T0,
respectively.
4 Methods
Our results are based on running the distributed algorithm of Section 2.2 for
a fixed geometry (i.e., fixed number of dimensions D ∈ {1, 2, 3}, fixed value of
the cube side ℓ, and fixed positioning of N nodes in the cube) and a fixed set of
values for the parameters (p−, τ , and p→). Each run of the algorithm terminates
either when no more message is in transit (so none will ever be thenceforth, given
the algorithm's reactive nature) or when a preestablished maximum number of
messages in transit has been reached, whichever comes first. Imposing the latter
upper bound is important because it serves to size the data structures where
messages in transit are kept for later processing.
Node positioning is achieved uniformly at random, so multiple runs are
needed for each fixed configuration (D, ℓ, N, p−, τ, p→). Each run leaves a trace
of all events taking place as it unfolds, each event referring to the arrival of a
message at a node and comprising the node's identification and the message's
arrival time. A series of values for w is then considered and for each w each of the
traces is analyzed, yielding the total correlation produced by the corresponding
run. The average total correlation over all the runs is then reported.
Following our discussion at the end of Section 3, for each value of w we gauge
Eq. (12) against the approximation to it given by M0(T0) = M0T0, according
to which M0 messages get sent at time t = 0 and received at time T0. We do
this by postulating a proportionality constant α > 0 between them, that is, by
assuming
Mt(w) = αM0T0.
(20)
Doing this allows us to express w as a function of α for each t and, whenever
possible, to characterize traffic regimes giving rise to substantial amounts of
total correlation.
9
4.1 Supporting analysis
We denote the value of t upon termination of a run by T and the total number of
messages sent by M . An approximation to Eq. (12) similar to the one above can
be used to relate T and M as M0(T ) = M T0, whose right-hand side quantifies
what would be expected to happen if all M messages were sent at time t = 0
and received at time T0. This leads to
T =
T0
I
ln(cid:18)1 +
M I
M0 (cid:19) .
Solving Eq. (20) for w = w(t) given α yields
w(t) =
T0
I
ln(cid:16)1 + αIe−It/T0(cid:17) ,
whose value for t = 0 is the duration of the first window,
w0 = w(0) =
T0
I
ln(1 + αI).
(21)
(22)
(23)
As for the duration of the last window, which we denote by w∗, it can likewise
be found by solving Eq. (20) for w, now letting the window's start time be
t = T − w, then letting w∗ = w. We obtain
w∗ =
T0
I
ln(cid:18) 1 + M I/M0
1 + (M/M0 − α)I(cid:19) .
The average window duration between time t = 0 and time
T − w∗ =
T0
I
ln(cid:18)1 +(cid:18) M
M0
− α(cid:19) I(cid:19) ,
denoted by ¯w, is also of interest and comes from the indefinite integral
I (cid:19)2
f (t) = Z w(t)dt = (cid:18) T0
Li2(cid:16)−αIe−It/T0(cid:17) ,
(24)
(25)
(26)
where Li2(z) = Pk≥1 zk/k2 is the dilogarithm of z. This given, we obtain
¯w =
f (T − w∗) − f (0)
T − w∗
=
T0
I
Li2(−αI/[1 + (M/M0 − α) I]) − Li2(−αI)
ln(1 + (M/M0 − α) I)
.
(27)
While for I = 0 we have T = M T0/M0 and
w0 = ¯w = w∗ = αT0,
(28)
for I 6= 0 everything depends on the sign of I. If I < 0, then we need M/M0 <
−1/I in order for T to be well defined. Moreover, we have
0 < w0 < ¯w < w∗ < T,
(29)
10
Table 1: Parameter values for each of settings I -- IV and each of the three possibil-
ities for the value of I in relation to 0. Values of p→ are given by p→ = µ/(N −1)
for I = 0, as per Eq. (7).
Setting N
τ
p−
I
II
III
IV
500
300
500
500
0.30
0.30
0.18
0.30
5
5
5
10
I < 0
0.01
0.01
0.01
0.01
p→
I = 0
0.02135
0.03563
0.01478
0.04634
I > 0
0.06
0.06
0.06
0.06
where the first inequality holds if α < M/M0, this being necessary and sufficient
for the last inequality to hold as well. For I > 0, on the other hand, T is always
well defined and we get
T > w0 > ¯w > w∗ > 0.
(30)
In this case, the constraint α < M/M0 is necessary and sufficient for both the
first and the last inequality to hold.
4.2 Computational experiments
We organize our computational experiments into settings, numbered I -- IV, each
comprising all configurations (D, ℓ, N, p−, τ, p→) for which N , p−, and τ are
fixed.
In each setting, there are three possibilities for the value of p→, one
ensuring I < 0 (p→ = 0.01), one for I = 0 (extracted from Eq. (7)), and one
for I > 0 (p→ = 0.06). The four settings are summarized in Table 1. Each of
settings II -- IV is derived from setting I by a change in the value of N , p−, or τ ,
respectively.
Each setting entails 2 100 runs of the algorithm for each p→ value, this total
comprising 100 independent trials for each combination of D ∈ {1, 2, 3} and ℓ ∈
{10−3, 10−2, . . . , 103}. Each run starts by placing the N nodes anew, uniformly
at random, and proceeds therefrom subject to the further indeterminacies that
characterize the algorithm through the probabilities p− and p→. Messages are
assumed to travel at the speed σ = 1. Each of the 100 traces resulting from the
same configuration is then analyzed and the corresponding value of C(X(w)) is
computed, the average over the traces being reported at the end. This is done
for each w ∈ {2−33, 2−32.5, . . . , 210}.
One last important aspect of a run has to do with the determination of the
maximum number of messages ever allowed to be in transit as the run unfolds.
We determine this number via the formula 1 000N µ, where 1 000µ can be in-
terpreted as the expected number of messages a node must receive if it is to
fire 1 000 times during the run. This latter number is in many ways arbitrary,
though, having to do only with ensuring that the computational resources re-
quired by all runs remain manageable. As a result, taken as a whole the 25 200
runs have generated a total of about 394 gigabytes of trace in compressed format.
Moreover, processing these data for the determination of the various C(X(w))
11
averages has required several weeks of computation on an Intel Xeon E5-1650
core clocked at 3.2 GHz and having exclusive access to 30 gigabytes of RAM.
5 Results
The average value of the total correlation C(X(w)) resulting from the compu-
tational experiments described in Section 4.2 are presented in normalized form
(i.e., relative to the maximum N − 1) in Figures 1, 2, and 3, respectively for
the I < 0 cases, the I = 0 cases, and the I > 0 cases. Each figure comprises
four panels, numbered I -- IV to match the four settings of Table 1. Each panel
contains 21 plots, each corresponding to one of the possible variations in the
number of dimensions D and in the cube side ℓ.
Each plot in the three figures is given against the rescaled version of the
window size w given by w/T0, where we recall from Eq. (9) that T0 is the
delay a message incurs when traversing the distance ∆(D)
at speed σ. That
is, the abscissas in Figures 1 -- 3 are all relative to the geometric and kinetic
underpinnings to which each plot refers. This rescaling reveals that, for each
fixed combination of a setting with an imbalance profile (i.e., for each panel in
the three figures), the behavior of the average C(X(w)) is essentially invariant
with respect to how the various D and ℓ values are paired. Our choice in
Section 2.3 of T0/I as the system's time constant is then clearly justified, as
the value of I is fixed in each of the figures' panels. Moreover, such invariance
also backs up our choice of σ = 1 for all simulations (see Section 4.2), since the
choice of any other value would simply alter the rescaling factors.
ℓ
Nevertheless, there are signs in Figures 3(I) and (III) that such invariance
may not hold quite as well when D = 1. We attribute this to the fact that ∆(D)
is only an expected distance and as such affects the role of T0 as a rescaling
factor differently for each number of dimensions D. In particular, recalling from
Section 2.1 that the corresponding variances get larger as D is decreased, it
seems clear that what is affecting invariance in the two figures in question is
precisely the poor representativeness of ∆(1)
. Even so, we see in Figure 3 that
this affects setting III much more severely than it does setting I. The reason
for this has to do with the value of imbalance I in each case, as we discuss in
Section 6.
ℓ
ℓ
In all but the case of Figure 1(IV), the average total correlation starts off at
some negligibly small value, then slowly climbs toward a peak along an increase
in w/T0 by some orders of magnitude, and finally recedes as w/T0 is made
to vary by further orders of magnitude. The plots in each of the panels of
Figures 1 -- 3 are given against a backdrop of two vertical lines, the leftmost one
marking the smallest w/T0 at which total correlation peaks for some (D, ℓ) pair,
the rightmost marking the largest such value. (The smallest w/T0 value in the
setting-III panel of Figure 1 does not take into account the cases D = 2, 3 with
ℓ = 103, whose peaks seem to occur past w = 210, the largest window used in
the simulations, and are therefore unknown.)
12
I
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
II
l
n
o
i
t
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
III
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
IV
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
10-4
0
10-5
10-2
10-3
100
Rescaled window size
10-1
101
102
Figure 1: Normalized total correlation (C(X(w))/(N − 1)) against rescaled win-
dow size (w/T0) for the I < 0 cases.
In the keys to
the plots, u = log10 ℓ. Vertical lines mark the smallest and largest w/T0 values
given in Table 2 (the smallest such value for setting IV is off range).
I -- IV identify settings.
13
I
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
II
l
n
o
i
t
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
III
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
IV
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
10-4
0
10-5
10-2
10-3
100
Rescaled window size
10-1
101
102
Figure 2: Normalized total correlation (C(X(w))/(N − 1)) against rescaled win-
dow size (w/T0) for the I = 0 cases.
In the keys to
the plots, u = log10 ℓ. Vertical lines mark the smallest and largest w/T0 values
given in Table 2.
I -- IV identify settings.
14
I
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
II
l
n
o
i
t
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
III
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
IV
n
o
i
t
l
a
e
r
r
o
c
l
a
t
o
t
.
m
r
o
N
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
10-5
10-4
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 1, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 2, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
D = 3, u = -3
u = -2
u = -1
u = 0
u = 1
u = 2
u = 3
101
10-2
10-3
100
Rescaled window size
10-1
102
Figure 3: Normalized total correlation (C(X(w))/(N − 1)) against rescaled win-
dow size (w/T0) for the I > 0 cases.
In the keys to
the plots, u = log10 ℓ. Vertical lines mark the smallest and largest w/T0 values
given in Table 2.
I -- IV identify settings.
15
The backdrop lines' locations are detailed in Table 2, where the (D, ℓ) giving
rise to each one is also shown. The table also contains for each such location a
value for α, the proportionality constant that through Eq. (20) is used to relate
the global message output Mt(w) inside a size-w window beginning at time t
to the reference output M0T0. For I = 0, we have found in Section 4.1 that
the w = w(α) necessary for Eq. (20) to hold is time-invariant and given as in
Eq. (28), whence α = w/T0. This is reflected in Table 2.
For I 6= 0, on the other hand, satisfying Eq. (20) for some fixed α requires
w to vary with time according to Eq. (22). This is what one would expect,
since the dwindling message output that results from an I < 0 scenario re-
quires an ever larger window size to accommodate the amount of traffic given
by αM0T0, and likewise the output expansion caused by I > 0 requires pro-
gressively smaller window sizes. This dependency on time is reflected in the
inequalities of Eqs. (29) and (30), respectively for I < 0 and I > 0, where the
earliest window size (w0), as well as the average one ( ¯w) and the latest (w∗) are
put in perspective. The values of α given in Table 2 for the I 6= 0 cases come
from Eq. (27) by letting each value of w/T0 reported in the table be such that
w = ¯w(α).
This use of Eq. (27) requires the expected value of M (the number of mes-
sages sent in a run) to be estimated from the 100 independent trials for each
configuration of interest. We do this by first averaging the number of messages
received during each run (let E denote this average) and then recalling that
each message received entails an expected number of messages sent equal to
(N − 1)p→/µ. Our estimate of the expected M is then3
M = M0 +
E(N − 1)p→
µ
.
(31)
An illustration highlighting the use of Eq. (27) on the setting-III cases of Table 2
for I 6= 0 is given in Figure 4, where the plots relative to Eq. (24) rely on the
same estimate of the expected M as above.
6 Discussion
Even though our notation for total correlation, C(X(w)), may suggest that it
is a function solely of the N random variables in X(w), it is in fact a function
of the joint probability mass function associated with those variables as well.
Therefore, changing the allotment of probability mass to the various points in
{0, 1}N must have an impact on the value of C(X(w)) as a matter of principle.
Thus, while we know that such variation must never lead total correlation to
fall below zero or above N − 1, figuring out the resulting expected value has the
3Our simulator is event-based, each event being the reception of a message. The total we
have available at the end of each simulation is that of messages received, not sent, and the
two may differ because not every message sent gets received, by virtue of the preestablished
maximum number of messages in transit having been reached (see Section 4). We then need
this workaround to find the expected M through E.
16
Table 2: Smallest and largest w/T0 at which C(X(w)) peaks for each combination of a setting with a possibility for I relative
to 0, all configurations in each combination considered. Each of the two extremes occurs for the configuration to which the
D and ℓ values shown refer. For each of the two extremes, the value of α reported alongside it is such that w = ¯w(α). The
smallest w/T0 reported for the I < 0, setting-III case does not take into account the peaks for D = 2, 3 with ℓ = 103, which
have remained undiscovered by the simulations.
1
7
Imbalance
Setting
I < 0
I = 0
I > 0
I
II
III
IV
I
II
III
IV
I
II
III
IV
ℓ
1
α
10−3
10−2
102
103
10−3
w/T0
1.47582
0.41743
2.45490
2 × 10−13
Smallest w/T0 for C(X(w)) peak
D
3
3
2
2
3
1
2
1
1
1
1
1
0.710
0.116
1.395
0.050
0.522
0.530
0.530
0.586
0.009
0.008
0.009
0.007
0.00129
0.00114
0.00129
0.00101
10−2
10−2
10−1
10−2
10−1
10−3
= α
= α
= α
= α
α
w/T0
2.17223
0.84853
3.41955
0.18729
Largest w/T0 for C(X(w)) peak
D
1
1
3
2
2
3
2
1
3
2
3
3
1.073
0.302
1.917
0.050
0.848
0.774
0.848
0.960
0.015
0.012
0.017
0.010
ℓ
103
10
10
10−2
10−1
103
10−1
102
103
102
10
1
0.00214
0.00170
0.00236
0.00148
= α
= α
= α
= α
A
i
i
e
z
s
w
o
d
n
w
d
e
a
c
s
e
R
l
B
i
i
e
z
s
w
o
d
n
w
d
e
a
c
s
e
R
l
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
10-1
10-2
10-3
10-4
10-5
10-6
10-7
0
0.5
1
0
0.004
0.008
Eq. (23)
Eq. (27)
Eq. (24)
1.5
2
2.5
Eq. (23)
Eq. (27)
Eq. (24)
0.012
0.016
0.02
α
α
Figure 4: Rescaled window size (w/T0) against α, as per Eqs. (23), (27),
and (24), respectively for w = w0, w = ¯w, and w = w∗. All plots refer to
setting III, with either I < 0 (A) or I > 0 (B). The corresponding (α, w/T0)
pairs in Table 2 are highlighted as bullets. Plot ranges are given by the inter-
vals [0.05, M/M0) in A (to within a difference of 10−4 on the upper side) and
[0.001, 0.019] in B.
18
potential of aiding in the interpretation of any particular figure one encounters
in some situation of interest. As far as we know, such expected value has so far
remained unknown, but resorting to results on the expected value of the joint
Shannon entropy (H(X(w)), in our case) [35], it is possible to ascertain that an
upper bound on the expected value of C(X(w)) is approximately 0.6 [20].
Once normalized to N − 1, this upper bound would be unnoticeable in any of
the panels in Figures 1 -- 3. In fact, it would fall two orders of magnitude below the
nearly flat values of Figure 1(IV), which are of the order of 10−1. Clearly, then,
in all of settings I -- IV and for all three imbalance scenarios, our model is seen to
be promoting the rise of probability mass functions leading to total correlations
significantly above the expected value. In most cases, this occurs over ranges
for the value of w, the temporal-window duration that provides meaning to the
random variables in X(w), spanning several orders of magnitude.
Not only this, but in most cases we have found normalized total correlation
to peak at substantial levels: between 0.22 and 0.67 in settings I and III when
I < 0 (Figure 1); 0.77 and 0.86 in all four settings when I = 0 (Figure 2); and
0.93 and 0.97 in all four settings when I > 0 (Figure 3). The peaks relating
to the I = 0 cases (and also those of the I > 0 cases, but to a more limited
extent) seem to occur largely independently of the setting in question, that is,
regardless of the number of nodes N or the parameters that control firing (the
probability p− and the threshold τ ).
Understanding this independence comes from focusing on the I = 0 cases,
since for I = 0 the value of probability p→ varies from setting to setting as
a function of N and µ (therefore a function of p− and τ as well, by Eq. (6));
see Table 1. Setting p→ in this way aims precisely to cause I = 0 and, as a
consequence, tends to compensate for any dependency of total-correlation peaks
on N , p−, or τ . It also suggests that the value of I, along with that of w/T0, is
a major player when it comes to determining how much total correlation can be
achieved. However, increasing the value of p→ from those used to ensure I = 0
to p→ = 0.06, thus ensuring I > 0, preserves nearly the same independence of
peak values as when I = 0 but fails to keep I constant from setting to setting
(see Table 3 for the specific value of I in each one). With the increased value of
p→ and the ever-growing barrage of messages that ensues, this independence is
now supported by window sizes at least two orders of magnitude lower (though
yielding taller peaks) but not significantly by the value of I itself. The value of
I, however, does make its influence felt, in the following manner. As mentioned
in Section 5, rescaling window size w through a division by T0 makes some of the
I > 0 cases stand out for D = 1 by failing to comply (though to a small degree)
with the invariance to the values of D and ℓ that seems to be the rule. In that
section we correctly attributed this to the increased variance of T0 (through that
of ∆(1)
ℓ ), but the deviation from invariance clearly increases with the value of I
as well (1.81 and 3.06, by Table 3, respectively for settings I and III).
On the other hand, the cases for which I < 0 are different in regard to
this independence issue. In fact, it is clear from Figure 1 that total-correlation
peaks are highly affected by the setting in question.
In particular, they are
19
Table 3: Nonzero imbalance values for each of settings I -- IV.
Setting
I for p→ = 0.01
I for p→ = 0.06
I
II
III
IV
−0.53
−0.72
−0.32
−0.78
1.81
0.68
3.06
0.29
increasingly lower than the corresponding peaks for I = 0, and also occur for
increasingly lower values of w/T0, as the value of I (see Table 3) becomes more
negative. For fixed T0, therefore, this consolidates the value of I as the main
driver defining the window duration for which total correlation peaks, provided
I ≤ 0.
The I = 0 cases are particularly interesting also because total correlation was
in all configurations of all settings found to peak roughly between w/T0 = 0.522
and w/T0 = 0.96 (see Table 2). As we noted in Section 5, this is also the
range of α, which for I = 0 is the fraction of the reference message output M0T0
yielding the message traffic inside a window of duration w beginning, essentially,
at any time t ≥ 0. That total correlation should peak with α between roughly
0.522 and 0.96 seems well aligned with our discussion at the end of Section 3
on maximizing total correlation, through which we found out that a condition
necessary (though not sufficient) to such maximization is that only half the
nodes receive messages inside the window.
Relativizing message traffic inside the window to M0T0 in the I 6= 0 cases
must be approached differently, because now fixing α requires the duration of
the window starting at time t to either expand with t (if I < 0) or shrink with
it (if I > 0). In these cases, taking each w in Table 2 to be the average duration
of these windows reveals that total correlation peaks for w/T0 between about
1.47582 and 3.41955 for I < 0 (therefore somewhat above the range recorded for
the I = 0 cases), ignoring the more or less degenerate cases of settings II and IV,
and between about 0.00101 and 0.00236 for I > 0 (therefore substantially below
the I = 0 ranges, as noted above). The corresponding values of α, given by
Eq. (27) and illustrated in Figure 4 for setting III, range between 0.71 and
1.917 for I < 0, between 0.007 and 0.017 for I > 0.
6.1 What of the brain?
Right at the opening of Section 1 we mentioned the brain as the most represen-
tative threshold-based system we know. It is only fitting, then, that we should
try and relate it more closely to the model we have developed and analyzed in
this paper. This is no simple task for at least two reasons. The first has to
do with the model itself, which as we also mentioned in Section 1 has not been
meant to faithfully represent the specifics of any one threshold-based system.
The second reason is that very little has so far been found out of the brain's so-
20
called microconnectome [36], that is, its network structure at the cellular level,
which is where thresholds work their influence.
Here we resort to the few glimpses that are available in order to estimate
parameter values within our model as well as possible. The largest microcon-
nectome to have been mapped to date is from the V1 area of the mouse visual
cortex [37].
It reveals a total of 1 278 excitatory neurons and 581 inhibitory
neurons, which following our discussion in Section 2.4 allows as to conclude
p− ≈ 581/(581 + 1 278) ≈ 0.31, thus agreeing with the commonly accepted
range of values for the ratio of inhibitory neurons [38]. It also reveals a total of
29 connected pairs involving 45 neurons, which again in the spirit of Section 2.4
leads to p→ ≈ 29/(45 × 44) ≈ 0.015. Using N = 45 and τ = 7.5 (halfway
between 5 and 10, which seem to delimit the range of the typical neuron's ratio
of threshold potential to synaptic potential [39 -- 41]), we obtain an imbalance
I = −0.96, close therefore to the minimum possible imbalance (I = −1).
Note that concluding p→ ≈ 0.015 reflects an effective, rather than merely
anatomical, view of the microconnectome. That is, the 29 axon-to-dendrite
connections leading to this estimate of the value of p→ do not result simply
from the possibility that action potentials travel from a neuron's axon toward
one of another's dendrites, but rather from actually observing such potentials.4
This distinction is of paramount importance in the present study, since the
parameter p→ is semantically tied to the actual sending of messages in our
model (once interpreted as a directed graph, as in Section 2.4), not merely to
the existence of a channel through which such messages could be sent.
To judge from the values of I for p→ = 0.01 in Table 3, and from the
conclusions we have drawn regarding the I < 0 cases of settings I -- IV, this
entire attempt to interpret cortical activity in the light of our model seems to
be suggesting that we view it as a threshold-based system severely imbalanced
on the negative side, therefore incapable of giving rise to any significant amount
of total correlation. The missing component, of course, is that the cortex is
continually subjected to new input, either external or originating spontaneously
from within [45], hence a better characterization of local imbalance in this case
would be based on
− 1,
(32)
(N − 1)p→ + µ0
I ′ =
µ
with µ0 standing for some core amount of external or spontaneous input ex-
pected to accompany every µ messages incoming from other cortical neurons.
Clearly, µ0 > 0 implies I ′ > I whenever all else remains constant. Local imbal-
ance would then be defined through I ′ 6= 0, but without the availability of some
4To further strengthen this notion that the value of p→ is highly network-dependent in
terms of effective signaling in the brain, we note that a completely different estimate of p→ can
be deduced from a slightly earlier (but similar in the sense of targeting effective connections)
study [42], whose authors report a neuron's expected out-degree to be in the order of 100. The
study in question is based on the rodent somatosensory cortex, but should a similar conclusion
hold for the human cortex taken as a whole, with its 16 billion neurons [43], then an estimate
of p→ would place its value at the order of 10−10 and that of I very close indeed to −1. This
is all rather speculative, though, especially if we consider the already established fact that
different cortical areas of the primate brain have different neuronal densities [44].
21
rationale on which to base an estimate of µ0 this is essentially as far as we can
go. In any event, such redefined imbalance might lie above the I < 0 values of
Table 3 while still being negative.5 In this case, we could expect C(X(w)) to
peak for w/T0 a few orders of magnitude above 3.41955 (see Table 2).
We can go further and estimate the value of T0 as well. Because the cerebral
cortex takes up 82% of all brain mass in humans [43], and assuming a uniform
density in addition to a brain volume of 1.5 × 10−3 m3 [46, BNID 112053], we
can take the cortical volume to be roughly 1.23 × 10−3 m3. In our model, this
volume corresponds to the three-dimensional cube of side ℓ ≈ 0.107 m, whence
∆(3)
ℓ ≈ 0.071 m (see Section 2.1). Assuming further that action potentials
propagate on myelinated axons at a speed (σ) between 50 and 100 m/s [46, BNID
107125] [47] leads, by Eq. (9), to a value of T0 roughly between 0.71 and 1.42 ms.
This would place w, the time window duration for peak total correlation, a few
orders of magnitude above some value between 2.4 and 4.9 ms. Significant
total correlation also arises for w/T0 values below or above the peak's location
by a small factor, so window durations of a few hundred milliseconds would
support information integration as defined. These, as it turns out, are time
lapses within range of the commonly accepted delay before a percept can be
rendered conscious [48].
7 Conclusion
Our approach to quantifying the integration of information to which interacting,
distributed threshold-based units give rise has relied on a very general model,
tailored to no specific system in particular but built around three fundamen-
tal notions: that the units interact with one another by passing positively or
negatively tagged messages among them; that a unit sends messages out as a
function of how the tag balance of incoming messages relates to a threshold; and
that this sending out of messages is selective, in the sense of being addressable
to a subset of the units only. Assuming that the units are positioned in a one-,
two-, or three-dimensional cube uniformly at random, and moreover considering
a temporal window of duration w, we have demonstrated by means of extensive
computational experiments that information does get integrated in significant
amounts inside the window, depending chiefly on the value of w, on the average
delay incurred by a message, and on local message imbalance (how much gets
transmitted vis-`a-vis what is received, itself dependent on the combined effect
of the message-passing and threshold parameters). We have analyzed situations
of peak information integration and related the results to cortical dynamics by
fixing the model's parameters accordingly. This has served to suggest a vali-
dation of the model, despite its purposeful generality, highlighting its potential
usefulness for systemic studies of threshold-based interacting units.
5Letting I ∗ denote the greatest of these negative values for I, the purported I ′ such that
I ∗ < I ′ < 0 would require 0 < µ0 < µ − (N − 1)p→, assuming for the former of these
inequalities that N and all parameters remain unchanged between I ∗ and I ′.
22
Given the window duration w and the set of random variables X(w) (one
variable per unit, each related to the reception of messages by the unit in ques-
tion within the window), our measure of information integration has been the
total correlation of the variables in X(w), here denoted by C(X(w)). This mea-
sure seeks to quantify the interdependence of the variables on one another and
can be interpreted as information gain that is attained in excess of the total gain
the units achieve at the local level. Total correlation can also be interpreted as
the Kullback-Leibler (KL) divergence of one probability mass function relative
i = xi). That is,
to another, viz., of Pr(X(w) = x) relative to QN
C(X(w)) can be rewritten as
i=1 Pr(X (w)
C(X(w)) = Xx∈{0,1}N
Pr(X(w) = x) log2
Pr(X(w) = x)
i=1 Pr(X (w)
QN
i = xi)
.
(33)
This expression highlights the well-known fact that the KL divergence is asym-
metric with respect to the two mass functions it applies to. Note, however, that
this is of no import in our context, because total correlation corresponds to the
very specific case in which the KL divergence applies to the two mass functions
above and in the direction indicated only. We mention this issue because the
current version of the all-partitions theory of integrated information mentioned
in Section 1 abandons the KL divergence in favor of the so-called earth-mover's
distance because of the latter's symmetry and consequent status as a distance
between two probability mass functions [49]. This provides further distinction
between the two approaches.
We finalize by noting that extensions to our approach are certainly possible,
especially by attempting to encompass those systems that, despite embodying
explicit references to a threshold-based dynamics, do so in a manner that is
not completely aligned with the one our model assumes. One example involves
the role of thresholds in the grouping and synchronization of the interacting
units [50]. Perhaps our approach can provide useful insight in such contexts as
well.
Acknowledgments
The author acknowledges partial support from CNPq, CAPES, and a FAPERJ
BBP grant.
References
[1] M. F. Bear, B. W. Connors, and M. A. Paradiso. Neuroscience. Wolters
Kluwer, Philadelphia, PA, fourth edition, 2015.
[2] W. S. McCulloch and W. Pitts. A logical calculus of the ideas immanent
in nervous activity. B. Math. Biophys., 5:115 -- 133, 1943.
23
[3] A. Pistorello, C. Romoli, and S. Crespi-Reghizzi. Threshold nets and cell-
assemblies. Inform. Control, 49:239 -- 264, 1981.
[4] A. Ilachinski. Cellular Automata. World Scientific, Singapore, 2001.
[5] J. J. Hopfield. Neural networks and physical systems with emergent col-
lective computational abilities. Proc. Natl. Acad. Sci. USA, 79:2554 -- 2558,
1982.
[6] J. J. Hopfield. Neurons with graded response have collective computational
properties like those of two-state neurons. Proc. Natl. Acad. Sci. USA,
81:3088 -- 3092, 1984.
[7] D. H. Ackley, G. E. Hinton, and T. J. Sejnowski. A learning algorithm for
Boltzmann machines. Cognitive Sci., 9:147 -- 169, 1985.
[8] J. Pearl. Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann,
San Mateo, CA, 1988.
[9] T. Hrycej. Gibbs sampling in Bayesian networks. Artif. Intel., 46:351 -- 363,
1990.
[10] R. Kinderman and J. L. Snell. Markov Random Fields and Their Applica-
tions. American Mathematical Society, Providence, RI, 1980.
[11] S. Geman and D. Geman. Stochastic relaxation, Gibbs distributions, and
IEEE T. Pattern Anal., 6:721 -- 741,
the Bayesian restoration of images.
1984.
[12] V. C. Barbosa. Massively Parallel Models of Computation. Ellis Horwood,
Chichester, UK, 1993.
[13] D. Koller and N. Friedman. Probabilistic Graphical Models. The MIT Press,
Cambridge, MA, 2009.
[14] P. Smolensky. Information processing in dynamical systems: foundations of
harmony theory. In D. E. Rumelhart, J. L. McClelland, and PDP Research
Group, editors, Parallel Distributed Processing, volume 1, pages 194 -- 281.
The MIT Press, Cambridge, MA, 1986.
[15] A. Fischer and C. Igel. Training restricted Boltzmann machines: an intro-
duction. Pattern Recogn., 47:25 -- 39, 2014.
[16] G. E. Hinton. Deep belief networks. Scholarpedia, 4:5947, 2009.
[17] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Belle-
mare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen,
C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra,
S. Legg, and D. Hassabis. Human-level control through deep reinforcement
learning. Nature, 518:529 -- 533, 2015.
24
[18] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521:436 -- 444,
2015.
[19] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den
Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot,
S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lill-
icrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis. Master-
ing the game of Go with deep neural networks and tree search. Nature,
529:484 -- 489, 2016.
[20] A. Nathan and V. C. Barbosa. Network algorithmics and the emergence of
information integration in cortical models. Phys. Rev. E, 84:011904, 2011.
[21] K. K. Cassiano and V. C. Barbosa. Information integration in elementary
cellular automata. J. Cell. Autom., 10:235 -- 260, 2015.
[22] S. Watanabe. Information theoretical analysis of multivariate correlation.
IBM J. Res. Dev., 4:66 -- 82, 1960.
[23] D. W. Zhou, D. D. Mowrey, P. Tang, and Y. Xu. Percolation model of
sensory transmission and loss of consciousness under general anesthesia.
Phys. Rev. Lett., 115:108103, 2015.
[24] D. Balduzzi and G. Tononi.
Integrated information in discrete dynami-
cal systems: motivation and theoretical framework. PLoS Comput. Biol.,
4:e1000091, 2008.
[25] J. N. D. Kerr, D. Greenberg, and F. Helmchen. Imaging input and output of
neocortical networks in vivo. Proc. Natl. Acad. Sci. USA, 102:14063 -- 14068,
2005.
[26] S. Shoham, D. H. O'Connor, and R. Segev. How silent is the brain:
is
J. Comp. Physiol. A,
there a "dark matter" problem in neuroscience?
192:777 -- 784, 2006.
[27] A. M. Mathai, P. Moschopoulos, and G. Pederzoli. Random points associ-
ated with rectangles. Rend. Circ. Mat. Palermo, 48:163 -- 190, 1999.
[28] D. P. Robbins and T. S. Bolis. Average distance between two points in a
box. Am. Math. Mon., 85:277 -- 278, 1978.
[29] A. Zilinskas. On the distribution of the distance between two points in a
cube. Random Oper. Stoch. Equ., 11:21 -- 24, 2003.
[30] W. Feller. An Introduction to Probability Theory and Its Applications,
volume I. John Wiley & Sons, New York, NY, third edition, 1968.
[31] M. Penrose. Random Geometric Graphs. Oxford University Press, Oxford,
UK, 2003.
25
[32] P. Erdos and A. R´enyi. On random graphs. Publ. Math. (Debrecen), 6:290 --
297, 1959.
[33] R. M. Karp. The transitive closure of a random digraph. Random Struct.
Algor., 1:73 -- 93, 1990.
[34] T. S. Han. Multiple mutual informations and multiple interactions in fre-
quency data. Inform. Control, 46:26 -- 45, 1980.
[35] L. L. Campbell. Averaging entropy. IEEE T. Inform. Theory, 41:338 -- 339,
1995.
[36] J. M. L. Budd and Z. Kisv´arday. Communication and wiring in the cortical
connectome. Front. Neuroanat., 6:42, 2012.
[37] W.-C. A. Lee, V. Bonin, M. Reed, B. J. Graham, G. Hood, K. Glattfelder,
and R. C. Reid. Anatomy and function of an excitatory network in the
visual cortex. Nature, 532:370 -- 374, 2016.
[38] B. Rudy, G. Fishell, S. Lee, and J. Hjerling-Leffler. Three groups of in-
terneurons account for nearly 100% of neocortical GABAergic neurons.
Dev. Neurobiol., 71:45 -- 61, 2011.
[39] M. Steriade, I. Timofeev, and F. Grenier. Natural waking and sleep states:
a view from inside neocortical neurons. J. Neurophysiol., 85:1969 -- 1985,
2001.
[40] A. Hasenstaub, Y. Shu, B. Haider, U. Kraushaar, A. Duque, and D. A. Mc-
Cormick. Inhibitory postsynaptic potentials carry synchronized frequency
information in active cortical networks. Neuron, 47:423 -- 435, 2005.
[41] Y. Shu, A. Hasenstaub, A. Duque, Y. Yu, and D. A. McCormick. Modula-
tion of intracortical synaptic potentials by presynaptic somatic membrane
potential. Nature, 441:761 -- 765, 2006.
[42] M. Shimono and J. M. Beggs. Functional clusters, hubs, and communities
in the cortical microconnectome. Cereb. Cortex, 25:3743 -- 3757, 2015.
[43] S. Herculano-Houzel. The human brain in numbers: a linearly scaled-up
primate brain. Front. Hum. Neurosci., 3:31, 2009.
[44] C. E. Collins, D. C. Airey, N. A. Young, D. B. Leitch, and J. H. Kaas.
Neuron densities vary across and within cortical areas in primates. Proc.
Natl. Acad. Sci. USA, 107:15927 -- 15932, 2010.
[45] A. Luczak, P. Bartho, and K. D. Harris. Gating of sensory input by spon-
taneous cortical activity. J. Neurosci., 33:1684 -- 1695, 2013.
[46] R. Milo, P. Jorgensen, U. Moran, G. Weber, and M. Springer.
BioNumbers -- the database of key numbers in molecular and cell biology.
Nucl. Acids Res., 38:D750 -- D753, 2010.
26
[47] H. A. Swadlow and S. G. Waxman. Axonal conduction delays. Scholarpedia,
7:1451, 2012.
[48] M. H. Herzog, T. Kammer, and F. Scharnowski. Time slices: what is the
duration of a percept? PLoS Biol., 14:e1002433, 2016.
[49] M. Oizumi, L. Albantakis, and G. Tononi. From the phenomenology to
the mechanisms of consciousness: integrated information theory 3.0. PLoS
Comput. Biol., 10:e1003588, 2014.
[50] K. P. O'Keeffe, P. L. Krapivsky, and S. H. Strogatz. Synchronization as
aggregation: cluster kinetics of pulse-coupled oscillators. Phys. Rev. Lett.,
115:064101, 2015.
27
|
1807.10374 | 1 | 1807 | 2018-07-26T21:37:02 | Towards Identifying the Systems-Level Primitives of Cortex by In-Circuit Testing | [
"q-bio.NC"
] | The hypothesis considered here is that cognition is based on a small set of systems-level computational primitives that are defined at a level higher than single neurons. It is pointed out that for one such set of primitives, whose quantitative effectiveness has been demonstrated by analysis and computer simulation, emerging technologies for stimulation and recording are making it possible to test directly whether cortex is capable of performing them. | q-bio.NC | q-bio | Towards Identifying the Systems-Level Primitives of Cortex by In-Circuit Testing
Leslie G. Valiant
[email protected]
Harvard University
July 22, 2018
Summary
The hypothesis considered here is that cognition is based on a small set of systems-level
computational primitives that are defined at a level higher than single neurons. It is pointed out
that for one such set of primitives, whose quantitative effectiveness has been demonstrated by
analysis and computer simulation, emerging technologies for stimulation and recording are
making it possible to test directly whether cortex is capable of performing them.
1. Introduction
In both biology and computer science the virtue of modularity in complex systems is widely
understood. In contrast, the idea that complex systems need to have many levels of description,
while well appreciated in computer science, has been much less discussed in biology. We believe
that for a computational theory of cortex to be useful, it would need to give an account of
algorithmic tasks at levels intermediate between those of neurons and behavior. Just as we cannot
understand the workings of a cell phone by studying its operations only at the single bit level, we
cannot expect to understand the brain at the single neuron level.
The suggestion we explore here is that cognitive computation in cortex is built on a small
collection of base primitives at the systems level. This contrasts with Hebbian plasticity (Hebb,
1949), which is a single primitive at the single neuron level. Here we discuss one candidate for
such a collection of base primitives, one distinguished by the existence of demonstrations, by
analysis and computer simulations, that it offers a scheme in which the capacity of cortex for
basic cognitive tasks can be given a quantitative explanation (Valiant 1994, 2005, 2006; Feldman
& Valiant 2009). This collection consists of: association, memorization, inductive learning, and
hierarchical memory assignment, all suitably defined. The purpose of this note is to point out that
because of technological advances over the last decades, especially in optogenetics (Miyawaki et
al. 1997, Boyden et al. 2005, Deisseroth et al. 2006, Shemesh et al. 2017), it is now becoming
feasible to test systematically whether cortex can indeed perform these primitives.
1
These four primitives are defined below in terms of their actions on randsets, random sets of
neurons of a certain size. The nature of the proposed experiments is analogous to "in-circuit
testing" of a traditional electronic printed circuit board, where instead of testing the input-output
behavior of the whole board, the testing is done on smaller parts, using only probes. In the
proposed experiments there will be no sensory stimuli or behavioral responses -- the experiments
will consist entirely of stimulating and recording from appropriate sets of neurons. The goal is to
find out what changes in computational function in a subcircuit of cortex can be achieved by
specific neural stimulation. Existing experimental evidence that the firing patterns of cortical
neurons can be altered by suitable behavioral training (Jackson et al. 2006, Rebesco et al. 2010) is
relevant and encouraging, but does not directly address the question of which computational
primitives are being executed. Behaviors prompted by sensory inputs cause all kinds of activity in
the brain extraneous to the execution of any one primitive. The in-circuit methodology is
suggested to isolate the operation of a subcircuit as much as possible from such interference.
The goal therefore is to test, by direct neural stimulation and recording, whether cortex is
capable of performing certain basic computational tasks. The paradigmatic first such task is
"association of a set A to set B", abbreviated as A→B, which here is defined as follows: for
neuron sets A and B we want to induce the new functionality that, after the training, whenever set
A is stimulated then the newly modified circuit will cause set B to become active also. Training
here would involve stimulating A and B according to some timed protocol. Testing would involve
stimulating A again and now recording from the set B that had been stimulated in training. This
particular task may be viewed as a direct systems level analog of Hebbian plasticity.
We note that our notion of randsets is not the same as the notion of assemblies, in the
sense of Hebb (1949). That traditional notion of assemblies is associated with the idea that the
neurons in these assemblies are better connected to each other than to other neurons (e.g. Carillo-
Read et al. 2017). Randsets account for the claimed computational capabilities simply by virtue of
being randomly chosen, and probably with weak criteria of randomness being sufficient. There is
no requirement that they be well connected to each other. The theory basically asserts that for
random sets in a randomly connected network there will be a large enough bandwidth of reliable
communication from one randset to another, that simple local updates at the neurons that result
from training will be sufficient to realize the claimed functionalities. There is no comparable
theory that shows that the high interconnectivity assumption of traditional assemblies provides
any comparable computational capabilities. Fortunately, also, it is the randset theory that appears
to be one that is readily testable experimentally for the sets of operations we contemplate, since
random sets are more easily identified than highly interconnected ones. (Of course, it may be that
our randsets, once selected, do become more strongly connected to each other over time, but that
is not our concern here.)
An experiment relevant to what we are contemplating here is reported in Seeman et al.
(2017), where the setting up of such associations was attempted by electronic stimulation, but did
not succeed reliably. The experiment was done for one value of r. Possible interpretations are that
2
the value of r used there was too small, or that it was not a part of cortex that performs this
function. We are suggesting a systematic exploration of the relevant parameter space, and in
particular the set size r. Another relevant result, one obtained by optogenetic means, is by
Carrillo-Reid et al. (2016), which can be viewed as also an example of "in-circuit" testing, but the
task considered there was pattern completion, as in auto-associative nets (e.g. Hopfield, 1982).
We believe that pattern completion is a less fundamental operation for our purposes than the ones
we consider here, since there is no clear definition of what it should be, and any reasonable
system offers some version of it as an epiphenomenon.
2. The Four Tasks and their Potential Validation
Each instance of the tasks described here can be induced by an associated training
protocol, and is realized as local changes in the neurons within the given randomly connected
network, giving rise to a subcircuit having the functionality of that task instance. Computer
simulations (Feldman & Valiant, 2009) show that thousands of instances of such tasks can be
realized with realistic network parameters without having the continued effectiveness of the
earlier ones acquired degraded by later ones. We note that having multiple task types, which
between them may be adequate to form a basis for cognition, is a much more exacting
requirement than single-task models such as traditional "auto-associative nets" which perform the
single task of pattern completion. The longer term motivation is that of identifying a "cognitively
adequate'' set of primitives, perhaps our four forming the core, and showing that higher level
cognitive operations can be efficiently implemented in terms of these.
The following detail the four tasks that we are suggesting as the core, and for which we are here
suggesting validation by in-circuit testing. An example of some more detailed specifications still,
for one incarnation, is described by Feldman and Valiant (2009).
Association: For sets A, B train so that result is: If in future A is stimulated then B (but not random
other neurons) will become active.
Supervised Memorization of Conjunctions: For sets A, B, C train so that the result is: If in
future A and B are stimulated then C will become active also (but not if just one of A, B is
stimulated.)
Inductive Learning of Simple Threshold Functions: For sets A, B, C, D (say) train so that the
resulting circuit generalizes a classification at target A, beyond the examples seen, according to a
linear separator on the variables B, C, D. In particular, if in future some subset of B, C, D is
stimulated then the target A will become active according to whether some linear separator
consistent with the examples holds. For example such a linear separator may be B + C + 2D ≥ 2,
where now these variables have a 1/0 interpretation corresponding to whether the randsets are
active or not.
3
Hierarchical Memory Formation: For sets A, B, train so that result is: a set C (not specified by
the experimenter) has been allocated by the computation internally so that the effect "supervised
memorization of conjunctions" is achieved for A, B and that, now unknown, C. This realizes the
fundamental psychological task of "chunking", where a new compound concept, here A & B,
becomes equal citizen in the circuit with earlier ones.
While the first three of these tasks can be tested directly by in-circuit testing, the last task
appears more difficult. One possible approach is to glean information using the timing mechanism
at work.
Verifying how these tasks are realized in cortex, if indeed they are, is a further area of
challenge. The suggestion has been made (Valiant, 2012) that the hippocampal system is
responsible for determining the identity of the randset C in hierarchical memory formation, for
given A and B. Such a statement, which assigns responsibility for performing a particular task to a
particular brain area, may be easier to test.
3. General Considerations
How large sets? In the analysis the parameter r is dependent on the number of neurons in the
system, the number of connections, the synaptic strengths, and the algorithms used for realizing the
tasks. For the association task, for example, having r too small will fail to train the target B as
desired, while having r too large will train too many spurious neurons in addition. Hence, we expect
that considerable experimentation may be needed to find the right value of r in any one cortical
area. Note that in this case of associations, the size of the source set A appears to be critical, but this
is the easier set in that it only needs stimulation. The size of the target set B, which has to be both
stimulated and recorded from, may be much less critical.
Which brain areas? To test our theory one needs to perform the experiment in a cortical area
where the corresponding task is implemented in biology. We have behaving mammals in mind. It
may be that primates are the best at some of these tasks. There is some uncertainty as to which parts
of cortex perform these various tasks, and some experimentation there may be necessary.
How sets chosen? Since the hypothesis requires that the experiment work for randomly chosen sets
of a certain size in the appropriate brain areas, it is reasonable to go with whatever bias the
experimental technique used imposes. Neurons in particular layers may need to be found. The
theory aims to explain how large numbers of task instances can be processed. Hence, experiments
need to be done in parts of cortex that are large enough that it is plausible that so many tasks are
indeed supported. The various sets A, B, C, etc., may be in the same cortical area, or in different
areas.
What timing protocols for training? Timing may be critical. When stimulating a pair of sets A, B
for association we need an asymmetric protocol, such as B being stimulated with one millisecond
delay after A.
4
Comparisons? Comparative studies for different cortical areas and different species, including
humans, would be of great interest.
4. Conclusion
The hypothesis to be tested is that cortex is able to perform a specific set of tasks on random sets of
neurons. The new opportunity is offered by the remarkable recent advances in stimulation and
recording technologies. The experiments would determine whether these tasks can be performed in
appropriate cortices, but not how. However, positive results would demonstrate that the brain has
impressive computational capabilities that are relevant to cognitive computation. These systems-
level capabilities, if present, are unlikely to have arisen by accident. It would be a remarkable
coincidence if the brain had these capabilities but did not use them.
5. Acknowledgement
This work was supported in part by a grant from the National Science Foundation CCF 1509178.
6. References
Hebb, D. O. (1949). The Organization of Behavior: A Neuropsychological Theory. New York:
Wiley and Sons.
Valiant, L.G. (1994). Circuits of the Mind, Oxford University Press: New York.
Valiant, L.G. (2005). Memorization and association on a realistic neural model. Neural
Computation, 17:3, 527-555.
Valiant, L. G. (2006). A quantitative theory of neural computation. Biological Cybernetics, 95:3
205-211
Feldman, V. & Valiant, L. G. (2009). Experience-induced neural circuits that achieve high
capacity. Neural Computation, 21:10, 2715-54.
Miyawaki, A., Llopis, J., Heim, R., McCaffery, J. M., Adams, J. A., Ikura, M., and Tsien, R. Y.
(1997). Fluorescent indicators for Ca2+ based on green fluorescent proteins and calmodulin.
Nature. 388 (6645): 882 -- 887.
Boyden, E. S.; Zhang, F.; Bamberg, E.; Nagel, G.; Deisseroth, K. (2005). Millisecond-timescale,
genetically targeted optical control of neural activity. Nat. Neurosci. 8 (9): 1263 -- 8.
5
Deisseroth, K., Feng, G., Majewska, A. K., Miesenbock, G., Ting, A., and Schnitzer, M. J.
(2006). Next-Generation Optical Technologies for Illuminating Genetically Targeted Brain
Circuits. Journal of Neuroscience. 26 (41): 10380 -- 6.
Shemesh, Or A, Tanese, D., Zampini, V., Linghu, C., Piatkevich, K., Ronzitti, E.,
Papagiakoumou, E., Boyden, E.S., and Emiliani, V. (2017). Temporally precise single-cell
resolution optogenetics, Nature Neuroscience 20:1796 -- 1806.
Jackson, A., Mavoori, J. and Fetz, E. E., (2006) Long-term motor cortex plasticity induced by an
electronic neural implant. Nature 444, 56 -- 60.
Rebesco, J. M., Stevenson, I. H., Körding, K. P., Solla, S. A., and Miller, L. E., (2010) Rewiring
neural interactions by micro-stimulation, Frontiers in Systems Neuroscience, 4:39, pp. 1-15.
Carillo-Reid L., Yang, W., Miller, J. K., Peterka, D. S., and Yuste R. (2017) Imaging and
Optically Manipulating Neuronal Ensembles, Annual Review in Biophysics, 46:271-293.
Jackson, A., Mavoori, J. and Fetz, E. E., (2006) Long-term motor cortex plasticity induced by an
electronic neural implant. Nature 444, 56 -- 60.
Rebesco, J. M., Stevenson, I. H., Körding, K. P., Solla, S. A., and Miller, L. E., (2010) Rewiring
neural interactions by micro-stimulation, Frontiers in Systems Neuroscience, 4:39, pp. 1-15.
Seeman, S.C., Mogen, B.J., Fetz, E.E., and Perlmutter, S.I. (2017) Paired Stimulation for Spike-
Timing-Dependent Plasticity in Primate Sensorimotor Cortex, Journal of Neuroscience,
37(7):1935 -- 1949.
Carillo-Reid, L., Yang, W., Bando, Y., Peterka, D. S., and Yuste, R. (2016) Imprinting and
recalling cortical ensembles, Science 353:6300, 691-694.
Hopfield, J. J., (1982) Neural networks and physical systems with emergent collective
computational abilities", Proceedings of the National Academy of Sciences of the USA, 9:8, 2554 --
2558.
Valiant L. G. (2012) The hippocampus as a stable memory allocator for cortex, Neural
Computation, 24:11, 2873-2899.
6
|
1607.05530 | 2 | 1607 | 2016-09-15T21:58:12 | Verbal Perception and the Word Length Effect | [
"q-bio.NC"
] | A theoretical framework is proposed for the understanding of verbal perception -- the conversion of words into meaning, modeled as a compromise between lexical demands and contextual constraints -- and the theory is tested against experiments on short-term memory. The observation that lists of short words are recalled better than lists of long ones has been a long-standing subject of controversy, further complicated by the apparent inversion of the effect for mixed lists. In the framework here proposed, these behaviors emerge as an effect of the different level of localization of short and long words in semantic space. Events corresponding to the recognition of a nonlocal word have a clustering property in phase space, which facilitates associative retrieval. The standard word-length effect arises directly from this property, and the inverse effect from its breakdown. An analysis of data from the PEERS experiments (Healey and Kahana, 2016) confirms the main predictions of the theory. Further predictions are listed and new experiments are proposed. Finally, an interpretation of the above results is presented. | q-bio.NC | q-bio | Verbal Perception and the Word Length Effect
Francesco Fumarola
(Dated: October 6, 2018)
A theoretical framework is proposed for the understanding of verbal perception
-- the conversion of words into meaning, modeled as a compromise between lexical
demands and contextual constraints -- and the theory is tested against experiments
on short-term memory. The observation that lists of short words are recalled better
than lists of long ones has been a long-standing subject of controversy, further com-
plicated by the apparent inversion of the effect for mixed lists. In the framework here
proposed, these behaviors emerge as an effect of the different level of localization of
short and long words in semantic space. Events corresponding to the recognition of a
nonlocal word have a clustering property in phase space, which facilitates associative
retrieval. The standard word-length effect arises directly from this property, and the
inverse effect from its breakdown. An analysis of data from the PEERS experiments
(Healey and Kahana, 2016) confirms the main predictions of the theory. Further
predictions are listed and new experiments are proposed. Finally, an interpretation
of the above results is presented.
I.
INTRODUCTION
A. Free recall: lexical and serial-position effects
In 1894, with their pioneering work on free-recall experiments, Binet and Henry intro-
duced a key tool for the controlled investigation of short-term memory (Binet and Henry,
1894). In its traditional form, a free-recall experiment is performed by presenting the subject
with a list of words and then requesting him or her to recall it in any order (Murdock, 1960,
1962; Roberts, 1972; Standing, 1973).
Several types of effects have been reported:
1) Effects depending on the lexical properties of individual words. In particular, lists of
short words are recalled better than lists of long ones, a fact known in the literature as the
word-length effect (Baddeley et al., 1975; Russo and Grammatopoulou, 2003; Tehan and
6
1
0
2
p
e
S
5
1
]
.
C
N
o
i
b
-
q
[
2
v
0
3
5
5
0
.
7
0
6
1
:
v
i
X
r
a
2
Tolan, 2007; Bhatarah et al., 2009).
2) Effects in which the recall probability depends on the absolute position of words in
the list. It has been observed that the first and last words in the list are recalled more easily
("primacy" and "recency" effects).
3) Effects depending on the relative position of words with respect to each other. Most
notably, the recall probabilities of contiguous words correlate positively, a fact known as the
contiguity effect (Murdock, 1960, 1962).
The need to understand serial-position effects led to the devising of retrieved-context
models, such as the Temporal Context Model (Howard and Kanaha, 2002). In these models
the recall process, rather than retrieving a word directly, retrieves the context associated to
the word first.
Within this scenario, recency effects appear because the context at the time of the "mem-
ory test" is most similar to the context associated with recent items. When an item is re-
trieved, it reinstates the context active when that item was presented. Because this context
overlaps with the encoding context of the items' neighbors, a contiguity effect results.
Through these models, serial-position effects have been substantially understood over
the past fifteen years (Howard and Kahana, 2002, 2002b; Sederberg, Howard, and Kahana,
2008; Polyn, Norman, and Kahana, 2009, 2009b; Lohnas, Polyn, and Kahana, 2015; Kahana,
2012). The same cannot be said, however, about the word-length effect.
B. Riddles of the word length effect
The word-length effect (WLE) has been a traditional testing ground for models of short-
term memory (Campoy, 2011; Jalbert et al., 2011), and it has played a key role in establishing
the working-memory paradigm and the phonological loop hypothesis (Baddeley and Hitch.,
1974).
The standard account of the effect (Baddeley, 2007) relies on a trade-off between memory
decay (in the phonological store) and subvocal rehearsal via an articulatory control process.
Because shorter words take less time to rehearse, more decaying traces of them can be
refreshed than decaying traces of long items, and, therefore, more short items can be recalled.
This picture, however, is not able to account for all experimental observations concerning
this effect, and has been repeatedly called into question.
3
In (Neath et al., 2003), it was shown that with words having the same number of syllables
but different pronunciation times, no unambiguous WLE arises. This result (extended in
Jalbert et al. 2011) suggests that the effect depends on the number of syllables, and not on
the time it takes to pronounce them.
Experiments have also been performed in conditions where there was a delay between
lists, making subvocal rehearsal possible in the interval. No appreciable difference in recall
probabilities was found (Campoy, 2008).
In the same study, experiments were performed in which subvocal rehearsal was prevented
by a high presentation rate. No delay was allowed between the presentation of word lists
and the memory test. Yet, the WLE occurred unperturbed.
In the 2011 paper I just cited, Jalbert et al.
concluded: "the WLE may be better
explained by the differences in linguistic and lexical properties of short and long words
rather than by length per se".
C. Semantic predictors of word length
In the meanwhile, within the fields of experimental and computational
linguistics,
progress has been made in understanding the role of word length in verbal processing. Over
the years, it has emerged that words with different lengths tend to have different semantic
properties.
The idea was first put forth in pedagogical studies (Klare, 1988).
In (Elts, 1995), a
correlation coefficient of 0.96 was found between a noun's length and its average tendency
to be used as a technical term ("terminologicality"). Mikk et al. (2000), using data on the
human-assessed complexity of a large sample of words, found a correlation coefficient 0.86
between words' length and their semantic complexity.
Pinning down the precise semantic property that correlates to word length has proven
difficult. Already in (Greenberg, 1966) it was argued that a word's length correlates pos-
itively to its conceptual "markedness" of meaning. Various notions of markedness have
subsequently been discussed in the literature (Haspelmath, 2006).
Piantadosi et al. (Piantadosi et al., 2011, 2011B) and later Mahowald et al., (Mahowald
et al., 2012) reported that the length of words correlates positively with their contextual
information rate. More recently, Lewis and Frank (2016) have carried out a comprehensive
4
experimental study across 80 languages. They found that, in all the languages considered,
judgments of conceptual complexity for a sample of real words correlate highly with their
length, and they even control for frequency, familiarity, imageability, and concreteness. Their
conclusion is: "While word lengths are systematically related to usage − both frequency
and contextual predictability − our results reveal a systematic relationship with meaning as
well".
In the light of these findings, it would be a natural step to attempt an explanation of the
WLE in terms of the semantic differences among words. However, no such approach seems
to have been attempted in the literature.
D. The inverse word length effect
Recently, new aspects of the WLE have emerged through the analysis of a large set of data
from experiments by Miller and al (Miller et al., 2012). The data analysis was performed
by Katkov et al. (Katkov et al., 2014), who found no negative correlation between total
length of presented items and number of recalled words, thus disproving both rehearsal-time
theories and hypotheses based on the increasing complexity of longer items.
Moreover, they reported an inversion of the effect in mixed lists, that is, lists where words
are selected irrespectively of their length. They observed that, in this type of lists, the mean
values of recall probabilities allow to establish an increasing trend. Long words are recalled
better than short ones.
An "inverse" WLE had been previously reported by at least two groups, but in somewhat
less general circumstances: one of them (Hulme et al., 2006) embedded strictly pure lists
with a single word of a different type, while the results of Xu et al. (Xu et al., 2009), may
not bear direct comparison with data in languages other than Chinese.
If the inversion of the WLE for mixed lists will be confirmed by further experiments, it
will have to be taken into account by every general theory of the standard WLE. Let us
consider, therefore, what requirements a model should fulfill to explain both phenomena
simultaneously.
Call γ the fraction of long words in the list; call Pl(γ) the probability of recalling suc-
cessfully a given long word from a list in which a fraction γ of words are long; and let Ps(γ)
be the probability of recalling successfully a short word, from a list with a fraction γ of long
5
words.
Obviously, the function Pl(γ) is only defined for γ > 0, and the function Ps(γ) only for
γ < 1. For γ ∈]0, 1[, both functions are defined.
Theorists would have to reconcile two observations on the curves Ps(γ), Pl(γ):
1. Pl(γ = 1) < Ps(γ = 0)
2. Pl(γ) > Ps(γ) for γ ∈]0, 1[.
The only way these two inequalities can be simultaneously satisfied is if both Pl and Ps
are, on the whole, decreasing functions of γ. The simplest choice of these curves compatible
with experiments is one where both are monotonously decreasing, that is:
dPl
dγ
< 0
dPs
dγ
< 0
(1)
Pl
Ps
γ = 0
γ = 1
FIG. 1: Back-of-envelope sketch of the possible behavior of recall probabilities for short and long
words, as a function of the fraction of long words in the list.
This means that, whenever we replace a short word of the list with a long word, we are
lowering the recall probability of all the words in the list, both long and short. A higher
number of long words makes every single word in the list harder to recall.
It is difficult to imagine how this could ensue from the different duration of long and short
words. The question is then: can equation (1) result directly from the different semantic
properties of long and short words?
6
In this paper, I will show that the answer is positive as long as one models carefully the
process of verbal perception.
A suitable way of doing so is demonstrated in the next section. In section III, I employ a
retrieved-context description of verbal recall to derive both WLEs (standard and inverse). In
section IV, I test two key predictions of the theory against data from the PEERS experiment
of Kahana et al. (Lohnas and Kahana, 2013; Healey and Kahana, 2016). In section V, I
list five experiments designed to test further predictions of the theory. In the Conclusions,
I sketch a possible interpretation of the results.
II. VERBAL PERCEPTION
A. Structure of Semantic Space
Verbal perception is the conversion of words into meaning, and any theory of the phe-
nomenon must begin by defining the space in which the mental trajectory takes place. I will
refer to this as "semantic space" and will represent "semantic states" as integers belonging
to a segment X = (−A, A), with A ∈ N. Measurable quantities are to be computed in the
limit A → ∞, and a generalization to higher dimensions will be introduced below.
Since X ⊆ Z, it follows that, between two discrete times t0 and t0 + N , the psychological
trajectory can be represented by a sequence of integers x0, x1, . . . , xN . As memory plays a
crucial role in thought process, the laws of motion governing the trajectory will be in general
highly non-markovian.
Call W the vocabulary available to the system. Naturally, a given word w ∈ W may be
appropriate to describing more than one state, though with a varying degree of appropri-
ateness. For each word w, thus, one can define its semantic range Xw ⊂ X as the set of all
states described by that word. Meanings should be seen as distributed in an universal way
over X, as in Figure 2. The function w → Xw matches each word with locations associated
to meanings suitable for that word.
Not all states are equally fit to be verbalizable. Hence, there will be a varying probability
qx that the state x will match the word describing it. The value of qx is the degree of
verbalizability of state x; or conversely, given x ∈ Xw, qx is the fitness of word w as a
descriptor of state x.
7
FIG. 2: Position dependence of word distribution in semantic space. The space is represented
as an array of boxes. Each box represents a state, each color a word. Three of the words are
typed above next to circles the color that represents them. The images underneath show meaning
in different regions. States are filled with color to the extent to which they are verbalizable. The
white boxes correspond to qx = 0.
(cid:83)
If the sets {Xw}w∈W do not overlap, we can define a verbalization function vx :
w∈W Xw → W such that vx = w whenever x ∈ Xw. When the system tries to ver-
balize state x ∈ X, it produces word vx with probability qx, and the silent word 0 with
probability 1 − qx. We will define qx so that it is only null for nonverbalizable states.
Finally, we can take the trajectory to become markovian whenever it is driven by a
specific verbal input: at the instant in which a new word is presented, the system hops into
the nearest state described by that word.
Accordingly, let the function ξw : X → Xw be defined by the condition:
(2)
with the understanding that, if two x ∈ Xw satisfy this condition, ξw(x) is chosen by
ξw(x) − x ≤ x − x ∀x ∈ Xw
flipping a coin.
Given a verbal input (cid:126)w = (w1, w2, . . . , wN ), with the wi's ∈ W, and a starting point x0,
the trajectory of the system while perceiving the input will be given by :
= "tail"= "tusk"= "neck" 8
FIG. 3: Representation of the trajectory induced by the perception of a verbal input. The input
consists of five words, depicted here as circles. As the system perceives each new word, it reaches
for the nearest state described by that word.
xn = ξwn ◦ ξwn−1 ◦ . . . ◦ ξw1(x0)
(3)
∀n ∈ (1, N ) .
Obviously, the ordering of words in the input is essential. Suppose for instance that the
vocabulary contains two words, W = {A, B}, with XA = {0, 3} and XB = {2}, and the
starting point is x0 = 0. The verbal input (A, B) yields the trajectory (0, 0, 2), while the
input (B, A) yields the trajectory (0, 2, 3). Word A corresponds to two different states in
the two inputs.
FIG. 4: Noncommutative nature of verbal perception. Example of trajectories induced by the
permutation of two words.
B. Effect of the Voronoi length distribution
The verbalizable states associated to word w can be seen as seeds of a Voronoi partition
of X into cells {cx}x∈Xw, where states located on the boundaries are understood to belong
to each of two cells with probability 1
2.
9
FIG. 5: Voronoi cells of the "red" and "green" word within the same system.
During the perception of verbal inputs, the distance travelled by the system is controlled
by the function d(x, w) := ξw(x)− x, the "reaching distance" of word w from state x. This
is just the distance of x from the closest state belonging to Xw.
The level of nonlocality of a word w may be measured by its reaching distance averaged
over all starting points:
dw =
1
X
(cid:88)
x∈X
ξw(x) − x
(4)
Obviously dw1 > dw2 does not necessarily imply Xw1 > Xw2. Consider two words w1,
w2 such that Xw1 = {−A, A} and Xw2 = {−A/2, A/2}, with A even and > 2. We have
Xw1 = Xw2, but dw1 = A2
2A+1 > dw2 = A2+A
2(2A+1).
A direct relation can be shown to exist between a word's average reaching distance and
the size of its Voronoi cells. Call λi the size of the i-th Voronoi cell of word w. Assuming
¯λ (cid:29) 1, we may neglect boundary effects and write
(cid:88)
i∈Xw
λi/2(cid:88)
d=1
2
dw ∼ 1
X
(cid:32)
d ∼ ¯λ
4
(λ − ¯λ)2
¯λ2
1 +
(cid:33)
(5)
Thus, the average reaching distance of a word depends solely on the first two moments
of its Voronoi length distribution.
If (λ − ¯λ)2 (cid:38) ¯λ2 for a word w, its word structure contains strong fluctuations, so one may
separate X into regions were states described by w are denser, and regions were they are
sparser. Word w is effectively "localized" inside the regions were such states are dense, that
is, it expresses those semantic areas better than those where its states are sparse.
The degree of word localization has arguably a strong effect on the dynamics, as shown
in Figure 6.
10
FIG. 6: Example of two trajectories induced by the same input (shown as an array of circles).
The two trajectories begin when the system in different states. The red (localized) word creates a
narrow trapping region in semantic space, marked by the dashed ellipse. Once the localized word
appears in the input, all trajectories are trapped within the ellipse.
C. Analysis of word-length related properties
In most experiments, the relevant number of word-lengths is four, since words with more
than four syllables are rare in English. For simplicity, here we will consider the existence of
only two word-lengths, short and long.
Call dα (with α = s, l) the reaching distance averaged over space and over all words of
the same length (short or long). From eq. (5), we see that the space-averaged reaching
distance is the product of two factors, involving respectively the first and second moment
of the Voronoi length distribution. Let us consider how these two moments depend on the
word's length.
The average Voronoi length ¯λ may be related to the frequency of a given word in a
corpus of the language.
Indeed, if the system, in its 'speaking mode', explores semantic
space ergodically and uniformly, the frequency of a word is νw = λw
−1.
In a typical corpus of the English language, the frequency ν(S) of words with S syllable
is monotonously decreasing. As a consequence, we expect the average of the Voronoi length
λw to be larger for short words than for long words. While this is correct for most languages,
notice that there are exceptions, such as Turkish and Arabic, where the function ν(S) is
peaked at S = 2 (Fucks, 1956; Grzybek, 2007).
11
Let us now look at the relative fluctuations of the Voronoi length, described by (λ−¯λ)2
2
λ
.
We will surmise their magnitude through a qualitative argument.
As mentioned in the Introduction, various approaches have been taken to prove that
long words are on average more 'technical', 'specialized', 'distinctive' or 'marked' than short
words. Several claims made in (Elts, 1995), (Mikk et al., 2000), and (Lewis and Frank,
2016) may be rephrased as the statement that long words are, on average, conceptually
more specific. A word is conceptually specific if it is localized in certain areas of semantic
space. A correlation exists, therefore, between word-length and semantic localization.
Localization, in turn, will occur if the scale over which the Voronoi length fluctuates is
comparable or greater than its average value. This leads to the conclusion that the relative
fluctuations of the Voronoi length will be larger for longer words.
Thus, both factors in eq. (5) take a greater value if the word is long. It follows that
dl > ds.
D. Higher dimensions
A one-dimensional modeling of semantic space is of course unrealistic, and may not suffice
for every application of the theory. If X is taken to be a connected subset of Zn, the definition
we have given for the word structure {Xw}w∈W applies all the same. The points in space
are now vectors, and equations (3) and (4) are still valid, the distance in eq. (4) being the
Euclidean distance in n dimensions.
The Voronoi cells, however, become less simple to treat as they can be arbitrary polyhedra
(for a complete treatement, see Aurenhammer et al., 2013). Formula (5) for the reaching
distance must be modified, and it takes a geometry-dependent form.
The Voronoi structure, yet, affects directly only the process of verbal perception, not the
process of memory retrieval, which will be the subject of the next section. Thus, while in
the figures I will refer to the one-dimensional case, the mathematical results will apply to
any number of dimensions.
12
III. VERBAL RECALL
A. Retrieved-context primer
We have seen that the rules of motion of the system become markovian during the
perception of verbal input.
In retrieved-context models of short-term memory, the rules
of motion are also markovian during the search for memories, which is conducted through
the principle of free association. In these models, the retrieval of memories per se is not a
measurable phenomenon. What can be observed is the retrieval of words describing those
memories. Therefore, a recall experiment must be seen as the composition of two processes:
a processes of memory retrieval, and a process of memory verbalization.
The markovian process of memory retrieval must be described here, in keeping with the
spirit of context-driven models, as a random walk on X that effects a retrieval whenever it
meets a state corresponding to the experience to be recalled. The verbalization process, on
the other hand, depends on the verbalizability of memories: a memory x, once retrieved,
has a probability qx of leading the system to produce the word describing it.
The following mathematical problem arises. Supposing one is given
1) the structure {vx, qx}x of the vocabulary;
2) a word list (cid:126)w = (w1, w2, . . . , wN ) presented to the system;
3) the state of the system when the word list begins to be presented, that is, a probability
distribution χ[y0] on its position y0;
4) the state of the system when the retrieval process begins, that is, a probability distri-
bution ψ[x0] on the new position x0;
one wants to predict the probability that the i-th word will be among those recalled by
the system.
In the next section, this program will be carried out for the particular case of a vocabulary
containing words of two different lengths.
B. Application to the double word-length scenario
We will begin by defining the probability Ph(t) that a memory placed at distance h from
x0 is met by the retrieving random walk for the first time after a time t. In one dimension,
this is given by
13
y0
x0
FIG. 7: Trajectory during presentation and recall of a three-word list. Starting from a random
position (stage A), the system moves under the effect of verbal input (stage B), and its discontinuous
path leaves memory traces (stage C) that are pursued by a random walk in the retrieval stage (stage
D). The points in space-time corresponding to retrieval are boxed.
(cid:88)
(cid:126)n: (cid:80)h
1 ni=t
Ph(t) =
fn1fn2...fnh
where f2n = 0 and f2n−1 = (2n−3)!!
n!2n .
ABDtC The probability that a memory will be retrieved is therefore
T(cid:88)
pretrieval(h) =
Ph(t)
14
(6)
where the cutoff T is needed to obtain meaningful results in one or two dimensions, and
0
can otherwise be let to infinity.
Suppose a list (cid:126)w = (w1, w2, . . . , wN ) has been presented. Call pi( (cid:126)w; y0, x0) the probability
of retrieving the memory created by the i-th word in the list, given a certain initial position
y0 for the trajectory during presentation, and a certain initial position x0 for the trajectory
during retrieval.
We have:
(cid:18)(cid:12)(cid:12)(cid:12) Ξi(y0) − x0
(cid:12)(cid:12)(cid:12)(cid:19)
pi( (cid:126)w; y0, x0) = pretrieval
(7)
where Ξi := ξwi ◦ . . . ◦ ξw2 ◦ ξw1.
Our goal is to compute the average recall probability for an arbitrary list composed by S
short words and L long words arranged into a given order. This amounts to averaging eq.
(7) over all lists of the same type (cid:126)α = (α1, . . . , αN ) where αi ∈ {s, l} for i = 1, . . . , N :
pi((cid:126)α; y0, x0) =
1
s W L
W S
l
pi( (cid:126)w; y0, x0) :=
pretrieval
(cid:42)
(cid:18)(cid:12)(cid:12)(cid:12) Ξi(y0) − x0
(cid:12)(cid:12)(cid:12)(cid:19)(cid:43)
(8)
(cid:126)α
(cid:88)
(cid:126)w∈W N
α(wi)=αi
i=1,...,N
where α(w) is the type of word w and Ws (Wl) is the number of short (long) words in
the vocabulary.
We will perform the averaging over lists through a mean-field approach. Mean-field
approaches, widely employed in physics, consist in inverting the order of the two steps: the
computation of observables and the averaging. Instead of averaging the final probabilities,
one averages an intermediate, non-observable quantity usually called the "field".
In this
case, we may average the functions ξw themselves.
From section IIC, we know that the average displacement induced by the function ξw is
equal to dα(w) with ds < dl, while no constraints emerged concerning the direction of this
displacement as a function of word type. Therefore, to average the functions Ξi over all
word lists of one type, we replace ξw with the function ¯ξα(w) defined by ¯ξα(x) = x + dαe,
where e is a unity vector randomly chosen from a suitable distribution Ω0[e].
We can thus write Ξi(y0) ∼ y0 +(cid:80)j
(cid:42)
pretrieval(cid:16)(cid:12)(cid:12)(cid:12)
pi((cid:126)α; y0, x0) =
where Ω(e1, . . . , ei) =(cid:81)i
k=1 Ω0[ ek].
i(cid:88)
k=1
k=1 dαk ek, and eq. (8) becomes
(cid:12)(cid:12)(cid:12)(cid:17)(cid:43)
Ω
dαk ek + y0 − x0
15
(9)
The initial positions y0 and x0 are not measurable quantities. In eq. (9), therefore, they
must be averaged over through two suitable distributions χ(y0) and ψ(x0), yielding
i(cid:88)
(cid:12)(cid:12)(cid:12)(cid:17)(cid:29)
(cid:28)
pretrieval(cid:16)(cid:12)(cid:12)(cid:12)
k=1 dαk ek and p(yx) := pretrieval(cid:0)(cid:12)(cid:12)y − x(cid:12)(cid:12)(cid:1).
dαk ek + y0 − x0
k=1
χ,ψ,Ω
pi((cid:126)α) =
where yi :=(cid:80)i
The function we need to average may be rewritten as
(cid:28)
(cid:29)
p(yix0)
=
(10)
χ,ψ,Ω
p(yix0) = P(cid:2)i1 = i(cid:3) +
P(cid:2)i1 = j(cid:3) p(yiyj)
(cid:88)
j(cid:54)=i
where P(cid:2)i1 = i(cid:3) is the probability that the i-th memory, yi, will be the first one to be
found, and(cid:80)N
P(cid:2)i1 = i(cid:3) < 1. Substituting eq. (11) into eq. (10), we find
(cid:43)
χ,ψ p(yiyj)
(cid:42)(cid:10)P(cid:2)i1 = i(cid:3)(cid:11)
(cid:10)P(cid:2)i1 = j(cid:3)(cid:11)
(cid:88)
pi((cid:126)α) =
χ,ψ +
(12)
(11)
i=1
j(cid:54)=i
Ω
The distribution ψ refers to the state x0 of the system after the so-called retention interval,
during which the subject freely elaborates the information gathered during presentation. A
full understanding of such elaboration would require modeling the free motion of this system,
but we have only been able to markovianize the equations of motion during a progressive
searching task or in the presence of a driving input. Thus, a reasonable ansatz is necessary.
Neglecting recency effects, we can suppose ψ to contain N similar peaks at the locations
y1, y2, . . . , yN explored during presentation. In this picture, the dependence of(cid:10)P[i1 = i](cid:11)
on the index i will be negligeable, so we can approximate
value p0.
Substituting this into eq. (12), and rewriting the argument of pretrieval, we find
(cid:34)
(cid:42)
pretrieval(cid:16)(cid:12)(cid:12)(cid:12) max(i,j)(cid:88)
(cid:88)
j(cid:54)=i
pi((cid:126)α) = p0
1 +
k=min(i,j)+1
Ω
(cid:69)
(cid:68)P[i1 = i]
(cid:12)(cid:12)(cid:12)(cid:17)(cid:43)
dαk ek
χ,ψ
(cid:35)
χ,ψ
with a constant
(13)
where Ω is now the product of j − i copies of the distribution Ω0, and the value of the
summand depends solely on the number of long and short words located between word i
and word j.
Using pretrieval(0) = 1 to include the j = i term into the sum, and noticing that the labels
16
of the ek's are interchangeable, we can finally rewrite eq. (13) as
(cid:88)
pi((cid:126)α) = p0
π(Sij, Lij)
where we have introduced the quantities
j
max(i,j)(cid:88)
min(i,j)+1
Lij :=
π(m, q) :=
1(αk = l)
(cid:28)
pretrieval(cid:16)(cid:12)(cid:12)(cid:12)ds
m(cid:88)
k=1
Sij := i − j − Lij
(cid:12)(cid:12)(cid:12)(cid:17)(cid:29)
Ω
q(cid:88)
k=1
ek + dl
hk
(14)
(15)
(16)
and the unit vectors gk, hk are independently distributed according to m + q copies of
the distribution Ω0.
For pure lists, eq. (14) becomes
ppure
i,α := pi(α, α, . . . , α) = p0
(cid:20)
1 +
(cid:16) N−i(cid:88)
i−1(cid:88)
(cid:17)
π(cid:0)δαsh, δαlh(cid:1)(cid:21)
+
h=1
h=1
(17)
C. Coexistence of word length effects
As mentioned above, the probability of recalling the i-th word of the list is equal to
the probability of retrieving the i-th memory, multiplied by the factor qwi. In the previous
section, we averaged the retrieval probability over all words of the same type; similarly, we
must now average the verbalizability qx, which we take to be distributed independently of
the word structure {vx}x. Defining qα as the average verbalizability of words of type α, we
obtain the full recall probability
Pi(α1, . . . , αN ) = qαi pi(α1, . . . , αN )
(18)
As mentioned in the Introduction, the classical WLE is the experimental fact that pure
lists made of shorter words are easier to remember. Of course such behavior may always
17
be prevented, in principle, by making the verbalizability ratio qs/ql sufficiently low. Yet, if
the reported inversion of the WLE exists within this model, it must rely on the opposite
requirement -- namely, that the verbalizability ratio is sufficiently high.
The questions we are therefore supposed to answer are: first, whether there exists for this
model a range of values of qs/ql where both the classical and the inverse WLE occur; second,
under what conditions on the parameters this may happen, and whether such conditions are
relevant to current experiments.
Let us describe the typical experimental situation.
In the experiments, word lists are
generated by drawing words at random from a vocabulary W.
scenario, this vocabulary will contain Ws short words and Ws long ones. We may consider,
therefore, an ensemble of lists of length N , where each word has a probability γ := Wl
of being long, and a probability 1 − γ of being short.
In a double-word length
Ws+Wl
The recall probability for the i-th word of the list can be averaged over all lists whose
i-th word is of type α, yielding pi,α(γ) := (cid:104)pi((cid:126)α)(cid:105)γ. Substituting eq. (14), this becomes:
(cid:34) (cid:88)
(cid:68)
S,L≥0
0≤S+L≤N−i
(cid:69)
(cid:88)
(cid:68)
S,L≥0
0≤S+L≤i−1
(cid:35)
(cid:69)
pi,α(γ) = p0
π(S, L)
+
γ
π(S + δαs, L + δαl)
(19)
γ
In the first sum, which is the contribution from j ≥ i in eq. (18), S and L stand for the
number of short or long words in positions k such that i + 1 ≤ k ≤ j; in the second sum (the
contribution from j < i) S and L stand for the number of short or long words in positions
k such that j + 1 ≤ k ≤ i.
In eq. (19), the notation(cid:10) . . .(cid:11)
γ has come to denote an averaging over S, L performed
for each separate value of S + L and summed together. This is done through the binomial
distribution
(cid:18)L + S
L
(cid:19)
γL(1 − γ)S
Φ(S, L) =
(20)
Given that the total recall probability is Pi,α(γ) = qαpi,α(γ), the standard WLE effect
(Pi,s(0) > Pi,l(1)) will occur if ql
qs
will occur if Pi,s(γ) < Pi,l(γ), that is, if ql
qs
effects coexist if θinv < ql
qs
rewritten as
< θcl, where θcl = pi,s(0)/pi,l(1), and the inverse effect
> θinv, where θinv = pi,s(γ)/pi,l(γ). The two
< θcl , which can only happen if θinv < θcl. This condition can be
pi,l(1)pi,s(γ) < pi,s(0)pi,l(γ)
18
(21)
Now, it can be seen that pi,α(γ) a decreasing function of γ, because by increasing γ one
transfers weight from the first to the second argument of π inside both terms of eq. (19),
which reduces the value of the function π. Hence, we have pi,l(1) < pi,l(γ) and pi,s(γ) <
pi,s(0), from which it follows that the inequality (21) is identically satisfied.
We conclude that, in this model, the WLE can undergo an inversion for any γ ∈]0, 1[, as
sketched in Figure 1.
D. Formulas for recall probabilities: slow-diffusion regime
Consider a list of the type (α1, . . . , αN ), where αi ∈ {s, l}, containing L words and S
short ones. The trajectory during presentation is illustrated in Figure 8. The system begins
from a random position y0, and at each new word wi of type αi, its position is shifted forth
by the operator ξwi.
The distance travelled at step i is, in the mean field approach, equal to dαi. So a memory
produced by the presentation of a short word will be formed in the vicinity of the latest
memory (that is, within a distance of order ds) whereas a memory produced by the pre-
sentation of a long word will be formed at a longer distance (of order dl) from the memory
preceding it. Consequently, memories are divided into clusters separated by a distance of
order dl from each other. Each cluster spreads over a width of order ds.
FIG. 8: Trajectory during the presentation of a mixed list. The list structure corresponding to the
example is lsslsssslsssslss, where l stands for long words and s for short ones. The longer jumps
correspond to the presentation of long (localized) words.
These memory clusters correspond to different "segments" of the list. Call {li}L
i=1 the
index values corresponding to long words within a given list. If l1 = 1, the segments are
19
(cid:126)si = (wli, wli+1, . . . , wli+1−1) for i = 1, . . . , L − 1, and (cid:126)sL = (wlL, wlL+1, . . . , wN ). If w1 is
short, there is an additional segment (cid:126)s0 = (w1, w2, . . . , wl1−1), and the number of segments
is L + 1. Call ai the length of segment (cid:126)si. All the ai's must be positive except a0, which
may be null, and(cid:80)L
i=0 ai = N .
The direction in which the trajectory moves at each step is defined by the unknown
distribution Ω0, which depends on the details of the word structure {vx}x. For a generic
choice of Ω0, clusters formed at longer time intervals from each other will lie further apart
processes: a clustering process and a diffusion process.
in semantic space. Hence, the trajectory during presentation is the composition of two
Notice that in eq. (13) the argument of pretrieval is of the order of (cid:112)Sijd2
the list is short enough (that is, if pretrieval((cid:112)Sd2
l . If
√
Sds) ∼
√
pretrieval(ds) and pretrieval(dl) (cid:28) pretrieval(
Sds)), the summand has two orders of magnitude:
one of the order of pretrieval(dl) = pl and one of the order of pretrieval(ds) = ps. This
s + Ld2
l ) ∼ pretrieval(dl), pretrieval(
s + Lijd2
corresponds to the fact that the diffusion process is slow on the scale of the trajectory
during presentation. In this regime, formula (13) for pi((cid:126)α) may be estimated by replacing
the summand with pl whenever there is at least one long word between min(i, j) + 1 and
max(i, j), and with ps otherwise. This is equivalent to approximating the matrix elements
π(m, q) of eq. (16) with π(1, 0) if q = 0 and m > 0, and with π(0, 1) if q > 0, thus ignoring
the dependence of the retrieval process on the distance between segments of the list.
Call yi1 the first memory to be retrieved. The conditional retrieval probability for memory
yj is pretrieval(yj − yi1). In the slow-diffusion limit, this is of the order of ps if memories yi1
and yj belong to the same cluster, and is of the order of pl otherwise. Averaging over the
first retrieval, we find
(cid:104)
p(i) = p0
1 + (ci − 1)ps + (N − ci)pl
(cid:105)
where ci is the length of the segment of the list to which word i belongs.
The average recall probabilty P α for words of type α will thus be equal to:
(22)
(23)
(24)
L(cid:88)
i=0
Ps =
qsp0
S
(cid:16)
(cid:17)(cid:104)
(cid:104)
L(cid:88)
i=1
ai − 1 + δ0i
1 + (ai − 1)ps + (N − ai)pl
Pl =
qlp0
L
1 + (ai − 1)ps + (N − ai)pl
(cid:105)
(cid:105)
Defining µ = N−a0
L
and ∆ = 1
L
(cid:80)L
i=0 a2
i , we can rewrite this as
(cid:104)
Ps = qsp0
N pl − ps + 1 +
(cid:104)
L
N − L
(∆ − µ)(ps − pl)
N pl − ps + 1 + µ(ps − pl)
Pl = qlp0
20
(25)
(26)
(cid:105)
(cid:105)
All the dependence on the ordering of words in the list, therefore, enters the recall prob-
abilities through the parameteres µ and ∆. The values of these parameters are shown in
table I for simple lists.
TABLE I: Values of the observables µ, ∆, and Ps/Pl for simple lists
List Structure
l . . . l
l . . . l s l . . . l
l l l . . .
S
s . . . s
(cid:124)(cid:123)(cid:122)(cid:125)
(cid:124)(cid:123)(cid:122)(cid:125)
(cid:124)(cid:123)(cid:122)(cid:125)
l s . . . s
s . . . s
M
m
l s . . . s
. . . l s . . . s
(cid:124)(cid:123)(cid:122)(cid:125)
m
(cid:124)(cid:123)(cid:122)(cid:125)
l s . . . s
m
µ
1
1
∆
1
1 + S2
N−S
Ps/Pl
--
(N−S)pl+(S−1)ps+1
(N−1)pl+1
qs
ql
N
N−1
N − M
N +2
N−1
N 2 − 2M (N − M )
qs
ql
N 2ps−2N ps+ps+N−1+M (2M−2N +1)(ps−pl)
N pl+ps+1
N 2pl+N−2N pl+ps−1
qs
ql
(N−1)[(N−M )ps+M pl−ps+1]
m + 1
(m + 1)2
qs
ql
For a pure list, entirely composed of words of type α, eqs. (25) and (26) become:
(cid:2)1 + (N − 1)pα
(cid:3)
P pure
α = qαp0
(27)
Data for mixed lists may be interpolated with formulas (25) and (26) to test the theory
and fix the values of the internal parameters.
Finally, let us look at the range of occurrence of the WLEs in the slow-diffusion regime.
The classical WLE (P pure
s
> P pure
l
) emerges for ql
qs
< θcl where
θcl =
1 + (N − 1)ps
1 + (N − 1)pl
(28)
The ratio Pl/Ps for mixed lists can be estimated as follows. Given a fixed value of L < N ,
µ ranges between µmin = 1 (for a0 = N − L) and µmax = N/L (for a0 = 0). The minimum
21
value of ∆ is obtained by starting the list with a short word and having N − (L + 1)(cid:100) N−L−1
L+1 (cid:101)
L+1 (cid:101)) − N segments of 1 + (cid:98) N−L−1
L+1 (cid:99)
segments of 1 + (cid:100) N−L−1
words. The maximal ∆ is obtained by setting a0 = 0 and lumping all the short words into
one segment: ∆max = (N−L+1)2+L−1
L+1 (cid:101) words and (L + 1)(1 + (cid:100) N−L−1
L
.
Substituting ∆max and µmin in (25), (26), one obtains a strict upper bound on the ratio
Ps/Pl: Ps
Pl
< qs
ql
θinv, where
θinv =
L2(ps − pl) + N L(pl − 2ps) + psN 2 + (N − L)(ps − 2pl + 1)
(N − L)[1 + (N − 1)pl]
(29)
Once again, the inverse WLE (Ps/Pl < 1) will emerge under the sufficient condition
> θinv; hence, the possibility that the classical and inverse effects may coexist requires
ql
qs
θcl > θinv.
If we rewrite this inequality by means of eq. (28) and (29), the "microscopic" probability
parameters ps and pl cancel out and we obtain the general condition L2−LN +2(N −L) < 0,
which is identically satisfied for any L > 2.
It follows that, in the slow-diffusion regime, the system displays an inversion of the
WLE for all mixed lists containing more than two long words. This statement, unlike the
conclusions of the previous section, is not only true on average, but holds true regardless of
the order in which short and long words are arranged within the list.
E. Formulas for recall probabilities: fast-diffusion regime
In the previous section we have considered the case where the diffusion process was much
slower than the clustering process -- which translates into upper bounds on the order of
magnitude of S and L. Notice that we defined these bounds in terms of the parameters ds
and dl, whose value may vary from subject to subject. Therefore, the very same list may be
experienced in a fairly stationary regime by one subject, and in a regime of fast diffusion by
a more easily distracted one.
Here, we will consider the opposite case of very fast diffusion, defined as the regime
where pretrieval(ds + x) goes down fast on the length scale of ds, so the matrix elements of
π decay quickly as the value of the indices grows, and the recall dynamics is dominated
by the contiguity effect. If a long word causes a lesser diffusion than two short words, we
may neglect all but the top-left elements of the π-matrix: π(0, 0) = 1, π(1, 0) = ps, and
22
π(0, 1) = pl. Otherwise, we can work in the lowest approximation by neglecting also pl.
Eq. (19) becomes
(cid:104)
1 + ps + pα − (ps − pl)γ
(cid:105)
Pα(γ) ∼ p0qα
(30)
where I dropped the pedix i because all dependence on i vanishes as long as i is neither
1 nor N .
The thresholds for the verbalizability ratio are
θcl =
1 + 2ps
1 + 2pl
θinv =
1 + 2ps + (pl − ps)γ
1 + ps + pl + (pl − ps)γ
(31)
and for θinv < ql
qs
< θcl, eq. (30) is nothing but a linear version of Figure 1.
When experiments yield a near-linear curve Pα(γ), thus, it may be taken as a sign that
the system is operating in a very fast diffusion mode. When the observed curve is nonlinear,
one must infer that terms of the type π(h, 0) for h > 2 are relevant, and a polynomial
interpolation can be performed, by truncating the sums in eq. (19).
The crossover toward linearity of the curves Pα(γ) may be explored by tuning experi-
mentally the amount of diffusion in the system, as we will see in section V.
IV. DATA ANALYSIS
A. The Data
The foregoing analysis has proven that, within this model of verbal perception, the recall
process can display consistently both the classical and the inverse WLE, as observed in the
experiments.
There may be, however, other mechanisms leading to a similar prediction. Katkov et al,
for instance, propose a lexical explanation for both the direct and inverse WLE, based on
possible differences in the long-term neural representation of long and short words (Katkov
et al., 2014). In order to tease out which mechanism is really responsible for the effects, one
needs to extract further information from the data, going beyond the computation of mere
recall probabilities.
23
I will do so by using data from PEERS (Penn Electrophysiology of Encoding and Retrieval
Study), a large study conducted at the University of Pennsylania and devoted to assembling
a database on the electrophysiological correlates of memory (Lohnas and Kahana, 2013).
The sample I have considered corresponds to Experiment I of PEERS. It includes data
from trials on 156 college students (age range: 18−30) and on 38 older adults (age range:
61−85 years). In each trial, 16 words were presented one at a time on a computer screen.
Each word was drawn from a pool of 1638 words with length varying between S = 1 and
S = 6 syllables. The word-length distribution was peaked at S = 2, reflecting that of a
typical English lexicon (but not of a typical English corpus).
Each item was kept on the screen for 3000 ms, followed by an interstimulus interval of
800−1200 ms. After the last item in the list, there was a delay of 1200− 1400 ms, after which
the participant was given 75 s to attempt to recall aloud any of the just-presented items.
Multiple trials were performed on each subject, summing up to 3744 trials for the students
sample and to 912 for older adults. For more details on the experimental procedure, see
Healey and Kahana, 2016.
Because the word lengths involved are more than two, the formulas we derived in the
previous section may not be applied verbatim. Nonetheless, two key consequences of the
theory afford a direct comparison with the data.
B. Recall by contiguity
We say that a word is recalled "by contiguity" when its recall occurs immediately after
the recall of a word contiguous to it within the list. In the model we have developed, no
matter how high the amount of diffusion, recall by contiguity will be more frequent for short
words than for long words. This follows from the fact that two consecutive short words
belong necessarily to the same segment, whereas a short word and a long word, though
contiguous, may belong to different segments, and two long words are sure to belong to
different segments.
While this prediction has been derived in a two-length model, it generalizes immediately
to models with multiple lengths. We have proven in section II that shorter words are
characterized by a shorter reaching distance; hence, memories formed by shorter words have
a higher chance of being located in the proximity of memories formed by contiguous words.
Contiguity will play therefore a stronger role in the recall of a shorter word: the shorter the
words involved, the higher the chances of recall by contiguity.
24
Pcon
S
FIG. 9: Probability of recall by contiguity for words with S syllables (blue dots); probability of
recall by contiguity (black); and mean probability of recall by contiguity (red). All values have
been computed from PEERS data.
Let us call Pcon(w) the probability that a given word w, if recalled, will be recalled by
contiguity. The theory predicts that Pcon should be larger for shorter words. If so, computing
Pcon from the data and averaging it over all words of the same length would yield a decreasing
curve Pcon(S).
In Figure 9, Pcon is shown for all words having the same number of syllables (blue dots).
The data have been aggregated from all trials. The distribution of Pcon (black line) is wide
for all word lengths; nevertheless, the mean probability of recall by contiguity (shown in
red) decreases monotonically with the number of syllables, in agreement with the theory.
The 1638-element wordpool used in PEERS contains only four 4 five-syllable words, and
a single 6-syllable word ("encyclopedia"). Hence, the statistics for these two lengths may
not be considered reliable.
While Figure 9 refers to Experiment 1 of PEERS, the robustness of the effect has been
12345600.10.20.30.40.50.60.7PconPconhistogrammean(Pcon) 25
checked for by repeating the analysis on the databases from Experiments 2 and 3 (see Lohnas
et al., 2015). The decreasing trend has proven invariant across databases.
C. Distribution of Jumps
Let (i1, i2, . . . , iM ) be the serial positions of the words recalled by the subject during a
certain trial. At each step n in the recall process, the system performs a serial-position
jump of a magnitude δn = in − in+1. In our terminology, longer jumps will require the
exploration of a larger portion of semantic space. Therefore, the distribution of jumps should
be monotonously decreasing as a function of their size (as appears indeed from the PEERS
data, Figure 10).
Pt
δ
FIG. 10: In black, distribution of serial-position jumps in the recall process as a function of jump
size δ, computed from PEERS data.
In colors, distribution of jumps for a fixed size b of the
word-length barrier, shown in the legend. As the lists contain 16 words, δ ranges between 1 and
15.
From our analysis of the two-length scenario, we know that these jumps cover greater
distances in semantic space if the memory created by the word of departure and the memory
created by the word of arrival belong to different clusters. Since it is long words that have
246810121400.050.10.150.20.250.30.350.40.450.5b=1b=2b=3b=4 the ability to break clusters, a long word located in a position k such that i < k ≤ j plays
effectively the role of a recall barrier, hindering direct transitions between wi and wj.
26
FIG. 11: To the left, depiction of the hyerarchical segmentation of a 16-word list. Word lenght
is encoded in colors as shown in the vertical sidebar. At each level of segmentation, clusters are
displayed as horizontal bars covering the given portions of the list. To the right, a plot of the mean
cluster length (blue) and of the number of clusters per list (red), as a function of the hierarchical
level S of clusters, computed from PEERS data by averaging over all trials.
S
Similarly, in experiments with multiple word lengths, longer words create memories at
longer distances. As clustering in semantic space will occur then over multiple length scales,
clusters assume a hyerarchical structure. So does the segmentation of the list, with mod-
erately long words marking off smaller segments within the larger segments delimited by
longer words. This is shown in Figure 11, together with the average length of clusters as
computed from the data, and the average number of clusters per list.
The probability for the transition wi → wj (a jump of size i− j) will be affected mainly
by the hyerarchical level of the clustering involved. Since longer words break clusters at a
higher level, we conclude that the effect of clustering on such a jump will be controlled by
the length of the longest word located between the positions i and j:
max{L(wj+1), L(wj+2), . . . , L(wi)} if i > j
max{L(wi+1), L(wi+2), . . . , L(wj)} if j > i
bij =
(32)
where L(w) is the length of word w and, once again, the lowest index has been excluded
for the same reasons why it was not counted in eq. (13) of section III.
1234560246810121416cluster lengthnumber of clusters per list123456 27
FIG. 12: Serial-position jump from word 8 to word 3 during the recall of a nine-word list (with
words from the PEERS wordpool). The red bars describe word length: the "barrier word" for this
transition is marked in yellow.
We will call bij the "barrier size" for the i → j transition. The role played by this
quantity is depicted in Figure 12. If the clustering mechanism does exist, it will enhance
the transition probability for small values of bij, and suppress it for higher values.
To test this prediction, we must first study the problem in the absence of clustering,
calculating the transition probability Pt(δ) for jumps of size δ. The resulting formula must
be evaluated using the experimental values of parameters, and can then be compared with
the values of Pt(δ) extracted from the data.
If the clustering occurs, we will observe an
enhancement of Pt for small values of the barrier size b and a suppression at high values.
Let us suppose that the clustering does not occur. The system then will not "feel" the
word-length barriers, and the probability Pt of a jump will be independent on the size of
the barrier to be overcome. Conversely, the probability distribution of barriers over the
jumps performed by the system will depend only on the jumps' size, and on the probability
distribution of word length with the list.
The probability that a jump of size δ involves a word-length barrier b is equal to the
probability that the longest word out of a random sequence of δ words has length b. This is
equal to
ZebraSidingTubeRefrigeratorNewspaper (cid:33)δ
f (L)
(cid:32) b(cid:88)
L=1
(cid:32) b−1(cid:88)
L=1
−
(cid:33)δ
f (L)
Pt(bδ) =
28
(33)
where f (L) is the word-length distribution with the lists. Formula (33) describes the
no-clustering barrier-size distribution, that is, the distribution of barrier sizes for jumps of
a given length if the jumps are not influenced by the barriers.
I have evaluated formula (33) using values of f (L) computed directly from the lists
recorded in the PEERS database. In Figure 13 (see the end-page), the results are compared
with the values of Pt(bδ) extracted directly from data. The no-clustering curves are dashed,
the experimental curves solid.
The effect of clustering is apparent. All experimental curves reveal an enhancement of
the transition probability for small values of b, and a suppression at large values. Morevoer,
the effect persists across the whole range of possible jump sizes.
We would expect that the influence of clustering should be small for jumps shorter than
the typical size of the clusters, and grow as the jumps get longer. This is indeed what we
notice in Figure 13. The influence of clustering can be observed already for jumps of size one,
with an enhancement of probability for small barriers and suppression for larger barriers.
Yet, the clustering becomes more prominent as the jumps get longer.
A noticeable increase in the importance of clustering is observed after δ = 6. This may
be due to the fact that the S = 3 clusters are becoming important, as the jumps have gotten
longer than the average length of the clusters of the third level (see Figure 11).
Since the PEERS lists have length 16, there is no available statistics for values of δ larger
than 15. Nonetheless, we can predict that this growth will saturate when the probability of
meeting a maximal bareer approaches near-certainty. For jumps longer than that, b is no
longer a controlling parameter; from that point on, the continued suppression of recall will
derive from the need to find the next memory multiple clusters away.
V. FURTHER TESTS
A. Predictions on average recall probabililties
Most free-recall experiments performed so far have employed either a wordpool with an
arbitrary number of syllables, as in PEERS, or a pool characterized by a fixed length that
29
is usually 2 syllables (as in the Toronto Wordpool, see Friendly et al., 1982).
Further tests on this theory, however, would be easiest to perform through experiments
in which the words presented for recall are drawn from either of two sets: a set of very long
words (e.g. number of syllables S = 4) and a set of very short ones (S = 1).
When the lists are sufficiently short, the system operates in a slow-diffusion (or "cluster-
ing") regime. In this case, three parameters should be kept track of in the experiment: ∆,
µ, and the number L of long words in each list (see section III). While the recall probability
of individual words depends on the details of the list, the average recall probability for short
or long words should be affected by the list structure only through those three parameters,
which completely characterize the list for the purposes of this type of experiment.
It is then feasible to test qualitatively the following predictions:
1. P s correlates positively with ∆ and negatively with µ.
2. P l correlates positively with µ.
These are straightforward consequences of (25), (26).
In experiments with words of two lengths, data may be interpolated with formulas (25)
and (26) to test the theory and to fix the values of the internal parameters.
B. Predictions depending on serial position
Serial position effects related to word-length can be isolated experimentally by using word
lists constructed on two or more basic templates of segment-structure (cid:126)α = (α1, . . . , αN ), with
αi = s, l. The experimenter would present equivalent lists to a number of participants and
would correlate the recall probability of words with their positions within each template
structure.
Large data sets from such experiments should be able to corroborate or to rule out the
mechanisms I have described. Competing effects of familiar types, such as primacy and
recency, may be easily subtracted.
Two simple predictions can be made, easier to check in a slow-diffusion regime but equally
valid with faster diffusion:
1. Words from longer segments have a higher recall probability than words of the same
type from shorter segments;
2. Successful recall of words from one segment hinders the recall of words from others.
30
The first prediction follows from eq. (22), and should be easy to test.
The second prediction must be compared with data by computing the correlation func-
tions Cij, that is, the joint probability for the recall of the i-th and j-th words in the list.
From there, it is straightforward to obtain the in-segment and cross-segment correlation
functions:
(cid:42)
(cid:43)
(cid:42)
(cid:43)
C in
α1α2(d) =
Cij
i−j=d
i,j∈ same segment
α(wi)=α1,α(wj )=α2
C cross
α1α2 (d) =
Cij
i−j=d
i,j∈ different segments
α(wi)=α1,α(wj )=α2
(34)
and to verify whether C in
α1α2(d) > C cross
α1α2 (d), as the theory suggests.
We may add to these predictions a third one -- namely, the fact that the long word of
each segment is the easiest one to recall. This results directly from the bounds we derived
in the previous section on the ratio qs/ql.
C.
Inter-response intervals
A further prediction of the theory regards inter-response intervals -- that, is the time elaps-
ing between one recalled item and the next -- whose measurement in free-recall experiments
dates back to (Murdock and Okada, 1970).
During retrieval from long-term memory, it was shown (Gruenewald and Lockhead, 1980)
that clusters occur due to stable semantic associations between objects: a subject who is
asked to list some animals, for instance, may recall first a set of farm animals, then a number
of house pets, then several birds. The inter-response intervals are shorter within clusters
than between clusters.
The retrieval of examples in the experiment of Gruenewald and Lockhead depends entirely
on the long-term representation of items. In the situation we have described, on the contrary,
short-term memory is at play, and the retrieval process has to locate the vanishing traces of
a recent experience.
Nonetheless, it is easy to see that the time interval elapsing between the retrieval of two
memories will be longer between two memories belonging to different clusters, and shorter
for memories belonging to the same cluster.
Hence, the same is true among items of the list that are successfully verbalized. The
inter-response intervals will be longer for the consecutive recall of two words belonging to
31
different segments of the list, and shorter for the consecutive retrieval of two words belonging
to the same segment.
D. Experiments with varying presentation rate
As mentioned in the Introduction, the WLE has been reported in experiments where
the time interval between the presentation of consecutive items was a controlled parameter.
Experiments on the WLE with rapid presentation of the stimuli were first performed by
Coltheart and Langdon (1998), who found the WLE by presenting an item every 114 ms,
every 157 ms, and every 243 ms. In (Campoy, 2008), somewhat lower presentation rates
were used (between 300 and 400 ms), and again the persistence of the effect was proven over
different rehearsal times.
Here, I will argue that such experiments may offer an ideal tool to study the crossover
between the "diffusive" and the "clustering" regime.
Indeed, if we modify the foregoing computation to allow the system to random-walk on
its own for a time τ between the presentation of the i-th and i + 1-th items, this will be
equivalent to increasing the average distance d travelled between the memory yi generated
by word wi and the memory yi+1 generated by word yi+1. This amounts to rescaling time
while replacing dl and ds with larger effective distances. The matrix elements of π will decay
faster as their indices grow. Therefore, the system will move closer to the diffusive regime.
On the other hand, the theory predicts (section IIIE) that the curve Pα(γ) will become
linear in the fast-diffusion regime. Hence, by reducing the presentation rate, one should see
the two curves in Fig. 1 becoming progressively linearized, at least up to values of τ so
large that not only the clustering, but also the contiguity effect breaks down. In this limit,
moreover, eq. (30) predicts that the curves Ps(γ) and Pl(γ) will become parallel.
Other ways of controlling the crossover between clustering and diffusing regime (e.g. by
pharmacological means, or through distractor tasks such as those of Bjork and Whitten,
1974) can be similarly applied.
32
E. Subject-dependent variability of the classical WLE
We have just examined some ways of testing the theory that are going to require ad-hoc
experiments. Let us conclude by mentioning one type of test that may be performed on
already available databases -- namely, data from the experiments performed so far on the
classical WLE.
It is known that different people employ different strategies in order to exploit the struc-
ture of semantic space during memorization and recall (Healey et al., 2014). When it comes
to the clustering described above, we expect its strength to be just as dependent on the
subject considered.
On the other hand, the strength of the classical WLE is also a function of the particular
subject. The difference between the recall probabilities of all-short lists and all-long lists,
while being positive on average, will vary in magnitude from subject to subject.
If the
mechanism at the root of the of the WLE is indeed a clustering phenomenon, we expect
that it will be stronger for subjects for whom the clustering is stronger.
Let (i1, i2, . . . , iM ) be again the serial positions of the words that have been recalled by
(cid:80)M−1
the subject during a certain trial. The average serial-position jump between consecutively
n=1 in − in+1. For each subject, we can define
δsubject as the average of δtrial over all trials performed on him or her. The parameter δsubject
recalled words for this trial is δtrial = 1
M−1
may serve as a simple measure of the subject's inclination toward clustering. Indeed, δsubject
is close to unity for subjects strongly inclined toward clustering, and δsubject (cid:29) 1 if the
subject's tendency toward clustering is low.
One way of gaining insight into the likelihood of the theory consists in correlating δsubject
with the strength of the classical WLE effects, controlled by ps− pL, where ps and pL are the
recall probabilities calculated for pure lists of short and long words. A negative correlation
between these two variables would be a strong confirmation that the WLE is indeed due to
a clustering phenomenon.
VI.
INTERPRETATION AND CONCLUSIONS
I have proposed a theory of verbal perception, extracted some of its properties, and
validated it through a comparison with free-recall data.
33
The theory is based on the notion that a word does not have, in general, a single meaning.
A human subject exposed to a stream of verbal input will decide on the meaning of each
new word on the basis of both the structure of its vocabulary and the meaning he/she has
given to the words preceding it. This also applies to a list of random words, because our
mind strives to interpret them as parts of a meaningful discourse.
It may be instructive to think of such discourses as "narratives". Common experience tells
us that a two-word list is already capable of creating a strong narrative sense (e.g.: picnic,
lightning). When a word in the list has no semantic connection to the context created by
the words preceding it, the mind perceives a "change of scenery" and assumes that a new
narrative is beginning.
A list of words is thus perceived as a collection of distinct "stories". When prompted
to recall the list, the subject remembers each story as a separate experience, and needs to
re-experiences a given story before retrieving the words responsible for creating it.
Words that have specific meanings have obviously less probability of fitting into a ran-
domly generated story. Otherwise said, the words most likely to break the narrative are
those with the highest level of localization in semantic space. We have argued that this
correlates positively with word length.
Hence, a list of N long words is likely to break into as many one-word stories, whereas
a list of short words is more likely to be perceived as a single continuous narrative. Since a
single narrative is easier to recall than many unrelated ones, the standard word length effect
ensues.
The clustering property of short words is at play in mixed lists as well. But its effect is
hindered by the presence of long words breaking the narrative. As our analysis has shown,
this can lead to an inversion of the WLE, encountered in experimental observations.
In this scenario, the behavior depicted by Figure 1 becomes quite logical. By replacing
a short word with a long word, one splits the list into a larger number of narratives, which
makes every single word in the list (whether short or long) harder to reach during the
retrieval process.
The interplay between the trajectory of the system during the presentation of lists and
the trajectory during the memory test produces a nontrivial spectrum of behaviors, highly
dependent both on the structure of lists and on the amount of "diffusion" that interferes
with clustering.
34
A telltale symptom of these mechanisms is that a short word is more likely than a longer
one to be recalled right after a word contiguous to it within the list. An analysis of data
from the PEERS experiments (Healey and Kahana, 2016) confirms this prediction.
Another prediction of the theory is that the serial-position jumps performed during recall
will be enhanced by clustering for shorter words, and suppressed for longer ones. This is
also confirmed by an analysis of PEERS data, across the whole available spectrum of jump
lengths.
Several directions stand open for experimental and theoretical work on this model. Ex-
perimentally, further tests may be obtained by performing the five types of measurements
listed in section V. Open directions for theoretical work include: 1. generalizing the recall
probability formulas to the case where the word lengths available are more than two; 2.
including possible competition between words for the verbalization of a given state; 3. sin-
gling out extra effects from the fluctuations around the mean field behavior; 4. accounting
for primacy and recency effects; 5. applying the same technique to predicting the genesis of
false memories.
Finally, by positing a suitable mechanism for spontaneous language production, it would
be useful to derive equations linking the underlying word structure to emerging verbal pat-
terns, thus providing a direct link between the hidden variables and the observables of the
model.
I am grateful to Michael J. Kahana and his University of Pennsylvania research group
for making their raw data publicly available. I would like to thank heartily people at the
Neurotheory Center of Columbia University (Ken Miller, Misha Tsodyks) for hosting me
there in the fall of 2015 and for responding with enthusiasm to this theory. Thanks also to
Yashar Ahmadian of the University of Oregon, for reading an early draft of the paper and
suggesting several improvements.
VII. BIBLIOGRAPHY
Aurenhammer F., Klein R., and Lee D.T., 2013. Voronoi Diagrams and Delaunay Trian-
gulations. Singapore: World Scientific Publishing Company.
Baddeley, A.D., Hitch, G., 1974. Working memory. In G.H. Bower (Ed.), The psychology
of learning and motivation: Advances in research and theory, Vol. 8, pp. 47-89. New York:
35
Academic Press.
Baddeley A.D., Thomson N., Buchanan M., 1975. Word length and the structure of
short-term memory. Journal of Verbal Learning and Verbal Behavior, 14:575-589.
Baddeley, A.D., 2007. Working memory, thought and action. Oxford: Oxford University
Press.
Bhatarad P., Ward G., Smith J. Hayes L., 2009. Examining the relationship between free
recall and immediate serial recall: similar patterns of rehearsal and similar effects of word
length, presentation rate and articulatory suppression. Memory and Cognition 37: 689-713.
Binet A. and Henry. V., 1894. La memoire des mots. L'annee psychologique, Bd.
I
1:1-23.
Bjork, R. A. and Whitten, W. B., 1974. Recency-sensitive retrieval processes in long-term
free recall. Cognitive Psychology, 6: 173189.
Campoy G., 2008. The effect of word length in short-term memory: Is rehearsal neces-
sary? Quarterly Journal of Experimental Psychology, 61:5, 724-734.
Campoy, G., 2011. Retroactive interference in short-term memory and the word-length
effect. Acta Psychol. 138, 135-142.
Coltheart, V., Langdon, R., 1998. Recall of short word lists presented visually at fast
rates: Effects of phonological similarity and word length. Memory and Cognition, 26,
330342.
Elts, J., 1995. Word length and its semantic complexity, in Family and textbooks: 115-126.
Tartu: University of Tartu.
Friendly M., Franklin P.E., Hoffman D., Rubin D.C., 1982. The Toronto Word Pool:
Norms for imagery, concreteness, orthographic variables, and grammatical usage for 1,080
words. Behavior Research Methods And Instrumentation, Vol. 14(4), 375-399.
Fucks W., 1956. Die Mathematischen Gesetze der Bildung von Sprachelementen aus
Ihren Bestandteilen, Nachrichtentechnische Fachberichte 3:7-21.
Greenberg, J., 1966. Universals of language. Cambridge, MA: MIT Press.
Gruenewald, P.J. and Lockhead, G.R., 1980. The free recall of category examples. J.
Exp. Psychol. [Hum- Learn]. 6, 225-240.
Grzybek P., 2007. History and methodology of word-length studies, in "Contributions to
the Science of Text and Language", Dordrecht: Springer, pp. 15-90.
Haspelmath, M., 2006. Against markedness (and what to replace it with). Journal of
36
Linguistics, 42 (01), 25-70.
Healey M., Crutchley P., Kahana MJ., 2014. Individual differences in memory search
and their relation to intelligence. J Exp Psychol 143: 15531569.
Healey, M. K. and Kahana, M. J., 2016. A four-component model of age-related memory
change. Psychological Review, 123(1), 23-69.
Howard, M. W. and Kahana, M. J., 2002a. A distributed representation of temporal
context. Journal of Mathematical Psychology, 46, 269-299.
Howard, M. W. and Kahana, M. J., 2002b. When does semantic similarity help episodic
retrieval? Journal of Memory and Language, 46, 85-98.
Hulme, C., Suprenant, A. M., Bireta, T. J., Stuart, G., and Neath, I., 2004. Abolishing
the word-length effect. J. Exp. Psychol. Learn. Mem. Cogn. 30, 98-106.
Jalbert, A., Neath, I., Bireta, T. J., and Surprenant, A. M., 2011. When does length
cause the word length effect? J. Exp. Psychol. Learn. Mem. Cogn. 37, 338-353.
Kahana M. J., 1996. Associative retrieval processes in free recall. Memory and Cognition
24:103-9.
Kahana M. J., 2012. Foundations of Human Memory, Oxford University Press.
Katkov M., Romani S., Tsodyks M., 2014. Word length effect in free recall of randomly
assembled word lists. Frontiers of Computational Neuroscience 8:129.
Klare, G.R., 1988. The formative years.
In B.L. Zakaluk and S. J. Samules (Eds.),
Readability, its past, present and future, Newark, Delaware: IRA, pp. 14-34.
Kruglanski A.M. and Tory Higgin E., 2007. Social Psychology: Handbook of Basic
Principles, The Guilford Press; Second Edition edition (p.642).
Lohnas, L.J., Polyn S.M.,Kahana, M.J. (2015) Expanding the scope of memory search:
Modeling intralist and intralist effects in free recall, Psychological Revier, Vol. 122, No. 2,
337363.
Lohnas, L. J. and Kahana, M. J. (2013). Parametric effects of word frequency effect in
memory for mixed frequency lists. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 39, 19431946.
Mahowald K., Fedorenko E., Piantadosi S.T., Gibson E., 2012. Info/information theory:
speakers actively choose shorter words in predictable contexts, Cognition, 126: 313-318.
Lewis, M. L. and Frank M.C., 2016. The length of words reflects their conceptual com-
plexity. Cognition 153: 182-195.
37
Mikk. J., Heli U., and Elts J., 2001. Word length as an indicator of semantic complexity,
in Text as a linguistic paradigm:
levels, constituents, constructs, Festschrift in honour of
Ludek Hrebcek. Trier, 187-195.
Miller J.F., Weidemann C.T., Kahana M.J., 2012. Recall termination in free recall.
Memory and Cognition 40: 4, Pages: 540 - 550.
Murdock B-B, 1960. The immediate retention of unrelated words. Journal of Experi-
mental Psychology 60:222-234.
Murdock B. B., 1962. The serial position effect of free recall. Journal of Experimental
Psychology, 64(5)M:482-488.
Murdock B.B. and Okada R., 1970. Interresponse times in single-trial free recall. Journal
of Experimental Psychology 86:263-267.
Neath I., Bireta T.J., Surprenant A.M., 2003. The time-based word length effect and
stimulus set specificity, Psychonomic Bulletin Rev. Jun;10(2):430-4.
Neath I., Brown G. D. A., 2006. SIMPLE: further applications of a local distinctiveness
model of memory, in The Psychology of Learning and Motivation, ed. Ross B. H., editor.
(San Diego, CA: Academic Press), 201-243.
Piantadosi S. T., Tily H. and Gibson E., 2011. Word lengths are optimized for efficient
communication, Proceedings of the National Academy of Sciences, 108, 9:3526.
Piantadosi S. T., Tily H. and Gibson E., 2011B. Reply to Reilly and Kean: Clarifications
on word length and information content, Proceedings of the National Academy of Sciences,
108, 20: E109.
Polyn, S. M., Norman, K. A., and Kahana, M. J., 2009a. A context maintenance and
retrieval model of organizational processes in free recall. Psychological Review, 116, 129-156.
Polyn, S. M., Norman, K. A., and Kahana, M. J., 2009b. Task context and organization
in free recall. Neuropsychologia, 47, 2158-2163.
Roberts W.A., 1972. Free recall of word lists varying in length and rate of presentation:
a test of total-time hypotheses. Journal of Experimental Psychology 92:365-372.
Romani S., Pinkoviezky I., Rubin A., Tsodyks M., 2013. Scaling laws of associative
memory retrieval. Neural Computation 25:2523-2544.
Russo R. and Grammatopoulou N., 2003. Word length and articulatory suppression affect
short-term and long-term recall tasks. Memory and Cognition 31:728-737.
Sederberg, P. B., Howard, M. W., and Kahana, M. J., 2008. A context-based theory of
38
recency and contiguity in free recall. Psychological Review, 115(4), 893-912.
Standing L., 1973. Learning 10.000 pictures. Quarterly Journal of Experimental Psy-
chology 25:207-222.
Tehan G. and Tolan G.A., 2007. Word length effects in long-term memory. Journal of
Memory and Language 56:35-48.
Xu Zhan and Li Bi-Qin, 2009. The Mechanism of Reverse Word Length Effect of Chinese
in Working Memory. Acta Psychologica Sinica, Vol. 41 Issue (09): 802-811.
39
δ = 2
δ = 4
δ = 6
δ = 8
δ = 10
δ = 12
δ = 14
δ = 1
δ = 3
δ = 5
δ = 7
δ = 9
δ = 11
δ = 13
δ = 15
FIG. 13: Probability Pt of serial-position jumps in the recall process, plotted as a function of the
size b of word-length barriers, for all possible values of jump size δ. The red curves have been
calculated in the absence of clustering, as per formula (33). The blue curves are extracted from
PEERS data. Clustering manifests as an enhancement of transition probability for low barriers
(small b) and as a suppression where larger barriers are encountered.
12345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.512345600.5 |
1507.07422 | 1 | 1507 | 2015-07-24T01:05:16 | Analysis of Pain Hemodynamic Response Using Near-Infrared Spectroscopy (NIRS) | [
"q-bio.NC"
] | Despite recent advances in brain research, understanding the various signals for pain and pain intensities in the brain cortex is still a complex task due to temporal and spatial variations of brain hemodynamics. In this paper we have investigated pain based on cerebral hemodynamics via near-infrared spectroscopy (NIRS). This study presents a pain stimulation experiment that uses three acupuncture manipulation techniques to safely induce pain in healthy subjects. Acupuncture pain response was presented and hemodynamic pain signal analysis showed the presence of dominant channels and their relationship among surrounding channels, which contribute the further pain research area. | q-bio.NC | q-bio | The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
ANALYSIS OF PAIN HEMODYNAMIC RESPONSE
USING NEAR-INFRARED SPECTROSCOPY (NIRS)
Raul Fernandez Rojas1, Xu Huang1, Keng Liang Ou2, Dat Tran1 and
Sheikh Md. Rabiul Islam1
1Faculty of Education, Science, Technology and Mathematics,
University of Canberra, Australia
2College of Oral Medicine, Taipei Medical University, Taiwan
ABSTRACT
Despite recent advances in brain research, understanding the various signals for pain and pain intensities
in the brain cortex is still a complex task due to temporal and spatial variations of brain haemodynamics.
In this paper we have investigated pain based on cerebral hemodynamics via near-infrared spectroscopy
(NIRS). This study presents a pain stimulation experiment that uses three acupuncture manipulation
techniques to safely induce pain in healthy subjects. Acupuncture pain response was presented and
Haemodynamic pain signal analysis showed the presence of dominant channels and their relationship
among surrounding channels, which contribute the further pain research area.
KEYWORDS
Near Infrared Spectroscopy, Non-invasive Brain Imaging, Hemodynamic, Acupuncture
1. INTRODUCTION
Pain is a subjective experience and every individual experiences it in different ways. Pain
perception can be range from emotional pain such as the loss of a love one to physical pain such
as a broken leg. In addition, pain sensation as the product of dedicated neural mechanisms
continues to be a topic of debate. But what is pain? It seems to be very hard to define. According
to Perl [1] “With the benefit of the past two centuries of scientific work and thought, can one
define pain? Considering the evidence, it seems reasonable to propose pain to be both a specific
sensation and an emotion, initiated by activity in particular peripheral and central neurons. Pain
shares features with other sensations, but the strong association with disposition is special.”
There is a number of publications about pain research [2-4] to answer questions such as: how to
interpret it, how to control it, or how to use it to indicate human body health status. In this paper
we focus on the instance that an unconscious patient under medical examination cannot express
level of pain experience in that moment. For this reason, this research tries to find a relationship
between the pain intensity and the hemodynamic response to understand the patient’s feelings. By
finding this “silent communication” between pain and brain hemodynamics, medical services will
be able to provide a better solution for such pain sensation.
Near infrared spectroscopy (NIRS) is employed in our experiments. NIRS can offer simultaneous
measurements from dynamic changes of Oxyhemoglobin (HbO) and Deoxyhemoglobin (HbR) in
the brain cortex with a reasonable temporal resolution (less than one second) and spatial
resolution (less than 3 cm), while Total haemoglobin (HbT) is obtained by the difference between
HbO and HbR. Our experimental study makes use of changes in HbO to represent neural activity
DOI : 10.5121/ijma.2015.7203 31
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
in the human brain. In addition, acupuncture is also used to induce pain to patients in a safe
manner.
This paper is organised as follows. Section 2 gives a review of the equipment and experimental
setup. The related analysis methods will be discussed responding to the experimental setup. In
section 3, the experimental results will be investigated, including dynamic NIRS response to
acupuncture stimulation, pain responses of HbO2 and HbR to acupuncture stimulation, pain
responses to different acupuncture manipulations, dominant pattern analysis, and direction of pain
information. In section 4, conclusions are presented.
2. MATERIALS AND METHODS
2.1. Equipment
Hemodynamic measurement was performed using an optical topography system Hitachi ETG-
4000 (Hitachi Medical Corporation). The topography system uses near-infrared spectroscopy
(NIRS) to investigate cerebral hemodynamics. Optical topography makes use the different
absorption spectra of oxygenated and deoxygenated haemoglobin in the near infrared region. The
NIRS system produces 695 and 830 nm NIR signals through frequency-modulated laser diodes.
These NIR signals are transmitted to the brain using optical fibre emitters (Figure 1, red circles).
Near-infrared light penetrates head tissue and bone, generally it reaches 2 to 3 cm into the
cerebral cortex. Once NIR light reaches the cerebral cortex, it is absorbed by haemoglobin, while
the non-absorbed NIR signals are reflected to the source, where it is sampled by a high-sensitivity
photodiode (Figure 1, blue circles). NIR light between emitters and detectors is sampled at a
given time point named channels (Figure 1, numbered squares). Since Oxyhemoglobin (HbO) and
Deoxyhemoglobin (HbR) absorb NIR light differently, two wavelengths of light (695 and 830
nm) are used. In this way, it is possible to read these two types of haemoglobin simultaneously.
The sample frequency used in this experiment was of 10Hz (10 samples per second). The
configuration for this experiment was using two probes of 12 channels to measure neurologic
activity of the head (parietal/temporal region). Figure 1 shows the 24 channel configuration used
in the study, channels 1 to 12 represent the right hemisphere, while channels 13 to 24 represent
left hemisphere. Signals in between channels were computed from the surrounding channels.
2
1
5
4
3
7
6
10
9
8
17
22
12
14
19
24
16
21
11
13
18
23
15
20
Emitter
Detector
Measurement
Channel
Figure 1. Channel configuration, right hemisphere (channels 1-12) and left hemisphere (channels 13-24).
2.2. Experiment Setup and Subjects
The experiment was designed by the School of Oral Medicine of Taipei Medical University
(TMU) in collaboration with the University Of Canberra (UC). In the present study, six healthy
individuals (2 females, 4 males) participated in the experiment, aged 25 to 35 years old. All
participants provided written consent and the experiment was approved by the Ethics Committee
of TMU. The experiment was carried out in the Brain Research Laboratory at TMU in a quiet,
temperature (22-24oC) and humidity (40-50%) controlled room. The experiments were done in
the morning (10:00am-12:00pm) and each experiment lasted around 30 minutes. Quantitative
32
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
data was collected using the ETG-4000 with the patients sat down in an ergonomic chair near the
topography system.
The experiment was designed to recognise pain stimulation and pain release in patients through
hemodynamic responses. In order to identify the pain through NIRS, acupuncture was used to
induce pain stimulation in a safe manner. Brand new acupuncture needles were used for each
experiment, and using traditional Chinese acupuncture techniques that were performed by an
acupuncturist of TMU Hospital. The puncture point used for stimulation was the “Hegu Point”,
located on top of the hand, between the thumb and forefinger. This acupuncture point (acupoint)
is known by its property to relieve pain, especially headaches and toothaches. This acupoint is
also used to reduce fever, eliminate congestion in the nose, stop spasms, and decrease toothache.
A western name for this acupoint is “the dentist’s point” because it can stop tooth pain while
moist the throat and tongue. This point was used because it is an area of easy access and the hand
can be set aside while the patient is relaxed on the chair. Each patient was punctured on both
hands (Figure 2), one hand on a day and the opposite hand on another day; each hand was treated
as a separated experiment. The data base was organised of 12 data sets of changes in
Haemoglobin files, two sets (right and left hands) per each subject.
Figure 2. Patient with acupuncture needle in Hegu point.
Three types of pain stimulations (acupuncture techniques) were applied in the experiment, needle
insertion (Task 1, T1), needle twirl (Task 2, T2), and needle removal (Task 3, T3). The first type
of stimulation (T1) was carried out 30 seconds after the start of experiment and it lasted for 6
seconds. The second type of stimulation (T2) was applied for 30 seconds (rest time) after T1 and
it lasted for 10 seconds, this stimulation was repeated three times after 30 second resting time.
The last stimulation (T3), was carried out after the third application of task 2, and it lasted for 5
seconds with a 30 seconds recovery time and 15 seconds post-time to finish the experiment.
The patients were explained about the acupuncture procedure and experiment. In addition,
patients were given a brief explanation of side effects of acupuncture in case they had any
symptoms during the experiment; no side effects were reported during and after experiment. After
the briefing, the patient was told to sit down on the chair, told to relax and close the eyes to
reduce visual evoked stimulation. The optical topography system was set to record after all
conditions were met and patient was relaxed.
33
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
2.3. Stages of Experiment
Figure 3 exhibits the three types of acupuncture stimulations applied in the experiment. The
image shows the changes of HbO, HbR, and (HbT) concentration during the different stages of
the experiment; for illustration purposes data was taken from subject 4 in channel 19. The first
stage of the study is the pre-time which is used to adapt the patient to the experiment and obtain
baseline for the haemoglobin signals, the pre-time lasted for 30 seconds. The second stage (T1,
task 1) is started immediately after pre-time stage and it begins the first pain stimulus (needle
insertion), and it is done within 6 seconds. After the pain stimulation, patient is given a resting
time of 30 seconds (Rt1), which allows the haemoglobin concentration to return to baseline for
the next stimulation. The next task (T2) is needle twirl, it lasted 10 seconds and is followed by a
30 second resting time (Rt2); this task was repeated two more times including resting time (Rt3,
Rt4) of 30 seconds each. Task three (T3) is started immediately after Rt4 and it lasted for 5
seconds; in this task the needle is removed from the hand. The last recovery time (Rt5) allows the
patient to recover from the previous pain stimulus and also haemoglobin signals return to
baseline, it lasts for 30 seconds. The last stage in the experiment is the Post-time and lasts for 15
seconds, after the end of this time the equipment stops recording data.
)
m
m
M
m
*
i
l
(
n
b
o
g
o
m
e
H
0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
0
Start
Finish
Rt1
Rt3
Rt4
Pre-time
Rt2
Oxy
DeOxy
Total
Rt5
T2
Post-time
T1
T2
T2
T3
500
1000
1500
2000
Samples (10/sec)
Figure 3. Description of tasks of the experiment, Rt=rest time, T=task.
2.4. NIRS colour representation
Figure 4 shows a colour representation of a typical response from a sample of HbO intensity with
the 24 channels of subject number 4. The colour representation helped to distinguish pattern
trends and focus on areas where the dynamic response was more active. Left figure represents the
channel intensities of the right hemisphere, while the right figure represents the channel
intensities of the left hemisphere. The colour intensity runs from -0.3 (blue/cold) to +0.5
(red/hot); being the blue colour the lower measurements of HbO, while the red colour represents
the higher values of HbO. For visual analysis, values above +0.5 and below -0.3 were represented
within this range.
2.5. Data Analysis
Signal Analysis was performed by two signal processing techniques. The first method was the
cross correlation analysis of raw data to identify time-related channel similarities. While the
second method was the Horn and Schunk optical flow (OF) algorithm, this method was used to
identify patterns within both hemispheres. Both methods are described as follows.
34
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
05
10
17
22
02
07
12
14
19
24
04
09
16
21
01
06
11
13
18
23
0.3
0.2
0.1
0
-0.1
-0.2
03
08
15
20
Figure 4. Channel intensity, right (channels 1-12) and left (channels 13-24) hemispheres.
2.5.1. Cross Correlation Analysis
Data analysis was first performed by comparing channels on both sides of the head. Cross
correlation function was computed between channels 1-12 in right probe and 13-24 in left probe.
The cross correlation between two input signals (channels) is a kind of template matching. This
measure of temporal similarity of two signals was done by computing a time-shifting along one of
the input signals. Equation 1 defines the cross correlation between two waveforms x(t) and y(t) as
follows:
ݎ௫௬ሺ߬ሻ = ࢣ ݔሺݐሻݕሺݐ −߬ሻ
ஶିஶ
(1)
Where ߬ is the time-lag between x(t) and y(t), the value of ݎ௫௬ denotes the difference (lag/lead)
between channel signal y(t) and channel signal x(t). The cross correlation value between two
channels in the same probe is done from -40 sec to +40 sec at a rate of 10 samples per second.
2.5.2. Optic Flow Analysis
The optical flow is defined as the “flow” of pixel values at the image plane in time varying
images [5]. Ideally the optical flow corresponds to the motion field, however in some situations
the optical flow is not always equal to the motion field. To overcome any problems in the
calculation of the optical flow, assumptions have to be defined. Firstly, it is assumed that the
surface of the image being observed is flat to avoid variations in brightness due to shading effects.
Secondly, the incident illumination is uniform across the surface. Thirdly, it is also assumed that
the optical flow varies smoothly and has no spatial discontinuities.
We calculate optical flow from the apparent motion at the image plane based on visual perception
with the dimensions of the image velocity of such as motion. We define the optical flow vector
on the image. As the optical flow is estimated from two consecutive images, it appears as a
ݒԦ = ሺݑ,ݒሻ, where u and v represent the x and y components of the optical flow vector at a point
displacement vector ݀Ԧ = ሺ݀௫,݀௬ሻ [6]. Where ݀௫ and ݀௬ denote the x and y components
respectively of the displacement vector at a point.
The Horn and Schunk algorithm [5] implemented in this work is based on differential solving
schemes. It computes image velocity from numerical evaluation of spatiotemporal derivatives of
image intensities. Let denote image intensity (or brightness) at the point (x,y) in the image plane
at time t by E(x,y,t). Thus, the image domain is consequently assumed to be differentiable in
space and time. The basic assumption in measuring image motion is that the brightness of a
particular point in the time-varying image is constant, as a result,
ܧሺݔ,ݕ,ݐሻ = ܧሺݔ + ߜݔ,ݕ +ߜݕ,ݐ +ߜݐሻ = 0,
(2)
35
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
where ߜݔ and ߜݕ are the displacement in x direction and y direction, respectively, after a short
interval of time ߜݐ. This assumption brings to the following condition, known as the “Optical
Flow Constraint Equation”. If brightness varies smoothly with x, y, and t, Equation (2) can be
expanded in a Taylor series and obtain
ܧሺݔ,ݕ,ݐሻ+ ߜݔ డா
డ௫ +ߜݕ డா
డ௬ + ߜݐ డா
డ௧ + ݁ = ܧሺݔ,ݕ,ݐሻ,
(3)
where e contains second and higher-order terms in ߜݔ, ߜݕ, and ߜݐ. Subtracting ܧሺݔ,ݕ,ݐሻ on both
sides, dividing by ߜݐ, and taking the limit as ߜݐ → 0, we have
݀ݔ
݀ݐ +߲ܧ
߲ܧ
݀ݐ +߲ܧ
݀ݕ
߲ݐ = 0
߲ݕ
߲ݔ
ௗ௧ = ݒ, and using the abbreviations ܧ௫ = డா
ௗ௧ = ݑ and ௗ௬
where ௗ௫
ܧ௫ݑ +ܧ௬ݒ + ܧ௧ = 0.
finally obtain Equation (5), which is called the Optical Flow Constraint Equation
(4)
డ௫, ܧ௬ = డா
డ௬, and ܧ௧ = డா
డ௧, we
(5)
The components u and v are used to approximate the velocity component in the direction of the
spatial gradient of the image intensity. Therefore, this is an under-constrained equation, since
only the motion component in the direction of the local gradient of the image intensity function
may be estimated, this is known as “the aperture problem”. It can be seen rewriting Equation (5)
in the form,
ሺܧ௫,ܧ௬ሻ∙ሺݑ,ݒሻ = −ܧ௧,
(6)
thus the component of the movement in the direction of the brightness gradient ሺܧ௫,ܧ௬ሻ is
− ா
ටாೣమାாమ
.
(7)
From Equation (7), it is seen that we cannot determine the component of the optical flow at right
angles to the brightness gradient. As a result, the flow velocity (u,v) cannot be computed locally
without introducing additional constraints [5].
Horn and Schunck introduced a smoothness constraint, supposing that the motion field is smooth
over the entire image domain [5]. This constraint, measures the departure from smoothness in the
velocity flow. To express the smoothness constraint is required to minimize the square of the
magnitude of the gradient of the optical flow velocity as follows,
݁௦ = ඵሺሺݑ௫ଶ +ݑ௬ଶሻ+ሺݒ௫ଶ +ݒ௬ଶሻሻ݀ݔ݀ݕ
the error in the optical flow constraint equation can be expressed as,
݁ = ඵሺܧ௫ݑ + ܧ௬ݒ +ܧ௧ሻଶ݀ݔ݀ݕ
(8)
(9)
Horn and Schunck’s algorithm computes an estimation of the velocity field (u, v) that minimizes
the sum of the errors (Equation 9) for the rate of change of image brightness in equation (3), and
the measure of the departure from smoothness in the velocity flow, equation (8) [5]. By
combining equation (3) and (8) with a suitable weighting factor, and by using calculus of
variation to minimize the total error, it is formulated an iterative solution for optical flow:
ݑାଵ = ݑത − ܧ௫൫ܧ௫ݑത + ܧ௬ݒı + ܧ௧൯
∝ଶ+ܧ௫ଶ + ܧ௬ଶ
(10)
36
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
ݒାଵ = ݒı −ܧ௬൫ܧ௫ݑത +ܧ௬ݒı + ܧ௧൯
Where superscripts indicate the iteration number, subscripts refer to derivation, and α is a positive
constant know as smoothness factor.
ଶ
ଶ ܧ௬
Ýଶ ܧ௫
(11)
3. EXPERIMENTAL RESULTS
3.1. Dynamic NIRS response to acupuncture stimulation
The hemodynamic response was visualised by computing the colour representation in both
probes. Ten serial samples from subject 4 based on HbO data from the 24 channels are presented
in two separated images. Left image represents channels 13 to 24, while right image represents
channels 1 to 12; the ten samples were taken every 22 seconds. Figure 5B shows the
hemodynamic response in channel 1 while Figure 5C shows the hemodynamic response in
channel 17. Red concentrations in Figure 5A indicate higher hemodynamic response to
acupuncture stimulation. The increase and decrease of HbR was less significant than the response
of HbO2, this effect was observed on both sides of the head. In addition, all patients showed a
similar response pattern however different lags and magnitudes among channels were obtained
for each subject. Hemodynamic response to acupuncture stimulation is explained as follows.
1
A
2
8
0
3
0
1
5
2
0
7
8
0
3
0
1
5
2
0
1
3
1
8
2
3
1
3
1
8
2
3
1
6
2
1
1
6
2
1
1
4
1
9
2
4
1
4
1
9
2
4
1
7
2
2
1
7
2
2
0
1
5
0
0
1
5
0
2
1
7
0
2
0
2
1
7
0
2
0
9
0
4
0
9
0
4
0
1
1
6
0
1
0
1
1
6
0
1
0
3
1
5
2
0
8
1
5
2
0
1
3
1
8
2
3
1
3
1
8
2
3
1
6
2
1
1
6
2
1
1
4
1
9
2
4
1
4
1
9
2
4
1
7
2
2
1
7
2
2
0
1
5
0
0
1
5
0
2
1
7
0
2
0
2
1
7
0
2
0
9
0
4
0
9
0
4
0
1
1
6
0
1
0
1
1
6
0
1
0
8
0
3
0
8
0
3
0
4
0
1
5
0
2
1
7
0
2
0
9
0
4
0
1
1
6
0
1
0
8
0
3
0
1
5
2
0
1
3
1
8
2
3
9
0
1
5
0
2
1
7
0
2
0
9
0
4
0
1
1
6
0
1
0
8
0
3
0
1
5
2
0
1
3
1
8
2
3
1
6
2
1
1
6
2
1
1
4
1
9
2
4
1
4
1
9
2
4
1
7
2
2
1
7
2
2
5
1
5
2
0
1
3
1
8
2
3
10
1
5
2
0
1
3
1
8
2
3
8
0
3
0
8
0
3
0
1
6
2
1
1
6
2
1
1
4
1
9
2
4
1
4
1
9
2
4
1
7
2
2
1
7
2
2
0
1
5
0
0
1
5
0
2
1
7
0
2
0
2
1
7
0
2
0
9
0
4
0
9
0
4
0
1
1
6
0
1
0
1
1
6
0
1
0
1
3
1
8
2
3
1
3
1
8
2
3
1
6
2
1
1
6
2
1
1
4
1
9
2
4
1
4
1
9
2
4
1
7
2
2
1
7
2
2
0
1
5
0
0
1
5
0
2
1
7
0
2
0
2
1
7
0
2
0
9
0
4
0
9
0
4
0
1
1
6
0
1
0
1
1
6
0
1
0
8
0
3
0
8
0
3
0
C
1
5
2
0
6
1
5
2
0
0.5
0.4
0.3
0.2
0.1
0
B
)
m
m
M
m
*
i
l
(
n
b
o
g
o
m
e
H
1 2 3 4 5 6 7 8 9 10
)
m
m
*
M
m
(
-0.1
-0.2
220
440
660
880
1100 1320 1540 1760 1980 2200
Samples (10/sec)
Oxy
DeOxy
Total
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
1 2 3 4 5 6 7 8 9 10
220
440
660
880
1100
1320
Samples (10/sec)
1540
1760
1980
2200
Oxy
DeOxy
Total
i
l
n
b
o
g
o
m
e
H
Figure 5. Dynamic haemoglobin response during experiment.
(A) Ten sequential NIRS images, obtained every 22 seconds, and numbered from 1 to 10. The
images show the dynamic change of HbO trough 24 channels. The two images represent
each side of the cerebral hemisphere.
(B) The dynamic trace recorded from channel 1 on the right side of the head. The number
along upper Y axis corresponds to the numbered images in A.
37
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
(C) The dynamic trace recorded from channel 17 on the left side of the head. The number
along upper Y axis corresponds to the numbered images in A.
Figure 6 exhibits the pain dynamic changes of haemoglobin (Hb) of Oxygenated Hb (HbO, in
blue), Deoxygenated Hb (HbR, in red), and Total Hb (HbT, in green); For illustration purposes,
data was taken from channel 3. Markers (dot lines) show the start and end of needle insertion
stimulation (task 1), with a duration of 6 seconds. HbT is the combination of HbO and HbR, in
other words is the sum of HbO and HbR. In a typical response after each acupuncture stimulation,
the HbO presented a slight drop followed immediately for a constant rise to reach a peak, then a
gradual decrease to resting level. On the other hand, HbR presented a slight increase then a
moderate decrease to reach its lowest point; this variation was contrasting the reaction of HbO. In
addition, HbT was dependent of both HbO and HbR, therefore it varied according to the
behaviour of these two signals. It is to note that the estimations of HbO are more accurate than
HbR and HbT [7, 8], for this reason HbO was used to present and analyse the data from this
section onwards.
)
m
m
*
M
m
(
i
l
n
b
o
g
o
m
e
H
0.4
0.3
0.2
0.1
0
-0.1
-0.2
Oxy
DeOxy
Total
B
C
A
300
350
400
450
500
550
600
Samples (10/sec)
Figure 6. Distinctive dynamic response obtained after acupuncture stimulation. A) small drop of HbO after
stimulation, B) followed by a constant increase until reach a peak, C) then a gradual decrease to resting
level.
3.2. Pain responses of HbO to Acupuncture Stimulation
Needle insertion (task 1) was the first acupuncture manipulation in the experiment. It was applied
at thirty seconds after commencement of experiment and lasted for six seconds. Each acupuncture
manipulation was applied within the time indicated by the dot lines (markers) in Figure 7, four
channels were used only for illustration purposes. Once the insertion was done (within samples
301 and 361), the needle was twirled until de-qi (Chinese for obtaining the qi or arriving at the qi)
was reached. Once the patient manifested the numbness/heaviness in the arm, the acupuncturist
stopped the stimulation. This de-qi sensation was approximately reached three seconds after
needle insertion. Figure 7 shows the acupuncture pain response after needle insertion (T1).
Immediately after reaching the highest pain response, approximately after 14 seconds, HbO
dropped towards baseline. It is worth mentioning that in all subjects the highest peak was
exhibited after this first stimulus (needle insertion). This drop-rise-drop behaviour was exhibited
on both sides of the head (Figure 5B,C). Needle insertion was followed by a resting time of 30
seconds. This time was used for the patient to rest after the stimulus and let the haemoglobin
signals go back to base line. It is worth mentioning that there is delay in reaction time between the
physical stimulation and the NIRS signal among all the stimulations, and this is consistent with
previous research using NIRS [9, 10].
38
)
m
m
*
M
m
i
l
(
n
b
o
g
o
m
e
H
y
x
O
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
T1
T2
T2
T2
T3
301
361
661 761
1161
1061
Samples (10/sec)
1461 1561
Ch12
Ch1
Ch7
Ch10
1861
1911
Figure 7. Pain Response after needle insertion.
Needle twirl (task 2) by the acupuncturist and it aims to increase the de-qi sensation. The time
frame for this stimulation is ten seconds (samples 661-761) and was carried out immediately after
resting time (Rt1), refer to Figure 6. The Acupuncturist stopped the twirl manipulation early when
the patient expressed full numbness before the ten seconds. Similarly, the signal pain behaviour
exhibited a small drop first before increasing constantly for 8 seconds. At approximately 14
seconds after the start of the twirl stimulation a large peak of HbO was obtained. Again, this
signal dropped towards baseline after reaching its highest peak. After the stimulus application, the
patient was given 30 seconds to rest.
The second (within samples 1061-1161) and third (within samples 1461-1561) twirl stimulation
exhibited a different pattern than the first stimulation. These responses were smaller than the first
response (see Figure 6), exhibiting a small increase and decrease of HbO after the stimulation. In
contrast to the first needle twirl, the peak was reached after only three seconds in the second
needle twirl; while for the last twirl, the stimulation response was slightly bigger than the
previous twirl. The fact that the hemodynamic response decreased (compared to the first twirl
stimulation) can be described by two factors: 1) the patient was familiar with the twirl
manipulation and became adapted to this sensation, 2) by applying pressure in the hogu point
(well known by its analgesic properties) the patient experienced a painkiller effect that decrease
the effect of pain by needle manipulation. After the stimulus application, the patient was given 30
seconds to rest.
Hemodinamics of needle removal (task 3) are similar to task 1, in this stage the removal was done
within a time frame of five seconds (samples 1861-1911). This stimulation was carried out
immediately after the fourth resting time (Rt4). In this manipulation, the acupuncturist removed
the needle with a fast abrupt pull. Similarly, to all previous stimulations, the pain response to
needle manipulation exhibited a small drop of HbO followed by an increase to reach a peak in
approximately four seconds after needle removal. After needle removal patient was given 40
seconds to rest which was used to let the signals go back to base line and to end the experiment.
3.3. Dominant pattern analysis
After visual analysis, lags between channels were observed which suggested dominant channels
on both sides of the head. As shown in Figure 8, images taken after acupuncture stimulation show
that around channel 7 (dotted lines) the magnitude of HbO is stronger and appears to be the centre
of radiation towards surrounding channels. This is a strong indication that the area near channel 7
39
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
(Ch7) is a dominant area for this particular patient. Figure 8 also reveals that surrounding
channels (Ch4, Ch5, Ch9 and Ch10) had a stronger relationship with Ch7 since their magnitude
increased similarly. On the other hand, channels Ch1, Ch6 and Ch11 had a week relationship
since their magnitude increased later at a slower rate.
05
10
05
10
05
10
02
07
12
02
07
12
02
07
12
04
09
04
09
04
09
01
06
11
01
06
11
01
06
11
03
08
03
08
03
08
Figure 8. Images taken every five seconds after acupuncture stimulation showing dominant channel.
Based on the assumption that there is a relationship among channels, the cross correlation
analysis was computed. Figure 9 shows the time-dependant cross correlation between Ch7 and
other channels for needle insertion stimulus. The results show that channel 7 and surrounding
channels have a small lag time that confirms the results of the visual analysis. An example of this
is the lag between Ch7 and Ch9 (τ=+0.2 sec), or Ch7 and Ch6 (τ=+2.5 sec); these results reflect
the lag time between this dominant region and the time delay among peripheral channels. In
addition, no delay was observed between Ch7 and surrounding channels Ch5 and Ch10. These
results suggest that the hemodynamic of stimulation signals have a progressive movement from
dominant regions towards peripheral regions. It is also important to note that following
stimulation tasks showed a similar pattern, however the lag time is less significant, it suggests that
lag/led time between channels is directly related to the stimulation intensity.
7
h
C
s
v
n
o
i
t
a
e
r
r
o
c
s
s
o
r
C
l
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
Stimulus 1 and rest time
Ch1 (lag=1.0sec)
Ch2 (lag=0.1sec)
Ch4 (lag=0.0sec)
Ch5 (lag=0.0sec)
Ch6 (lag=2.5sec)
Ch9 (lag=0.2sec)
Ch10 (lag=0.0sec)
Ch12 (lag=0.1sec)
5
10
15
20
25
30
35
40
Lag time (sec)
Figure 9. Cross correlation comparison between dominant channel and surrounding channels.
3.4. Direction of Pain Signal Information.
Following the results obtained by cross correlation analysis, optical flow was used to determine
the direction of NIRS signals through the channels. Figure 10 shows the optical field computed
with two consecutive images from channels 1-12 after needle insertion of patient four. It can be
seen that flow of NIRS data goes from channel 7 and it radiates towards surrounding channels. It
supports the previous finding using cross correlation analysis about the existence of dominant
40
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
channels. This “dominant area” or “dominant channel” behaviour was found on both cerebral
hemispheres and is present in all patients.
05
10
05
10
02
07
12
02
07
12
04
09
04
09
01
06
11
01
06
11
03
08
03
08
30
25
20
15
10
5
0
0
5
10
15
20
25
30
Figure 10. Optic flow computed from two consecutive images (samples 300 and 301) of patient four.
It is also important to note that based on optic flow analysis, pain relationship can be made.
Figure 11 presents optic flow fields of four pair of samples taken from the complete acupuncture
manipulation process. In this image, only four channels (Ch2, Ch7, Ch10, and Ch12) are showed
in the image for demonstration purposes. Figure 11A shows the optic flow at four different stages
through the experiment, the first two OF images exhibit the radiation pattern from Ch7 to
peripheral channels; this outer movement is directly related to the slope increase of the pain data.
On the other hand, the last two OF images represent an internal movement from peripheral
channels to Ch7; this inner movement is related to the slope decrease in pain stimulation. In other
words, it is possible to link the OF result with the intensity and progression of pain through the
experiment.
30
25
20
15
10
5
0
0
1
A
B
2
30
25
20
15
10
5
0
0
5
10
15
20
25
30
3
30
25
20
15
10
5
0
0
5
10
15
20
25
30
4
30
25
20
15
10
5
0
0
5
10
15
20
25
30
5
10
15
20
25
30
)
*
m
m
M
m
(
i
l
n
b
o
g
o
m
e
H
y
x
O
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
upward slopes
downward slopes
301
361
661 761
1161
1061
Samples (10/sec)
1461 1561
Ch12
Ch1
Ch7
Ch10
1861
1911
Figure 11. Optic flow results from four consecutive pair of samples, note the slope direction where the
sample where taken. A) Optic flow fields showing the outer and inner flow of information between Ch7
and peripheral channels. B) NIRS data show the slope direction that reflects the intensity level of pain
stimulation.
41
The International Journal of Multimedia & Its Applications (IJMA) Vol.7, No.2, April 2015
4. CONCLUSIONS
In this paper, the distinctive hemodynamic curve response to three types of acupuncture
stimulation was presented. It was showed that the typical pain signal presents a small drop first to
continue with a constant rise until it reaches a maximum to drop again towards resting level. This
behaviour was consistent in all patients after the three types of acupuncture manipulations: needle
insertion, needle twirl, needle removal.
We have also implemented cross correlation analysis and optical flow analysis within the context
of diffuse optical imaging of acupuncture stimulation. Cross correlation confirmed visual analysis
by identifying time-dependant relationship among channels. Regions where the neural activity is
dominant, such as Ch7 in patient four, were also identified. This dominant neural activity was
present in all patients, however magnitude levels varied among patients.
Optical flow corroborated pattern flow from dominant channels to peripheral channels. This
method also distinguished the relationship between the pain stimulation and direction of pattern
flow. It was found that an increase of pain stimulation was reflected as flow of NIRS information
moving towards peripheral channels, while decrease of pain stimulation was reflected as flow of
NIRS signals moving from peripheral channels to dominant channel.
REFERENCES
[1] E. R. Perl, "Ideas about pain, a historical view," Nature Reviews Neuroscience, vol. 8, pp. 71-80,
2007.
[2] M. Bartocci, L. L. Bergqvist, H. Lagercrantz, and K. Anand, "Pain activates cortical areas in the
preterm newborn brain," Pain, vol. 122, pp. 109-117, 2006.
[3] C.-H. Lee, T. Sugiyama, A. Kataoka, A. Kudo, F. Fujino, Y.-W. Chen, et al., "Analysis for distinctive
activation patterns of pain and itchy in the human brain cortex measured using near infrared
spectroscopy (NIRS)," PloS one, vol. 8, p. e75360, 2013.
[4] M. Verner, M. J. Herrmann, S. J. Troche, C. M. Roebers, and T. H. Rammsayer, "Cortical oxygen
consumption in mental arithmetic as a function of task difficulty: a near-infrared spectroscopy
approach," Frontiers in human neuroscience, vol. 7, 2013.
[5] B. K. Horn and B. G. Schunck, "Determining optical flow," in 1981 Technical Symposium East, 1981,
pp. 319-331.
[6] B. D. Lucas and T. Kanade, "An iterative image registration technique with an application to stereo
vision," in IJCAI, 1981, pp. 674-679.
[7] Y. Otsuka, E. Nakato, S. Kanazawa, M. K. Yamaguchi, S. Watanabe, and R. Kakigi, "Neural
activation to upright and inverted faces in infants measured by near infrared spectroscopy,"
NeuroImage, vol. 34, pp. 399-406, 2007.
[8] M. Peña, A. Maki, D. KovacIJ ić, G. Dehaene-Lambertz, H. Koizumi, F. Bouquet, et al., "Sounds and
silence: an optical topography study of language recognition at birth," Proceedings of the National
Academy of Sciences, vol. 100, pp. 11702-11705, 2003.
[9] Y. Honda, E. Nakato, Y. Otsuka, S. Kanazawa, S. Kojima, M. K. Yamaguchi, et al., "How do infants
perceive scrambled face?: A near-infrared spectroscopic study," Brain research, vol. 1308, pp. 137-
146, 2010.
[10] S. Boden, H. Obrig, C. Köhncke, H. Benav, S. Koch, and J. Steinbrink, "The oxygenation response to
functional stimulation: is there a physiological meaning to the lag between parameters?,"
Neuroimage, vol. 36, pp. 100-107, 2007.
42
|
1502.02764 | 1 | 1502 | 2015-02-10T03:07:56 | The Modeling and Quantification of Rhythmic to Non-rhythmic Phenomenon in Electrocardiography during Anesthesia | [
"q-bio.NC",
"cs.CE"
] | Variations of instantaneous heart rate appears regularly oscillatory in deeper levels of anesthesia and less regular in lighter levels of anesthesia. It is impossible to observe this "rhythmic-to-non-rhythmic" phenomenon from raw electrocardiography waveform in current standard anesthesia monitors. To explore the possible clinical value, I proposed the adaptive harmonic model, which fits the descriptive property in physiology, and provides adequate mathematical conditions for the quantification. Based on the adaptive harmonic model, multitaper Synchrosqueezing transform was used to provide time-varying power spectrum, which facilitates to compute the quantitative index: "Non-rhythmic-to-Rhythmic Ratio" index (NRR index). I then used a clinical database to analyze the behavior of NRR index and compare it with other standard indices of anesthetic depth. The positive statistical results suggest that NRR index provides addition clinical information regarding motor reaction, which aligns with current standard tools. Furthermore, the ability to indicates the noxious stimulation is an additional finding. Lastly, I have proposed an real-time interpolation scheme to contribute my study further as a clinical application. | q-bio.NC | q-bio | Graduate Institute of Biomedical Electronics and Bioinformatics
College of Electrical Engineering and Computer Science
National Taiwan University
Doctoral Dissertation
The Modeling and Quantification of
Rhythmic to Non-rhythmic Phenomenon
in Electrocardiography during Anesthesia
Author : Yu-Ting Lin
Advisor : Jenho Tsao, Ph.D.
February 2015
5
1
0
2
b
e
F
0
1
]
.
C
N
o
i
b
-
q
[
1
v
4
6
7
2
0
.
2
0
5
1
:
v
i
X
r
a
"All composite things are not constant. Work hard to gain your own enlightenment."
Siddh¯artha Gautama
Abstract
Variations of instantaneous heart rate appears regularly oscillatory in deeper levels of
anesthesia and less regular in lighter levels of anesthesia. It is impossible to observe this
"rhythmic-to-non-rhythmic" phenomenon from raw electrocardiography waveform in cur-
rent standard anesthesia monitors. To explore the possible clinical value, I proposed the
adaptive harmonic model, which fits the descriptive property in physiology, and provides
adequate mathematical conditions for the quantification. Based on the adaptive har-
monic model, multitaper Synchrosqueezing transform was used to provide time-varying
power spectrum, which facilitates to compute the quantitative index: "Non-rhythmic-
to-Rhythmic Ratio" index (NRR index). I then used a clinical database to analyze the
behavior of NRR index and compare it with other standard indices of anesthetic depth.
The positive statistical results suggest that NRR index provides addition clinical infor-
mation regarding motor reaction, which aligns with current standard tools. Furthermore,
the ability to indicates the noxious stimulation is an additional finding. Lastly, I have
proposed an real-time interpolation scheme to contribute my study further as a clinical
application.
Keywords: instantaneous heart rate; rhythmic-to-non-rhythmic; Synchrosqueezing trans-
form; time-frequency analysis; time-varying power spectrum; depth of anesthesia; electro-
cardiography
Acknowledgements
First of all, I would like to thank Professor Jenho Tsao for all thoughtful lessons, discus-
sions, and guidance he has provided me in the last five years. His immense knowledge
in electrical engineering has also been a great addition to assisting me in my research. I
thank Professor Hsiao-Lung Chan, Professor Kung-Bin Sung, Professor Nien-Tsu Huang,
and Professor Hau-tieng Wu for joining my dissertation defense committee. I would also
like to thank Professor Chih-Ting Lin for his great advice in my proposal defense.
I would like to thank Professor Yi Chang and all my colleagues in the department of
anesthesia, Shin Kong Wu Ho-Su Memorial Hospital, for supporting me on seeking the
answer to my clinical question. I am particularly grateful to Cheng-Hsi Chang M.D. for
his constant assistance and advice in regards to my clinical practice and my research.
I would like to thank Professor Hau-tieng Wu for his inspiration and collaboration.
It
has been a great blessing that we were both able to work together and understand the
problems from the perspective as a scientist, engineer, and physician. This has helped us
drawn wonderful and meaningful results. I am very grateful for the valuable assistance
and discussion from all the members in Signal Processing Laboratory. I am particularly
grateful to Pei-Feng Lin, Ming-huang Chen, Zongmin Lin for all their valuable assistance
and inspiring discussions during my stay at the NTU campus.
I would like to express my sincere gratitude to Hui-Hsun Huang M.D. in the department of
anesthesia, National Taiwan University Hospital. His pioneer work on using time-frequency
analysis to attack the IHR problem in anesthesia hugely inspired me since my residency.
I would also like to thank Professor Charles K. Chui for his guidance in seeking a solution
on real-time interpolation.
My acknowledgements would be incomplete without recognition of Shu-Shya Hseu M.D.
and Huey-Wen Yien M.D. for their extraordinary mentorship since I was a medical student.
My greatest appreciation are reserved for my family. My parents and brother have been
the constant support of my life. My daughter and my son have always brought me joy.
Lastly, I am most grateful to my wife's consistent dedication to me since we have met.
iii
Contents
Abstract
Acknowledgements
Contents
List of Figures
List of Tables
Abbreviations
1 Introduction
1.1 Field of Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 An Obscure Phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Quantification for Depth of Anesthesia . . . . . . . . . . . . . . . . . . . . .
1.4 Project Motivation and Goals . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Thesis Outline
2 Background and Previous Work
2.1 Historical Background in Anesthesia . . . . . . . . . . . . . . . . . . . . . .
2.2 Current View on Anesthesia . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Perspective from Anesthesiology . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Anesthesia and Long-Term Mortality . . . . . . . . . . . . . . . . . . . . . .
2.5 Perspective from Physiology . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 Physiology of Amplitude and Frequency Modulation . . . . . . . . . . . . .
2.7
Instantaneous Heart Rate and Heart Rate Variability . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
Integration of Multidisciplinary Backgrounds
2.8
ii
iii
iv
vii
viii
ix
1
1
1
1
2
3
4
4
4
5
6
6
7
8
9
3 Theory and Modeling
10
3.1 Proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Description of Proposed Methodology . . . . . . . . . . . . . . . . . . . . . 10
3.3 Adaptive Harmonic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4 Non-rhythmic to Rhythmic Ratio . . . . . . . . . . . . . . . . . . . . . . . . 12
3.5 Non-rhythmic Component vs. Stochastic Process . . . . . . . . . . . . . . . 15
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
iv
Contents
v
4 Time-Frequency Analysis, Reassignment, and Synchrosqueezing
17
4.1 Time-Frequency Analysis
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Reassignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3 Synchrosqueezing Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4 Multitaper Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.5 Multitaper Synchrosqueezing Spectrogram . . . . . . . . . . . . . . . . . . . 23
4.6 Adaptive Filter and Parametric Estimation . . . . . . . . . . . . . . . . . . 24
. . . . . . . . . . . . . 26
4.7 Common Condition in Physiology and Mathematics
5 Performance Evaluation using Clinical Data
27
5.1 Clinical Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.2 Anesthesia Protocol
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.3 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.4 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.5 Serial Prediction Probability Analysis in Predicting Anesthetic Events . . . 30
5.6 Correlations with Sevoflurane Concentration . . . . . . . . . . . . . . . . . . 31
5.7 Algorithm of Serial Prediction Probability (sPK) Analysis . . . . . . . . . . 31
6 Results from Clinical Database
33
6.1 Visual Information in TF Plane . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.2 Multiple Component Phenomenon . . . . . . . . . . . . . . . . . . . . . . . 34
6.3 Ability to Predict Anesthetic Events . . . . . . . . . . . . . . . . . . . . . . 37
6.4 Correlation with Sevoflurane Concentration . . . . . . . . . . . . . . . . . . 41
6.5 Quantitative Results of Noxious Stimulation . . . . . . . . . . . . . . . . . . 42
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7 Implications to Anesthesiology and Physiology
49
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
7.1 Main Findings
7.2 Genuineness of NRR index . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
7.3 Existing Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.4 Physiologic Interpretation of NRR . . . . . . . . . . . . . . . . . . . . . . . 50
7.5 Clinical Application of the NRR Index . . . . . . . . . . . . . . . . . . . . . 52
7.6 Potential Index in Noxious Stimulation . . . . . . . . . . . . . . . . . . . . . 52
7.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
8 Real-time Processing Using the Blending Operator
54
8.1 Real-time Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
8.2 Technique Obstacles for Real-time Processing . . . . . . . . . . . . . . . . . 55
8.3 B-spline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
8.4 Quasi-interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
8.5 Local Interpolation Operator
. . . . . . . . . . . . . . . . . . . . . . . . . . 59
8.6 Blending Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
8.7 Real-time tvPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
9 Conclusion
63
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
9.1 Research Findings
9.2 Accomplishments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
9.3 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Contents
vi
9.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Bibliography
66
List of Figures
1.1 Rhythmic-to-Non-rhythmic phenomenon . . . . . . . . . . . . . . . . . . . .
2.1 R-R peak interval and instantaneous heart rate . . . . . . . . . . . . . . . .
2
8
3.1 The Calculation of Rhythmic Power and Non-rhythmic Power . . . . . . . . 14
4.1 Color Graph: STFT spectrogram vs. Reassigned spectrogram . . . . . . . . 19
4.2 STFT spectrogram vs. Reassigned spectrogram . . . . . . . . . . . . . . . . 21
4.3 The time-varying power spectrum: multitaper Synchrosqueezing spectrogram 24
4.4 Demonstration of multitaper Synchrosqueezing using artificial data . . . . . 25
6.1 Dynamic Feature in Anesthesia . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.2 Rhythmic-to-Non-rhythmic phenomenon, NRR index and depth of anesthe-
sia varying during controlled ventilation . . . . . . . . . . . . . . . . . . . . 35
6.3 Rhythmic-to-Non-rhythmic phenomenon, NRR index and depth of anesthe-
sia during spontaneous breathing . . . . . . . . . . . . . . . . . . . . . . . . 36
6.4 Skin Incision during Anesthesia . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.5 Rhythmic-to-Non-rhythmic Phenomenon Demo no.1 . . . . . . . . . . . . . 38
6.6 Rhythmic-to-Non-rhythmic phenomenon Demo no.2 . . . . . . . . . . . . . 38
6.7 Rhythmic-to-Non-rhythmic phenomenon Demo no.3 . . . . . . . . . . . . . 39
6.8 Multiple Components in IHR during Anesthesia . . . . . . . . . . . . . . . . 39
6.9 Two Components in IHR during Anesthesia . . . . . . . . . . . . . . . . . . 40
6.10 Multiple components and rhythmic-to-non-rhythmic phenomenon . . . . . . 40
6.11 Multiple components and rhythmic-to-non-rhythmic phenomenon no.2 . . . 41
6.12 Serial PK Analysis: Loss of Consciousness . . . . . . . . . . . . . . . . . . . 42
6.13 Serial PK Analysis: Skin Incision . . . . . . . . . . . . . . . . . . . . . . . . 43
6.14 Serial PK Analysis: First Reaction . . . . . . . . . . . . . . . . . . . . . . . 44
6.15 Serial PK Analysis: Return of Consciousness . . . . . . . . . . . . . . . . . . 45
6.16 Serial PK Analysis: Endotracheal Intubation . . . . . . . . . . . . . . . . . 46
9.1 Diagram of "rhythmic-to-non-rhythmic phenomenon" . . . . . . . . . . . . 64
vii
List of Tables
4.1 Comparison of TF analysis methods
. . . . . . . . . . . . . . . . . . . . . . 26
6.1 Representative Value of Indices . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.2 Correlation to Estimated Sevoflurane Effect-site Concentration . . . . . . . 48
viii
Abbreviations
NRR
HR
IHR
HRV
ECG
Non-Rhythmic-to-Rhythmic
Heart Rate
Instantaneous Heart Rate
Heart Rate Variability
ElectroCardioGraphy
PK analysis Prediction probability analysis
BIS index
Bispectral index
PS
tvPS
STFT
CWT
TF
RRI
EDR
Power Spectrum
time-varying Power Spectrum
Short Time Fourier Transform
Continuous Wavelet Transform
Time-Frequency
R-peak to R-peak Intevral
ECG-Derived Respiration
HF power
High Frequency power
LF power
Low Frequency power
tvHF power
time-varying High Frequency power
tvLF power
time-varying Low Frequency power
ix
To My Family and to the Memory of My Grandparents.
x
Chapter 1
Introduction
1.1 Field of Project
The human body undergoes unique conditions when under anesthesia. Anesthesiologists
administer anesthetics, also known as anesthetic medicine, dependent on the patient's
physiological information. The physiological information may include the patient's heart
rate, blood pressure, and electroencephalography (EEG) as this information reflects the
inner dynamics of the human body. My study will present a new perspective on the
physiological information that could contribute to the clinical practice of anesthesia.
1.2 An Obscure Phenomenon
As a practicing anesthesiologist, I have noticed that the instantaneous heart rate (IHR)
fluctuates dependent on the various levels of anesthesia -- it became regularly oscillatory in
deeper level of anesthesia, whereas it became less regularly in lighter level.(see Fig.1.1) This
phenomenon, I refer to as "rhythmic-to-non-rhythmic phenomenon", has been described
qualitatively in the past. However, to the best of my knowledge, no quantitative method
has been proposed. Also, the underlying physiological mechanism has not been addressed.
1.3 Quantification for Depth of Anesthesia
This phenomenon has induced three questions -- 1) how to quantify this phenomenon into
a useful index, 2) whether the quantitative index can reflect the level of anesthesia, and
3) what are the underlying physiological mechanisms.
1
Chapter 1. Introduction
2
Figure 1.1:
The tracing of R-R peak interval (RRI) represents instantaneous heart rate (IHR) ex-
hibiting "rhythmic" or "non-rhythmic" patterns during anesthesia. This phenomenon is
difficult to be read from the raw electrocardiography (ECG) waveform. Interestingly, the
ECG waveform is mandatory in standard anesthesia monitor, but IHR is not.
For an anesthesiologist, information from this phenomenon is unavailable in the clini-
cal perspective. From Fig.1.1, it is apparent that the "rhythmic-to-non-rhythmic" phe-
nomenon can be seen from the tracing of instantaneous heart rate, but not from raw
electrocardiographic (ECG) waveform. Although real-time monitoring of ECG is manda-
tory in the modern anesthesia monitor, the tracing of instantaneous heart rate is not dis-
played. Even though an anesthesiologist was aware of this phenomenon, he cannot obtain
"rhythmic-to-non-rhythmic" phenomenon information from modern anesthesia monitor to
improve the anesthetic management.
If NRR index can provide additional clinical value, I believe that it could help in future
developments of standard anesthesia monitor.
1.4 Project Motivation and Goals
My clinical observation leads to an intuition that this phenomenon arises from the in-
ner physiologic dynamics under surgery and anesthesia. Hence, the quantification could
provide valuable information to help anesthesiologists in administering anesthetics.
Pursuing these questions involves model construction based on knowledge of three distinct
fields: signal processing, clinical anesthesiology and physiology. Then, I will propose
a quantitative index, referred to as "non-rhythmic-to-rhythmic ratio"(NRR) to quantify
the phenomenon. The computation of NRR index requires multitaper Synchrosqueezing
transform, a recently developed time-frequency analysis technique.
Chapter 1. Introduction
3
I used a clinical database to explore the property of NRR index. Meanwhile, NRR index
was compared with standard depth-of-anesthesia indices used in current clinical practicing
to understand its performance.
1.5 Thesis Outline
My thesis will begin with a literature review of the clinical anesthesiology, neural physiol-
ogy, and time-frequency analysis. Bringing multidisciplinary knowledge up to date leads to
a theory for the possible mechanism of rhythmic-to-non-rhythmic phenomenon. I proposed
a model as the foundation of further analysis using time-frequency analysis. The adaptive
harmonic model is compatible in both ways: fitting the descriptive property in physiol-
ogy, and providing adequate mathematical conditions for the quantification. Based on the
adaptive harmonic model and multitaper Synchrosqueezing transform, NRR index was
developed to quantify rhythmic-to-non-rhythmic phenomenon. I used a clinical database
to analyze the behavior of NRR index, and compare it with other current standard indices
of anesthetic depth. The statistical result shows positive value in clinical: NRR index
provides information regarding motor reaction earlier and better than all the other depth-
of-anesthesia indices. The result suggests a possible anatomic area and functional part
in the brain that mediate the rhythmic-to-non-rhythmic phenomenon. To benefit each
individual patient undergoing surgery and anesthesia, NRR index should be implemented
in real-time. An optimal spline interpolation technique is proposed for the development
of "real-time NRR index".
Chapter 2
Background and Previous Work
2.1 Historical Background in Anesthesia
More than seventy years ago, Guedel made a remark on the connection between anesthesia
and respiratory patterns[1]. It appears that there is a direct relation between anesthetic
depth and respiratory patterns. A deeper level of anesthesia results in regular respiratory
patterns while a lighter level of anesthesia results in irregular respiratory patterns. Dur-
ing that era, depth of anesthesia were concluded from physical signs, such as the changes
in pupil size and movement of the eyeballs. However, the introduction of muscle relax-
ant in clinical anesthesia has hindered the observation of muscle movement and Guedel's
assessment method.
Even though Guedel's observations became less known, this "rhythmic-to-non-rhythmic"
phenomenon has been documented in other literature. About twenty years ago, Kato used
power spectrum to analyze IHR in anesthesia[2]. His study showed that the spectrum
appears to be concentrated during deeper level of anesthesia; whereas the spectrum is
dispersing in lighter level of anesthesia. Kato also mentioned the respiratory effect on
this phenomenon. It is possible that the power spectrum technique helps to capture the
obscure phenomenon in IHR at that era.
2.2 Current View on Anesthesia
Today, we would like to further explain this phenomenon through the advancement of
various scientific fields in signal processing, anesthesiology, and neural physiology.
Because of the advancement in anesthesiology, we now understand that anesthetics exert
effect mainly on the brain. Distinct anatomic location mediates different functions, with
4
Chapter 2. Background
5
differential susceptibility to same anesthetics. Furthermore, each patient has his own
unique genetic profile that renders diverse response to anesthetics. The inter-individual
genetic variability, plus the dynamic influence of surgical procedure, make it difficult to
determine appropriate anesthetic levels based on the dosage of anesthetics or single index
of anesthetic depth. There is strong evidence that anesthetic management can improve the
long-term health of patients, it is important to have monitoring instruments facilitating
a comprehensive anesthetic management specific for each unique patient. Hence, the
quantification of "rhythmic-to-nonrhythmic" phenomenon should be worthwhile.
Current advancements in signal processing has resulted in the development of time-frequency
analysis. In particular, reassignment and Synchrosqueezing technique were not readily ac-
cessible in the past. Furthermore, current advancements in neural physiology has also
provided more insight into the underlying mechanism of the "rhythmic-to-non-rhythmic"
phenomenon, which has not been addressed according to my best knowledge. A more
detail literature review regarding three main disciplines is as follows.
2.3 Perspective from Anesthesiology
Anesthesia comprises several components, including hypnosis, analgesia, immobility, am-
nesia, and autonomic nerve system stability[3 -- 6]. For each patient, the dosage for different
anesthetics is tailored to the patient's specific surgical needs. Therefore, anesthesiologists
®
need to continuously monitor the patient's response to the anesthesia. Bispectral Index
(BIS), an index computed from electroencephalography(EEG), is the current standard
tools used to monitor depth-of-anesthesia for either the hypnotic component or the level
of sleepiness. However, analgesia component, commonly known as the inhibition of pain
sensation in anesthesia, requires an indirect interpretation of physiologic information such
as heart rate (HR) and blood pressure[7 -- 11]. It is impossible to achieve optimal anesthe-
sia using only one kind of anesthetics. Also, anesthesiologists are unable to monitor the
depth of anesthesia by using merely one index, consequently inadequately measuring anes-
thetic effects. For a more comprehensive administration of anesthesia, it may be helpful
to integrate different types of information to come to a decision.
It is important to understand what is wrong with simply relying on BIS monitor. BIS is
a processed EEG, which measures the activity of the frontal cortical area only. On the
other hand, amnesia is the suppression of memory, which is governed by medial temporal
cortical area. Analgesia is the suppression of pain sensation, which is governed by various
subcortical areas. Immobility is the inability of movement which is typically controlled
by the thalamus and spinal cord[7, 12, 13]. Lastly, the entire autonomic nerve system is
controlled by the brainstem[14 -- 16]. Although BIS index is the current gold standard of
Chapter 2. Background
6
anesthetic depth monitor, BIS sensor is unable to measure the activities from all above
listed anatomical areas except the frontal cortex. Literature in recent years indicates that
anesthetics exert differential effects on different brain areas[3, 7, 17]. Thus, there is a
considerable risk measuring frontal cortex to conclude the anesthetic effect on other brain
areas. For example, researchers have reported that the BIS monitor cannot reduce the rate
of intra-operative awareness, which requires adequate memory function suppression[18].
2.4 Anesthesia and Long-Term Mortality
Anesthesia is more than merely falling asleep. Evidences support the connection between
anesthesia and long-term mortality. Too deep of hypnotic index (BIS index) for too long
in surgery is associated with worse one-year mortality[19]. It has also been reported that
different anesthetics cause different immune responses, which affect tumor metastasis[20].
Evidence suggests incorporating "stress reduction" strategy into anesthetic management
leads to longer life expectancy. First, infants who needed surgery for their congenital
heart disease had lower long term mortality rate if they received adequate analgesics[21].
Second, more adult patients who underwent major surgeries survived one year later if they
receive adequate medication to stabilize autonomic nerve system[22, 23]. Lastly, when
epidural anesthesia, a regional anesthesia technique, is combined with general anesthesia
for patients undergoing major surgeries, studies have showed that epidural anesthesia
reduces stress and pain, leads to faster recovery, and overall better outcome[24]. There
are more compelling studies that support the benefit of anesthesia management on long
term outcome. These evidences suggest that a comprehensive optimization of anesthesia
is worthwhile.
In summary, anesthesia comprises several components. A comprehensive anesthesia man-
agement that covers several components could benefit patient for many years. Liteurature
survey also brings hope that quantitative analysis of "rhythmic-to-non-rhythmic" phe-
nomenon could be useful to monitor the anesthetic effect on subcortical regions.
2.5 Perspective from Physiology
From a physiological perspective, the underlying mechanism of the "rhythmic-to-non-
rhythmic" phenomenon could be partially explained by breathing mechanism and car-
diopulmonary coupling[14]. It is known that the neural respiratory control comprises of
two systems, 1) the involuntary automatic control system and 2) the voluntary control
system. The involuntary automatic control system is controlled by the respiratory center
Chapter 2. Background
7
in the brainstem. The voluntary control system is controlled by the forebrain[25]. The
anatomies of these two systems are distinct from one another. The respiratory neural
signals from these two systems compete with each other but are both integrated at the
spinal cord to control the respiratory motor neuron. During spontaneous respiration, the
involuntary automatic respiratory pattern generated in the preBotzinger complex [25 -- 29],
is rhythmic. On the contrary, the breathing pattern of the voluntary respiratory motor
control is non-rhythmic, which involves larger cerebral areas, including cortical processing
and thalamic integration[25, 30 -- 34].
The suppression effect of anesthetics on the forebrain, including the cortex and the tha-
lamus, is stronger than that on the brainstem[35]. Meanwhile, studies have shown that
the preBotzinger complex, the involuntary respiratory control center, is less susceptible
to the anesthetics[36, 37]. These relations suggest that under deeper level of anesthesia,
the non-rhythmic respiratory is more suppressed than the rhythmic respiratory center.
Hence, the IHR exhibits oscillation that is more rhythmic during deeper anesthetic level.
Contrarily, the non-rhythmic respiratory center, being less suppressed during lighter level
of anesthesia, causes activity more non-rhythmic in IHR. This concludes that IHR exhibits
more nonr-hythmic. In summary, literature in physiology supports the correlation that
the differential anesthetic effects between the involuntary control system and voluntary
control system induces the "rhythmic-to-non-rhythmic" phenomenon in IHR.
Thus, I hypothesize that IHR exhibits the central respiratory activity via cardiopulmonary
effect[14], and it reflects the integration of rhythmic and non-rhythmic respiratory activ-
ities. Although more evidence is necessary to clarify this hypothesis, I propose using the
NRR quantification methodology as a potential tool to evaluate the depth of anesthesia
from a different aspect than using EEG-based monitoring.
2.6 Physiology of Amplitude and Frequency Modulation
The human body constantly regulates the respiration to meet the requirement of metabolism.
In other words, the human body has to inhale oxygen (O2), and expel carbon dioxide (CO2)
adequately. The respiratory control center in the brainstem receives the feedback infor-
mation of oxygen concentration and the acidity to modulates the volume and the rate of
breath. The acidity is due to the fact that CO2 exists as hydrogen ion (H+) and carbonic
acid in human body[16, 38].
Since CO2 are the metabolic products of human body, the H+ concentration and O2
concentration in the human body are gradually, but consistently, changing. Thus, we
can describe that the respiration signal in the brainstem is oscillatory with amplitude
Chapter 2. Background
8
modulation (AM) and frequency modulation (FM). The AM and FM are continuous and
slow change in time. These physiologic features can be depicted in mathematical language
for subsequent modeling and quantification in the next chapter.
In summary, literature from physiology supports an integrated rhythmic and non-rhythmic
neural activity reflecting the level of anesthesia. From a clinical anesthesiology perspective,
technological advancements on monitoring various aspects of anesthetic depth can help
improve the quality of overall patient care.
2.7 Instantaneous Heart Rate and Heart Rate Variability
Figure 2.1: A schematic diagram for the R-R interval (RRI) signal derived from the
raw lead III electrocardiographic (ECG) signal. We refer to RRI signal as instantaneous
heart rate. The diagram also showed the ECG-derived respiration (EDR)
The term heart rate is important to all anesthesiologists and is usually referred to as the
number of heartbeats within one minute. To an anesthesiologist, an ongoing heart rate
indicates the stability of the heart condition, or the stress responses to surgical stimulation.
The heart rate is mainly controlled by both sympathetic nerve and parasympathetic nerve
system from the brainstem to the heart[16]. Due to the dynamic nature of patient's under
surgery and anesthesia, we do not feel comfortable with the heart rate information at
an one-minute interval. Instead, a standard requirement in anesthesia practice includes
a continuous real-time display of the heart rate.
In this thesis, the "rhythmic-to-non-
rhythmic" phenomenon is only observable at the instantaneous level. Conversely, the
Chapter 2. Background
9
average heart rate of a normal routine excludes the information of "rhythmic-to-non-
rhythmic" phenomenon.
Both the Instantaneous heart rate (IHR) and ECG derived respiration (EDR),(Fig.2.1) can
be derived from the raw ECG waveform[39 -- 41]. Although we can say that the "rhythmic-
to-non-rhythmic" phenomenon in IHR is a kind of "generalized" heart rate variability
(HRV), a medical guideline has set up the standard method of HRV analysis[42]. Un-
fortunately, the standard analysis techniques stated in this guideline cannot reveal the
information of the "rhythmic-to-non-rhythmic" phenomenon. In particular, the SDNNx,
pNNx or the power spectrum technique measures information in a specified period, which
inevitably neglects the dynamic and transient changes within this period. The guideline
discusses but does not resolve the potential issue of stationarity in the data. This also
supports my desire to rely on the current scientific advancements on signal processes to
resolve the current medical shortcomings in my project.
2.8 Integration of Multidisciplinary Backgrounds
Integration of multidisciplinary backgrounds provides partially theoretical background for
the mechanism of "rhythmic-to-non-rhythmic" phenomenon. Below is a list of key facts
from various disciplines:
1. In respiratory control, non-rhythmic activity originates from the forebrain, whereas
rhythmic activity originates from the brainstem.
2. The forebrain has far more neurons and synapses when compared to the brainstem.
3. Neural structures with more synapses are more susceptible to anesthetics
4. Respiration have an coupling effect on IHR
Fact 1 and 2 are knowledge in neurology and neural anatomy. Fact 3 is from anesthesiology,
and fact 4 is physiology knowledge. Each single fact seems ordinary when considered in
its own disciplines. However, when I combine them together, the theoretical ground is
justified. Then, we can proceed to the modeling problem in the next chapter.
Chapter 3
Theory and Modeling
3.1 Proposal
The "rhythmic-to-non-rhythmic" phenomenon in IHR[2, 43] suggests that IHR during
anesthesia consists of two components: one more rhythmic, less affected by anesthesia,
and one more non-rhythmic, suppressed by deep anesthesia, and the strength of these two
components vary according to the anesthesia depth.
I hypothesize that quantifying the "rhythmic-to-non-rhythmic" phenomenon will reflect
the level of anesthesia. In order to quantify this change in IHR, I will introduce a novel
index which I referred to as Non-rhythmic to Rhythmic Ratio (NRR) index.
3.2 Description of Proposed Methodology
My hypothesis will be tested through a four step method. I will capture the changes in
rhythm through IHR using the adaptive harmonic model. Next, I will use the multitaper
Synchrosqueezing transform to extract the dynamic behavior from the ECG signals. In
order to quantify "rhythmic-to-non-rhythmic" phenomenon, I will use the NRR index.
Lastly, I will use a clinical database to compare the NRR index with other anesthetic
depth indices.
3.3 Adaptive Harmonic Model
It is important to consider which appropriate model to utilize to adequately describe
and quantify this rhythmic-to-non-rhythmic phenomenon. Specifically, it is important
10
Chapter 3. Modeling
11
to recall the high difficulty on describing the inner dynamic of the human body through
mathematical formulas. This difficulty is due to the fact that the human body is a complex
system creating non-linear and unpredictable interactions with the outer environment. The
drastic influences of surgery and anesthetic medication complicates the difficulty further.
It is fairly simple to ruin the capability of a model to describe the reality by imposing
unrealistic conditions. That is the most challenging task currently at hand when we try to
model the biological phenomenon of our study. In order to do so we have to be extremely
cautious on what conditions we choose and impose on our study model.
After careful observation from a time-varying spectrum, I propose a relatively conservative
model to analyze the rhythmic part of IHR. As is shown in Fig.1.1, the definition of
rhythmic implies that the oscillation exhibits time-varying frequency and time-varying
amplitude. However, the modulations are in a relatively slow rate. This presents the notion
that the relative predictable oscillatory activity can be perceived as a rhythmic movement.
Hence, a more conservative phenomenological modeling, also known as adaptive harmonic
model, will be proposed as follows.
The adaptive model integrates with the Synchrosqueezing transform creating a whole new
scheme starting from the theoretical grounds to the clinical phenomenon. We start from
motivating the technical model we consider to analyze IHR(t). It is commonly observed in
clinical settings that stronger anesthetic levels are associated with more regular patterns
in the instantaneous heart rate IHR(t), which is caused by the cardiopulmonary coupling
effect under anesthesia [2, 43]. Based on this major observation, our treatment of the
IHR(t) signal will be purely phenomenological; that is, the parameters and indices we will
derive from the IHR(t) signal will be based solely on these signals themselves, and not
on explicit, quantitative models of the underlying mechanisms. Since we mainly extract
IHR from ECG signal as RRI, we will consider the following phenomenological model for
RRI(t):
RRI(t) = T (t) + A(t) cos(2πφ(t)),
(3.1)
where T (t) > 0, A(t) > 0 and φ(cid:48)(t) > 0. We call T (t) the trend, which captures the
"average heart rate" over a long period; we require the trend to be positive, but not to
be constant; we allow it to vary in time, as long as the variations are small enough. We
call the derivative φ(cid:48)(t) of the function φ(t) the instantaneous frequency (IF) of RRI(t);
we require IF to be positive, but not to be constant; we allow it to vary in time, as long
as the variations are slight from one period to the next, i.e. φ(cid:48)(cid:48)(t) ≤ φ(cid:48)(t) for all time t,
where is some small, pre-assigned positive number. Likewise, We call A(t) the amplitude
modulation (AM) of R(t), which should be positive and smaller than T (t), but is allowed to
vary slightly as well, i.e. A(cid:48)(t) ≤ φ(cid:48)(t) for all time t. In summary, we have the following
Chapter 3. Modeling
conditions for the model 3.1 [44, 45]:
0 < A(t) < T (t), φ(cid:48)(t) > 0,
A(cid:48)(t) ≤ φ(cid:48)(t), φ(cid:48)(cid:48)(t) ≤ φ(cid:48)(t), T (cid:48)(t) ≤ for all t.
12
(3.2)
(3.3)
If a RRI signal satisfied the model (3.1), we call it rhythmic; otherwise we call it non-
rhythmic. It is possible that a function can be composed of both the rhythmic and non-
rhythmic components.
This seeming complicated model is actually a direct generalization of the Fourier harmonic
model [44, 45]. Close inspecting the waveform of IHR, we can see that the conditions on
φ(cid:48)(cid:48)(t) and A(cid:48)(t) are reasonable.
lation with "small enough" frequency modulation and amplitude modulation. Without
In other words, the rhythmic RRI exhibits an oscil-
adequate knowledge of underlying mechanism, the adaptive harmonic model fulfills the
phenomenological behavior of rhythmicity feature.
3.4 Non-rhythmic to Rhythmic Ratio
The amount of variance inside the RRI signal are known as Heart Rate Variability (HRV).
HRV has been proven to be closely related to the autonomic activity [42]. The most
commonly used tool to quantify HRV is the power spectrum (PS) [42]. As useful as the
PS is, however, in this study the PS cannot capture the momentary dynamics in the RRI
time series. This issue limits the application to human under highly dynamic situation
-- in particular, the "rhythmic-to-non-rhythmic" phenomenon in anesthesia. The main
technical focus of this paper is to resolve this limitation.
In this paper I will quantify the rhythmic-to-non-rhythmic pattern of the RRI series by us-
ing a novel index referred to as Non-rhythmic to Rhythmic Ratio (NRR). NRR is motivated
by the clinical finding that deeper anesthetic levels are associated with the appearance of
more regular oscillatory patterns in IHR. Thus, under deeper anesthesia, IHR oscillates
more regularly. We call this regular oscillatory IHR "rhythmic", which appears as a sharp
peak on the PS. On the other hand, when the subject is under lighter anesthesia, IHR
varies irregularly, and we call this IHR "non-rhythmic". The non-rhythmic IHR exhibits
irregular and random-like behavior that appears as a plateau on the PS. As a result from
this observation, the RRI time series is composed of a non-rhythmic component and a
rhythmic component with various ratios. Based on the above facts, we hypothesize that
the time-varying ratio change of the non-rhythmic and rhythmic components may reflect
the anesthetic effect on the brain.
Chapter 3. Modeling
13
To quantify the ratio of the rhythmic component and non-rhythmic component, we will use
a new signal processing tool to fully capture the momentary RRI time series change. We
will replace the PS with the time-varying power spectrum (tvPS) by applying a recently
developed time-frequency analysis technique [43, 46], which is referred to as multitaper
Synchrosqueezing transform [44, 45, 47]. The tvPS is a non-parametric generalization
of the PS providing instantaneous PS information. We will then be able to trace the
momentary strength of the rhythmic component and non-rhythmic component through
analyzing the tvPS.
By using the tvPS we will further extend the definition of the classical frequency do-
mains HRV parameters. Recall that traditionally, the high frequency (HF) power, the low
frequency (LF) power, and the low frequency to high frequency power ratio (LHR), are
determined from the PS of the RRI time series. Vagal activity mainly contributes to HF
power, meanwhile, sympathetic activity influences both LF power and LHR [42, 48]. By
using the tvPS of RRI, we are able to obtain time-varying HRV parameters, including the
time-varying high frequency power (tvHF), the time-varying low frequency power (tvLF)
and the time-varying low frequency to high frequency power ratio (tvLHR). All these pa-
rameters are suitable to provide a better understanding of the dynamical situations of
anesthesia.
The main index NRR is defined as the ratio between momentary non-rhythmic power to
momentary rhythmic power. In mathematical terms, we compute the rhythmic component
power by identifying the location of peak on tvPS. The non-rhythmic power is then defined
as tvHF subtracted by the momentary rhythmic power. Finally, NRR is defined as
(cid:18) Non-rhythmic power
(cid:19)
Rhythmic power
,
NRR = log10
where the value of NRR is high with non-rhythmic RRI and the value of NRR is low with
rhythmic RRI. As previously discussed, the dynamical behavior of RRI(t) is captured by
tvPS. Please see Figure 3.1
In practical calculation, the non-rhythmic power is calculated as the rhythmic power sub-
tracted from the high frequency power of RRI. The rhythmic component power is calcu-
lated as the sum of the power near the IF of the rhythmic component within the range
of high frequency on the tvPS. We define the IF of the rhythmic component, denoted as
fr(t), as the time varying position of the maximal power within the high frequency range
on the tvPS of RRI(t):
fr(t) := argmax
0.1≤ξ≤1
MRRI,K(t, ξ),
(3.4)
Chapter 3. Modeling
14
Figure 3.1: Fifty-second electrocardiography (ECG) waveforms, time tracings of R-R
interval (RRI), ECG-derived respiration (EDR), and the corresponding spectra in lighter
level of anesthesia (A, B, C) and deeper level of anesthesia (D, E, F). Whether in lighter
(A) or deeper level of anesthesia (D), it is hard to read the IHR information from raw
ECG waveform by naked eyes. Time tracings (B, E) of RRI and EDR and their power
spectra (C, F) show that RRI is more non-rhythmic in lighter level of anesthesia then in
deeper level. The frequency bands around the instantaneous frequency f r and its multiple
frequencies (the unshaded area) constitute the rhythmic component. The rest (shaded
area) constitutes the non-rhythmic component. dB = decibel
Chapter 3. Modeling
15
which is implemented by the curve extraction algorithm considered in [49]. Next, we will
calculate the rhythmic component power, denoted as Pr(t), as the sum of the power inside
the bands around fr(t) on the tvPS of RRI(t). In particular,
(cid:90) nfr(t)+0.01Hz
nfr(t)−0.01Hz
Pr(t) :=
MRRI,K(t, ξ)dξ.
Here the width of the band is chosen to be 0.01Hz in an ad hoc way. We mention that when
the signal is of broadband spectrum, there is no dominant curve that can be visualized,
as shown in Fig.3.1.
In this case, the most dominant curve defined in (3.4) from the
time-frequency representation is viewed as the rhythmic component.
The non-rhythmic component power is defined as the power within the high frequency
subtracted by the rhythmic component power. Lastly, NRR is defined intuitively as the
ratio of the non-rhythmic component power to the rhythmic component power:
(cid:18) tvHF(t) − Pr(t)
(cid:19)
.
Pr(t)
NRR(t) = log10
Through defining power within absolute boundary as Pr(t), NRR can differentiate the level
between "rhythmic" and "non-rhythmic". In this thesis, the variables affixing (t) mean
that they are functions of time. It is important to notice that these sorts of time-varying
parameters describe the momentary behavior of the signal. Since this instantaneous in-
formation cannot be captured by the classical power spectral analysis, we would expect to
gain more insight into the IHR signal by considering our definitions.
3.5 Non-rhythmic Component vs. Stochastic Process
For the purpose of this thesis, the term non-rhythmic is defined as any signal that appears
"irregular", "noisy", or "as random". This poses two interesting questions. First, does
this mean that we should treat all non-rhythmic signals as a stochastic process? Second,
will we be able to model the non-rhythmic signal as an auto-regressive moving-average
(ARMA) model or other stochastic process in order to analyze it using parametric meth-
ods? However, I would like to also remind that we should consider these issues using our
medical knowledge.
One of my hypotheses in the present study is that the non-rhythmic signal reflects highly
structured and complex activity of the human brain. Although it is too early to determine
that the human brain is a "deterministic" or "stochastic" system, when awake, the human
brain can retain memory for longer periods of time to execute complex tasks. The thought
or idea in the brain lasts longer too. Therefore, the brain exhibits non-rhythmic activity
Chapter 3. Modeling
16
when its inner state is longer lasting. However, the above statements would contradict the
stochastic process behavior that presents no memory effect in a Gaussian random process.
That is, the autocorrelation function is a Dirac's delta function:
(cid:88)
l∈Z
Ryy(l) =
y(n)¯y(n − l),
E[RRyy(0)] = σ and E[RRyy(k)] = 0 for k > 0 when y is a Gaussian process.
Although non-rhythmic signals may appear similar to a stochastic process, the human
brain that shows non-rhythmic activities should result in a more structured state. There-
fore, a non-rhythmic signal in its natural state is not completely "stochastic". This is why
I would like to use the term "non-rhythmic".
My second hypothesis is that there are two components, non-rhythmic and rhythmic,
that are contained in the "rhythmic-to-non-rhythmic" phenomenon. It is unknown how
these two components coexist but one simple way to model their co-existence would be
superposition. Regardless if we fully understand how the two components coexist, the fact
that stochastic process is fundamentally a "white-noise-driven" process means it doesn't
match the "rhythmic component plus non-rhythmic component" concept.
The above technical issues regarding parametric modeling for "rhythmic-to-non-rhythmic"
phenomenon needs to be carefully addressed in the future work. To narrow down the
problems I have to solve in my thesis, I decided to choose a non-parametric method as
a starting point. Still, statistical signal processing technique is a viable direction for my
future studies.
3.6 Summary
In summary, the proposed adaptive harmonic model (3.2) describes a fundamental physi-
ologic property: the rhythmic component possesses frequency modulation and amplitude
modulation. However, the law of physiology and chemistry sets a limit to these modula-
tions.
From the previously mentioned theory to the modeling and quantification of the NRR
index, I keep seeking the time-varying ability to capture the momentary dynamics, which
is essential to the dynamic nature of clinical anesthesia. This feature is also fundamentally
distinct to theories of other signal analysis methods. After establishing the theoretic part,
I will focus on the next problem: the time-varying spectrum.
Chapter 4
Time-Frequency Analysis,
Reassignment, and
Synchrosqueezing
4.1 Time-Frequency Analysis
Chapter 3 provided strong evidence that in order to execute the quantification of NRR
index within anesthesia the adaptive harmonic model requires a strong time-varying spec-
trum. A recently developed time-frequency analysis technique is called the multitaper
Synchrosqueezing transform. I believe that this technique is the most useful for my pro-
posed study.
Time-frequency analysis is a type of signal processing tool that generalizes the power
spectrum to provide time-varying ability. This type of analysis generates a series of power
spectra that has the ability to represent instantaneous spectra. Short-time Fourier trans-
form (STFT) and continuous wavelet transform (CWT) are two types of well known tools
for biomedical signal analysis[43, 50 -- 53]. Both of these tools are able to capture the instan-
taneous frequency information of the biological signal. Huang and Chan were the first to
employ time-frequency analysis to analyze the dynamic changes in IHR during anesthesia
[46].
Both STFT and CWT require window functions to "focus" or "localize" the signal. Take
STFT as example:
(cid:90) ∞
−∞
V (h)
X (t, ξ) =
X(s)h(t − s)e−i2πξsds
(4.1)
17
Chapter 4. Synchrosqueezing Transform
18
where h is a window function chosen by the user, for example, the Gaussian function
h(t) = e−t2. Note that X(τ )h(t − τ ) is the piece of signal around t chosen by the window
function. We then define the spectrogram of X(t) [50] as
X (t, ξ) =(cid:12)(cid:12)V (h)
X (t, ξ)(cid:12)(cid:12)2.
G(h)
(4.2)
From equation 4.1, it is apparent that choosing different window functions causes different
results. The influence of the chosen window function in relation to the STFT is no less
than the signal per se. In particular, the window function imposes a resolution limit which
relates to the well-known Heisenberg's uncertainty principal. The larger the time window
h is, the higher the spectrum resolution is, but with low temporal resolution, and vise
versa.
The window function also influences other TF methods, such as CWT, and Smoothed
Pseudo Wigner-Ville Distribution, to name a few. Rather than fiddling with various TF
methods, various window functions, and different sizes, it is highly advised to choose a
technique that alleviates the above issues.
4.2 Reassignment
To cope with the resolution limitation of time-frequency analysis, it is possible to further
use the phase information ϕ(t, ξ), which is discarded in the classical STFT spectrogram
(4.2) and scalogram[50, 51]. Kodera et al. first proposed this technique[54] by calculating
the instantaneous frequency ∂tϕ(t, ξ) and group delay ∂ξϕ(t, ξ). Auger and Flandrin coined
this technique as the reassignment and further generalized it to "Fourier transform based"
time-frequency analysis[50]. The reassignment technique provides a sharper representa-
tion, which improves the readability of the STFT spectrogram. Reassignment is helpful
in visualizing the dynamics of the biological signal that contains multiple nonstationary
components in the TF plane (Fig. 4.1).
The phase information hidden in STFT contains information that explains why time-
frequency domains are spreaded out.[55]. To further eliminate this issue, we can shift the
STFT coefficients accordingly to the reassignment rules determined by the phase informa-
tion. The temporal reassignment rule and frequency reassignment rule are as follows:
Chapter 4. Synchrosqueezing Transform
19
Figure 4.1: A side-by-side comparison between STFT spectrogram (A) and Reassign-
ment spectrogram (B). The tracing of original R-R interval signal recorded in anesthesia
is denoted as red line.
Chapter 4. Synchrosqueezing Transform
20
(cid:40)
(cid:98)tt,ξ := t + (cid:60)
(cid:40)
(cid:98)ξt,ξ := ξ − (cid:61)
(cid:41)
(cid:41)
,
V (T h)
X (t, ξ)
V (h)
X (t, ξ)
V (Dh)
(t, ξ)
X
V (h)
X (t, ξ)
where (cid:60) denotes the real part and (cid:61) denotes the imaginary part and h, T h, and Dh
are the three related window functions. Here T h is the time-ramped window used to
determine the temporal reassignment rule and Dh is the differential window used to de-
termine the frequency reassignment rule. The reassignment is then achieved by applying
the reassignment rules:
RS(h)
X (t, ξ) :=
(cid:90) ∞
(cid:90) ∞
−∞
−∞
X (s, η)δ(s −(cid:98)ts,ξ)δ(η −(cid:98)ξs,ξ)dsdη,
V (h)
where we move the spectrogram coefficient V (h)
location(cid:98)ts,ξ and (cid:98)ξs,ξ according to the reassignment rules.
X (s, η) at time s and frequency η to a new
The reassigned spectrogram is a valuable tool for signals that contain dynamic information
of anesthesia in my current study[43]. However, aside from the visual information repre-
sented in the TF plane, it is also desirable to execute further analysis of the signals, such
as isolating a component in TF plane, or performing reconstruction in time domain. Syn-
chrosqueezing transform is a recently developed reassignment technique that can provide
this possibility.
4.3 Synchrosqueezing Transform
Synchrosqueezing transform is a special kind of reassignment technique[44, 45, 56]. This
technique uses a light-weight mathematical model (4.5). The light-weight mathematical
model helps with further processing of signals, such as isolating a component in the TF
plane or reconstruction in time domain.
For my present study, Synchrosqueezing transform can be used to present non-parametric
generalizations of the power spectrum. The Synchrosqueezing transform is able to provide
the time-varying power spectrum (tvPS) which reveals the instantaneous changes that
happen in IHR. In addition, comparing with traditional STFT and wavelet, it is visually
more informative, more resistant to several kind of noise, and more resistant to the influ-
ence of chosen window function [44, 45, 49]. Mathematical works have supported these
abilities, as well as the ability of inverse transform. Luckily we do not need to impose too
Chapter 4. Synchrosqueezing Transform
21
Figure 4.2: A side-by-side comparison between STFT spectrogram (A) and Reassign-
ment spectrogram (B). The tracing of original R-R interval signal recorded in anesthesia
is superimposed as red line in panel B
much effort on choosing parameters as using the Synchrosqueezing transform technique
can easily capture the oscillatory activity.
The detailed information on the algorithm leading to tvPS is as follows. It has been known
that the phase information hidden in STFT contains more information that explains why
the STFT spectrogram is spreaded out. A solution that could help remedy this issue is by
shifting the STFT coefficients according to reassignment rules determined by the phase
information. The frequency reassignment rule is as follows:
Chapter 4. Synchrosqueezing Transform
Ω(t, ξ) :=
−i∂tV (h)
2πV (h)
X (t, ξ)
X (t, ξ)
22
(4.3)
where ∂t means the partial derivative with related to t. The Synchrosqueezing transform
is then achieved by applying the reassignment rule:
(cid:90) ∞
(cid:90) ∞
X (t, η)δ(ξ − Ω(t, η))dη,
V (h)
S(h)
X (t, ξ) :=
−∞
−∞
where for each fixed time t, we move the STFT coefficient V (h)
new location according to the reassignment rules Ω(t, ξ).
X (t, η) at frequency η to a
One way that the Synchrosqueezing transform is unique to the reassignment technique
is that the Synchrosqueezing transform has the ability to reconstruct back to the time
domain. Suppose the signal X(t) analyzed by Synchrosqueezing transform is:
where T (t) > 0, A(t) > 0 and φ(cid:48)(t) > 0.
X(t) = T (t) + A(t) cos(2πφ(t)),
(4.4)
In other words, X(t) comprises of an oscillatory component and a trend component.
In addition, the amplitude modulation and frequency modulation of the oscillatory com-
ponent is under the condition as follows:
0 < A(t) < T (t), φ(cid:48)(t) > 0,
A(cid:48)(t) ≤ φ(cid:48)(t), φ(cid:48)(cid:48)(t) ≤ φ(cid:48)(t), T (t)(cid:48) ≤ for all t.
(4.5)
(4.6)
Relevant results of the Synchrosqueezing transform can be found in literatures[49, 57 --
60]. A recent review article discussed in depth on the relation of the TF analysis, the
reassignment and Synchrosqueezing technique[61].
4.4 Multitaper Estimation
From a non-parametric spectral estimation perspective, Thompson proposed a powerful
multi-window technique that improves spectral estimation of the random process[62, 63].
Thompson uses the discrete prolate Spheroidal sequence to define a series of window func-
tion, which are orthogonal to each other, that have the same length in time domain.
The average of the spectra from multiple windows reduces the spectrum variance but
also prevents the bias of spectral estimation induced by varying window size. One strong
Chapter 4. Synchrosqueezing Transform
23
advantage of Thompson's spectrum is that we are able to estimate the spectrum of a
stationary random process without resort to parametric estimations.
The principle of Thompson's technique has been extended to the field of nonstationary
signal. Multitaper in time-frequency analysis is designed to be used with Hermite functions
as the sequence of multiple windows that are orthogonal to each other in time-frequency
plane. The final spectrogram estimation can then be obtained from the average of multiple
realizations. Taking the STFT spectrogram as an example, the multitaper spectrogram
can be defined as follows:
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
K
K(cid:88)
k=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2
V (hk)
X (t, ξ)
.
QX,K(t, ξ) :=
The multitaper spectrogram and multitaper scalogram were shown to make success for
nonstationary stochastic signal[64]. To achieve both high resolution for deterministic non-
stationary signal and low variance for stochastic nonstationary signal, Xiao and Flandrin
combined the reassignment technique and multitaper technique[47].
From the background information mentioned in Chapter 2 and the model in Chapter 3, it
can be seen that the IHR under anesthesia should contains both non-stationary component
and stochastic component. We have reported the advantage of using multitaper time-
frequency reassignment in extracting these dynamic features[43].
4.5 Multitaper Synchrosqueezing Spectrogram
Synchrosqueezing transform responds well to signals from different types of stochastic
process. One dilemma we face is understanding the true mechanism in biological signal.
Since the quantification of the non-rhythmic component is an important part of our current
model(3.4), we have decided to apply the multitaper method to further optimize the
performance of the Synchrosqueezing transform X(t). We have chosen the multitaper
Synchrosqueezing spectrogram to achieve the tvPS as follows:
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
K
K(cid:88)
k=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2
MX,K(t, ξ) :=
S(hk)
X (t, ξ)
,
where K > 0 is the number of windows chosen. We choose the the k-th Hermite function
as hk. For my current study, I will use K = 10 and the first ten Hermite window functions
will be hk, k = 1, . . . , 10 when deriving tvPS.
Chapter 4. Synchrosqueezing Transform
24
Figure 4.3: The time-varying power spectrum of instantaneous heart rate (IHR) as data
used in figure 4.2, realized by multitaper Synchrosqueezing spectrogram, showing rich
dynamic features of human body in anesthesia. As shown in figure 4.2, the readability is
better than STFT spectrogram, and also better than the reassigned spectrogram.
The advantage of using a multitaper Synchrosqueezing spectrogram is summarized as
follows:
1. Robust to several different kinds of noise, which might be slightly nonstationary.
2. Visually informative. When the signal is rhythmic, that is, satisfying the model (3.1),
a dominant curve following the IF can be seen on the multitaper Synchrosqueezing
spectrogram; otherwise the multitaper Synchrosqueezing spectrogram is blurred.
3. Adaptive to the data so that its dependence on the chosen window function is weak.
We compared various methods of TF analysis in Table4.1 and in simulated data analy-
sis(Fig.4.4).
4.6 Adaptive Filter and Parametric Estimation
Another kind of techniques based on parametric modeling for time series data analysis
is the adaptive filter, including the famous Kalman filter, and the parametric spectral
estimation[65]. The adaptive filter is designed to use a cost function to represent the
weighted mean square error. The goal is finding a set of parameters ai to minimize the
cost function.
ai = argmin E{e2[n]}
Chapter 4. Synchrosqueezing Transform
25
Figure 4.4: An artificial data, as a summation of two time-varying components and
a Gaussian white noise, and its TF analysis realized by STFT spectrogram, Syn-
chrosqueezed spectrogram and the multitaper Synchrosqueezed spectrogram. From 125 s
to 250 s, it is apparent that Synchrosqueezing technique improves the readability of the
two components, and multitaper technique refines further.
If we have adequate knowledge about the signal, we can obtain the optimal point on the
performance surface of the cost function by using partial differential.
E{e2[n]} = 0
∂
∂ai
It is apparent that TF reassignment is fundamentally different from adaptive filter. TF
reassignment is a non-parametric deterministic signal processing technique, whereas the
adaptive filter is a parametric technique for stochastic signal processing. Regardless of the
difference, both techniques use the same mathematical tools to achieve their goals.
Kalman filter in essence is a real-time prediction-correction algorithm that provides para-
metric estimation in the sense of weighted least-square fitting[62, 65]. As the physiologic
mechanism underlying rhythmic-to-non-rhythmic phenomenon is not adequately under-
stood, a technique which requires both elaborate model construction and careful parame-
ter selection is not my first-line choice for my research problem. Besides, From Chapter 3,
it shows that the NRR index has a high relation to oscillation. Therefore, for my study,
Chapter 4. Synchrosqueezing Transform
26
Table 4.1: Summary of abilities of TF methods. Details can be found in references cited
in this chapter. X: unable; V: feasible; ?: questionable
Reassignment SST STFT CWT
Instantaneous frequency estimation good
good fair
fair
Visual readability
good
good fair
fair
Component reconstruction
Robust to noise
X
?
V
V
V
good fair
fair
the output of the Kalman filter as a matrix of state variables may not be as intuitive as
the spectrum technique for data exploration and interpretation at this stage.
Parametric spectral estimation has a complete theory of modeling based on stochastic
process[62]. One critical issue of the parametric estimation is that the mechanism of the
signal we are facing in the present study is unknown, and is constantly changing. The para-
metric spectral estimation has a lower tolerance to incorrect modeling. In other words, it is
crucial to ensure that the correct model are constructed to generate reasonable parametric
spectral estimation. In addition, parametric spectral estimation does not behave well when
the signal is a superposition of multiple components. The rigidity of model construction
contrasts well with the flexibility of Synchrosqueezing: flexible in window selection and
adaptive to signal. Certainly, it is possible that the technique realizing time-varying power
spectrum in the future may belong to the family of parametric spectrum.
4.7 Common Condition in Physiology and Mathematics
The most unique characteristic of the Synchrosqueezing transform that is relevant to this
project is the mathematical condition 3.2 which shows that the frequecy modulation and
amplitude modulation are limited. The condition used on the Synchrosqueezing transform
is exactly the same condition as the adaptive harmonic model in order to describe this
rhythmic physiological behavior. Therefore, the Synchrosqueezing transform matches per-
fectly with the adaptive harmonic model for us to quantify the "rhythmic-to-non-rhythmic"
phenomenon that was previously discussed.
Chapter 5
Performance Evaluation using
Clinical Data
5.1 Clinical Database
In my hopes to investigate the clinical value of NRR index, I have collected and analyzed
a clinical database. The prospective observational study spanned the whole period of
typical surgical anesthesia. ECG, BIS index, and the end-tidal concentration of sevoflurane
were all recorded continuously. Anesthetic events, including loss of consciousness (LOC),
skin incision, first reaction of motor movement during emergence period, and return of
consciousness (ROC), were registered with precise time stamps.
After receiving approval from the Institutional Review Board (Taipei Veterans General
Hospital, Taipei, Taiwan), we enrolled 31 female patients. All patients were American
Society of Anesthesiologists (ASA) physical status of I or II and were scheduled for la-
paroscopic gynecological surgery undergoing general anesthesia. We obtained informed
consent in paper from each of our patients. All ASA I or II status patients are generally
under healthy conditions. It is important to investigate the changes of the NRR index
under the normal physiological conditions. We are aware that some diseases can affect
HRV (and IHR) untowardly. Thus, patients that are excluded were those younger than 20
years old or older than 55 years old, have a body mass index below 19 or higher than 30
and have either cardiovascular disease, arrhythmia, intake of medication that affects the
neurological and cardiovascular system, potential cancer, diabetes mellitus, or anticipated
difficult airway. All these surgeries were performed between the hours of 8 a.m. and 7
p.m.
27
Chapter 5. Performance Evaluation
28
5.2 Anesthesia Protocol
The standard anesthetic monitoring (HP agilent patient monitoring system) was applied to
all patients. This included ECG, pulse oximeter (SpO2) and non-invasive blood pressure. A
BIS electrode (BIS-XP sensor) and the ECG recorder (MyECG E3-80; Micro-Star Int'l Co.,
New Taipei City, Taiwan) were applied to the patients prior to anesthetic induction. After
the initial step of cleaning the skin with rubbing alcohol, the BIS sensor is applied to the
right side of the forehead. As a standard practice, all anesthetic managements, including
the administration of medications, and airway management were under the discretion of
the anesthesiologist. After intravenous infusion of lactated Ringer's solution begins, we
started anesthetic induction with fentanyl 2.5 -- 3 µg/kg. Next, right after preoxygenation,
hypnosis was induced with propofol 2 -- 2.5 mg/kg. LOC was then assessed by no response
to verbal command. Cisatracurium 0.15mg/kg was used to facilitate tracheal intubation.
Subsequently, mechanical ventilation was started from volume control mode with oxygen-
gas-sevoflurane mixture as low flow anesthesia. The respiratory rate and tidal volume of
the ventilator was adjusted to maintain an end-tidal carbon oxide (ETCO2) that ranges
between 35 and 40 mmHg, and to keep the peak airway pressure lower than 25 mm Hg.
The laparoscopic skin incision was made by a surgical blade. The adequacy of anesthetic
depth (during the whole period of surgery) was determined by the BIS index (< 60) and
based on the anesthesiologist's judgment. If the anesthesiologist determined an inadequate
level of analgesia, a bolus dose of fentanyl would be given to the patient. Toward the end
of the surgery, the peritoneal gas was deflated and the patient's body position recovered
from Trendelenburg's position to level-supine position. As the wound closure began, the
anesthetic gas was on a consistent decrease. Muscle relaxation was reversed with a com-
bination of neostigmine 0.05 mg/kg and glycopyrrolate 0.01 mg/kg during controlled ven-
tilation. Once the patient regained spontaneous breathing the controlled ventilation was
halted. Patients that exhibited inadequate spontaneous breathing (ETCO2 >50 mmHg
or SpO2 <95%) were assisted with manual positive ventilation, otherwise the patients
continued to breath spontaneously until regaining consciousness without the interference
of positive-pressure ventilation.ECG and BIS are the main electrophysiological signals we
analyze for this project. When we observed the emergence period, physical contact to
the patient was avoided and minimized if necessary to prevent the interference on the
sensors of ECG and BIS monitor. Any unnecessary interference to patient's respiration
from the manual (bag) ventilation was avoided also. When the patient showed adequate
consistency in their spontaneous breathing, we removed the endotracheal tube. If in any
case there was inadequate ventilation or upper airway obstruction, corrective actions were
taken that include either mask ventilation or nasopharyngeal airway insertion. During the
emergence period, I carefully assessed first reaction as any first visible motor reaction such
Chapter 5. Performance Evaluation
29
as movement of the arms or legs, coughing, or grimace. ROC was assessed through the
patient's ability to open their eyes and follow simple procedural commands.
5.3 Data Acquisition
The BIS index was continuously recorded from an Aspect A-2000 BIS monitor (version
XP, Host Rev:3.21, smoothing window: 15 seconds; Aspect Medical Systems, Nattick,
CA, USA) connected to a laptop computer (Asus Corp., Taipei, Taiwan). During surgery,
corrective measures were taken to improve signal quality of the BIS sensor when the signal
quality index was lower than 50; otherwise the BIS data during this period was discarded.
The raw limb lead ECG sampled at 1000Hz and 12-bit resolution, was recorded for off-line
analysis. The clocks on the laptop and ECG recorder were synchronized with a time accu-
racy of ±1 second. Time stamps of BIS record were provided by the laptop. The inhaled
and end-tidal concentrations of anesthetic gas, which was sampled from the connection
piece close to the endo-tracheal tube and detected by the gas analyzer on a Datex-Ohmeda
S/5 anesthesia machine (GE Health Care, Helsinki, Finland), were also recorded on the
laptop. All data was recorded without any interruption until the patient regained ROC. All
data collected for analysis of sevoflurane concentration correlation and anesthetic events,
including skin incision, first reaction and ROC, are under single ventilation mode. There
were no transitions from the mechanical ventilation to the spontaneous ventilation or vice
versa in the data. However, during LOC, the transition from the spontaneous ventilation
to mechanical ventilation is possible.
The offline ECG signal was analyzed in several steps. The R peak detection was auto-
matically determined by taking lead I, II and III ECG signals into account to improve
accuracy. The ectopic beats were removed and interpolated from the data to eliminate
incorrect R peaks and ectopic beats through visual verification. When electrocauteriza-
tion is severely interfered by the ECG, the data segment were also discarded. Whenever
electro-cauterization severely interfered the ECG, this data segment was discarded. IHR
was then derived from the R peaks through the cubic spline interpolation. RRI were
resampled to be equally spaced at 4 Hz from IHR for subsequent analyses (Fig.2.1).
In order to correct the hysteresis between the end-tidal sevoflurane concentration and the
anesthetic effect on the brain, the estimated effect-site sevoflurane concentration (Ceff)
was derived from the end-tidal sevoflurane concentration (Cet) by the following first order
differential equation. This equation is based on the pharmacokinetic-pharmacodynamic
modeling[66]:
Chapter 5. Performance Evaluation
30
dCeff
dt
= Ke0(Cet − Ceff)
where the constant Ke0 was assumed to be constant for all patients and defined as 0.20
/min according to previous studies [8].
5.4 Statistical Analysis
We treated BIS and ECG-derived continuous indices, including NRR, tvHF, tvLF, tvLHR
and HR, as anesthetic depth indices. The performances of the indices were evaluated in two
ways: the ability for these indices to predict anesthetic events and their correlations with
effect-site sevoflurane concentration during emergence period. Because the concentration
of anesthetic gas consistently decreased and the influence from surgery was relative minor,
we considered the emergence period for correlation analysis. A p value less than 0.05 would
be considered as statistically significant. Multiple significant tests were calculated by using
Bonferroni correction. Statistical results are expressed as mean (standard deviation [SD]).
Prediction probability (PK) analysis is a versatile statistical method for measuring the
performance of an anesthetic depth index [67]. A value of one means that the observation
is always correctly predicted, a value of 0.5 indicates that the observation is predicted at
a 50/50 chance. The ability to predict anesthetic events can be investigated by using the
serial prediction probability analysis[67]. The correlation with sevoflurane concentration
employed PK analysis and Spearman rank correlation. Estimation of PK and its standard
error was obtained through the Jackknife method. Prior to using the null hypothesis, any
PK value less than 0.5 would be changed into one minus the PK value[67].
5.5 Serial Prediction Probability Analysis in Predicting Anes-
thetic Events
To evaluate the performance of the anesthetic indices in predicting the anesthetic events,
serial PK analysis was designed as follows. The following timestamps are used as the
baseline: one minute before LOC, five seconds before skin incision, three minutes before
the first reaction and three minutes before ROC. The serial PK analysis is performed
index on baseline timestamp versus indices on
as successive PK analyzes of data pairs:
subsequent successive timestamps spaced by five seconds. As a result, the out of the serial
PK analysis is a sequence of time-varying PK values which reveals the temporal relation
between anesthetic events and the indices. By plotting the serial PK value and its standard
error bar, it becomes possible for us to do hypothesis test simply with the naked eye. If
Chapter 5. Performance Evaluation
31
there is significant difference (p < 0.05), it can be established if 1.5 times of their standard
error bars do not overlap [67, 68].
5.6 Correlations with Sevoflurane Concentration
Correlation between the above indices and sevoflurane concentration was investigated in
the emergence period. We chose the time interval where the sevoflurane concentration
is monotonically and continuously decreased. The period of spontaneous breathing is
defined from the start of adequate spontaneous breath to ROC. We employed the PK
analysis and Spearman rank correlation to analyze performance of indices sampled every
4 seconds. The results were tabulated as weighted averages according to each patient's
data length. The bootstrap method was used to calculate the 95% confidence interval of
Spearman correlation based on 10,000 samplings. A value of Spearman correlation closer
to 1 is better since we expect that the indices will increase to correlate with the decrease
of sevoflurane concentration.
5.7 Algorithm of Serial Prediction Probability (sPK) Anal-
ysis
In this section we detail a variation of the Prediction probability (PK) analysis [67], referred
to as Serial Prediction Probability (sPK) analysis. sPK is employed in this study to evaluate
the prediction performance of the BIS and ECG-derived continuous indices, including
NRR(t), tvHF(t), tvLF(t), tvLHR(t) and HR.
We consider K different measurement times, which are the times when we record the
anesthetic indicators. Then we determine two special timestamps, one is referred to as
the event time, indicating the happening of the event, for example, the first reaction, and
one is referred to as the base time, which is T > 0 seconds before the event time. Here T
depends on the event we are studying. Collect N subjects and record the measurement
times and the event time of each individual. Note that the measurement times and the
event time and are different among individuals, and we denote the k-th measurement time
of the n-th subject as tk,n, k = 1, . . . , K, and the base time as t0,n. Denoted as xk
n the
anesthetic indicator (e.g. BIS) of the n-th subject recorded at time tk,n, and denoted as
yk
n the associated outcome (e.g. first reaction) at time tk,n.
To evaluate the prediction power of an anesthetic indicator at the k-th measurement time,
we take the following two datasets. At the measurement time tk,n, we record the sam-
ple anesthetic indicators {xk
N}, where each
N} and the sample outcomes {yk
1, . . . , xk
1 , . . . , yk
Chapter 5. Performance Evaluation
32
n, yk
pair (xk
n) is recorded from the n-th subject. Denote Zk := {(xk
N )}
as the measurement data at the k-th measurement time. Moreover, at the base time
t0,n, we record the sample anesthetic indicators {x0
N} and the sample outcomes
{y0
1), . . . , (x0
as the data at the base time.
N}, where each pair is recorded at the base time. Denote Z0 := {(x0
1, yk
1 ), . . . , (xk
1, . . . , x0
1, . . . , y0
N , yk
1, y0
N )}
N , y0
For a given individual, there are five possible relationships between pairs of anesthetic
indicator and outcome at the k-th measurement time and those at the base time:
1. (xk
n, yk
n) and (x0
m, y0
m) are concordant if they are rank ordered in the same direction;
n) and (x0
m, y0
m) are disconcordant if they are rank ordered in the reverse di-
2. (xk
n, yk
rection.
3. (xk
n, yk
n) and (x0
m, y0
m) tie in x if xk
n = x0
m while yk
4. (xk
n, yk
n) and (x0
m, y0
m) tie in y if yk
n = y0
m while xk
n (cid:54)= y0
m;
n (cid:54)= x0
m;
5. (xk
n, yk
n) and (x0
m, y0
m) tie in both x and y if xk
n = x0
m and yk
n = y0
m.
Usually the anesthetic indicator is of finer scale and the outcome is of coarser scale. Thus,
unlike the usual notion of correlation, in the PK analysis the tie in x is undesirable but
we tolerate the tie in y and the tie in both x and y. Based on this notion, we define
the serial PK index in the following way. Denote Pc(k), Pd(k) and Ptx(k) the respective
probabilities that two pairs of (x, y) independently drawn from Zk and Z0 with replacement
are concordant, disconcordant and tie in x at the k-th measurement time. The serial PK
index at the k-th measurement time is denoted as
PK(k) =
Pc(k) + 1
2 Ptx(k)
Pc(k) + Pd(k) + Ptx(k)
.
The PK(k) value is interpreted in the same way as that in the PK analysis. A value of
one means that the indicator always correctly predicts the observed depth of anesthesia,
a value of 0.5 means that the indicator predicts no better than 50/50 chance, and a PK(k)
value less than 0.5 means that the indicator predicts inversely. To be more precise, a zero
value also means a perfect prediction but in a reversed direction.
The estimation of the serial PK and its standard error was obtained by the Jackknife
method. Before the null hypothesis, the serial PK value less than 0.5 was converted into
one minus the serial PK value.
Chapter 6
Results from Clinical Database
6.1 Visual Information in TF Plane
Abundant information can be visually read from the tvPS of IHR. Various features were
determined to reflect the physiologic dynamics of human body responding to various anes-
thetic events (Fig. 6.1).
It is demonstrated (Fig. 6.2,6.3) that the tvPS is more concentrated on deeper anesthetic
levels and more scattered on lighter anesthetic levels. This then created the hypothesis
on trying to understand the "rhythmic-to-non-rhythmic" phenomenon. This phenomenon
is consistently seen in most study cases. During spontaneous breathing (Fig.6.3), the
dominant curves can be seen in the first portion (usually on the left hand side) of the tvPS
graph. Meanwhile, less dominant curves can be identified in the second portion (usually
the right hand side). The dominant curve represents the typical rhythmicity feature.
It is also important to observe that the tvPS becomes more non-rhythmic prior to the
appearance of first reaction. (Fig. 6.3)
Painful surgical stimulations during anesthesia are referred to as noxious stimulation be-
cause an unconscious patient under anesthesia cannot report pain sensation. Two of the
most intense noxious stimulationthat occur in anesthesia and surgery are endotracheal in-
tubation and skin incision. Clinical anesthesiologist uses the increase of heart rate and
blood pressure as indicators for the level of noxious stimulation. In my study, we have
found the visual information of tvPS that reveals the influence of noxious information
where the low frequency (LF) power increases immediately after noxious stimulation, and
wanes soon. The "LF surge phenomenon" appeared in all study cases. Although the main
goal of my study is on investigating NRR index, it can be seen that the potential value of
LF power as an "noxious index" deserves to be investigated too.
33
Chapter 6. Results
34
Figure 6.1: The time-varying spectrum of R-R peak interval (RRI) recorded during
anesthesia showing rich dynamic features of human body in anesthesia. The RRI tracing
is superimposed as red line
These visual findings are quantified by NRR and HRV indices, such as tvHF, tvLF and
tvLHR for subsequent analyses.
6.2 Multiple Component Phenomenon
Occasionally, multiple independent oscillatory components hidden in IHR can be seen
from the tvPS. To further expand on this statement, the term independent is defined
as the frequency of one component of which the modulation is unrelated to other com-
ponents. Therefore, this is definitely not the harmonics (Fig.6.8, 6.9, 6.10, 6.11). The
"multiple component phenomenon" are an interesting finding, which demonstrates the
dynamic physiology of human body undergoing anesthesia.
Apparently, it is nearly impossible to identify the appearance of multiple component from
raw IHR waveform using naked eye. The tvPS is good at presenting "multiple compo-
nent phenomenon", but it is a difficult task for classical power spectrum. The "multiple
component phenomenon" appeared in one thirds of the study cases. The moment that it
appeared and disappeared seems random and unpredictable. Even so, this phenomenon
Chapter 6. Results
35
Figure 6.2: A representative data from a patient under anesthesia and controlled venti-
lation. The 1200 seconds R-R interval (RRI) is superimposed on its time-varying power
spectra (A) as a blue solid line. The corresponding electrocardiography derived respira-
tion (EDR) is superimposed on its time-varying power spectra (B) as a green line. Panel
C shows simultaneous recorded Bispectral index (BIS), effect-site sevoflurane concentra-
tion (Sevo) and non-rhythmic to rhythmic ratio (NRR). Arrow 1:
inadequate level of
anesthesia was noted and the sevoflurane concentration was raised. Sevoflurane concen-
tration reached top level after 300-second. Arrow 2: both BIS index and NRR reached
minimum. There is a dominant straight line around 0.167 Hz in the panel A, whose
fundamental frequency is the same as the rate of the ventilator, representing the en-
trainment of RSA by the controlled ventilation. The time-varying power spectra of RRI
showed rhythmic-to-non-rhythmic transition, corresponding to varying anesthetic depth,
whereas the time-varying power spectra of EDR was almost unchanged under controlled
ventilation.
Chapter 6. Results
36
Figure 6.3: A representative data obtained from the same patient under spontaneous
breathing in emergence period. The 1200-second R-R interval (RRI) is superimposed on
its time-varying power spectra (A) as a green solid line. The corresponding electrocar-
diography derived respiration (EDR) is superimposed on its time-varying power spectra
(B) as a blue solid line. Panel C shows simultaneous recorded Bispectral index (BIS),
effect-site sevoflurane concentration (Sevo) and non-rhythmic to rhythmic ratio (NRR).
Arrow 1: endotracheal suction and extubation, Arrow 2: first reaction, Arrow 3: regain
of consciousness. Under spontaneous breath, RRI and EDR are similar in high frequency
region of spectrograms. The rhythmic component and its harmonics are time-varying in
both time-varying power spectra of RRI and EDR. With the overall decrease of sevoflu-
rane, both RRI and EDR exhibited a rhythmic to non-rhythmic transition. The RRI and
EDR became non-rhythmic before the first reaction.(From ref.[69])
Chapter 6. Results
37
Figure 6.4: A representative data from a patient under anesthesia receiving skin incision
for her surgery. The 750-second R-R interval (RRI) is superimposed on its time-varying
power spectra (A) as a green solid line. The corresponding electrocardiography derived
respiration (EDR) is superimposed on its time-varying power spectra (B) as a blue solid
line. Panel C shows simultaneous recorded Bispectral index (BIS), effect-site sevoflurane
concentration (Sevo) and non-rhythmic to rhythmic ratio (NRR).
is related to specific anesthetic events like only appearing when the patient is undergoing
anesthesia. Although mechanism underlying this phenomenon is beyond the goal of my
present study, it still presents an interesting research topic for the future.
6.3 Ability to Predict Anesthetic Events
In the serial PK analysis for the first reaction (Fig. 6.14), the NRR is ahead of BIS
(p < 0.05 for 20 seconds). The NRR was able to predict the first reaction 30 seconds
Chapter 6. Results
38
Figure 6.5: tvPS of IHR during emergence period from the last three consecutive cases
(no.1) showing rhythmic-to-non-rhythmic phenomenon. BIS index (golden yellow tracing)
and IHR (blue tracing) are superimposed.
Figure 6.6: tvPS of IHR during emergence period from the last three consecutive cases
(no.2) showing rhythmic-to-non-rhythmic phenomenon. BIS index (golden yellow tracing)
and IHR (blue tracing) are superimposed.
Chapter 6. Results
39
Figure 6.7: tvPS of IHR during emergence period from the last three consecutive cases
(no.3) showing rhythmic-to-non-rhythmic phenomenon. BIS index (golden yellow tracing)
and IHR (blue tracing) are superimposed.
Figure 6.8: The time-varying power spectrum of instantaneous heart rate (IHR)
recorded during anesthesia showing three independent oscillatory components, revealing
rich dynamic features of human body in anesthesia.
Chapter 6. Results
40
Figure 6.9: The IHR (blue tracing) was recorded during spontaneous breathing. The
second component appeared shortly (200 -- 300 s).
Figure 6.10: The IHR during spontaneous breathing showed "rhythmic-to-non-
rhythmic" changes with BIS index (golden yellow tracing), accompanied with multiple
components during "rhythmic" period.
Chapter 6. Results
41
Figure 6.11: Another IHR data (blue tracing) during spontaneous breathing showed
"rhythmic-to-non-rhythmic" changes with BIS index (golden yellow tracing), accompanied
with multiple components during "rhythmic" period.
in advance (PK > 0.90). At the instance of first reaction (0 seconds), both the NRR
and BIS (both PK maxima > 0.95) were significantly better than other parameters. The
time-varying HRV indices and HR (PK < 0.83) provides significantly worse results than
the NRR.
In the serial PK analysis for skin incision (Fig. 6.4), tvLF reaches maximum (PK > 0.95)
roughly at 30 seconds after skin incision. It reflects skin incision best, followed by tvLHR
(PK < 0.85). The tvLF is significantly better than BIS (PK < 0.55) and NRR (PK < 0.65).
In the serial PK analysis for LOC (Fig. 6.12), BIS reflects perfectly (PK = 1) 50 seconds
after LOC. LOC is also associated with a decrease in tvLF, tvHF and HR. Nowever, NRR
does not reflects LOC well. In the serial PK analysis for ROC (Fig. 6.15), BIS is the best
index, and surpasses NRR, HR and the time-varying HRV indices significantly.
Due to the inadequate data quality, 1 patient has been eliminated from our data in regards
to the serial Pk analysis of the initial reaction. For our study, when we analyzed the serial
Pk of skin incision, LOC and ROC, we also eliminated 3, 5, and 4 patients respectively.
6.4 Correlation with Sevoflurane Concentration
The correlation with the estimated effect-site sevoflurane concentration is evaluated for
the spontaneous breathing (SB) period (table 6.2). The data number of SB is 5810 (23240
s) and the average duration for each patient is 750 ± 322 s. The overall indices ranking of
Chapter 6. Results
42
Figure 6.12: Tracings and their faded error bars about the prediction probability (PK)
values of the parameters and their standard errors in serial PK analysis for loss of con-
sciousness (LOC). The baseline is one minute before LOC. NRR = non-rhythmic to
rhythmic ratio; HF = high frequency power; LF = low frequency power; LHR = LF to
HF ratio; BIS = Bispectral Index; HR = heart rate. (From ref.[69])
the PK analysis and Speareman rank correlation (R) are consistent. While BIS correlates
with effect-site sevoflurane concentration best (SB: PK = 0.839, R = −0.836, p < 0.0001),
the NRR index is second (PK = 0.736, R = −0.619, p < 0.0001). The tvLF is the best
HRV indices ( PK = 0.669, r = −0.416, p < 0.0001). Compared to tvLF, the NRR index
is significantly better ( p < 0.01). BIS surpasses NRR significantly in SB (p < 0.001). The
NRR index is significantly better than HR (p < 0.0001).
6.5 Quantitative Results of Noxious Stimulation
When observing the serial PK analysis for skin incision (Fig. 6.4), tvLF reaches the
maximum amount (PK > 0.95) 30 seconds after skin incision. This shows that the tvLF is
the best index of the skin incision, followed by the heart rate (PK < 0.85). Furthermore,
tvLF is significantly better than BIS (PK < 0.55) and the NRR index (PK < 0.65).
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 -50 -25 0 25 50 75 100 PK time (s) Serial PK for LOC LHR NRR LF HR HF BIS Chapter 6. Results
43
Figure 6.13: Tracings and their faded error bars about the prediction probability (PK)
values of the parameters and their standard errors in serial PK analysis for skin incision.
The baseline is 5 seconds before skin incision. NRR = non-rhythmic to rhythmic ratio;
HF = high frequency power; LF = low frequency power; LHR = LF to HF ratio; BIS =
Bispectral Index; HR = heart rate. *P < 0.05, LF versus BIS and LF versus NRR. (From
ref.[69])
In the serial PK analysis for endotracheal intubation (Fig. 6.16), I compared different
indices calculated from multitaper reassignment spectrogram (HF, LF, LHR) with the in-
dices calculated from STFT spectrogram (HFSTFT , LFSTFT , LHRSTFT ). This analysis
signifies that the LF power and LHR provide the best results for the noxious stimulation
of endotracheal intubation. Meanwhile, the heart rate is represented the second best and
the BIS index is represented the worst. The indices that are calculated from multita-
per reassignment spectrogram performs better than the corresponding indicators that are
calculated from the STFT spectrogram.
6.6 Summary
Several features on TF plane can be easily seen conveniently via the tvPS, including
the "rhythmic-to-non-rhythmic phenomenon", "multiple component phenomenon", and
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 30 60 90 PK Time (s) Serial PK for Incision LF LHR HF HR NRR BIS * * * * * * * * Chapter 6. Results
44
Figure 6.14: Tracings and their faded error bars about the prediction probability (PK)
values of the parameters and their standard errors in serial PK analysis for first motor
movement reaction. The baseline is 3 minutes before first motor movement reaction.
NRR = non-rhythmic to rhythmic ratio; HF = high frequency power; LF = low frequency
power; LHR = LF to HF ratio; BIS = Bispectral Index; HR = heart rate. * P < 0.05,
NRR versus BIS. (From ref.[69])
"LF surge phenomenon". The "rhythmic-to-non-rhythmic phenomenon" and "LF surge
phenomenon" are apparent in the majority of my study cases. This universal consistency
foreshadows the subsequent positive statistical results.
The quantitative analysis proves that the NRR index indicates the first motor movement
reaction better than other clinical indices[69]. BIS index is the best indiator for LOC
and ROC. NRR index correlates with the concentration of sevoflurane. Furthermore, LF
power indicates noxious stimulation best, and better than the heart rate. It is also shown
the statistical benefit of the multitaper reassignment spectrogram over the classical STFT
spectrogram.
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 -180 -150 -120 -90 -60 -30 0 30 60 90 120 150 PK time (s) Serial PK for First Reaction NRR BIS HR LHR LF HF * * * * * * Chapter 6. Results
45
Figure 6.15: Tracings and their faded error bars about the prediction probability (PK)
values of the parameters and their standard errors in serial PK analysis for return of
consciousness (ROC). The baseline is 3 minutes before ROC. NRR = non-rhythmic to
rhythmic ratio; HF = high frequency power; LF = low frequency power; LHR = LF to
HF ratio; BIS = Bispectral Index; HR = heart rate. * P < 0.05, BIS versus NRR. (From
ref.[69])
0.3 0.4 0.5 0.6 0.7 0.8 0.9 -180 -150 -120 -90 -60 -30 0 PK Time (s) Serial PK for ROC BIS HR NRR LF LHR HF * * * * * * * Chapter 6. Results
46
Figure 6.16: Tracings and their faded error bars about the prediction probability (PK)
values of the parameters and their standard errors in serial PK analysis for endotra-
cheal intubation. The baseline is 5 seconds before endotracheal intubation. HF = high
frequency power calculated by multitaper reassigned spectrogram; LF = low frequency
power calculated by multitaper reassigned spectrogram; LHR = LF to HF ratio calcu-
lated by multitaper reassigned spectrogram; HF stft = high frequency power calculated
by STFT spectrogram; LF stft = low frequency power calculated by STFT spectrogram;
BIS = Bispectral Index; HR = heart rate.
0.4 0.5 0.6 0.7 0.8 0.9 0 25 50 75 100 PK Time (s) Serial PK for Intubation HF HF_stft LF LF_stft LHR LHR_stft BIS HR Chapter 6. Results
47
Table 6.1: Representative values of the Indices during important anesthetic events.
Values are presented as median values (lower and upper quartiles). HF, LF, and LHR are
expressed in common logarithm (10 base) of millisecond square. Heart rate is expressed
in beat per minute. BIS = Bispectral Index; NRR = non-rhythmic to rhythmic ratio; HF
= high frequency power; LF = low frequency power; LHR = LF to HF ratio; LOC = loss
of consciousness; ROC = return of consciousness; T0= 60 s before LOC; T1= 60 s after
LOC; T2= 5 s before incision; T3= 30s after incision; T4= 180 s before first reaction;
T5= 5 s after first reaction; T6= 180 s before ROC; T7= 5 s after ROC. (From ref.[69])
LOC
Skin incision
First reaction
ROC
BIS
NRR
HF
LF
LHR
T0
97
(94,98)
-0.26
(-0.42,
0.21)
2.37
(2.05,
2.76)
1.84
(1.50,
2.27)
-0.42
(-0.84,
-0.27)
T1
54
68)
T2
T3
(42,
39
(36,44)
41
(37,44)
T4
61
47)
T5
(70,
73
(67,80)
T6
80
83)
T7
(77,
86
(81,94)
0.09
(-0.09,
0.31)
1.50
(0.97,
1.66)
1.37
(1.01,
1.60)
-0.03
(-0.29,
0.19)
-0.14
(-0.51,
0.02)
0.04
(-0.19,
0.21)
1.21
(0.94,
1.55)
0.84
(0.46,
0.94)
-0.50
(0.46,
0.84)
1.42
(1.08,
1.81)
1.59
(1.25,
2.25)
0.24
(0.01,
0.50)
-0.80
(-1.1,
-0.38)
1.44
(1.09,
1.83)
0.09
(-0.37,
0.56)
-1.33
(-1.67,
-0.66)
0.47
(0.29,
0.61)
1.07
(0.78,
1.23)
0.79
(0.75,
1.31)
0.36
(0.12,
0.51)
1.11
(0.91,
1.40)
0.68
(0.01,
1.27)
0.43
(0.19,
0.62)
1.25
(0.93,
1.60)
0.91
(0.37,
1.26)
-0.25
(-0.73,
0.06)
-0.44
(-0.79,
0.10)
-0.34
(-0.66,
-0.12)
HR
(65,
74
91)
64
72)
(58,
(58,
65
72)
(60,
70
76)
(62,
71
80)
(69,
79
85)
76
83)
(69,
84
91)
(75,
Chapter 6. Results
48
Table 6.2: Prediction Probability and Spearman Rank Correlation for Indices vs. Es-
timated Sevoflurane Effect-site Concentration in Controlled Ventilation and Spontaneous
Breathing. (cid:63) p < 0.05 between non-rhythmic to rhythmic ratio (NRR) and other indices.
Indices with prediction probability (PK) value smaller than 0.5 was corrected by 1 − PK
before comparison. BIS = Bispectral Index; tvHF = high frequency power; tvLF = low
frequency power; tvLHR = LF to HF ratio; HR = heart rate. (From ref.[69])
Controlled Ventilation
PK (SE)
Spearman rank correla-
tion (95% CI)
0.716 (0.020)
-0.575 (-0.661, -0.476)
0.670 (0.025)
-0.467 (-0.583, -0.337)
0.479 (0.027)(cid:63)
0.073 (-0.069, 0.211)(cid:63)
0.582 (0.025)
-0.233 (-0.359, -0.101)(cid:63)
BIS
NRR
tvHF
tvLF
tvLHR
0.581 (0.025)
-0.233 (-0.364, -0.096)(cid:63)
HR
0.423 (0.024)(cid:63)
0.152 (0.034, 0.269)(cid:63)
Spontaneous Breathing
PK (SE)
Spearman rank correla-
tion (95% CI)
BIS index
0.839 (0.014)(cid:63)
-0.836 (-0.881, -0.778)(cid:63)
NRR
tvHF
tvLF
0.736 (0.019)
-0.619 (-0.688, -0.536)
0.489 (0.023)(cid:63)
0.001 (-0.117, 0.113)(cid:63)
0.669 (0.022)(cid:63)
-0.416 (-0.506, -0.317)(cid:63)
tvLHR
0.666 (0.022)(cid:63)
-0.420 (-0.514, -0.319)(cid:63)
HR
0.532 (0.022)(cid:63)
0.027 (-0.075, 0.133)(cid:63)
Chapter 7
Implications to Anesthesiology
and Physiology
7.1 Main Findings
There are two important clinical results in regards to the NRR index. First, the NRR
index is able to predict first reaction during emergency period. Second, the NRR index
correlates with the sevoflurance concentration during the patient's spontaneous breathing.
Even so, the NRR index is not the best predictor for LOC and ROC. These findings suggest
that there is a distinct physiological interpretation of the anesthetic depth level when
compared to the hypnosis measured by surface EEG. Furthermore, tvLF correlates well
with skin incision and endotracheal intubation. This momentary information is currently
unavailable to clinical anesthesiologists as both the IHR and tvPS are not implemented in
current monitor.
7.2 Genuineness of NRR index
The NRR index generally has a sharp increase nearly before the onset of the patient's
first reaction. This then makes it highly unlikely that the motion artifacts interfere the
NRR index. Furthermore, the observed NRR index performance in not confounded by
the ventilation mode change. In particular, aside from the LOC, there are no transitions
that occur from the mechanical ventilation to the spontaneous breathing in the data we
have obtained. Also, there is no positive pressure ventilation for all the first reaction
data. Furthermore, past literature on the association between "rhythmic-to-non-rhythmic"
transition of the oscillatory pattern in RRI index and the different level of anesthesia has
49
Chapter 7. Implications
50
been mentioned but no quantitative works have been done. By using the classical power
spectral analysis, Kato et al. has described that the high frequency component of RRI
disperses when the patient awakens and concentrates when the patient is under anesthetics
and controlled ventilation[2]. Based on the above facts, I am more than confident to state
that the NRR is convincing in its distinct and novel information that it is capable of
providing us.
7.3 Existing Findings
The results we have obtained from the BIS (Fig.6.12, Table 6.2) are in agreement with
previous studies showing that BIS is capable of monitoring awake status vs. anesthesia,
the first reaction and the decrease of anesthetic gas concentration[5, 8, 10]. Our BIS results
(Fig.6.13, Fig.6.16) also agree with past studies that provide evidience that BIS or other
EEG-derived indices cannot indicate the response to noxious stimulation[6, 7, 9].
The results from our analysis (Table 6.1) are also consistent with previous reported associ-
ation between anesthetic depth and traditional HRV parameters: HRV is decreased during
general anesthesia and increased during recovery[2, 46, 70, 71]. Lastly, our results are also
consistent with the reported finding that combining classical HRV indices with BIS does
not outperform BIS in discriminating awake from asleep during anesthetic induction[72].
It has been reported that classical HRV parameters do not outperform heart rate when
predicting noxious stimulation[4, 73]. Contrarily, our results demonstrate that tvLF pre-
dicts skin incision better than HR. It seems that the noxious stimuli of the skin incision
elicited a momentary increase of sympathetic activity, which leads to this finding. Al-
though the study is not designed for noxious stimulation, the finding that tvLF correlates
with skin incision shows the technical advantage of our approach and its potential as an
index for noxious stimulation.
7.4 Physiologic Interpretation of NRR
The association between NRR and anesthetic depth can be partially explained by the
cardiopulmonary coupling[14] and the respiratory physiology. When the coupling effect
is higher, the respiration pattern is reflected in the RRI index. The respiratory mecha-
nism also contributes to the differential influence that occur from anesthetics. The neural
respiratory control comprises of two systems, the involuntary automatic control system
and the voluntary control system. The involuntary automatic control system is mainly
Chapter 7. Implications
51
controlled by the respiratory center in the pontomedullary area. The voluntary control sys-
tem is mainly controlled by the forebrain[25]. These two systems are also distinct in their
neural anatomy and their functions. Respiratory efferent signals from these two systems
compete with each other and are integrated at the spinal level to control the respiratory
motor neuron. However, during spontaneous respiration, studies have also shown that
the involuntary automatic respiratory pattern generated in the pontomedullary center[25],
specifically the preBotzinger complex[26 -- 28], is rhythmic.
On the other hand, the breathing pattern of the voluntary respiratory motor control is
non-rhythmic and involves the cortical processing and thalamic integration which responds
to the peripheral inputs and the descending inputs[25, 30 -- 32]. The thalamus also ac-
tively participates in different motor functions, including respiration[31, 32], speech[33]
and cough[34] to name a few[74]. Our NRR index findings suggest that the non-rhythmic
respiratory activity involves the nearly entire cerebral regions is more susceptible to sevoflu-
rane than the rhythmic respiration generated in the medulla. This suggestion is supported
by literature. Guedel made the classical observation on respiratory pattern[1]. Bimer et
al denoted that the respiratory irregularities accompany electroencephalographic arousal
reaction in spontaneous breathing[75]. Also, Studies have also shown that the respiratory
activity of preBotzinger complex is less depressed by sevoflurane[36, 37], and the effect of
volatile anesthetics is less prominent to the brainstem compared to the cortical regions
that are more abundant in synaptic transmissions.[37].
The ability for the NRR index to predict first reaction suggests a connection between the
"rhythmic-to-non-rhythmic" phenomenon and motor reactions. Antognini et al. demon-
strated that immobility to surgical stimulation by volatile anesthetics in goats is mainly
modulated by the spinal cord, which also indirectly affects the thalamic response to nox-
ious stimulation[12, 13]. Another study by Velly et al. also presented significant results[7].
He recorded the human subcortical electrophysiological activity and demonstrated that
subcortical activity (possible the thalamic activicy) predicts suppression of movement to
noxious stimuli, but not changes of consciousness, whereas cortical EEG predicts loss
of consciousness, but not motor suppression. Based on the above, NRR might reflect
the subcortical activity since it better reveals information about motor reaction, but not
consciousness, compared with BIS. Since immobility to noxious stimulation is similar to,
although not the same as the first reaction under surgical wound pain in the present study,
this relationship suggests the possible role of the thalamus in "rhythmic-to-non-rhythmic"
phenomenon. Besides the NRR index, BIS also responded to first reaction in our serial
PK analysis, which is in agreement with other studies[10]. Therefore, it is possible that a
cortical mechanism is also related to the first reaction.
Chapter 7. Implications
52
From these results, it is possible to see that sevoflurane may affect non-rhythmic respiration
level of the spinal cord, subcortical supraspinal regions, or the cortex in the human brain.
Although we are currently unable to pinpoint the exact anatomic location, it is likely that
the subcortial supraspinal area, possibly the thalamus, is the most plausible region. The
above statement is sound as the NRR index generates different results in the serial PK
analysis and is also closely connected to the human body's respiration responses. Thus, we
hypothesize that the "rhythmic-to-non-rhythmic" phenomenon in IHR exhibits the central
respiratory activity via either central or peripheral mechanisms[14]. The IHR also reflects
the final integration of this "rhythmic-to-non-rhythmic" respiratory activity. Although
more evidence is necessary to further clarify these hypotheses, we propose to use the NRR
quantification methodology as a potential tool to evaluate the depth of anesthesia that
differs from EEG-based monitoring.
7.5 Clinical Application of the NRR Index
The first strength of this study is a new quantitative approach, referred to as the NRR
index, to analyze "rhythmic-to-non-rhythmic" phenomenon. The NRR index has the po-
tential to reflect different levels of anesthesia. Literature partially supports that there
is an underlying physiological mechanism. Second, the signal processing technique, mul-
titaper Synchrosqueezing transform, has been well studied in the literature, so we have
adequate theoretical support to resolve the potential limitations of the PS. The momentary
dynamics in IHR can be captured by the tvPS, which leads to the NRR index.
7.6 Potential Index in Noxious Stimulation
Although we did not design the study for noxious stimulation, from the data analysis
result, we found that the indices we apply are potential in detecting pain reaction under
anesthesia. In particular, the tvLF index correlates with the noxious stimulation better
than heart rate as an additional finding. More study will need to be conducted on the
tvLF to provide sound results.
It can also be helpful to compare this data with other
well known indices like the Surgical Stress Index
®
(SSI, GE healthcare) and Analgesia
Nociception Index
®
(ANI, Metrodoloris), which have been used on monitoring surgical
stress [11]. SSI is derived from the pulse wave amplitude and pulse beat interval of the
photoplethysmography whereas ANI indicates the parasympathetic tone derived from ECG
signal. It is possible that various indices can provide complementary information regarding
noxious stimulation that is gathered from different modalities and concepts.
Chapter 7. Implications
53
NRR index measures the rhythmic-to-non-rhythmic transition and quantifies a new kind of
information different from the above two instruments in both physiologic and mathematical
senses. Further investigations may be necessary to understand the clinical value and
application of tvPS.
7.7 Summary
In conclusion, we are able to extract hidden information regarding the anesthetic levels
from the routine ECG monitor. The quantitative results of the NRR index supports our
initial hypothesis that the "rhythmic-to-non-rhythmic" transition correlates with motor
reactions during emergence period earlier than BIS index, and correlates with sevoflu-
rance concentration. In addition, the potential of the time-varying HRV indices provides
observations of the relationship between tvLF index and the noxious stimulation. The
notion of these dynamics is rooted in the recently developed signal processing technique.
Without tvPS, the above quantification cannot be easily achieved. Overall, my clinical
study suggests that the ECG signal contains complementary information to the EEG-based
depth-of-anesthesia index.
Chapter 8
Real-time Processing Using the
Blending Operator
8.1 Real-time Processing
From the previous chapters, my study has concluded that the IHR signal during anesthesia
contains a wealth of information when viewing with tvP S.
It is possible to derive the
IHR signal from ECG, the waveform of pulse oximetry, and also the invasive intra-arterial
pressure waveform. Although we mainly use ECG signal in the present project, all possible
sources of IHR are mandatory or frequently used functions in modern standard anesthesia
monitors. A practical question is how do we apply the NRR index proposed in the present
study to individual patients.
The processing steps of NRR index starts from the identification of each heart beat from
ECG signal. Next, an interpolation is used to convert the irregularly-sampled heart beats
data into regularly-sampled IHR data. Next, the calculation of the tvP S is based on the
multitaper Synchrosqueezing tranform. And lastly the NRR index is computed based on
the tvP S. Thus, the positive statistical results in previous chapter, are all obtained from
off-line calculations.
In anesthesia practices, the main purpose of anesthesia monitor is to provide real time
monitoring of the patient's responses in order to provide timely responses. This purpose
then turns ordinary measurements like heart rate, blood oxygen saturation, and blood
pressure into vital information during anesthesia. The real-time processing of NRR in-
dex is necessary if we would like to "steer" the anesthetic depth of patients based on
their instantaneous information from IHR. The potential benefits include adequate dose
54
Chapter 8. Real-time Processing
55
of anesthetics, possibility to divert critical situation, reduce stress response, and improve
long-term welfare.
In this chapter, my main focus is the real-time processing from the irregular heart beat
sample to the tvP S. The goal is to interpolate the irregular data samples in real time.
Meanwhile, the polynomials property of the interpolation also serves as the vanishing-
moment and minimum-supported wavelets.
8.2 Technique Obstacles for Real-time Processing
Due to the irregular interval of each heartbeat, the heartbeat data is irregularly sampled
and this irregularity is unavoidable. Almost all continuous information are digitized in
equally-spaced samples in time or in space before further analysis for convenience, and for
efficiency. However, some situations provide non-uniform sampling and could be either
unavoidable or intended[76, 77].
In order to derive the continuous instantaneous heart
rate from the distinct heartbeats (Fig.2.1) is an example of the rich IHR information that
entirely relies on the irregularity of the heartbeats. Hence, the initial challenge is the
real-time processing of the non-uniform data samples.
When analyzing data the scope of continuous wavelet transform (CWT), the theoretical
background of Synchrosqueezing transform presents the optimal performance when us-
ing compactly supported windows in frequency domain[44]. That means we need whole
data converted in frequency domain for subsequent calculation. This is inefficient and
also highly unlikely for real-time data processing. Moreover, for real-time processing, an
wavelet transform (or STFT) better has a short, and compact support window in time.
This type of window may hamper the benefit of wavelet transform: an adequate vanishing
moment can eliminate trend component in terms of corresponding polynomial order to
preserve the oscillatory components we desire more efficiently.
A possible workaround is by cutting down a segment of data, windowing it, and then
choosing a suboptimal wavelet function empirically. Next, we would observe the effect of
suboptimal window and then adjust the parameters empirically until we finally accept the
suboptimal effect in real-time display.
There is a better solution to consider before we proceed with the above workaround. The
solution is a special spline interpolation, which has the interpolating ability to handle the
unequally-spaced heart beat sample, and the polynomial property to match the vanishing
moment property of wavelet transform. The resulting time-scale analysis is compactly
supported in time and provides better performance due to the vanishing moment property.
Chapter 8. Real-time Processing
56
Based on this scheme, the real-time tvPS can help remove the above mentioned technical
obstacles.
8.3 B-spline
We start from a brief introduction of the spline function. For digital samples g(tn) are
sampled at time points
t : a = t0 < t1 < t2 < . . . ,
(8.1)
where {tk} may be irregular, or non-uniform signal, which means that we allow tk+1− tk (cid:54)=
tj+1 − tj for k (cid:54)= j. Let m ≥ 3 be any desired integer and Πm−1 denote the space of all
polynomials of degree less than m.
The spline space is then defined as:
s : s−m+1 = s−m+2 = . . . = a = s0 < s1 < s2 < . . . ,
(8.2)
and the notation Ss,m := Ss,m[a,∞) denotes the spline space of order m with the knot
sequence s.
To formulate a locally supported basis of Ss,m, we now introduce the truncated powers
xm−1
+
:= (max{0, x})m−1.
The truncated power functions (sk − t)m−1
+ , k = 0, 1, . . ., are in Ss,m. Since (xk − t)m−1
+ ,
k ≥ 1, have global support, we apply the m-th order divided differences to form locally
supported functions. The divided differences are
[ u, . . . , u ]f :=
f (l)(u)
l!
if there are l + 1 entries in [u, . . . , u], and
[ u0, . . . , un ]f :=
[ u1, . . . , un ]f − [ u0, . . . , un−1 ]f
un − u0
if u0 ≤ u1 ≤ . . . ≤ un and un > u0.
Now we consider the knot sequence s of the spline space Ss,m. The normalized B-splines[78]
are the divided difference of the truncated power (sk − t)m−1
+ , namely,
Nm,k(t) = Ns,m,k(t) = (sm+k − sk)[ sk, . . . , sm+k ](· − t)m−1
+ ,
(8.3)
for k = −m + 1, . . . , 0, 1, 2, . . ..
Chapter 8. Real-time Processing
57
We would apply the recursive formula to implement Nm,k [52, page 143 (6.6.12a)]. The
B-splines are a locally supported.
However, if we choose the irregularly-spaced sequence t of time positions in (8.1) as the
knot sequences s in (8.2) by attaching t−m+1 = . . . = t0 = a to t as in (8.2), it can be
difficult to compute, or the computation cost will be higher for the non-uniform {tj} and
large values of n.
8.4 Quasi-interpolation
On the other hand, de Boor and Fix [79] proposed a spline representer f (t), which ap-
proximates the data instead of providing exact interpolating the target data function g(t)
at t = tj, j = 0, 1, . . . , n (that is, if f (tj) (cid:54)= g(tj) is allowed). This quasi-interpolation has
a "polynomial preservation" property: g(t) = p(t) ∈ Πm−1.
This polynomial preservation is important to improve the performance of the continuous
wavelet transform (CWT), and hence the Synchrosqueezing transform (SST), to reveal
the oscillatory components. Specifically, if the analysis wavelet is compactly supported by
the spline of order m, then the m-th order vanishing moment annihilates the m-th order
Taylor polynomial expansion of g(t). This then facilitates the trend removal of the the
signal g(t) as a polynomial. However, the quasi-interpolation scheme de Boor proposed
[78, 79] requires derivative data values of g(t). Others [80, 81] proposed to replace the
derivatives of g(t) by divided differences of {g(ti)}. However, these quasi-interpolations[82]
do not address the real-time issue at stake. In addition, quasi-interpolation introduced the
error g(ti) − f (ti) for i = 0, 1, . . . , n.
Chui et al. proposed a real-time quasi-interpolation[83]. Chui and his college also intro-
duced a "local interpolation"[84] to correct the error of "quasi-interpolation".
The real-time quasi-interpolant [83] is introduced as follow: for each k = 0, 1, . . ., consider
the Vandermonde determinant
D(tk, . . . , tk+m+1) := det
1
1
tk
...
tm−1
k
tk+1
...
tm−1
k+1
. . .
. . .
...
. . .
1
tk+m−1
...
tm−1
k+m−1
,
and the determinant D(tk, . . . , tk+j−1, ξj, tk+j+1, . . . , tk + m− 1) obtained by replacing the
(j + 1)-st column in the definition of D(tk, . . . , tk+m+1) by the column vector
ξj := [ ξ0(j, m), . . . , ξm−1(j, m) ]T ,
Chapter 8. Real-time Processing
58
where ξ0(j, m) = 1 and
ξi(j, m) =
σi(tj+1, . . . , tj+m−1)
,
(cid:32)
(cid:33)
for i = 1, . . . , m − 1 with σi(r1, . . . , rm−1) being the classical symmetric functions defined
by σ0(r1, . . . , rm−1) = 1 and for i = 1, . . . , m − 1,
σi(r1, . . . , rm−1) =
1≤l1<...<li≤m−1
rl1 . . . rli.
m − 1
i
(cid:88)
2m−2(cid:88)
l=m−1
(cid:88)
Using the determinants introduced above, we apply the below spline coefficients
ak,l :=
D(tk, . . . , tk+j−1, ξj, tk+j+1, . . . , tk+m−1)
D(tk, . . . , tk+m−1)
,
to formulate the compactly supported spline function
Mt,m,k(t) :=
ak,l−m+1Nt,m,k+l−m+1(t),
with suppMt,m,k = [tk−m+1, tk+m]. These basis functions provide a real-time implementa-
tion of the quasi-interpolation operator
(Qmg)(t) =
g(tk)Mt,m,k(t).
(8.4)
k
The quasi-interpolation operator Qm (8.4) possesses the polynomial preservation property.
For any m ≥ 1,
(Qmp)(t) = p(t),
for all t ≥ a and for all p ∈ Πm−1, provided that the summation (8.4) account for all non-
negative integers k = 0, 1, . . .. Furthermore, in view of the support of Mt,m,k, it follows
that
v+m−2(cid:88)
p(tk)Mt,m,k(t) = p(t),
t ∈ [tu, tv]
for all p ∈ Πm−1.
k=v−m+1
This local polynomial preservation property allows the CWT, and hence the SST, to
annihilate the (m − 1)-th degree Taylor polynomial approximation of the signal at tj,
where u < j < v.
Chapter 8. Real-time Processing
59
8.5 Local Interpolation Operator
The local interpolation operator, denoted by Rm, satisfies the interpolation property. To
define Rm, we will insert knots to t by considering a new knot sequence s ⊃ t (8.2). For
simplicity, we will introduce even order variable m :
For even m ≥ 4, we set
smk/2 = tk,
k = 0, 1, 2, . . . .
That is, we insert (m/2− 1) knots in between two consecutive knots in t. For convenience,
we will choose the new knots to be equally spaced between every pair of two consecutive
knots.
Fix even m ≥ 4. Let Ns,m,j be the m-th order B-spline with knot sequence s. Then the
completely local spline basis function can be defined by
Ls,m,j(t) :=
Ns,m,m(j−1)/2(t)
Ns,m,m(j−1)/2(tj)
.
Since tj = smj/2 is the "centered" knot and
we have
suppNs,m,m(j−1)/2 = [ sm(j−1)/2, sm(j+1)/2 ] = [ tj−1, tj+1 ],
Ls,m,j(tj) = 1
suppLs,m,j = [ tj−1, tj+1 ].
(8.5)
It is clear that Ls,m,j(tk) = δj−k, where the Kronecker delta notation is used (8.5).
The above preparation provides a real-time implementation of the local interpolation op-
erator, which satisfies the interpolation property (8.5).
Fix m ≥ 3. For a given function g ∈ C(R), the local interpolation operator Rm is defined
as
(cid:88)
(Rmg)(t) :=
g(tk)Ls,m,k(t).
(8.6)
k
8.6 Blending Operator
We are now ready to apply (8.4) to obtain the blending operator, denoted by Rm ⊕ Qm.
Chapter 8. Real-time Processing
60
Fix m ≥ 3 and g ∈ C(R). The blending operator is defined as Pm := Rm ⊕ Qm, where
Rm ⊕ Qm := Qm + Rm(I − Qm) = Qm + Rm − RmQm,
and I is the identity operator. In particular, we have
(cid:88)
(cid:88)
(cid:2)g(tk) −(cid:88)
g(tj)Mt,m,j(tk)(cid:3)Ls,m,k(t).
(8.7)
(Pmg)(t) :=
g(tk)Mt,m,k(t) +
k
k
j
We remark that in the definition of Pm, the two operators Rm and Qm are not commutative.
Let us summarize the two key properties of the blending operator in the following theorem.
1. The blending operator Pm possesses both the polynomial preservation property of
Qm and the interpolatory property of Rm.
2. The error of spline interpolation by the blending operator is both small and bounded.
In conclusion, the blending operator, as a local spline interpolation operator, achieves the
optimal interpolation error rate when compared with the traditional spline interpolation
operator.
In addition, the error depends only on the local data profile, which allows
real-time implementation with optimal error rate.
8.7 Real-time tvPS
We provide a real-time computational algorithm to compute f (t) for the upcoming data
samples g(t0), g(t1), . . . etc. For convenience, we only consider even order m. The formu-
lation for the odd order is similar but slightly more complicated.
The blending operator leads to the real-time spline interpolation. For real-time issue, the
wavelets have to be minimum supported in time. We can design the wavelets with the
same number of vanishing moments as the spline. Meanwhile the wavelets have minimum
support and maximum order of vanishing moments.
Let m, n ≥ 1 be arbitrary integers, and x an arbitrary knot sequence. The spline basis
functions is then as follows:
ψx,m;n,k(x) := N (n)
x,m+n,k(x),
k ∈ Z,
(8.8)
Chapter 8. Real-time Processing
61
Algorithm 1 Real-time implementation of the blending operator
First, pre-compute the B-spline values Ns,m,l(tj) =: nl,j. Then for each k, since Mt,m,k ∈
Ss,m, there exists a finite sequence {bk,l} in the formulation of
Mt,m,k(t) =
bk,lNs,m,l(t).
Also pre-compute dk,j = Mt,m,k(tj).
Now, while the data sequence {g(tk)} is acquired, compute
(cid:88)
l
gl =
bk,lg(tk)
(cid:88)
g(tl) −(cid:80)
k
and simultaneously compute
and then up-sample {g∗
g∗
l =
j dl,jg(tj)
,
nm(l−1)/2, l
l } by m(l − 1)/2; that is, set
and g#
l = 0.
m(k−1)/2 = g∗
g#
k,
n(cid:88)
f (t) =
l=−m+1
(gl − g#
l )Ns,m,l(t)
Then, we have an on-line computational scheme for the quasi-interpolation spline inter-
polation:
for increasing number of samples from g(tn) to g(tn+1), . . .; and this can be implemented
for real-time D/A conversion.
to be called VM wavelets on x, where Nx,m+n,k(x) is defined in (8.3), which satisfies the
moment conditions
(cid:90) ∞
(cid:90) ∞
−∞
−∞
xlψx,m;n,k(x)dx = 0,
l = 0, . . . , n − 1
xnψx,m;n,k(x)dx (cid:54)= 0.
(8.9)
That means the arbitrary chosen n in the interpolator leads to the vanishing moment order
of the wavelets if we use this spline basis as the wavelet function. The wavelet transform
can successfully eliminate the n order polynomial as a trend component.
Let m, n ≥ 1 be arbitrary integers. Then the derivative of the wavelet ψx,m;n,k on an
arbitrary knot sequence x, is given by
ψ(cid:48)
x,m;n,k(x) = ψx,m−1;n+1,k(x).
(8.10)
Chapter 8. Real-time Processing
62
This property permits us to conveniently calculate the derivative of the wavelet, which
is required to calculate the frequency reassignment in Synchrosqueezing transform.(see
equation 4.3)
Then, we can proceed to the Synchrosqueezed CWT to obtain the real-time time-frequency
analysis as real-time tvPS. Using this tvPS, we can apply the research in previous chap-
ters to quantify the "rhythmic-to-non-rhythmic" phenomenon in real-time. Using the
clinical data to "emulate" the real-time calculation of NRR index, we obtained a compa-
rable performance of the correlation with sevoflurane concentration during spontaneous
respiration: The PK value of real-time NRR index is 0.711 ± 0.021 (PK of off-line NRR
index:0.732 ± 0.018)(6.2).
Chapter 9
Conclusion
9.1 Research Findings
Thanks to the advancement of science in anesthesiology and neural physiology, I can pro-
pose a theory and modeling to investigate the "rhythmic-to-non-rhythmic" phenomenon.
Furthermore, because of the advancement of signal processing technique, I am able to use
multitaper Synchrosqueezing technique to compute the NRR index. Anesthesia creates
dynamic changes to the human body, and due to the time-varying ability and good TF
resolution, the NRR index is a particularly suitable tool reflecting the constant changes
during anesthesia[43, 61].
The numeric result from clinical database is positive [69]. NRR index performs better and
faster than BIS index when detecting motor movement reaction. This also supports the
fact that different brain regions mediating various functions, with differential susceptibility
to anesthetics may require multi-modal assessment of these anesthetic effects. NRR index
is therefore able to provide additional value. In addition, LF index provides good results
as an indicator for noxious stimulation, for both skin incision and endotracheal intubation.
ECG waveform is standard information in current anesthesia, while the IHR is not. My
study shows that these new indices derived from ECG waveform can provide additional
information that could be beneficial to patients undergoing anesthesia in the future.
9.2 Accomplishments
1. The adaptive harmonic model is a method of modeling the rhythmicity of the "rhythmic-
to-non-rhythmic" phenomenon
63
Chapter 9. Conclusion
64
Figure 9.1: The diagram represents the overall framework from the human brain to
signal analysis.
2. A time-varying power spectrum that is based on multitapering Synchrosqueezed
spectrogram.
3. Methodology of the NRR index
4. A methodology of serial PK analysis was developed to evaluate the dynamic perfor-
mance of anesthetic depth index as time-varying PK value.
5. Positive clinical value of the NRR index was obtained from a clinical database
6. Improved performance of classical HRV parameters was obtained using tvP S
7. A real-time scheme that provides real-time tvP S and real-time NRR index.
9.3 Future Directions
The results of my project indicates several future research possibilities.
In clinical anesthesiology, more studies should be conducted to answer questions regard-
ing the value of NRR index in different kinds of anesthetics (i.e. propofol, ketamine,
dexmedetomidine or xenon), in different types of patient (i.e. children, lying-in women,
Chapter 9. Conclusion
65
elderly patient, or patient with severe illness). Since the NRR index may indicate the
activity of subcortical areas, this poses questions on if the application of the NRR index
can lead to less stress response and better long-term outcomes, and if it is possible, how
can we obtain those results. On the other hand, the IHR is derived from the ECG signal,
I wonder whether it is possible to calculate the NRR index from photoplethsmography or
intra-arterial pressure waveform. It also needs more clinical study to better understand
the clinical role of NRR index and the tvLF index.
It is possible that a thorough understanding of the basic mechanism can lead to better
clinical benefit. The origin of "rhythmic-to-non-rhythmic" phenomenon should be in the
brain. However, the neural mechanism has not been addressed according to my best knowl-
edge. For example, where exactly is the anatomical location that creates this phenomenon?
What could cause the multiple component phenomenon as previously mentioned? Lastly,
how do the neurons coordinate with one another during this phenomenon? It is foreseeable
that the answers of these questions can directly contribute to the brain science. Thus, the
neurophysiological study based on the animal model is a potential future research.
The signal processing technique directly contributes to the performance of these clinical
indices. The current result of tvPS is helpful in revealing various features in the TF plane,
but there are still a lot of room for improvements. The time-frequency analysis is based
mostly on the Fourier transform. I am interested in whether it is possible to discard the
idea of frequency in order to have a better understanding of the signal's features. The real-
time calculation and display of the clinical indices provides benefit to individual patients.
Technique developments in real-time processing is imperative.
9.4 Summary
My study focuses on the qualifications of the "rhythmic-to-non-rhythmic" phenomenon.
I have verified that my proposed model and quantitative index through using clinical
database. The positive results support my proposal on using the adaptive harmonic model
and time-varying power spectrum based on the multitaper Synchrosqueezing transform.
The NRR index also presents unique values in clinical anesthesia. The achievements in
this thesis fill a gap in the fields of medicine and provide a new perspective on many of
the issues that could be understood in the future.
Bibliography
[1] Guedel AE. Inhalational Anesthesia: A Fundamental Guide. New York, Macmillan,
1937.
[2] Kato M, Komatsu T, Kimura T, Sugiyama F, Nakashima K, and Shimada Y. Spectral
analysis of heart rate variability during isoflurane anesthesia. Anesthesiology, 77(4):
669 -- 674, 1992.
[3] Miller RD, Eriksson LI, Fleisher L, Wiener-Kronish JP, and Cohen NH. Miller's
anesthesia. Elsevier Health Sciences, 2014.
[4] Seitsonen ERJ, Korhonen IKJ, Van Gils MJ, Huiku M, Lotjonen JMP, Korttila KT,
and Yli-Hankala AM. EEG spectral entropy, heart rate, photoplethysmography and
motor responses to skin incision during sevoflurane anaesthesia. Acta Anaesthesiolog-
ica Scandinavica, 49(3):284 -- 292, 2005.
[5] Vakkuri A, Yli-Hankala A, Talja P, Mustola S, Tolvanen-Laakso H, Sampson Tl, and
Viertio-Oja H. Time-frequency balanced spectral entropy as a measure of anesthetic
drug effect in central nervous system during sevoflurane, propofol, and thiopental
anesthesia. Acta Anaesthesiologica Scandinavica, 48(2):145 -- 153, 2004.
[6] Storm H, Støen R, Klepstad P, Skorpen F, Qvigstad E, and Raeder J. Nociceptive
stimuli responses at different levels of general anaesthesia and genetic variability. Acta
Anaesthesiologica Scandinavica, 57(1):89 -- 99, 2013.
[7] Velly LJ, Rey MF, Bruder NJ, Gouvitsos FA, Witjas T, Regis JM, Peragut JC, and
Gouin FM. Differential dynamic of action on cortical and subcortical structures of
anesthetic agents during induction of anesthesia. Anesthesiology, 107(2):202 -- 212,
2007.
[8] Soehle M, Ellerkmann RK, Grube M, Kuech M, Wirz S, Hoeft A, and Bruhn J.
Comparison between bispectral index and patient state index as measures of the
electroencephalographic effects of sevoflurane. Anesthesiology, 109(5):799 -- 805, 2008.
66
Bibliography
67
[9] Ekman A, Brudin L, and Sandin R. A comparison of bispectral index and rapidly
extracted auditory evoked potentials index responses to noxious stimulation during
sevoflurane anesthesia. Anesthesia & Analgesia, 99(4):1141 -- 1146, 2004.
[10] Schmidt GN, Bischoff P, Standl T, Hellstern A, Teuber O, and Schulte Esch J. Com-
parative evaluation of the datex-ohmeda s/5 entropy module and the bispectral index
(r) monitor during propofol-remifentanil anesthesia. Anesthesiology, 101(6):1283 --
1290, 2004.
[11] Wennervirta J, Hynynen M, Koivusalo AM, Uutela K, Huiku M, and Vakkuri A. Sur-
gical stress index as a measure of nociception/antinociception balance during general
anesthesia. Acta anaesthesiologica Scandinavica, 52(8):1038 -- 1045, 2008.
[12] Antognini JF and Schwartz K. Exaggerated anesthetic requirements in the preferen-
tially anesthetized brain. Anesthesiology, 79(6):1244 -- 1249, 1993.
[13] Antognini JF, Wang XW, and Carstens E. Isoflurane action in the spinal cord blunts
electroencephalographic and thalamic -- reticular formation responses to noxious stim-
ulation in goats. Anesthesiology, 92(2):559, 2000.
[14] Eckberg DL. Point: counterpoint: respiratory sinus arrhythmia is due to a cen-
tral mechanism vs. respiratory sinus arrhythmia is due to the baroreflex mechanism.
Journal of applied physiology, 106(5):1740 -- 1742, 2009.
[15] Poyhonen M, Syvaoja S, Hartikainen J, Ruokonen E, and Takala J. The effect of
carbon dioxide, respiratory rate and tidal volume on human heart rate variability.
Acta Anaesthesiologica Scandinavica, 48(1):93 -- 101, 2004.
[16] Hall JE. Guyton and Hall Textbook of Medical Physiology: Enhanced E-book. Elsevier
Health Sciences, 2010.
[17] Detsch O, Kochs E, Siemers M, Bromm B, and Vahle-Hinz C. Differential effects of
isoflurane on excitatory and inhibitory synaptic inputs to thalamic neurones in vivo.
British Journal of Anaesthesia, 89(2):294 -- 300, 2002.
[18] Avidan MS, Jacobsohn E, Glick D, Burnside BA, Zhang L, Villafranca A, Karl L,
Kamal S, Torres B, O'Connor M, Evers AS, Gradwohl S, Lin N, Palanca BJ, and
Mashour GA. Prevention of intraoperative awareness in a high-risk surgical popula-
tion. New England Journal of Medicine, 365(7):591 -- 600, 2011.
[19] Monk TG, Saini V, Weldon BC, and Sigl JC. Anesthetic management and one-year
mortality after noncardiac surgery. Anesthesia & Analgesia, 100(1):4 -- 10, 2005.
Bibliography
68
[20] Melamed R, Bar-Yosef S, Shakhar G, Shakhar K, and Ben-Eliyahu S. Suppression of
natural killer cell activity and promotion of tumor metastasis by ketamine, thiopen-
tal, and halothane, but not by propofol: mediating mechanisms and prophylactic
measures. Anesthesia & Analgesia, 97(5):1331 -- 1339, 2003.
[21] Anand KJS and Hickey PR. Halothane -- morphine compared with high-dose sufentanil
for anesthesia and postoperative analgesia in neonatal cardiac surgery. New England
Journal of Medicine, 326(1):1 -- 9, 1992.
[22] Wallace AW, Galindez D, Salahieh A, Layug EL, Lazo EA, Haratonik KA, Boisvert
DM, and Kardatzke D. Effect of clonidine on cardiovascular morbidity and mortality
after noncardiac surgery. Anesthesiology, 101(2):284 -- 293, 2004.
[23] Mangano DT, Layug EL, Wallace A, and Tateo I. Effect of atenolol on mortality and
cardiovascular morbidity after noncardiac surgery. New England Journal of Medicine,
335(23):1713 -- 1721, 1996.
[24] Bardram L, Funch-Jensen P, Jensen P, Kehlet H, and Crawford ME. Recovery after
laparoscopic colonic surgery with epidural analgesia, and early oral nutrition and
mobilisation. The Lancet, 345(8952):763 -- 764, 1995.
[25] Mitchell RA and Berger AJ. Neural regulation of respiration. The American review
of respiratory disease, 111(2):206 -- 224, 1975.
[26] Elsen FP and Ramirez JM. Postnatal development differentially affects voltage-
activated calcium currents in respiratory rhythmic versus nonrhythmic neurons of
the pre-botzinger complex. Journal of neurophysiology, 94(2):1423 -- 1431, 2005.
[27] Rekling JC and Feldman JL. Prebotzinger complex and pacemaker neurons: hypothe-
sized site and kernel for respiratory rhythm generation. Annual Review of Physiology,
60(1):385 -- 405, 1998.
[28] Ramirez JM, Quellmalz UJA, Wilken B, and Richter DW. The hypoxic response
of neurones within the in vitro mammalian respiratory network. The Journal of
Physiology, 507(2):571 -- 582, 1998.
[29] Shea SA. Behavioural and arousal-related influences on breathing in humans. Exper-
imental physiology, 81(1):1 -- 26, 1996.
[30] McFarland NR and Haber SN. Thalamic relay nuclei of the basal ganglia form both
reciprocal and nonreciprocal cortical connections, linking multiple frontal cortical
areas. The Journal of neuroscience, 22(18):8117 -- 8132, 2002.
[31] Ramsay SC, Adams L, Murphy K, Corfield DR, Grootoonk S, Bailey DL, Frackowiak
RS, and Guz A. Regional cerebral blood flow during volitional expiration in man:
Bibliography
69
a comparison with volitional inspiration. The Journal of physiology, 461(1):85 -- 101,
1993.
[32] Mckay LC, Evans KC, Frackowiak RSJ, and Corfield DR. Neural correlates of volun-
tary breathing in humans. Journal of Applied Physiology, 95(3):1170 -- 1178, 2003.
[33] Murphy K, Corfield DR, Guz A, Fink GR, Wise RJS, Harrison J, and Adams L.
Cerebral areas associated with motor control of speech in humans. Journal of Applied
Physiology, 83(5):1438 -- 1447, 1997.
[34] Simonyan K, Saad ZS, Loucks TMJ, Poletto CJ, and Ludlow CL. Functional neu-
roanatomy of human voluntary cough and sniff production. Neuroimage, 37(2):401 --
409, 2007.
[35] Sloan TB. Anesthetic effects on electrophysiologic recordings. Journal of clinical
neurophysiology, 15(3):217 -- 226, 1998.
[36] Kuribayashi J, Sakuraba S, Kashiwagi M, Hatori E, Tsujita M, Hosokawa Y, Takeda
J, and Kuwana S. Neural mechanisms of sevoflurane-induced respiratory depression
in newborn rats. Anesthesiology, 109(2):233 -- 242, 2008.
[37] Takita K and Morimoto Y. Effects of sevoflurane on respiratory rhythm oscillators in
the medulla oblongata. Respiratory physiology & neurobiology, 173(1):86 -- 94, 2010.
[38] Fink BR, Hanks EC, Ngai SH, and Papper EM. Central regulation of respiration
during anesthesia and wakefulness. Annals of the New York Academy of Sciences,
109(2):892 -- 900, 1963.
[39] Moody GB, Mark RG, Zoccola A, and Mantero S. Derivation of respiratory signals
from multi-lead ECGs. Computers in cardiology, 12:113 -- 116, 1985.
[40] Houtveen JH, Groot PF, and de Geus EJ. Validation of the thoracic impedance
derived respiratory signal using multilevel analysis.
International Journal of Psy-
chophysiology, 59(2):97 -- 106, 2006.
[41] Flaherty JT, Blumenschein SD, Alexander AW, Gentzler RD, Gallie TM, Boineau
JP, and Spach MS. Influence of respiration on recording cardiac potentials: Isopo-
tential surface-mapping and vectorcardiographic studies∗. The American journal of
cardiology, 20(1):21 -- 28, 1967.
[42] Task Force of the European Society of Cardiology, Task Force of the European Soci-
ety of Cardiology, et al. the north american society of pacing and electrophysiology.
heart rate variability: standards of measurement, physiological interpretation and
clinical use. Circulation, 93(5):1043 -- 1065, 1996.
Bibliography
70
[43] Lin YT, Hseu SS, Yien HW, and Tsao J. Analyzing autonomic activity in electrocar-
diography about general anesthesia by spectrogram with multitaper time-frequency
reassignment. In Biomedical Engineering and Informatics (BMEI), 2011 4th Inter-
national Conference on, volume 2, pages 630 -- 634. IEEE, 2011.
[44] Daubechies I, Lu J, and Wu HT. Synchrosqueezed wavelet transforms: an empirical
mode decomposition-like tool. Applied and computational harmonic analysis, 30(2):
243 -- 261, 2011.
[45] Chen YC, Cheng MY, and Wu HT. Non-parametric and adaptive modelling of dy-
namic periodicity and trend with heteroscedastic and dependent errors. Journal of
the Royal Statistical Society: Series B (Statistical Methodology), 76(3):651 -- 682, 2014.
[46] Huang HH, Chan HL, Lin PL, Wu CP, and Huang CH. Time-frequency spectral
analysis of heart rate variability during induction of general anaesthesia. British
journal of anaesthesia, 79(6):754 -- 758, 1997.
[47] Xiao J and Flandrin P. Multitaper time-frequency reassignment for nonstationary
spectrum estimation and chirp enhancement. Signal Processing, IEEE Transactions
on, 55(6):2851 -- 2860, 2007.
[48] Mazzeo AT, La Monaca E, Di Leo R, Vita G, and Santamaria LB. Heart rate vari-
ability: a diagnostic and prognostic tool in anesthesia and intensive care. Acta Anaes-
thesiologica Scandinavica, 55(7):797 -- 811, 2011.
[49] Thakur G, Brevdo E, Fuckar NS, and Wu HT. The synchrosqueezing algorithm for
time-varying spectral analysis: robustness properties and new paleoclimate applica-
tions. Signal Processing, 93(5):1079 -- 1094, 2013.
[50] Flandrin P. Time-frequency/time-scale analysis, volume 10. Academic Press, 1998.
[51] Daubechies I et al. Ten lectures on wavelets, volume 61. Philadelphia: SIAM, 1992.
[52] Chui CK. Wavelets: A Mathematical Tool for Signal Analysis, volume 1. SIAM, 1997.
[53] Chui CK. An introduction to wavelets, volume 1. Academic press, 1992.
[54] Kodera K, Gendrin R, and Villedary C. Analysis of time-varying signals with small bt
values. Acoustics, Speech and Signal Processing, IEEE Transactions on, 26(1):64 -- 76,
1978.
[55] Auger F and Flandrin P. Improving the readability of time-frequency and time-scale
representations by the reassignment method. Signal Processing, IEEE Transactions
on, 43(5):1068 -- 1089, 1995.
Bibliography
71
[56] Daubechies I and Maes S. A nonlinear squeezing of the continuous wavelet transform
based on auditory nerve models. Boca Raton, FL, USA: CRC Press, 1996.
[57] Wu HT, Flandrin P, and Daubechies I. One or Two Frequencies? The Synchrosqueez-
ing Answers. Advances in Adaptive Data Analysis, 3:29 -- 39, 2011.
[58] Thakur G and Wu HT. Synchrosqueezing-based recovery of instantaneous frequency
from nonuniform samples. SIAM Journal on Mathematical Analysis, 43(5):2078 -- 2095,
2011.
[59] Wu HT. Instantaneous frequency and wave shape functions (i). Applied and Compu-
tational Harmonic Analysis, 35(2):181 -- 199, 2013.
[60] Wu HT, Chan YH, Lin YT, and Yeh YH. Using synchrosqueezing transform to
discover breathing dynamics from ecg signals. Applied and Computational Harmonic
Analysis, 36(2):354 -- 359, 2014.
[61] Auger F, Flandrin P, Lin YT, McLaughlin S, Meignen S, Oberlin T, and Wu HT. A
coherent overview of time-frequency reassignment and synchrosqueezing. IEEE Signal
Processing Magazine, 30(6):32 -- 41, 2013.
[62] Steven MK. Modern Spectral Estimation: Theory and Application. Prentice-Hall,
N.J., 1988.
[63] Thomson DJ. Spectrum estimation and harmonic analysis. Proceedings of the IEEE,
70(9):1055 -- 1096, 1982.
[64] Bayram M and Baraniuk RG. Multiple window time-varying spectrum estimation.
Nonlinear and Nonstationary signal processing, pages 292 -- 316, 2000.
[65] Chui CK and Chen G. Kalman filtering: with real-time applications. Springer, N.Y.,
2009.
[66] Sheiner LB, Stanski DR, Vozeh S, Miller RD, and Ham J. Simultaneous modeling
of pharmacokinetics and pharmacodynamics: application to d-tubocurarine. Clinical
pharmacology and therapeutics, 25(3):358 -- 371, 1979.
[67] Smith WD, Dutton RC, and Smith NT. Measuring the performance of anesthetic
depth indicators. Anesthesiology, 84(1):38 -- 51, 1996.
[68] Wolfe R and Hanley J. If we're so different, why do we keep overlapping? when 1
plus 1 doesn't make 2. Canadian Medical Association Journal, 166(1):65 -- 66, 2002.
[69] Lin YT, Wu HT, Tsao J, Yien HW, and Hseu SS. Time-varying spectral analysis re-
vealing differential effects of sevoflurane anaesthesia: non-rhythmic-to-rhythmic ratio.
Acta Anaesthesiologica Scandinavica, 58(2):157 -- 167, 2014.
Bibliography
72
[70] Blues CM and Pomfrett CJ. Respiratory sinus arrhythmia and clinical signs of anaes-
thesia in children. British journal of anaesthesia, 81(3):333 -- 337, 1998.
[71] Galletly DC, Westenberg AM, Robinson BJ, and Corfiatis T. Effect of halothane,
isoflurane and fentanyl on spectral components of heart rate variability. British journal
of anaesthesia, 72(2):177 -- 180, 1994.
[72] Sleigh JW and Donovan J. Comparison of bispectral index, 95% spectral edge fre-
quency and approximate entropy of the eeg, with changes in heart rate variability
during induction of general anaesthesia. British journal of anaesthesia, 82(5):666 --
671, 1999.
[73] Luginbuhl M, Ypparila-Wolters H, Rufenacht M, Petersen-Felix S, and Korhonen I.
Heart rate variability does not discriminate between different levels of haemodynamic
responsiveness during surgical anaesthesia†. British journal of anaesthesia, 98(6):
728 -- 736, 2007.
[74] Tanaka M and Kunimatsu J. Contribution of the central thalamus to the generation
of volitional saccades. European Journal of Neuroscience, 33(11):2046 -- 2057, 2011.
[75] Bimer J and Bellville JW. Arousal reactions during anesthesia in man. Anesthesiology,
47(5):449 -- 454, 1977.
[76] Gunnarsson F and Gustafsson F. Frequency analysis using non-uniform sampling with
application to active queue management. In Acoustics, Speech, and Signal Processing,
2004. Proceedings.(ICASSP'04). IEEE International Conference on, volume 2, pages
ii -- 581. IEEE, 2004.
[77] Vio R, Strohmer T, and Wamsteker W. On the reconstruction of irregularly sampled
time series. Publications of the Astronomical Society of the Pacific, 112(767):74 -- 90,
2000.
[78] de Boor C. A practical guide to splines. Springer-Verlag, 1978.
[79] de Boor C and Fix GJ. Spline approximation by quasi-interpolants. Journal of
Approximation Theory, 8:96 -- 110, 1973.
[80] Lyche T and Schumaker LL. Local spline approximation methods. Journal of Ap-
proximation Theory, 15(4):294 -- 325, 1975.
[81] Schumaker LL. Spline functions: Basic theory. Wiley-Interscience, 1981.
[82] de De Villiers J. Mathematics of Approximation. Atlantis Press, 2012.
[83] Chen G, Chui CK, and Lai M. Construction of real-time spline quasi-interpolation
schemes. Approx. Th. and its Appl., 4:61 -- 75, 1988.
Bibliography
73
[84] Chui CK and Diamond H. A general framework for local interpolation. Numerische
Mathematik, 58(1):569 -- 581, 1990.
|
1911.04585 | 1 | 1911 | 2019-11-11T22:21:22 | Modulation of oculomotor control during reading of mirrored and inverted texts | [
"q-bio.NC"
] | The interplay between cognitive and oculomotor processes during reading can be explored when the spatial layout of text deviates from the typical display. In this study, we investigate various eye-movement measures during reading of text with experimentally manipulated layout (word-wise and letter-wise mirrored-reversed text as well as inverted and scrambled text). While typical findings (e.g., longer mean fixation times, shorter mean saccades lengths) in reading manipulated texts compared to normal texts were reported in earlier work, little is known about changes of oculomotor targeting observed in within-word landing positions under the above text layouts. Here we carry out precise analyses of landing positions and find substantial changes in the so-called launch-site effect in addition to the expected overall slow-down of reading performance. Specifically, during reading of our manipulated text conditions with reversed letter order (against overall reading direction), we find a reduced launch-site effect, while in all other manipulated text conditions, we observe an increased launch-site effect. Our results clearly indicate that the oculomotor system is highly adaptive when confronted with unusual reading conditions. | q-bio.NC | q-bio |
Modulation of oculomotor control during
reading of mirrored and inverted texts
Johan Chandra, André Krügel, Ralf Engbert
Deparment of University of Potsdam, Potsdam, Germany
November 13, 2019
Abstract
The interplay between cognitive and oculomotor processes during
reading can be explored when the spatial layout of text deviates from
the typical display. In this study, we investigate various eye-movement
measures during reading of text with experimentally manipulated lay-
out (word-wise and letter-wise mirrored-reversed text as well as in-
verted and scrambled text). While typical findings (e.g., longer mean
fixation times, shorter mean saccades lengths) in reading manipulated
texts compared to normal texts were reported in earlier work, little
is known about changes of oculomotor targeting observed in within-
word landing positions under the above text layouts. Here we carry
out precise analyses of landing positions and find substantial changes
in the so-called launch-site effect in addition to the expected overall
slow-down of reading performance. Specifically, during reading of our
manipulated text conditions with reversed letter order (against overall
reading direction), we find a reduced launch-site effect, while in all
other manipulated text conditions, we observe an increased launch-
site effect. Our results clearly indicate that the oculomotor system is
highly adaptive when confronted with unusual reading conditions.
1 Introduction
Visual acuity is greatest at the center of visual field (the fovea) and declines
sharply on the periphery, which limits the information extraction process
of visual input from the environment. To compensate the limitation, the
eyes generate short and rapid movements, saccades, to shift the fovea to
the regions of interests for high-acuity information processing [11]. Similar
principle is observed during processing of writtten text or reading. During
reading, the eyes typically move forward about 6-7 character spaces dur-
ing saccades and fixate on a word for about 200-250 ms to support word
processing. The control of (saccadic) eye movements requires the coordina-
tion of several fundamental cognitive subsystems such as word recognition,
attention [8], and oculomotor control [45]. While the cognitive system is
responsible for selecting which word to be fixated next, it is the oculomotor
system that is responsible for shifting the fovea to the regions of interests for
high-acuity information processing [11]. Thus, our reading ability depends
on the oculomotor performance, whose properties are reflected most clearly
in the statistics of within-word fixations, typically the eyes' landing position
on words after saccades.
Unlike the well-documented effects of cognitive modulation on temporal
aspects (e.g. fixation duration) of eye movement measures [e.g. 14, 40], small
effects of cognitive modulation on spatial aspects (e.g. within-word landing
position) of eye movement measures were reported. For example, ortho-
graphic familiarity and regularity influence landing positions [12, 55, 56, 57].
Furthermore, corpus analyses showed a significant effect of word frequency on
mean fixation position: Saccades landed further into the (3- to 6-letter) tar-
get word, when it was a high-frequency word as compared to a low-frequency
word [29]. Lavigne [24] reported a shift of initial fixation location toward
the end of high predictable words, but only when the words were seen more
frequently and for saccades lauched near the word beginning. On the other
hand, Rayner [41] reported that word predictability had little influence on
initial landing position on word and suggested that landing position effects
in reading were primarily modulated by low-level processing. Finally, data
from z-string scanning (where all letters were replaced by the letter "z" or
"Z") indicated that within-word landing position distributions are very sta-
ble and do not critically depend on meaningful content ([31]; see also [27]).
Moreover, most effects of higher-level processing on mean fixation position
are small (< 0.5 character spaces).
Interestingly, the within-word landing position were reported to influence
the "higher-level" processes. In several studies, it was reported that (isolated)
word recognition time was at the minimum when the eyes land at the center
of the word compared to when the eyes land on the word's periphery, termed
2
the optimal viewing position [33, 34]. Similar but weaker effect was observed
on refixation probability in reading: readers were less likely to refixate the
words if fixation land near the word center [53]. Oppositely, mean single
fixation duration is the greatest when the eyes land at word center, termed
the inverted optimal viewing position in Vitu et al.
[52]. To explain the
effect, Nuthmann and colleagues [30] argued that fixations landed on the
word egdes were typically not intended to land on the target (mislocated
fixations), reflected on the short durations.
The observed evidences in reading researches are the result of the inte-
gration of cognitive processes in word selection and oculomotor processes in
shifting the eyes to the area of interest. To isolate the underlying process
affected the observation, one should manipulate factors associated to one
process while controling the remaining factors. In fact, Kolers and Perkins
[15] used geometric rotations, reflections and other transformations of text
as the physical variation to study the recognizability of the texts and the
influence that practice in reading one type of transformation applied on the
recognition of others. They found that the transformations being tested var-
ied in difficulty and transferability. Likewise, Kowler and Anton [19] applied
similar types of transformation to test the effects on global saccade lengths
and fixation durations. By observing eye movement patterns of two partic-
ipants, they reported that the directional pattern of saccades had relatively
modest effects on reading speed under the instruction to read accurately.
They argued that the reading time was affected by longer time needed to
generate short saccades observed in reading difficult texts.
In a separate
test, they found negative relationship between saccade length and saccade
latency: short saccades (less than 30') have longer latency than long saccades.
Additionally, as a response to the Internet myth, Rayner et al.
[42] tested
different types of word transposition (internal, beginning, and end of word)
on reading time and reported that although participants were able to read
the text, reading time was slower for some transposition types, especially
when the word beginning was transposed. Hence they concluded the impor-
tance of word beginning (see also [54] for similar conclusion). Following the
above approach of presenting texts in unfamiliar representation, we designed
a study with four different experimental conditions and a control condition
to systematically investigate the possible modulations of oculomotor pro-
cesses in response to variations of the spatial layout of texts. Furthermore,
the current study employs various eye movement measures to describe the
oculomotor performances during reading.
3
In current study, letter positions and word representations were exper-
imentally manipulated in the following ways: We used texts composed of
mirror-inverted letters (mL), mirror-inverted words (mW), inverted words
(iW), where regular letters are printed in reverse order, and scrambled letters
(sL) (see Figure 1a for an example). In mirrored-words (mW) and mirrored-
letter (mL) conditions, either the complete word or the constituting letters
were mirror-inverted with respect to the vertical axis. In contrast, no mirror-
ing was involved in the construction of inverted-word (iW) and scrambled-
word (sL) text conditions. In inverted words (iW) condition, letter represen-
tation was normal, but the position was inverted in the iW condition to mimic
the letter position in the mW condition. In scrambled letters (sL) condition,
the positions of the first and the last letter of a word were maintained, while
the letter positions in between were randomized (i.e., there was no change in
words with length less than 4 letters). The condition mL, mW and iW are
equivalent to the condition NNV, NRV and NRN conditions in [19] while the
condition sL is equivalent to the internal transposition manipulation in [42].
The following sections describe the robust finding on within-word landing
position distribution during reading and the proposed models to explain the
observed phenomena. Furthermore, we will discuss about current reading
models and their predictions, particularly on saccade generation and "where"
the eyes land, in relation to the manipulations in current study.
Within-word landing positions
Regarding where the eyes land during reading, a robust finding is that within-
word landing position approximately follows a Gaussian density function,
with a pronounced peak, typically located halfway between word beginning
and word center, but with a surprisingly large variance [39, 28, 32]. The
landmark study by McConkie et al.
[28] identified two independent oculo-
motor error components in reading, which we will denote as the range-error
model throughout this article: (i) The random placement error is assumed to
reflect perceptuo-oculomotor inaccuracy in the execution of saccades, which
can be approximated by a Gaussian distribution. (ii) The saccadic range
error represents a systematic, launch-site contingent shift of mean landing
positions and is typically explained as a general response bias of the human
motor system [36].
Specifically, McConkie and colleagues [28] found that during reading, the
within-word landing positions varied systematically with the launch-site dis-
4
(a)
(b)
Figure 1: Experimental sentence stimuli and hypotheses on saccade lengths.
(a) In the control condition, normal German text (N) is presented; for mir-
rored letters (mL) and scrambled words (sL), we expect shorter saccade
lengths, on average. For mirrored words (mW) and inverted words (iW),
word beginnings and word ends are exchanged, so that we expect longer
mean saccade lengths and more regressions due to more frequent re-readings
of the same string. (b) The distribution of saccade lengths (solid lines) and
the expected changes due to the experimental manipulations.
5
JedeSprachederWeltbesitzteineGrammatik.(N)(mL)(mW)(iW)(sL)JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.WeltbesitzteineGrammatik.Legende:!!!!!!!!!Vorwärtssakkade,!!!!!!!!!Refixa1on,!!!!!!!!!Skipping,!!!!!!!!!!!Regression!!(N)!!!(G)!!!(bG)!JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.JedeSprachederWeltbesitzteineGrammatik.edeJehcarpSredtleWtztisebeniekitammarG.JdeeScrahpederWletbsizettenieGmartimak.-10-505101520Saccade length [char]Relative frequencyForward saccadesRegressionsShorter saccades (all)More Regressions (mW, iW) tance, i.e., the distance between fixation location (before the saccade) and
the beginning of the target word. The stable observation in reading is that
each letter increment of the saccade launch-site generates a shift of mean
landing-site with a magnitude of about half a character space to the left, the
launch-site effect, which is independent of the target word length. If the dis-
tances between launch-sites and landing sites are measured relative to word
centers, then the within-word mean landing position can be described by a
linear landing-position function [see also 38] of the form
∆P V P = λ · (L0 − L) ,
(1)
where ∆P V P denotes the average shift of the within-word mean landing po-
sition from the word center and L is the distance between launch site and
the center of the target word. While a negative value of ∆P V P indicates that
the within-word mean landing position shifts to the left of the word center,
a positive value indicates a rightward shift. The parameter L0 represents
the center-based launch-site distance, where saccades land precisely on the
word center, on average. The strength of the launch-site effect is represented
by the slope parameter λ. An estimated slope of λ ≈ 0.5 was observed in
readers of English [28] and German [30] texts.
While the range-error model was generally a successful first description
of the eye-movement data, experimental studies demonstrated effects that
could not be explained within this model. First, the presence of additional
stimuli could influence the saccadic landing positions [3, 4, 10, 49, 50, 51].
Second, Krügel and Engbert [20] demonstrated that saccade type (i.e. word
skipping) could influence the saccade landing positions during reading [see
also 22, 37]. These findings challenged the range-error model in explaining
observed saccadic landing positions during reading.
It is important to note, however, that the range-error model is purely
descriptive, since it does not include more fundamental computational prin-
ciples for oculomotor control. Furthermore, the slope parameter of λ repre-
sents quantification of the strength of the launch-site effect, without direct
inferences on what processes underlying the observed phenomenon. As a
consequence, it is not surprising that integrating new experimental evidence
in the range-error model is difficult. In the next section, we discuss a process-
oriented Bayesian model of within-word fixation position that was developed
over the last 10 years.
6
A Bayesian model of oculomotor control in reading
The framework of Bayesian decision theory has been proposed as a principled
approach to the optimal control of human behavior in the context of integra-
tion of sensorimotor and cognitive processes [16, 17, 18]. Since saccadic eye
movements require both sensorimotor (i.e., moving the eyes to foveate words)
and cognitive processes, Engbert and Krügel [5] proposed that eye movement
control during reading could be explained using Bayesian estimation. Accord-
ing to Bayes rule [e.g., 58], the optimal estimate of a target position x, given
a sensory observation at x0, can be calculated as the conditional probability
(posterior)
π(xx0) ∼ q(x0x)p(x) ,
(2)
where p(x) is the previously learned prior distribution of the target, inde-
pendent of current sensory input, and the conditional probability q(x0x) is
the sensory likelihood of the observation at position x0 given a target at x.
The relation in Eq. (2) determines the dependence of the posterior proba-
bility from x. The missing constant of proportionality can be obtained by
normalization.
Engbert and Krügel [20] proposed that the dependence of landing po-
sitions within a word on the saccadic launch site is a special case of the
Bayesian principles, Eq. (2), for saccade planning in reading. In a mathe-
matical model, the likelihood q(x0x) was modeled as an unbiased, normally
distributed probability density centered at the intended target word and with
0, which represents the degree of sensory uncertainty. Assuming
variance σ2
that the prior distribution is also normally distributed, the product of the
two Gaussian densitiy functions results in another normal density function,
the posterior probability density π(xx0). The posterior probability provides
a natural explanation of the launch-site effect, since the position of its maxi-
mum falls between the maximum of the prior distribution and the maximum
of the sensory likelihood. As a result, the posterior reproduces the systematic
tendencies of saccades (i) to overshoot the center of close target words and
(ii) to undershoot the center of distant target words [5].
The shift of the mean of the posterior µP from the observation x0 can be
calculated as
∆Bayes = µP − x0 =
σ2
0
0 + σ2
σ2
T
(µT − x0) ,
(3)
where σ2
T is the variance of the prior probability of center-based launch-
site distances. Comparing the equation for the launch-site effect, Eq. (1),
7
with the predicted effect in the Bayesian theory, Eq. (3), thus assuming
∆P V P ≡ ∆Bayes,we obtain
.
(4)
λ =
σ2
0
σ2
0 + σ2
T
Recently, Krügel and Engbert [21] introduced an advanced model, which
includes an explicit model for the computation of the word center from sen-
sory estimates of word boundaries. Therefore, Bayesian models provide a
robust theoretical framework to explain where the eyes move during reading.
One advantage to model within-word landing position distribution using an
explicit Bayesian model is that the interpretability of the results. The slope
parameter λ estimated from Bayesian model represents weighting of optimal
behavior during reading: maintaining constant saccade length while target-
ing word center. For extreme case where λ → 0, it can be interpreted that
the optimal oculomotor control in reading puts more weight on the precision
in landing on target precision, hence minimizing range error. On the other
hand, the value of λ → 1 means that optimal reading behavior rely on main-
taining constant saccade length, reducing the importance of target location.
With the typical empirical value of λ ≈ 0.5, we obtain the important result
that the sensory variance of the target location is approximately the same
as the variance of the prior distribution, i.e., σ2
T , in optimal reading
behavior.
0 ≈ σ2
Hypothesis and predictions from various reading models
Reading models were typically developed to help understanding the complex
processes underlying reading processes. Despite the fact that eye movement
control during reading requires both cognition and oculomotor systems, as
mentioned above, empirical studies found that cognition had small effects
on within-word landing positions. Interestingly, adding additional visual cue
[? ] or changing sentence presentation, i.e. texts are read from top to bot-
tom [13] did not notably change the landing position distributions on words.
Consequently, most mathematical models of saccade generation during read-
ing assume that oculomotor control is dissociated from cognitive processes.
Cognitive-based reading models (e.g., E-Z Reader:
[44, 43, 46]; SWIFT:
[6, 8, 47]) assume that cognitive processes related to language processing are
responsible for eye movements without distinguished effects on oculomotor
within-word targeting process. Moreover, most cognitive models based their
8
implementation of saccadic errors on the range-error model ([28]) with rel-
atively fixed values. Since the manipulation types tested in current study
maintain the spatial information such as word length, cognitive-based mod-
els will not predict substantial changes in within-word landing positions since
they asumme that oculomotor process is mainly affected by "low level" in-
formation. However, since text manipulations will increase processing loads,
hence affecting saccade generating time, these models will generate more
refixation saccades but less skipping saccades.
Most relevant to current study is Mr.Chips [26, 25], an ideal-observer
model that combine visual, lexical, and oculomotor information optimally
to read simple texts in the minimum number of saccades. The model op-
erates according to an entropy-minimization principle, generating saccades
that minimize uncertainty about the current word or saccades that move the
visual span furtherst to the right. Note that for Mr.Chips, a word is said to be
fixated if the central slot of the visual span (a linear array of character slots
with each slot can be either high or low resolution) falls on one of the letters
of the word and this central slot does not have preferred status. Mr. Chips'
skipping rate and global landing position distribution were similar to human
data. Mean saccade length decreases as the lexion size increases. Further-
more, it generated less refixations and reduced launch-site effect (λ = 0.21)
in reading normal text. Given that word identification played a key role for
lexical processing in the model, we speculate that if the words were written
from right to left (e.g.
in mW and iW conditions), Mr. Chips should gen-
erate more saccades that land on the second half of a word to capture more
information about word identity, assuming that the first half of the word
were identified beforehand and letter mirroring and inversion do not affect
it's lexical access process.
Our hypothesis for eye-movement measures on reading unfamiliar typog-
raphy are derived from the predicted increase in perceptual difficulty and
additional oculomotor demand. Longer fixation durations and shorter sac-
cade amplitudes can be expected for more difficult texts in all four conditions.
Increased perceptual difficulty should also result in less skipping cases, but
more refixations; both of these predictions are compatible with a reduced
average saccade length (Fig. 1b).
On the level of within-word landing postions, our hypotheses are more
specific for the different experimental conditions. Reading words with mirror-
inverted letters (mL) or with scrambled letters (sL) will produce within-
word landing positions similar to normal reading, since information on letter
9
positions did not change (mL) or did not deviate systematically from nor-
mal reading (sL). However, due to inversion of letter positions in reading
mirrored-words (mW) and inverted words (iW), we expect readers to shift
their eyes further to the right of the word string in the initial saccade and to
generate a regressive refixation after that inital saccade (see Fig. 1a).
Our study employs a learning paradigm, where participants are required
to train reading text in one of the four experimental conditions. The mo-
tivation for the learning paradigm is to check the stability of the resulting
eye-movement patterns and to exclude the possibility that the results are a
short-lived effect due to first exposure to an unfamiliar layout.
In general, we expect that the training will lead to improved performance
on the level of global reading measures (average fixation times and mean
saccade lengths). Since the scrambled letters (sL) condition represents the
least systematic variation of the text layout, we expect that the learning
effects are smaller than in the other three experimental conditions. On the
level of within-word landing positions, we expect that stable shift of the initial
landing positions towards the end of the word strings will be established
during learning in the conditions with mirrored words (mW) and inverted
words (iW), since the word beginnings are at the end of the manipulated
string in these two conditions. For the other two conditions (mL, sL), we
expect that within-word landing postions are very similiar to normal reading
after training.
2 Results
Reading text with manipulated layout due to mirrored-reversed and inverted
letter arrangement produced changes in behavioral measures on a global level
(e.g., mean fixation duration, average saccade length) and on the oculomtor
level (i.e., within-word fixation locations/saccadic landing positions). While
the primary goal of our study was to investigate the possible modulations
of oculomotor processes in response to variations of the spatial layout of
texts, we start with the presentation of results on global summary statistics
to evaluate the overall effects on reading performance as characterized by
eye-movement measures.
10
Global summary statistics
This section highlights the significant results for each manipulation across
training session in comparison with normal reading. If not otherwise men-
tioned, results refer to comparisons between the first experimental session
(in one of the four conditions mW, iW, mL, or sL) and the control condition
(normal reading).
First of all, global summary statistics from reading normal texts are repli-
cated. In control condition, most of the first-pass fixations move forward to
the next neighbouring words (forward saccades 41%) or skip (skipping sac-
cades 26%). About 22% of the fixations move within a word (refixation
saccades) and 12% of them move backward (regression saccades). On av-
erage, readers move their eyes 7.82 character spaces forward, 4.23 character
spaces backward and fixated on words for 245 ms during reading normal text.
Compared to control condition, percentage of forward saccades in experimen-
tal condition did not show strong deviation (31% - 44%). The percentage
of skipping saccades was reduced to less than 5% in conditions where texts
were written against the reading direction (e.g. in mW and iW conditions)
and between 10-15% in other conditions. Interestingly, half of the first-pass
fixations observed in conditions where texts were written against reading
direction came from refixation saccades, while only a third of first-pass fixa-
tions were generated from refixation saccades. Less than 12% of fixations in
reading manipulated text were moving backwards (e.g. regression saccades)
with the lowest observed in iW condition. When reading the manipulated
texts for the first time, readers fixated on words, on average, about 53-137 ms
longer and generated shorter saccades (5-6 character spaces for progressive
saccades and 3 character spaces) than those observed in reading normal text.
Note that mean fixation duration in sL condition is about 50 ms longer than
those reported in the study by White and colleagues [54]. Global summary
statistics from all sessions are presented in Table A1.
On a qualitative level, we compare the resulting saccade length distribu-
tions (Fig. 2) with our hypotheses (Fig. 1b). As expected, the distributions of
foward saccades length during reading manipulated text are generally shifted
to the left of those from control condition, indicating a qualitatively shorter
saccade length generated. Futhermore, more regressive saccades were ob-
served only in reading mirrored-word (mW) and inverted-word (iW) condi-
tions.
11
Figure 2: Distribution of saccade amplitudes across all sessions. Saccades
observed in the control condition (normal reading) are marked with red color.
Dark blue lines represent data from first session of reading manipulated text.
Data from sessions after the two trainings are marked with lighter blue hues.
12
Inverted Words (iW)Scrambled Letters (sL)Mirrored Words (mW)Mirrored Letters (mL)−20−1001020−20−10010200.000.050.100.000.050.10Saccade Length [letters]Relative FrequencySessionControlSession 1Session 2Session 3 Single fixation duration and refixation probability
Results from statistical modeling using linear mixed-effects models for single
fixation duration and refixation probability are summarized in Tables A2-A4.
For specification of linear mixed-effects models, see section 4.
No significant word length effect was observed in single fixation duration
measure for control condition, which in line with the finding reported in study
conducted by Kliegl et al. [14]. Compared to control condition, word-based
single fixation durations observed in experimental conditions are significantly
longer (all t value > 2). For example, on medium words (5-7 characters),
estimated mean single fixation durations in the first experimental sessions
were 48-104 ms longer than in control condition (estimated mean: 238 ms).
In line with the study by Kowler and Anton [19], texts written against reading
direction were more difficult to process. This processing difficulties seemed to
be the property of the corresponding type of text manipulation and were more
obvious after participants have learned how to read the texts. When readers
fixated on medium words only once in the last experimental sessions, the
fixation durations in mW (estimated mean: 309 ms) and and sL (estimated
mean: 326 ms) conditions were shorter compared to those observed in mW
and iW conditions with estimated mean of 570 ms and 455 ms respectively.
Similar to single fixation measures, there was no significant word length
effect observed in first of multiple fixation duration measures for control con-
dition. The mean of first fixation duration in sL condition did differ signifi-
cantly across all sessions compared to those observed in control condition. In
mL condition, the duration of first fixation measures was significantly longer
when readers read the manipulated text for the first time, but did not differ
significantly after trainings. Interestingly, the mean of first fixation duration
was significantly longer in mW and iW conditions compared to control condi-
tion. For example, on medium words, estimated mean durations of the first
fixation in both conditions were about 90 ms longer than those estimated for
control condition. The difference remained significant even after trainings
(estimated means session 3: mW = 465.32; iW = 414.38).
On refixation probability measure, we replicated the word length effect
in both control and experimental conditions (all p < 0.01). Long words were
more likely to be refixated than short words. However, when the texts were
manipulated, readers were more likely to refixated the words. For example,
the estimated refixation probability on medium words in control condition
was 0.16. However, readers in experimental conditions were twice more likely
13
to fixate the words with the same lengths (estimated RFP: mL = 0.42;
sL = 0.65; mW = 0.44; iW = 0.56). Interestingly, after training to read
texts with words written against the reading direction (e.g. mW and iW
conditions), readers were almost always fixated on medium and long words
(estimated RFP above 0.87 on the third session).
Do these manipulation types have something in common or do they gen-
erate different effect on fixation duration and probability measures presented
above? The results from separate models which estimated the different effect
size of manipulation types (see Table A5) confirmed that some manipulation
type generated greater effect than the others. All of the three models showed
similar trends. The effect in mL condition was significantly different from
control condition, but adding the effect from sL conditon to the mean of the
two conditions (control and mL conditions) did not yield a significant gain on
effect size. However, adding the effect from mW condition and iW conditions
yielded significant gains on the effect size. The results from various measures
showed that of all manipulation types tested in the current study, the iW
condition generated the largest effect.
The effects of word lengths on duration of single fixation and first of multi-
ple fixations, as well as on skipping and refixation probabilites are visualized
in Fig. 3.
Landing positions distributions and launch-site effect
A first glance at the resulting distributions for within-word landing positions
indicates differences between normal reading and manipulated texts (Ap-
pendix A3-A4). Except for short words (i.e., word length up to 4 chars), we
found increased leftward shifts across all experimental conditions, where with
mirrored words (mW) and inverted words (iW) showed stronger shifts com-
pared to the other two manipulations (mL, sL). Since shorter saccades were
generated in experimental conditions, we analyses only saccades launched up
to 5 character away from the word beginning, resulting in more overshoots.
Undershoots were typically observed on longer saccades launched furhter
than 7 characters away and more prominent on long words [28, 30, 20]. Based
on Bayesian fits of the distribution (see Methods), we obtained mean landing
positions that were used for further analyses.
In Figure 4a, center-based mean landing sites are plotted against center-
based launch sites for forward saccades and different lengths of the target
words. The slope parameters from landing position functions are plotted for
14
Figure 3: Upper row. Mean fixation duration of single fixation (left) and
first of multiple fixations as the function of word length classes. Lower row.
Skipping (left) and refixation probabilities as the function of word length
classes. Word with length less than 3 characters are grouped in word length
class 3; long words (> 8characters) are grouped in word length class 8. Error
bars represent the standard error of the means. Red line and dots represent
data from reading normal text. The dark blue color represents data from the
first experimental session. The lighter blue hues indicate data from the last
two experimental sessions.
15
llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllInverted Words (iW)Scrambled Letters (sL)Mirrored Words (mW)Mirrored Letters (mL)345678345678200300400500200300400500Word length classMean Fixation Duration [ms]Single FixationllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllInverted Words (iW)Scrambled Letters (sL)Mirrored Words (mW)Mirrored Letters (mL)345678345678200300400500200300400500Word length classMean Fixation Duration [ms]First of Multiple FixationsllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllInverted Words (iW)Scrambled Letters (sL)Mirrored Words (mW)Mirrored Letters (mL)3456783456780.000.250.500.751.000.000.250.500.751.00Word length classSkipping probabilitySkipping Probability on Word LengthllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllInverted Words (iW)Scrambled Letters (sL)Mirrored Words (mW)Mirrored Letters (mL)3456783456780.000.250.500.751.000.000.250.500.751.00Word length classRefixation probabilityRefixation Probability on Word LengthSessions:llllControlSession 1Session 2Session 3 (a)
(b)
Figure 4: Analysis of the launch-site effect. (a) Landing position function:
Mean center-based landing position as a function of center-based launch-
site distance for all experimental conditions. Red line and dots represent
data from normal reading session. Data from the first experimental sessions
are presented in dark blue color. The lighter blue hues represent the last
two experimental sessions. (b) The estimated slope parameter λ (and the
95 % confidence interval) for all experimental conditions as a function of the
experimental session. The red horizontal dashed line represents the estimated
slope for reading normal text (baseline). Results for reading mirrored words
(mW) and inverted words (iW) are presented with magenta and cyan lines,
resp. Green and yellow lines represent estimations for reading texts with
mirrored letters (mL) and scrambled letters (sL).
16
Inverted Words (iW)Scrambled Letters (sL)Mirrored Words (mW)Mirrored Letters (mL)−9−8−7−6−5−4−3−9−8−7−6−5−4−3−10123−10123Launch−site distance to word centerCenter−based mean landing positionSession:ControlSession 1Session 2Session 3Landing Position FunctionControlControlControlControlControlControlControlControlControlControlControlControlControlControlControlControl0.20.30.40.50.6Session 1Session 2Session 3SessionSlopeCondition:mWmLiWsLEffect on Slope of Landing Position Function all sessions in Figure 4b. The estimated slope of (λ0 = 0.37) was observed
in reading normal text. Compared to the value from control condition, both
conditions iW (λ1 = 0.31,λ2 = 0.25,λ3 = 0.23) and mW (λ1 = 0.24, λ2 =
0.26, λ3 = 0.29) generated shallow slopes, which indicates a tendency to
reduced oculomotor control. In contrast, conditions mL (λ1 = 0.44, λ2 =
0.38, λ = 0.37) and sL (λ1 = 0.52, λ2 = 0.48, λ = 0.41) generated greater
slope values (λ) compared to reading normal text, indicating a tendency to
increased oculomotor control. Interestingly, except in the mW condition, the
λ values approach the value of normal reading after trainings.
0) and standard deviation of prior distribution (σ2
In the Bayesian model, the slope parameter λ represents the strength of
oculomotor control and the relation between observational error of the target
T ). In extreme
location (σ2
0 → 0, the eyes always land on the
cases, a value of slope near to 0, where σ2
target (i.e., the word center) regardless of where the saccade started. In the
other extreme case, the a slope value of λ → 1 indicates the absence of a
target selection process in saccade planning, so that the eyes generate random
constant saccade lengths (from a uniform distribution). Therefore, our data
show that saccades from reading text composed of words with reversed letter
sequences (i.e., mW and iW) tend to land precisely on the target location
(word center) on average, while reading manipulated text with normal letter
order (i.e., mL and sL) tend to generate saccades similar length.
Comparing the results with our hypotheses, there is no dramatic effect of
a shift of the mean landing position towards the word ends. Therefore, we
ran a post-hoc analyses for single and two-fixation cases in the next section.
Landing positions for single vs. two-fixation cases
Single fixation cases. Only cases where exactly one fixation land on a word
were considered. Relative frequency of fixation landing position were cal-
culated based on word length. To obtain estimates for the mean µSF and
standard deviation σSF for the landing position distribution of each word
length, a grid search method (in steps of 0.01) with a minimum-χ2 criterion
was applied. Analysis of variance was conducted to statistically compare the
manipulation effects on mean (µSF ) and standard deviation (σSF ) of landing
position distributions.
No significant difference was observed for the mean (µSF ) of the landing
position distribution of single fixation cases between control and experimen-
tal conditions (F [1, 30] = 2.364, p = 0.135). The main effect on standard
17
deviation (σSF ) of within-word landing position distribution of single fixation
cases yielded an F -ratio of F [1, 30] = 32.7,p < 0.000, indicating a significant
difference between reading normal text (M=2.95,SD=0.84) and manipulated
text (M=1.51,SD=0.55). When words were fixated only once, readers' mean
landing position did not change across the manipulation types. However, the
precision of landing on selected target increased as the text display deviated
from normal presentation.
Among our key results is that the two conditions where the word begin-
nings were located at the end of the manipulated word strings (i.e., conditions
mW and iW) produced only a slight shift of the average landing position.
This finding turned out to be robust and remained observable even after
trainings. Figure 5 presents data for single fixations. The finding did not
support our hypothesis that readers targeted the second half of the word
strings during reading mirrored-word (mW) and inverted words (iW). On a
qualitative level, there is little adaptivity of the oculomotor system.
Since in our hypothesis, an initial saccade towards the second half of the
word would require an additional refixation in the first half of the word, we
ran a post-hoc analysis of all two-fixation cases. The corresponding plot
in Figure 6 presents data for the initial landing position (first saccade into
the word) from all cases, where the word was fixated exactly two times. In
contrast to typical OVP effect on refixation probability, plotting the initial
of two fixations give us a better understanding of word targeting in saccade
planning.
Two-fixation cases. For this analysis, we considered cases when exactly
two fixations on a word were observed. The landing position distribution of
the initial fixation was fitted to a quadratic polynomial, i.e.,
y = A + B · L(x − C)2 ,
(5)
where x denotes the initial landing position and y is the relative frequency of
the fixation. The parameter A represents the actual relative frequency of the
initial landing position. The parameter B is the slope of the parabolic curve,
representing the within-word maximum or minimum relative frequency of the
landing position. The parameter C reflects within-word position where the
relative frequency was at the minimum or maximum, depending on the value
of parameter B. In general, the value of the parameter B is assumed to be
positive, resulting in a distribution of landing position qualitatively similar to
the optimal viewing position (OVP) curve but with different interpretation:
18
Figure 5: Within-word landing position distributions for single-fixation
cases. Columns relate to different word lenghts. Rows refer to experimental
manipulations. Different colors indicate control condition (red) and experi-
mental conditions (blue hues).
19
4−letterwords5−letterwords6−letterwords7−letterwordsMirrored Words (mW)Mirrored Letters (mL)Inverted Words (iW)Scrambled Letters (sL)02460246024602460.00.10.20.30.40.00.10.20.30.40.00.10.20.30.40.00.10.20.30.4Landing PositionRelative Frequency Session: ControlSession 1Session 2Session 3Single Fixation Cases when a word is fixated exactly twice, the initial fixations tend to on word
edges more often than on the word center. The estimation of the three
parameters representing the characteristics of the landing position of the
first of two fixation cases was conducted based on a maximum likelihood
method using the bbmle package [2] for R studio. The estimated parameters
are summarized in Table A6-A10.
In two-fixation cases under normal reading conditions, readers tend to ini-
tially land their eyes close to the word beginning with a subsequent secondary
fixation further into the word (most refixations are forward directed). Sim-
ilar patterns were observed in the condition where manipulated texts were
written in typical direction from right to left (mL and sL). However, what
is most remarkable for our study, in the two conditions (mW and iW) with
reversed letter sequences, the initial landing positions deviated from those of
normal reading (for detailed estimated parameters, see Table A7 and A9).
When a word was fixated more than once, the first saccade landed mostly
behind the location of the word center in the second half of the word strings;
thus, there is clear top-down adaptivity in the two conditions where letter
sequence is reversed from right to left. The effect was stronger for longer
words (see Fig. 6).
These findings support our hypothesis that readers first target the right
part of the word strings and then moved the eyes backward, following the
writing direction. Topical subheadings are allowed.
3 Discussion
Natural text reading requires efficient coordination of cognitive and oculo-
motor control processes. Cognitive principles are essential for the selection of
an upcoming target word, however, it is the oculomotor system that provides
the machinery to move the gaze to the region of the identified target word.
In the current study, we set out to investigate whether ongoing cognition is
able to overwrite default oculomotor control when the reader is confronted
with manipulated (mirrored-reversed, inverted, and scambled) text layouts.
On a global level, changing the display and positions of letters in a word
modulates eye-movement statistics. Since mirrored-reversed and other ma-
nipulated words are more difficult to process, a general tendency of longer
average fixation durations and shorter mean saccade amplitudes is observed.
These findings are in line with previous studies [19, 42]. Interestingly, sim-
20
Figure 6: Within-word landing position distribution for the initial landing
position (first saccade) in two-fixation cases.
21
4−letter Words5−letter Words6−letter Words7−letter WordsMirrored Words (mW)Mirrored Letters (mL)Inverted Words (iW)Scrambled Letters (sL)02460246024602460.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.5Initial Landing PositionRelative Frequency Session:ControlSession 1Session 2Session 3Saccade 1 of Two−Fixation Cases ilar with findings from Kowler and Anton [19] and Kolers and Perkins [15],
orthographical manipulations from our current study can be grouped into
two categories based on their writing directions. Difficulties increased dra-
matically, when readers read texts written against the reading direction as in
mirrored words (mW) and inverted words (iW) conditions, where the letter-
sequence of words is written from right to left within the displayed string due
to mirroring or inversion. Our interpretation is supported by the fact that the
observed deviations remained stable even after trainings. In contrast, during
the manipulations that kept the letter-sequence in the reading direction (i.e.,
from left to right) in mirrored-letter word (mL), the patterns observed in
eye movement measures (e.g., word-length effect) remained similar to those
of normal reading and, after training, approached the behavior observed in
normal reading. Finally, in our fourth experimental condition, we analyzed
reading text composed of scrambled-letter (sL) words. Since letters in a word
were randomly scrambled, there was no systematic way that the oculomotor
behavior could adapt to process the new text layout.
One of our motivations to run the current study was to obtain a detailed
picture on within-word landing positions as the most important signature for
oculomotor control. Mean landing position distribution of single fixations in
reading manipulated texts does not significantly differ from reading normal
text, although there was a slight shift observed in mirrored words (mW) and
inverted words (iW) conditions. However, readers tend to increase their pre-
cision in landing on word center when texts presentation was manipulated.
The finding supports the idea that word center serves as saccade target lo-
cation [5]. Analysis of mean landing position of forward saccades based
on different launch sites demonstrated a general left shifts in experimental
conditions compared to those observed in control condition. Especially in
conditions where words are written against reading direction such as in mW
and iW conditions, there is no clear preference for the eyes to initially land in
the second half of the word strings (where the beginning of a word is located
in this manipulation). This is contrary to our hypothesis. Interestingly, the
peaks of the landing position in all manipulation types were shifted toward
word centers, not necessaryly toward the beginning of word strings, compared
to the peaks observed in reading normal text. In some rare undershoot cases
in our data (e.g. launch site: -4, word length: 7 in iW and mW conditions),
the peaks in experimental condition shift rightward of the peak observed in
reading normal text. Furthermore, the variance of the initial landing position
distribution was much smaller, meaning that the eyes landed more precisely
22
on the word center, which served as the theoretical target position. Given
the precision of the landing position distributions and the highest likelihood
to land on word center, we speculate that the leftward shifts were actually
moving toward the word centers, not word beginning. Regarding the general
improved precision of landing position distributions, a possible explanation
is that with longer average fixation durations under our manipulations, the
oculomotor system had more time to prepare the next saccade, which could
result in a reduced saccadic error [28]. An alternative explanation is that
unusual presentation of text was more salient and popped up parafoveally to
enable precise saccadic targeting [12].
Additionally, modulation of manipulation types on launch-site effect was
observed. In conditions where texts were written against reading direction
(from right to left) we observed a reduced launch-site effect, while an in-
creased launch-site effect was observed in conditions where texts were writ-
ten normally. However, we are not sure if the difference effects are caused
by the change in reading direction or increased processing difficulties. Nev-
ertheless, the finding could also be explained in the theoretical framework
of a Bayesian model of sensorimotor integration that we applied to reading
[5, 21]. According to Bayes' rule, the observed landing position distribu-
tion is the posterior distribution in a sensorimotor transform based on prior
knowledge of typical target positions. If we assume that the prior knowledge
is uninformative under unfamiliar text layout, then the posterior distribu-
tion would shift toward the likelihood distribution, which is assumed to be
unbiased with respect to the target word center. Therefore, in the Bayesian
model, we could have expected a more centered landing-position distribution
if we assumed that the experimental manipulations induce less prior knowl-
edge on the possible target positions. Our finding demonstrated that by
simply changing the letter-sequence information of words, we could obtain
estimations for the slope parameter λ other than the typical value of 0.5.
As a consequence, reading models development should aim for integrating a
process-oriented model in generating saccade lengths.
The most striking finding in the current study is the effect of text-display
manipulation on initial landing position of two-fixation cases. When a word
shows reversed letter ordering (hence a change in their spatial information)
as a consequence of the experimental manipulation, we observed a clear shift
into the second half of the word string when two fixations were generated.
This finding demonstrates that the oculomotor system is able to adapt to
display changes. Given that our hypothesis requires more than two fixations,
23
analysing the initial landing positions in refixation cases is more reasonable
to test the hypothesis.
Furthermore, our study also demonstrates that the typical usage of mean
fixation position as a dependent variable to characterize oculomotor control
is not specific enough to characterize control processes underlying reading.
Since reading is a complex process, data delivered are usually complicated
and require various analysis methods and inferences. Further systematic
analyses and advanced mathematical modeling are required to investigate
the dynamical processes underlying principles of oculomotor control and their
interaction with ongoing cognition during reading.
4 Methods
Participants
A group of 37 participants (27 females, 10 males), aged between 16 and
39 years, received a total of EUR 70 for taking part in four 45-minute lab
sessions and six 30-minute training sessions. They were all naive with respect
to the purpose of the experiment. Participants reported normal or corrected-
to-normal vision and declared their informed consent, p. The experiment
conformed to the Declaration of Helsinki. Informed consent was obtained for
experimentation by all participants. According to the standards of Deutsche
Forschungsgemeinschaft (German Research Foundation) and German Society
for Psychological Research, ethics committee approval was not required for
this study.
Apparatus, Material & Procedure
Participants were assigned to four different groups based on four types of text
manipulation, namely mirrored-word text (mW), mirrored-letter text (mL),
inverted-word text (iW), and scrambled-word text (sL). Each participant did
a total of four lab sessions and six training sessions at home via web-based
interface.
For lab sessions, participants read an excerpt from the German version
of the novel "The Adventure of the Empty House" [? ]. They were seated
at a viewing distance of 70cm in front of a 19-inch Mitsubishi Diamond
Pro 2070 Monitor (screen resolution 1,280×1,024 pixels, refresh rate 100 Hz)
24
with the head supported by a chin rest. The stimuli (Courier font, size 18,
black) were presented on vertical center line of the computer display with
gray background color. Eye movements from both eyes were recorded using
an EyeLink 1000 System (SR Research, Osgoode/Ontario, Canada) with a
sampling rate of 1000 Hz and spatial resolution better than 0.01◦. At the
end of the session, participants had to answer three questions related to the
text they have just read. A maximum of 600 lines of text were presented
over all lab sessions.
For training sessions, participants read an excerpt from the novel "Small
World" [48] by visiting a website created for the experiment1. After logging
in, participants could read the manipulated text, which was presented as a
line of max. 85 characters at the center of the screen, without time limit.
When finished, they could go to the next line by clicking the right arrow or
return to the previous page by clicking the left arrow. One training session
should last at least 30 minutes. After logging out, participants received three
questions via E-mail, which should be answered as soon as possible. No limit
of the number of lines presented in training sessions was enforced.
The complete procedures of the experiment went as follows: During the
first lab session, participants read normal text (a total of 150 lines of maxi-
mum 85 characters). After a two-hour break, participants read manipulated
text in the second session, which lasted up to 45 minutes or when 150 lines
were read. At home, participants were required to read manipulated text
for two 30-minute sessions on the website before the third lab session. At
the third lab session, participants continue reading manipulated text from
where they left off at the previous session. Participants should conduct four
30-minute training sessions before taking part in the last lab session.
Data Preparation
Data containing blinks were discarded from the analysis. Saccades and fix-
ations were detected using a velocity-based algorithm developed by Engbert
et al. [9]. As a result, a total of 380, 292 fixations were detected. From this
data set we excluded fixations based on the following criteria (i) fixations
on the first and last words of a sentence as well as the first and the last of
participant's fixation sequence, (ii) fixations with duration less than 20 ms
or longer than 1000 ms, fixations landing outside the text rectangle, and sac-
1Using ShinyApps by RStudio accessible via http://www.shinyapps.io
25
cades shorter than one character space (12 pixels) or longer than 25 character
spaces were removed from the analysis. Trials containing fixation duration
longer than 2000 ms and less than three fixation points after filtering were ex-
cluded from analyis. The remaining 236, 937 fixations are the valid fixations
(see Table A1).
Summary statistics
In a global analysis, we computed statistics of fixation durations and saccade
lengths of all valid fixations. For local processing, fixation duration and
probability were grouped based on word length. Word lengths were grouped
into three classes: short (≤ 4 characters), medium (5-7 characters) and long
(> 7 characters) for further statistical analysis.
Fixation durations. Visual information processing in reading is marked
by the time spent on fixating words. Statistical analysis on single fixation
For complimentary analysis on possible differences in processing All cases in
which a word received exactly one fixation generates single fixation duration
(SFD). In the cases in which a word received more than one fixations, the
duration of the first and the second of those fixations were calculated for
evaluating the first of multiple fixations (FMD) and second of multiple fixa-
tions (SMD). All of those fixation duration measures were conducted on the
first-pass reading data set. The sum of all fixations on a word, regardless the
saccadic types, generated the total viewing time (TVT). For further statisti-
cal analysis, the fixation duration measures on each word were transformed
into their logarithmic values.
Fixation probability. As measures of fixation probability we computed
skipping probability (SKP), single fixation probability (SFP), and refixation
probability (RFP) on first-pass reading data. Additionally, we calculated
regression probability (RGP). For further statistical analysis, each word was
assigned a logical value of 1 if it is skipped (SKP), fixated only once (SFP),
fixated more than once (RFP) or a target of regressive saccades (RGP).
Otherwise, a logical value of 0 was assigned to the word.
To capture the changes of the manipulation effects on word processing
difficulties across sessions,
linear-mixed model analysis (3x4 factors) was
conducted on the dependent variables single fixation duration (SFD) and
refixation probability (RFP) as a function fixed word lenght effect and ran-
dom effects generated by participants and sentences. The model was con-
trasted against the word length effect during normal reading (sess = 0). We
26
used the lme4 package for R-Language of Statistical Computing [1] for esti-
mating fixed (word length classes) and random (participant and sentence)
coefficients. Specifically, each fixation duration and probability measure
is modeled as a function of fixed word length effect (WLC across sessions
(sess), with a fully parameterized variance-covariance matrix for participant
(1 + WLC + sesspID) assuming participants may generate different slopes for
each word length and sentence/line number (1sID) with the assumption
that the intercept for each sentence do not vary accross sessions and word
lengths. These model used the default treatment contrast. For example, the
log-transformed measure of single fixation duration (SFD) is modeled as
lmer(log(SFD) = WLC ∗ sess + (1 + WLC + sesspID) + (1sID)),
(6)
while refixation probability (RFP) are modeled as
glmer(RFP = WLC ∗ sess + (1 + WLC + sesspID) + (1sID)),
(7)
because the glmer() function allows the statistical analysis of binary out-
come. The models were estimated based on restricted maximum likelihood
(REML). Furthermore, a 3x5-factor model using helmert constrast was con-
ducted separately for each dependent measure with word lengths and manip-
ulation types (type) as fixed main effects to investigate the different effect
sizes among different manipulation types. The predictor type for reading
normal text was used as baseline (type = 0).
Initial Landing Position
Within-word landing-position distributions in reading are broad and over-
lap with neighboring words. As a consequence, observations of word-based
landing positions in a reading experiment are truncated at word boundaries
[28, 7]. In order to derive estimates of the means and standard deviations of
the landing-position distribution, truncated normal curves for the distribu-
tions of within-word fixation positions were fitted using Bayesian parameter
inference [23] provided by the package rjags [35] in the R environment.
Here we considered only cases where readers made forward saccades. Fix-
ation data for each experimental condition and each experimental session
were grouped based on word length and launch-site distance leading to 16
different word-length and launch-site specific data subsamples per condition
and reading session. For each data subset Si, we estimated a two-dimensional
27
posterior distribution over the parameters mean µ and standard deviation σ
of the underlying Gaussian landing-position distribution according to
p(Siµ, 1/σ2)p(µ, 1/σ2)
(cid:82)(cid:82) p(Siµ, 1/σ2)p(µ, 1/σ2) dµd1/σ2 .
p(µ, 1/σ2Si) =
(8)
We assumed that observations of landing positions are generated by a normal-
density likelihood function p(Siµ, 1/σ2) with mean µ and precision τ = 1/σ2.
Furthermore, with p(µ, 1/σ2) we specified a normally distributed prior on
the mean µ with mean M and precision T and a prior on τ distributed as a
gamma density distribution with shape parameter A and rate B (see [23]).
The parameters M, T and A, B for the prior over µ and τ, resp., were derived
from the data in the following steps: For the control condition, landing-
position distributions for each word length and launch-site distance were
independently fitted by a truncated Gaussian function and the parameters
of these fits were used to estimate the parameters of the empirical prior
distribution. For reading manipulated text we also used the parameters of
the prior of the control condition for estimating landing distributions of the
first reading session and updated the prior for the later sessions systematically
based on the results of the previous reading session.
5 References
[1] Douglas Bates, Martin Mächler, Ben Bolker, and Steve Walker. Fitting
linear mixed-effects models using lme4. Journal of Statistical Software,
67(1):1 -- 48, 2015.
[2] Ben Bolker. bbmle: Tools for General Maximum Likelihood Estimation,
2017. R package version 1.0.20.
[3] Christian Coëffé and J. Kevin O'Regan. Reducing the influence of non-
target stimuli on saccade accuracy: Predictability and latency effects.
Vision Research, 27(2):227 -- 240, 1987.
[4] H. Deuble, W. Wolf, and G. Hauske. The evaluation of the oculomotor
error signal. In Alastair G. Gale and Frank Johnson, editors, Theoretical
and Applied Aspects of Eye Movement Research, volume 22 of Advances
in Psychology, pages 55 -- 62. North-Holland, 1984.
28
[5] Ralf Engbert and André Krügel. Readers use bayesian estimation for
eye movement control. Psychological Science, 21(3):366 -- 371, 2010.
[6] Ralf Engbert, André Longtin, and Reinhold Kliegl. A dynamical model
of saccade generation in reading based on spatially distributed lexical
processing. Vision Research, 42(5):621 -- 636, 2002.
[7] Ralf Engbert and Antje Nuthmann. Self-consistent estimation of mislo-
cated fixations during reading. PLOS ONE, 3(2):1 -- 6, 02 2008.
[8] Ralf Engbert, Antje Nuthmann, Eike M. Richter, and Reinhold Kliegl.
Swift: A dynamical model of saccade generation during reading. Psy-
chological Review, 112(4):777 -- 813, 2005.
[9] Ralf Engbert, Petra Sinn, Konstantin Margenthaler, and Hans Truken-
bord. Microsaccade toolbox.
[10] J. M. Findlay. Global visual proceeding for saccadic eye movements.
Vision Research, 22(8):1033 -- 1045, 1982.
[11] John M Findlay and Iain D Gilchrist. Active vision: The psychology of
looking and seeing. Oxford University Press, 2003.
[12] Jukka Hyönä. Do irregular letter combinations attract readers' atten-
tion? evidence from fixation locations in words. Journal of Experimental
Psychology: Human Perception and Performance, 21(1):68 -- 81, Febru-
ary 1995.
[13] Rebecca L. Johnson and Emma L. Starr. The preferred viewing lo-
cation in top-to-bottom sentence reading. The Quarterly Journal of
Experimental Psychology, 71(1):220 -- 228, 2017.
[14] Reinhold Kliegl, Ellen Grabner, Rolfs Martin, and Ralf Engbert. Length,
frequency, and predictability effects of words on eye movements in read-
ing. European Journal of Cognitive Psychology, 16(1 -- 2):262 -- 284, 2004.
[15] P. A. Kolers and D. N. Perkins. Spatial and ordinal components of form
perception and literacy. Cognitive Psychology, 7(2):228 -- 267, 1975.
[16] Konrad P. Körding. Decision theory: What "should" the nervous system
do? Science, 318(5850):606 -- 610, 2007.
29
[17] Konrad P. Körding and Daniel M. Wolpert. Bayesian integration in
sensorimotor learning. Nature, 427(6971):244 -- 247, 2004.
[18] Konrad P. Körding and Daniel M. Wolpert. Bayesian decision theory
in sensorimotor control. Trends in Cognitive Sciences, 10(7):319 -- 326,
2006.
[19] Eileen Kowler and Stanley Anton. Reading twisted text: Implications
for the role of saccades. Vision Research, 27(1):45 -- 60, 1987.
[20] André Krügel and Ralf Engbert. On the launch-site effect for skipped
words during reading. Vision Research, 50(16):1532 -- 1539, 2010.
[21] André Krügel and Ralf Engbert. A model of saccadic landing positions
in reading under the influence of sensory noise. Visual Cognition, 22(3-
4):37 -- 41, 2014.
[22] André Krügel, Françoise Vitu, and Ralf Engbert. Fixation positions af-
ter skipping saccades: A single space makes a large difference. Attention,
Perception, & Psychophysics, 74(8):1556 -- 1561, 2012.
[23] John K. Kruschke. Doing Bayesian data analysis: A tutorial with R,
JAGS, and Stan. Academic Press, 2014.
[24] Frederic Lavigne, Francoise Vitu, and Gery d'Ydewalle. The infuence of
semantic context on initial eye landing sites in words. Acta Psychologia,
104(191-214), 2000.
[25] Gordon E. Legge, Thomas A. Hooven, Timothy S. Klitz, J. Stephen
Mansfield, and Bosco S. Tjan. Mr. chips 2002: new insights from an
ideal-observer model of reading. Vision Research, 42(18):2219 -- 2234,
2002.
[26] Gordon E. Legge, Timothy S. Klitz, and Bosco S. Tjan. Mr. chips: An
ideal-observer model of reading. Psychological Review, 104(3):524 -- 553,
1997.
[27] Steven G Luke and John M Henderson. Oculomotor and cognitive con-
trol of eye movements in reading: evidence from mindless reading. At-
tention, Perception, & Psychophysics, 75(6):1230 -- 1242, 2013.
30
[28] G. W. McConkie, P. W. Kerr, M.D. Reddix, and D Zola. Eye movement
control during reading: I. the localisation of initial eye fixations in words.
Visual Research, 28(10):1107 -- 1118, 1988.
[29] Antje Nuthmann. The "where" and "when" of eye fixations in reading.
Unpublished Doctoral Dissertation, University of Potsdam, Germany
(online available
from ttp://nbn-resolving.de/urn:nbn:de:kobv:517-
opus-7931), 2006.
[30] Antje Nuthmann, Ralf Engbert, and Reinhold Kliegl. Mislocated fix-
ations during reading and the inverted optimal viewing position effect.
Vision Research, 45(17):2201 -- 2217, 2005.
[31] Antje Nuthmann, Ralf Engbert, and Reinhold Kliegl. The iovp ef-
fect in mindless reading: Experiment and modeling. Visual Research,
47(7):990 -- 1002, 2007.
[32] J. Kevin O'Regan. Eye movements and reading. In E. Kowler (Ed.),
Eye movements and their role in visual and cognitive processes (pp. 395-
453). Elsevier, Amsterdam, 1990.
[33] J. Kevin O'Regan and A. Levy-Schoen. Eye movement strategy and
tactics in word recognition and reading. In M.Coltheart (Ed.), Attention
and performance XII: The psychology of reading(pp. 363-383). Erlbaum,
Hillsdale, NJ, 1987.
[34] J. Kevin O'Regan, A. Levy-Schoen, J. Pynte, and B. Brugaillere. Con-
venient fixation location within isolated words of different length and
structure. Journal of Experimental Psychology: Human Perception and
Performance, 10(2):250257, 1984.
[35] Martyn Plummer.
rjags: Bayesian Graphical Models using MCMC,
2016. R package version 4-6.
[36] E. C. Poulton. Human manual control.
In Supplement 2: Handbook
of Physiology, The Nervous System, Motor Control, pages 1337 -- 1389.
Wiley-Blackwell, 1981.
[37] Ralph Radach. Blickbewegungen beim Lesen: Psychologische Aspekte
der Determination von Fixationspositionen. Waxmann Verlag, 1996.
31
[38] Ralph Radach and G. W. McConkie. Determinants of fixation positions
in words during reading. Eye Guidance in Reading and Scene Perception,
pages 77 -- 100, 1998.
[39] Keith Rayner. Eye guidance in reading: Fixation locations within words.
Perception, 8(1):21 -- 30, 1979.
[40] Keith Rayner. Eye movements and attention during reading, scene per-
ception, and visual search. The Quarterly Journal of Experimental Psy-
chology, 62(8):1457 -- 1506, 2009.
[41] Keith Rayner, K. S. Binder, J. Ashby, and Alexander Pollatsek. Eye
movement control in reading: word predictability has little influence on
initial landing positions in words. Vision Research, 41(7):943 -- 54, 2001.
[42] Keith Rayner, Sarah J. White, Rebecca L. Johnson, and Simon P. Liv-
ersedge. Raeding wrods with jubmled lettres: There is a cost. Psycho-
logical Science, 17(3):192 -- 193, 2006.
[43] E. D. Reichle, Alexander Pollatsek, and Keith Rayner. E -- z reader:
A cognitive-control, serial-attention model of eye-movement behavior
during reading. Cognitive Systems Research, 7(1):4 -- 22, 2006.
[44] Erik D. Reichle, Alexander Pollatsek, Donald L. Fischer, and Keith
Rayner. Toward a model of eye movement control in reading. Psycho-
logical Review, 105(1):125 -- 157, 1998.
[45] Erik D. Reichle, Keith Rayner, and Alexander Pollatsek. Eye movement
control in reading: accounting for initial fixation locations and refixa-
tions within the e-z reader model. Vision Research, 39(26):4403 -- 4411,
1999.
[46] Erik D. Reichle, Tessa Warren, and Kerry McConnell. Using e-z reader
to model the effects of higher level language processing on eye movements
during reading. Psychonomic Bulletin & Review, 16(1):1 -- 21, 2009.
[47] Daniel Schad and Ralf Engbert. The zoom lens of attention: Simulat-
ing shuffled versus normal text reading using the swift model. Visual
Cognition, 20(4-5):391 -- 421, 04 2012.
[48] Martin Suter. Small World. Diogenes, Zürich, 1997.
32
[49] Françoise Vitu. The existence of a center of gravity effect during reading.
Vision Research, 31(7-8):1289 -- 1313, 1991.
[50] Françoise Vitu. About the global effect and the critical role of retinal
eccentricity: Implications for eye movements in reading. Journal of Eye
Movement Research, 2(3):1 -- 18, 2008.
[51] Françoise Vitu, D. Lancelin, A. Jean, and F. Farioli. Influence of foveal
distractors on saccadic eye movements: A dead zone for the global effect.
Vision Research, 46(28):4684 -- 4708, 2006.
[52] Francoise Vitu, George W. McConkie, Paul Kerr, and J. Kevin O'Regan.
Fixation location effects on fixation durations during reading: an in-
verted optimal viewing position effect. Vision Research, 41(25-26):3513 --
3533, 2001.
[53] Francoise Vitu, J. Kevin O'Regan, and M. Mittau. Optimal landing
position in reading isolated words and continuous text. Perception &
Psychophysics, 47(6):583 -- 600, 1990.
[54] Sarah J. White, Rebecca L. Johnson, Simon P. Liversedge, and Keith
Rayner. Eye movements when reading transposed text: The importance
of word-beginning letters. Journal of Experimental Psychology: Human
Perception and Performance, 34(5):1261 -- 1276, 2008.
[55] Sarah J. White and Simon P. Liversedge. Orthographic familiarity in-
fluences initial eye fixation positions in reading. European Journal of
Cognitive Psychology, 16(1-2):52 -- 78, 2004.
[56] Sarah J. White and Simon P. Liversedge. Foveal processing does not
modulate non-foveal orthographic influences on fixation positions. Vi-
sion Research, 46(3):426 -- 437, 2006.
[57] Sarah J. White and Simon P. Liversedge. Linguistic and nonlinguistik
influences on the eyes' landing positions during reading. The Quarterly
Journal of Experimental Psychology, 59(4):760 -- 782, 2006.
[58] Daniel M. Wolpert and Zoubin Ghahramani. Bayes rule in perception,
action and cognition. In The Oxford Companion to the Mind, pages 1 -- 4.
Oxford University Press, 2005.
33
6 Acknowledgements
This work was supported by Deutsche Forschungsgemeinschaft (grant EN
471/15-1 to R. E.).
7 Author contributions statement
R.E. and A.K. developed the study concept and the study design. Testing
and data collection were performed by J.C.. J.C. and A.K. performed the
data analysis and all authors contributed to the interpretation. J.C. and A.K.
drafted the manuscript, and R.E. provided critical revisions. All authors
approved the final version of the manuscript for submission.
8 Additional information
Competing interests The authors declare no competing interests.
A Appendix
34
)
L
s
(
s
r
e
t
t
e
L
d
e
l
b
m
a
r
c
S
)
W
i
(
s
d
r
o
W
d
e
t
r
e
v
n
I
.
s
c
i
t
s
i
t
a
t
s
y
r
a
m
m
u
s
l
a
b
o
l
G
:
1
A
e
l
b
a
T
)
L
m
(
s
r
e
t
t
e
L
d
e
r
o
r
r
i
M
)
W
m
(
s
d
r
o
W
d
e
r
o
r
r
i
M
l
o
r
t
n
o
C
n
o
i
t
i
d
n
o
C
3
2
6
0
4
2
2
8
4
6
6
2
1
9
4
5
5
2
3
8
2
7
5
2
2
4
0
1
4
2
1
1
3
5
2
2
3
6
4
2
2
2
2
5
0
2
4
2
1
7
7
6
6
2
3
7
5
9
6
2
2
0
0
5
7
2
1
8
9
5
1
2
0
7
8
4
2
8
N
7
9
5
0
1
4
0
.
4
4
5
6
4
3
1
6
9
.
5
5
6
3
8
8
2
6
.
5
6
3
4
8
3
9
4
.
3
4
7
0
8
1
5
4
.
0
2
0
7
1
3
8
8
.
5
3
6
6
9
3
9
.
0
1
4
0
3
0
5
3
1
1
9
5
.
2
4
8
9
2
5
1
1
4
.
7
5
4
3
0
0
1
9
5
.
5
6
9
1
2
4
5
0
.
2
4
9
9
9
1
2
9
.
9
1
5
0
8
3
2
9
.
7
3
5
8
9
2
8
.
9
0
0
3
1
6
3
1
1
7
4
.
4
4
8
8
1
4
1
3
5
.
5
5
9
6
2
0
1
8
3
.
2
7
7
8
8
4
9
5
.
7
4
3
7
9
1
1
2
.
9
1
0
0
4
3
1
1
.
3
3
2
9
9
6
6
.
9
6
0
3
3
3
9
0
1
9
4
.
2
4
5
9
7
4
1
1
5
.
7
5
2
1
9
1
1
1
5
.
0
8
3
6
3
5
2
0
.
5
4
4
7
7
5
.
6
2
7
7
5
6
4
.
8
4
9
4
6
5
4
.
5
9
2
3
6
6
7
0
1
6
6
.
4
4
8
3
3
3
1
4
3
.
5
5
8
2
4
0
1
8
1
.
8
7
3
5
6
4
2
6
.
4
4
4
7
6
6
4
.
6
6
9
0
5
7
8
.
8
4
4
8
5
6
.
5
4
4
3
0
4
3
0
1
9
8
.
5
4
1
9
1
2
1
1
1
.
4
5
9
7
4
9
5
7
.
7
7
0
5
0
4
3
7
.
2
4
1
3
4
5
5
.
4
9
8
9
4
3
6
.
2
5
9
4
5
9
7
.
5
3
8
3
1
7
2
1
1
7
6
.
0
5
5
7
9
0
1
3
3
.
9
4
9
0
8
8
6
2
.
0
8
9
0
9
3
8
3
.
4
4
4
6
9
1
3
.
2
2
0
3
9
2
6
2
.
3
3
8
3
6
4
2
.
7
6
6
2
9
3
6
1
1
9
0
.
8
4
6
6
5
2
1
1
9
.
1
5
1
6
0
0
1
7
0
.
0
8
9
7
3
4
2
5
.
3
4
3
9
1
2
8
.
1
2
6
8
4
3
5
6
.
4
3
5
9
6
1
9
.
6
7
7
2
3
4
1
2
1
2
5
.
5
4
4
3
5
4
1
8
4
.
4
5
0
0
2
1
1
6
0
.
7
7
6
3
6
5
2
3
.
0
5
6
7
4
1
8
1
.
3
1
2
8
0
4
5
4
.
6
3
4
3
7
5
5
.
6
9
9
2
1
9
6
1
1
7
3
.
3
4
6
6
2
5
1
3
6
.
6
5
0
2
6
1
1
2
1
.
6
7
5
2
7
5
7
2
.
9
4
1
3
9
1
0
.
8
7
5
9
4
6
6
.
2
4
4
5
8
5
3
.
7
7
3
3
1
5
2
2
1
5
5
.
4
4
9
4
2
5
1
5
4
.
5
5
6
6
1
1
1
2
2
.
3
7
6
8
3
5
4
2
.
8
4
4
3
8
7
4
.
7
3
3
9
4
8
1
.
4
4
8
0
8
4
2
.
7
1
5
3
1
5
3
0
1
3
9
.
7
4
7
4
2
1
1
7
0
.
2
5
5
2
9
6
7
5
.
1
6
0
6
9
2
4
7
.
2
4
6
1
5
5
4
.
7
6
3
4
3
2
6
.
9
4
3
9
6
1
0
.
0
1
4
5
3
7
7
3
0
4
5
9
.
8
4
0
1
1
2
4
5
0
.
1
5
3
4
3
2
3
1
8
.
6
7
6
5
9
4
1
4
2
.
6
4
3
6
9
9
8
.
0
3
7
5
3
7
5
7
.
2
2
2
9
6
3
2
4
.
1
1
5
4
2
8
3
.
6
5
1
.
6
8
1
.
6
1
8
.
5
5
5
.
5
5
2
.
5
6
8
.
6
6
5
.
6
3
5
.
5
1
8
.
5
8
5
.
5
9
1
.
5
2
8
.
7
N
%
N
%
N
%
N
%
N
%
N
%
N
%
n
a
e
m
]
s
m
[
n
a
e
m
]
r
a
h
c
[
5
2
.
4
5
9
.
3
6
6
.
4
9
7
.
2
7
5
.
2
3
6
.
2
7
6
.
2
8
8
.
2
1
0
.
3
3
0
.
3
7
9
.
2
2
8
.
2
3
2
.
4
n
a
e
m
]
r
a
h
c
[
d
e
t
c
e
t
e
D
s
n
o
i
t
a
x
fi
n
o
i
s
s
e
S
d
e
d
r
a
c
s
i
D
n
o
i
t
a
x
fi
s
s
a
p
-
t
s
r
i
F
n
o
i
t
a
x
fi
d
i
l
a
V
d
r
a
w
r
o
F
s
e
d
a
c
c
a
s
i
g
n
p
p
i
k
S
s
e
d
a
c
c
a
s
n
o
i
t
a
x
fi
s
n
o
i
t
a
x
fi
e
R
s
n
o
i
s
s
e
r
g
e
R
n
o
i
t
a
r
u
d
n
o
i
t
a
x
fi
d
r
a
w
r
o
f
e
d
a
c
c
a
s
e
d
u
t
i
l
p
m
a
d
r
a
w
k
c
a
b
e
d
a
c
c
a
s
e
d
u
t
i
l
p
m
a
35
Table A2: Results from Linear Mixed-Effects: Single Fixation Duration
Condition
Mirrored Letters (mL)
Mirrored Words (mW)
Inverted Words (iW)
Scrambled Letters (sL)
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
b
5.47
0.02
0.18
0.10
0.09
0.06
0.07
0.06
5.44
0.02
0.31
0.31
0.29
-0.14
-0.07
-0.01
5.48
0.01
0.36
0.27
0.22
-0.21
-0.10
-0.06
5.46
0.01
0.11
0.11
0.10
0.10
0.12
0.09
SE
0.04
0.02
0.02
0.02
0.02
0.01
0.01
0.01
0.04
0.02
0.07
0.06
0.06
0.01
0.01
0.01
0.04
0.02
0.05
0.03
0.03
0.01
0.01
0.01
0.03
0.02
0.04
0.05
0.04
0.01
0.01
0.01
t value
124.53
1.15
9.69
5.32
4.22
7.79
9.38
8.30
142.66
0.87
4.54
5.11
5.10
-12.13
-7.08
-0.99
128.67
0.45
7.48
7.84
6.17
-18.10
-9.56
-5.71
186.58
0.24
2.91
2.21
2.34
11.95
14.61
10.80
Note: Fixation duration value is log-transformed. WLC = word length. Non-significant values are marked in bold.
36
Table A3: Results from Linear Mixed-Effects model: Duration of First of
Multiple Fixations
Conditions
Mirrored Letters (mL)
b
5.43
-0.01
0.16
0.04
0.03
0.09
0.09
0.08
5.37
0.04
0.34
0.32
0.24
-0.16
-0.20
-0.10
5.46
0.03
0.33
0.25
0.20
-0.16
-0.14
-0.11
5.44
0.01
0.06
0.06
0.06
0.15
0.22
0.15
SE
0.05
0.02
0.03
0.03
0.03
0.01
0.02
0.02
0.05
0.03
0.04
0.03
0.05
0.02
0.01
0.01
0.04
0.02
0.05
0.05
0.05
0.02
0.02
0.02
0.03
0.02
0.04
0.03
0.03
0.02
0.02
0.02
t value
114.86
-0.31
4.56
1.23
0.93
6.04
5.86
4.68
101.59
1.56
9.23
10.98
4.36
-10.31
-14.05
-6.60
122.19
1.22
6.32
4.83
4.14
-10.58
-9.11
-7.32
162.02
0.27
1.54
1.84
1.77
8.80
12.83
8.52
Mirrored Words (mW)
Inverted Words (iW)
Scrambled Letters (sL)
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Note: Fixation duration value is log-transformed. WLC = word length. Non-significant values are marked in bold.
37
Table A4: Results from Linear Mixed-Effects model: Refixation Probabilities
Condition
Mirrored Letters (mL)
b
Mirrored Words (mW)
Inverted Words (iW)
Scrambled Letters (sL)
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
Intercept
WLC
Session 1
Session 2
Session 3
WLC:Session 1
WLC:Session 2
WLC:Session 3
-1.63
1.15
1.29
0.76
0.48
0.37
0.41
0.38
-1.35
1.15
1.39
1.35
1.19
0.55
0.64
0.69
-1.51
1.21
1.87
1.69
1.65
0.64
0.52
0.62
-1.51
1.46
0.78
0.57
0.67
0.32
0.42
0.31
SE
0.12
0.09
0.18
0.14
0.17
0.05
0.06
0.06
0.18
0.11
0.23
0.14
0.21
0.05
0.05
0.05
0.17
0.07
0.15
0.17
0.17
0.05
0.05
0.05
0.15
0.10
0.23
0.20
0.18
0.05
0.06
0.06
z value Pr(>z)
-13.75
12.52
7.32
5.53
2.75
6.86
7.33
6.48
-7.62
10.50
6.15
9.78
5.71
10.00
12.98
13.85
-9.03
18.14
12.49
10.06
9.98
11.76
9.96
12.17
-9.73
13.95
3.48
2.90
3.64
5.92
7.61
5.52
0.000
0.000
0.000
0.000
0.006
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.004
0.000
0.000
0.000
0.000
Note: WLC = word length. Non-significant values are marked in bold.
38
b
5.65
Intercept
-0.03
WLC
0.09
mL
0.01
sL
0.05
mW
0.04
iW
0.03
WLC:mL
WLC:sL
0.02
WLC:mW -0.05
WLC:iW -0.04
5.60
Intercept
0.00
WLC
mL
0.08
-0.01
sL
0.06
mW
0.04
iW
WLC:mL
0.04
WLC:sL
0.05
WLC:mW -0.06
WLC:iW -0.04
Intercept
-0.40
1.59
0.62
mL
0.04
sL
0.21
mW
0.21
iW
0.17
WLC:mL
WLC:sL
0.08
WLC:mW 0.07
WLC:iW 0.05
SE
0.03
0.01
0.02
0.02
0.01
0.01
0.00
0.00
0.00
0.00
0.03
0.01
0.02
0.01
0.01
0.01
0.01
0.01
0.00
0.00
0.08
0.04
0.08
0.06
0.04
0.04
0.02
0.02
0.01
0.01
z value
213.36
-3.57
3.95
0.46
3.90
4.07
6.97
8.38
-19.63
-19.74
212.41
-0.29
4.08
-0.42
5.47
4.45
5.82
9.21
-17.12
-14.35
-5.03
36.17
7.33
0.66
4.80
5.91
7.17
4.47
5.05
4.54
First of
Multiple
Fixations
duration
(FMF)
Refixation
Probability WLC
(RFX)
Table A5: Results from Linear Mixed-Effects model: Effect sizes accross
manipulation types
Measures
Single
Fixation
Duration
(SFD)
Pr(>z)
0.000
0.000
0.000
0.506
0.000
0.000
0.000
0.000
0.000
0.000
Note: LMM model with helmert contrast. SFD and FMD values are log-transformed (base 10).
WLC = word length. Non-significant values are marked in bold.
39
Table A6: Quadratic fit to initial landing position curve (two-fixation cases)
for reading normal text: Estimates of parameters A, B and C
Session Word Length Parameters Estimate Std. Error
0
4
4
4
5
5
5
6
6
6
7
7
7
0.020
0.008
0.101
0.015
0.004
0.121
0.012
0.003
0.168
0.016
0.003
0.251
z value Pr(z)
0.013
2.466
0.000
7.814
24.916
0.000
0.134
1.497
0.000
9.139
0.000
27.468
1.516
0.130
0.000
8.859
0.000
25.043
-0.518
0.604
0.000
7.115
19.834
0.000
A
B
C
A
B
C
A
B
C
A
B
C
0.051
0.065
2.528
0.023
0.040
3.336
0.019
0.023
4.21
-0.008
0.018
4.98
40
Table A7: Quadratic fit to initial landing position curve (two-fixation cases)
for reading mirrored-word (mW) text: Estimates of parameters A, B and
C
Session Word Length Parameters Estimate Std. Error
1
2
3
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
0.080
0.048
2.670
0.144
0.007
2.787
0.181
-0.009
3.523
0.173
-0.009
3.267
0.074
0.019
4.152
0.099
0.012
4.202
0.056
0.000
17.788
0.166
-0.007
3.057
0.043
0.063
2.718
0.0845
0.011
4.708
-0.008
0.001
17.752
0.176
-0.009
3.101
41
0.022
0.009
0.173
0.027
0.007
0.757
0.028
0.005
0.602
0.019
0.003
0.314
0.039
0.010
1.145
0.027
0.008
1.221
0.086
0.000
0.000
0.025
0.004
0.526
0.011
0.005
0.069
0.042
0.009
2.078
0.076
0.000
0.001
0.028
0.004
0.465
z value
3.684
5.309
15.650
5.288
1.037
3.684
6.533
-1.680
5.852
9.009
-3.265
10.392
1.897
1.9497
3.627
3.633
1.518
3.443
0.650
1.049
38076.98
6.692
-2.078
5.806
3.861
13.502
39.444
1.998
1.119
2.265
-0.101
2.050
6.275
-2.315
6.666
23601.67
Pr(z)
0.000
0.000
0.000
0.000
0.300
0.000
0.000
0.093
0.000
0.000
0.001
0.000
0.058
0.051
0.000
0.000
0.129
0.001
0.516
0.294
0.000
0.000
0.038
0.000
0.000
0.000
0.000
0.046
0.263
0.023
0.919
0.040
0.000
0.000
0.021
0.000
Table A8: Quadratic fit to initial landing position curve (two-fixation cases)
for reading mirrored-letter (mL) text: Estimates of parameters A, B and
C
Session Word Length Parameters Estimate Std. Error
1
2
3
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
-0,007
0,078
2,802
0,009
0,038
3,593
0,028
0,016
4,800
0,015
0,007
6,717
0,002
0,060
3,150
1,316
-0,002
-20,033
0,002
0,022
4,544
0,011
0,010
5,933
0,041
0,054
2,980
0,010
0,037
3,654
0,002
0,024
4,363
-0,002
0,017
4,990
42
0,022
0,009
0,120
0,017
0,005
0,172
0,006
0,001
0,161
0,016
0,002
0,787
0,004
0,002
0,035
0,331
0,001
0,018
0,019
0,004
0,324
0,010
0,002
0,405
0,023
0,010
0,211
0,021
0,006
0,227
0,020
0,004
0,290
0,020
0,003
0,329
z value Pr(z)
0.753
-0,314
0.000
8,332
23,424
0.000
0.621
0,494
0.000
7,626
0.000
20,842
4,569
0.000
0.000
12,389
0.000
29,770
0,943
0.346
0.000
4,283
0.000
8,540
0,446
0.656
0.000
36,706
0.000
89,4066
0.000
3,974
-3,487
0.000
0.000
-1087,91
0.924
0.000
0.000
0.282
0.000
0.000
0.066
0.000
0.000
0.631
0.000
0.000
0.936
0.000
0.000
0.902
0.000
0.000
0,095
5,463
14,023
1,076
6,450
14,661
1,838
5,436
14,152
0,487
6,021
16,110
0,080
5,568
15,042
-0,123
5,455
15,169
Table A9: Quadratic fit to initial landing position curve (two-fixation cases)
for reading inverted-word (iW) text: Estimates of parameters A, B and
C
Session Word Length Parameters Estimate Std. Error
1
2
3
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
-0.460
0.002
22.504
0.213
-0.012
1.510
0.214
-0.018
2.940
0.197
-0.013
3.064
0.058
0.030
3.677
0.006
0.002
11.017
0.042
0.000
24.665
0.167
-0.008
3.522
0.021
0.053
3.185
0.080
0.014
4.295
0.057
0.000
25.887
0.165
-0.008
3.685
43
0.116
0.000
0.004
0.022
0.006
0.653
0.026
0.005
0.243
0.023
0.003
0.276
0.014
0.005
0.309
0.054
0.001
0.001
0.116
0.000
0.001
0.025
0.004
0.439
0.010
0.004
0.106
0.029
0.008
1.099
0.114
0.000
0.000
0.022
0.003
0.426
z value
-3.976
5.742
5642.92
9.688
-1.884
2.312
8.132
-3.570
12.093
8.550
-3.954
11.102
4.153
5.758
11.908
0.115
3.198
11956.68
0.361
0.881
41339.44
6.785
-2.279
8.024
2.219
12.494
30.049
2.767
1.763
3.908
0.502
0.767
7.515
-2.388
8.654
51952.32
Pr(z)
0.000
0.000
0.000
0.000
0.060
0.021
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.909
0.001
0.000
0.718
0.378
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.006
0.078
0.000
0.616
0.443
0.000
0.000
0.017
0.000
Table A10: Quadratic fit to initial landing position curve (two-fixation cases)
for reading scrambled-letter (sL) text: Estimates of parameters A, B and
C
Session Word Length Parameters Estimate Std. Error
1
2
3
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
4
4
4
5
5
5
6
6
6
7
7
7
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
A
B
C
-0.034
0.087
2.833
0.017
0.031
3.861
0.022
0.019
4.563
0.620
0.000
-20.736
0.002
0.086
2.538
-0.009
0.042
3.580
0.021
0.014
5.149
0.618
-0.001
-20.966
1.376
-0.002
-22.215
0.025
0.029
3.918
0.020
0.015
5.046
-0.027
0.003
10.061
44
0.034
0.014
0.169
0.008
0.0024
0.118
0.008
0.002
0.155
0.043
7.177
0.001
0.026
0.010
0.095
0.024
0.007
0.210
0.015
0.003
0.492
0.055
0.000
0.001
0.758
0.001
0.041
0.009
0.003
0.145
0.020
0.004
0.601
0.011
0.000
0.000
z value
-1.020
6.051
16.776
2.123
13.118
32.791
2.859
11.561
29.519
14.332
-11.598
-19706.55
0.081
8.373
26.578
-0.363
6.196
17.015
1.348
4.705
10.456
11.303
-9.141
-17086.51
1.815
-1.553
-547.09
2.778
10.999
27.020
0.984
3.698
8.399
-2.442
16.467
49984.09
Pr(z)
0.301
0.000
0.000
0.034
0.000
0.000
0.004
0.000
0.000
0.000
0.000
0.000
0.935
0.000
0.000
0.716
0.000
0.000
0.178
0.000
0.000
0.000
0.000
0.000
0.070
0.121
0.000
0.005
0.000
0.000
0.325
0.000
0.000
0.015
0.000
0.000
Figure A1: Within-word landing position distribution group by launch-site
distance and word length for reading mirrored-word (mW) texts. Red line
and dots represent data from normal reading session. Data from the first
experimental session are presented in dark blue color. The lighter blue hues
represent the last two experimental sessions.
45
3−letterwords4−letterwords5−letterwords6−letterwords7−letterwordslaunch site−5launch site−4launch site−3launch site−2launch site−102468024680246802468024680.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.5Initial Landing Position [char]Relative FrequencySession:ControlSession 1Session 2Session 3Landing Position Distribution: Mirrored Words (mW) Figure A2: Within-word landing position distribution group by launch-site
distance and word length for reading mirrored-letter (mL) texts. Red line
and dots represent data from normal reading session. Data from the first
experimental session are presented in dark blue color. The lighter blue hues
represent the last two experimental sessions.
46
3−letterwords4−letterwords5−letterwords6−letterwords7−letterwordslaunch site−5launch site−4launch site−3launch site−2launch site−102468024680246802468024680.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.5Initial Landing Position [char]Relative FrequencySession:ControlSession 1Session 2Session 3Landing Position Distribution: Mirrored Letters (mL) Figure A3: Within-word landing position distribution group by launch-site
distance and word length for reading inverted-word (iW) texts. Red line
and dots represent data from normal reading session. Data from the first
experimental session are presented in dark blue color. The lighter blue hues
represent the last two experimental sessions.
47
3−letterwords4−letterwords5−letterwords6−letterwords7−letterwordslaunch site−5launch site−4launch site−3launch site−2launch site−102468024680246802468024680.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.5Initial Landing Position [char]Relative FrequencySession:ControlSession 1Session 2Session 3Landing Position Distribution: Inverted Words (iW) Figure A4: Within-word landing position distribution group by launch-site
distance and word length for reading scrambled-letter (sL) texts. Red line
and dots represent data from normal reading session. Data from the first
experimental session are presented in dark blue color. The lighter blue hues
represent the last two experimental sessions.
48
3−letterwords4−letterwords5−letterwords6−letterwords7−letterwordslaunch site−5launch site−4launch site−3launch site−2launch site−102468024680246802468024680.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.50.00.10.20.30.40.5Initial Landing Position [char]Relative FrequencySession:ControlSession 1Session 2Session 3Landing Position Distribution: Scrambled Letters (sL) |
1208.6041 | 3 | 1208 | 2012-10-03T19:21:32 | Response Selection Using Neural Phase Oscillators | [
"q-bio.NC",
"physics.bio-ph"
] | In a recent paper, Suppes et al. (2012) [arXiv:arXiv:1010.3063] used neural oscillators to create a model, based on reasonable neurophysiological assumptions, of the behavioral stimulus-response (SR) theory. In this paper, we describe the model in a less mathematical and more physical and intuitive way. | q-bio.NC | q-bio |
Response Selection Using Neural Phase
Oscillators
J. Acacio de Barros∗, and G. Oas†
∗Liberal Studies, San Francisco State University, San Francisco, CA, 94132
†CSLI, Stanford University, Stanford, CA 94305
Abstract
In a recent paper, Suppes, de Barros, & Oas (2012) used neural oscillators to cre-
ate a model, based on reasonable neurophysiological assumptions, of the behavioral
stimulus-response (SR) theory. In this paper, we describe the main characteristics of
the model, emphasizing its physical and intuitive aspects.
1 Introduction
It is an honor for Acacio de Barros and Gary Oas to participate in a festschrift for
Pat Suppes. It is especially rewarding to do so with a paper where we discuss our
most recent work with Pat, a model of brain processes using neural oscillators.
We are, as Pat would say, "true blue physicists." So, for us, collaborating with
Pat on this truly interdisciplinary paper is an example not only of his intellectual
influence, but also of his friendship and mentorship. We are happy to dedicate
this paper to Pat. Happy Birthday Pat!
The work we present here started more than ten years ago, when JAB and
Pat begun thinking about how to model in a physically plausible way collections
of neurons in terms of oscillators.
In one of his known intuitions, Pat kept
insisting that the brain "gotta use oscillators." Of course, as is often the case,
his "intuition" was based on hard work and detailed empirical data that he
collected working on the EEG of words and sentences. Nevertheless, as we
kept trying to make our model work (and we had many failures, and some
successes; see Vassilieva et al. (2011) for an example), Pat kept insisting: we
should understand the brain computations with oscillators.
I am pleased to
say that, despite my initial skepticism, we now have a model that we feel is not
only grounded on neurophysiologically sound evidence, but that also reproduces
quite well some empirical behavioral data (Suppes, de Barros, & Oas, 2012).
In this paper we attempt to describe the main features this model by fo-
cusing on the physical processes underlying the neural computations. We chose
to do so for the following reasons. First, because of its interdisciplinarity, our
1
1 Introduction
2
model requires concepts from many different areas (neurophysiology, physics,
psychology, etc). Such concepts are not complex, but are often unfamiliar to
most researchers. Second, we are confident that our model is relevant to cogni-
tive psychologists, as it may explain some mathematical models showing good
empirical fit (de Barros, 2012a,b). So, we believe that this paper can provide a
clearer and intuitive view of the main physical features of our model for those
thinking about applying it, supplementing the discussions found in Suppes, de
Barros, & Oas (2012).
Let us start our discussion with the broad problem of understanding how
the brain processes information. This is perhaps the most challenging cur-
rent scientific endeavors, mainly due to the fact that our brain is tremendously
complicated, as it is constituted of many different components that are, by
themselves, complex, but that also seem to sometimes interact holistically with
each other. Among the approaches to try and understand the brain, the most
prominent ones are the top-down and bottom-up. In the top-down approach,
we start with the higher-level functions and go to their underlying mechanisms.
An example of such approach would be the field of cognitive neuroscience, where
often one starts with experiments in cognitive psychology and tries to under-
stand them from principles in neuroscience (Adolphs, 2003). In the bottom-up
approach, one tries to start with neurophysiology, and by studying how each
elementary component works, one tries to see how higher functions arise from
such components or their interaction (Kandel et al., 2000).
Each of those approaches have their shortcomings. For example, one of the
main issues is what we may call a problem of scale. When trying to understand
a complex system, the first question that arises is how detailed we need to be.
In the case of the brain, some researchers say that we need to go all the way
down to the chemical reactions in the synapses. Others argue that individual
neurons hold the key to understanding brain computation. Yet another view is
that collections of neurons are important. So, when trying to understand how
the brain works, our first problem is where to begin. Regardless of what scale is
chosen and where we start, ultimately we would need to understand the whole
process if we were to claim to have understood the brain.
The main problem with connecting a higher scale with a lower one is due
to its complexity. For example, evidence exists that higher cognitive processes
involve tens to hundreds of thousands of neurons, interacting with each other in
very complex ways. Modeling such processes require the use of powerful com-
puters. But, even when a model is shown to work from the underlying neuronal
dynamics, the use of massive computer simulations helps little in understanding,
in an intuitive or conceptual way, what is actually happening. The system is
simply too complex.
To deal with the issue of complexity, different approaches can be taken. One
possible route is to find physically plausible arguments that impose constraints
on the system's dynamics, therefore reducing it to fewer degrees of freedom. This
is the approach taken by Suppes, de Barros, & Oas (2012). In their paper, a large
number of independent neurons was modeled by a single dynamical parameter
determined by the phase of a neural oscillator. They then showed that under
2 A Brief Review of SR theory
3
certain reasonable assumptions, the main characteristics of behavioral stimulus-
response (SR) theory could be described by neural oscillators. The use of neural
oscillators thus provided a significant reduction on the number of degrees of
freedom, allowing for the physical interpretation of many different parameters
in the model.
In this paper we present the work of Suppes, de Barros, & Oas (2012),
with emphasis on the physics and intuition behind the model. Our goal is to
make this model more understandable, as many of the concepts used in our
previous paper are not well-known to certain audiences. For example, while all
physicists have an excellent knowledge of oscillations and interference and could
easily follow the arguments leading from neurons to oscillators, only a few would
feel comfortable with the mathematical learning theories used. Neuroscientists,
on the other hand, would probably feel at home with neurons and learning
theories, but not so much with oscillators and interference. Neither would most
psychologists. Here we focus on the intuitions behind the physics, with the hopes
that, in conjunction with the oscillator model, psychologists and neuroscientists
could benefit more from the insights gained.
2 A Brief Review of SR theory
Stimulus-response theory (or SR theory; see Suppes and Atkinson (1960)) is one
of the most successful behavioral learning theories in psychology. Though it has
decreased in importance in current psychology, we chose to model SR theory for
the following reasons. First, it is based on a rigid trial structure, which permits
its concepts to be formally axiomatized, resulting in many important non-trivial
but illuminating representation theorems (Suppes, 2002). In fact, the theory is
rich enough to represent language in it. Second, despite its few parameters (the
learning probability c and the number of stimuli), it has been shown to fit well
to empirical data in a variety of experiments. Finally, as we showed in Suppes,
de Barros, & Oas (2012), SR theory seems to have natural counterparts at a
neuronal level, and is, in some sense still used by neuroscientists (though, sadly,
not in its mathematical form).
Here we present the mathematical version of SR theory for a continuum of
responses, formalized in terms of a stochastic process (we follow Suppes, de
Barros, & Oas, 2012). Let (Ω,F , P ) be a probability space, and let Z, S, R,
and E be random variables, with Z : Ω → ES S : Ω → S, R : Ω → R, and
E : Ω → E, where S is the set of stimuli, R the set of responses, and E the set
of reinforcements. Then a trial in SR theory has the following structure:
Zn → Sn → Rn → En → Zn+1.
(1)
The trial structure works the following way. Trial n starts with a certain state
of conditioning and a sampled stimulus. Once a stimulus is sampled, a response
is computed according to the state of conditioning. Then, reinforcement fol-
lows, which can lead (with probability c) to a new state of conditioning for
3 Oscillator model
4
trial n + 1 . In more detail, at the beginning of a trial, the state of condition-
1
i
1
m
, . . . , z(n)
1
, its corresponding z(n)
i
determines
on trial n. Once a stimulus Sn = si is sam-
m (cid:17). The vector
m (cid:17) associates to each stimuli si ∈ S, i = 1, . . . , m, where m = S
, . . . , z(n)
ing is represented by the random variable Zn = (cid:16)z(n)
(cid:16)z(n)
is the cardinality of S, a value z(n)
pled with probability P (Sn = sisiǫS) =
the probability of responses in R by the probability distribution K (cid:16)rz(n)
P (cid:16)a1 ≤ Rn ≤ a2Sn = si, Zn,i = z(n)
the i-th component of the vector (cid:16)z(n)
tion K(cid:16)rz(n)
ance and mode z(n)
with probability c, i.e. P (cid:16)Zn+1,i = ySn = si, En = y, Zn,i = z(n)
P (cid:16)Zn+1,i = z(n)
i (cid:17), i.e.
i (cid:17)
i (cid:17) dx, where k(cid:16)xz(n)
k(cid:16)xz(n)
m (cid:17). The probability distribu-
i (cid:17) is the smearing distribution, and it is determined by its vari-
i (cid:17) = c and
i (cid:17) = 1 − c. The trial ends with a
is the probability density associated to the distribution, and where Zn,i
. The next step is the reinforcement En, which is effective
Sn = si, En = y, Zn,i = z(n)
i (cid:17) = R a2
a1
, . . . , z(n)
1
is
i
i
new (with probability c) state of conditioning Zn+1.
3 Oscillator model
In this section we will describe intuitively the oscillator model. We start by
arguing for the use of neural oscillators as a way to model the brain at a system
level. We then discuss how we can represent in a mathematically sensible way
these oscillators. Finally, we show how response computations and learning can
be modeled using this theoretical apparatus. Readers interested in more detail
are referred to Suppes, de Barros, & Oas (2012).
There are many different ways in which researchers try to figure out how the
brain works. For example, in cognitive neuroscience, among the most popular
research techniques are fMRI (functional magnetic resonance imaging), MEG
(magnetoencephalogram), and EEG (electroencephalogram). MEG and EEG
measure the electrical activities in the brain, whereas fMRI measures changes
in blood flow associated with higher metabolic rates. While fMRI's popularity is
due to its better spatial resolution, MEG and EEG present significantly better
time resolution. However, what these techniques have in common is that, in
order to measure a signal from the brain, they require a large numbers of neurons
to fire synchronously. To make our point, let us focus on EEG (though MEG
would be adequate too). There are many experiments (see Carvalhaes et al.
(2012) and references) showing that the EEG data allow a good representation
of language or visual imagery. Thus, neurophysiological evidence points toward
language being an activity involving large collections of synchronizing neurons,
and we will center our model exactly on this.
Before we show how to describe such collections of synchronizing neurons
mathematically, it is useful to think about the physical mechanisms of synchro-
3 Oscillator model
5
V (a.u.)
TA
t0
t0+TA
t0+2TA
t (a.u.)
V (a.u.)
TB
t0
t0+TB
t0+2TAB
t (a.u.)
Fig. 1: Approximate shape of the action potentials VA and VB as a function of
time t for two uncoupled neurons nA and nB firing periodically, with
periods TA and TB. For simplicity, we chose a t0 when both neurons fire
simultaneously.
nization. Let us look first at individual neurons, and then think about ensembles
of neurons. Figure 1 shows the qualitative behavior of two neurons nA and nB
firing periodically, with TB < TA. What happens if we now couple nA to an
excitatory synapse coming from neuron nB? Because t0 + TB < t0 + TA, the
excitatory coupling will increase the membrane potential of neuron nA before
t0 +TA, causing nA to fire a little earlier than it would if it were not connected to
nB. So, excitatory synaptic couplings between neurons can change the timing
of coupling, and this timing is changed such that the firings of both neurons
approach (in this case, the firing of nA approaches that of nB). In other words,
excitatory couplings push nA and nB toward synchronization. In fact, it is pos-
sible to prove mathematically that if the number of neurons is large enough, the
sum of the many weak synaptic interactions can cause a strong effect, making all
neurons fire closer together (Izhikevich, 2007); even when weakly coupled, en-
sembles of periodically firing neurons synchronize. It is interesting to note that
the argument shown above can be scaled up to distinct collections of neurons.
Imagine we have two ensembles of neurons, NA and NB, such that neurons in
them synchronize. If neurons in NA and NB become coupled, then the same
mechanism as discussed above will be at play, and the ensembles will synchro-
nize among themselves. We will come back to this point later, when we talk
about response mechanisms.
We are now in good shape to introduce the intuition behind the mathemati-
cal description for the dynamics of synchronization. One of the main simplifying
3 Oscillator model
6
assumptions we make is that the relevant information coded in the brain is rep-
resented by the synchronization of an ensemble of neurons. This ensemble may
include tens of thousands of neurons, but because they are synchronized, we can
represent them, at least in first approximation, by a single dynamical variable.
To understand this, let us think about the simplest case, where an oscillator
O (t) can be represented by a sine function1:
O (t) = A sin ωt,
(2)
where ω = ω (t) is its time-dependent frequency. Since ω may be a function of
time, the value of O (t) is completely determined by the argument of the sine, i.e.
by ϕ = ωt. The quantity ϕ is the phase of the oscillator O (t) = A sin ϕ (t). Since
collections of firing neurons have very little variability in its intensity (except,
as we see below, when they interfere), we can describe a neural oscillator by its
phase. The interaction of a neural oscillator with other neural oscillators may
change the evolution of its phase.
We emphasize that there is a certain invariance of scale in the above argu-
ment: it somehow does not matter how many neurons we have; all that matters
is that their amplitude does not vary, that their couplings are strong enough
to produce synchronization, and that their dynamics is encoded in the phase.
Furthermore, in the same way that individual oscillating neurons synchronize to
each other, a collection of coherent neurons can also synchronize to another col-
lection of coherent neurons. Since neurons firing coherently may be described
approximately by their phase, we can focus on the phase dynamics, instead
of being concerned about the full description of the very complex dynamical
system.
Now, let us look a little more into the details of the mathematics of two
synchronizing oscillators. Let us start with two oscillators, O1 (t) and O2 (t),
described by their phases ϕ1 and ϕ2. If the two oscillators are uncoupled and
their frequency ω is constant, then it is clear from equation (2) that they should
satisfy the following set of differential equations,
dϕ1
dt
dϕ2
dt
= ω1,
= ω2,
(3)
(4)
where ωi, i = 1, 2, are their natural frequencies. However, if they are weakly
coupled, such that their interaction does not affect the overall form of the os-
cillations given by O1 (t) and O2 (t) but affects their phase, then equations (3)
and (4) need to be modified to include changes to the phase. Furthermore, if
the underlying interaction is such that it will make the phases approach each
other, such as in the case of synaptically coupled neurons, then it is possible to
1 We use a sine function for simplicity, but the following argument is valid for periodic
functions.
3 Oscillator model
7
show that, in first approximation, the modified dynamical equations become
dϕ1
dt
dϕ2
dt
= ω1 − k12 sin (ϕ1 − ϕ2) ,
= ω2 − k21 sin (ϕ2 − ϕ1) ,
(5)
(6)
where kij are the phase coupling strengths. If we extend this to allow for N
oscillators, equations (5) and (6) then become
dϕi
dt
= ωi −Xj6=i
kij sin (ϕi − ϕj ) .
(7)
Equation (7) is known as Kuramoto equation (Kuramoto, 1984), and it is widely
used to describe complex systems with emergent synchronization. The strength
and usefulness of Kuramoto's equation comes from two main points. First, it
can be solved under certain symmetric conditions and in the limit of large N ,
yielding significant insight into the nature of emerging synchronization. Second,
a set of weakly-coupled oscillating dynamical systems close to a Andronov-Hopf
bifurcation can be described, in first approximation, by Kuramoto-like equations
(see Izhikevich (2007)). For our purpose, Kuramoto's equations are a good
approximation for the dynamics of coupled neural oscillators.
So, we now turn into the discussion of how we can think of stimulus and
response as modeled by oscillators, and in particular by Kuramoto's equations.
The basic idea is simple. Once a distal stimulus is presented, the perceptual
system activates an ensemble of brain neurons, Ns, associated with it. This
system itself is described by Kuramoto's equations, and, because it synchronizes,
we use its average phase to describe its mean dynamics. If this stimulus elicits a
response, the activation of the response neurons via synaptic couplings follows.
Responses, as stimuli, are also represented by synchronously firing ensemble
of neurons. The selection of a particular response happens when the stimulus
oscillator synchronizes in phase with it, and such phase is determined by the
relative couplings between stimulus and response oscillators. Let us now look
more into its detail.
The simplest stimulus-response neural oscillator model requires three os-
cillators, Os, Or1 , and Or2 . Os is the oscillator representing firing neurons
corresponding to the sampling of a stimulus, and Or1 and Or2 are the response
oscillators. Their phases are ϕs, ϕr1 , and ϕr2 . Before we describe their dy-
namics, let us go through the process of a response computation. Whenever Os
is activated, and subsequently Or1 and Or2 , then the intensity of firings (i.e.,
the rate of firing, as the individual neuron amplitudes are reasonably stable) in
each response oscillator is not only due to its firing, but also to the firings of
Os. As we mentioned earlier, a collection of firing neurons may interfere, and
in this case, interference means stronger firing rates when in phase, and weaker
firing rates when off of phase. Let us analyze this with a mathematically simple
3 Oscillator model
example of equal intensity harmonic oscillators, given by
Os(t) = A cos (ω0t) = A cos (ϕs(t)) ,
Or1 (t) = A cos (ω0t + δφ1) = A cos (ϕr1(t)) ,
Or2 (t) = A cos (ω0t + δφ2) = A cos (ϕr2(t)) .
8
(8)
(9)
(10)
Equations (8) -- (10) represent the case where the oscillators are already syn-
chronized with the same frequency ω0 but with relative but constant phase
differences δφ1 and δφ2. The mean intensity give us a measure of the excitation
carried by the oscillations, and for the superposition of Os(t) and Or1 (t) it is
given by
I1 = D(Os(t) + Or1 (t))2Et
where hf (t)it0 = 1
computation yields
t0
= (cid:10)Os(t)2(cid:11)t +(cid:10)Or1 (t)2(cid:11)t + h2Os(t)Or1 (t)it ,
∆T R t0+∆T
f (t) dt (∆T ≫ 1/ω0) is the time average . A quick
I1 = A2 (1 + cos (δφ1)) ,
and, similarly for I2,
I2 = A2 (1 + cos (δφ2)) .
Therefore, the intensity depends on the phase difference between the response-
computation oscillators and the stimulus oscillator.
Now, the maximum intensity of I1 and I2 is 2A2, whereas their minimum
intensity is zero. If we think of I1 and I2 as competing possible responses, the
maximum difference between them happens when one of their relative phases
(with respect to the stimulus oscillator) is zero while the other is π. It is standard
to use the contrast, defined by
b =
I1 − I2
I1 + I2
,
(11)
as a measure of how different the intensities are. From its definition, b takes
values between −1 and 1. When I1 and I2 are as different as possible, b = 1;
if, on the other hand, I1 and I2 are the same, b = 0.
The contrast provides us with a useful way to think about responses that
are between r1 and r2. To see this, let us impose
which results in
and
δφ1 = δφ2 + π ≡ δφ,
I1 = A2 (1 + cos (δφ)) ,
I2 = A2 (1 − cos (δφ)) .
(12)
(13)
(14)
In this case, the single parameter δφ is sufficient to determine the contrast, as
b = cos (δφ) ,
(15)
3 Oscillator model
9
0 ≤ δϕ ≤ π. So, the phase difference δφ between stimulus and response oscilla-
tors codes a continuum of responses between −1 and 1 (more precisely, because
δϕ is a phase, the interval is in the unit circle T, and not in a compact interval
in R). For arbitrary intervals (ζ1, ζ2), all that is required is a re-scaling of b.
To summarize the above arguments. When a stimulus and response os-
cillators activate, they fire periodically,leading to their synchronization with
constant phase relation. This phase relation causes interference, which in turn
determines the relative strength of the intensities for each response. Thus, re-
sponses are determined by the interference of oscillators, which is itself affected
by the neural oscillators' couplings.
We now examine in more detail the mathematics of the stimulus and response
model. Let us look at each step of (1).
Sampling
When a stimulus sn is sampled, a collection of neurons start firing synchronously,
corresponding to the activation of a neural oscillator, Osn . Such activation leads
to a spreading of activation to oscillators coupled to the stimulus oscillator, in-
cluding the response Or1 and Or2 . Since the selection and activation of Osn
involves the perceptual system, we do not attempt to model with neural oscil-
lators this step, but simply assume their activation in a way that is consistent
with the stochastic process represented in SR theory by the random variable
Sn. Furthermore, though it would be important to develop a detailed theory of
spreading activation, we do not, as for our current purposes it suffices to simply
assume the activation of Or1 and Or2 .
Response
After the stimulus sn is sampled, the active oscillators evolve for the time in-
terval ∆tr, the time it takes to compute a response, according to the following
set of Kuramoto differential equations.
dϕi
dt
= ωi −Xi6=j
kij sin (ϕi − ϕj + δij) ,
(16)
where kij is the coupling constant between oscillators i and j, and δij is an
anti-symmetric matrix representing phase differences, and i and j can be either
Osn , Or1 , or Or2 . Here we use the notation where Oi corresponds to a neural
oscillator and ϕi to its phase. Equation (16) can be rewritten as
dϕi
dt
= ωi −Xj (cid:2)kE
ij sin (ϕi − ϕj) + kI
ij cos (ϕi − ϕj)(cid:3) ,
(17)
ij = kij cos (δij) and kI
where kE
ij = kij sin (δij ), which has an immediate phys-
ical interpretation: kE
ij corre-
sponds to inhibitory ones. These are the 4N excitatory and inhibitory coupling
ij corresponds to excitatory couplings, whereas kI
3 Oscillator model
strengths between oscillators.
dϕi
dt
= ω0 −Xi6=j (cid:2)kE
i,j sin (ϕi − ϕj ) − kI
i,j cos (ϕi − ϕj )(cid:3) ,
10
(18)
where ω0 is their natural frequency. The solutions to (18) and the initial con-
ditions randomly distributed at activation give us the phases at time tr,n =
ts,n + ∆tr. The coupling strengths between oscillators determine their relative
phase locking, which in turn corresponds to the computation of a given response,
according to equation (11).
Reinforcement and Conditioning
As we saw above, the computation of a response depends on the inhibitory and
excitatory couplings between neural oscillators. Therefore, when an effective
reinforcement Yn corresponding to changes in the conditioning Zn+1 occurs,
the coupling strengths change. As with stimulus and responses, we represent
a reinforcement by a neural oscillator. Such oscillator, with frequency ωe, is
activated during reinforcement, and we assume that it forces the reinforced
response-computation and stimulus oscillators to synchronize with the same
phase difference of δϕ, while the two response-computation oscillators are kept
synchronized with a phase difference of π. Let the reinforcement oscillator be
activated on trial n at time te,n, tr,n+1 > te,n > tr,n, for an interval of time
∆te. Let K0 be the coupling strength between the reinforcement oscillator
and the stimulus and response-computation oscillators. In order to match the
probabilistic SR axiom governing the effectiveness of reinforcement, we also
assume that there is a normal probability distribution governing the coupling
strength K0 between the reinforcement and the other active oscillators with
probability density
f (K0) =
1
σK0√2π
exp(cid:26)−
1
2σ2
K0 (cid:0)K0 − K 0(cid:1)2(cid:27) .
(19)
When a reinforcement is effective, all active oscillators phase-reset at te,n, and
during reinforcement the phases of the active oscillators evolve according to the
following set of differential equations.
dϕi
dt
= ω0 −Xi6=j (cid:2)kE
i,j sin (ϕi − ϕj) − kI
i,j cos (ϕi − ϕj)(cid:3)
−K0 sin (ϕi − ωet + Φi) ,
(20)
where Φsn − Φr1 = δϕ and Φr1 − Φr2 = π. The excitatory couplings are rein-
forced if the oscillators are in phase with each other, according to the following
equations.
dkE
i,j
dt
= ǫ (K0)(cid:2)α cos (ϕi − ϕj) − kE
i,j(cid:3) .
(21)
4 Final remarks
11
Similarly, for inhibitory connections, if two oscillators are perfectly off sync,
then we have a reinforcement of the inhibitory connections.
dkI
i,j
dt
= ǫ (K0)(cid:2)α sin (ϕi − ϕj) − kI
i,j(cid:3) ,
In the above equations,
ǫ (K0) = (cid:26) 0 if K0 < K ′
ǫ0 otherwise,
(22)
(23)
where ǫ0 ≪ ω0, α and K0 are constant during ∆te, and K ′ is a threshold
constant throughout all trials. We can think of K ′ as a threshold below which
the reinforcement oscillator has no (or very little) effect on the stimulus and
response-computation oscillators. For large enough values of ∆te, the behavioral
probability parameter c of effective reinforcement mentioned above is, from (19)
and (23), reflected in the equation:
c = Z ∞
K ′
f (K0) dK0.
(24)
This relationship comes from the fact that, if K0 < K ′, there is no effective
learning from reinforcement, since there are no changes to the couplings due to
(21) -- (22), and (18) describing the oscillators' behavior. Intuitively K ′ is the
effectiveness parameter: the larger it is, the smaller the probability of effective
reinforcement.
4 Final remarks
In this paper we described the neural oscillator model presented in Suppes, de
Barros, & Oas (2012), with particular emphasis to the physics and intuition
behind many of the processes represented by equations (18). To summarize
it, the coded phase differences were used to model a continuum of responses
within SR theory in the following way. At the beginning of a trial a stimulus
oscillator is activated, and with it the response oscillators. Then, the coupled
oscillator system evolves according to (18) if no reinforcement is present, and
according to (20) -- (22) if reinforcement is present. The coupling constants and
the conditioning of stimuli are not reset at the beginning of each trial, and
changes to couplings correspond to effective reinforcement. Because of the finite
amount of time for a response, the probabilistic characteristics of the initial
conditions lead to the smearing of the phase differences after a certain time,
with an effect similar to that of the smearing distribution in the SR model for
a continuum of responses (Suppes, 1959).
We emphasize that in this paper we focused mainly on the physical basis
of our model, and did not go much into mathematical detail. Furthermore, in
Suppes, de Barros, & Oas (2012) we applied the neural oscillator model to many
different experimental situations illustrated in the literature, whereas here we
4 Final remarks
12
did not address in detail any empirical data. Interested readers are referred to
our original paper.
SR theory has enjoyed tremendous success in the past, and, in a certain
sense, its main features are still present in modern day neuroscience. We be-
lieve that by showing how neurons may result in theoretical structures that are
somewhat similar to SR ones, as done in Suppes, de Barros, & Oas (2012), we
can provide the basis for an extension of SR theory that could be considered
more realistic. For example, in our model, many parameters, such as time of
response, frequency of oscillations, coupling strengths, etc., were fixed based on
reasonable assumptions. However, a more detailed and systematic study should
be able to relate such parameters to either underlying physiological constraints
or to behavioral variations, thus opening up the possibilities for new empiri-
cal studies that go beyond SR theory. Also, in our model we postulated many
features without showing or proving their dynamics from underlying neuronal
dynamics. This was the case for the activation of a stimulus and the spreading of
activation of a stimulus and responses. A more detailed theory based on neural
oscillators of such dynamics would certainly provide interesting empirical tests.
Finally, the use of neural oscillators and interference may also help explain
certain aspects of cognition that are considered "non-classical." The distinction
between classical and quantum behavior is a subtle one, and still not yet un-
derstood. For example, a well studied quantum-like decision making process is
the violation of Savage's sure-thing principle, shown in a series of experiments
by Tversky and Shaffir (Shafir and Tversky, 1992; Tversky and Shafir, 1992).
Similar violations do not need any quantum-like representation in the form of
a Hilbert space, as proposed in the literature, but instead can be obtained by
interference of neural oscillators (de Barros, 2012b). Furthermore, the use of
neural oscillator interference even leads to predictions that are not compatible
with a Hilbert space structure (de Barros, 2012a), suggesting that the use of
quantum-like processes is not as quantum as many would wish.
References
Adolphs, R. (2003). Cognitive neuroscience of human social behaviour. Nature
Reviews Neuroscience, 4(3):165 -- 178.
Carvalhaes, C. G., de Barros, J. A., Perrau-Guimaraes, M., and Suppes, P.
(2012). Using the scalp electric field to improve recognition of brain processes.
In preparation.
de Barros, J. A. (2012a). Joint probabilities and quantum cognition. In Khren-
nikov, A., editor, Proceedings of Quantum Theory: Reconsiderations of Foun-
dations - 6, Växjö, Sweden. Institute of Physics. To appear.
de Barros, J. A. (2012b). Quantum-like model of behavioral response computa-
tion using neural oscillators. arXiv:1207.0033. Submitted to BioSystems.
4 Final remarks
13
de Barros, J. A. and Suppes, P. (2009). Quantum mechanics, interference, and
the brain. Journal of Mathematical Psychology, 53(5):306 -- 313.
Izhikevich, E. M. (2007). Dynamical Systems in Neuroscience: The Geometry
of Excitability and Bursting. The MIT Press, Cambridge, Massachusetts.
Kandel, E. R., Schwartz, J. H., and Jessell, T. M. (2000). Principles of neural
science. McGraw-Hill, New York, NY, 4th edition.
Kuramoto, Y. (1984). Chemical Oscillations, Waves, and Turbulence. Dover
Publications, Inc., Mineola, New York.
Shafir, E. and Tversky, A. (1992). Thinking through uncertainty: Nonconse-
quential reasoning and choice. Cognitive Psychology, 24(4):449 -- 474.
Suppes, P. (1959). Stimulus sampling theory for a continuum of responses.
In Arrow, K., Karlin, S., and Suppes, P., editors, Mathematical Methods
in the Social Sciences, pages 348 -- 365. Stanford University Press, Stanford,
California.
Suppes, P. (2002). Representation and Invariance of Scientific Structures. CSLI
Publications, Stanford, California.
Suppes, P. and Atkinson, R. C. (1960). Markov learning models for multiperson
interactions. Stanford University Press, Stanford, California.
Suppes, P. and de Barros, J. (2007). Quantum mechanics and the brain. Quan-
tum Interaction: Papers from the AAAI Spring Symposium, Technical Report
SS-07, 8:75 -- 82.
Suppes, P., de Barros, J. A., and Oas, G. (2012). Phase-oscillator computations
as neural models of stimulus -- response conditioning and response selection.
Journal of Mathematical Psychology, 56(2):95 -- 117.
Tversky, A. and Shafir, E. (1992). The disjunction effect in choice under uncer-
tainty. Psychological Science, 3(5):305 -- 309.
Vassilieva, E., Pinto, G., de Barros, J., and Suppes, P. (2011). Learning pat-
tern recognition through quasi-synchronization of phase oscillators. Neural
Networks, IEEE Transactions on, 22(1):84 -- 95.
|
1102.0166 | 1 | 1102 | 2011-02-01T14:32:38 | Hebbian learning of recurrent connections: a geometrical perspective | [
"q-bio.NC",
"nlin.AO"
] | We show how a Hopfield network with modifiable recurrent connections undergoing slow Hebbian learning can extract the underlying geometry of an input space. First, we use a slow/fast analysis to derive an averaged system whose dynamics derives from an energy function and therefore always converges to equilibrium points. The equilibria reflect the correlation structure of the inputs, a global object extracted through local recurrent interactions only. Second, we use numerical methods to illustrate how learning extracts the hidden geometrical structure of the inputs. Indeed, multidimensional scaling methods make it possible to project the final connectivity matrix on to a distance matrix in a high-dimensional space, with the neurons labelled by spatial position within this space. The resulting network structure turns out to be roughly convolutional. The residual of the projection defines the non-convolutional part of the connectivity which is minimized in the process. Finally, we show how restricting the dimension of the space where the neurons live gives rise to patterns similar to cortical maps. We motivate this using an energy efficiency argument based on wire length minimization. Finally, we show how this approach leads to the emergence of ocular dominance or orientation columns in primary visual cortex. In addition, we establish that the non-convolutional (or long-range) connectivity is patchy, and is co-aligned in the case of orientation learning. | q-bio.NC | q-bio |
Hebbian learning of recurrent connections: a
geometrical perspective
Mathieu N. Galtier ∗
Olivier D. Faugeras †
Paul C. Bressloff ‡
May 29, 2018
Abstract: We show how a Hopfield network with modifiable recurrent connec-
tions undergoing slow Hebbian learning can extract the underlying geometry of
an input space. First, we use a slow/fast analysis to derive an averaged system
whose dynamics derives from an energy function and therefore always converges
to equilibrium points. The equilibria reflect the correlation structure of the in-
puts, a global object extracted through local recurrent interactions only. Second,
we use numerical methods to illustrate how learning extracts the hidden geomet-
rical structure of the inputs. Indeed, multidimensional scaling methods make it
possible to project the final connectivity matrix on to a distance matrix in a high-
dimensional space, with the neurons labelled by spatial position within this space.
The resulting network structure turns out to be roughly convolutional. The resid-
ual of the projection defines the non-convolutional part of the connectivity which
is minimized in the process. Finally, we show how restricting the dimension of
the space where the neurons live gives rise to patterns similar to cortical maps.
We motivate this using an energy efficiency argument based on wire length min-
imization. Finally, we show how this approach leads to the emergence of ocular
∗Corresponding author: [email protected]. NeuroMathComp Project Team, INRIA
Sophia-Antipolis M´editerran´ee, 2004 route des Lucioles-BP 93, 06902 Sophia Antipolis, France
†NeuroMathComp Project Team, INRIA Sophia-Antipolis M´editerran´ee, 2004 route des
Lucioles-BP 93, 06902 Sophia Antipolis, France
‡Department of Mathematics, University of Utah, 155 South 1400 East, Salt Lake City, Utah
84112, USA. Mathematical Institute, University of Oxford, 24-29 St. Giles', Oxford OX1 3LB,
UK
1
dominance or orientation columns in primary visual cortex. In addition, we es-
tablish that the non-convolutional (or long-range) connectivity is patchy, and is
co-aligned in the case of orientation learning.
Keywords:
averaging, energy minimization, multidimensional scaling, cortical maps
correlation-based Hebbian learning, Hopfield networks, temporal
1 Introduction
Activity-dependent synaptic plasticity is generally thought to be the basic cellular
substrate underlying learning and memory in the brain. Donald Hebb [Hebb, 1949]
postulated that learning is based on the correlated activity of synaptically con-
nected neurons: if both neurons A and B are active at the same time, then the
synapses from A to B and B to A should be strengthened proportionally to the
product of the activity of A and B. However, as it stands, Hebb's learning rule
diverges. Therefore, various modification of Hebb's rule have been developed,
which basically take one of three forms (see [Gerstner and Kistler, 2002] and
[Dayan and Abbott, 2001]): first, a decay term can be added to the learning rule
so that each synaptic weight is able to "forget" what it previously learned. Sec-
ond, each synaptic modification can be normalized or projected on different sub-
spaces. These constraint–based rules may be interpreted as implementing some
form of competition for energy between dendrites and axons, see [Miller, 1996,
Miller and MacKay, 1996] and [Ooyen, 2001] for details. Third, a sliding thresh-
old mechanism can be added to Hebbian learning. For instance, a post-synaptic
threshold rule consists in multiplying the presynaptic activity and the subtraction
of the average postsynaptic activity from its current value, which is referred as
covariance learning ([Sejnowski and Tesauro, 1989]). Probably the best known
of these rules is the BCM rule [Bienenstock et al., 1982].
It should be noted
that history-based rules can also be defined without changing the qualitative dy-
namics of the system: instead of considering the instantaneous value of the neu-
rons' activity, these rules consider its weighted mean over a time window (see
[Foldi´ak, 1991, Wallis and Baddeley, 1997]). Recent experimental evidence sug-
gests that learning may also depend upon the precise timing of action potentials
[Bi and Poo, 2001]. Contrary to most Hebbian rules that only detect correlations,
these rules can also encode causal relationships in the patterns of neural activa-
tion. However, the mathematical treatment of these spike timing dependent rules
is much more difficult than rate based ones.
2
Hebbian-like learning rules have often been studied within the framework
of unsupervised feedfoward neural networks [Oja, 1982, Bienenstock et al., 1982,
Miller and MacKay, 1996, Dayan and Abbott, 2001]. They also form the basis of
most weight-based models of cortical development, assuming fixed lateral con-
nectivity (e.g. mexican hat) and modifiable vertical connections (see the review
of [Swindale, 1996])1.
In these developmental models, the statistical structure
of input correlations provides a mechanism for spontaneously breaking some un-
derlying symmetry of the neuronal receptive fields leading to the emergence of
feature selectivity. When such correlations are combined with fixed intracorti-
cal interactions, there is a simultaneous breaking of translation symmetry across
cortex leading to the formation of a spatially periodic cortical feature map. A re-
lated mathematical formulation of cortical map formation has been developed in
[Takeuchi and Amari, 1979, Bressloff, 2005] using the theory of self–organizing
neural fields. Although very irregular, the two-dimensional cortical maps ob-
served at a given stage of development, can be unfolded in higher dimensions
to get smoother geometrical structures. Indeed, [Bressloff et al., 2001] suggested
that the network of orientation pinwheels in V1 is a direct product between a circle
for orientation preference and a plane for position, based on a modification of the
icecube model of Hubel and Wiesel [Hubel and Wiesel, 1977]. From a more ab-
stract geometrical perspective, Petitot [Petitot, 2003] has associated such a struc-
ture to a 1-jet space and used this to develop some applications to computer vision.
More recently, [Bressloff and Cowan, 2003] and [Chossat and Faugeras, 2009] have
considered more complex geometrical structures such as spheres and hyperbolic
surfaces that incorporate additional stimulus features such as spatial frequency
and textures, respectively.
In this paper, we show how geometrical structures related to the distribution
of inputs can emerge through unsupervised Hebbian learning applied to recur-
rent connections in a rate-based Hopfield network. Throughout this paper, the
inputs are presented as an external non-autonomous forcing to the system and not
an initial condition as is often the case in Hopfield networks. It has previously
been shown that, in the case of a single fixed input, there exists an energy func-
tion that describes the joint gradient dynamics of the activity and weight variables
[Dong and Hopfield, 1992]. This implies that the system converges to an equilib-
rium during learning. We use averaging theory to generalize the above result to
the case of multiple inputs, under the adiabatic assumption that Hebbian learning
1There have only been a few computational studies that consider the joint development of
lateral and vertical connections [Bartsch and Van Hemmen, 2001, Miikkulainen et al., 2005].
3
occurs on a much slower time scale than both the activity dynamics and the sam-
pling of the input distribution. We then show that the equilibrium distribution of
weights, when embedded into Rk for sufficiently large integer k, encodes the geo-
metrical structure of the inputs. Finally, we numerically show that the embedding
of the weights in two dimensions (k = 2) gives rise to patterns that are quali-
tatively similar to experimentally observed cortical maps, with the emergence of
features columns and patchy connectivity. Although the mathematical formalism
we introduce here could be extended to most of the rate-based Hebbian rules in
the literature, we present the theory for Hebbian learning with decay because of
the simplicity of the resulting dynamics.
Note that the use of geometrical objects to describe the emergence of connec-
tivity patterns has previously been put forward by Amari in a different context.
Based on the theory of information geometry, Amari considers the geometry of
the set of all the networks and defines learning as a trajectory on this manifold for
perceptron networks in the framework of supervised learning [Amari, 1998] or
for unsupervised Boltzmann Machines [Amari et al., 1992]. He uses differential
and Riemannian geometry to describe an object which is at a larger scale than the
cortical maps this paper is focusing on.
Moreover, Zucker and colleagues are currently developing a non-linear dimen-
sionality reduction approach to caracterize the statistics of natural visual stimuli
(see [Lawlor and Zucker, 2010, Coifman et al., 2005]). Although they do not use
learning neural networks and stay closer to the field of computer vision than this
paper, it turns out their approach is similar to the geometrical embedding approach
we are using.
The structure of the paper is as follows. In section 2, we formally introduce
the model. We derive the averaged system in section 3, which then allows us
to study the stability of the learning dynamics in the presence of multiple inputs
by constructing an appropriate energy function. We adress stability in section
4. In section 5 we determine the geometrical structure of the equilibrium weight
distribution and show how it reflects the structure of the inputs. We also relate this
approach to the emergence of cortical maps. Finally, the results are discussed in
section 6.
4
2 Model
2.1 Neural network evolution
A neural mass corresponds to a mesoscopic coherent group of neurons. It is con-
venient to consider them as building blocks, first for computational simplicity,
second for their direct relationship to macroscopic measurements of the brain
(EEG, MEG and Optical imaging) which average over numerous neurons, and
third because one can functionally define coherent groups of neurons within cor-
tical columns. For each neural mass i ∈ {1..N}, define the mean membrane
potential Vi(t) at time t. The instantaneous population firing rate νi(t) is linked to
the membrane potential through the relation νi(t) = s(cid:0)Vi(t)(cid:1), where s is a smooth
sigmoid function. In the following, we choose
s(v) =
Sm
1 + exp(cid:0) − 4S′
m(v − φ)(cid:1) ,
where Sm, S′
and the offset of the sigmoid.
m and φ are respectively the maximal firing rate, the maximal slope
Consider a Hopfield network of neural masses described by the equation
(1)
(2)
dVi
dt
(t) = −αVi(t) +
NXj=1
Wij(t) s(cid:0)Vj(t)(cid:1) + Ii(t).
The first term roughly corresponds to the intrinsic dynamics of the neural mass:
it decays exponentially to zero at a rate α if it receives neither external inputs nor
spikes from the other neural masses. We will fix the units of time by setting α =
1. The second term corresponds to the rest of the network sending information
through spikes to the given neural mass i, with Wij(t) the effective synaptic weight
from neural mass j. The synaptic weights are time–dependent because they evolve
according to a continuous time Hebbian learning rule (see below). The third term
Ii(t) corresponds to an external input to neural mass i, e.g. information extracted
by the retina or thalamo-cortical connections. We take the inputs to be piecewise
constant in time, that is, at regular time intervals a new input is presented to the
network. In this paper, we will assume that the inputs are chosen by peridodically
cycling through a given set of M inputs. An alternative approach would be to
randomly select each input from a given probability distribution [Geman, 1979].
It is convenient to introduce vector notation by representing the time–dependent
membrane potentials by V ∈ C 1(R+, RN ), the time–dependent external inputs
5
by I ∈ C 0(R+, RN ), and the time–dependent network weight matrix by W ∈
C 1(R+, RN ×N ). We can then rewrite the above system of ordinary differential
equations as a single vector-valued equation
dV
dt
= −V + W · S(V ) + I,
(3)
where S : RN → RN corrresponds to the term by term application of the sigmoid
S, i.e. S(V )i = s(Vi).
2.2 Correlation-based Hebbian learning
The synaptic weights are assumed to evolve according to a correlation–based Heb-
bian learning rule of the form
dWij
dt
= ǫ(s(Vi)s(Vj) − µWij),
(4)
where ǫ is the learning rate, and we have included a decay term in order to stop the
weights from diverging. In order to rewrite the above equation in a more compact
vector form, we introduce the tensor (or Kronecker) product S(V ) ⊗ S(V ) so that
in component form
[S(V ) ⊗ S(V )]ij = S(V )iS(V )j,
(5)
where S is treated as a mapping from RN to RN . The tensor product implements
Hebb's rule that synaptic modifications involve the product of postynaptic and
presynaptic firing rates. We can then rewrite the combined voltage and weight
dynamics as the following non–autonomous (due to time–dependent inputs) dy-
namical system:
Σ :
dV
dt
dW
dt
= −V + W · S(V ) + I
= ǫ(cid:16)S(V ) ⊗ S(V ) − µW(cid:17).
(6)
Let us make few remarks about the existence and uniqueness of solutions.
First, boundedness of S implies boundedness of the system Σ. More precisely, if
I is bounded, then the solutions are bounded. To prove this, note that the right
hand side of the equation for W is the sum of a bounded term and a linear decay
6
term in W . Therefore, W is bounded and hence the term W ·S(V ) is also bounded.
The same reasoning applies to V . S being Lipschitz continuous implies that the
right hand side of the system is Lipschitz. This is sufficient to prove existence
and uniqueness of the solution by applying the Cauchy-Lipschitz theorem. In the
following, we will derive an averaged autonomous dynamical system Σ′, which
will be well-defined for the same reasons.
3 Averaging the system
We will show that system Σ can be approximated by an autonomous Cauchy prob-
lem which will be much more convenient to handle. This averaging method makes
the most of multiple time–scales in the system. First, it is natural to consider that
learning occurs on a much slower time–scale than the evolution of the membrane
potentials (as determined by α), i.e.
ǫ ≪ 1.
(7)
Second, an additional time-scale arises from the rate at which the inputs are sam-
pled by the network. That is, the network cycles periodically through M fixed
inputs, with the period of cycling given by T . It follows that I is T –periodic,
piecewise constant. We assume that the sampling rate is also much slower than
the evolution of the membrane potentials,
M
T
≪ 1.
(8)
Finally, we assume that the period T is small compared to the time-scale of the
learning dynamics,
ǫ ≪
.
(9)
1
T
We can now simplify the system Σ by applying Tikhonov's theorem for slow/fast
systems, and then classical averaging methods for periodic systems.
3.1 Tikhonov's theorem
Tikhonov's theorem ([Tikhonov, 1952] and [Verhulst, 2007] for a clear introduc-
tion) deals with slow/fast systems. It says the following:
7
Theorem 3.1. Consider the initial value problem
x = f (x, y, t), x(0) = x0, x ∈ Rn, t ∈ R+
ǫ y = g(x, y, t), y(0) = y0, y ∈ Rm
Assume that:
1. A unique solution of the initial value problem exists and we suppose, this
holds also for the reduced problem
x = f (x, y, t), x(0) = x0
0 = g(x, y, t)
with solutions ¯x(t), ¯y(t).
2. The equation 0 = g(x, y, t) is solved by ¯y(t) = φ(x, t), where φ(x, t) is a
continuous function and an isolated root. Also suppose that ¯y(t) = φ(x, t)
is an asymptotically stable solution of the equation dy
dτ = g(x, y, τ ) that is
uniform in the parameters x ∈ Rn and t ∈ R+.
3. y(0) is contained in an interior subset of the domain of attraction of ¯y.
Then we have
limǫ→0 xǫ(t) = ¯x(t), 0 ≤ t ≤ L
limǫ→0 yǫ(t) = ¯y(t), 0 ≤ d ≤ t ≤ L
with d and L constants independent of ǫ.
In order to apply Tikhonov's theorem directly to the system Σ, we first need
to rescale time according to t → ǫt. This gives
ǫ
dV
dt
dW
dt
= −V + W · S(V ) + I
= S(V ) ⊗ S(V ) − µW.
Tikhonov's theorem then implies that solutions of Σ are close to solutions of the
reduced system (in the unscaled time variable)
( V (t) = W · S(cid:0)V (t)(cid:1) + I(t)
W = ǫ(cid:16)S(V ) ⊗ S(V ) − µW(cid:17),
8
(10)
provided that the dynamical systems Σ in equation (6), and equation (10) are well
defined. It is easy to show that both systems are Lipschitz because of the properties
of S. Following [Faugeras et al., 2008], we know that if
S′
mkW k < 1,
(11)
then there exists an isolated root ¯V : R+ → RN of the equation V = W · S(V ) + I
and ¯V is asymptotically stable. Equation (11) corresponds to the weakly con-
nected case. Moreover, the initial condition belongs to the basin of attraction of
this single fixed point. Note that we require M
T ≪ 1 so that the membrane po-
tentials have sufficient time to approach the equilibrium associated with a given
input before the next input is presented to the network. In fact, this assumption
make it reasonable to neglect the transient activity dynamics due to the switching
between inputs.
3.2 Periodic averaging
The system given by equation (10) corresponds to a differential equation for W
with T -periodic forcing due to the presence of V on the right–hand side. Since
T ≪ ǫ−1, we can use classical averaging methods to show that solutions of (10)
are close to solutions of the following autonomous system on the time-interval
[0, 1
ǫ ] (which we suppose large because ǫ << 1)
V (t) = W · S(V (t)) + I(t)
Σ0 :
dW
dt
T Z T
= ǫ(cid:16) 1
0
S(V (s)) ⊗ S(V (s))ds − µW (t)(cid:17).
(12)
It follows that solutions of Σ are also close to solutions of Σ0. Finding the explicit
solution V (t) for each input I(t) is difficult and requires fixed points methods,
e.g. a Picard algorithm. Therefore, we will consider yet another system Σ′ whose
solutions are also close to Σ0 and hence Σ. In order to construct Σ′ we need to
introduce some additional notation.
Let us label the M inputs by I (a), a = 1, . . . , M and denote by V (a) the fixed
point solution of the equation V (a) = W · S(V (a)) + I (a). Given the periodic
9
sampling of the inputs, we can rewrite (12) as
V (a) = W · S(V (a)) + I (a)
dW
dt
= ǫ(cid:16) 1
M
MXa=1
S(V (a)) ⊗ S(V (a)) − µW (t)(cid:17).
(13)
If we now introduce the N × M matrices V and I with components Via = V (a)
and Iia = I (a)
, then we can eliminate the tensor product and simply write (13) in
the matrix form
i
i
V
= W · S(V) + I
dW
dt
= ǫ(cid:16) 1
M
S(V) · S(V)T − µW (t)(cid:17),
where S(V) ∈ RN ×M such that [S(V)]ia = s(V (a)
). A second application of
Tikhonov's theorem (in the reverse direction) then establishes that solutions of the
system Σ0 (written in the matrix form (14)) are close to solutions of the matrix
system
i
(14)
(15)
Σ′ :
dV
dt
dW
dt
= −V + W · S(cid:0)V(cid:1) + I
= ǫ(cid:16) 1
S(V) · S(V)T − µW (t)(cid:17)
M
m
µ . Therefore, equation (11) says that if S2
In the remainder of the paper we will focus on the system Σ′ whose solutions
are close to those of the original system Σ provided condition (11) is satisfied,
i.e. the network is weakly connected. Clearly, the fixed points (V ∗, W ∗) of sys-
tem Σ satisfy kW ∗k ≤ S2
mS ′
µ < 1 then
m
Tikhonov's theorem can be applied and systems Σ and Σ′ can be reasonably con-
sidered as good approximations of each other. The advantage of the averaged sys-
tem Σ′ is that is given by autonomous ordinary differential equations. Moreover,
since it is Lipschitz continuous, it leads to a well-posed Cauchy problem. Finally,
note that it is straighforward to extend our approach to time-functional rules (e.g.
sliding threshold or BCM rules as described in [Bienenstock et al., 1982]) which,
in this new framework, would be approximated by simple ordinary differential
equations (as opposed to time-functional differential equations) provided S is re-
defined appropriately.
10
1
5
10
25
50
100
250
error in %
105
90
75
60
45
30
15
0
Figure 1: Percentage of error between final connectivities for the exact and aver-
aged system.
3.3 Simulations
To illustrate the above approximation, we simulate a simple network with both
exact, i.e. Σ, and averaged ,i.e. Σ′, evolution equations. For these simulations,
the network consists of N = 10 fully-connected neurons and is presented with
M = 10 different random inputs taken uniformly in the intervals [0, 1]N. For this
simulation we use s(x) =
1+e−4(x−1) , and µ = 10. Figure 1. shows the percentage
of error between final connectivities for different values of ǫ and T /M. Figure 2
shows the temporal evolution of the norm of the connectivity for both the exact
and averaged system for T = 103 and ǫ = 10−3.
1
4 Stability
4.1 Liapunov function
In the case of a single fixed input (M = 1), the systems Σ and Σ′ are equivalent
and reduce to the neural network with adapting synapses previously analyzed by
[Dong and Hopfield, 1992]. Under the additional constraint that the weights are
symmetric (Wij = Wji), these authors showed that the simultaneous evolution of
the neuronal activity variables and the synaptic weights can be re-expressed as a
11
0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
averaged system
exact system
0
5000
10000
15000
20000
Figure 2: Temporal evolution of the norm of the connectivities of the exact system
Σ and averaged system Σ′.
gradient dynamical system that minimizes a Liapunov or energy function of state.
We can generalize their analysis to the case of multiple inputs (M > 1) and non-
symmetric weights using the averaged system Σ′. That is, following along similar
lines to [Dong and Hopfield, 1992], we introduce the energy function
E(U, W ) = −
1
2
where U = S(V), kW k2 = hW, W i =Pi,j W 2
ij,
hU, W · Ui =
hU, W · Ui − hI, Ui + h1, S−1(cid:0)U(cid:1)i +
MXa=1
MXa=1
h1, S−1(cid:0)U(cid:1)i =
NXi=1Z U (a)
U (a)
i WijU (a)
NXi=1
MXa=1
S−1(ξ)dξ.
,
j
hI, Ui =
i
0
Mµ
2
kW k2
(16)
I (a)
i U (a)
i
(17)
NXi=1
(18)
and
In contrast to [Dong and Hopfield, 1992], we do not require a priori that the
weight matrix is symmetric. However, it can be shown that the system always
converges to a symmetric connectivity pattern. More precisely, A =n(V, W ) ∈
RN ×M × RN ×N : W = W To is an attractor of the system Σ′. A proof can be
found in appendix 8.1. It can then be shown that on A (symmetric weights), E is
a Liapunov function of the dynamical system Σ′, that is,
dE
dt
≤ 0,
and
dE
dt
= 0 =⇒
dY
dt
= 0, Y = (V, W )T .
12
The boundedness of E and the Krasovskii-LaSalle invariance principle then im-
plies that the system converges to an equilibrium [Khalil and Grizzle, 1996]. We
thus have
Theorem 4.1. The initial value problem for the system Σ′ with Y(0) ∈ H, con-
verges to an equilibrium state.
Proof. See appendix 8.2
It follows that neither oscillatory nor chaotic attractor dynamics can occur.
4.2 Linear stability
Although we have shown that there are stable fixed points, not all of the fixed
points are stable. However, we can apply a linear stability analysis on the system
Σ′ to derive a simple sufficient condition for a fixed point to be stable. The method
we use in the proof could be extended to more complex rules. The proof reveals
the significant role played by the Kronecker product in Hebbian learning.
Theorem 4.2. The equilibria of system Σ′ satisfy:
(cid:26) V ∗ = 1
W ∗ = 1
µM S(V ∗) · S(V ∗)T · S(V ∗) + I
µM S(V ∗) · S(V ∗)T
and a sufficient condition for stability is
3S′
mkW ∗k < 1
provided 1 > ǫµ which is most probably the case since ǫ << 1.
Proof. See appendix 8.3
(19)
(20)
This condition is strikingly similar to that derived in [Faugeras et al., 2008]. In
fact, condition (20) is stronger than the contracting condition (11). It says the
network may converge to a weakly connected situation. It justifies the averag-
ing method by saying that we remain in the domain of validity of the averaging
method. It also says that the dynamics of V is likely (because the condition is
only sufficient) to be contracting and therefore subject to no bifurcations: a fully
recurrent learning neural network is likely to have a "simple" dynamics.
13
5 Geometrical structure of equilibrium points
5.1 Learning the correlation matrix of the inputs
It follows from equation (19) that the equilibrium weight matrix W ∗ is given by
the correlation matrix of the firing rates. Moreover, in the case of sufficiently
large inputs, the matrix of equilibrium membrane potentials satisfies V ∗ ≈ I.
More precisely, if S(I (a)
for all a = 1, . . . , M and i = 1, . . . , N, then
we can generate an iterative solution for V ∗ of the form
) ≪ I (a)
i
i
V ∗ = I +
1
µ
S(cid:0)I(cid:1) · S(cid:0)I(cid:1)T · S(cid:0)I(cid:1) + h.o.t.
On the other hand, if the inputs are comparable in size to the synaptic weights,
then there is no explicit solution for V ∗. Roughly speaking, we observe that the
connection term has the role of "smoothing" the solution. Therefore, if a Gaussian
is presented to the network (as the only input), the membrane potential is likely
to be another Gaussian with a larger variance.
If no input is presented to the
network (I = 0), then S(0) 6= 0 implies that the activity is non-zero, that is, there
is spontaneous activity. Combining these observations, we see that the network
roughly extracts and stores the correlation matrix of the strongest inputs within
the weights of the network.
5.2 From a symmetric connectivity matrix to a convolutional
network
So far neurons have been identified by a label i ∈ {1..N}; there is no notion
of geometry or space in the preceding results. However, as we show below, the
inputs may contain a spatial structure that can be encoded by the connectivity. In
this section, we propose a mechanism to unveil the hidden geometrical structure
within the connectivity. More specifically, we want to find an integer k ∈ N and N
points in Rk, denoted by xi, i ∈ {1, . . . , N}, so that the connectivity can roughly
be written as W ∗
ij ≃ exp(−kxi − xjk2). In other words, we interpret the final
connectivity as a matrix describing the distance between the neurons living in a
k-dimensional space. However, W ∗ is not always a distance matrix, therefore, it
is natural to project it on the set of distance matrices. Finding the best fit of W ∗ to
a distance matrix is usually called multidimensional scaling. This set of methods
is reviewed in [Borg and Groenen, 2005].
14
max the
max. Second, define a bijective kernel function on x ∈ R+ such that
largest component of W ∗: W ∗
Wii/W ∗
K(x) = W ∗
First, definecW ∈ RN ×N as W ∗ whose diagonal terms are set to W ∗
ij = NijcWij with Nij = 1 if i 6= j and Nii =
maxe−x/σ2. Given thatcW is non-negative, we define the matrix D =
K −1(cW ) corresponding to the application of the inverse of K to each component
ofcW . As said before, we want to find k ∈ N and xi ∈ Rk for i ∈ {1, . . . , N}
However, we can compute the projection of the symmetric matrixcW onto the set
of distance matrices by applying multidimensional scaling methods as described
in [Borg and Groenen, 2005]. We use the stress majorization or SMACOF algo-
rithm for the stress1 cost function throughout the article. In other words, we can
find the distance matrix D⊥ such that kDqk = kD − D⊥k is minimal. Therefore,
so that D is a distance matrix, Dij = kxi − xjk2. In general, this is not possible.
cWij = K(Dq)ijK(D⊥)ij. Define M such that M(xi, xj) = K(Dq)ij Nij and Gσ
a Gaussian with a standard deviation equal to σ. In spatial coordinates
W ∗(xi, xj) = M(xi, xj) Gσ(kxi − xjk)
(21)
Multidimensional scaling methods consist in minimizing the contribution of M in
the preceding equation. Hence, we refer to M as the non-convolutional connec-
tivity.
A position xi ∈ Rk is associated to each neuron i ∈ {1, . . . , N} such that
(cid:0)W · S(V )(cid:1)i =
W ∗
ijS(Vj) =
NXj=1
NXj=1
M(xi, xj) Gσ(kxi − xjk) S(V (xj)) (22)
In particular, we can assume that k is large enough for kDqk to be very small.
Moreover, if the neurons are equally excited on average (i.e. the diagonal of W ∗
is already equal to W ∗
maxId), then it is reasonable to consider that M(xi, xj) = 1
leading to the following convolutional product
W · S(V ) = Gσ(k.k) ∗ S(V )
Therefore, in the space defined by the xi ∈ Rk the connectivity is close to being
convolutional.
5.3 Unveiling the geometrical structure of the inputs
We hypothesize that the space defined by the xi reflects the underlying geometrical
structure of the inputs. We have not found a way to prove this so we only provide
15
numerical examples that illustrate this claim. In the following examples, we relate
the geometry of the manifold suggested by the xi to the network inputs. Thus we
feed the network with inputs having a defined geometrical structure and then show
how this structure can be extracted from the connectivity by the method above.
However, it is by extracting the structure from unknown inputs that these networks
might reveal themselves useful. Therefore, the following is only a (numerical)
proof of concept.
We assume the inputs to be uniformly distributed over a manifold Ω with fixed
geometry. This strong assumption amounts to considering that the feedforward
connectivity (which we do not consider here) has already properly filtered the
information coming from the sensory organs. More precisely, define the set of
inputs by the matrix I ∈ RN ×M such that I (a)
i = f (kyi − zakΩ) where the za are
uniformly distributed points over Ω, the yi are the positions on Ω that "label" the
ith neuron, and f is a decreasing function on R+. The norm k.kΩ is the natural
norm defined over the manifold Ω. For simplicity, assume f (x) = fσ(x) = Ae− x2
so that the inputs are localized bell-shaped bumps on the shape Ω.
σ2
5.3.1 Planar retinotopy
0
0
20
40
60
80
20
40
60
80
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0
0
20
40
60
80
20
40
60
80
0.0032
0.0028
0.0024
0.0020
0.0016
0.0012
0.0008
0.0004
Figure 3: Plot of planar retinotopic inputs on Ω = [0, 1]×[0, 1] (left) and final con-
nectivity matrix of the system Σ′ (right). The parameters used for this simulation
are s(x) =
1+e−4(x−1) , l = 1, µ = 10, ǫ = 0.001, N = M = 100, σ = 4.
1
16
We consider a set of Gaussian inputs uniformly distributed over a two-dimensional
plane, e.g. Ω = [0, 1] × [0, 1]. For simplicity, we take N = M = K 2 and set
za = yi for i = a, a ∈ {1, . . . , M}. (The numerical results show an identical
structure for the final connectivity when the yj correspond to random points, but
the analysis is harder). In the simpler case of one-dimensional Gaussians with
N = M = K, the input matrix takes the form I = Tfσ, where Tf is a symmetric
Toeplitz matrix:
f (0)
f (1)
f (2)
...
f (1)
f (0)
f (1)
...
f (2)
f (1)
f (0)
...
f (K) f (K − 1) f (K − 2)
· · ·
f (2)
f (1)
...
· · ·
· · ·
· · ·
· · ·
...
· · ·
f (K)
f (K − 1)
f (K − 2)
...
f (0)
(23)
Tf =
It follows that I (a)
In the two-dimensional case, we set y = (u, v) ∈ Ω and introduce the labeling
yk+(l−1)K = (uk, vl) for k, l = 1, . . . K.
i ∼ exp(−(uk −
uk′)2) exp(−(vl − vl′)2) for i = k + (l − 1)K and a = k′ + (l′ − 1)K. Hence,
we can write I = Tfσ ⊗ Tfσ, where ⊗ is the Kronecker product; the Kronecker
product is responsible for the K ×K sub-structure we can observe in figure 3 with
K = 10. Note that if we were interested in a n-dimensional retinotopy, then the
input matrix could be written as a Kronecker product between n Toeplitz matrices.
As previously mentioned, the final connectivity matrix roughly corresponds to the
correlation matrix of the input matrix. It turns out that the correlation matrix of I
is also a Kronecker product of two Toeplitz matrix generated by a single Gaussian
(with a different standard deviation). Thus, the connectivity matrix has the same
basic form as the input matrix when za = yi for i = a. The inputs and stable
equilibrium points of the simulated system are shown in figure 3. The positions
xi of the neurons after multidimensional scaling are shown in figure 4.
5.3.2 Toroıdal retinotopy
We now assume that the inputs are uniformly distributed over a two-dimensional
torus, i.e. Ω = T2. That is, the input labels za are randomly distributed on the
torus. The neuron labels yi are regularly and uniformly distributed on the torus.
The inputs and final stable weight matrix of the simulated system are shown in
figure 5. The positions xi of the neurons after multidimensional scaling for k = 3
are shown in figure 6, and appear to form a cloud of points distributed on a torus.
In order to confirm this, we have used a numerical method from computational co-
homology [Zomorodian and Carlsson, 2005] to construct the so–called persistent
17
Figure 4: Positions xi of the neurons after having applied classical multidimen-
sional scaling to the final connectivity matrix shown in figure 3 for k = 2 (left)
and k = 3 (right). The regular spacing of the neurons for k = 2 shows that
the planar structure of the inputs has been recovered, although the corner of the
square appear less regular due to boundary effects. In the case k = 3, there is an
embedding of the plane into three dimensions; the saddle–like shape accounts for
the corner irregularity observed when k = 2.
Figure 5: Plot of retinotopic inputs on Ω = T2 (left) and the final connectivity
matrix (right) for the system Σ′. The parameters used for this simulation are
s(x) =
1+e−4(x−1) , l = 1, µ = 10, ǫ = 0.001, N = 1000, M = 10, 000, σ = 2.
1
18
Figure 6: Left: Positions xi of the neurons for k = 3 after having applied multi-
dimensional scaling methods presented in part 5.2 to the final connectivity matrix
shown in figure 5. Right: Persistent cohomology barcodes of the cloud of points xi
computed using the Jplex software package of [Sexton and Vejdemo-Johansson, ].
(See section 8.4 for a short introduction to persistent cohomology). The triplet of
betti numbers (1,2,1) appear stable confirming that the points lie on a 2 dimen-
sional torus.
19
cohomology barcodes of the neurons' positions. These determine certain topolog-
ical invariants of the underlying space. (See section 8.4 for a short introduction
to persistent cohomology and barcodes). The results are also shown in figure 6,
and establish that the network has learnt the underlying toroidal geometry of the
inputs.
5.4 Links with neuroanatomy
The brain is subject to energy constraints which are completely neglected in the
above formulation. These constraints most likely have a significant impact on the
positions of real neurons in the brain. Indeed, it seems reasonable to assume that
the positions and connections of neurons reflect a trade-off between the energy
costs of biological tissue and their need to process information effectively. For
instance, it has been suggested that a principle of wire length minimization may
occur in the brain [Swindale, 1996, Chklovskii et al., 2002]. In our neural mass
framework, one may consider that the stronger two neural masses are connected,
the larger the number of real axons linking the neurons together. Therefore, mini-
mizing axonal length can be read as: the stronger the connection the closer, which
is consistent with the convolutional part of the weight matrix. However, the un-
derlying geometry of natural inputs is likely to be very high-dimensional, whereas
the brain lies in a three-dimensional world. In fact, the cortex is so flat that it is
effectively two-dimensional. Hence, the positions of real neurons are different
from the positions xi ∈ Rk in a high dimensional vector space; since the cor-
tex is roughly two-dimensional, the positions could only be realized physically if
k = 2. Therefore, the three-dimensional toric geometry or any higher dimensional
structure could not be perfectly implemented in the cortex without the help of non-
convolutional long-range connections. Indeed, we suggest that the cortical con-
nectivity is made of two parts: i) a local convolutional connectivity corresponding
to the convolutional term Gσ in (21), which is consistent with the requirements of
energy efficiency, and ii) a non-convolutional connectivity corresponding to the
factor M in equation (21), which is required in order to represent various stimulus
features. If the cortex were higher-dimensional (k ≫ 2) then M ≡ 1.
We illustrate the above claim by considering two examples based on the func-
tional anatomy of the primary visual cortex: the emergence of ocular dominance
columns and orientation columns, respectively. We proceed by returning to the
case of planar retinotopy (section 5.3.1) but now with additional input structure.
In the first case, the inputs are taken to be binocular and isotropic, whereas in
the second case they are taken to be monocular and anisotropic. The details are
20
presented below. Given a set of prescribed inputs, the network evolves accord-
ing to equation (15) and the lateral connections converge to a stable equilibrium.
The resulting weight matrix is then projected onto the set of distance matrices for
k = 2 (as described in section 5.2) using the stress majorization or SMACOF
algorithm for the stress1 cost function as described in [Borg and Groenen, 2005].
We thus assign a position xi ∈ R2 to the ith neuron, i = 1, . . . , N. (Note that the
position xi extracted from the weights using multidimensional scaling is distinct
from the "physical" position yi of the neuron in the retinocortical plane; the latter
determines the center of its receptive field). The convolutional connectivity (Gσ
in equation 21) is therefore completely defined: on the planar map of points xi,
neurons are isotropically connected to their neighbors; the closer the neurons are
the stronger is their convolutional connection. Moreover, since the stimulus fea-
ture preferences (orientation, ocular dominance) of each neuron i, i = 1, . . . , N,
are prescribed, we can superimpose these feature preferences on to the planar
map of points xi. In both examples, we find that neurons with the same ocular
or orientation selectivity tend to cluster together (see figures 7 and 8): interpolat-
ing these clusters then generates corresponding feature columns. It is important
to emphasize that the retinocortical positions yi do not have any columnar struc-
ture, that is, they do not form clusters with similar feature preferences. Thus, in
contrast to standard developmental models of vertical connections, the columnar
structure emerges from the recurrent weights following Hebbian learning and an
application of multidimensional scaling. It follows that neurons coding for the
same feature tend to be strongly connected; indeed, the multidimensional scaling
algorithm has the property that it positions strongly connected neurons close to-
gether . Equation (21) also suggests that the connectivity has a non-convolutional
part, M, which is a consequence of the low-dimensionality (k = 2). In order to
illustrate the structure of the non-convolutional connectivity, we select a neuron i
in the plane and draw a link from it at position xi to the neurons at position xj for
which M(xi, xj) is maximal. We find that M tends to be patchy, i.e. it connects
neurons having the same feature preferences. In the case of orientation, M also
tends to be co-aligned, i.e. connecting neurons with similar orientation preference
along a vector in the plane of the same orientation.
5.4.1 Ocular dominance columns and patchy connectivity
In order to construct binocular inputs, we partition the N neurons into two sets
i ∈ {1, . . . , N/2} and i ∈ {N/2 + 1, . . . , N} that code for the left and right eyes,
respectively. The ith neuron is then given a retinocortical position yi ∈ [0, 1] ×
21
[0, 1], with the yi uniformly distributed across the plane. We do not assume a
priori that there exist any ocular dominance columns, that is, neurons with similar
retinocortical positions yi do not form clusters of cells coding for the same eye.
We then take the ath input to the network to be of the form
i = (1 + γ(a))e− (yi−za)2
I (a)
,
i = (1 − γ(a))e− (yi−za+s)2
I (a)
σ′2
σ′2
i = 1, . . . , N/2
,
i = N/2 + 1, . . . , N,
where s ∈ R2 represents some form of binocular disparity, za and γ(a) are
randomly generated from [0, 1]2 and [−1, 1], respectively, see [Bressloff, 2005].
Thus, if γ(a) > 0 (γ(a) < 0) then the corresponding input is predominantly from
the left (right) eye. In our simulations we take σ = σ′ = 0.1 and s = 0.005. The
results of our simulations are shown in figure 7. In particular, we plot the points xi
obtained by performing multidimensional scaling on the final connectivity matrix
for k = 2, and superimposing upon this the ocular dominance map obtained by
interpolating between clusters of neurons with the same eye preference. We also
illustrate the non-convolutional connectivity by linking one selected neuron to the
five neurons labeled j it is most strongly connected to (with M(xi, xj) > 1), with
i the label of the central neuron. This clearly shows that long–range connections
tend to link cells with the same ocular dominance.
5.4.2 Orientation columns and colinear connectivity
In order to construct oriented inputs, we partition the N neurons into four groups
Σθ corresponding to different orientation preferences θ = {0, π
4 }. Thus,
if neuron i ∈ Σθ then its orientation preference is θi = θ. For each group, the
neurons are randomly assigned a retinocortical position yi ∈ [0, 1] × [0, 1]. Again,
we do not assume a priori that there exist any orientation columns, that is, neurons
with similar retinocortical positions yi do not form clusters of cells coding for the
same orientation preference. Each cortical input I (a)
is generated by convolving a
thalamic input consisting of an oriented Gaussian with a Gabor–like receptive field
[Miikkulainen et al., 2005]. Let Rθ denote a 2-dimensional rigid body rotation in
the plane with θ ∈ [0, 2π). Then
2 , 3π
4 , π
i
where
I (a)
i =Z Gi(ξ − yi)Ia(ξ − za)dξ,
Gi(ξ) = G0(Rθiξ)
22
(24)
(25)
Figure 7: Plot of the positions xi of neurons for k = 2 in red (right eye) and green
(left eye). We have used an interpolation method to highlight the areas dominated
by the right eye in black. These ocular dominance columns have fractal borders
which are less regular than those observed in optical imaging experiments. The
convolutional connectivity (Gσ in equation (21)) is implicitly described by the po-
sition of the neurons: the closer the neurons, the stronger their connections. The
strongest components of the non-convolutional connectivity (M in equation (21))
from a central red neuron are also shown by drawing links from this neuron to
the target neurons. The color of the link refers to the color of the target neuron.
Therefore, we see that it is mainly connected to neurons of its same ocular dom-
inance resulting in a patchy distribution. The parameters used for this simulation
are s(x) =
1+e−4(x−1) , l = 1, µ = 10, ǫ = 0.01, N = 800 M = 3200.
1
23
and G0(ξ) is the Gabor–like function
G0(ξ) = A+e−ξT .Λ−1.ξ − A−e−(ξ−e0)T .Λ−1.(ξ−e0) − A−e−(ξ+e0)T .Λ−1.(ξ+e0)
with e0 = (0, 1) and
Λ =(cid:18)σlarge
0
0
σsmall(cid:19) .
The amplitudes A+, A− are chosen so thatR G0(ξ)dξ = 0. Similarly, the thalamic
aξ) with I(ξ) the anisotropic Gaussian
input Ia(ξ) = I(Rθ′
I(ξ) = e−ξT .Λ′−1.ξ,
Λ′ =(cid:18)σ′
large
0
0
small(cid:19) .
σ′
a and za are randomly generated from [0, π) and [0, 1]2
The input parameters θ′
large = 0.266... and
respectively. In our simulations we take σlarge = 0.133..., σ′
small = 0.0333.... The results of our simulations are shown in figure 8.
σsmall = σ′
In particular, we plot the points xi obtained by performing multidimensional scal-
ing on the final connectivity matrix for k = 2, and superimposing upon this the
orientation preference map obtained by interpolating between clusters of neurons
with the same orientation preference. To avoid border problems we have zoomed
on the center on the map. We also illustrate the non-convolutional connectiv-
ity by linking one selected neuron to all other neurons for which M is maximal.
The patchy, anisotropic nature of the long–range connections is clearly seen. The
anisotropic nature of the connections is further quantified in the histogram of fig-
ure 9.
6 Discussion
In this paper, we have shown how a neural network can learn the underlying geom-
etry of a set of inputs. We have considered a fully recurrent neural network whose
dynamics is described by a simple non-linear rate equation, together with unsu-
pervised Hebbian learning with decay that occurs on a much slower time scale.
Although several inputs are periodically presented to the network, so that the re-
sulting dynamical system is non-autonomous, we have shown that such a system
has a fairly simple dynamics: the network connectivity matrix always converges
to an equilibrium point. We have then demonstrated how this connectivity matrix
can be expressed as a distance matrix in Rk for sufficiently large k, which can be
24
Figure 8: Plot of the positions xi of neurons for k = 2 obtained by multidi-
mensional scaling of the weight matrix. Neurons are clustered in orientation
columns represented by the colored areas, which are computed by interpolation.
The strongest components of the non-convolutional connectivity (M in equation
(21)) from a particular neuron in a yellow area are illustrated by drawing black
links from this neuron to the target neurons. Since the yellow color corresponds
to an orientation of 3π
4 , the non-convolutional connectivity shows the existence
of a co-linear connectivity as exposed in [Bosking et al., 1997]. The parameters
used for this simulation are s(x) =
1+e−4(x−1) , l = 1, µ = 10, ǫ = 0.01, N = 900
M = 9000.
1
25
140
120
100
80
60
40
20
0
Figure 9: Histogram of the 5 largest components of the non-convolutional connec-
tivity for 80 neurons randomly chosen among those shown in Fig. 8. The abcissa
corresponds to the difference in radian between the direction preference of the
neuron and the direction of the links between the neuron and the target neurons.
This histogram is weighted by the strength of the non-convolutional connectiv-
ity. It shows a preference for co-aligned neurons but also a slight preference for
perpendicularly-aligned neurons (e.g. neurons of the same orientation but parallel
to each other).
26
related to the underlying geometrical structure of the inputs. If the connectivity
matrix is embedded in a lower two-dimensional space (k = 2), then the emerg-
ing patterns are similar to experimentally observed cortical feature maps. That
is, neurons with the same feature preferences tend to cluster together forming
cortical columns within the embedding space. Moreover, the recurrent weights
decompose into a local isotropic convolutional part, which is consistent with the
requirements of energy efficiency, and a longer–range non-convolutional part that
is patchy. This suggest a new interpretation of the cortical maps: they correspond
to two-dimensional embeddings of the underlying geometry of the inputs.
One of the limitations of applying simple Hebbian learning to recurrent corti-
cal connections is that it only takes into account excitatory connections, whereas
20% of cortical neurons are inhibitory. Indeed, in most developmental models of
feedforward connections, it is assumed that the local and convolutional connec-
tions in cortex have a Mexican hat shape with negative (inhibitory) lobes for neu-
rons that are sufficiently far from each other. From a computational perspective, it
is possible to obtain such a weight distribution by replacing Hebbian learning with
some form of covariance learning ([Sejnowski and Tesauro, 1989]). However, it
is difficult to prove convergence to a fixed point in the case of the covariance learn-
ing rule, and multidimensional scaling method cannot be applied directly unless
the Mexican hat function is truncated so that it is invertible. Another limitation
of rate-based Hebbian learning is that it does not take into account causality, in
contrast to more biologically detailed mechanisms such as spike timing dependent
plasticity.
The approach taken here is very different from standard treatments of cortical
development [Miller et al., 1989, Swindale, 1996], in which the recurrent connec-
tions are assumed to be fixed and of convolutional Mexican hat form whilst the
feedforward vertical connections undergo some form of correlation-based Heb-
bian learning. In the latter case, cortical feature maps form in the physical space
of retinocortoical coordinates yi, rather than in the more abstract planar space
of points xi obtained by applying multidimensional scaling to recurrent weights
undergoing Hebbian learning in the presence of fixed vertical connections. A par-
ticular feature of cortical maps formed by modifiable feedforward connections is
that the mean size of a column is determined by a Turing-like pattern forming in-
stability, and depends on the length scales of the Mexican hat weight function and
the two-point input correlations [Miller et al., 1989, Swindale, 1996]. No such
Turing mechanism exists in our approach so that the resulting cortical maps tend
to be more fractal-like (many length scales) compared to real cortical maps. Nev-
ertheless, we have established that the geometrical structure of cortical feature
27
maps can also be encoded by modifiable recurrent connections. This should have
interesting consequences for models that consider the joint development of feed-
forward and recurrent cortical connections. One possibility is that the embedding
space of points xi arising from multidimensional scaling of the weights becomes
identified with the physical space of retinocortical positions yi. The emergence of
local convolutional structures together with sparser long-range connections would
then be consistent with energy efficiency constraints in physical space.
Our paper also draws a direct link between the recurrent connectivity of the
network and the positions of neurons in some vector space such as R2. In other
words, learning corresponds to moving neurons or nodes so that their final position
will match the inputs' geometrical structure. Similarly, the Kohonen algorithm
[Kohonen, 1990] describes a way to move nodes according to the inputs presented
to the network. It also converges toward the underlying geometry of the set of
inputs. Although not formally equivalent, it seems that both of these approaches
have the same qualitative behaviour. However, our method is more general in the
sense that no neighborhood structure is assumed a priori; such a structure emerges
via the embedding into Rk.
Finally, note that we have used a discrete formalism based on a finite num-
ber of neurons. However, the resulting convolutional structure obtained by ex-
pressing the weight matrix as a distance matrix in Rk, equation (21), allows us
to take an appropriate continuum limit. This then generates a continuous neural
field model in the form of an integro-differential equation whose integral kernel is
given by the underlying weight distribution. Neural fields have been used increas-
ingly to study large–scale cortical dynamics (see [Coombes, 2005] for a review).
Our geometrical learning theory provides a developmental mechanism for the for-
mation of these neural fields. One of the useful features of neural fields from a
mathematical perspective is that many of the methods of partial differential equa-
tions can be carried over. Indeed, for a general class of connectivity functions
defined over continuous neural fields, a reaction-diffusion equation can be de-
rived whose solution approximates the firing rate of the associated neural field
[Degond and Mas-Gallic, 1989, Cottet, 1995, Edwards, 1996]. It appears that the
necessary connectivity functions are precisely those that can be written in the
form (21). This suggests that a network that has been trained on a set inputs with
an appropriate geometrical structure behaves as a diffusion equation in a high-
dimensional space together with a reaction term corresponding to the inputs.
28
7 Acknowldegments
MNG and ODF were partially funded by the ERC advanced grant NerVi nb
227747. MG was partially funded by the r´egion PACA, France. This publica-
tion was based on work supported in part by the National Science Foundation
(DMS-0813677) and by Award No KUK-C1-013-4 made by King Abdullah Uni-
versity of Science and Technology (KAUST). PCB was also partially supported
by the Royal Society Wolfson Foundation.
8 Appendix
8.1 Proof of the convergence to the symmetric attractor A
We need to prove the 2 points: (i) A is an invariant set, and (ii) for all Y(0) ∈
RN ×M × RN ×N , Y(t) converges to A as t → +∞. Since RN ×N is the direct sum
of the set of symmetric connectivities and the set of anti-symmetric connectivies,
we write W (t) = WS(t) + WA(t), ∀t ∈ R+, where WS is symetric and WA is
anti-symetric.
(i) In (15), the right hand side of the equation for W is symmetric. Therefore,
if ∃t1 ∈ R+ such that WA(t1) = 0, then W remains in A for t ≥ t1.
(ii) Projecting the expression for W in equation (15) on to the anti-symmetric
component leads to
dWA
dt
= −ǫµWA(t)
(26)
whose solution is WA(t) = WA(0) exp(−ǫµt), ∀t ∈ R+. Therefore,
0. The system converges exponentially to A.
lim
t→+∞
WA(t) =
8.2 Proof of theorem 4.1
Consider the following Lyapunov function (see equation (16))
where µ = µM, such that if W = WS + WA, where WS is symmetric and WA is
anti-symmetric.
E(U, W ) = −
1
2
hU, W · Ui − hI, Ui + h1, S−1(cid:0)U(cid:1)i +
− ∇E(U, W ) =(cid:18)WS · U + I − S−1(cid:0)U(cid:1)
(cid:19)
U · U T − µW
29
µ
2
kW k2,
(27)
(28)
Therefore, writing the system Σ′, equation (15), as
dY
dt
= γ(cid:18)WS · S(V) + I − S−1(cid:0)S(V)(cid:1)
S(V) · S(V)T − µW
(cid:19) + γ(cid:18)WA.S(V)
0
(cid:19) ,
where Y = (V, W )T , we see that
dY
dt
= −γ(cid:18)∇E(cid:0)σ(V, W )(cid:1)(cid:19) + Γ(t)
(29)
where γ(V, W )T = (V, ǫW/M)T , σ(V, W ) = (S(V), W ) and Γ : R+ → H such
that kΓk →
0 exponentially (because the system converges to A). It follows
that the time derivative of E = E ◦ σ along trajectories is given by:
t→+∞
d E
dt
=(cid:28)∇ E,
dY
dt(cid:29) =(cid:28)∇V E,
dV
dt(cid:29) +(cid:28)∇W E,
dW
dt (cid:29).
(30)
Substituting equation (29) then yields
d E
dt
= −(cid:28)∇ E, γ(cid:0)∇E ◦ σ(cid:1)(cid:29) +(cid:28)∇ E, Γ(t)(cid:29)
}
{z
M(cid:28)∇W E ◦ σ, ∇W E ◦ σ(cid:29) + Γ(t).
= −(cid:28)S′(V)∇U E ◦ σ, ∇U E ◦ σ(cid:29) −
Γ(t)
ǫ
(31)
We have used the chain–rule of differentiation, whereby
∇V ( E) = ∇V (E ◦ σ) = S′(V)∇U E ◦ σ,
and S′(V)∇U E (without dots) denotes the Hadamard (term by term) product, that
is,
[S′(V)∇U E]ia = s′(V (a)
i
)
∂E
∂U (a)
i
Note that Γ →
t→+∞
0 exponentially because ∇ E is bounded, and S′(V) > 0
because the trajectories are bounded. Thus, there exists t1 ∈ R+ such that ∀t > t1,
∃k ∈ R∗
+ such that
d E
dt
≤ −kk∇E ◦ σk2 ≤ 0.
(32)
As in [Cohen and Grossberg, 1983] and [Dong and Hopfield, 1992], we apply the
Krasovskii-LaSalle invariance principle [Khalil and Grizzle, 1996]. We check that:
30
• E is lower bounded. Indeed, V and W are bounded. Given that I and S are
also bounded it is clear that E is bounded.
d E
dt
•
is negative semidefinite on the trajectories as shown in equation (32).
Then the invariance principle tells us that the solutions of the system Σ′ approach
the set M = nY ∈ H :
H : ∇E ◦ σ = 0o. Since
d E
dt
dY
dt
(Y) = 0o. Equation (32) implies that M = nY ∈
= −γ(cid:16)∇E ◦ σ(cid:17) and γ 6= 0 everywhere, M consists
of the equilibrium points of the system. This completes the proof.
8.3 Proof of theorem 4.2
Denote the right–hand side of system Σ′, equation (15) by
F (V, W ) =(
ǫ
−V + W · S(cid:0)V(cid:1) + I
M(cid:0)S(V).S(V)T − µMW(cid:1)
ǫ
The fixed points satisfy the condition F (V, W ) = 0 which immediately leads to
equations (19). Let us now check the linear stability of this system. The differen-
tial of F at V ∗, W ∗ is
where S′(V ∗)Z denotes a Hadamard product, that is, [S′(V ∗)Z]ia = s′(V ∗
i
−Z + W ∗ ·(cid:0)S′(V ∗)Z(cid:1) + J · S(V ∗)
M(cid:16)(cid:0)S′(V ∗)Z(cid:1) · S(V ∗)T + S(V ∗) ·(cid:0)S′(V ∗)Z(cid:1)T
− µMJ(cid:17),!
dF(V ∗,W ∗)(Z, J) =
J(cid:19) =
Assume that there exist λ ∈ C∗, (Z, J) ∈ H such that dF(V ∗,W ∗)(cid:18)Z
J(cid:19). Taking the second component of this equation and computing the dot
λ(cid:18)Z
(a))Z (a)
.
i
product with S(V ∗) leads to
(λ + ǫµ)J · S =
ǫ
M (cid:0)(S′Z) · ST · S + S · (S′Z)T · S(cid:1)
where S = S(V ∗), S′ = S′(V ∗). Substituting this expression in the first equation
leads to
M(λ + ǫµ)(λ + 1)Z = (
λ
µ
+ ǫ)S · ST · (S′Z) + ǫ(S′Z) · ST · S + ǫS · (S′Z)T · S
(33)
31
Observe that setting ǫ = 0 in the previous equation leads to an eigenvalue
equation for the membrane potential only:
(λ + 1)Z =
1
µM
S · ST · (S′Z).
µM(cid:0)S · ST(cid:1), this equation implies that λ + 1 is an eigenvalue of
Since W ∗ = 1
the operator X 7→ W ∗.(S′X). The magnitudes of the eigenvalues are always
smaller than the norm of the operator. Therefore, we can say that if 1 > kW ∗kS′
m
then all the possible eigenvalues λ must have a negative real part. This sufficient
condition for stability is the same as in [Faugeras et al., 2008]. It says that fixed
points sufficiently close to the origin are always stable.
Let us now consider the case ǫ 6= 0. Recall that Z is a matrix. We now
"flatten" Z by storing its rows in a vector called Zrow. We use the following result
in [Brewer, 1978]: the matrix notation of operator X 7→ A·X ·B is A⊗BT , where
⊗ is the Kronecker product. In this formalism the previous equation becomes
M(λ+ǫµ)(λ+l)Zrow =(cid:18)(
λ
µ
+ǫ)S·ST ⊗Id+ǫId⊗ST ·S+ǫS⊗ST(cid:19)·(S′Z)row
(34)
where we assume that the Kronecker product has the priority over the dot product.
We focus on the linear operator O defined by the right hand side and bound its
kXk which is equal
norm. Note that we use the following norm kW k∞ = supX
to the largest magnitude of the eigenvalues of W .
kW.Xk
kOk∞ ≤ S′
m(cid:18)
λ
µ
kS · ST ⊗ Idk∞ + ǫkS · ST ⊗ Idk∞ + ǫkId ⊗ ST · Sk∞
+ ǫkS ⊗ ST k∞(cid:19).
(35)
Define, νm to be the magnitude of the largest eigenvalue of W ∗ = 1
µM (S · ST ).
First, note that S · ST and ST · S have the same eigenvalues (µM)νi but different
eigenvectors denoted by ui for S · ST and vi for ST · S. In the basis set spanned by
the ui ⊗ vj, we find that S · ST ⊗ Id and Id ⊗ ST · S are diagonal with (µM)νi as
eigenvalues. Therefore, kS·ST ⊗Idk∞ = (µM)νm and kId⊗ST ·Sk∞ = (µM)νm.
Moreover, observe that
(ST ⊗S)T ·(ST ⊗S)·(ui⊗vj) = (S·ST ·ui)⊗(ST ·S·vj) = (µM)2νiνj ui⊗vj (36)
32
Therefore, (ST ⊗ S)T · (ST ⊗ S) = (µM)2diag(νiνj). In other words, ST ⊗ S
is the composition of an orthogonal operator (i.e. an isometry) and a diagonal
matrix. Immediately, it follows that kST ⊗ Sk ≤ (µM)νm.
Compute the norm of equation (34)
(λ + ǫµ)(λ + 1) ≤ S′
m(λ + 3ǫµ)νm.
(37)
Define fǫ : C → R such that fǫ(λ) = (λ + ǫµ)(λ + 1) − (λ + 3ǫµ)S′
mνm.
We want to find a condition such that fǫ(C+) > 0, where C+ is the right half
complex plane. This condition on ǫ, µ, νm, and S′
m will be a sufficient condition
for linear stability. Indeed, under this condition we can show that only eigenvalues
with a negative real part can meet the necessary condition (37). Complex number
of the right half plane cannot be eigenvalues and thus the system is stable. The
case ǫ = 0 tells us that f0(C+) > 0 if 1 > S′
mνm, compute
∂fǫ
∂ǫ
(λ) = µ(ℜ(λ) + µǫ)
(λ + 1)
(λ + ǫµ)
− 3µS′
mνm
If 1 ≥ ǫµ, which is most probably true given that ǫ << 1, then (λ+1)
Assuming λ ∈ C+ leads to:
(λ+ǫµ) ≥ 1.
∂fǫ
∂ǫ
(λ) ≥ µ(µǫ − 3S′
mνm) ≥ µ(1 − 3S′
mνm)
Therefore, the condition 3S′
fǫ(C+) > 0.
mνm < 1, which implies S′
mνm < 1, and leads to
8.4 A very short introduction to computational cohomology
In algebraic topology, topological spaces (which are continuous objects) can be
classified by roughly counting their number of holes. This coordinate-invariant
description of a topological space is called its homology (or cohomology, the
difference between them is beyond the scope of this paper). In fact the homology
can be summarized by giving the betti numbers of the topological state. The
sequence of betti number is made of positive integers. The first three betti numbers
have the following definition: the first is the number of connected components,
the second is the number of two-dimensional or "circular" holes and the third is
the number of 3-dimensional holes or "voids". See [Hatcher, 2002] for a more
rigorous approach.
33
However, in the example of toroidal retinotopy (see section 5.3.2 and figure 6).
we are dealing with a discrete cloud of points. Therefore, one needs to extend the
definition of the betti numbers to discrete objects in order to find the underly-
ing topology of the space within which the points are distributed. This is called
computational or persistent cohomology. One reconstructs the topological space
by considering balls of a given radius centered on each point in the cloud. For
each radius (the abscissa of the right picture of figure 6), one can compute the
betti numbers of the resulting topological space. A barcode graph, e.g. the right
picture of figure 6, is constructed by drawing a horizontal bar for each connected
component, 2-dimensional hole or 3-dimenional void etc. This is done for a range
of radii. Finding the persistent cohomology of a cloud of points consists in ob-
serving the set of betti numbers that are stable through a significantly large range
of radii. One then assumes that the points most likely lie on a topological space
of a given homology if the corresponding betti numbers are stable enough. See
[Zomorodian and Carlsson, 2005] for details.
In section 5.3.2, we used the Jplex software package of [Sexton and Vejdemo-Johansson, ]
to compute the barcodes of the points corresponding to learning from inputs uni-
formly distributed over a 2-dimensional torus. We used 200 landmarks spread
according to the maxminlandmark method to build simplices in Jplex, which re-
turned the maximum radius (beyond which all the betti numbers except the first
vanish) we used in the simulation. In figure 6, we see that for a wide range of radii
the triplet (1, 2, 1) is stable. This corresponds to a 2 dimensional torus.
References
[Amari, 1998] Amari, S. (1998). Natural gradient works efficiently in learning.
Neural computation, 10(2):251–276.
[Amari et al., 1992] Amari, S., Kurata, K., and Nagaoka, H. (1992). Information
IEEE Transactions on Neural Networks,
geometry of Boltzmann machines.
3(2):260–271.
[Bartsch and Van Hemmen, 2001] Bartsch, A. and Van Hemmen, J. (2001).
Combined Hebbian development of geniculocortical and lateral connectivity
in a model of primary visual cortex. Biological Cybernetics, 84(1):41–55.
34
[Bi and Poo, 2001] Bi, G. and Poo, M. (2001). Synaptic modification by cor-
related activity: Hebb's postulate revisited. Annual review of neuroscience,
24:139.
[Bienenstock et al., 1982] Bienenstock, E., Cooper, L., and Munro, P. (1982).
Theory for the development of neuron selectivity: orientation specificity and
binocular interaction in visual cortex. J Neurosci, 2:32–48.
[Borg and Groenen, 2005] Borg, I. and Groenen, P. (2005). Modern multidimen-
sional scaling: Theory and applications. Springer Verlag.
[Bosking et al., 1997] Bosking, W., Zhang, Y., Schofield, B., and Fitzpatrick, D.
(1997). Orientation selectivity and the arrangement of horizontal connections
in tree shrew striate cortex. Journal of neuroscience, 17(6):2112.
[Bressloff, 2005] Bressloff, P. (2005). Spontaneous symmetry breaking in self–
organizing neural fields. Biological Cybernetics, 93(4):256–274.
[Bressloff et al., 2001] Bressloff, P., Cowan, J., Golubitsky, M., Thomas, P., and
Wiener, M. (2001). Geometric visual hallucinations, euclidean symmetry and
the functional architecture of striate cortex. Phil. Trans. R. Soc. Lond. B,
306(1407):299–330.
[Bressloff and Cowan, 2003] Bressloff, P. C. and Cowan, J. D. (2003). A spher-
ical model for orientation and spatial frequency tuning in a cortical hypercol-
umn. Philosophical Transactions of the Royal Society B.
[Brewer, 1978] Brewer, J. (1978). Kronecker products and matrix calculus in
system theory. IEEE Transactions on Circuits and Systems, 25(9).
[Chklovskii et al., 2002] Chklovskii, D., Schikorski, T., and Stevens, C. (2002).
Wiring optimization in cortical circuits. Neuron, 34(3):341–347.
[Chossat and Faugeras, 2009] Chossat, P. and Faugeras, O. (2009). Hyperbolic
planforms in relation to visual edges and textures perception. PLoS Computa-
tional Biology, 5(12):367–375.
[Cohen and Grossberg, 1983] Cohen, M. and Grossberg, S. (1983). Absolute sta-
bility of global pattern formation and parallel memory storage by competitive
neural networks.
In IEEE Transactions on Systems, Man, and Cybernetics,
SMC-13, pages 815–826.
35
[Coifman et al., 2005] Coifman, R., Maggioni, M., Zucker, S., and Kevrekidis, I.
(2005). Geometric diffusions for the analysis of data from sensor networks.
Current opinion in neurobiology, 15(5):576–584.
[Coombes, 2005] Coombes, S. (2005). Waves, bumps, and patterns in neural field
theories. Biological Cybernetics, 93(2):91–108.
[Cottet, 1995] Cottet, G. (1995). Neural networks: Continuous approach and
applications to image processing. Journal of Biological Systems, 3:1131–1139.
[Dayan and Abbott, 2001] Dayan, P. and Abbott, L. (2001). Theoretical Neuro-
science : Computational and Mathematical Modeling of Neural Systems. MIT
Press.
[Degond and Mas-Gallic, 1989] Degond, P. and Mas-Gallic, S. (1989). The
Weighted Particle Method for Convection-Diffusion Equations. Part 1: The
Case of an Isotropic Viscosity. Mathematics of Computation, pages 485–507.
[Dong and Hopfield, 1992] Dong, D. and Hopfield, J. (1992). Dynamic prop-
erties of neural networks with adapting synapses. Network: Computation in
Neural Systems, 3(3):267–283.
[Edwards, 1996] Edwards, R. (1996). Approximation of neural network dynam-
ics by reaction-diffusion equations. Mathematical methods in the applied sci-
ences, 19(8):651–677.
[Faugeras et al., 2008] Faugeras, O., Grimbert, F., and Slotine, J.-J. (2008). Abo-
lute stability and complete synchronization in a class of neural fields models.
SIAM J. Appl. Math, 61(1):205–250.
[Foldi´ak, 1991] Foldi´ak, P. (1991). Learning invariance from transformation se-
quences. Neural Computation, 3(2):194–200.
[Geman, 1979] Geman, S. (1979). Some averaging and stability results for ran-
dom differential equations. SIAM J. Appl. Math, 36(1):86–105.
[Gerstner and Kistler, 2002] Gerstner, W. and Kistler, W. M. (2002). Mathemati-
cal formulations of hebbian learning. Biological Cybernetics, 87:404–415.
[Hatcher, 2002] Hatcher, A. (2002). Algebraic topology. Cambridge Univ Pr.
36
[Hebb, 1949] Hebb, D. (1949). The organization of behavior: a neuropsycholog-
ical theory. Wiley, NY.
[Hubel and Wiesel, 1977] Hubel, D. H. and Wiesel, T. N. (1977). Functional
architecture of macaque monkey visual cortex. Proc. Roy. Soc. B, 198:1–59.
[Khalil and Grizzle, 1996] Khalil, H. and Grizzle, J. (1996). Nonlinear systems.
Prentice hall Upper Saddle River, NJ.
[Kohonen, 1990] Kohonen, T. (1990). The Self-Organizing Map. Proceedings of
the IEEE, 78(9).
[Lawlor and Zucker, 2010] Lawlor, M. and Zucker, S. (2010). Third-Order Edge
Statistics Reveal Curvature Dependency. In Snowbird workshop on learning.
[Miikkulainen et al., 2005] Miikkulainen, R., Bednar, J., Choe, Y., and Sirosh, J.
(2005). Computational Maps in the Visual Cortex. Springer, New York.
[Miller, 1996] Miller, K. (1996). Synaptic economics: competition and coopera-
tion in synaptic plasticity. Neuron, 17:371–374.
[Miller and MacKay, 1996] Miller, K. and MacKay, D. (1996). The role of con-
straints in hebbian learning. Neural Comp, 6:100–126.
[Miller et al., 1989] Miller, K. D., Keller, J. B., and Stryker, M. P. (1989). Ocular
dominance column development: analysis and simulation. Science, 245:605–
615.
[Oja, 1982] Oja, E. (1982). A simplified neuron model as a principal component
analyzer. J. Math. Biology, 15:267–273.
[Ooyen, 2001] Ooyen, A. (2001). Competition in the development of nerve con-
nections: a review of models. Network: Computation in Neural Systems,
12(1):1–47.
[Petitot, 2003] Petitot, J. (2003). The neurogeometry of pinwheels as a sub-
Riemannian contact structure. Journal of Physiology-Paris, 97(2-3):265–309.
[Sejnowski and Tesauro, 1989] Sejnowski, T. and Tesauro, G. (1989). The Hebb
rule for synaptic plasticity: algorithms and implementations. Neural models of
plasticity: Experimental and theoretical approaches, pages 94–103.
37
[Sexton and Vejdemo-Johansson, ] Sexton,
and
Johansson,
complex
http://comptop.stanford.edu/programs/jplex/.
M.
H.
JPlex
simplicial
Vejdemo-
library.
[Swindale, 1996] Swindale, N. (1996). The development of topography in the
visual cortex: a review of models. Network: Computation in Neural Systems,
7(2):161–247.
[Takeuchi and Amari, 1979] Takeuchi, A. and Amari, S. (1979). Formation of
topographic maps and columnar microstructures in nerve fields. Biological
Cybernetics, 35(2):63–72.
[Tikhonov, 1952] Tikhonov, A. (1952). Systems of differential equations with
small parameters multiplying the derivatives. Matem. sb, 31(3):575–586.
[Verhulst, 2007] Verhulst, F. (2007). Singular perturbation methods for slow–fast
dynamics. Nonlinear Dynamics, 50(4):747–753.
[Wallis and Baddeley, 1997] Wallis, G. and Baddeley, R. (1997). Optimal, un-
supervised learning in invariant object recognition. Neural computation,
9(4):883–894.
[Zomorodian and Carlsson, 2005] Zomorodian, A. and Carlsson, G. (2005).
Computing persistent homology. Discrete and Computational Geometry,
33(2):249–274.
38
|
1003.5557 | 1 | 1003 | 2010-03-29T14:51:39 | A ratchet mechanism for amplification in low-frequency mammalian hearing | [
"q-bio.NC",
"physics.bio-ph"
] | The sensitivity and frequency selectivity of hearing result from tuned amplification by an active process in the mechanoreceptive hair cells. In most vertebrates the active process stems from the active motility of hair bundles. The mammalian cochlea exhibits an additional form of mechanical activity termed electromotility: its outer hair cells (OHCs) change length upon electrical stimulation. The relative contributions of these two mechanisms to the active process in the mammalian inner ear is the subject of intense current debate. Here we show that active hair-bundle motility and electromotility can together implement an efficient mechanism for amplification that functions like a ratchet: sound-evoked forces acting on the basilar membrane are transmitted to the hair bundles whereas electromotility decouples active hair-bundle forces from the basilar membrane. This unidirectional coupling can extend the hearing range well below the resonant frequency of the basilar membrane. It thereby provides a concept for low-frequency hearing that accounts for a variety of unexplained experimental observations from the cochlear apex, including the shape and phase behavior of apical tuning curves, their lack of significant nonlinearities, and the shape changes of threshold tuning curves of auditory nerve fibers along the cochlea. The ratchet mechanism constitutes a general design principle for implementing mechanical amplification in engineering applications. | q-bio.NC | q-bio | A ratchet mechanism for amplification
in low-frequency mammalian hearing
Tobias Reichenbach and A. J. Hudspeth
∗Howard Hughes Medical Institute and Laboratory of Sensory Neuroscience, The Rockefeller University, New York, New York 10065-6399, U.S.A.
∗
Submitted to Proceedings of the National Academy of Sciences of the United States of America
0
1
0
2
r
a
M
9
2
]
.
C
N
o
i
b
-
q
[
1
v
7
5
5
5
.
3
0
0
1
:
v
i
X
r
a
The sensitivity and frequency selectivity of hearing result from tuned
amplification by an active process in the mechanoreceptive hair cells.
In most vertebrates the active process stems from the active motil-
ity of hair bundles. The mammalian cochlea exhibits an additional
form of mechanical activity termed electromotility:
its outer hair
cells (OHCs) change length upon electrical stimulation. The rela-
tive contributions of these two mechanisms to the active process in
the mammalian inner ear is the subject of intense current debate.
Here we show that active hair-bundle motility and electromotility can
together implement an efficient mechanism for amplification that
functions like a ratchet: sound-evoked forces acting on the basilar
membrane are transmitted to the hair bundles whereas electromotil-
ity decouples active hair-bundle forces from the basilar membrane.
This unidirectional coupling can extend the hearing range well below
the resonant frequency of the basilar membrane. It thereby provides
a concept for low-frequency hearing that accounts for a variety of
unexplained experimental observations from the cochlear apex, in-
cluding the shape and phase behavior of apical tuning curves, their
lack of significant nonlinearities, and the shape changes of threshold
tuning curves of auditory nerve fibers along the cochlea. The ratchet
mechanism constitutes a general design principle for implementing
mechanical amplification in engineering applications.
auditory system cochlea hair cell
T he mammalian cochlea acts as a frequency analyzer in
which high frequencies are detected at the organ's base
and low frequencies at more apical positions. This frequency
mapping is thought to be achieved by a position-dependent
resonance of the elastic basilar membrane separating two fluid-
filled compartments (Fig. 1A) [1, 2, 3]. When sound evokes
a pressure wave that displaces the basilar membrane, the re-
sultant traveling wave gradually increases in amplitude as it
progresses to the position where the basilar membrane's reso-
nant frequency coincides with that of the stimulus. Aided by
mechanical energy provided by the active process, the wave
peaks at a characteristic place slightly before the resonant po-
sition and then declines sharply (Fig. 1B). This mechanism is
termed critical-layer absorption [1], for a wave cannot travel
beyond its characteristic position on the basilar membrane,
but peaks and dissipates most of its energy there. The mech-
anism displays scale invariance: different stimulation frequen-
cies induce traveling waves that display a common, strongly
asymmetric form upon rescaling of the amplitude and spatial
coordinate [4, 5].
Two important aspects of the cochlea's mechanics remain
problematical. First, the basilar membrane's resonant fre-
quency apparently cannot span the entire range of audible fre-
quencies. Experimental measurements of basilar-membrane
stiffness suggest that high-frequency resonances are feasible
but low-frequency ones are inaccessible [6]. This result accords
with the analysis of threshold tuning curves for auditory-
nerve fibers [7, 8, 9] and measurements of basilar-membrane
displacement [8, 10, 11, 12], both of which indicate that a
peaked traveling wave occurs for high-frequency stimulation
but not for low-frequency stimulation. Indeed, threshold tun-
ing curves of auditory-nerve fibers [7, 8, 9] show that high-
frequency curves are scale-invariant and possess a sharp cut-
off at frequencies above their characteristic frequencies, re-
1 -- 16
flecting the mechanism of critical-layer absorption. However,
low-frequency curves lack this sharp cutoff and, in contra-
diction of the expectation for critical-layer absorption, are
instead characterized by an approximately symmetric shape
around their characteristic frequencies. Recent experiments
in the chinchilla have shown that the shape change between
the tuning curves of high- and low-frequency fibers occurs
between two crossover frequencies of about 5 kHz and 1.5
kHz [8, 9]. Measurements of basilar-membrane displacement
yield similar conclusions. Experiments from the cochlear base
confirm the existence at high frequencies of peaked travel-
ing waves that result in strongly asymmetric tuning curves
and a pronounced nonlinearity at the characteristic frequen-
cies [13, 14, 8, 12]. Apical measurements of basilar-membrane
displacement, however, produce symmetric tuning curves that
lack a sharp cutoff for frequencies higher than the character-
istic frequency [10, 15, 11]. Moreover, only small nonlinearity
has been measured [15, 16, 10], raising the question whether
amplification occurs at the apex. All available experimental
results therefore indicate that low-frequency hearing does not
function through critical-layer absorption but must rely on
another mechanism to achieve frequency selectivity.
The second key uncertainty is the nature of the active
process that operates in the mammalian cochlea.
In addi-
tion to amplifying weak signals, the active process produces
increased frequency selectivity, compressive nonlinearity, and
spontaneous otoacoustic emission. In the mammalian cochlea
amplification is provided by specialized OHCs located in the
organ of Corti along the basilar membrane (Fig. 2A). Each
OHC displays two forms of motility. Like those in the hear-
ing organs of other vertebrates [17, 18], the hair bundle of
an OHC can produce mechanical force [17, 18, 19, 20, 21]
(see [22] for a review). But an OHC also exhibits electro-
motility: when its membrane potential changes as a result
of sound-evoked hair-bundle deflection, the entire cell under-
goes a length change owing to conformational rearrangement
of the membrane protein prestin [23, 24, 25]. Direct measure-
ments of the respective roles of the two forms of motility are
complicated, for it is difficult to determine the micromechan-
ical responses of the organ of Corti while the cochlea remains
intact.
Here we have taken a theoretical approach to investigate
a possible mechanism for amplification by the synergistic in-
terplay of active hair-bundle motility and electromotility. We
show that they can operate together to achieve low-frequency
selectivity in the absence of basilar-membrane resonance and
critical-layer absorption. The model advances theoretical un-
derstanding of cochlear mechanics in three ways: it reproduces
previous theoretical results insofar as they coincide with ex-
periments; it accounts naturally for a variety of unexplained
findings from the cochlear apex; and it makes robust, experi-
mentally testable predictions.
Fig. 1. Principles of cochlear mechanics. (A) In a schematic diagram of the mam-
malian cochlea, the basilar membrane (BM) is displaced by sound stimuli acting on
the stapes (top left). (B) In the classical theory of cochlear mechanics, sound evokes
a pressure wave that causes a longitudinal traveling wave of basilar-membrane dis-
placement (thick line). The motion of the basilar membrane and the displacements
of the associated hair bundles are approximately equal. As the wave approaches the
position where its frequency matches the basilar membrane's resonant frequency, the
wave's amplitude (thin line) increases and its wavelength and velocity decline. The
wave peaks at a characteristic place slightly before the resonant position and then de-
clines sharply, yielding a strongly asymmetric envelope of the traveling wave (shading).
Experiments confirm this behavior in the basal, high-frequency part of the cochlea.
(C) We propose an alternative theory for the cochlea's mechanics at low frequencies.
The basilar membrane near the cochlear apex does not resonate, but the traveling
wave on the basilar membrane propagates along the entire cochlea without a strong
variation in amplitude, wavelength, and velocity (black). However, the interplay of
electromotility and active hair-bundle motility fosters an independent resonance of
the complex formed by the hair bundles, reticular lamina, and tectorial membrane.
The hair-bundle displacement (red) at the characteristic place can therefore exhibit
an approximately symmetric peak, exceeding basilar-membrane motion by orders of
magnitude.
Results
The Ratchet Mechanism. Consider a transverse element of the
cochlear partition comprising the basilar membrane, an inner
hair cell (IHC) and three OHCs, and the overlying tectorial
membrane (Fig. 2A). When a sound-evoked force displaces
the basilar membrane upward, the resultant shearing motion
between the tectorial membrane and the top of the OHCs,
or reticular lamina, deflects the hair bundles in the positive
direction. Active hair-bundle motility in the OHCs increases
the amplitude of deflection. Without electromotility, this ad-
ditional displacement would couple back to the basilar mem-
brane and augment the movement there. If electromotility is
adjusted such that the OHCs elongate just as much as the
tectorial membrane and reticular lamina move upward, how-
ever, the basilar membrane does not experience the active
force and thus undergoes no additional displacement (see also
Movie S1).
A mathematical formulation of this amplification mech-
anism clarifies its operation. Consider a two-mass model
in which the two degrees of freedom are the motion of the
basilar membrane, XBM, and that of the complex formed by
2
Fig. 2. The ratchet mechanism. (A) The organ of Corti rests upon the basilar
membrane (BM). Three OHCs are connected to Deiters' cells (DC), which together
couple the basilar membrane to the reticular lamina (dark green, top of the OHCs) and
through the hair bundles to the overlying tectorial membrane (TM). Sound-evoked
external forces (black arrow) displace the basilar membrane, here upwards, and pro-
duce shearing (black arrow) of the hair bundles of OHCs (red asterisk) and the inner
hair cell (IHC) (cyan asterisk). Two forms of motility underlie the active process:
active hair-bundle motility (single-headed red arrow) and membrane-based electro-
motility (double-headed red arrow). (B) The two fundamental degrees of freedom
are the basilar-membrane displacement XBM and the displacement XHB of the
hair-bundle complex (circle), which comprises the hair bundles, reticular lamina (RL),
and tectorial membrane. Coupling stems from the impedance ZD of the combined
OHCs and Deiters' cells as well as the impedance ZC of the remaining organ of
Corti. (C) In the ratchet mechanism displacements of the basilar membrane caused
by external forces are communicated to the hair-bundle complex. Internal forces (red
arrow) in the hair bundles increase the shearing motion, which decouples from basilar-
membrane displacement through appropriate length changes of the OHCs (dotted red
arrow). For an animated representation of the model, see Movie S1.
the hair bundles, reticular lamina, and tectorial membrane,
XHB (Fig. 2B). A sound stimulus of frequency f produces
an oscillating external force Fext(t) = Fexte2πif t + c.c. act-
ing on the basilar membrane, in which c.c denotes the com-
plex conjugate. The evoked oscillations of the basilar mem-
brane as well as the hair-bundle complex occur predomi-
nantly at the same frequency, XBM = XBMe2πif t + c.c. and
XHB = XHBe2πif t + c.c..
Internal forces Fint arise within
hair bundles, where they provide negative damping and intro-
duce nonlinearities (Supporting Information). The cell bod-
ies of OHCs can be described as piezoelectric elements [26].
For small, physiologically relevant motions their electrically
evoked displacement XEE = XEEe2πif t + c.c. is proportional
to the hair-bundle displacement, which triggers changes in
the membrane potential, so XEE = −α XHB + c.c. with a com-
plex mechanomotility coefficient α [27]. The hair-bundle and
basilar-membrane displacements then depend linearly on the
forces:
.
[ 1 ]
(cid:18) XHB
(cid:19)
XBM
A
=
(cid:18) Fint
(cid:19)
Fext
1
2πif
The matrix A contains the impedances ZHB and ZBM of the
hair-bundle complex and the basilar membrane as well as the
coupling impedances ZD and ZC (Fig. 2B):
(cid:18)ZHB + (1 + α)ZD + ZC
−ZD − ZC
(cid:19)
A =
−(1 + α)ZD − ZC
ZBM + ZD + ZC
.
[ 2 ]
high frequencylow frequencyBMApexBaseADisplacementBDisplacementRelative position01CA*DC*TMIHCBMOHCBCHair−bundle.TMcomplexRLBMXBMXHBXEEZDZCcoupledecoupleExternal force.Internal forceHBBM A key feature of this relation is that the matrix ele-
ment A21, which describes the coupling of the internal force
to the basilar membrane, vanishes at a critical value α∗ ≡
−1 − ZC/ZD. At the same time, the coupling of the external
force to the hair bundle, which is represented by element A12,
remains nonzero. For the critical value α∗ the displacements
become
XHB =
XBM =
1
2πif ZHB
Fint +
1
2πif (ZBM + ZD + ZC)
Fext .
ZD + ZC
2πif ZHB(ZBM + ZD + ZC)
Fext ,
[ 3 ]
The hair-bundle displacement depends on both the internal
and external forces, whereas the basilar membrane experiences
only the external force. In other words, the sound-evoked ex-
ternal force on the basilar membrane is transmitted to the
hair bundle, but the internal force from active hair-bundle
motility does not feed back onto the motion of the basilar
membrane (Fig. 2C). This symmetry-breaking mode has the
characteristics of a ratchet in the sense that information flows
unidirectionally:
information applied as a force against the
basilar membrane can be detected in the form of hair-bundle
displacement, but force acting on the hair bundle cannot be
detected at the basilar membrane. The basilar-membrane dis-
placement equals that occuring when the hair-bundle complex
is fixed at XHB = 0 and electromotility is absent (α = 0). The
ratchet mechanism therefore resembles an ideal operational
amplifier that neither feeds back on nor draws energy from
the input [28].
The Mechanomotility Coefficient α. The magnitude and
phase of the critical value α∗ depend on the coupling
impedances ZC and ZD. Under realistic assumptions ZC is
similar to or smaller than ZD and α∗ is therefore near unity.
Such a value for the mechanomotility coefficient α has been
measured experimentally for apical OHCs [27]. The phase of α
is controlled by the complex network of ion channels that reg-
ulates the membrane potential depending on the hair-bundle
displacement [27, 29]. Although experiments on the phase of
α are not available, theoretical considerations confirm that
OHCs have sufficient flexibility through ion-channel regula-
tion to adjust the phase of α to a variety of values, and in
particular to the phase required for α∗ (Supporting Informa-
tion).
A Concept for Low-Frequency Hearing. As its most impor-
tant characteristic, the ratchet mechanism can explain fre-
quency selectivity near the cochlear apex in the absence of a
peaked traveling wave. The resonant frequency of the hair-
bundle complex is determined solely by the impedance ZHB
and the internal forces, and is therefore defined by the proper-
ties of the hair bundles and tectorial membrane (Equation 3)
[30, 31, 32].
In particular, hair-bundle resonance can occur
in the absence of basilar-membrane resonance. The ratchet
mechanism therefore permits the resonant frequency of the
hair bundles to follow a logarithmic law along the cochlea,
whereas the resonant frequency of the basilar membrane re-
mains significantly greater near the apex (Fig. 3B) [6].
Further evidence points to the occurrence of the ratchet
mechanism at the cochlear apex but not at the base. First,
the basilar membrane at the base is narrow and presumably
tuned to the characteristic frequencies at which auditory-
nerve fibers are most sensitive (Fig. 3B). Amplification of
basilar-membrane motion there is feasible and need not be
avoided.
Indeed, experiments have demonstrated a strong
compressive nonlinearity of basilar-membrane motion at high
3. Cochlear model.
Fig.
(A) The ratchet mechanism operates when the
mechanomotility coefficient α (green) coincides with the critical value α∗ (grey).
Although electromotility is negligible to the basal side of fH, it underlies the ratchet
mechanism apical to the position of fL. (B) The resonant frequency of the hair-
bundle complex (HB, red) agrees with that of the basilar membrane (BM, blue) only
for frequencies above fL. (C) A high-frequency sound stimulus (f1 = 8 kHz) in-
duces a traveling wave that peaks in the basal region. The displacements of the
hair bundles (red) coincide with that of the basilar membrane (blue). Elimination of
active hair-bundle motility decreases the sensitivity by a factor of 90, 000 (green,
hair bundles; black, basilar membrane) indicative of a strong nonlinearity. A low-
frequency stimulus (f2 = 200 Hz) triggers a traveling wave that does not peak on
the basilar membrane, but the hair-bundle displacement exhibits a resonance enabled
by the ratchet mechanism (same color code). Without active hair-bundle motion the
hair-bundle displacement decreases by a factor of only ten indicative of a weak non-
linearity. (D) The phase of the basilar-membrane displacement for f1 has a strongly
increasing slope near the resonant position and thus shows a wave traveling to the
resonant position but not beyond. For f2 the slope of the phase remains almost
constant, corresponding to a wave traveling beyond the characteristic place.
frequencies. However, the apex exhibits a wider basilar mem-
brane that experiences stronger viscous forces and whose
resonant frequencies deviate from the characteristic frequen-
cies [6]. Amplification of basilar-membrane motion near the
apex would therefore be highly inefficient. The absence of
a strong compressive nonlinearity in experimental measure-
ments of basilar-membrane displacement in the apical region
confirms the lack of basilar-membrane amplification there.
Next, the membrane time constant of OHCs restricts elec-
tromotility's ability to work on a cycle-by-cycle basis to fre-
quencies below a few kilohertz [33, 34], disabling the ratchet
mechanism for higher frequencies. It is noteworthy that the
ion channels of apical OHCs differ significantly from those of
basal OHCs [35] and that apical OHCs are considerably longer
than basal ones and can accordingly produce greater length
changes [25].
3
0123Coefficientα0123Coefficientα0123Coefficientα102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)102103104Resonantfrequency(Hz)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)10−1100101102103104Sensitivity(nm·Pa−1)-6-4-2000.20.40.60.81Phase(cycles)RelativepositionABCDPeakedtravelingwaveRatchetmechanismIIIIIIα∗fHfLf1f2HBBMHBBM×90,000×10-6-4-2000.20.40.60.81Phase(cycles)RelativepositionABCDPeakedtravelingwaveRatchetmechanismIIIIIIα∗fHfLf1f2HBBMHBBM×90,000×10-6-4-2000.20.40.60.81Phase(cycles)RelativepositionABCDPeakedtravelingwaveRatchetmechanismIIIIIIα∗fHfLf1f2HBBMHBBM×90,000×10-6-4-2000.20.40.60.81Phase(cycles)RelativepositionABCDPeakedtravelingwaveRatchetmechanismIIIIIIα∗fHfLf1f2HBBMHBBM×90,000×10 A Unified Model for Cochlear Mechanics. We have quantified
these considerations in a one-dimensional model of the cochlea
(Supporting Information). Studies of threshold tuning curves
of auditory-nerve fibers in the chinchilla have delineated three
distinct regimes that are separated by two crossover frequen-
cies, a high frequency fH ≈ 5 kHz and a low frequency
fL ≈ 1.5 kHz (Fig. 3) [9]. In our model for the basal regime
I, above fH, electromotility is negligible owing to the mem-
brane time constant; each segment of the basilar membrane
is tuned to its characteristic frequency.
In the transitional
regime II, between fH and fL, unidirectional coupling pro-
vided by electromotility occurs but every segment of the basi-
lar membrane still resonates at its characteristic frequency. In
the apical regime III, below fL, electromotility combines with
active hair-bundle motility to implement the ratchet mecha-
nism for amplification. The resonant frequency of the basilar
membrane is therefore of minor importance and remains ap-
proximately constant at a value significantly above the range
of characteristic frequencies.
In accordance with the theory of critical-layer absorp-
tion [1], in regime I the traveling wave produced by a pure-
tone stimulus peaks at the characteristic place and sharply
decreases beyond that point (Fig. 3C). The characteristic
behavior of critical-layer absorption -- decrease of the travel-
ing wave's speed and wavelength near the characteristic place
-- appears in the phase behavior as an increase of the slope
Fig. 4. Threshold tuning curves of auditory-nerve fibers. (A) The tuning curve at
each position along the cochlea has a characteristic frequency f0 corresponding to
the resonant frequency of the hair-bundle complex. When tuning curves are rescaled
such that the frequency is measured in octaves relative to the characteristic frequency
f0 and the threshold is measured relative to that at the characteristic frequency, char-
acteristic shape changes are seen to occur between curves of different characteristic
(B) Tuning curves for high characteristic frequencies, above fH, fall
frequencies.
onto a universal curve that exhibits the strongly asymmetric form and high-frequency
cutoff characteristic of a peaked traveling wave. (C) As the characteristic frequency
declines from fH to fL, the left limb falls (arrow), indicating the emerging influence
of electromotility and the ratchet mechanism. (D) As the characteristic frequency
diminishes below fL, the right limb falls steeply (arrow), pointing to the breakdown
of the peaked-wave mechanism and the dominance of ratchet amplification.
4
(Fig. 3D). Active hair-bundle motility amplifies the peak dis-
placement. A strong compressive nonlinearity arises because
amplification by active hair-bundle motility both increases the
hair-bundle and basilar-membrane motion per unit pressure
difference and enhances the amplitude of the pressure wave
itself (Fig. 3C). The combination of the two effects yields
a compressive nonlinearity that extends over a significantly
broader range of sound intensities than the nonlinearity in
hair-bundle motion itself.
In regime II electromotility influences the micromechan-
ics and causes hair-bundle displacement to exceed basilar-
membrane displacement. The theory of critical-layer ab-
sorption still applies. The basilar-membrane tuning curves,
phase behavior, and nonlinearity consequently parallel those
in regime I.
In regime III the ratchet mechanism leads to very dif-
ferent behavior.
Sound evokes a low-frequency pressure
wave that traverses the entire cochlea, evoking only a small
basilar-membrane displacement throughout, for the resonant
frequency of the basilar membrane is higher everywhere
(Fig. 3B). At the characteristic place, however, the hair bun-
dles exhibit an independent resonance amplified through ac-
tive hair-bundle motility and facilitated by the unidirectional
coupling provided by electromotility. The hair-bundle dis-
placement there can exceed the basilar-membrane response
by orders of magnitude (Fig. 3C). In further contrast to the
theory of critical-layer absorption, the traveling wave does not
slow and its wavelength does not vanish at the characteristic
place, for the basilar membrane does not resonate. This be-
comes apparent in the behavior of the traveling wave's phase,
whose slope remains low and nearly constant across the char-
acteristic place (Fig. 3D). The phase behavior also confirms
that the wave reaches the helicotrema. Because amplification
through the ratchet mechanism does not act on the basilar
membrane, the latter exhibits approximately linear behavior
(Fig. 3C). The pressure wave is therefore unaffected by the
active process. Hair-bundle displacement is amplified and ex-
hibits a moderate compressive nonlinearity (Fig. 3C).
Threshold Tuning Curves of Auditory-Nerve Fibers. Strong
support for the proposed model comes from studies of thresh-
old tuning curves for auditory-nerve fibers [7, 8, 9]. Recent
measurements have shown characteristic shape changes oc-
curring around the crossover frequencies fH and fL [9]. To
compare these data to our model, we have computed tuning
curves and -- despite their complexity and the simplicity of
our model -- found striking agreement with the measurements
(Fig. 4). High-frequency fibers, those tuned above fH, dis-
play the strongly asymmetric shape characteristic of critical-
layer absorption. Moreover, they fall onto a universal curve
when the frequency and threshold are measured relative to the
values at the characteristic frequency. This accords with ex-
perimental findings and the scaling symmetry of the peaked
traveling-wave mechanism [4, 5, 9]. For intermediate char-
acteristic frequencies between fL and fH, the scaling law is
violated as the ratchet mechanism starts to influence the mi-
cromechanics, causing the left limb of the threshold tuning
curves to fall as the characteristic frequency decreases [9]. A
second violation of the scaling arises for characteristic frequen-
cies below fL, for which the right limb falls. Also observed
experimentally [9], this second violation of scaling indicates a
breakdown of critical-layer absorption, which predicts a steep
increase for threshold tuning curves at frequencies above the
characteristic frequency.
0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfL0102030405060708090100100010000Threshold(dBSPL)Frequency(Hz)AfHfLf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequencyf04f02f02f0Frequency0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH0102030405060f04f02f02f0Rel.thresh.(dBSPL)FrequencyDf0≤fLCfL≤f0≤fHBf0≥fH Discussion
The classical theory of cochlear mechanics assumes that the
basilar membrane resonates at a characteristic place for each
frequency in the auditory range. The resulting mechanism
of critical-layer absorption is characterized by (i) strongly
asymmetric, scale-invariant tuning curves with steep high-
frequency cutoffs, (ii) an increasing slope of the traveling
wave's phase upon approaching the characteristic place, (iii)
basilar-membrane displacement similar to hair-bundle dis-
placement, and (iv) pronounced compressive nonlinearity at
the characteristic frequency. This picture is consistent with di-
verse experimental findings from the cochlear base, including
direct measurements of basilar-membrane motion and studies
of threshold tuning curves for auditory-nerve fibers. However,
as summarized in the Introduction, experimental results from
the cochlea's apex are in qualitative disagreement with the
first, second, and fourth characteristics.
Here we have proposed a concept for low-frequency hear-
ing that employs a ratchet mechanism involving the interplay
of active hair-bundle motility and electromotility. Whereas
sound-evoked forces displace the basilar membrane and the
hair bundles, the active forces within the hair bundle can de-
couple from the basilar membrane through appropriate elon-
gation and contraction of OHCs. This mechanism allows hair
bundles to resonate independently of the basilar membrane
and thus explains how the ear's hearing range can extend
to values well below the resonant frequency of the basilar
membrane. We have shown that amplification by the ratchet
mechanism leads to characteristic behavior that is distinct
from that associated with critical-layer absorption: it exhibits
(i(cid:48)) approximately symmetric tuning curves that display no
sharp high-frequency cutoff, (ii(cid:48)) a constant slope of the trav-
eling wave's phase across the characteristic place, (iii(cid:48)) tuned
hair-bundle displacement that exceeds the untuned basilar-
membrane displacement by orders of magnitude at resonance,
and (iv(cid:48)) approximately linear basilar-membrane displacement
and only moderate compressive nonlinearity in hair-bundle
motion at the characteristic frequency.
Comparison to Experimental Results. The proposed model
yields the classical theory of critical-layer absorption for high-
frequency sounds (Figs.
1B, 3 and 4A). Experiments in
the base have confirmed (i) the asymmetric form of tuning
curves [13, 14], (ii) the increasing slope of the phase [36, 37],
and (iv) the strong compressive nonlinearity [13, 37, 14]. Ow-
ing to difficulties in accessing the motion of the organ of Corti
at the base, the relation (iii) of hair-bundle motion to basilar-
membrane motion has not yet been measured there.
Because the membrane time constant of OHCs is ex-
pected to limit the operation of electromotility to low frequen-
cies [33, 34], we have chosen in our model to consider only ac-
tive hair-bundle motility in the basal part of the cochlea. How-
ever, studies in prestin knockout mice suggest that electro-
motility is necessary for high-frequency amplification [38, 39].
Electromotility near the base may adjust the operating point
of the hair-bundle motor on a time-scale slower than the pe-
riod of oscillation or provide a fast electrical signal in the form
of membrane-potential change upon mechanical stimulation.
Our theory accounts for a number of unexplained results
from the cochlear apex. Experiments have measured (i(cid:48))
approximately symmetric tuning curves without sharp high-
frequency cutoffs of tectorial-membrane and therefore hair-
bundle displacement [10, 15, 11]. In our model the approxi-
mate symmetry arises naturally, for it reflects the resonance of
the hair bundles, reticular lamina, and tectorial membrane in-
dependently of resonance by the basilar membrane (Figs. 1C
and 3C,D). Experiments have consistently reported (ii(cid:48)) a con-
stant phase slope across the characteristic place [10, 40, 11].
This feature emerges in our model (Fig. 3D) because the basi-
lar membrane is untuned at low frequencies and the pressure
wave therefore reaches the helicotrema. There remains a con-
troversy about the existence of nonlinearity at the cochlear
apex (reviewed in Ref.
[12]): although some investigators
have found no nonlinearites [15], others have reported a small
compressive [10, 16] or expansive nonlinearity [11]. However,
all agree about (iv(cid:48)) the absence of a strong compressive non-
linearity.
The cochlear apex allows observation of different parts
of the organ of Corti -- such as the tectorial membrane and
Hensen's cells as well as the basilar membrane -- and thus per-
mits tests of characteristic (iii(cid:48)) of our model. Experiments
on a temporal-bone preparation [15, 41] have shown that vi-
brations of the reticular lamina and tectorial membrane are
orders of magnitude larger than the basilar-membrane dis-
placement at the characteristic place. These findings are in
perfect agreement with our theory. However, the in vitro ex-
periments have been criticized for an uncertainty about which
constituents of the organ of Corti were measured [12] and are
in contradiction with studies using other preparations [10, 16].
Definite conclusions must therefore be postponed until further
experimental results become available.
Perhaps the most reliable comparison of our model to
experimental measurements comes through tuning curves of
auditory-nerve fibers, which yield information about cochlear
mechanics that is least disturbed by experimental intervention
and represent responses from the whole length of the cochlea.
These tuning curves exhibit progressive shape changes from (i)
the strongly asymmetric form typical of high-frequency fibers
to (i(cid:48)) the nearly symmetric form found for low-frequency
fibers [7, 8, 9]. Our modeled responses of tuning curves of
auditory-nerve fibers exhibit the same shape changes (Fig. 4),
thus providing a conceptual explanation for the three distinct
cochlear regimes.
Experimentally Testable Predictions. Our theory yields ex-
perimentally testable predictions based on the characteris-
tics (i(cid:48))-(iv(cid:48)). As discussed above, (i(cid:48)) the approximately
symmetric tuning curves as well as (ii(cid:48)) the constant phase
slope have already been observed.
In contrast, characteris-
tic (iii(cid:48)), which predicts hair-bundle motion that exceeds the
basilar-membrane motion at the characteristic frequency by
orders of magnitude, remains controversial. Future exper-
iments are therefore required to test this prediction. Such
studies should also determine whether, as our theory implies,
the complex formed by the hair bundles, reticular lamina,
and tectorial membrane exhibits a resonance independent of
the basilar membrane, whose response is untuned. A further
test is feasible through experimental studies of the nonlin-
ear behavior at the apex, for which our theory predicts (iv(cid:48)) a
moderate compressive nonlinearity in hair-bundle motion, and
therefore in tectorial-membrane motion, but no more than a
weak nonlinearity in the basilar membrane's response. Fi-
nally, and beyond the theory presented in this article, the
ratchet mechanism should allow for traveling waves in the
tectorial membrane [42] that are unidirectionally coupled to
basilar-membrane waves.
The Ratchet Principle. The ratchet mechanism constitutes a
general design principle for mechanical amplification whereby
the output does not feed back onto the input. In this way it
represents a mechanical analogue of the operational amplifier
from electrical engineering [28]. Although the current technol-
ogy of signal detectors such as microphones relies on electrical
amplification, mechanical amplification could improve signal-
5
to-noise ratios and thereby greatly advance sensitivity and
detection of weak signals. The ratchet mechanism thus opens
a path for implementing controlled mechanical amplification
in engineering applications.
Materials and Methods
Hydrodynamics.
By combining continuity equations and fluid-momentum equa-
tions, we describe the cochlea's hydrodynamics with the partial differential equation
ρ∂2
t XBM + Λ∂tXBM =
1
2L2 ∂r (h∂rp) .
[ 4 ]
Here ρ denotes the density of liquid in the cochlea, L the length of the cochlea, h
the height of the scalae, and p and XBM respectively the pressure across and the
displacement of the basilar membrane at position r and time t. The coefficient Λ
accounts for friction due to fluid motion. Position is measured in units of the cochlear
length, such that r = 0 corresponds to the basal and r = 1 to the apical end.
The pressure translates into an external force acting on the basilar membrane and
yields a displacement that we compute employing the model of Fig. 2B (Supporting
Information). Temporal Fourier transformation yields an ordinary differential equa-
tion that we solve numerically with the shooting method in Mathematica 7 (Wolfram
Research). We apply two boundary conditions. First, p = p0 at r = 0: a sound-
evoked pressure p0 acts at the stapes. And second, p = 0 at r = 1: because
the two scalae communicate at the helicotrema, the pressure difference between them
vanishes at the apical end of the cochlea.
Tuning Curves of Auditory-Nerve Fibers. Tuning curves of auditory-nerve fibers
are computed by assuming that the hearing threshold corresponds to a root-mean-
square deflection of the hair bundles of IHCs by 0.3 nm. These bundles are thought
to be coupled by fluid motion to the shearing between the reticular lamina and tecto-
rial membrane, and thus to the displacement of the hair bundles of OHCs (Supporting
Information).
ACKNOWLEDGMENTS. We thank S. Leibler for discussion, the two reviewers for
valuable suggestions, and the members of our research group for comments on the
manuscript. This research was supported by grant DC00241 from the National In-
stitutes of Health and by a fellowship to T. R. from the Alexander von Humboldt
Foundation. A. J. H. is an Investigator of Howard Hughes Medical Institute.
26. X. Dong, M. Ospeck, and K. H. Iwasa. Piezoelectric reciprocal relationship of the
membrane motor in the cochlear outer hair cell. Biophys. J., 82:1254 -- 1259, 2002.
27. B. N. Evans and P. Dallos. Stereocilia displacement induced somatic motility of
cochlear outer hair cells. Proc. Natl. Acad. Sci. U.S.A., 90:8347 -- 8351, 1993.
28. P. Horowitz and W. Hill. The Art of Electronics. Cambridge University Press, second
29.
edition, 1989.
I. Sziklai. The significance of the calcium signal in the outer hair cells and its possible
role in tinnitus of cochlear origin. Eur. Arch. Otorhinolaryngol., 261:517 -- 525, 2004.
30. J. J. Zwislocki and E. J. Kletsky. Tectorial membrane: a possible effect on frequency
analysis in the cochlea. Science, 204:639 -- 641, 1979.
31. A. W. Gummer, W. Hemmert, and H. P. Zenner. Resonant tectorial membrane motion
in the inner ear: its crucial role in frequency tuning. Proc. Natl. Acad. Sci. U.S.A.,
93:8727 -- 8732, 1996.
32. H. Cai, B. Shoelson, and R. S. Chadwick. Evidence of tectorial membrane radial
motion in a propagating mode of a complex cochlear model. Proc. Natl. Acad. Sci.
U.S.A., 101:6243 -- 6248, 2004.
33. J. Santos-Sacchi. On the frequency limit and phase of outer hair cell motility: effects
of the membrane filter. J. Neurosci., 12:1906 -- 1916, 1992.
34. G. D. Housley and J. F. Ashmore. Ionic currents of outer hair cells isolated from the
guinea-pig cochlea. J. Physiol., 448:73 -- 98, 1992.
35. J. Engel, C. Braig, L. Ruttiger, S. Kuhn, U. Zimmermann, N. Blin, M. Suasbier,
H. Kalbacher, S. Munkner, K. Rohbock, P. Ruth, H. Wintner, and M. Knipper. Two
classes of outer hair cells along the tonotopic axis of the cochlea. Neurosci., 143:837 --
849, 2006.
36. W. S. Rhode. Observations of the vibration of the basilar membrane in squirrel mon-
keys using the Mossbauer technique. J. Acoust. Soc. Am., 49:1218 -- 1231, 1971.
37. A. L. Nuttall and D. F. Dolan. Steady-state sinusoidal velocity responses of the basilar
membrane in guinea pig. J. Acoust. Soc. Am., 99:1556 -- 1565, 1996.
38. M. M. Mellado Lagarde, M. Drexl, V. A. Lukashkina, A. N. Lukashkin, and I. J. Rus-
sell. Outer hair cell somatic, not hair bundle, motility is the basis of the cochlear
amplifier. Nat. Neurosci., 11:746 -- 748, 2008.
39. P. Dallos, X. Wu, M. A. Cheatham, J. Gao, J. Zheng, C. T. Anderson, S. Jia, X. Wang,
W. H.Y. Cheng, S. Sengupta, D. Z. Z. He, and J. Zuo. Prestin-based outer hair cell
motility is necessary for mammalian cochlear amplification. Neuron, 58:333 -- 339,
2008.
40. W. S. Rhode and N. P. Cooper. Fast traveling waves, slow traveling waves and their
interactions in experimental studies of apical cochlear mechanics. Auditory Neurosci.,
2:289 -- 199, 1996.
International Team for Ear Research. Cellular vibration and motility in the organ of
Corti. Acta Oto-Laryngol. Suppl., 467:1 -- 279, 1989.
41.
42. R. Ghaffari, A. J. Aranyosi, and D. M. Freeman. Longitudinally propagating trav-
eling waves of the mammalian tectorial membrane. Proc. Natl. Acad. Sci. U.S.A.,
104:16510 -- 16515, 2007.
1. J. Lighthill. Energy flow in the cochlea. J. Fluid Mech., 106:149 -- 213, 1981.
2. F. Mammano and R. Nobili. Biophysics of the cochlea: Linear approximation. J.
Acoust. Soc. Am., 93:3320 -- 3332, 1993.
3. C. A. Shera, A. Tubis, and C. L. Talmadge. Do forward- and backward-traveling waves
occur within the cochlea? Countering the critique of Nobili et al. JARO, 5:349 -- 359,
2004.
4. W. M. Siebert. Stimulus transformations in the peripheral auditory system. In P. A.
Kolers and M. Eden, editors, Recognizing Patterns, pages 104 -- 133. MIT, Cambridge,
1968.
5. G. Zweig. Basilar membrane motion. Cold Spring Harb Symp Quant Biol, 40:619 -- 633,
1976.
6. R. C. Naidu and D. C. Mountain. Measurements of the stiffness map challenge a basic
tenet of cochlear theories. Hear. Res., 124:124 -- 131, 1998.
7. N. Y. S. Kiang, M. C. Liberman, W. F. Sewell, and J. J. Guinan. Single unit clues to
cochlear mechanics. Hear. Res., 22:171 -- 182, 1986.
8. M. Ulfendahl. Mechanical responses of the mammalian cochlea. Progr. Neurobiol.,
53:331 -- 380, 1997.
9. A. N. Temchin, N. C. Rich, and M. A. Ruggero. Threshold tuning curves of chinchilla
auditory-nerve fibers. I. Dependence on characteristic frequency and relation to the
magnitudes of cochlear vibrations. J. Neurophysiol., 100:2889 -- 2898, 2008.
10. N. P. Cooper and W. S. Rhode. Nonlinear mechanics at the apex of the guinea-pig
cochlea. Hear. Res., 82:225 -- 243, 1995.
11. C. Zinn, H. Maier, H. P. Zenner, and A. W. Gummer. Evidence for active, nonlin-
ear, negative feedback in the vibration response of the apical region of the in vivo
quinea-pig cochlea. Hear. Res., 142:159 -- 183, 2000.
12. L. Robles and M. A. Ruggero. Mechanics of the mammalian cochlea. Physiol. Rev.,
81:1305-1352, 2001.
13. N. P. Cooper and W. S. Rhode. Basilar membrane mechanics in the hook region of
cat and guinea-pig cochlea: sharp tuning and nonlinearity in the absence of baseline
position shift. Hear. Res., 63:191 -- 196, 1992.
14. M. A. Ruggero, N. C. Rich, A. Recio, S. S. Narayan, and L. Robles. Basilar-membrane
J. Acoust. Soc. Am.,
responses to tones at the base of the chinchilla cochlea.
101:2151 -- 2163, 1997.
15. S. M. Khanna and L. F. Hao. Reticular lamina vibrations in the apical turn of a living
guinea pig cochlea. Hear. Res., 132:15 -- 33, 1999.
16. W. S. Rhode and N. P. Cooper. Nonlinear mechanics in the apical turn of the chinchilla
cochlea in vivo. Auditory Neurosci., 3:101 -- 121, 1996.
17. P. Martin and A. J. Hudspeth. Active hair-bundle movements can amplify a hair cell's
response to oscillatory mechanical stimuli. Proc. Natl. Acad. Sci. U.S.A., 96:14306 --
14311, 1999.
18. P. Martin, A. J. Hudspeth, and F. Julicher. Comparison of a hair bundle's spontaneous
oscillations with its response to mechanical stimulation reveals the underlying active
process. Proc. Natl. Acad. Sci. U.S.A, 98:14380 -- 14385, 2001.
19. H. J. Kennedy, A. C. Crawford, and R. Fettiplace. Force generation by mammalian
hair bundles supports a role in cochlear amplification. Nature, 433:880 -- 883, 2005.
20. D. K. Chan and A. J. Hudspeth. Ca2+ current-driven nonlinear amplification by the
mammalian cochlea in vitro. Nat. Neurosci., 8:149 -- 155, 2005.
21. B. Nadrowski, P. Martin, and F. Julicher. Active hair-bundle motility harnesses noise
to operate near an optimum of mechanosensitivity. Proc. Natl. Acad. Sci. U.S.A.,
101:12195 -- 12200, 2004.
22. A. J. Hudspeth. Making an effort to listen: Mechanical amplification in the ear.
Neuron, 59:530 -- 545, 2009.
23. J. Zheng, W. Shen, D. Z. Z. He, K. B. Long, L. D. Madison, and P. Dallos. Prestin
is the motor protein of cochlear outer hair cells. Nature, 405:149 -- 155, 2000.
24. P. Dallos, J. Zheng, and M. A. Cheatham. Prestin and the cochlear amplifier. J.
Physiol., 576:37 -- 42, 2006.
25. J. Ashmore. Cochlear outer hair cell motility. Physiol. Rev., 88:173 -- 210, 2008.
6
SupportingInformationARatchetMechanismforAmplificationinLow-FrequencyMammalianHearingTobiasReichenbachandA.J.HudspethHowardHughesMedicalInstituteandLaboratoryofSensoryNeuroscience,TheRockefellerUniversity,NewYork,NewYork10065-6399,U.S.A.Herewefurtherdescribeourmodelingapproach.Weprovideadetaileddescriptionofthetwo-massmodelfortheorganofCortianddiscusstheenergybalance.Wethenturntothehydrodynamicsofthecochlea.Becauseourgoalistodemonstrateaprincipleratherthantodescribethecochleainfulldetail,wefocusonabasic,one-dimensionalmodel.WeconcludewithamodelfortheregulationoftheOHCs'membranepotentialwhichdemonstrateshowthecriticalmechanomotilitycoefficientα∗canresult.MicromechanicsoftheorganofCortiTodescribethemicromechanicsoftheorganofCortiweemployamodelwithtwodegreesoffreedom:themotionofthebasilarmembraneandthemotionofthehair-bundlecomplex(Figure2B).TheimpedanceZBMofthebasilarmembraneresultsfromamassmBM,viscousdampingλBM,andstiffnessKBM.Similarly,theimpedanceofthecomplexformedbythehairbundles,reticularlamina,andtectorialmembraneinvolvesmassmTM,viscousdampingλHB,andstiffnessKHB.Themotionsofthebasilarmembraneandofthehair-bundlecomplexarecoupledthroughtheimpedanceZCoftheorganofCorti,whichpossessesviscousandelasticcontributionsλCandKC.FurthercouplingarisesthroughthecombinedOHCsandDeiters'cells(Fig.S1).AnOHCcanbedescribedasapiezoelectricelement[1].ItslengthchangeδLdependsbothonachangeδVinthemembranepotentialaswellasontheappliedforceF.Consideringoscillatorymotionsatangularfrequencyω=2πfwecanwriteδL=fδLeiωt+c.c.,δV=fδVeiωt+c.c.,andF=eFeiωt+c.c.inwhichc.cdenotesthecomplexconjugate.ThelengthchangethenfollowsasfδL=−cfδV+c2eF(5)withcoefficientsc,c2.WedefineXEE=−cfδVasthepartofthelengthchangethatresultsfromvoltagechangesalone.Thelengthchangec2eFresultsfromtheOHC'sviscoelasticityZOHC=−iω−1c−12inserieswithXEE(Fig.S1).ItliesinserieswiththeDeiters'cellwhichcanbedescribedasanotherviscoleasticelementZDC.ThetwoviscoelasticelementsZOHCandZDCinseriescombineintotheviscoelasticelementZD=(Z−1OHC+Z−1DC)−1(6)1 .OHCRLDCBMZDCZOHCXHBδLXBMZCXEEFigureS1:Couplingbetweenthereticularlamina(RL)andthebasilarmembrane(BM).ThecouplingarisesfromOHCsinserieswithDeiters'cells(DC)andanimpedanceZCfromtheremainingorganofCorti.SeethetextforadescriptionofthecoupledOHCsandDeiters'cells.suchthattheschematicofFig.2Bresults.WedenotebyKDtheelasticcomponentandbyλDtheviscouspartoftheimpedanceZD.Theequationsofmotionforhair-bundleandbasilar-membranedisplacementXHBandXBMthenreadmTM∂2tXHB+λHB∂tXHB+KHBXHB+(KD+λD∂t)(XHB−XEE−XBM)+(KC+λC∂t)(XHB−XBM)=Fint,mBM∂2tXBM+λBM∂tXBM+KBMXBM−(KD+λD∂t)(XHB−XEE−XBM)−(KC+λC∂t)(XHB−XBM)=Fext.(7)AsoundstimulusatangularfrequencyωprovidesanexternalforceFextonthebasilarmembranewhosedependenceontimetmaybewrittenasFext(t)=Fexteiωt+c.c.withtheFouriercoefficientFext.Theevokedoscillationsofthebasilarmembrane,XBM,andofthehair-bundlecomplex,XHB,occurdominantlyatthesamefrequencyf;additionalfrequenciesmayarisefromnonlineareffectsowingtointernalforcesinthehairbundle.WeconsiderthecorrespondingFouriercoefficientsXBMandXHB.TheelectricallyevokedlengthchangeXEEofOHCsdependslinearlyonthehair-bundledisplacement,forthemembrane-potentialchangedependslinearlyonhair-bundlemotion(seeSupplementarySectionontheelectricalregulationoftheOHCmembranepotential):XEE=−αXBMwithacomplex,frequency-dependentmechanomotilitycoefficientα.SupplementaryEquation(7)thenyieldsEquations(1)and(2)withZHB=iωmTM+λHB−iω−1KHB,ZBM=iωmBM+λBM−iω−1KBM,ZC=λC−iω−1KC,ZD=λD−iω−1KD.(8)Theforceswithinthehairbundledependonitsdisplacement;theireffectistocounterviscousdampingaswellastoinfluencetheresonantfrequencyofthehair-bundlecomplex.Wemaydecomposetheseforcesintolinearand2 nonlinearparts:Fint=iωZactHBXHB+Fnonlinint(XHB).(9)Herethelineartermdominatesthedynamicsforsmalldisplacementsensuingfromweaksoundstimulianddescribesthefullyactivescenario.Thenonlineartermconnectstheactivetothepassivecasethatarisesforstrongsoundstimuli.Atthecriticalvalueα∗,thedisplacementsaregivenbyEquations(3),whichcanberewrittenusingSupple-mentaryEquation(9)asXHB=1iω(ZHB−ZactHB)Fnonlinint(XHB)+ZD+ZCiω(ZHB−ZactHB)(ZBM+ZD+ZC)Fext,XBM=1iω(ZBM+ZD+ZC)Fext.(10)Resonanceinhair-bundlemotionoccursatthefrequencyforwhichZHB−ZactHB=0.Activehair-bundlemotilitymakesthissituationpossible.Asdiscussedinthemaintext,thisresonanceconditionisindependentoftheprop-ertiesofthebasilarmembrane.CochlearhydrodynamicsandtravelingwavesWenowincorporatethedescriptionforthemicromechanicsoftheorganofCortiintoamodelforthewholecochlea.Inthesimplestform,thecochleaisconsideredasone-dimensional,exhibitingaslowpressurewave.Combiningcontinuityequationsandfluid-momentumequations,thispressurewaveobeysthepartialdifferentialequationρ∂2tXBM+Λ∂tXBM=12L2∂r(h∂rp).(11)Hereρdenotesthedensityofliquidinthecochlea,Lthelengthofthecochlea,htheheightofthescalae,andpandXBMrespectivelythepressureacrossandthedisplacementofthebasilarmembraneatpositionrandtimet.ThecoefficientΛaccountsforfrictionduetofluidmotion.Positionismeasuredinunitsofthecochlearlength,suchthatr=0correspondstothebasalandr=1totheapicalend.Weapplytwoboundaryconditions.First,p=p0atr=0:asound-evokedpressurep0actsatthestapes.Andsecond,p=0atr=1:becausethetwoscalaecommunicateatthehelicotrema,thepressuredifferencebetweenthemvanishesattheapicalendofthecochlea.ConsideringtheFouriercoefficientsofangularfrequencyω,weobtain−ω2ρXBM+iωΛXBM=12L2∂r(h∂rp).(12)Thebasilar-membranedisplacementXBMandpressurepareconnectedthroughEquation(1),withp=Fext/ABM.ABMdenotestheareaofatransversestripofthebasilarmembranethathasthewidthofonecell,about8µm;allparametersemployedinthetwo-massmodelrefertoatransversesegmentofthecochlearpartitionofthatwidth.WesolveSupplementaryEquation(12)numericallywiththeshootingmethodinMathematica7(WolframResearch).3 Theheighthofthescalaevariesfromabout1mmatthebasetoabout200µmattheapex.Measurementsofthefluidpressureinthebasalregionindicate,however,thatthepenetrationdepthofthewaveissignificantlysmaller[2].Inourone-dimensionalsimulations,wethereforeuseaneffectiveheightof100µm.Concerningthemapofcharacteristicfrequenciesf0(r)alongthecochlea,weconsiderthechinchilla'shearingrangefromamaximalfrequencyofabout30kHzattheextremebasetoaminimumofabout50Hzattheextremeapex[3].WeassumethattheresonantfrequencyofthebasilarmembranecannotcoverthiswholerangeandisinsteadgivenbyfBM(r)=qf2BM,apex+[1−(fBM,apex)/fmax)2]f0(r)2.(13)Herefmax=30kHzdenotesthehighestcharacteristicfrequencyandfBM,apex=1kHzthelowestresonantfrequencythatthebasilarmembraneexhibits.fBMcoincideswithf0forhighfrequenciesbutdeviatesforfrequenciesbelowfLandapproachesfBM,apexintheapicalregionofthecochlea(Figure3B).TheresonantfrequencyfBMofthebasilarmembraneissetbyitsmassandstiffnessaccordingtofBM=(2π)−1pKBM/mBM.WeconsiderKBMtovaryproportionaltofBMandmBMtovaryinversely:KBM(r)=KmaxBMfBM(r)/fmaxandmBM(r)=KBM(r)/[2πfBM(r)]2,withKmaxBM=2N·m−1.ThemassmTMofthetectorialmembraneisassumedtofollowthesamespatialvariationasthebasilarmembranemass,albeitwithasmallervalue:mTM=mBM/5.Theincreaseinthebasilar-membraneandtectorial-membranemassfrombasetoapexpresumablyreflectstheincreasingsizeoftheorganofCortitowardstheapex.Forthevariationofthemechanomotilitycoefficientαalongthecochlea,anditsdependenceonthestimulusfrequencyf,wemaketheansatzα(r,f)=f4cf4c+f0(r)4×[δf(r)f0]2[f2−f0(r)2]2+[δff0(r)]2×α∗.(14)Thefirstfactoraccountsforthehigh-frequencycutoffabovewhichthemembranepotential,andthuselectromotility,cannolongerfollowhair-bundledisplacementonacycle-by-cyclebasis;weusefc=4kHz.Toyieldtheratchetmechanismatlowerfrequencies,wechooseαtobeα∗atthenaturalfrequency.Thesecondfactordescribeshowtheresponsechangeswhenthefrequencyfofstimulationdeviatesfromthenaturalfrequencyf0;weassumethatαisclosetoα∗aslongasbothfrequenciesaresimilar,butotherwisedeclinesinmagnitude.Thewidthofthecorrespondingcurveisdeterminedbytheparameterδ,whichwesetatδ=2.Themechanomotilitycoefficientα(r,f=f0)isshowninFigure3A.Thelinearpartoftheactivehair-bundleforcecountersviscousdampingandprovidesavanishingimpedanceatthenaturalfrequency.UsingEquations(1),(2),andSupplementaryEquation(9),thistranslatesintoZHB−ZactHB(cid:12)(cid:12)(cid:12)f=f0=−ZBM[(1+α)ZD+ZC]ZBM+ZC+ZD(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)f=f0.(15)WeassumethattheimaginarypartofZHBistunedtothenaturalfrequencyf0suchthattheactivecontributionpossessesonlyarealpartcorrespondingtonegativedamping,ZactHB≡λactHB.Notethatintheratchetmechanism,whenα=α∗,theright-handsideofSupplementaryEquation(15)vanishes,soactivehair-bundleforcesneedtoovercomeonlydampingassociatedwithhair-bundlemotion.4 ParameterDescriptionValueReferenceLlengthofthecochlea20mm[4]Λfluidfriction5×105N·s·m−4wBM(r)widthofthebasilarmembrane(50+200r)µm[5]ABM(r)areaofthebasilar-membranestripwBM(r)×8µmλHBfrictionofthehair-bundlecomplex1µN·s·m−1[6]KIHBstiffnessofthehairbundlesofIHCs2mN·m−1[7]λIHBfrictionofthehairbundlesofIHCs1µN·s·m−1[6]λBM(r)frictionofthebasilarmembrane0.03×wBM(r)N·s·m−2KCstiffnessoftheorganofCorti10mN·m−1λCfrictionoftheorganofCorti10µN·s·m−1KDstiffnessofOHCsandDeiters'cells20mN·m−1[8]λDfrictionofOHCsandDeiters'cells10µN·s·m−1S1:Parametersusedinthecochlearmodel.Whenαandactivehair-bundleforcesobeySupplementaryEquations(14)and(15)andnonlinearitiesfromthehair-bundleforcesareignored,thehair-bundledisplacementdivergesatresonance.Intheactualcochlea,noiseaswellasnonlinearitiessuperveneandleadtoalargebutfinitedisplacement.Inoursimulations,weincorporatethiseffectbyconsideringvaluesforαandZactHBthatdeviateby1%fromtheiridealvaluesgivenbySupplementaryEquations(14)and(15).Tuningcurvesofauditory-nervefibersTuningcurvesofauditory-nervefibersarecomputedbyassumingthatthehearingthresholdcorrespondstoaroot-mean-squaredeflectionofthehairbundlesofIHCsby0.3nm.Thesebundlesarethoughttobecoupledbyfluidmotiontotheshearingbetweenthereticularlaminaandtectorialmembrane,andthustothedisplacementofthehairbundlesofOHCs.AssumingthatthedisplacementXIHBofthehairbundlesofinnerhaircellsisdependentonfrictionλIHBandelasticityKIHB,itcouplestotheOHCs'bundlemotionthroughtherelationλIHB∂t(XIHB−XHB)+KIHBXIHB=0.(16)UponFouriertransformation,weobtainXIHB=iωλIHBiωλIHB+KIHBXHB.(17)Forhighfrequencies,viscouscouplingisstrongandthehair-bundledisplacementsofIHCsandOHCscoincide.Forlowfrequencies,thecouplingdecreasesandthehair-bundlemotionofIHCsremainssmallerthanthatofOHCs.Thiseffectunderliestherisingthresholdsatlowfrequenciesseenontheleftlimbsofhigh-frequencytuningcurves(Figure5)aswellasinexperimentalmeasurements.Thevalueof0dBSPLisdefinedbyaroot-mean-squaresound-pressurestimulusof20µPa.Becausethis5 ABNa+Ca2+,K+K+XHBVVERaCbVPRbFigureS2:RegulationoftheOHCmembranepotential.(A)Hair-bundledeflectionopensionchannels,K+andCa2+flowintothecell.ThecellmembranecontainsvariousadditionalionchannelsthatareregulatedbythemembranepotentialandtheCa2+concentration.Inourmodel,weconsidervoltage-regulatedoutwardK+channelsandinwardNa+channels.(B)AnequivalentcircuitofanOHCwithavariableresistanceRacontrolledbyhair-bundledisplacement,amembranecapacitanceCb,andabasoleteralmembraneconductanceRb.externalpressureisenhancedbythemiddleearbeforeactingonthestapes,weassumeanincreasebyafactorof20(26dB)[9].Iftheyhavenotbeendetailedinthetextabove,theparametersofthetwo-massmodelaswellastheone-dimensionaldescriptionsofthecochleaandhairbundledynamicsofIHCsaregiveninTableS1.ElectricalregulationoftheOHCmembranepotentialDeflectionXHBoftheOHCshairbundlesevokesachangeδV≡V−V0inthemembranepotentialfromitsrestingvalueV0,whichleadstoalengthchangeXEE=−cδV(seeSupplementaryEquation5andbelow)withacoefficientcofabout20µm·V−1[1].Theratchetmechanismreliesonacriticalvalueα∗=−1−ZC/ZDoftheratioα=−XEE/XHBbetweenhair-bundledisplacementandcelllengthchange.WhileitsexactvaluedependsonthecouplingimpedancesZCandZD,therealpartofα∗isnegative.AtypicalsituationZC=ZDleadstothecriticalvalueα∗=−2.Hereweshowhowsuchavalueofα∗canemergewithinabasicmodelfortheregulationoftheOHCsmembranepotential.Themodelreliesonvoltage-regulatedoutwardK+channels[10]thatcounteractthecurrentthroughthemembranecapacitance,aswellasavoltage-regulatedinwardsodiumcurrent[11;12]thatcanleadtohyperpolarizationuponpositivehair-bundledeflection.TheapicalsurfaceofanOHC,includingthehairbundle,isbathedinendolymphatapotentialVEofabout80mV(FigureS2).ThebasolateralpartissurroundedbyperilymphatgroundpotentialVP.Theelectricalbehaviorcanberepresentedbyasimplecircuitdescribedbyanequationforcurrentconservation:IRa=IRb+ICb.(18)6 00.511.522.533.5101102103104060120180240300360MechanomotilitygainαPhaseofα(◦)Frequency(Hz)00.511.522.533.5101102103104060120180240300360MechanomotilitygainαPhaseofα(◦)Frequency(Hz)FigureS3:Themodeledmechanomotilitycoefficientα.Weshowthedependenceofthemagnitude(black)andphase(red)ofαonthesoundfrequency.Forf0=200Hzα=−2emergeswhichrepresentsatypicalvalueforthecriticalα∗.Themagnitudeofαisstronglyattenuatedforhigherfrequencies.ByvaryingthevoltagedependencyofthebasalK+andNa+channelsthecriticalvalueα∗canemergeatdifferentfrequenciesf0belowafewkilohertz.HereIRadenotesthecurrentthroughtheresistanceRaandIRbandICbdenotethecurrentsthroughRbandCb,respectively.WeassumethatRaresultsfromtheconductivitygaofthemechanotransductionchannelsinthehairbundle;thecurrentismainlycarriedbyK+ions[13].Theconductanceofthebasolateralcellmembranestemsfromvoltage-regulatedK+channels(conductancegKb)aswellasvoltage-regulatedNa+-channels[11;12](conductancegNab).Weobtainga(VE−V)=Cb∂tV+gKb(V−VP)−gNab(VNa−V)(19)inwhichVNadenotestheNa+reversalpotentialandCbthecellmembrane'scapacitance.ThepotentialV0insidethecellatvanishinghair-bundledisplacementXHB=0followsasV0=g(0)aVE+gNa(0)bVNa+gK(0)bVPg(0)a+gNa(0)b+gK(0)b(20)inwhichthesuperscript(0)denotestherespectiveconductancesatvanishinghair-bundledisplacement.Hair-bundledisplacementXHBchangestheconductancegaofthecell'sapicalmembrane.Tolinearoderweobtainga=g(0)a+∂ga∂XHB(cid:12)(cid:12)(cid:12)(cid:12)(0)XHB.(21)ThealteredapicalconductanceevokesachangesδV=V−V0inthemembranepotentialwhichinfluencesthebasalK+andNa+conductances.TheK+channelsarevoltage-regulatedatanactivationtimeτ.ThedynamicsofthechangeδgKb=gKb−gK(0)bcanbedescribedbyτ∂tδgKb=∂gKb∂V(cid:12)(cid:12)(cid:12)(cid:12)(0)δV−δgKb(22)7 ParameterDescriptionValueReferenceVEpotentialoftheendolymph80mV[14]VPpotentialoftheperilymph−100mV[14]VNaNa+reversalpotential22mV[11]gK+aapicalconductivityofK+channels5nS[15]gKbbasoleteralK+conductivity10nS[10]Cbmembranecapacitance40pF[10]∂gK+a/∂XHBdependenceofapicalK+conductivityonXHB5×10−3S·m−1[16]τactivationtimeofvoltage-regulatedK+channels1ms[10]S2:ParametersformodelingtheregulationoftheOHCsmembranepotential.whichresultsineδgKb=11+iωτ∂gKb∂V(cid:12)(cid:12)(cid:12)(cid:12)(0)fδV(23)foroscillatorystimuliXHB=XHBeiωt+c.c.atangularfrequencyω.TheNa+channelsarerapidlyactivated;tolinearordertheirconductancefollowsasgNab=gNa(0)b+∂gNab∂V(cid:12)(cid:12)(cid:12)(cid:12)(0)δV.(24)LinearizationofSupplementaryEquation(19)yields"iωCb+ga+gKb+gNab+1−iωτ1+ω2τ2∂gKb∂V(cid:12)(cid:12)(cid:12)(cid:12)(0)(V−VP)−∂gNab∂V(cid:12)(cid:12)(cid:12)(cid:12)(0)(VNa−V)#fδV=∂ga∂XHB(cid:12)(cid:12)(cid:12)(cid:12)(0)XHB.(25)Themechanomotilitycoefficientαfollowsasα=cfδV/XHB.FromSupplementaryEquation(25)wenotethatdelayedvoltage-regulatedK+channelscanreducetheeffectofthemembranecapacitance.AlargeenoughvoltagedependenceoftheinwardNa+channelscanthenyieldhyperpolarizationofthecelluponpositivedeflectionofthehairbundle,asrequestedforthecriticalvalueα∗.FigureS3showstheresultingmagnitudeandphaseofαfortypicalparametersforapicalOHCs(TableS2)independenceonthefrequencyofhair-bundlestimulation.Thevalues∂gKb/∂V=1.36µS·V−1and∂gNab/∂V=1.85µS·V−1havebeenchosentoyieldthecriticalvalueα∗=−2atthefrequencyf0=200Hzandlieintherangeofmeasuredvalues[10;11].8 References[1]X.Dong,M.Ospeck,andK.H.Iwasa.Piezoelectricreciprocalrelationshipofthemembranemotorinthecochlearouterhaircell.Biophys.J.,82:1254 -- 1259,2002.[2]E.Olson.Directmeasurementofintra-cochlearpressurewaves.Nature,402:526 -- 529,1999.[3]R.S.HeffnerandH.E.Heffner.Behavioralhearingrangeofthechinchilla.Hear.Res.,52:13,1991.[4]P.A.SantiandD.C.Muchow.Morphometryofthechinchillaorganofcortiandstriavascularis.J.Histochem.andCytochem.,27:1539 -- 1542,1979.[5]L.I.ProkofjevaandA.G.Chernyi.Quantitativecharacteristicsofthebasilarmembraneinthemammaliancochlea.NauchnyeDokiVyssShkolyBiolNauki.,11:44 -- 50,1986.[6]P.Martin,A.J.Hudspeth,andF.Julicher.Comparisonofahairbundle'sspontaneousoscillationswithitsresponsetomechanicalstimulationrevealstheunderlyingactiveprocess.Proc.Natl.Acad.Sci.U.S.A,98:14380 -- 14385,2001.[7]D.K.ChanandA.J.Hudspeth.MechanicalresponsesoftheorganofCortitoacousticandelectricalstimulationinvitro.Biophys.J.,89:4382 -- 4395,2005.[8]J.Ashmore.Cochlearouterhaircellmotility.Physiol.Rev.,88:173 -- 210,2008.[9]S.Puria,W.T.Peake,andJ.J.Rosowski.Sound-pressuremeasurementsinthecochlearvestibuleofhuman-cadaverears.J.Acoust.Soc.Am.,101:2754 -- 2770,1997.[10]F.MammanoandJ.F.Ashmore.Differentialexpressionofouterhaircellpotassiumcurrentsintheisolatedcochleaoftheguinea-pig.J.Physiol.,469:639 -- 646,1996.[11]C.M.Witt,H.Y.Hu,W.E.Brownell,andD.Bertrand.Physiologicallysilentsodiumchannelsinmammalianouterhaircells.J.Neurophysiol.,72:1037 -- 1040,1994.[12]H.P.ZennerD.Oliver,P.PlinkertandJ.P.Ruppersberg.Sodiumcurrentexpressionduringpostnataldevelopmentofratouterhaircells.PflugersArch.Eur.J.Physiol.,434:1432 -- 2013,1997.[13]A.J.Hudspeth.Howtheear'sworkswork.Nature,341:397 -- 404,1989.[14]P.Mistrik,C.Mullaley,F.Mammano,andJ.Ashmore.Three-dimensionalcurrentflowinalarge-scalemodelofthecochleaandthemechanismofamplificationofsound.J.Roy.Soc.Int.,6:279 -- 291,2009.[15]A.J.Ricci,A.C.Crawford,andR.Fettiplace.Tonotopicvariationintheconductanceofthehaircellmechanotransducerchannel.Neuron,40:983 -- 990,2003.[16]M.Beurg,M.G.Evans,C.M.Hackney,andR.Fettiplace.ALarge-ConductanceCalcium-SelectiveMechan-otransducerChannelinMammalianCochlearHairCells.J.Neurosci.,26:10992 -- 11000,2006.9 |
1602.08299 | 3 | 1602 | 2017-02-01T01:21:09 | Theory of population coupling and applications to describe high order correlations in large populations of interacting neurons | [
"q-bio.NC",
"physics.bio-ph"
] | To understand the collective spiking activity in neuronal populations, it is essential to reveal basic circuit variables responsible for these emergent functional states. Here, I develop a mean field theory for the population coupling recently proposed in the studies of visual cortex of mouse and monkey, relating the individual neuron activity to the population activity, and extend the original form to the second order, relating neuron-pair's activity to the population activity, to explain the high order correlations observed in the neural data. I test the computational framework on the salamander retinal data and the cortical spiking data of behaving rats. For the retinal data, the original form of population coupling and its advanced form can explain a significant fraction of two-cell correlations and three-cell correlations, respectively. For the cortical data, the performance becomes much better, and the second order population coupling reveals non-local effects in local cortical circuits. | q-bio.NC | q-bio |
Theory of population coupling and applications to describe high order correlations in
large populations of interacting neurons
RIKEN Brain Science Institute, Wako-shi, Saitama 351-0198, Japan
(Dated: September 24, 2018)
Haiping Huang
To understand the collective spiking activity in neuronal populations, it is essential to reveal
basic circuit variables responsible for these emergent functional states. Here, I develop a mean field
theory for the population coupling recently proposed in the studies of visual cortex of mouse and
monkey, relating the individual neuron activity to the population activity, and extend the original
form to the second order, relating neuron-pair's activity to the population activity, to explain the
high order correlations observed in the neural data.
I test the computational framework on the
salamander retinal data and the cortical spiking data of behaving rats. For the retinal data, the
original form of population coupling and its advanced form can explain a significant fraction of
two-cell correlations and three-cell correlations, respectively. For the cortical data, the performance
becomes much better, and the second order population coupling reveals non-local effects in local
cortical circuits.
PACS numbers: 87.19.L-, 02.50.Tt, 75.10.Nr
I.
INTRODUCTION
To uncover the neural circuit mechanisms underlying animal behavior, e.g., working memory or decision making,
is a fundamental issue in systems neuroscience [1, 2]. Recent developments in multi-neuron recording methods make
simultaneous recording of neuronal population activity possible, which gives rise to the challenging computational
tasks of finding basic circuit variables responsible for the observed collective behavior of neural populations [3]. The
collective behavior arises from interactions among neurons, and forms the high dimensional neural code. To search
for a low dimensional and yet neurobiologically plausible representation of the neural code, thus becomes a key step
to understand how the collective states generate behavior and cognition.
Correlations among neurons' spiking activities play a prominent role in deciphering the neural code [4]. Various
models were proposed to understand the pairwise correlations in the population activity [5 -- 7]. Modeling these
correlations sheds light on the functional organization of the nervous system [8]. However, as the population size
grows, higher order correlations have to be taken into account for modeling synchronous spiking events, which are
believed to be crucial for neural information transmission [9 -- 11].
In addition, the conclusion drawn from small
size populations may not be correct for large size populations. Theoretical studies have already proved that high
order interactions among neurons are necessary for generating widespread population activity [12, 13]. However,
introduction of high order multi-neuron couplings always suffers from a combinatorial explosion of model parameters
to be estimated from the finite neural spike train data.
To account for high order correlations, various models with different levels of approximation were proposed, for
example, the reliable interaction model [14] with the main caveat that the rare patterns are discarded during inference
of the coupling terms, the dichotomized Gaussian model [7, 15] in which correlations among neurons are caused by
common Gaussian inputs to threshold neurons, the K-pairwise model [16, 17] in which an effective potential related
to the synchronous firing of K neurons was introduced, yet hard to be interpreted in terms of functional connectivity,
and the restricted Boltzmann machine [18] where hidden units were shown to be capable of capturing high order
dependences but their number should be predefined and difficult to infer from the data [19]. One can also take into
account part of the statistical features of the population activity (e.g., simultaneous silent neural pattern) and assume
homogeneity for high order interactions among neurons due to the population size limitation [20].
In this paper,
I provide a low dimensional neurobiological model for describing the high order correlations and extracting useful
information about neural functional organization and population coding.
In this study, I interpret correlations in terms of population coupling, a concept recently proposed to understand
the multi-neuron firing patterns of the visual cortex of mouse and monkey [21]. The population coupling characterizes
the relationship of the activity of a single neuron with the population activity; this is because, the firing of one neuron
is usually correlated with the firing pattern of other neurons. I further generalize the original population coupling to
its higher order form, i.e., the relationship of pairwise firing with the population activity. I then derive the practical
dimensionality reduction method for both types of population couplings, and test the method on different types of
neural data, including ganglion cells in the salamander retina onto which a repeated natural movie was projected [17],
and layer 2/3 as well as layer 5 cortical cells in the medial prefrontal cortex (MPC) of behaving rats [22].
In this paper, I develop a theoretical model of population coupling and its advanced form, to explain higher order
correlations in the neural data. Methodologically, I propose the fast mean field method not only to learn the population
couplings but also to evaluate the high order correlations. Note that this is computationally hard in a traditional
maximum entropy model by using sampling-based method. Conceptually, I generalize the normal population coupling
by introducing the second-order population coupling, which reveals interesting features from the data. First, it can
explain a significant amount of three-cell correlations, and it works much better in cortical data than in retinal data.
Second, the second-order population coupling matrix has distinct features in retinal and cortical data. The cortical
one shows clear stripe-like structure while the retinal one has no such apparent structure. Altogether, this work marks
a major step to understand the low-order representation of complex neural activity in both concepts and methods.
2
II. MODEL DESCRIPTION AND MEAN-FIELD METHODS
For a neuronal population of size N , the neural spike trains of duration T are binned with temporal resolution τ ,
yielding M = (cid:100)T /τ(cid:101) samples of N -dimensional binary neural firing patterns. I use si = +1 to denote firing state of
neuron i, and si = −1 for silent state. Neural responses to repeated stimulus (or the same behavioral tasks) vary
substantially (so-called trial-to-trial variability) [23, 24]. To model the firing pattern statistics, I assign each firing
pattern s a cost function (energy in statistical physics jargon) E(s), then the probability of observing one pattern s
can be written as P (s) ∝ exp(−E(s)), where
E(s) = −(cid:88)
hisi −(cid:88)
i
i
(cid:33)
(cid:32)(cid:88)
j(cid:54)=i
Jisi
sj
.
(1)
This is the first low dimensional representation to be studied. High energy state s corresponds to low probability of
observation. hi is the firing bias constraining the firing rate of neuron i, while Ji characterizes how strongly neuron
i's spiking activity correlates with the population activity measured by the sum of other neurons' activity. I name Ji
the first order population coupling (PopC1). Thus, only 2N parameters needs to be estimated from the neural data.
This number of model parameters is much less than that in conventional maximum entropy model [17].
To model the high order correlation (e.g., three neuron firing correlation), I further generalize PopC1 to its advanced
form, i.e., the second order population coupling, namely PopC2, describing the relationship of pairwise firing with the
population activity, and the corresponding energy is given by
E(s) = −(cid:88)
hisi −(cid:88)
i
i<j
(cid:33)
(cid:32)(cid:88)
k(cid:54)=i,j
wijsisj
sk
,
(2)
where wij characterizes how strongly the firing state of the neuron pair (ij) correlates with the firing activities of
other neurons. This term is expected to increase the prediction ability for modeling high order correlations in the
neural data. Under the framework of PopC2, the total number of parameters to be estimated from the data is
N + N (N − 1)/2. PopC1 and PopC2 have a clear neurobiological interpretation (for PopC1, see a recent study [21],
and the results obtained under the PopC2 can also be experimentally tested), and moreover they can be interpreted
in terms of functional interactions among neurons (as shown later).
To find the model parameters as a low dimensional representation, I apply the maximum likelihood learning principle
corresponding to maximizing the log-likelihood ln P (s) with respect to the parameters. The learning equation for
PopC1 is given by
(cid:32)(cid:68)
(cid:69)
(cid:32)(cid:88)
(cid:68)
si
j(cid:54)=i
data
sisj
−(cid:68)
(cid:69)
si
(cid:33)
(cid:69)
(cid:68)
−(cid:88)
model
,
sisj
data
j(cid:54)=i
ht+1
i = ht
i + η
J t+1
i = J t
i + η
(cid:33)
,
(cid:69)
model
(3a)
(3b)
where t and η denote the learning step and learning rate, respectively. The maximum likelihood learning shown here
has a simple interpretation of minimizing the Kullback-Leibler divergence between the observation probability and
the model probability [25, 26]. In an analogous way, one gets the learning equation for PopC2,
(cid:32)(cid:88)
(cid:68)
k(cid:54)=i,j
(cid:69)
(cid:68)
− (cid:88)
k(cid:54)=i,j
(cid:33)
(cid:69)
wt+1
ij = wt
ij + η
sisjsk
data
sisjsk
model
.
(4)
In the learning equations Eq. (3) and Eq. (4), the data dependent terms can be easily computed from the binned
neural data. However, the model expectations of the firing rate (magnetization in statistical physics) and correlations
are quite hard to evaluate without any approximations. Here I use the mean field method to tackle this difficulty.
First, I write the energy term into a unified form,
E(s) = −(cid:88)
hisi −(cid:88)
i
a
(cid:89)
i∈∂a
Γa
si,
(5)
where a denotes the interaction index and ∂a denotes the neuron set involved in the interaction a. a = (ij) for
PopC1 and (ijk) for PopC2. Therefore, PopC1 introduces the pairwise interaction as Γij = Ji + Jj, while PopC2
introduces the triplet interaction as Γijk = wij + wjk + wik. The multi-neuron interaction in the conventional Ising
model is decomposed into first order or second order population coupling terms. This decomposition still maintains
the functional heterogeneity of single neurons or neuron pairs, but reduces drastically the dimensionality of the neural
representation for explaining high order correlations. In principle, one can combine PopC1 and PopC2 to predict both
pairwise and triplet correlations. However, in this work, I focus on the pure effect of each type of population coupling.
In fact, the conventional Ising model [27] can be recovered by setting Γij = Jij, which is pairwise interaction. The
learning equation is derived similarly, and is run by reducing the deviation between the model pairwise correlation
and the clamped one (computed from the data) [28].
Second, the statistical properties of the model (Eq. 5) can be analyzed by the cavity method in the mean field
theory [29]. The self-consistent equations are written in the form of message passing (detailed derivations were given
in Refs [30, 31], see also Appendix A) as
hi +
(cid:89)
(cid:88)
b∈∂i\a
j∈∂b\i
mi→a = tanh
tanh
−1 mb→i
mb→i = tanh Γb
mj→b,
,
3
(6a)
(6b)
(7)
where ∂b\i denotes the member of interaction b except i, and ∂i\a denotes the interaction set i is involved in with a
removed. mi→a is interpreted as the message passing from the neuron i to the interaction a it participates in, while
mb→i is interpreted as the message passing from the interaction b to its member i. This iteration equation is also
called the belief propagation (BP), which serves as the message passing algorithm for the statistical inference of the
model parameters. Iteration of the message passing equation on the inferred model would converge to a fixed point
corresponding to a global minimum of the free energy (in the cavity method approximation [31])
F ≡ − ln Z = −(cid:88)
(cid:88)
(cid:1). I define the function Hi(x) ≡ exhi(cid:81)
ln Zi +
a
i
(∂a − 1) ln Za,
ron is − ln Zi = − ln(cid:80)
ln(cid:0)1 + tanh Γa
(cid:81)
where Z is the normalization constant of the model probability P (s). The free energy contribution of one neu-
x=±1 Hi(x) and the free energy contribution of one interaction is − ln Za = − ln cosh Γa −
b∈∂i cosh Γb(1 + x mb→i). At the same time, the
i∈∂a mi→a
model firing rate and multi-neuron correlation can be estimated as
mi = tanh
tanh
−1 mb→i
(cid:32)
(cid:88)
tanh Γa +(cid:81)
(cid:81)
hi +
b∈∂i
(cid:33)
Magnetization and correlation are defined as mi = (cid:104)si(cid:105) and Ca =(cid:10)(cid:81)
1 + tanh Γa
Ca =
i∈∂a mi→a
i∈∂a mi→a
.
i∈∂a si
,
(cid:11), respectively.
(8a)
(8b)
A brief derivation of Eq. (8) is given in Appendix A. Here the multi-neuron correlation is calculated directly from
the cavity method approximation and expected to be accurate enough for current neural data analysis [28]. This is
because, correlations under the model are evaluated taking into account nearest-neighbor interactions, rather than
naive full independence among neurons. This approximation is expected to work well in a weakly-correlated neural
population [5, 8], where long-range strong correlations do not develop. A similar application of this principle revealed
a non-trivial geometrical structure of population codes in salamander retina [28]. Another advantage is the low
computation cost. Both the free energy and the pairwise correlations can be estimated by the time complexity of the
order O(N 2) for PopC1, and O(N 3) for triplet correlations in PopC2, which is one order of magnitude lower than the
as S = −(cid:80)
tractable model of PopC1 recently proposed in Ref. [32]. A more accurate expression could be derived from linear
response theory [33] with much more expensive computational cost (increased by an order of magnitude (O(N )).
To estimate the information carried by a neural population, one needs to compute the entropy, which is defined
s P (s) ln P (s), and it measures the capacity of the neural population for information transmission. The
more obvious variability the neural responses have, the larger the entropy value is. The entropy of the model can
be estimated from the fixed point of the message passing equation. Based on the standard thermodynamic relation,
S = −F + E, where E is the energy of the neural population and given by
4
(9a)
(9b)
(9c)
(9d)
mj→b
(cid:89)
j∈∂b\i
a
i
hi
∆Ei +
(∂a − 1)∆Ea,
x=±1 Gi(x)
E = −(cid:88)
(cid:88)
(cid:80)
x=±1 xHi(x) +(cid:80)
(cid:80)
tanh Γa +(cid:81)
x=±1 Hi(x)
(cid:81)
Γb sinh Γb(1 + x mb→i) + xΓb cosh Γb(1 − tanh2 Γb)
i∈∂a mi→a
i∈∂a mi→a
1 + tanh Γa
,
,
Gi(x) =
exhi
cosh Γa(1 + x ma→i).
∆Ei =
∆Ea = Γa
(cid:88)
× (cid:89)
b∈∂i
a∈∂i\b
The basic procedure to infer population couplings is given as follows. At the beginning, all model parameters are
assigned zero value. It is followed by three steps: (i) Messages are initialized randomly and uniformly in the interval
(−1, 1). (ii) Eq. (6) are then run until converged, and the magnetizations as well as multi-neuron correlations are
estimated using Eq. (8). (iii) The estimated magnetizations and correlations are used at each gradient ascent learning
step (Eq. (3) or Eq. (4)). When one gradient learning step is finished, another step starts by repeating the above
procedure (from (i) to (iii)). To learn the higher order population coupling, the damping technique is used to avoid
oscillation behavior, i.e., wnew = γwnew + (1 − γ)wold where γ is the damping factor taking a small value.
The inferred model can also be used to generate the distribution of spike synchrony, i.e., the probability of K
simultaneous spikes. This distribution can be estimated by using Monte Carlo (MC) simulation on the model. The
standard procedure goes as follows. The simulation starts from a random initial configuration s0, and tries to search
for the low energy state, then the energy is lowered by a series of elementary updates, and for each elementary
update, N proposed neuronal state flips are carried out. That is, the transition probability from state s to s(cid:48) with
only si flipped (s(cid:48)
j∈∂a\i sj. The equilibrium samples
are collected after sufficient thermal equilibration. These samples (a total of 20000 samples in simulations) are finally
used to estimate the distribution of spike synchrony.
i = −si) is expressed as e−2siHi where Hi = hi +(cid:80)
a∈∂i Γa
(cid:81)
III. RESULTS
By using the mean field method, I first test both types of population couplings on the retina data, which is the spike
train of 160 ganglion cells in a small patch of the salamander retina [17]. The retina was stimulated with a repeated
natural movie. The spike train data is binned with the bin size equal to 20ms reflecting the temporal correlation time
scale, yielding about 282744 binary firing patterns for data modeling.
I then test the same concepts on the cortical data of behaving rats. Rats performed the odor-place matching
working memory during one task session, and spiking activities of 117 cells in both superficial layer 2/3 and deep layer
(layer 5) of medial prefrontal cortex were simultaneously recorded (for detailed experiments, see Ref. [22]). One task
session consists of about 40 trials, yielding a spike train of these cortical cells binned with the temporal resolution
τ = 20ms (a total of 140596 firing patterns).
A.
Inference performances on the retinal data
Fig. 1 reports the inference result on a network example of 40 neurons selected randomly from the original dataset.
The firing rate is predicted faithfully by the model using either MC or BP (Fig. 1 (A)). Inferring only PopC1, one
could predict about 74.10% of entire pairwise correlation (a precision criterion is set to 7.0 × 10−3 in this paper)
(Fig. 1 (B)). This means that 74.10% of the whole correlation set have the absolute value of the deviation between
5
FIG. 1: (Color online) Inference performances of PopC1 on the retinal data (N = 40, one typical example). (A) Firing rate
in the data is reproduced by the model. (B) Two-cell correlations are explained partially by PopC1 (74.10%). A Monte-Carlo
(MC) sampling of the model yields similar results to belief propagation (BP), which is much faster. (C) From the MC samples,
three-cell correlations can also be estimated. (D) Probability of K synchronous spiking under the model is compared with that
of the data.
a ) smaller than the precision criterion. Using the sampled
the predicted correlation and measured one (C pred
configurations of neural firing activity from the MC simulation, one could also predict three-cell correlations (Fig. 1
(C)), whereas, the prediction fraction can be improved by a significant amount after introducing PopC2, as I shall
show later.
In addition, fitting only 2N model parameters in PopC1 analysis could not predict the tail of spike
synchrony distribution (Fig. 1 (D)); this is expected as no higher order interaction terms are included in the model,
and rare events of large K spikes are also difficult to observe in a finite sampling during MC simulations.
a − C ms
The inference results of PopC2 are given in Fig. 2. Note that, by considering the correlation between the pairwise
firing activity and the global population activity, i.e., the second order population coupling, the three-cell correlation
could be predicted partially (64.44%), and this fraction is much larger than that of PopC1 (Fig. 1 (C)). This is
due to the specific structure of PopC2, which incorporates explicitly three-cell correlations into the construction of
couplings (Eq. (4)). Technically, the mean-field theory for PopC2 avoids the slow sampling and evaluates the high
order correlations in a fast way. Alternatively, one could fit the data using the conventional Ising model [27] with
the same number of model parameters as PopC2, whereas, the three-cell correlations are hard to predict using MC
samplings, and a similar phenomenon was also observed in a previous work for modeling pairwise correlations [33].
Therefore I speculate that PopC2 acts as a key circuit variable for third order correlations.
The interaction matrix of {wij} reveals how important each pair of neurons is for the entire population activity
(emergent functional state of the whole network). As shown in Fig. 2 (C), PopC2 matrix has no apparent structure
of organization, i.e., each neuron can be paired with both positive and negative couplings. Some pairs have large
negative PopC2, suggesting that these components are anti-correlated with the population activity. That is to say, the
activity of these neuron-pairs is not synchronized to the population activity characterized by the summed activity over
all neurons except these pairs. In the network, there also exist positive PopC2s, which shows that these neuron-pairs
are positively correlated with the population in neural activity. The interaction matrix shown here may be related to
two-cell correlation(model)0.70.750.80.850.90.951two-cell correlation (data)0.70.750.80.850.90.951BP 74.10%MCequality(B)P(K)10−510−410−310−210−1100K024681012141618DataMC(D) mean firing rate (model)−0.95−0.9−0.85−0.8mean firing rate (data)−1−0.95−0.9−0.85−0.8BPMCequality(A)three-cell correlation(model)−1−0.9−0.8−0.7−0.6three-cell correlation (data)−1−0.9−0.8−0.7−0.6MC 46.31% equality(C) 6
FIG. 2: (Color online) Inference performances of PopC2 on the retinal data (N = 40, one typical example). (A) Firing rate in
the data is reproduced by the model. (B) Three-cell correlations are explained partially by PopC2 (64.44%). (C) Interaction
matrix for PopC2. (D) Probability of K synchronous spiking under the model is compared with that of the data.
FIG. 3: (Color online) Inference performances of Ising model on the retinal data (N = 40, one typical example) compared
with population coupling. (a) Two-neuron interaction matrix. (b) Probability of K synchronous spiking under the model is
compared with that of the data.
the revealed overlapping modular structure of retinal neuron interactions [8, 14]. In this structure, neurons interact
locally with their adjacent neurons, and in particular this feature is scalable and applicable for larger networks. It
seems that one individual neuron does not impact directly the entire population, and a small group of neighboring
neurons have similar visual feature selectivity [34]. This result is also consistent with two-neuron interaction map of
the conventional Ising model (Fig. 3 (a)). Note that in functional interpretation, these two-neuron interactions are
inherently different from PopC2, which is designed to explain high-order correlations by using less model parameters
than necessary.
three-cell correlation (model)−1−0.95−0.9−0.85−0.8−0.75−0.7−0.65three-cell correlation (data)−1−0.95−0.9−0.85−0.8−0.75−0.7−0.652F1BP 64.44%equality(B) mean firing rate (model)−1−0.95−0.9−0.85 mean firing rate (data)−1−0.95−0.9−0.852F1F2(A) P(K)10−510−410−310−210−1100K02468101214DataMC(D)(C) neuron index510152025303540 −0.01−0.008−0.006−0.004−0.00200.002neuron index510152025303540510152025303540(a) neuron index0510152025303540 −0.100.10.20.30.4neuron index05101520253035400510152025303540 (b)P(K)10−510−410−310−210−1100 K024681012 dataPopC1PopC2 Ising 7
FIG. 4: (Color online) Multi-information and prediction fraction of correlations under the simplified model for the retinal
network. The result is averaged over 10 network samples for each N . (a) Multi-information (in bits) versus the network size
N . (b) The prediction fraction ρ versus the network size. PopC1 is used to predict two-cell correlations, and PopC2 is used to
predict three-cell correlations.
PopC2 behaves better than PopC1 in predicting the spike synchrony distribution (Fig. 2 (D)) in the small K regime
(the prediction is improved from K = 4 for PopC1 to K = 8 for PopC2). An intuitive explanation is that PopC2
introduces equivalently triplet interactions among neurons, and it is known that high order interactions are necessary
for generating widespread population activity [12]. However, PopC2 overestimates the distribution when rare events of
synchronous spiking are considered. This may be related to the difficulty of obtaining sufficient equilibrium samples of
the model, especially those samples with large population activity. The spike synchrony distribution is also compared
with that obtained under Ising model (Fig. 3 (b)). Different performances are related to the multi-information measure
of neural population explained below.
Sind−Smodel, in which Sind =(cid:80)
The amount of statistical structure in the neural data due to introducing interactions among neurons can be
(cid:80)
measured by the multi-information [5]. I first introduce an independent model where only the firing rates of individual
neurons are fitted and the corresponding entropy is defined as Sind. The multi-information is then defined as I(N ) =
x=±1 S((1+mix)/2), where S(u) = −u ln u, and Smodel is assumed to be an upper
bound to the true entropy. The true entropy for large populations is difficult to estimate since it requires including
all possible interactions among neurons. However, the model entropy with low order interaction parameters could be
an approximate information capacity for the neural population, which depends on how significant the higher order
correlations are in the population.
i
Fig. 4 (a) shows the multi-information as a function of the network size. PopC1 and PopC2 are compared with
the Ising model [33], which reconstructs faithfully the pairwise correlations. PopC2 improves significantly over PopC1
in capturing the information content of the network, but its multi-information is still below that of the Ising model,
which is much more evident for larger network size. This is expected, because only part of third order correlations
are captured by PopC2, while the Ising model describes accurately the entire pairwise correlation profile which may
be the main contributor to the collective behavior observed in the population. However, PopC2 provides us an easy
way to understand the higher order correlation, while in the Ising model, it is computationally difficult to estimate
the higher order correlations. The average prediction fraction of correlations by PopC1 and PopC2 is plotted in Fig. 4
(b). PopC1 predicts more than 75% of the pairwise correlations, while PopC2 predicts more than 54% of the triplet
correlations. The prediction fraction changes slightly with the network size.
B.
Inference performances on the cortical data
To show the inference performance of both types of population couplings on the cortical data, I randomly select
a typical network example of 40 neurons from the original dataset, and then apply the computation scheme to this
typical example. Results are shown in Fig. 5. Surprisingly, the simplified PopC1 is able to capture as high as 99.23%
of pairwise correlations, implying that when a rat performed working memory tasks, there exists a simplified model
to describe emergent functional states in the medial prefrontal cortical circuit. Moreover, MC sampling of the PopC1
model also predicts well the spike synchrony distribution (Fig. 5 (D)). This is very different from that observed in the
Multi-Information00.20.40.60.811.21.41.6 N30405060708090IsingPopC1PopC2(a) ρ0.50.550.60.650.70.750.8 N30405060708090PopC1PopC2(b) 8
FIG. 5: (Color online) Inference performances of PopC1 on the cortical data (N = 40, one typical example). (A) Firing rate
in the data is reproduced by the model. (B) Two-cell correlations are explained partially by PopC1 (99.23%). A Monte-Carlo
(MC) sampling of the model yields similar results to belief propagation (BP). (C) From the MC samples, three-cell correlations
can also be estimated. (D) Probability of K synchronous spiking under the model is compared with that of the data.
retinal data. In this sense, the MPC circuit is simple in its functional states when the subject is performing specified
tasks.
More interesting circuit features are revealed by PopC2, which is shown in Fig. 6. About 94.79% of three-cell
correlations are explained by PopC2 in the MPC circuit. The interaction matrix of PopC2 in Fig. 6 (C) shows a
clear non-local structure in the cortical circuit (stripe-like structure). That is, some neurons interact strongly with
nearly all the other neurons in the selected population, and these interactions have nearly identical strength of PopC2.
Such neurons having stripe-like structure in the PopC2 matrix may receive a large number of excitatory inputs from
pyramidal neurons [22], and thus play a key role in shaping the collective spiking behavior during the working memory
task. The non-local effects are consistent with findings reported in the original experimental paper (cross-correlogram
analysis) [22] and the two-neuron interaction map under Ising model (Fig. 7 (a)). Thus, to some extent, PopC2
may reflect intrinsic connectivity in the cortical circuit, although the relationship between functional connections
and anatomical connections has not yet been well established [35]. Lastly, PopC2 overestimates the tail of the spike
synchrony distribution (Fig. 6 (D)), which may be caused by the sampling difficulty of the inferred model (a model
with triplet interactions among its elements). The spike synchrony distribution of Ising model is also compared (Fig. 7
(b)).
Multi-information versus the cortical network size is plotted in Fig. 8 (a). In the cortical circuit, PopC2 behaves
comparably with the Ising model; even for some network size (N = 50), it reports a higher information content than
the Ising model in the randomly selected subpopulations, which may be caused by the nature of the selected neurons
(e.g., inhibitory interneurons [22], and they have stripe-like structure in the PopC2 matrix). Note that PopC1 gives
an information close to zero for small network sizes, suggesting that by introducing PopC1, one could not increase
significantly the amount of statistical structure in the network activity explained by the model. However, the multi-
information of PopC1 grows with the network size, indicating that the role of PopC1 would be significant for larger
neural populations. Fig. 8 (b) reports the prediction fraction of the correlation profile by applying PopC1 and PopC2.
Both population couplings can capture over 90% of correlations, which is significantly different from that observed in
the retinal data.
mean firing rate (model)−0.9−0.8−0.7−0.6−0.5−0.4mean firing rate (data)−1−0.9−0.8−0.7−0.6−0.5−0.423F1BPMCequality(A) two-cell correlation(model)0.20.40.60.81two-cell correlation (data)0.20.40.60.81BP 99.23%MCequality(B)three-cell correlation(model)−1−0.8−0.6−0.4−0.2three-cell correlation (data)−1−0.8−0.6−0.4−0.2MC 89.51% equality(C)P(K)10−510−410−310−210−1100K0123456789DataMC(D) 9
FIG. 6: (Color online) Inference performances of PopC2 on the cortical data (N = 40, one typical example). (A) Firing rate in
the data is reproduced by the model. (B) Three-cell correlations are explained partially by PopC2 (94.79%). (C) Interaction
matrix for PopC2. (D) Probability of K synchronous spiking under the model is compared with that of the data.
FIG. 7: (Color online) Inference performances of Ising model on the cortical data (N = 40, one typical example) compared
with population coupling. (a) Two-neuron interaction matrix. (b) Probability of K synchronous spiking under the model is
compared with that of the data.
IV. DISCUSSION
The emergent properties of the neural code arise from interactions among individual neurons. A complete charac-
terization of the population activity is difficult, because on the one hand, the number of potential interactions suffers
from a combinatorial explosion, on the other hand, the collective behavior at the network level would become much
more complex as the network size grows. In this paper, I develop a theoretical framework to understand how pairwise
or higher order correlations arise and the basic circuit variables corresponding to these correlation structures. The
three-cell correlation (model)−1−0.9−0.8−0.7−0.6−0.5−0.4−0.3−0.2three-cell correlation (data)−1−0.9−0.8−0.7−0.6−0.5−0.4−0.3−0.22F1BP 94.79%equality(B) P(K)10−510−410−310−210−1100K02468DataMC(D) mean firing rate (model)−1−0.9−0.8−0.7−0.6−0.5−0.4 mean firing rate (data)−1−0.9−0.8−0.7−0.6−0.5−0.42F1F2(A)(C) neuron index510152025303540 −0.0015−0.001−0.000500.0005neuron index510152025303540510152025303540(b)P(K)10−510−410−310−210−1100 K02468 dataPopC1PopC2 Ising(a)neuron index0510152025303540 00.020.040.060.08neuron index05101520253035400510152025303540 10
FIG. 8: (Color online) Multi-information and prediction fraction of correlations under the simplified model for the cortical
network. The result is averaged over 10 network samples for each N . (a) Multi-information (in bits) versus the network size
N . (b) The prediction fraction ρ versus the network size. PopC1 is used to predict two-cell correlations, and PopC2 is used to
predict three-cell correlations.
model is based on the concept of population coupling, characterizing the relationship between local firing activity of
individual neuron or neuron-pair and the global neural activity. An advantage is that, it provides a low dimensional
and neurobiologically interpretable representation to understand the functional interaction between neurons and their
correlation structures. In particular, the concept of population coupling and the associated mean field method used in
this paper offer an easy way to evaluate higher order correlations, while the usual sampling method is computationally
hard and traditional models (e.g., Ising model) lack a direct interpretation of higher order correlations in terms of
simplified (population) couplings.
With the mean field method, the concept of population coupling is tested on two different types of neural data.
One is the firing neural activities of retinal ganglion cells under natural movie stimuli. The other is the population
activities of medial prefrontal cortex when a rat was performing odor-place matching working memory tasks.
For the retinal data, on average PopC1 accounts for more than 75% of pairwise correlations, and PopC2 accounts
for over 54% of three-cell correlations. The interaction matrix of PopC2 contains information about the functional
interaction features in the retinal circuitry. It seems that a retinal neuron can be paired with not only negatively
strong couplings, but also slightly positive couplings. Only a few pairs of neurons have strong correlations with the
global activity of the population. To describe the spike synchrony distribution, PopC2 performs better than PopC1,
nevertheless, both of them could not capture the trend of the tail (rare events related to higher order interactions
existing in the network). This is not surprising, because PopC1 and PopC2 are simplified descriptions of the orig-
inal high dimensional neural activity, taking the trade-off between the computation complexity and the description
goodness.
To extract the statistical structure embedded in the neural population, PopC2 improves significantly over PopC1,
and has further additional benefit of describing the third-order correlations observed in the data, as PopC2 could be
used to construct triplet interactions among neurons, although direct constructing all possible triplet interactions is
extremely computationally difficult.
Unlike the retinal circuit, the cortical circuit yields a much smaller absolute value of the multi-information, implying
that no significant higher order correlations (interactions) were present in the neural circuit when the circuit was car-
rying out task-related information processing rather than encoding well-structured stimuli (as in the retinal network).
This also explains why a simplified description such as PopC1 and PopC2 is accurate enough to capture the main
features of the population activity, including the spike synchrony distribution. The inferred model on the cortical
data reveals a different interaction map from that of the retinal circuit.
In the cortical circuit, neurons form the
stripe-like structure in the interaction matrix, suggesting that these neurons may receive a large number of excitatory
inputs [22]. These inputs may come from different layers of cortex, and they can execute top-down or bottom-up
information processing, thus modulate the global brain state in the target cortex during behavioral tasks.
Before summary of this work, I made some discussions about two relevant recent studies on population coupling
(see notes added). Ref. [32] modeled directly the joint probability distribution of individual neural response and
population rate (the number of neurons simultaneously active) by linear coupling and complete coupling models. The
linear coupling reproduces separately the distribution of individual neuronal state and the population rate distribution,
and their couplings, while PopC1 introduced in my work reproduces mean firing rate and the correlation between
Multi-information−0.0200.020.040.060.080.1 N30405060708090 IsingPopC1PopC2(a) ρ0.880.90.920.940.960.98 N30405060708090PopC1PopC2(b) 11
individual neuronal state and the background population activity (except the neuron itself). Note that PopC1 does
not model population rate distribution explicitly, which is hard to interpret in terms of functional connectivity. The
complete-coupling model reproduces the joint probability distributions between the response of each neuron and
the population rate, from which it is hard to conclude that the high-order interactions responsible for high-order
correlations can be interpreted and tested. However, PopC2 reproduces mean firing rate and the correlation between
neuron-pair activities and the background population activity (except the neuron-pair itself), and thus explains high-
order correlations by an energy model. Furthermore, in this sense, this work overcame a weakness pointed out in
another independent later work of population coupling [36], which fitted directly the population rate distribution and
the firing probability for each neuron conditioned on the population rate, and analogously the corresponding model
parameters can not be readily interpretable in a biological setting. Due to intrinsic difference in model definitions,
these two relevant works have nice properties of studying tuning curves of individual neurons to the population rate,
and sampling from the model to reproduce the population synchrony distribution.
In summary, I develop a theoretical model of population coupling and its advanced form, to relate the correlation
profile in the neural collective activity to the basic circuit variables. The practical dimensional reduction method is
tested on different types of neural data, and specific features of neural circuit are revealed. This model aiming at
describing high order correlations with a low order representation, is expected to be useful for modeling big neural
data. Note that the interaction matrices shown in Fig. 2 and Fig. 6 are qualitatively robust to changes of the data size
to only the first half (data not shown), verifying that the revealed features are not an artifact of overfitting. However,
it still deserves further studies by introducing regularization in the learning equation.
It is also very interesting
to incorporate more physiologically plausible parameters to explain how the collective spiking behavior arises from
the microscopic interactions among the basic units. Another interesting study is to clarify the role of higher order
correlations in decoding performances based on maximum likelihood principles [37, 38].
Acknowledgments
I am grateful to Shigeyoshi Fujisawa and Michael J Berry for sharing me the cortical and retinal data, respectively.
I also thank Hideaki Shimazaki and Taro Toyoizumi for stimulating discussions. This work was supported by the
program for Brain Mapping by Integrated Neurotechnologies for Disease Studies (Brain/MINDS) from Japan Agency
for Medical Research and development, AMED.
note added. -- After I submitted this work to arXiv:1602.08299, I became aware of Ref. [32] (arXiv:1606.08889), and
later Ref. [36] (bioRxiv, 2016). Discussions about these two recent relevant works are made in the last section of this
paper.
Appendix A: Derivation of mean-field equations
In this appendix,
(cid:81)
I give a simple derivation of mean-field equations given in Sec. II. More details can
First, after removing an interaction a, one defines the cavity probability
be obtained from Refs [30, 31].
Pi→a(si) = ehisi
Pb→i(si), where the product comes from the physical meaning of the second kind of
Zi
j∈∂b sj(cid:81)
cavity probability, namely Pb→i(si) defined as the cavity probability when only the connection from b to i is
(cid:81)
Pb→i(si) is thus formulated as
retained while other neighbors of i are removed (so-called cavity probability).
it follows that the cavity magnetization
j∈∂b\i Pj→b(sj). With these two probabilities,
b∈∂i\a
(cid:80)
(cid:16)
mi→a =(cid:80)
mb→i by mb→i = tanh(ub→i) =(cid:80)
siPi→a(si) = tanh
sj :j∈∂b\i eΓb
si
hi +(cid:80)
si Pb→i(si).
si
b∈∂i\a ub→i
, where ub→i is named cavity bias in physics [29]. It is related to
More specifically, the cavity magnetization is derived as follows,
(cid:17)
ehi(cid:81)
ehi(cid:81)
,
mi→a = Pi→a(+1) − Pi→a(−1) =
−hi−2(cid:80)
−hi−2(cid:80)
hi +
(cid:88)
ehi − e
ehi + e
=
= tanh
b∈∂i\a ub→i
b∈∂i\a ub→i
ub→i
b∈∂i\a
Pb→i(+1) − e−hi(cid:81)
Pb→i(+1) + e−hi(cid:81)
b∈∂i\a
b∈∂i\a
b∈∂i\a
b∈∂i\a
Pb→i(−1)
Pb→i(−1)
(A1a)
where I have used the definition ub→i ≡ 1
interaction,
2 ln
Pb→i(+1)
Pb→i(−1)
. Next, I show how to derive the cavity bias. First, for pairwise
12
ub→i =
=
1
2
1
2
ln
ln
= tanh
eΓb Pj→b(+1) + e−Γb Pj→b(−1)
e−Γb Pj→b(+1) + eΓb Pj→b(−1)
1 + tanh Γbmj→b
1 − tanh Γbmj→b
−1 (tanh Γbmj→b) ,
(A2a)
where I have used the parameterization Pj→b(sj) = 1+sj mj→b
for triplet interaction,
2
, and the mathematical identity e2z = 1+tanh z
1−tanh z . Similarly,
ub→i =
=
1
2
1
2
ln
ln
= tanh
eΓb [Pj→b(+1)Pk→b(+1) + Pj→b(−1)Pk→b(−1)] + e−Γb [Pj→b(+1)Pk→b(−1) + Pj→b(−1)Pk→b(+1)]
e−Γb [Pj→b(+1)Pk→b(+1) + Pj→b(−1)Pk→b(−1)] + eΓb [Pj→b(+1)Pk→b(−1) + Pj→b(−1)Pk→b(+1)]
1 + e−2Γb 1−mj→bmk→b
e−2Γb + 1−mj→bmk→b
−1 (tanh Γbmj→bmk→b) .
1+mj→bmk→b
1+mj→bmk→b
These results are written in a compact form as Eq. (6) in the main text.
Similarly, the single neuron magnetization mi is obtained via mi = 1
Zi
si:i∈∂a
Pb→i(si), where Zi is a
normalization constant, and the multi-neuron correlation Ca = 1
i∈∂a Pi→a(si), where
Za
Za is a normalization constant. Note that Zi and Za are also related to the free energy contribution of single neuron
and neuronal interaction [30], respectively. The full (non-cavity) magnetization can be derived in a similar manner to
(cid:1). In detail, the two-point correlation Ca (∂a = 2) is computed as follows,
Eq. (A1), as mi = tanh(cid:0)hi +(cid:80)
si
i∈∂a sieΓa
b∈∂i ub→i
(cid:80)
(cid:81)
siehisi(cid:81)
i∈∂a si(cid:81)
(cid:81)
b∈∂i
(cid:80)
(cid:80)
(cid:80)
si,sj
Ca ≡ (cid:104)sisj(cid:105) =
sisjeΓasisj Pi→a(si)Pj→a(sj)
eΓasisj Pi→a(si)Pj→a(sj)
si,sj
1+mj→ami→a
e2Γa − 1−mj→ami→a
e2Γa + 1−mj→ami→a
tanh Γa + mi→amj→a
1 + tanh Γami→amj→a
1+mj→ami→a
.
=
=
(cid:80)
(cid:80)
si,sj ,sk
si,sj ,sk
1+mj→ami→amk→a
e2Γa − 1−mj→ami→amk→a
e2Γa + 1−mj→ami→amk→a
tanh Γa + mi→amj→amk→a
1 + tanh Γami→amj→amk→a
1+mj→ami→amk→a
.
=
=
(A3a)
(A4a)
(A5a)
Analogously, three-point correlation (∂a = 3) can be evaluated as
Ca ≡ (cid:104)sisjsk(cid:105) =
sisjskeΓasisj sk Pi→a(si)Pj→a(sj)Pk→a(sk)
eΓasisj sk Pi→a(si)Pj→a(sj)Pk→a(sk)
These results are written in a compact form as Eq. (8b) in the main text.
Finally, the partition function Zi for pairwise interaction is computed as follows,
Zi = ehi (cid:89)
= ehi (cid:89)
b∈∂i
Pb→i(+1) + e−hi (cid:89)
cosh Γb(1 + tanh Γbmj→b) + e−hi (cid:89)
Pb→i(−1)
b∈∂i
b∈∂i
b∈∂i
cosh Γb(1 − tanh Γbmj→b),
(A6a)
where I have used Pb→i(si) = eΓbsiPj→b(+1) + e−ΓbsiPj→b(−1) and the magnetization parameterization of Pj→b.
For triplet interaction, Pb→i(si) = eΓbsi[Pj→b(+1)Pk→b(+1) + Pj→b(−1)Pk→b(−1)] + e−Γbsi [Pj→b(+1)Pk→b(−1) +
Pj→b(−1)Pk→b(+1)], and the corresponding Zi = ehi(cid:81)
b∈∂i cosh Γb(1 + tanh Γbmj→bmk→b) + e−hi(cid:81)
13
b∈∂i cosh Γb(1 −
tanh Γbmj→bmk→b). Following the same line, the partition function Za is evaluated for pairwise interaction as
Za =
eΓasisj Pi→a(si)Pj→a(sj)
= cosh Γa[1 + tanh Γami→amj→a].
(cid:88)
si,sj
(cid:88)
(A7a)
(A8a)
Similarly, the partition function Za for triplet interaction can be computed as
Za =
eΓasisj sk Pi→a(si)Pj→a(sj)Pk→a(sk)
which is exactly the compact equation given in the main text.
si,sj ,sk
= cosh Γa[1 + tanh Γami→amj→amk→a],
[1] R. Q. Quiroga and S. Panzeri. Extracting information from neuronal populations:
information theory and decoding
approaches. Nat Rev Neurosci, 10:173, 2009.
[2] R. Yuste. From the neuron doctrine to neural networks. Nat Rev Neurosci, 16:487, 2015.
[3] I. H. Stevenson and K. P. Kording. How advances in neural recording affect data analysis. Nat Rev Neurosci, 14:139, 2011.
[4] M. R. Cohen and A. Kohn. Measuring and interpreting neuronal correlations. Nat Neurosci, 14:811, 2011.
[5] E. Schneidman, M. J. Berry, R. Segev, and W. Bialek. Weak pairwise correlations imply strongly correlated network states
in a neural population. Nature, 440:1007, 2006.
[6] S. Cocco, S. Leibler, and R. Monasson. Neuronal couplings between retinal ganglion cells inferred by efficient inverse
statistical physics methods. Proc. Natl. Acad. Sci. USA, 106:14058, 2009.
[7] J. H. Macke, M. Opper, and M. Bethge. Common input explains higher-order correlations and entropy in a simple model
of neural population activity. Phys. Rev. Lett, 106:208102, 2011.
[8] Elad Ganmor, Ronen Segev, and Elad Schneidman. The architecture of functional interaction networks in the retina. The
Journal of Neuroscience, 31:3044, 2011.
[9] L. Martignon, G. Deco, K. Laskey, M. Diamond, W. Freiwald, and E. Vaadia. Neural coding: higher-order temporal
patterns in the neurostatistics of cell assemblies. Neural Computation, 12:2621, 2000.
[10] M. J. Schnitzer and M. Meister. Multineuronal firing patterns in the signal from eye to brain. Neuron, 37:499, 2003.
[11] I. E. Ohiorhenuan, F. Mechler, K. P. Purpura, A. M. Schmid, Q. Hu, and J. D. Victor. Sparse coding and high-order
correlations in fine-scale cortical networks. Nature, 466:617, 2010.
[12] S.-I. Amari, H. Nakahara, S. Wu, and Y. Sakai. Synchronous firing and higher-order interactions in neuron pool. Neural
Computation, 15:127, 2003.
[13] F. Montani, E. Phoka, M. Portesi, and S. R. Schultz. Statistical modelling of higher-order correlations in pools of neural
activity. Physica A: Statistical Mechanics and its Applications, 392:3066, 2013.
[14] E. Ganmor, R. Segev, and E. Schneidman. Sparse low-order interaction network underlies a highly correlated and learnable
neural population code. Proc. Natl. Acad. Sci. USA, 108:9679, 2011.
[15] Shan Yu, Hongdian Yang, Hiroyuki Nakahara, Gustavo S. Santos, Danko Nikoli´c, and Dietmar Plenz. Higher-order
interactions characterized in cortical activity. The Journal of Neuroscience, 31:17514, 2011.
[16] G. Tkacik, O. Marre, T. Mora, D. Amodei, M. J. Berry II, and W. Bialek. The simplest maximum entropy model for
collective behavior in a neural network. Journal of Statistical Mechanics: Theory and Experiment, 2013:P03011, 2013.
[17] G. Tkacik, O. Marre, D. Amodei, E. Schneidman, W. Bialek, and M. J. Berry II. Searching for collective behavior in a
large network of sensory neurons. PLoS Comput Biol, 10:e1003408, 2014.
[18] U. Koster, J. Sohl-Dickstein, C. M. Gray, and B. A. Olshausen. Modeling higher-order correlations within cortical micro-
columns. PLoS Comput Biol, 10:e1003684, 2014.
[19] Haiping Huang. Effects of hidden nodes on network structure inference. Journal of Physics A: Mathematical and Theoretical,
48:355002, 2015.
[20] H. Shimazaki, K. Sadeghi, T. Ishikawa, Y. Ikegaya, and T. Toyoizumi. Simultaneous silence organizes structured higher-
order interactions in neural populations. Scientific Reports, 5:9821, 2015.
[21] M. Okun, N. A. Steinmetz, L. Cossell, M. F. Iacaruso, H. Ko, P. Bartho, T. Moore, S. B. Hofer, T. D. Mrsic-Flogel,
M. Carandini, and K. D. Harris. Diverse coupling of neurons to populations in sensory cortex. Nature, 521:511, 2015.
[22] S. Fujisawa, A. Amarasingham, M. T. Harrison, and G. Buzsaki. Behavior-dependent short-term assembly dynamics in
the medial prefrontal cortex. Nat Neurosci, 11:823, 2008.
[23] Rob R. de Ruyter van Steveninck, Geoffrey D. Lewen, Steven P. Strong, Roland Koberle, and William Bialek. Repro-
ducibility and variability in neural spike trains. Science, 275:1805, 1997.
[24] Michael Okun, Pierre Yger, Stephan L Marguet, Florian Gerard-Mercier, Andrea Benucci, Steffen Katzner, Laura Busse,
Matteo Carandini, and Kenneth D Harris. Population rate dynamics and multineuron firing patterns in sensory cortex.
The Journal of Neuroscience, 32:17108, 2012.
14
[25] Simona Cocco and R´emi Monasson. Adaptive cluster expansion for the inverse ising problem: convergence, algorithm and
tests. J. Stat. Phys, 147:252, 2012.
[26] H. Huang. Sparse hopfield network reconstruction with (cid:96)1 regularization. Eur. Phys. J. B, 86:484, 2013.
[27] Y. Roudi, E. Aurell, and J. Hertz. Statistical physics of pairwise probability models. Front. Comput. Neurosci, 3:1, 2009.
[28] Haiping Huang and Taro Toyoizumi. Clustering of neural code words revealed by a first-order phase transition. Phys. Rev.
E, 93:062416, 2016.
[29] M. M´ezard and G. Parisi. The bethe lattice spin glass revisited. Eur. Phys. J. B, 20:217, 2001.
[30] H. Huang and H. Zhou. Cavity approach to the sourlas code system. Phys. Rev. E, 80:056113, 2009.
[31] M. M´ezard and A. Montanari. Information, Physics, and Computation. Oxford University Press, Oxford, 2009.
[32] Christophe Gardella, Olivier Marre, and Thierry Mora. A Tractable Method for Describing Complex Couplings between
Neurons and Population Rate. eneuro, 3:e0160 -- 15.2016, 2016. arXiv:1606.08889.
[33] H. Huang and H. Zhou. Counting solutions from finite samplings. Phys. Rev. E, 85:026118, 2012.
[34] Ko Ho, Hofer Sonja B., Pichler Bruno, Buchanan Katherine A., Sjostrom P. Jesper, and Mrsic-Flogel Thomas D. Functional
specificity of local synaptic connections in neocortical networks. Nature, 473:87, 2011.
[35] Dimitri Yatsenko, Josic Kresimir, Ecker Alexander S., Froudarakis Emmanouil, Cotton R. James, and Tolias Andreas S.
Improved Estimation and Interpretation of Correlations in Neural Circuits. PLoS Comput Biol, 11:e1004083, 2015.
[36] Cian O'Donnell, J Tiago Goncalves, Nick Whiteley, Carlos Portera-Cailliau, and Terrence J Sejnowski. The population
tracking model: A simple, scalable statistical model for neural population data. Neural Computation, 29:50 -- 93, 2017.
[37] Michael T. Schaub and Simon R. Schultz. The ising decoder: reading out the activity of large neural ensembles. J Comput
Neurosci, 32:101, 2012.
[38] Greg Schwartz, Jakob Macke, Dario Amodei, Hanlin Tang, and Michael J. Berry. Low error discrimination using a
correlated population code. J Neurophysiol, 108:1069, 2012.
|
1707.05952 | 1 | 1707 | 2017-07-19T06:32:34 | Criticality in the brain: A synthesis of neurobiology, models and cognition | [
"q-bio.NC"
] | Cognitive function requires the coordination of neural activity across many scales, from neurons and circuits to large-scale networks. As such, it is unlikely that an explanatory framework focused upon any single scale will yield a comprehensive theory of brain activity and cognitive function. Modelling and analysis methods for neuroscience should aim to accommodate multiscale phenomena. Emerging research now suggests that multi-scale processes in the brain arise from so-called critical phenomena that occur very broadly in the natural world. Criticality arises in complex systems perched between order and disorder, and is marked by fluctuations that do not have any privileged spatial or temporal scale. We review the core nature of criticality, the evidence supporting its role in neural systems and its explanatory potential in brain health and disease. | q-bio.NC | q-bio | Criticality in the brain
Cocchi et al. 1
Criticality in the brain: A synthesis of neurobiology, models and cognition
Luca Cocchi1*, Leonardo L. Gollo1, Andrew Zalesky2, Michael Breakspear1
1 Queensland Institute for Medical Research, Brisbane, Australia.
2 Melbourne Neuropsychiatry Centre, The University of Melbourne, Melbourne, Australia.
Abstract
Cognitive function requires the coordination of neural activity across many scales, from neurons and
circuits to large-scale networks. As such, it is unlikely that an explanatory framework focused upon
any single scale will yield a comprehensive theory of brain activity and cognitive function. Modelling
and analysis methods for neuroscience should aim to accommodate multiscale phenomena. Emerging
research now suggests that multi-scale processes in the brain arise from so-called critical phenomena
that occur very broadly in the natural world. Criticality arises in complex systems perched between
order and disorder, and is marked by fluctuations that do not have any privileged spatial or temporal
scale. We review the core nature of criticality, the evidence supporting its role in neural systems and
its explanatory potential in brain health and disease.
Keywords: Bifurcations; metastability; multistability; dynamics; power-law; cognition.
Highlights
Criticality is a wide-spread phenomenon in natural systems
Criticality provides a unifying framework to model and understand brain activity and
cognitive function
Substantial evidence now supports the hypothesis that the brain operates near criticality
We review the role of criticality in healthy and pathological brain dynamics
Caveats and pitfalls regarding the assessment of criticality in the brain are discussed
* Email: [email protected]
1
Cocchi et al. 2
Criticality in the brain
List of abbreviations
ATP: Adenosine Triphosphate
BS: Burst suppression
DC: Direct current
DFA: Detrended fluctuation analysis
ECoG: Electrocorticography
EEG: Electroencephalogram
fMRI: Functional magnetic resonance imaging
GABA: Gamma-Aminobutyric acid
IBI: Inter-burst interval
MEG: Magnetoencephalography
REM: Rapid eye movement sleep
SOC: Self-organised criticality
SWS: Slow-wave sleep
Tc: Critical temperature
2
Criticality in the brain
Cocchi et al. 3
Table of Contents
1. Introduction ......................................................................................................................................... 4
2. Criticality in physical systems ............................................................................................................ 5
2.2. Criticality and phase transitions ...................................................................................... 8
2.3. The conceptual appeal of criticality .............................................................................. 11
2.4. Self-organised criticality ............................................................................................... 12
3. Criticality in the brain ....................................................................................................................... 13
3.1. Rhythmic fluctuations, bifurcations and slowing down ................................................ 13
3.2. Neuronal avalanches and phase transitions ................................................................... 14
3.3. Computational aspects of neuronal criticality ............................................................... 16
3.4. Self-organised neuronal criticality ................................................................................ 18
4. Challenges and pitfalls of the criticality hypothesis ......................................................................... 19
5. Emerging role of criticality in cognition ........................................................................................... 23
5.1. Criticality in brain and behaviour.................................................................................. 23
5.2. Suppression of criticality during task performance....................................................... 24
6. Criticality in disease .......................................................................................................................... 25
6.1. Bifurcations and seizures .............................................................................................. 25
6.2. Crackling noise and neonatal burst-suppression ........................................................... 26
6.3. Criticality and neuropsychiatric disorders..................................................................... 27
7. Summary ........................................................................................................................................... 28
8. References ......................................................................................................................................... 30
Box 1: Ten-Point Summary .................................................................................................................. 42
Box 2: Glossary..................................................................................................................................... 43
3
Criticality in the brain
Cocchi et al. 4
1. Introduction
Enormous strides have been achieved in neuroscience across a hierarchy of scales of enquiry, from
the variety of neural cell types and their molecular biology, through the function of cortical circuits
and, in recent years, to the complex architecture of large-scale brain networks. Much of this success
has been achieved within research silos, with a focus on scale-specific phenomena, partly mandated
by the apertures of various imaging technologies and partly by the training and cultures within the
various neuroscientific disciplines. Research in neuroscience also proceeds within a largely
descriptive world-view, with increasing emphasis on the collation and statistical characterization of
"big data" (Biswal et al., 2010; Markram et al., 2015). Whilst specific mechanisms have been
elucidated across an array of basic and clinical neuroscience domains, important challenges remain to
be addressed: First, since correlations between behaviour and neuronal activity have been documented
at almost every scale of analysis, it seems unlikely that a description of the brain at any particular
scale will be sufficient to describe brain function. How is neural activity integrated across scales to
give rise to cognitive function? What are the mechanisms linking activity across scales? Second, brain
function does not only rely upon the execution of particular functions, but also on adaptive switching
from one function to another, depending on context and goals. What are the fundamental principles
underlying such complex, flexible neuronal dynamics? Third, what are the major theoretical
frameworks to explain and unify the properties of all the large volumes of data currently being
accrued? Fourth, how is information encoded by neurons – in the entropy of individual spikes, or via
likelihood functions encoded by the distributions of population activity?
The principles that unify brain function across spatial and temporal scales remain largely unknown.
However, comparable multi-scale challenges exist in other scientific disciplines. Meteorology, for
example, spans scales from local wind gusts through regional weather systems up to global climate
patterns. Each scale is nested within a larger scale, such that the local variance in wind gusts depends
upon the regional weather, which is likewise constrained by global trends such as El Niño.
Mathematicians and physicists have developed a considerable armoury of analytic tools to address
multi-scale dynamics in a host of physical, biological and chemical systems (Bak et al., 1987). Chief
amongst these is the notion of criticality, an umbrella term that denotes the behaviour of a system
perched between order (such as slow, laminar fluid flow) and disorder [such as the turbulence of a
fast-flowing fluid, (Shih et al., 2015)]. A critical system shows scale-free fluctuations that stretch
from the smallest to the largest scale, and which may spontaneously jump between different
spatiotemporal patterns. Despite their apparent random nature, the fluctuations in these systems are
highly structured, obeying deep physical principles that show commonality from one system to the
other (so-called universality). They can hence be subject to robust statistical analysis and modelling.
Critical systems thus display the type of cross-scale effects and dynamic instabilities linking activity
at different scales that is typical of brain functioning. An emerging literature suggests that brain
function may be supported by critical neural dynamics, with original research that continues to
flourish (Deco and Jirsa, 2012; Kelso et al., 1992; Priesemann et al., 2014; Scott et al., 2014) on the
background of an existing body of reviews and syntheses (Beggs and Timme, 2012; Chialvo, 2010;
Deco and Jirsa, 2012; He, 2013; Hesse and Gross, 2015; Kelso et al., 1992; Plenz and Thiagarajan,
2007; Priesemann et al., 2014; Schuster et al., 2014; Scott et al., 2014; Shew and Plenz, 2013). The
principles supporting the emergence of these patterns of activity are not yet fully understood but
recent studies using neuroimaging techniques such as functional magnetic resonance imaging (fMRI)
and electroencephalogram (EEG) (Deco et al., 2009; Linkenkaer-Hansen et al., 2001; Stam and de
Bruin, 2004) have added to earlier work in slice preparations (Beggs and Plenz 2003). Computational
4
Criticality in the brain
Cocchi et al. 5
models also show that neural systems have maximum adaptability to accommodate incoming
processing demands when they are close to a critical point (Friston et al., 2012b; Friston, 2000; Gollo
and Breakspear, 2014; Kastner et al., 2015; Shew et al., 2009; Yang et al., 2012). Conversely, brain
disorders, as diverse as epilepsy, encephalopathy, bipolar disorder and schizophrenia may correspond
to excursions from such an optimal critical point.
Despite the ubiquity of criticality in many branches of science, its application to neuroscience is
relatively recent and unknown to many neuroscientists. When it is used, it is often invoked
metaphorically; a practice which risks mixing distinct processes incorrectly into a rubric term.
Research into criticality has much to offer neuroscientists, but needs to be used in accordance with its
well-defined operational criteria. Accumulating evidence should also be viewed cautiously according
to emerging pitfalls. Here, we first revisit the core notion of critical phenomenon and provide
examples from the physical sciences. We then review the classic and recent studies of neuronal
criticality. We finally consider emerging applications that advance new theories of healthy and
maladaptive cognition using the innovative tools that criticality provides.
2. Criticality in physical systems
Criticality refers to the appearance of erratic fluctuations in a dynamical system that is close to losing
dynamic stability. Because the nature of the instability can vary (as we review below), criticality is a
broad umbrella term that subsumes several related phenomena but also excludes others. In this
section, we present a brief pedagogical account of criticality. We first consider critical fluctuations
that occur close to instability in systems consisting of only a few interacting components. This allows
us to introduce core signatures of criticality; namely the emergence of scale-free temporal
fluctuations, slowing down and multistability, defined below. We then consider criticality in complex
systems composed of many interacting parts. These extra degrees of freedom allow for the occurrence
of scale-free spatiotemporal fluctuations known as avalanches. Finally, we consider self-organized
criticality - that is, the process by which criticality emerges without the need for external tuning of a
control parameter.
2.1. Criticality and bifurcations
We first consider dynamical systems composed of only a few interacting components. Consider the
classic example in which two species interact as predator and prey. When the interactions amongst the
species are weak and alternative food sources are available, relatively simple models predict that the
populations of both species reach stable equilibria, and the processes of consumption and
reproduction occur at a steady rate (Berryman, 1992). However, if the interactions between the
species increase (i.e. the predators rely more heavily upon the prey population), there reaches a
critical point of interactivity above which the two populations begin to oscillate: When the predator
population is relatively low, the number of prey animals grows through unbalanced reproduction. This
then yields a ready food source for the surviving predators, whose population increases. However, as
the increasing numbers of predators consume the available prey species, the numbers of the latter then
decline, with a subsequent effect on the survival of the predators; the cycle then begins anew. This
transition from steady state to cyclic behaviour due to strong interactions is called a bifurcation
(Figure 1a). The strength of the interaction is called a control parameter and the point at which the
bifurcation occurs is denoted the critical point. For three or more interacting species, further
bifurcations to more complex dynamics can occur, leading from periodic to chaotic oscillations
(smooth and deterministic but aperiodic oscillations) (Arneodo et al., 1980; Vano et al., 2006).
5
Criticality in the brain
Cocchi et al. 6
There are two crucial variations on this simple example. First physical processes, such as the predator-
prey example, inevitably occur in the presence of small but unceasing random fluctuations. This noise
arises from a myriad of causes such as the probabilistic nature of individual predator-prey encounters
as well as influences not explicitly modelled (diseases, fluctuating environmental conditions etc).
When the interactions between predator and prey are weak, the equilibrium state is very stable and the
presence of such random fluctuations only have a minor impact on the observed steady state
populations. Likewise, if the interactions are strong, the cyclic oscillations in population numbers are
also very stable: The amplitude of the oscillations is relatively stable and the noise is again effectively
suppressed. More technically, away from the critical point, the system is said to be dynamically stable
and the fluctuations are strongly damped, dropping off quickly (with an exponential decay rate).
However, in the immediate vicinity of the critical point, the perturbations grow in magnitude,
dominating the observations because the system is less stable (or weakly stable). That is, the variance
of the observed fluctuations grows in magnitude. Moreover, the fluctuations decay slowly. To be
precise, the decay of fluctuations in time changes from a fast (exponential, Figure 1b, inset) to a slow
(power-law, Figure 1c, inset) process. Fluctuations with power-law correlations are scale-free
because they do not have a characteristic time scale. Upon further increases in the control parameter,
the fluctuations in the envelope of the oscillations quickly become stable (Figure 1d) and drop off
quickly (exponentially, likewise the inset of Figure 1b).
Hence near the critical point of a bifurcation, we encounter two central features of criticality: High
amplitude scale-free fluctuations (Figure 1h) and slowing down (i.e., longer autocorrelation, Figure
1i). These large, slow fluctuations are termed crackling noise after the sound they make if played
audibly (Sethna et al., 2001).
A second variation concerns the nature of the bifurcation itself. The predator-prey model exhibits a
classic bifurcation, whereby a single critical point separates two distinct behaviours (steady state and
oscillatory) in parameter space. This is denoted a supercritical bifurcation because the cyclic
oscillations occur for values of the control parameter strictly greater than the critical point. However,
subcritical bifurcations are also possible. In this setting, there exists a region where the steady state
solutions and the periodic oscillations co-exist. Outside of this zone, the system behaves in the same
way as the supercritical bifurcation (i.e., a single steady state or a periodic pattern of activity) (Figure
1e). However, within this zone, the two dynamic states co-exist (Figure 1e and f). Noise can then
push the system between these stable states, causing erratic jumps between low amplitude equilibrium
and high amplitude oscillations (Figure 1f). This type of behaviour is called multistability (Freyer et
al., 2011; Tognoli and Kelso, 2014). We consider examples of multistability in brain and behaviour
below.
Although critical fluctuations (near a supercritical bifurcation) and multistable fluctuations (due to a
subcritical bifurcation) are mathematically related, they yield quite distinct statistics: The former
(slow fluctuations) follow a scale-free power law distribution. This is the classic meaning of the term
criticality. Noise-driven switches between two or more multistable attractors do not occur with a
scale-free probability. In the setting of large additive noise, the transitions are akin to a Poisson
process and thus follow an exponential distribution. With smaller, state-dependent noise, the system
tends to get trapped near each state, with the transitions then following a heavier tailed stretched
exponential distribution (Freyer et al., 2011; Freyer et al., 2012). Either way, such multistable
switching does not possess scale-free (power law) properties and does not correspond to the classic
notion of criticality.
6
Criticality in the brain
Cocchi et al. 7
Figure 1: Criticality in a low dimensional system consisting of a few interacting components. (a) Super-critical
bifurcation diagram, depicting the amplitude of a system's state variable (y-axis) as a function of a control
parameter (such as the strength of interactions, x-axis). When the control parameter is increased, the activity
of the system switches from a damped equilibrium point (red circle) to oscillatory behaviour (yellow circle).
The point of change is known as the critical point (blue circle). (b) In the presence of noise, the damped system
(red circle) exhibits low amplitude, rapid fluctuations. The duration of these follows an exponential probability
distribution (red dots, inset). (c) At the critical point, the fluctuations have high variance and rise and fall
slowly, following a power law distribution, corresponding to a linear relationship between their duration and
their likelihood in double logarithmic coordinates (inset). In addition, the slope of this relation is described by a
critical exponent of α = −3/2. (d) Beyond the critical point, the system exhibits sustained oscillations. The
fluctuations in the amplitude envelope of the oscillations are fast and small. (e) Sub-critical bifurcation
7
Criticality in the brain
Cocchi et al. 8
diagram, with a zone of co-existence (or "bistability") between the fixed point and oscillatory behaviours. In
this case, system noise not only drives fluctuations around each attractor, but can also drive sudden and
erratic jumps between the two dynamic states as depicted by the double headed arrow. (f) Example bistable
time series. (g) Multistable bifurcation diagram as reproduced from Freeman (Freeman, 1987), proposed as a
model for perceptual activity in the olfactory system. In this case, the number and complexity of the attractors
is larger, however, the underlying principle is the same. Critical slowing down corresponds to a sharp increase
in the coefficient of variation of the mean amplitude across 200 trials (h) and the auto-correlation function (i)
at the critical point.
2.2. Criticality and phase transitions
We have thus far considered relatively simple systems composed of only a few components, or where
the elements of the system are lumped into a small number of variables (such as all predators and all
prey species each being considered a single entity). We now move to studying critical systems
composed of many interacting components such as magnetic spins in iron (Stanley, 1987), grains of
sand falling onto a pile (Bak et al., 1988), or neurons (Plenz and Thiagarajan, 2007). On top of the
example considered in Figure 1, these examples introduce a spatial dimension through which the
components of the system interact.
The emergence of a magnetic field in a ferromagnetic material (such as iron) cooled below a critical
temperature (Tc) is a classic example of a phase transition. Such materials have permanent magnetic
moments (dipoles) due to the spin of unpaired electrons in atomic or molecular electron orbits. These
dipoles interact through the mutual effects of the local fields that they impart on their immediate
neighbours (Figure 2a), causing neighbouring dipoles to align and form local domains of coherent
fields. These effects partially counter the influence of stochastic thermal and quantum effects that
cause random flips in the direction of the dipoles. At temperatures greater than Tc, stochastic flips
disrupt the formation of larger domains and, in the absence of an external field, the material will not
possess a macroscopic magnetic field. Slow cooling of the material allows domains of increasing size
to form, although domains at the very largest scales continue to disappear into the background noise.
However, when the material is cooled to a critical point (the Curie temperature), the smaller domains
coalesce into increasingly larger ones until they approach the size of the entire system. The
coalescence of small domains into those of successively larger size is called an avalanche (Figure
2a). At the Curie temperature, avalanches have no characteristic size and thus may intermittently
sweep through the entire system. These avalanches can be measured empirically using a large,
external pick-up device (Cote and Meisel, 1991; McClure Jr and Schroder, 1976; Meisel and Cote,
1992; Perković et al., 1995). Below the critical temperature, the mutual interactions amongst the spins
align into domains that encompass nearly every dipole; at this point, the material shows a coherent
ferromagnetic field (despite ongoing disorder at the atomic scales). The transition through the critical
point in such a high dimensional system is called a phase transition.
A phase transition in iron cooled below its Curie temperature is a classic example of how simple
internal interactions can overcome disorder and yield, through a critical point, a macroscopic field.
The imposition of an external field of sufficient strength on a ferromagnet below the critical point can
cause the field to suddenly switch directions to align with the applied field (Vojta et al., 2013): Here,
in contrast, the macroscopic order is imposed externally.
There are many similarities between bifurcations and phase transition, including the presence of
super- and subcritical varieties (Figure 2b): These are called continuous and discontinuous or
(second- and first-order) phase transitions in this context (Kim et al., 1997). The transition from a
para- to a ferromagnet due to cooling, outlined above, is an example of a continuous phase transition:
8
Criticality in the brain
Cocchi et al. 9
The sudden switching of that field due to an external field is a discontinuous one, as is water turning
into vapour in the presence of heat (Stanley 1987). When the phase transition is discontinuous, then
noise-driven multistability may also occur: Noise can induce switches between ordered and random
states (Figure 2b). Alternatively, as in the case of the (discontinuous) phase transition between water
and steam, there can arise complex mixtures of both. Just as in the case of a bifurcation, however,
multistability arising due to a discontinuous phase transition does not exhibit the scale-free, 'critical
phenomena' discussed above. Critical, power law scaling in the spatial and temporal domains is
unique to a continuous phase transition.
Whilst we here focused upon the canonical example of spins in a weak external field, the basic
ingredients (local interactions, noise, a large number of subsystems, an external influence that brings
weak coherence) occur widely and, as a result, phase transitions are ubiquitously observed in nature
(Stanley, 1987).
The temporal behaviour of a spatially-extended system near a phase transition mirrors the behaviour
of two interacting elements near a bifurcation – namely slow, high amplitude fluctuations. These
fluctuations, however, additionally exhibit complex spatiotemporal processes – avalanches – that also
show scale-free statistical properties. Away from the critical point, the likelihood of an avalanche
drops off quickly (exponentially) with size (the number of elements involved). In the classical
example of magnetism, the domains of coherent spins are very small at high temperatures. If the
temperature is tuned towards the critical point (Tc), the coherent domains sporadically increase in size
and the ensuing distribution of avalanche sizes decays slowly as a function of spatial scale. The size
of such domains measured over time converges toward a scale-free (power law) effect (Figure 2c). If
the temperature is further reduced, large and stable coherent domains appear corresponding to the
emergence of an internal magnetic field.
The correlation length is a useful concept in this setting. In the disordered phase of the system (i.e. for
weak external fields) only adjacent spins are correlated– distant spins are completely uncorrelated. As
larger avalanches begin to appear at low temperature, electron spins become correlated across the
scale of the corresponding coherent domains. The correlation length – the spatial scale at which pairs
of electrons are at least weakly correlated –increases. At the phase transition, as the size of the
coherent domains approaches that of the system, the correlation length diverges upwards. As a result,
external perturbations applied to any part of the system may lead to a change in the state of the whole
system. Put alternatively the system has maximum dynamic range because any small perturbation has
a chance of changing the electron spins of such a critical paramagnetic system.
We have focused upon phase transitions in spatially extended (embedded) systems, such as magnets
and water which are dominated by interactions or collisions between the neighbouring elements of the
system. However, the description of avalanches in large N systems does not inevitably refer to space.
A branching process is a simple model of a phase transition that describes how activated elements
may either decay (to inactive) or activate other elements with the progression of time (De Carvalho
and Prado, 2000). Criticality occurs when the decay and activation rates are in equal ratio. This
canonical model of a phase transition, which has been used to study criticality in neural systems [e.g.
(Beggs and Plenz, 2003)], does not require a spatial dimension; metrics such as the correlation length
do not make sense in these models. Although such abstract models may not take space into account,
complex multi-unit physical systems such as the brain must be embedded in space, and the
interactions between the elements are very often constrained by their spatial proximity (Roberts et al.,
2016).
9
Criticality in the brain
Cocchi et al. 10
Figure 2: Phase transitions and avalanches in spatially extended systems. (a) Continuous (or second order)
phase transition. The disordered (random) phase (red) shows lack of spatial order (red square) with randomly
oriented spins in a typical physical system such as a ferromagnet in a weak external field. The ordered phase
(yellow) shows large domains of co-aligned spins. At the critical point (blue), avalanches of complex partially-
ordered domains rise and dissolve across all scales (blue square), leading to a power law size distribution. (b)
Phase transitions can also be discontinuous (also called first order). As with a sub-critical bifurcation, system
noise then causes erratic switching between the disordered and ordered phases (double headed arrow). (c)
Cumulative probability distribution of the relationship between the size and likelihood of avalanches at
criticality. A scale-free processes yields a linear scaling relationship in double logarithmic coordinates (a power
law) with a critical exponent of α = −3/2. Note the slight exponential truncation at the right hand side, due to
finite size effects.
As we have seen, phase transitions can be considered a natural extension of the notion of a bifurcation
from systems with few components, to those with many. The underlying mathematics is very similar
10
Criticality in the brain
Cocchi et al. 11
(in fact, the so-called Landau equation, used to model generic phase transitions, is mathematically
identical to the Normal form equation used to describe bifurcations). Historically, however, the two
phenomena have been studied in different fields - bifurcations by applied mathematicians, whereas
phase transitions were classically the domain of physicists. This legacy has led to a difference in the
use of the central terms, sub- and super-criticality: In mathematics, "sub-" and "super-critical" qualify
bifurcations, denoting distinct instabilities that differ in their underlying mathematical nature (Figure
1a and 1e) and, as a result, the behaviour they yield. In physics, these terms are used to describe the
phases of the system. "Subcritical" phase is used to denote the stable, absorbing state below a phase
transition (Figure 1a, red dot and Figure 1b). The term "supercritical" is used to denote the ordered
state above the transition, whereby the amplitude of the order parameter (y-axis) is typically non-zero
(Figure 1a, yellow dot and Figure 1d). It is unlikely that an attempted synthesis of those terms here
would pervade both fields. For the remainder of this paper, we use the terms "sub-" versus "super-
critical bifurcation" to denote the type of instability, and "sub-" versus "super-critical state",
"activity" or "phase" to denote where a particular system lies with respects to the phase transition. In
general, since the interpretation is largely dependent on the audience, caution is required when
interpreting or using those terms more broadly.
Phase transitions and criticality have been documented in a very broad range of physical systems over
many decades (Kosterlitz and Thouless, 1973; Yang and Lee, 1952) and their study remains one of
the most active areas of research in branches of physics such as statistical mechanics (Papanikolaou et
al., 2011; Sethna et al., 2001; Zapperi et al., 2005). The signatures of criticality have been
documented in systems as diverse as flocking birds (Cavagna et al., 2010), earthquakes (Burridge and
Knopoff, 1967; Carlson and Langer, 1989; Rice and Ruina, 1983); solar flares (Lu et al., 1993),
armed conflict (Roberts and Turcotte, 1998), traffic jams (Nagel and Herrmann, 1993), and capital
wealth (Roberts and Turcotte, 1998) – even crumpled paper [(Houle and Sethna, 1996; Kramer and
Lobkovsky, 1996); for review see (Roberts and Turcotte, 1998)].
We turn to evidence for criticality in the brain in Section 3 after considering its underlying appeal. But
before moving, it is instructive to contextualize the importance of criticality as a theoretical
framework and how it may come about in many distinct systems.
2.3. The conceptual appeal of criticality
Criticality derives its basic appeal from a number of considerations. First, it speaks to the presence of
a relatively simple underlying process - the response of a weakly stable system to stochastic
perturbations - arising in very different settings. The processes that drive the system close to
instability can be diverse – a build-up of fuel; varying temperature; a driving external magnetic field;
strong interactions between species – but the collective response in generating slow, multiscale
fluctuations is shared. Likewise, the many specific details of the systems differ markedly (e.g.,
magnetic spins, moving tectonic plates, neurons) but can be unified by their core dynamic nature -
possessing interactions among their components that erratically amplify and damp microscopic
perturbations. This notion of "universality" is very appealing to scientists who seek unifying
principles across diverse systems. That is, the appearance of power-law and invariant scaling in
markedly different systems suggests the importance of processes that transcend their particular
incarnation. Of note, the characteristic exponent in the power-law scaling that describe the critical
fluctuations in many of these systems typically converges to a value of −3/2. Theoretical
considerations support the emergence of this value in systems at the cusp of a phase transition
(Zapperi et al., 1995). Thus, basic theoretical arguments unify diverse phenomena – this is the essence
of universality (Stanley, 1999).
11
Criticality in the brain
Cocchi et al. 12
Computational considerations also underlie the appeal of criticality. Here we review the
computational aspects of criticality in physical systems. In Section 3.3, we focus on the computational
advantages of criticality in the brain. The earliest demonstrations of the computational advantages of
the critical state were done in a very simple and idealized system called cellular automata (CA),
whose dynamics evolve discretely in space and time according to very simple interaction rules
(Langton, 1990). By changing an underlying interaction parameter, CA can be tuned to converge very
quickly to simple periodic (spatiotemporal) structures, or to unstructured, chaotic processes. In
between these scenarios – at the so-called "edge of chaos" – CA exhibit lengthy mixtures of ordered
and disorganized structures. Theoretical arguments show that the computational complexity of CA
diverge in this regime. That is, if one considers the information content of the system at each time
point, the number of iterations before CA converge onto a stable solution becomes very long for this
in-between state (although see (Mitchell et al., 1993) for an opposing position). Depending upon
one's viewpoint, the simplicity of CA is either conceptually appealing (since complexity arises from
very simple laws) or distracting, because the physical meaning of CA is unclear. However, the
implications of the proposal – complexity from simplicity - are tantalising, underlying the influence of
CA. Its catch-phrase "edge of chaos" became a very well-known way to refer to the computational
advantages offered by the critical state. Further research has shown that several other physical and
biological systems also have optimal computational properties at criticality (Crutchfield and Young,
1988; Kauffman and Johnsen, 1991; Mora and Bialek, 2011; Nykter et al., 2008).
The appeal of criticality also finds support from thermodynamic perspectives. In a stable, subcritical
system, random fluctuations arising from thermal energy and other sources of entropy are confined to
the microscopic scale, like a giant TV screen showing pixel-wise static. While these microscopic
fluctuations have high entropy, meso- and macroscopic scales are damped and are hence in a
featureless, low entropy state. Above the critical point, the macroscopic scale of the system can show
interesting features, such as periodic structures and oscillations. However, fluctuations at finer scales
are slaved to these large-scale features and thus do not express the potential entropy arising from the
smallest microscopic scales. In these two states – sub- and super-critical, respectively, high entropy
can be thought of as being trapped at one particular scale and unavailable at other scales. At the
critical state, microscopic fluctuations erratically disseminate to larger scales through avalanches and
crackles. These fluctuations introduce packets of disorder which accordingly increase the information
content of the system across all scales. That is, while criticality increases correlations – and thus
decreases entropy – at the smallest scale, it "transports" random fluctuations across scales increasing
the total complexity of the system (Tononi et al., 1994).
2.4. Self-organised criticality
Why is it that so many systems found in nature appear to be perched at a critical point? In theory the
critical point becomes confined to a very small range of values as the size of the system increases (see
Figure 1h). For an experimental system such as the paramagnetic material discussed above, the
external field can be carefully (manually) tuned until crackling noise and avalanches appear.
However, for other systems, such as earthquakes, forest fires and flocking birds the fingerprints of
criticality seem to arise internally without the need for careful tuning by an external observer.
The answer to this apparent dilemma is contained in the notion of self-organised criticality (SOC), a
process whereby a complex system is driven toward its critical point across a very wide set of starting
points and parameter values. The classic example of SOC was provided in the behaviour of sand-pile
avalanches by the work of Per Bak (Bak, 1990; Bak et al., 1987, 1988). In essence, the slow addition
12
Criticality in the brain
Cocchi et al. 13
of sand to the apex of a sand-pile leads to the gradual increase in the slope of its sides. At a critical
slope, scale-free avalanches of falling sand begin to occur. The slope angle decreases with each
avalanche as (gravitational) energy is released from the system, then increases again as new sand is
added. The slope does not need to be tuned by the experimentalist but naturally emerges from the
interplay of the interactions between the adjacent grains of sand, the external (gravitational) force and
the slow addition of sand.
Modelling SOC in complex systems with weak local interactions and noise is a very active field
(Marković and Gros, 2014). Two mechanisms appear sufficient for the appearance of SOC – firstly
the dissipation of energy and secondly some form of "memory" in the system. For example, a large
forest fire burns through a build-up of timber fuel: Energy has dissipated from the system and a period
of time must now pass until there is sufficient new fuel for fire of any appreciable magnitude.
Likewise, a large avalanche of sand in a slowly building sand pile changes the gradient of the pile.
Time and small avalanches must then accrue before the slope of the pile is sufficiently steep to trigger
another large avalanche. Again in seismology, tension from tectonic plates is released following a
large earthquake and its aftershocks, such that subsequent large earthquakes are unlikely to follow
immediately. Each of these systems is characterized by the build-up and subsequent dissipation of
energy or resources, whose release is 'remembered' by the system until sufficient resources have
recovered. We will encounter similar concepts when we discuss mechanisms of SOC in neuronal
systems (Section 3.3).
3. Criticality in the brain
The role of criticality and multistability in neurophysiological systems of the brain was first
articulated over 3 decades ago by Walter Freeman following detailed empirical analyses and
computational models of the rabbit olfactory bulb (Figure 1g). In particular, Freeman proposed that
the process of inhalation and exhalation acted, via modulation of the gain of excitatory neurons, to
sweep the activity of the olfactory bulb through a sub-critical bifurcation and hence through a zone of
multistability (Freeman, 1987; Freeman, 1991). Sensory inputs, arising from contact of inhaled
molecules with membrane receptors of olfactory neurons, then act to selectively perturb the system
onto one of several competing dynamic patterns. This dynamic pattern was proposed to encode the
olfactory input – the percept - until it destabilized during exhalation as the system passed again into
the zone where only the stable equilibrium solution exists. We now survey more recent examples of
criticality and multistability in neuronal systems that build upon the foresights contained in Freeman's
prescient papers.
3.1. Rhythmic fluctuations, bifurcations and slowing down
The mechanisms underlying motor coordination have been an intriguing area for the application of
dynamic systems theory (Bressler and Kelso, 2001; Kelso and Clark, 1982). One fruitful candidate
has been the study of rhythmic finger tapping. At slow frequencies, humans are able to tap in either of
two stable modes: syncopation and anti-syncopation. However, at high frequencies, the anti-
syncopation mode becomes unstable and only the in-phase syncopation pattern is expressed (Kelso et
al., 1986). The hallmarks of criticality are seen just prior to this transition, namely slowing, high
amplitude movement fluctuations (Kelso, 1984, 2014). To explain this, Haken, Kelso and Bunz used a
sub-critical bifurcation in a simple model of motor coordination between the left and right motor
cortices (Haken et al., 1985). The transition from a bimodal to unimodal pattern of behaviour
occurred at the critical value of the movement frequency. This framework has also been employed to
13
Criticality in the brain
Cocchi et al. 14
explain the transition between movement patterns induced by transcranial magnetic stimulation
(TMS) (Kelso, 2014).
More recent work on criticality has focused on the temporal fluctuations observed in the major
rhythms of EEG and MEG data. Employing an analysis called detrended fluctuation analysis (DFA),
Linkenkaer-Hauser and colleagues reported that the fluctuating amplitudes of the two dominant
oscillations of the human brain – the alpha and beta rhythms – exhibited scale-free temporal statistics
(Linkenkaer-Hansen et al., 2001). Fluctuating levels of synchrony between pairs of electrodes have
also been reported to show scale-free statistics (Stam and de Bruin, 2004). Likewise, the power
spectrum of human neocortical activity acquired from invasive ECoG data shows scale-free temporal
behaviour across a very broad range of frequencies (Miller et al., 2009). These frequencies also show
multi-scale nesting – that is, the amplitude of high frequencies is coupled to the phase of lower
frequencies; a pattern that is recursively repeated from very slow to very high frequencies (He et al.,
2010). Computational models of large-scale neuronal activity – neural field models – suggest that the
critical temporal statistics in these electrocortical recordings may arise from a subcritical bifurcation
of activity in corticothalamic loops (Freyer et al., 2009; Freyer et al., 2011). Such modelling proposes
that noise-driven switching between a low amplitude steady state and high amplitude oscillations
(Figure 1) yields the empirically observed critical fluctuations seen at rest (Freyer et al., 2012).
3.2. Neuronal avalanches and phase transitions
In 2003, Beggs and Plenz found evidence of critical behaviour in the erratic spontaneous activity
measured in in vitro neuronal cultures (Beggs and Plenz, 2003). They documented the two salient
features of criticality in a spatially extended critical system, namely power-law scaling in time (the
duration of bursts of activity) and space (the number of electrodes spanned by each burst). Together
with the earlier work of Freeman, this finding ushered in criticality as a term of clear relevance to
complex neuronal multi-scale phenomenon.
Since these initial reports of avalanches in in vitro slice preparations by Beggs and Plenz, research
into critical avalanche-like activity in spatiotemporal neural recordings has proceeded at great pace
(Plenz and Thiagarajan, 2007; Schuster et al., 2014; Shew, 2015). Observations of scale-free
spatiotemporal fluctuations in spontaneous, physiological data have progressed from in vitro slice
preparations (Beggs and Plenz, 2003), to in vivo recordings from superficial layers of cortex (Gautam
et al., 2015; Gireesh and Plenz, 2008) to awake non-human primates (Petermann et al., 2009) [for
review, see Shew and Plenz (2013)]. Scale-free avalanches have been reported in human whole brain
magnetoencephalographic (MEG) data (Shriki et al., 2013), and complex, scale-free spatial
dependences, consistent with avalanches, have been described in whole brain functional neuroimaging
(fMRI) data (Tagliazucchi et al., 2012). Notably, these recordings cross broad scales of aperture from
multi-unit recordings to macroscopic field potentials and whole brain functional neuroimaging data.
There is a growing focus in the imaging community on spontaneous (resting-state) fMRI data and the
reproducible structures these reveal in health (Damoiseaux et al., 2006; Zalesky et al., 2014) and
disease (Fornito et al., 2015). Whereas the number of channels in neurophysiological recordings, such
as MEG, has a modest upper bound (of several hundred), the high spatial resolution of fMRI yields
thousands of voxels (~100,000 voxels). These data hence contain the breadth of spatial scales required
to interrogate whether the spatial fluctuations are scale-free (Eguiluz et al., 2005) and thus whether
critical dynamics underlie the dynamic patterns seen at rest (Chialvo, 2012). Recent evidence from
both empirical (Tagliazucchi et al., 2012) and computational (Deco and Jirsa, 2012) analyses points in
14
Criticality in the brain
Cocchi et al. 15
favour of this proposal. Among the most intriguing findings are the recapitulation of the classic
resting-state networks (Yeo et al., 2011) by models of critical dynamics arising from primate (Honey
et al., 2007) and human (Haimovici et al., 2013) structural connectomes. Analysis of the temporal
statistics of resting-state fMRI and EEG also suggests that long-range, scale-free correlations may
indeed lie at the very heart of the slow fluctuations that are observed in these data (Van de Ville et al.,
2010).
This emerging view is schematically summarized in Figure 3. In the sub-critical zone, bursts of
cortical activity are sporadic uncoordinated (red box). Above the critical value, cortical activity is
coupled too tightly and conversely, inadequately segregated (yellow box). Resting-state networks
function at the critical value, where switching between network states occurs due to weak dynamic
instabilities (light blue box).
Figure 3: Proposed role of criticality in large-scale, resting-state brain dynamics. In the sub-critical region,
individual brain regions are effectively uncoupled, showing a lack of integration (red square). Conversely, in
the super-critical region, integration is too great and there is a lack of segregation (yellow). Near the critical
point, an emerging body of work in EEG, MEG and fMRI suggests that brain systems show a dynamic balance of
integration and segregation (blue square), fluctuating among the various resting-state networks (and EEG
rhythms). Cognitive function requires a slight incursion away from the critical regime leading to a stabilization
of one particular network, consistent with the earlier proposals of Freeman.
There is now a very well established relationship between resting state brain networks and the
underlying structural connectome from which they arise (Honey et al., 2007). Criticality arising in
simple systems does not require a complex spatial substrate: Rather as we have reviewed above, its
hallmark is the emergence of complex spatiotemporal processes from simple, local interactions.
However, complex networks may allow critical-like behaviour to occur in a region of parameter space
instead of a single point (Moretti and Muñoz, 2013). Moreover, in the primate brain, structural-
functional correlations argue for the existence of a relationship between critical dynamics and the
relatively static underlying structural connectome. The nature of this relationship between critical
states in resting-state fMRI data and the connectome is not well understood. Highly interconnected
15
Criticality in the brain
Cocchi et al. 16
cortical hubs, and brain regions comprising the so-called default-mode network may play a prominent
role in maintaining resting-state network dynamics (Gollo et al., 2015; Leech et al., 2012; Vasa et al.,
2015) and in facilitating the efficient spread of avalanche events through macroscopic brain networks
(Misic et al., 2015). The constellation of densely connected hub regions – the rich club – appear to
support a slow, stable dynamic "core" whereas peripheral sensory regions introduce (rapid) stimuli-
related variability in the system (Bassett et al., 2013; Gollo et al., 2015; Hasson et al., 2015). Such a
core-periphery organization of brain network dynamics speaks to a hierarchy of time-scale
fluctuations, in which hub regions integrate and regulate the network dynamics largely operating at
slow frequencies (Cocchi et al., 2016; Gollo et al., 2017; Gollo et al., 2015; Hasson et al., 2008;
Honey et al., 2012; Murray et al., 2014). Hub regions within the default-mode brain network may
represent a structural signature of near-critical behaviour. Regions comprising this network exhibit
coordinated activity in the resting-state when the coherence between nodes of other "task-positive"
networks is generally suppressed (Cocchi et al., 2013; Fox et al., 2005; Hearne et al., 2015).
These observations suggest mechanisms through which critical dynamics may adapt to, and reshape
the complex nervous systems in which they occur, particularly the relationship between synaptic
processes, functional connectivity and network topology (Rubinov et al., 2009; Zhigalov et al., 2017).
3.3. Computational aspects of neuronal criticality
The pioneering work of Freeman (on bifurcations and multistability), and Beggs and Plenz (on critical
avalanches) provided proof-of-principles that the science of criticality could be used to inform our
understanding of complex patterns of activity in the brain and, by extension, behaviour. Research in
these areas has accelerated dramatically and now yields a stream of important discoveries spanning
from the neuronal (Gal and Marom, 2013; Gollo et al., 2013) to the whole-brain scale (Kitzbichler et
al., 2009).
As reviewed above, the study of criticality in physical systems using simple models suggested that
systems at the critical state are endowed with optimal computational properties. Such advantages have
recently been demonstrated in models and empirical recordings of critical neuronal systems [for
review, see (Beggs, 2007; Shew and Plenz, 2013)]. Perhaps most crucially, optimal dynamic range –
the sensitivity of a neuronal system to respond to, and amplify, inputs across a broad spectrum of
intensities – was shown to be maximized in models of neuronal systems tuned to a critical state
(Kinouchi and Copelli, 2006; Larremore et al., 2011). In slice cultures grown on the surface of
multielectrode arrays, Shew and colleagues later provided empirical evidence for this proposal,
showing that the maximum dynamic range to electrical perturbation was indeed maximized when the
cultures were pharmacologically manipulated to be close to criticality (Shew et al., 2009). Recent
electrophysiological recordings from the anaesthetized rat provided the first in vivo evidence that
dynamic range in perceptual systems is maximized when background activity is at the critical point
(Gautam et al., 2015).
Another example of the computational advantages of criticality arises from simplified neuronal
models which predict high fidelity and optimal information transmission (maximum mutual
information between sender and receiver) at criticality (Beggs and Plenz, 2003; Greenfield and Lecar,
2001): This prediction was also later observed in vitro (Shew et al., 2011) and more recently in
awake, behaving mice (Fagerholm et al., 2016).
Research in this field has also suggested that information storage and capacity – the ability of a
system to encode a broad repertoire of complex states from which information can be decoded
16
Criticality in the brain
Cocchi et al. 17
⁄
𝑚)
, where Fmax corresponds to the saturated response, and 𝑆0
(Gatlin, 1972)– may be optimized at criticality. Again, this notion has been captured in simple neural
models (Bertschinger and Natschläger, 2004; Haldeman and Beggs, 2005; Yang et al., 2017) and also
demonstrated in empirical recordings, including those arising in unperturbed (resting state) recordings
(Breakspear, 2001; Deco and Jirsa, 2012) as well as through careful pharmacological manipulations of
in vivo neurophysiological recordings (Stewart and Plenz, 2006). Selective enhancement of weak (but
not strong) stimuli also occurs near a phase transition (Copelli, 2014). It is also interesting to note that
optimal computational properties can arise at both continuous and discontinuous phase transitions
(Gollo et al., 2012).
A compelling argument for the advantage provided by the critical state derives from the analysis of
psychophysics experiments, which quantify the relationship between physical stimuli and perceptual
responses. Psychophysics relations may represent the earliest documented evidence of critical
dynamics in the nervous system (Kello et al., 2010). Across a variety of sensory modalities, these
experiments showed power-law relations, known as Steven's laws, in which the perceived
psychophysical or neuronal response F is given by: 𝐹(𝑆) ∝ 𝑆𝑚, where S is the stimulus level and m is
the Stevens exponent (Stevens, 1975). To account for the saturation of the response that occurs for
extreme stimuli, the psychophysics response is modelled by a sigmoid Hill function: 𝐹(𝑆) ∝
the input level for
𝐹𝑚𝑎𝑥 𝑆𝑚 (𝑆𝑚 + 𝑆0
half-maximum response. Importantly, this function also retains the power-law regime governed by the
exponent m. This power-law behaviour has a key putative function: It allows animals to distinguish
stimulus intensity varying across many orders of magnitude. As such, the power-law regime can
compress decades of stimulus variation S into a single decade of response F. A standard means to
measure this coding performance is called the dynamic range. The larger the dynamic range, the
better the ability to detect changes in stimuli. Modelling the neuronal behaviour at the sensory
periphery, Kinouchi and Copelli showed that a large dynamic range emerges from the collective
response of a network of many interacting units, and, more importantly, the dynamic range is optimal
when the network is at the critical state (Kinouchi and Copelli, 2006). This work explains the long-
lasting psychophysical scaling relations (Steven's law), and provides a clear example in which a
meaningful biological feature is optimised at criticality. Similar to earlier proposals, the work of
Kinouchi and Copelli was also based on a fairly simple model of neuronal activity. Crucially,
however, the model generated a prediction that was subsequently verified experimentally (Shew et al.,
2009).
The explanation of the power-law regime of psychophysics laws in terms of the optimal sensitivity of
critical states was an important contribution. However, as usual in analytic approaches, some
simplifications were made, leading to at least two conundrums. The first challenge refers to the
important trade-off between sensitivity and specificity. The finding that optimal signal coding occurs
at criticality implies maximum sensitivity. Yet, the specificity of this state is compromised because of
increased levels of fluctuation. As illustrated in Figures 1 and 2, criticality corresponds to the state
with the largest macroscopic fluctuations. This is why the critical state allows for the amplification of
stimuli of small intensity, which enhances the ability to distinguish the stimulus intensity varying over
orders of magnitude (i.e., large sensitivity). However, the very same effect is also a limitation because
the high fluctuations of criticality reduce the specificity of the system. The issue is whether the
improved sensitivity remains beneficial to the system when the reduction in specificity is also taken
into account. Fortunately, a solution for the sensitivity-specificity conflict exists in the presence of
diversity amongst components of the system. Heterogeneous excitable systems exhibit recruitment
properties in which units are recruited following their order of excitability (Gollo et al., 2016).
Therefore, a state of optimal sensitivity has units forming a subpopulation in a critical regime as well
17
Criticality in the brain
Cocchi et al. 18
as subpopulations operating in a non-critical state with improved reliability. Hence, optimised systems
represent the coexistence of subpopulations of reliable units and poor sensitivity (poised away from
their critical state) with a subpopulation of unreliable units and great sensitivity (typical of the critical
state). In other words, optimal perceptual performance may rely on the contribution of critical and
non-critical units (Gollo, 2017).
The second challenge emerging from the explanation of the power-law regime of psychophysics laws
derives from to the fact that the critical state for optimal psychophysical response of Kinouchi and
Copelli separates an active state from an inactive state (Kinouchi and Copelli, 2006). The issue here is
that the inactive state corresponds to an absorbing state in which the system cannot escape unless
external stimuli are provided. Hence, if the system falls into this state, it will get trapped there.
Critical systems typically do fall into this state because of their enhanced fluctuations. If the system
corresponds to the brain or a part of the nervous system, such an inactive state would be expected to
occur rather frequently. However, such silent states are not observed in vivo. The solution to this issue
also derives from the incorporation of an essential, but often overlooked, ingredient: inhibition.
Larremore and colleagues showed that ceaseless activity and critical avalanches coexist when a
substantial fraction of the units are inhibitory (Larremore et al., 2014). Hence, evidence is gradually
accumulating for the previously idealised proposal that biological systems can exploit special features
of the critical state. Meanwhile, models supporting this proposal are incrementally incorporating
greater physiological detail.
Research in the auditory system has highlighted the advantages conferred by active responses of hair
bundles poised in a critical state (Camalet et al., 2000; Eguíluz et al., 2000). In response to sound-
wave input, hair bundles oscillate and their dynamics vary from a steady to an oscillatory regime.
Each bundle is tuned to a particular natural frequency and adjacent cells respond maximally to
successive pitches, giving rise to a tonotopic map in the cochlea (Romani et al., 1982). Crucially,
computational modelling suggests that active and nonlinear hair bundles that operate near a Hopf
bifurcation optimise the cochlea's performance and enhance the main features of auditory coding,
such as, amplification, frequency selectivity and compressive nonlinearity (Hudspeth, 2008;
Maoiléidigh et al., 2012).
3.4. Self-organised neuronal criticality
A growing body of empirical work has thus asserted the presence of scale-free statistics across a
diversity of in vitro and in vivo neural recordings, while computational models have highlighted its
computational advantages. Of note, the critical activity observed by Beggs and Plenz (2003) was
stable for many hours and did not require careful tuning of the parameters of their culture (i.e. the pH,
temperature, etc). Simple models of criticality classically rely upon fine-tuning of system parameters
to a critical value (Levina et al., 2014). What is the basis for a robustness that apparently eschews the
need for such a balancing act? As discussed above (Section 2.4), analyses of critical systems in
physical systems reconcile this paradox by recourse to self-organised criticality (SOC). In brief, SOC
arises when the interactions amongst system components are imbued with some form of plasticity,
such as when system energy accumulates and is then dissipated by a large-scale avalanche. During
periods of quiescence, energy slowly accumulates until it tips the system into (or above) criticality.
The consequent energy dissipation briefly renders the system sub-critical until further energy
accumulates. There is thus a time scale separation between the fast system dynamics and the slow
build up and dissipation of energy.
18
Criticality in the brain
Cocchi et al. 19
Several models of criticality in the brain incorporate such slow processes (Marković and Gros, 2014).
A considerable body of research has focused upon the role of various forms of synaptic plasticity,
including simple activity-dependent up- and down-regulation (de Arcangelis, 2008), activity-
dependent synaptic plasticity (de Arcangelis et al., 2006), synaptic potentiation (Stepp et al., 2015),
short-term synaptic depression through depletion of synaptic vesicles (Bonachela et al., 2010; Levina
et al., 2014; Mihalas et al., 2014; Millman et al., 2010), Hebbian (Van Kessenich et al., 2016) and
anti-Hebbian synaptic plasticity (Cowan et al., 2014; Magnasco et al., 2009), and spike-time
dependent plasticity (de Andrade Costa et al., 2015; Rubinov et al., 2011). As with physical systems,
the (relatively) slow synaptic plasticity serves to broaden the critical point to a broad, stable region.
Other neurobiological processes have also been proposed, including balanced excitation-inhibition
and network topology (Rubinov et al., 2011) and dynamic neuronal gain (Brochini et al., 2016). More
recently, the role of energy build-up and dissipation in physical systems has been recast in critical
neural systems as the replenishment and depletion of intracellular metabolic resources including
Adenosine Triphosphate (ATP); (Roberts et al., 2014b; Stramaglia et al., 2015; Virkar et al., 2016).
4. Challenges and pitfalls of the criticality hypothesis
Despite this recent emergence of criticality research in neuroscience, lessons learned in other branches
of science raise important pitfalls and caveats. First, inferring the presence of scale-free statistics in
neuroscience data has classically rested upon fitting a power-law (or Pareto) regression to the
probability distribution of the size of the temporal or spatial fluctuations (Figure 1 and Figure 2). The
statistical principles underlying this exercise were critiqued in a highly influential survey by Clauset
and colleagues (Clauset et al., 2009). While a linear regression in double logarithmic coordinates can
yield a fit that looks impressive, such a process is insensitive to the distribution of data at the right-
hand tail of the distribution – the very region where the presence of a heavy-tailed power law needs to
be rigorously tested. This is because the number of empirically measured samples found in the tail of
the distribution is often too limited for robust inference. Another concern is that the samples of a
cumulative distribution function are not independent, whereas regression assumes data independence.
Clauset and colleagues developed a more principled approach based on maximum likelihood
estimation to test and compare different statistical models of the data, including the power law, but
also other candidate heavy tailed distributions including the log normal and stretched exponential
forms (Clauset et al., 2009; Vuong, 1989). In brief, once the best fitting power-law parameters for an
empirical dataset have been determined with maximum likelihood estimation, the goodness-of-fit
between the fitted power law and empirical distribution is tested with the Kolmogorov-Smirnov (K-S)
statistic. To this end, Clauset and colleagues proposed to randomly sample data from the fitted power
law, independently fit a new power law to each of these new data samples and then evaluate the
goodness-of-fit between the new samples and the new power laws. This is repeated many times to
generate an empirical distribution of K-S statistics, which can then be used to compute a p-value for
the K-S statistic corresponding to the observed data. If this p-value is significant, the randomly
sampled data is a better fit to the power law than the observed data, and thus a power law should be
excluded as an appropriate model. Otherwise, if the p-value is not significant, power-law behaviour is
supported and the final step is to exclude other distributions as providing better evidence. Relative fits
are typically computed between candidate distributions (lognormal, stretched exponentials, etc.) using
log-likelihoods.
Using this approach, Clauset et al. (2009) revisited several physical phenomena thought to have scale-
free statistics and showed that several of these data were better explained by other long-tailed
distributions, not power laws. Using these methods shows that the same holds true for many neuronal
19
Criticality in the brain
Cocchi et al. 20
fluctuations (Roberts et al., 2014a). For example, it appears that fluctuating alpha rhythm follows a
stretched exponential distribution, not a power law (Freyer et al., 2009). In turn, biophysical models
suggest that the alpha rhythm arises from noise-driven multistability, rather than being generated by
classic (super-) criticality (Freyer et al., 2011).
A second caveat was issued by the empirical analyses and modelling work of Tomboul and Destexhe
(Touboul and Destexhe, 2010, 2015) who showed that under certain situations, the aggregate
behaviour of non-critical stochastic systems could yield irregular time series with power law statistics,
albeit over a limited range. This is an important issue which also highlights the importance of null
models for the different experimental methods (Farmer, 2015). These findings suggest that inferences
regarding criticality based on the observation of power law scaling in empirical data should be made
with caution, particularly if the scaling extends for less than two orders of magnitude or the slope of
the power law is steep (Miller et al., 2009) - i.e. the scaling exponent is greater than 2.5.
A third caveat concerns other classes of interesting, emergent phenomena. Non-trivial, emergent
dynamics can arise through other complex nonlinear phenomena. A classic example is that of so-
called winnerless competition (Melbourne et al., 1989; Rabinovich et al., 2001) that has been
proposed to underlie cognitive tasks such as animal gait (Golubitsky et al., 1999), perceptual rivalry
(Ashwin and Lavric, 2010), and sequential decision-making (Rabinovich et al., 2008). Winnerless
competition arises from metastable transitions along a sequence of unstable states (not unlike the
drawings of M.C. Esher). Unlike criticality, the successive states are not weakly stable, but are
unstable (Figure 4). In common with critical systems, noise plays a crucial role in a metastable
system. However, the statistics of a metastable system are not power laws (Figure 4b). Rather the
duration that the system dwells near each of its states varies in proportion to the logarithm of the noise
amplitude (Ashwin et al., 2006). The temporal statistics therefore have a characteristic (and relatively
short) time scale, and are not scale-free (Figure 4b). While multistable (Figure 4a) and metastable
(Figure 4b) systems are therefore related (particularly in name!), their statistics are distinct and the
underlying nonlinear causes differ. Unfortunately, the two terms are often used interchangeably.
A number of final caveats pertain to the practical aspects of empirical data. First, the algorithm
developed by Clauset et al. (2009) assumes that there is no upper bound to the empirical power-law
distribution. This is a flawed assumption for most experimental data, which inevitably derive from a
finite number of sensors, and may bias model selection. Recent work has revisited this assumption,
developing methods that test the likelihood of a power law with a simple cut-off (Langlois et al.,
2014; Shew et al., 2015). Second, as mentioned above, the range of many data tested for power law
scaling often span less than two orders of magnitude, yielding data that is particularly sparse in the
right-hand tail (precisely where power law scaling is most clearly expressed). Although the use of
model estimation rather than linear regression partly mitigates this, disambiguating amongst the
variety of candidate heavy-tailed distributions can only be reliably performed when the data scale
over more than two orders of magnitude. Lengthy acquisitions may help here. For example, free-
living activity patterns in humans derived from accelerometry recordings over seven consecutive days
scale across four orders of magnitude: These allow for disambiguation of composite exponential and
truncated power law distributions in active versus inactive periods of the day (Chapman et al., 2016).
20
Criticality in the brain
Cocchi et al. 21
Figure 4: Different expressions of instability lead to different types of complex dynamics. (a) In a multistable
system, noise drives a system erratically between different attractors. Because the system is briefly trapped in
each basin of attraction, the time series shows a relatively long-tailed (stretched exponential) dwell
distribution, here shown in linear-log coordinates (inset). (b) In a metastable system, there are no attractors,
but rather a sequence of linked unstable fixed points. Because these are only weakly unstable, the system
dwells in the neighbourhood of each, but does not show trapping. The sequential dwell times are therefore
not long-tailed but show a characteristic time-scale corresponding to the peak in a gamma function shown
here in linear coordinates (inset). In a critical system, a single fixed point is very weakly attracting or neutral.
System noise leads to long and unstructured excursions corresponding to scale-free fluctuations and
corresponding power-law statistics. Disambiguating these different underlying causes of complex dynamics
21
Criticality in the brain
Cocchi et al. 22
the
imperfect measurement process: In
can be achieved with careful analyses of the system statistics, together with inversion of corresponding
computational models.
Finally, noisy fluctuations in physiological data do not only arise from the system of interest (i.e.
neural activity), but also from
the setting of
neurophysiological data, these fluctuations consist of additive noise from physiological sources (such
as muscular activity, cardiovascular influences) as well as artefacts due to extraneous effects (e.g.
thermal scanner noise, head motion). In principle, such effects could lead to false positives in power-
law evaluation. However, these inputs are generally uncorrelated and their summation therefore
(according to the central limit theorem) likely to be Gaussian, not heavy-tailed. Nonetheless, care
should be taken to disambiguate their contribution to any putative heavy-tailed system statistics in
case one particular artefact (such as head movement in the scanner) dominates. Methods of doing this
include: (i) taking independent measurements of these artefacts, such as taking empty scanner room
recordings (Shriki et al., 2013) and ensuring that they do not possess the same statistics as attributed
to the underlying neuronal system (Kitzbichler et al., 2009); (ii) using algorithms such as independent
components analysis or source reconstruction to unmix neuronal fluctuations from physiological and
measurement noise (Freyer et al., 2009); (iii) and using a formal inversion framework that formally
accommodates measurements effects including spatial or temporal filtering and additive noise (Razi et
al., 2015).
These caveats highlight crucial points. While the application of criticality to neuroscience is an
exciting field, progress needs to proceed with due caution. Analyses of neuroscience data for power
laws first needs to consider other heavy-tailed candidate distributions. Second, inference should
ultimately be based upon models of the causes of the observed statistics and avoid a direct inference
of criticality that is based only upon data analysis. Third, computational models of neural systems that
are based upon criticality should be tested closely against empirical data using appropriate
frameworks (Daunizeau et al., 2009; Penny, 2012) that allow disambiguation against competing
models that invoke other nonlinear mechanisms [for review, see Roberts et al. (2015)].
While these caveats highlight important limitations, experimental manipulations and recent theoretical
developments offer new opportunities to explore the "criticality hypothesis". As noted at the outset of
this review, the notion of universality (properties that transcend the details of a particular system and
are thus found in many diverse settings) is one of the central appeals of criticality. Universal scaling
laws - such as the presence of a scaling function that inter-relates the common underlying shape of
critical fluctuations across temporal and spatial scales (Sethna et al., 2001; Zapperi et al., 2005) - can
be extracted from data and subject to null hypothesis testing (Friedman et al., 2012; Roberts et al.,
2014a). Relationships between the exponents of different (spatial and temporal) scaling laws, may
also be derived from empirical data (Friedman et al., 2012; Roberts et al., 2014a) and benchmarked
against the simple relationships predicted by the mathematical theory of phase transitions (Sethna et
al., 2001). Also, as reviewed above, the theoretical advantages of criticality in neuronal models was
demonstrated in a series of elegant empirical studies using pharmacological manipulation to sweep
systems from subcritical to critical to supercritical (Gautam et al., 2015; Shew et al., 2011; Shew et
al., 2009; Tagliazucchi et al., 2016): Showing a sudden change in a scaling law, corresponding to a
peak in information capacity or dynamic range, provides convergent evidence for the occurrence of a
phase transition in the underlying system.
22
Criticality in the brain
Cocchi et al. 23
5. Emerging role of criticality in cognition
Notwithstanding the aforementioned caveats, growing empirical and modelling research clearly
supports the view that neural dynamics likely occur near critical instabilities. The recognition of the
limitations of this new field simply shows that it has matured beyond the "proof of principle" stage
(Feyerabend, 1993). The scene is thus set for the translation of criticality into cognitive and clinical
brain research.
In Section 3.1, we noted a canonical example of critical fluctuations near the transition from anti-
syncopated to syncopated rhythmic finger tapping. Do critical dynamics generalize to other
behaviours? Accelerometer-based analyses of free behaviour in humans (going about their everyday
lives) shows that periods of inactivity exhibit power-law statistics (Nakamura et al., 2007), whose
scaling coefficients differ between wake and sleep (Chapman et al., 2016) and differ again in major
depression (Nakamura et al., 2008). Intriguingly, accumulating evidence suggests that periods of
activity – although long-tailed – do not fit a power law, but rather a stretched exponential (e.g., inset
Figure 4a, depicting the Weibull distribution). This shift in the statistical proprieties of a system as a
function of context is emerging as a powerful tool to understand the neural principles supporting
healthy and pathological brain functions.
5.1. Criticality in brain and behaviour
In traditional cognitive neuroscience experiments, separate trials are typically treated as independent.
However, they are not necessarily treated independently by research participants. Indeed it has been
shown that reaction times between sequential trials show long-range correlations across a wide
diversity of tasks (Palva and Palva, 2011; Thornton and Gilden, 2005). This is perhaps not surprising,
given that, outside the scanner, human activity (Sreekumar et al., 2016; Sreekumar et al., 2014) and
memory (Nielson et al., 2015) show a complex temporal structure that scales across many orders of
magnitude. Detailed analyses of reaction time data in typical psychophysics experiments support the
presence of a power law structure (Van Orden et al., 2005). However, definitive support for this
position, as well as putative underlying causes such as criticality, do remain contentious (Heathcote et
al., 2000; Heathcote et al., 1991; Wagenmakers et al., 2004).
A rich literature on critical dynamics in brain and behaviour exists. However, with few notable
exceptions (Jirsa et al., 1994; Kelso et al., 1992), these two streams of research have preceded largely
in parallel. Several recent papers have stepped toward a unifying framework. By acquiring high
density MEG/EEG data while participants performed an audiovisual threshold-stimulus detection task
Palva et al. (Palva et al., 2013) showed that the critical exponents of scale-free neuronal dynamics
correlate with the inter-individual variability in behavioural scaling laws. These correlations, show a
specific anatomical pattern, with the combined delta and alpha frequency bands correlations mapping
onto posterior parietal cortex. The same cortical areas appear
involved when
electrophysiological signals in the beta and gamma ranges are considered, with the addition of the
cuneus and inferotemporal brain areas. Interestingly, a significant association between neural and
behavioural long-range temporal correlations was also found in brain regions comprising the default
mode brain network (Greicius et al., 2003). Notably, the brain-behaviour association found in task
execution largely overlaps with an association between visual behavioural performances and scale-
free dynamics observed in resting-state MEG data. This suggests that the temporal structure of
endogenous neural dynamics are, to some extent, preserved during task performance. Similarly, the
analysis of auditory task data showing that cortical activations organize as neural avalanches in both
visual cognitive and resting-state contexts [(Arviv et al., 2015), see also (Shew et al., 2015)].
to be
23
Criticality in the brain
Cocchi et al. 24
Moreover, near-critical neural dynamics may determine fluctuations in perceptual and cognitive
processes. Overall, these findings highlight the strong interdependence between the near-critical
neural processes characterizing resting-state and task-specific processes.
5.2. Suppression of criticality during task performance
Using simultaneously acquired EEG and fMRI data, Fagerholm et al (Fagerholm et al., 2015) recently
recapitulated previous findings of scale-free cortical dynamics at rest [e.g., (Palva et al., 2013)].
However, as participants engaged in tasks of increasing attentional load, the statistics of cortical
activity appeared to shift increasingly further from the critical state. It was surmised that although
criticality was important for the unconstrained "exploratory" resting-state, executing specific tasks
required suppression of the associated variability under the influence of dorsal attentional networks
(Fagerholm et al., 2015; Hellyer et al., 2014). In keeping with this notion, complementary
computational analyses have suggested that altered anatomical connectivity, cognitive flexibility and
information processing following traumatic brain injury are linked to a reduction in the variance of
functional connectivity simulated by a model of coupled phase oscillators (Hellyer et al., 2014).
The proposition that near-critical neural dynamics are central to the emergence of conscious cognition
is consistent with the results of high-density electrocorticography recordings (ECoG) in primates,
showing that loss of consciousness is characterized by a reduction in the number of eigenmodes (the
number of "excited modes" in a dynamic system) that are close to instability (Solovey et al., 2015).
Conversely, the return of consciousness appears to be accompanied by a corresponding increase in the
number of eigenmodes close to instability. Recent detailed analyses of in vivo electrophysiological
and two-photon recordings in rodents provide direct support for a clear association between the
appearance of critical activity and the emergence of consciousness from anaesthesia (Bellay et al.,
2015; Scott et al., 2014). Likewise, induction of unconsciousness through propofol anaeasthesia is
accompanied by a loss of the signatures of criticality in fMRI data (Tagliazucchi et al., 2016).
Non-invasive and invasive human neuroimaging work assessing neural critical states by the mean of
neural avalanches analysis further support the link between near-critical neural regimes and behaviour
(Palva et al., 2013; Priesemann et al., 2013; Tagliazucchi et al., 2012). For example, ECoG data
collected from patients with refractory partial epilepsy were analysed as a function of vigilance
(Priesemann et al., 2013). In general, the probability distribution frequency of neural avalanches
followed a power law. However, slow wave sleep (SWS) was characterized by the most frequent
occurrence of avalanches; unconstrained wakefulness showed an intermediate occurrence, while rapid
eye movement sleep (REM) showed the fewest. These results are in line with previous findings from
invasive recording in rats (Ribeiro et al., 2010) and highlight the close relationship between changes
in near-critical dynamics and distinct mental states.
It has been suggested that near-critical dynamics in the default mode operate to increase the "dynamic
repertoire" of the brain when subjects are at rest (Deco and Jirsa, 2012; Deco et al., 2011) – that is, to
increase the number of proximal "cognitive sets" or stored memories (Scarpetta and de Candia, 2013).
This is consistent with modelling and empirical work, reviewed above, that reveals that criticality
optimizes the entropy (Fagerholm et al., 2016) and information capacity (Shew et al., 2011) of
spontaneous neuronal activity. Criticality at rest could also function to tune cortical dynamics to an
optimal state of "expected uncertainty" concerning the nature of as yet unencountered sensory inputs
(Friston et al., 2012b). That is, by facilitating a flexible, but also constrained degree of instability in
sensory-perceptual systems, criticality could endow the cortex with the ability to adapt to volatile and
non-stationary environmental fluctuations: Away from the critical point, internal fluctuations are
24
Criticality in the brain
Cocchi et al. 25
stable and strongly damped, hence shrinking toward a baseline (mean) value. Close to the critical
point, such fluctuations encompass a broad spectrum of scales, similar to the scale-invariant statistics
of natural scenes (Field, 1987; Ruderman and Bialek, 1994). At a very fundamental level, this
matching of internally-generated fluctuations to the statistics of external scenes, via criticality, revisits
the notion that spontaneous cortical activity encodes an optimal model of the environment (Berkes et
al., 2011) and uses this to deploy adaptive behavioural strategies (Fiser et al., 2010), including
saccadic (Friston et al., 2012a) and heavy-tailed fixational eye movements (Roberts et al., 2013).
Upon task execution - as uncertainty decreases - this unstable dynamic landscape could accordingly
be suppressed, allowing stabilization of a single task-related attractor. Such proposals are consistent
with the broader recognition that "brain noise" plays an adaptive role in health (Garrett et al., 2011)
and ageing (McIntosh et al., 2010), but decreases during task execution (Churchland et al., 2010;
Ponce-Alvarez et al., 2015).
By analysing the statistics and spatiotemporal scaling laws of neural avalanches from ex vivo local
field potential (LFP) recordings of visual cortex, Shew et al. (2015) showed that strong visual
stimulation initially engendered super-critical dynamics. However, with the continuation of the visual
input, these were quickly tuned to a critical state through neuronal adaptation. A simple neuronal
model with the adaptation to input mediated by short-term synaptic depression was able to capture the
switch from non-critical to critical activity with continuous input (Shew et al., 2015). These results
highlight the importance of synaptic plasticity in switching between critical and non-critical regimes
in order to facilitate perception and cognition.
6. Criticality in disease
6.1. Bifurcations and seizures
Whereas the role of criticality in cognition is relatively nascent, casting seizures as dynamic disorders
that arise out of critical instabilities is supported by an appreciable body of evidence (Da Silva et al.,
2003; Meisel et al., 2012). The primary generalized seizures of childhood - Absence seizures –
correspond to the presence of high amplitude 3 Hz spike-and-wave oscillations that appear and
terminate equally quickly. These seizures have been modelled as critical bifurcations in
corticothalamic loops by several groups (Destexhe and Sejnowski, 2001) although there is yet a lack
of consensus as to whether these bifurcations are subcritical (Fröhlich et al., 2010; Suffczynski et al.,
2004) or supercritical (Breakspear et al., 2006; Robinson et al., 2002). Generalised tonic-clonic
seizures, which are associated with the progression of various high amplitude waveforms, have been
modelled as a subcritical bifurcation (Breakspear et al., 2006). Accordingly, patients with epilepsy
"reside" in pathological multistable dynamic regimes, due to neurophysiological disturbances and are
occasionally perturbed into seizure dynamics (Breakspear et al., 2006). Fast epileptic activity
recorded invasively with cortical surface recordings have also been subject to detailed models using
critical bifurcations (Bartolomei et al., 2001; Wendling et al., 2005). These are proposed to reflect
aberrant interactions between excitatory and inhibitory cortical neurons due to GABA-ergic
dysfunction (Wendling et al., 2002). A detailed account of the complex progression of low frequency
(DC-like) shifts in the electrical baseline and the nested high frequency oscillations that superimpose
on these has recently been advanced using bifurcations: This "epileptor" - a set of equations that
express the corresponding nested bifurcations - represents a comprehensive integration of slow and
fast time scales underlying criticality and bifurcations in seizure activity (Jirsa et al., 2017; Jirsa et al.,
2014).
25
Criticality in the brain
Cocchi et al. 26
6.2. Crackling noise and neonatal burst-suppression
Burst suppression (BS) is a class of electrocortical activity that occurs in preterm neonates, full-term
newborn infants with encephalopathy, and following propofol anaesthesia. BS is characterised by
high amplitude, irregular bursts that erratically punctuate a flat EEG trace (Niedermeyer et al., 1999)
(Figure 5). Analyses of BS in preterm and encephalopathic infants show that the bursts are
characterized by long-tailed, scale-free properties that stretch across a remarkable six orders of
magnitude (Roberts et al., 2014a). While individual bursts are highly irregular, binning them
according to their size and then averaging all bursts within each bin yields very smooth shapes that
vary only slowly across scales (Roberts et al., 2014a) (Figure 5). The relationship between these
scale-specific shapes are described by a so-called scaling function that speaks to the nature of
criticality in all systems where it arises (Baldassarri et al., 2003; Papanikolaou et al., 2011). Notably,
the exponents that describe the scaling and shape of the bursts pre-empt clinical outcomes in these
neonates (Iyer et al., 2015b; Iyer et al., 2014). These shapes have been modelled by a term that
describes metabolic depletion and replenishment in these critically vulnerable newborn brains
(Roberts et al., 2014a). Despite the success in modelling BS in babies using a criticality framework,
bursts in propofol anaesthesia do not have scale-free properties (Ching et al., 2010). In contrast, the
effect of this type of anaesthesia in brain activity has been modelled with alternative (non-critical)
dynamical mechanisms (Bojak et al., 2015).
Figure 5: Critical statistics in neonatal hypoxic ischemic encephalopathy. (a) Scalp EEG channels show
characteristic bursting pattern of activity with bursts erratically punctuating a flat EEG trace. (b) The area
under each of these bursts (a measure of the energy discharged) shows power-law scaling over 3 orders of
magnitude, with an exponential truncation at the far right. (c) The relationship between the length and
duration of bursts also follows a power law (linear in double logarithmic coordinates. (d) Bayesian model
selection shows that a truncated power law is easily the best model for these data. Panels (a) adapted from
(Iyer et al., 2015a); panels (b), (c) and (d) adapted from (Roberts et al., 2014a).
26
Criticality in the brain
Cocchi et al. 27
Establishing the presence of critical instabilities in neurological disorders opens a number of novel
therapeutic and diagnostic windows [for review, see Coombes and Terry (2012)]. For example, in
concert with computational modelling, treating seizures as transitions across bifurcations permits
tracking the trajectory of the seizure through bifurcation space (Freestone et al., 2013; Nevado-
Holgado et al., 2012), with a longer-term objective of seizure control (Nelson et al., 2011).
Intriguingly, such control systems could use micro-stimulation to exploit the increased dynamic
responsiveness of a system close to criticality as an "early warning system" (Scheffer et al., 2009): To
our knowledge, incorporating this approach to seizure prediction within the framework of criticality
theory has not yet been achieved. Likewise, deriving measures of near-criticality from clinical, bed-
side recordings of burst suppression in preterm and encephalopathic neonates could be used as
prognostic markers of long-term outcome, hence opening novel therapeutic windows for early
intervention (Roberts et al., 2017).
6.3. Criticality and neuropsychiatric disorders
Criticality, bifurcations and phase transitions have provided increasingly nuanced understandings of
several major neurological conditions. It is intriguing to consider a possible role in neuropsychiatric
conditions such as psychotic and affective disorders, although these are likely to be more subtle in
their deviation from the "optimum critical point". Cognitive disturbances in schizophrenia, such as
working memory, have been captured in computational models by changing the influence of
fluctuations in a multistable landscape (Loh et al., 2007). Changes in large-scale resting-state
dynamics seen in this disorder – such as a decrease in the global brain signal – have also been
modelled by altering the balance of noise and stability in macroscopic neuronal activity (Yang et al.,
2014). Schizophrenia is associated with less suppression of activity in the default-mode network when
subjects engage in external tasks (Harrison et al., 2007; Meyer-Lindenberg et al., 2001; Nejad et al.,
2011; Whitfield-Gabrieli et al., 2009). Dynamic insights into this loss of efficiency could draw from
the theory of criticality. In particular, if adaptive cortical activity reflects a switch from sub-critical to
super-critical activity under the influence of attention and arousal, then it follows that the precision of
this process could be diminished in schizophrenia. Default activity would then be inadequately
suppressed during attention to external tasks, while super-critical activity would not be optimally
portioned into distinct spatiotemporal patterns. While this is clearly speculative, the tools of criticality
provide new methodological tools and innovative directions for research in this often-intractable
disorder.
Whereas schizophrenia is characterized by disorganization – suggesting too little stability – core
manifestations of other psychiatric disorders suggest too much stability. The ruminations of
depressive disorders and the maladaptive pre-occupations of obsessive-compulsive disorder may arise
in a system residing too deeply in the subcritical zone, preventing adaptive switching from
interospective to exteroceptive states. Analysis of resting-state fMRI from patients with melancholia
supports this view, showing a disturbance in the key dynamic parameters that modulate stability
(Hyett et al., 2015a). Notably, when patients with melancholia view emotionally salient material,
frontal attentional networks increase system stability, contrary to their influence in healthy controls
(Hyett et al., 2015b). As noted above, critical instability is needed for a system to respond to the
stream of stimuli. Here, too little instability is proposed to underlie the lack of reactivity.
Although these are preliminary findings, the potential for criticality to provide fresh insights into
tough neuropsychiatric research problems is again evident. It is quite possible that psychiatric
phenotypes do not only reflect a simple failure of critical brain dynamics, but also their replacement
by different dynamic processes in compensation (Breakspear et al., 2015).
27
Criticality in the brain
Cocchi et al. 28
7. Summary
Evidence for the widespread occurrence of criticality in nature, and its corresponding computational
advantages, has triggered the interest of scientists in many different fields. The list of advantages
associated with criticality spans many systems and different measurable quantities (Assis and Copelli,
2008; Boedecker et al., 2012; Deco et al., 2013; Gollo et al., 2013; Haldeman and Beggs, 2005;
Hidalgo et al., 2014; Kastner et al., 2015; Legenstein and Maass, 2007; Livi et al., 2016; Mosqueiro
and Maia, 2013; Publio et al., 2012; Schrauwen et al., 2009; Shew and Plenz, 2013). Despite the field
traditionally developing outside of neuroscience, many of the most exciting findings now focus on
brain dynamics. This body of work suggests that specific functions of the central nervous system may
exploit optimal properties observed at criticality, representing a general functional property of the
brain (Chialvo, 2006).
What are the mechanisms that establish criticality in the brain? The obvious lack of an externally
tunable parameter highlights the importance of self-organized criticality (SOC). As we reviewed
above, SOC arises when the interactions between the elements (neurons) of a system are endowed
with plasticity, or when there is a slow build-up and a fast release of energy. Synaptic processes such
as frequency adaptation or spike-time dependent plasticity are clear candidates for SOC at the
neuronal level. Homeostatic mechanisms, such as the hypothalamus-pituitary axis, govern metabolic
processes in the body, and reflect the slow time-scales of energy accumulation and utilization. In this
vein, neuronal criticality may extend beyond the central nervous system to incorporate the autonomic
nervous system and corresponding interoceptive feedback loops to the midbrain and insula. Although
criticality can arise from a fairly simple memory process, it seems unlikely to be confined to any
single mechanism in the brain, but may reflect multiple processes operating for different goals
(computation, adaptation, homeostasis). Likewise, changes in the environment, such as a scarcity of
food sources, which mandate an (allostatic) change in the balance of energy metabolism may also
engage SOC processes through a change in the critical set point. These more complex feedback
systems are not classically thought of as SOC, but rather homeostatic, self-regulatory systems of the
brain and body. Future work is required to understand if these broader homeo- and allostatic processes
do follow the principles of SOC.
Conversely, pathological failures of adaptive criticality may reflect disturbances in any of the
supporting mechanisms, such as when a failure of balanced inhibition leads to a seizure, or when
interoceptive disturbances lead to fatigue and depression (Petzschner et al., 2017). Criticality may
also arise in pathological settings, such as if metabolic disturbances lead to an abnormal build-up and
release of energy that would not normally occur: Burst suppression in the hypoxic newborn may be an
example of this scenario (Roberts et al., 2015).
The study of the brain as a complex, dynamic network is now flourishing. Modelling the inherently
unstable, multiscale neuronal dynamics that generate flexible cognitive states is crucial to this
endeavour. Although the notion of criticality is an attractive candidate, much work remains to be done
to clarify the relationship between cognition and criticality. The increasing sophistication of
theoretical models and their hypothesis-driven application to empirical data has to be tempered
against the caveats that have recently emerged. Research that incorporates these lessons will shed new
light on the dynamic brain networks that underlie action, perception and cognition in health and
disease.
28
Criticality in the brain
Cocchi et al. 29
Unifying models of brain function give sense to the results generated by the boom of explorative
studies assessing the association between cognitive functions and neural dynamics. Without such
models, there is a risk that the accumulation of experimental evidence will not substantially impact
upon our understanding of the neural underpinnings of human cognition in health and disease. Recent
studies suggest that the transition from critical to super-critical regimes may represent a general
principle underlying the emergence of goal-directed behaviour. Those transitions may be moderated
by slow integrators that are the highly connected hubs of the brain. In general, the framework of
criticality challenges earlier prevailing conceptualizations of normal and pathological cognitive
functions as emerging from discrete regions or static networks of brain regions.
The science of criticality provides a new armoury of analytic techniques for basic and translational
neuroscientists. The field is now sufficiently mature that inferring the presence of scale-free statistics
from time series data should derive from formal likelihood tests and comparisons to other heavy-
tailed processes. In addition, showing that the spatial and/or temporal statistics of a system are scale-
free is now only the first step in emerging "best practice". Inferring that the underlying system is
critical should ideally rest upon showing that the best model that can account for those statistics is one
in which the model itself is perched at the cusp of criticality. Criticality should not become a catch-all
term for everything that is complex or variable. It is only one amongst many possible causes of
complex dynamics and care should be taken to disambiguate amongst these.
Funding: L.C., L.L.G, A.Z and M.B. were supported by the Australian National Health Medical
Research Council (L.C. APP1099082, L.L.G. APP1110975, A.Z. APP1047648, M.B. APP1037196).
This work was also supported by the Australian Research Council Centre of Excellence for
Integrative Brain Function (M.B., ARC Centre Grant CE140100007).
Conflicts of Interest: The authors declare that the research was conducted in the absence of any
commercial or financial relationships that could be construed as a potential conflict of interest.
Acknowledgments: Dedicated to the memory of Walter Freeman whose seminal early work inspired
us and shaped the field.
29
Criticality in the brain
Cocchi et al. 30
8. References
Arneodo, A., Coullet, P., Tresser, C., 1980. Occurence of strange attractors in three-dimensional
Volterra equations. Physics Letters A 79, 259-263.
Arviv, O., Goldstein, A., Shriki, O., 2015. Near-Critical Dynamics in Stimulus-Evoked Activity of the
Human Brain and Its Relation to Spontaneous Resting-State Activity. The Journal of
Neuroscience 35, 13927-13942.
Ashwin, P., Burylko, O., Maistrenko, Y., Popovych, O., 2006. Extreme sensitivity to detuning for
globally coupled phase oscillators. Physical review letters 96, 054102.
Ashwin, P., Lavric, A., 2010. A low-dimensional model of binocular rivalry using winnerless
competition. Physica D: Nonlinear Phenomena 239, 529-536.
Assis, V.R., Copelli, M., 2008. Dynamic range of hypercubic stochastic excitable media. Physical
Review E 77, 011923.
Bak, P., 1990. Self-organized criticality. Physica A: Statistical Mechanics and its Applications 163,
403-409.
Bak, P., Tang, C., Wiesenfeld, K., 1987. Self-organized criticality: An explanation of the 1/f noise.
Physical review letters 59, 381-384.
Bak, P., Tang, C., Wiesenfeld, K., 1988. Self-Organized Criticality. Phys Rev A 38, 364-374.
Baldassarri, A., Colaiori, F., Castellano, C., 2003. Average shape of a fluctuation: Universality in
excursions of stochastic processes. Physical review letters 90, 060601.
Bartolomei, F., Wendling, F., Bellanger, J.-J., Régis, J., Chauvel, P., 2001. Neural networks involving
the medial temporal structures in temporal lobe epilepsy. Clinical neurophysiology 112, 1746-
1760.
Bassett, D.S., Wymbs, N.F., Rombach, M.P., Porter, M.A., Mucha, P.J., Grafton, S.T., 2013. Task-
based core-periphery organization of human brain dynamics. PLoS computational biology 9,
e1003171.
Beggs, J.M., 2007. Neuronal avalanche. Scholarpedia 2, 1344.
Beggs, J.M., Plenz, D., 2003. Neuronal avalanches in neocortical circuits. The Journal of
neuroscience : the official journal of the Society for Neuroscience 23, 11167-11177.
Beggs, J.M., Timme, N., 2012. Being critical of criticality in the brain. Frontiers in physiology 3, 163.
Bellay, T., Klaus, A., Seshadri, S., Plenz, D., 2015. Irregular spiking of pyramidal neurons organizes
as scale-invariant neuronal avalanches in the awake state. Elife 4, e07224.
Berkes, P., Orbán, G., Lengyel, M., Fiser, J., 2011. Spontaneous cortical activity reveals hallmarks of
an optimal internal model of the environment. Science 331, 83-87.
Berryman, A.A., 1992. The Orgins and Evolution of Predator‐Prey Theory. Ecology 73, 1530-1535.
Bertschinger, N., Natschläger, T., 2004. Real-time computation at the edge of chaos in recurrent
neural networks. Neural computation 16, 1413-1436.
Biswal, B.B., Mennes, M., Zuo, X.N., Gohel, S., Kelly, C., Smith, S.M., Beckmann, C.F., Adelstein,
J.S., Buckner, R.L., Colcombe, S., Dogonowski, A.M., Ernst, M., Fair, D., Hampson, M.,
Hoptman, M.J., Hyde, J.S., Kiviniemi, V.J., Kotter, R., Li, S.J., Lin, C.P., Lowe, M.J.,
Mackay, C., Madden, D.J., Madsen, K.H., Margulies, D.S., Mayberg, H.S., McMahon, K.,
Monk, C.S., Mostofsky, S.H., Nagel, B.J., Pekar, J.J., Peltier, S.J., Petersen, S.E., Riedl, V.,
Rombouts, S.A., Rypma, B., Schlaggar, B.L., Schmidt, S., Seidler, R.D., Siegle, G.J., Sorg,
C., Teng, G.J., Veijola, J., Villringer, A., Walter, M., Wang, L., Weng, X.C., Whitfield-
Gabrieli, S., Williamson, P., Windischberger, C., Zang, Y.F., Zhang, H.Y., Castellanos, F.X.,
Milham, M.P., 2010. Toward discovery science of human brain function. Proceedings of the
National Academy of Sciences of the United States of America 107, 4734-4739.
Boedecker, J., Obst, O., Lizier, J.T., Mayer, N.M., Asada, M., 2012. Information processing in echo
state networks at the edge of chaos. Theory in Biosciences 131, 205-213.
Bojak, I., Stoyanov, Z.V., Liley, D.T., 2015. Emergence of spatially heterogeneous burst suppression
in a neural field model of electrocortical activity. Frontiers in systems neuroscience 9.
Bonachela, J.A., De Franciscis, S., Torres, J.J., Munoz, M.A., 2010. Self-organization without
conservation: are neuronal avalanches generically critical? Journal of Statistical Mechanics:
Theory and Experiment 2010, P02015.
30
Criticality in the brain
Cocchi et al. 31
Breakspear, M., 2001. Perception of odors by a nonlinear model of the olfactory bulb. International
Journal of Neural Systems 11, 101-124.
Breakspear, M., Roberts, G., Green, M.J., Nguyen, V.T., Frankland, A., Levy, F., Lenroot, R.,
Mitchell, P.B., 2015. Network dysfunction of emotional and cognitive processes in those at
genetic risk of bipolar disorder. Brain, awv261.
Breakspear, M., Roberts, J., Terry, J.R., Rodrigues, S., Mahant, N., Robinson, P., 2006. A unifying
explanation of primary generalized seizures through nonlinear brain modeling and bifurcation
analysis. Cerebral Cortex 16, 1296-1313.
Bressler, S.L., Kelso, J.S., 2001. Cortical coordination dynamics and cognition. Trends in Cognitive
Sciences 5, 26-36.
Brochini, L., de Andrade Costa, A., Abadi, M., Roque, A.C., Stolfi, J., Kinouchi, O., 2016. Phase
transitions and self-organized criticality in networks of stochastic spiking neurons. Scientific
reports 6.
Burridge, R., Knopoff, L., 1967. Model and theoretical seismicity. Bull. Seismol. Soc. Am. 57, 3411–
3471.
Camalet, S., Duke, T., Jülicher, F., Prost, J., 2000. Auditory sensitivity provided by self-tuned critical
oscillations of hair cells. Proceedings of the National Academy of Sciences 97, 3183-3188.
Carlson, J., Langer, J., 1989. Mechanical model of an earthquake fault. Phys Rev A 40, 6470.
Cavagna, A., Cimarelli, A., Giardina, I., Parisi, G., Santagati, R., Stefanini, F., Viale, M., 2010. Scale-
free correlations in starling flocks. Proceedings of the National Academy of Sciences 107,
11865-11870.
Chapman, J.J., Roberts, J.A., Nguyen, V.T., Breakspear, M., 2016. Quantification of free-living
activity patterns using accelerometry in adults with mental illness. arXiv preprint
arXiv:1612.04471.
Chialvo, D.R., 2006. The brain near the edge. arXiv preprint q-bio/0610041.
Chialvo, D.R., 2010. Emergent complex neural dynamics. Nature Physics 6, 744-750.
Chialvo, D.R., 2012. Critical brain dynamics at large scale. arXiv preprint arXiv:1210.3632.
Ching, S., Cimenser, A., Purdon, P.L., Brown, E.N., Kopell, N.J., 2010. Thalamocortical model for a
propofol-induced α-rhythm associated with loss of consciousness. Proceedings of the
National Academy of Sciences 107, 22665-22670.
Churchland, M.M., Yu, B.M., Cunningham, J.P., Sugrue, L.P., Cohen, M.R., Corrado, G.S.,
Newsome, W.T., Clark, A.M., Hosseini, P., Scott, B.B., Bradley, D.C., Smith, M.A., Kohn,
A., Movshon, J.A., Armstrong, K.M., Moore, T., Chang, S.W., Snyder, L.H., Lisberger, S.G.,
Priebe, N.J., Finn, I.M., Ferster, D., Ryu, S.I., Santhanam, G., Sahani, M., Shenoy, K.V.,
2010. Stimulus onset quenches neural variability: a widespread cortical phenomenon. Nature
neuroscience 13, 369-378.
Clauset, A., Shalizi, C.R., Newman, M.E., 2009. Power-law distributions in empirical data. SIAM
review 51, 661-703.
Cocchi, L., Sale, M.V., L, L.G., Bell, P.T., Nguyen, V.T., Zalesky, A., Breakspear, M., Mattingley,
J.B., 2016. A hierarchy of timescales explains distinct effects of local inhibition of primary
visual cortex and frontal eye fields. Elife 5.
Cocchi, L., Zalesky, A., Fornito, A., Mattingley, J.B., 2013. Dynamic cooperation and competition
between brain systems during cognitive control. Trends Cogn Sci 17, 493-501.
Coombes, S., Terry, J.R., 2012. The dynamics of neurological disease: integrating computational,
experimental and clinical neuroscience. European Journal of Neuroscience 36, 2118-2120.
Copelli, M., 2014. Criticality at Work: How Do Critical Networks Respond to Stimuli? Criticality in
Neural Systems, 347-364.
Cote, P., Meisel, L., 1991. Self-organized criticality and the Barkhausen effect. Physical review letters
67, 1334.
Cowan, J.D., Neuman, J., van Drongelen, W., 2014. Self-organized criticality and near criticality in
neural networks. Criticality in Neural Systems, 465-484.
Crutchfield, J.P., Young, K. (1988) Computation at the onset of chaos. In: The Santa Fe Institute,
Westview. Citeseer.
31
Criticality in the brain
Cocchi et al. 32
Da Silva, F.L., Blanes, W., Kalitzin, S.N., Parra, J., Suffczynski, P., Velis, D.N., 2003. Epilepsies as
dynamical diseases of brain systems: basic models of the transition between normal and
epileptic activity. Epilepsia 44, 72-83.
Damoiseaux, J.S., Rombouts, S.A., Barkhof, F., Scheltens, P., Stam, C.J., Smith, S.M., Beckmann,
C.F., 2006. Consistent resting-state networks across healthy subjects. Proceedings of the
National Academy of Sciences of the United States of America 103, 13848-13853.
Daunizeau, J., Friston, K.J., Kiebel, S.J., 2009. Variational Bayesian identification and prediction of
stochastic nonlinear dynamic causal models. Physica D: Nonlinear Phenomena 238, 2089-
2118.
de Andrade Costa, A., Copelli, M., Kinouchi, O., 2015. Can dynamical synapses produce true self-
organized criticality? Journal of Statistical Mechanics: Theory and Experiment 2015,
P06004.
de Arcangelis, L. 2008. Activity-Dependent Model for Neuronal Avalanches. In: Aspects of Physical
Biology. pp. 215-230. Springer.
de Arcangelis, L., Perrone-Capano, C., Herrmann, H.J., 2006. Self-organized criticality model for
brain plasticity. Physical review letters 96, 028107.
De Carvalho, J.X., Prado, C.P., 2000. Self-organized criticality in the Olami-Feder-Christensen
model. Physical review letters 84, 4006.
Deco, G., Jirsa, V., McIntosh, A.R., Sporns, O., Kotter, R., 2009. Key role of coupling, delay, and
noise in resting brain fluctuations. Proceedings of the National Academy of Sciences of the
United States of America 106, 10302-10307.
Deco, G., Jirsa, V.K., 2012. Ongoing cortical activity at rest: criticality, multistability, and ghost
attractors. The Journal of neuroscience 32, 3366-3375.
Deco, G., Jirsa, V.K., McIntosh, A.R., 2011. Emerging concepts for the dynamical organization of
resting-state activity in the brain. Nature Reviews Neuroscience 12, 43-56.
Deco, G., Jirsa, V.K., McIntosh, A.R., 2013. Resting brains never rest: computational insights into
potential cognitive architectures. Trends in neurosciences 36, 268-274.
Destexhe, A., Sejnowski, T.J., 2001. Thalamocortical Assemblies: How Ion Channels, Single Neurons
and Large-Scale Networks Organize Sleep Oscillations.
Eguiluz, V.M., Chialvo, D.R., Cecchi, G.A., Baliki, M., Apkarian, A.V., 2005. Scale-free brain
functional networks. Physical review letters 94.
Eguíluz, V.M., Ospeck, M., Choe, Y., Hudspeth, A., Magnasco, M.O., 2000. Essential nonlinearities
in hearing. Physical review letters 84, 5232.
Fagerholm, E.D., Lorenz, R., Scott, G., Dinov, M., Hellyer, P.J., Mirzaei, N., Leeson, C., Carmichael,
D.W., Sharp, D.J., Shew, W.L., Leech, R., 2015. Cascades and cognitive state: focused
attention incurs subcritical dynamics. The Journal of neuroscience : the official journal of the
Society for Neuroscience 35, 4626-4634.
Fagerholm, E.D., Scott, G., Shew, W.L., Song, C., Leech, R., Knöpfel, T., Sharp, D.J., 2016. Cortical
Entropy, Mutual Information and Scale-Free Dynamics in Waking Mice. Cerebral Cortex.
Farmer, S., 2015. Comment on "Broadband Criticality of Human Brain Network Synchronization" by
Kitzbichler MG, Smith ML, Christensen SR, Bullmore E (2009) PLoS Comput Biol 5:
e1000314.
Feyerabend, P., 1993. Against method. Verso.
Field, D.J., 1987. Relations between the statistics of natural images and the response properties of
cortical cells. JOSA A 4, 2379-2394.
Fiser, J., Berkes, P., Orbán, G., Lengyel, M., 2010. Statistically optimal perception and learning: from
behavior to neural representations. Trends in cognitive sciences 14, 119-130.
Fornito, A., Zalesky, A., Breakspear, M., 2015. The connectomics of brain disorders. Nature reviews.
Neuroscience 16, 159-172.
Fox, M.D., Snyder, A.Z., Vincent, J.L., Corbetta, M., Van Essen, D.C., Raichle, M.E., 2005. The
human brain is intrinsically organized into dynamic, anticorrelated functional networks.
Proceedings of the National Academy of Sciences of the United States of America 102, 9673-
9678.
Freeman, W.J., 1987. Simulation of chaotic EEG patterns with a dynamic model of the olfactory
system. Biological cybernetics 56, 139-150.
32
Criticality in the brain
Cocchi et al. 33
Freeman, W.J., 1991. The physiology of perception. Scientific American 264, 78-85.
Freestone, D.R., Long, S.N., Frey, S., Stypulkowski, P.H., Giftakis, J.E., Cook, M.J. (2013) A method
for actively tracking excitability of brain networks using a fully implantable monitoring
system. In: Engineering in Medicine and Biology Society (EMBC), 2013 35th Annual
International Conference of the IEEE. pp. 6151-6154. IEEE.
Freyer, F., Aquino, K., Robinson, P.A., Ritter, P., Breakspear, M., 2009. Bistability and non-Gaussian
fluctuations in spontaneous cortical activity. The Journal of neuroscience : the official journal
of the Society for Neuroscience 29, 8512-8524.
Freyer, F., Roberts, J.A., Becker, R., Robinson, P.A., Ritter, P., Breakspear, M., 2011. Biophysical
mechanisms of multistability in resting-state cortical rhythms. The Journal of neuroscience :
the official journal of the Society for Neuroscience 31, 6353-6361.
Freyer, F., Roberts, J.A., Ritter, P., Breakspear, M., 2012. A canonical model of multistability and
scale-invariance in biological systems. PLoS computational biology 8, e1002634.
Friedman, N., Ito, S., Brinkman, B.A., Shimono, M., DeVille, R.L., Dahmen, K.A., Beggs, J.M.,
Butler, T.C., 2012. Universal critical dynamics in high resolution neuronal avalanche data.
Physical review letters 108, 208102.
Friston, K., Adams, R., Perrinet, L., Breakspear, M., 2012a. Perceptions as hypotheses: saccades as
experiments. Frontiers in psychology 3, 151.
Friston, K., Breakspear, M., Deco, G., 2012b. Perception and self-organized instability. Front Comput
Neurosci 6.
Friston, K.J., 2000. The labile brain. II. Transients, complexity and selection. Philosophical
Transactions of the Royal Society of London B: Biological Sciences 355, 237-252.
Fröhlich, F., Sejnowski, T.J., Bazhenov, M., 2010. Network bistability mediates spontaneous
transitions between normal and pathological brain states. The Journal of neuroscience 30,
10734-10743.
Gal, A., Marom, S., 2013. Self-organized criticality in single-neuron excitability. Physical Review E
88.
Garrett, D.D., Kovacevic, N., McIntosh, A.R., Grady, C.L., 2011. The importance of being variable.
The Journal of neuroscience 31, 4496-4503.
Gatlin, L.L., 1972. Information theory and the living system.
Gautam, S.H., Hoang, T.T., McClanahan, K., Grady, S.K., Shew, W.L., 2015. Maximizing sensory
dynamic range by tuning the cortical state to criticality. PLoS computational biology 11,
e1004576.
Gireesh, E.D., Plenz, D., 2008. Neuronal avalanches organize as nested theta-and beta/gamma-
oscillations during development of cortical layer 2/3. Proceedings of the National Academy of
Sciences 105, 7576-7581.
Gollo, L.L., 2017. Coexistence of critical sensitivity and subcritical specificity can yield optimal
population coding. arXiv preprint arXiv:1707.04484.
Gollo, L.L., Breakspear, M., 2014. The frustrated brain: from dynamics on motifs to communities and
networks. Philosophical transactions of the Royal Society of London. Series B, Biological
sciences 369.
Gollo, L.L., Copelli, M., Roberts, J.A., 2016. Diversity improves performance in excitable networks.
PeerJ 4, e1912.
Gollo, L.L., Kinouchi, O., Copelli, M., 2013. Single-neuron criticality optimizes analog dendritic
computation. Scientific reports 3.
Gollo, L.L., Mirasso, C., Eguíluz, V.M., 2012. Signal integration enhances the dynamic range in
neuronal systems. Physical Review E 85, 040902.
Gollo, L.L., Roberts, J.A., Cocchi, L., 2017. Mapping how local perturbations influence systems-level
brain dynamics. Neuroimage.
Gollo, L.L., Zalesky, A., Hutchison, R.M., van den Heuvel, M., Breakspear, M., 2015. Dwelling
quietly in the rich club: brain network determinants of slow cortical fluctuations.
Philosophical transactions of the Royal Society of London. Series B, Biological sciences 370.
Golubitsky, M., Stewart, I., Buono, P.L., Collins, J.J., 1999. Symmetry in locomotor central pattern
generators and animal gaits. Nature 401, 693-695.
33
Criticality in the brain
Cocchi et al. 34
Greenfield, E., Lecar, H., 2001. Mutual information in a dilute, asymmetric neural network model.
Physical Review E 63, 041905.
Greicius, M.D., Krasnow, B., Reiss, A.L., Menon, V., 2003. Functional connectivity in the resting
brain: a network analysis of the default mode hypothesis. Proceedings of the National
Academy of Sciences of the United States of America 100, 253-258.
Haimovici, A., Tagliazucchi, E., Balenzuela, P., Chialvo, D.R., 2013. Brain organization into resting
state networks emerges at criticality on a model of the human connectome. Physical review
letters 110, 178101.
Haken, H., Kelso, J.S., Bunz, H., 1985. A theoretical model of phase transitions in human hand
movements. Biological cybernetics 51, 347-356.
Haldeman, C., Beggs, J.M., 2005. Critical branching captures activity in living neural networks and
maximizes the number of metastable states. Physical review letters 94, 058101.
Harrison, B.J., Yucel, M., Pujol, J., Pantelis, C., 2007. Task-induced deactivation of midline cortical
regions in schizophrenia assessed with fMRI. Schizophrenia research 91, 82-86.
Hasson, U., Chen, J., Honey, C.J., 2015. Hierarchical process memory: memory as an integral
component of information processing. Trends Cogn Sci 19, 304-313.
Hasson, U., Yang, E., Vallines, I., Heeger, D.J., Rubin, N., 2008. A hierarchy of temporal receptive
windows in human cortex. The Journal of neuroscience : the official journal of the Society for
Neuroscience 28, 2539-2550.
He, B.J., 2013. Scale-free dynamics and critical phenomena in cortical activity. Frontiers E-books.
He, B.J., Zempel, J.M., Snyder, A.Z., Raichle, M.E., 2010. The temporal structures and functional
significance of scale-free brain activity. Neuron 66, 353-369.
Hearne, L., Cocchi, L., Zalesky, A., Mattingley, J.B., 2015. Interactions between default mode and
control networks as a function of increasing cognitive reasoning complexity. Human brain
mapping.
Heathcote, A., Brown, S., Mewhort, D., 2000. The power law repealed: The case for an exponential
law of practice. Psychonomic bulletin & review 7, 185-207.
Heathcote, A., Popiel, S.J., Mewhort, D., 1991. Analysis of response time distributions: An example
using the Stroop task. Psychological Bulletin 109, 340.
Hellyer, P.J., Shanahan, M., Scott, G., Wise, R.J., Sharp, D.J., Leech, R., 2014. The control of global
brain dynamics: opposing actions of frontoparietal control and default mode networks on
attention. The Journal of neuroscience : the official journal of the Society for Neuroscience
34, 451-461.
Hesse, J., Gross, T., 2015. Self-organized criticality as a fundamental property of neural systems.
Criticality as a signature of healthy neural systems: multi-scale experimental and
computational studies.
Hidalgo, J., Grilli, J., Suweis, S., Muñoz, M.A., Banavar, J.R., Maritan, A., 2014. Information-based
fitness and the emergence of criticality in living systems. Proceedings of the National
Academy of Sciences 111, 10095-10100.
Honey, C.J., Kotter, R., Breakspear, M., Sporns, O., 2007. Network structure of cerebral cortex
shapes functional connectivity on multiple time scales. Proceedings of the National Academy
of Sciences of the United States of America 104, 10240-10245.
Honey, C.J., Thesen, T., Donner, T.H., Silbert, L.J., Carlson, C.E., Devinsky, O., Doyle, W.K., Rubin,
N., Heeger, D.J., Hasson, U., 2012. Slow cortical dynamics and the accumulation of
information over long timescales. Neuron 76, 423-434.
Houle, P.A., Sethna, J.P., 1996. Acoustic emission from crumpling paper. Physical Review E 54, 278.
Hudspeth, A., 2008. Making an effort to listen: mechanical amplification in the ear. Neuron 59, 530-
545.
Hyett, M.P., Breakspear, M.J., Friston, K.J., Guo, C.C., Parker, G.B., 2015a. Disrupted effective
connectivity of cortical systems supporting attention and interoception in melancholia. JAMA
psychiatry 72, 350-358.
Hyett, M.P., Parker, G.B., Guo, C.C., Zalesky, A., Nguyen, V.T., Yuen, T., Breakspear, M., 2015b.
Scene unseen: Disrupted neuronal adaptation in melancholia during emotional film viewing.
NeuroImage: Clinical 9, 660-667.
34
Criticality in the brain
Cocchi et al. 35
Iyer, K.K., Roberts, J.A., Hellstrom-Westas, L., Wikstrom, S., Hansen Pupp, I., Ley, D., Breakspear,
M., Vanhatalo, S., 2015a. Early Detection of Preterm Intraventricular Hemorrhage From
Clinical Electroencephalography. Crit Care Med 43, 2219-2227.
Iyer, K.K., Roberts, J.A., Hellstrom-Westas, L., Wikstrom, S., Hansen Pupp, I., Ley, D., Vanhatalo,
S., Breakspear, M., 2015b. Cortical burst dynamics predict clinical outcome early in
extremely preterm infants. Brain 138, 2206-2218.
Iyer, K.K., Roberts, J.A., Metsaranta, M., Finnigan, S., Breakspear, M., Vanhatalo, S., 2014. Novel
features of early burst suppression predict outcome after birth asphyxia. Annals of clinical
and translational neurology 1, 209-214.
Jirsa, V.K., Friedrich, R., Haken, H., Kelso, J.S., 1994. A theoretical model of phase transitions in the
human brain. Biological cybernetics 71, 27-35.
Jirsa, V.K., Proix, T., Perdikis, D., Woodman, M.M., Wang, H., Gonzalez-Martinez, J., Bernard, C.,
Bénar, C., Guye, M., Chauvel, P., 2017. The virtual epileptic patient: individualized whole-
brain models of epilepsy spread. Neuroimage 145, 377-388.
Jirsa, V.K., Stacey, W.C., Quilichini, P.P., Ivanov, A.I., Bernard, C., 2014. On the nature of seizure
dynamics. Brain 137, 2210-2230.
Kastner, D.B., Baccus, S.A., Sharpee, T.O., 2015. Critical and maximally informative encoding
between neural populations in the retina. Proceedings of the National Academy of Sciences
112, 2533-2538.
Kauffman, S.A., Johnsen, S., 1991. Coevolution to the edge of chaos: Coupled fitness landscapes,
poised states, and coevolutionary avalanches. Journal of theoretical biology 149, 467-505.
Kello, C.T., Brown, G.D., Ferrer-i-Cancho, R., Holden, J.G., Linkenkaer-Hansen, K., Rhodes, T., Van
Orden, G.C., 2010. Scaling laws in cognitive sciences. Trends in cognitive sciences 14, 223-
232.
Kelso, J., 1984. Phase transitions and critical behavior in human bimanual coordination. American
Journal of Physiology-Regulatory, Integrative and Comparative Physiology 246, R1000-
R1004.
Kelso, J., 2014. The dynamic brain in action: Coordinative structures, criticality, and coordination
dynamics. Criticality in Neural Systems, 67-104.
Kelso, J., Bressler, S., Buchanan, S., DeGuzman, G., Ding, M., Fuchs, A., Holroyd, T., 1992. A phase
transition in human brain and behavior. Physics Letters A 169, 134-144.
Kelso, J., Scholz, J., Schöner, G., 1986. Nonequilibrium phase transitions in coordinated biological
motion: critical fluctuations. Physics Letters A 118, 279-284.
Kelso, J.S., Clark, J.E., 1982. The development of movement control and coordination. John Wiley &
Sons.
Kim, S., Park, S.H., Ryu, C.S., 1997. Noise-enhanced multistability in coupled oscillator systems.
Physical review letters 78, 1616.
Kinouchi, O., Copelli, M., 2006. Optimal dynamical range of excitable networks at criticality. Nature
physics 2, 348-351.
Kitzbichler, M.G., Smith, M.L., Christensen, S.R., Bullmore, E., 2009. Broadband criticality of
human brain network synchronization. PLoS computational biology 5, e1000314.
Kosterlitz, J.M., Thouless, D.J., 1973. Ordering, metastability and phase transitions in two-
dimensional systems. Journal of Physics C: Solid State Physics 6, 1181.
Kramer, E.M., Lobkovsky, A.E., 1996. Universal power law in the noise from a crumpled elastic
sheet. Physical Review E 53, 1465.
Langlois, D., Cousineau, D., Thivierge, J.-P., 2014. Maximum likelihood estimators for truncated and
censored power-law distributions show how neuronal avalanches may be misevaluated.
Physical Review E 89, 012709.
Langton, C.G., 1990. Computation at the edge of chaos: phase transitions and emergent computation.
Physica D: Nonlinear Phenomena 42, 12-37.
Larremore, D.B., Shew, W.L., Ott, E., Sorrentino, F., Restrepo, J.G., 2014. Inhibition causes ceaseless
dynamics in networks of excitable nodes. Physical review letters 112, 138103.
Larremore, D.B., Shew, W.L., Restrepo, J.G., 2011. Predicting criticality and dynamic range in
complex networks: effects of topology. Physical review letters 106, 058101.
35
Criticality in the brain
Cocchi et al. 36
Leech, R., Braga, R., Sharp, D.J., 2012. Echoes of the brain within the posterior cingulate cortex. The
Journal of neuroscience : the official journal of the Society for Neuroscience 32, 215-222.
Legenstein, R., Maass, W., 2007. Edge of chaos and prediction of computational performance for
neural circuit models. Neural Networks 20, 323-334.
Levina, A., Herrmann, J.M., Geisel, T., 2014. Theoretical Neuroscience of Self‐Organized Criticality:
From Formal Approaches to Realistic Models. Criticality in Neural Systems, 417-436.
Linkenkaer-Hansen, K., Nikouline, V.V., Palva, J.M., Ilmoniemi, R.J., 2001. Long-range temporal
correlations and scaling behavior in human brain oscillations. The Journal of neuroscience :
the official journal of the Society for Neuroscience 21, 1370-1377.
Livi, L., Bianchi, F.M., Alippi, C., 2016. Determination of the edge of criticality in echo state
networks through Fisher information maximization. arXiv preprint arXiv:1603.03685.
Loh, M., Rolls, E.T., Deco, G., 2007. A dynamical systems hypothesis of schizophrenia.
Lu, E.T., Hamilton, R.J., McTiernan, J., Bromund, K.R., 1993. Solar flares and avalanches in driven
dissipative systems. The Astrophysical Journal 412, 841-852.
Magnasco, M.O., Piro, O., Cecchi, G.A., 2009. Self-tuned critical anti-Hebbian networks. Physical
review letters 102, 258102.
Maoiléidigh, D.Ó., Nicola, E.M., Hudspeth, A., 2012. The diverse effects of mechanical loading on
active hair bundles. Proceedings of the National Academy of Sciences 109, 1943-1948.
Marković, D., Gros, C., 2014. Power laws and self-organized criticality in theory and nature. Physics
Reports 536, 41-74.
Markram, H., Muller, E., Ramaswamy, S., Reimann, M.W., Abdellah, M., Sanchez, C.A., Ailamaki,
A., Alonso-Nanclares, L., Antille, N., Arsever, S., Kahou, G.A., Berger, T.K., Bilgili, A.,
Buncic, N., Chalimourda, A., Chindemi, G., Courcol, J.D., Delalondre, F., Delattre, V.,
Druckmann, S., Dumusc, R., Dynes, J., Eilemann, S., Gal, E., Gevaert, M.E., Ghobril, J.P.,
Gidon, A., Graham, J.W., Gupta, A., Haenel, V., Hay, E., Heinis, T., Hernando, J.B., Hines,
M., Kanari, L., Keller, D., Kenyon, J., Khazen, G., Kim, Y., King, J.G., Kisvarday, Z.,
Kumbhar, P., Lasserre, S., Le Be, J.V., Magalhaes, B.R., Merchan-Perez, A., Meystre, J.,
Morrice, B.R., Muller, J., Munoz-Cespedes, A., Muralidhar, S., Muthurasa, K., Nachbaur, D.,
Newton, T.H., Nolte, M., Ovcharenko, A., Palacios, J., Pastor, L., Perin, R., Ranjan, R.,
Riachi, I., Rodriguez, J.R., Riquelme, J.L., Rossert, C., Sfyrakis, K., Shi, Y., Shillcock, J.C.,
Silberberg, G., Silva, R., Tauheed, F., Telefont, M., Toledo-Rodriguez, M., Trankler, T., Van
Geit, W., Diaz, J.V., Walker, R., Wang, Y., Zaninetta, S.M., DeFelipe, J., Hill, S.L., Segev, I.,
Schurmann, F., 2015. Reconstruction and Simulation of Neocortical Microcircuitry. Cell 163,
456-492.
McClure Jr, J.C., Schroder, K., 1976. The magnetic Barkhausen effect. Critical Reviews in Solid State
and Material Sciences 6, 45-83.
McIntosh, A.R., Kovacevic, N., Lippe, S., Garrett, D., Grady, C., Jirsa, V., 2010. The development of
a noisy brain. Arch Ital Biol 148, 323-337.
Meisel, C., Storch, A., Hallmeyer-Elgner, S., Bullmore, E., Gross, T., 2012. Failure of adaptive self-
organized criticality during epileptic seizure attacks. PLoS computational biology 8,
e1002312.
Meisel, L., Cote, P., 1992. Power laws, flicker noise, and the Barkhausen effect. Physical Review B
46, 10822.
Melbourne, I., Chossat, P., Golubitsky, M., 1989. Heteroclinic cycles involving periodic solutions in
mode interactions with O (2) symmetry. Proceedings of the Royal Society of Edinburgh:
Section A Mathematics 113, 315-345.
Meyer-Lindenberg, A., Poline, J.B., Kohn, P.D., Holt, J.L., Egan, M.F., Weinberger, D.R., Berman,
K.F., 2001. Evidence for abnormal cortical functional connectivity during working memory in
schizophrenia. The American journal of psychiatry 158, 1809-1817.
Mihalas, S., Millman, D., Iyer, R., Kirkwood, A., Niebur, E., 2014. Nonconservative Neuronal
Networks During Up‐States Self‐Organize Near Critical Points. Criticality in Neural Systems,
437-464.
Miller, K.J., Sorensen, L.B., Ojemann, J.G., den Nijs, M., 2009. Power-law scaling in the brain
surface electric potential. PLoS computational biology 5, e1000609.
36
Criticality in the brain
Cocchi et al. 37
Millman, D., Mihalas, S., Kirkwood, A., Niebur, E., 2010. Self-organized criticality occurs in non-
conservative neuronal networks during/up/'states. Nature physics 6, 801-805.
Misic, B., Betzel, R.F., Nematzadeh, A., Goni, J., Griffa, A., Hagmann, P., Flammini, A., Ahn, Y.Y.,
Sporns, O., 2015. Cooperative and Competitive Spreading Dynamics on the Human
Connectome. Neuron 86, 1518-1529.
Mitchell, M., Hraber, P., Crutchfield, J.P., 1993. Revisiting the edge of chaos: Evolving cellular
automata to perform computations. arXiv preprint adap-org/9303003.
Mora, T., Bialek, W., 2011. Are biological systems poised at criticality? Journal of Statistical Physics
144, 268-302.
Moretti, P., Muñoz, M.A., 2013. Griffiths phases and the stretching of criticality in brain networks.
Nature communications 4.
Mosqueiro, T.S., Maia, L.P., 2013. Optimal channel efficiency in a sensory network. Physical Review
E 88, 012712.
Murray, J.D., Bernacchia, A., Freedman, D.J., Romo, R., Wallis, J.D., Cai, X., Padoa-Schioppa, C.,
Pasternak, T., Seo, H., Lee, D., Wang, X.J., 2014. A hierarchy of intrinsic timescales across
primate cortex. Nature neuroscience 17, 1661-1663.
Nagel, K., Herrmann, H.J., 1993. Deterministic models for traffic jams. Physica A: Statistical
Mechanics and its Applications 199, 254-269.
Nakamura, T., Kiyono, K., Yoshiuchi, K., Nakahara, R., Struzik, Z.R., Yamamoto, Y., 2007.
Universal scaling law in human behavioral organization. Physical review letters 99, 138103.
Nakamura, T., Takumi, T., Takano, A., Aoyagi, N., Yoshiuchi, K., Struzik, Z.R., Yamamoto, Y.,
2008. Of mice and men--universality and breakdown of behavioral organization. PLoS ONE
3, e2050.
Nejad, A.B., Ebdrup, B.H., Siebner, H.R., Rasmussen, H., Aggernaes, B., Glenthoj, B.Y., Baare,
W.F., 2011. Impaired
in
antipsychotic-naive patients with first-episode schizophrenia. The world journal of biological
psychiatry : the official journal of the World Federation of Societies of Biological Psychiatry
12, 271-281.
temporoparietal deactivation with working memory
load
Nelson, T.S., Suhr, C.L., Freestone, D.R., Lai, A., Halliday, A.J., Mclean, K.J., Burkitt, A.N., Cook,
M.J., 2011. Closed-loop seizure control with very high frequency electrical stimulation at
seizure onset in the GAERS model of absence epilepsy. International journal of neural
systems 21, 163-173.
Nevado-Holgado, A.J., Marten, F., Richardson, M.P., Terry, J.R., 2012. Characterising the dynamics
of EEG waveforms as the path through parameter space of a neural mass model: application
to epilepsy seizure evolution. Neuroimage 59, 2374-2392.
Niedermeyer, E., Sherman, D.L., Geocadin, R.J., Hansen, H.C., Hanley, D.F., 1999. The burst-
suppression electroencephalogram. Clinical EEG (electroencephalography) 30, 99-105.
Nielson, D.M., Smith, T.A., Sreekumar, V., Dennis, S., Sederberg, P.B., 2015. Human hippocampus
represents space and time during retrieval of real-world memories. Proceedings of the
National Academy of Sciences 112, 11078-11083.
Nykter, M., Price, N.D., Larjo, A., Aho, T., Kauffman, S.A., Yli-Harja, O., Shmulevich, I., 2008.
Critical networks exhibit maximal information diversity in structure-dynamics relationships.
Physical review letters 100, 058702.
Palva, J.M., Palva, S., 2011. 22 Roles of multiscale brain activity fluctuations in shaping the
variability and dynamics of psychophysical performance. Progress in brain research 193,
335.
Palva, J.M., Zhigalov, A., Hirvonen, J., Korhonen, O., Linkenkaer-Hansen, K., Palva, S., 2013.
Neuronal long-range temporal correlations and avalanche dynamics are correlated with
behavioral scaling laws. Proceedings of the National Academy of Sciences 110, 3585-3590.
Papanikolaou, S., Bohn, F., Sommer, R.L., Durin, G., Zapperi, S., Sethna, J.P., 2011. Universality
beyond power laws and the average avalanche shape. Nature Physics 7, 316-320.
Penny, W.D., 2012. Comparing dynamic causal models using AIC, BIC and free energy. Neuroimage
59, 319-330.
Perković, O., Dahmen, K., Sethna, J.P., 1995. Avalanches, Barkhausen noise, and plain old criticality.
Physical review letters 75, 4528.
37
Criticality in the brain
Cocchi et al. 38
Petermann, T., Thiagarajan, T.C., Lebedev, M.A., Nicolelis, M.A., Chialvo, D.R., Plenz, D., 2009.
Spontaneous cortical activity in awake monkeys composed of neuronal avalanches.
Proceedings of the National Academy of Sciences of the United States of America 106, 15921-
15926.
Petzschner, F.H., Weber, L.A., Gard, T., Stephan, K.E., 2017. Computational Psychosomatics and
Computational Psychiatry: Towards a joint framework for differential diagnosis. Biological
Psychiatry.
Plenz, D., Thiagarajan, T.C., 2007. The organizing principles of neuronal avalanches: cell assemblies
in the cortex? Trends in neurosciences 30, 101-110.
Ponce-Alvarez, A., He, B.J., Hagmann, P., Deco, G., 2015. Task-Driven Activity Reduces the
Cortical Activity Space of the Brain: Experiment and Whole-Brain Modeling. PLoS
computational biology 11, e1004445.
Priesemann, V., Valderrama, M., Wibral, M., Le Van Quyen, M., 2013. Neuronal avalanches differ
from wakefulness to deep sleep--evidence from intracranial depth recordings in humans.
PLoS computational biology 9, e1002985.
Priesemann, V., Wibral, M., Valderrama, M., Propper, R., Le Van Quyen, M., Geisel, T., Triesch, J.,
Nikolic, D., Munk, M.H., 2014. Spike avalanches in vivo suggest a driven, slightly subcritical
brain state. Frontiers in systems neuroscience 8, 108.
Publio, R., Ceballos, C.C., Roque, A.C., 2012. Dynamic range of vertebrate retina ganglion cells:
Importance of active dendrites and coupling by electrical synapses. PloS one 7, e48517.
Rabinovich, M., Volkovskii, A., Lecanda, P., Huerta, R., Abarbanel, H., Laurent, G., 2001.
Dynamical encoding by networks of competing neuron groups: winnerless competition.
Physical review letters 87, 068102.
Rabinovich, M.I., Huerta, R., Varona, P., Afraimovich, V.S., 2008. Transient cognitive dynamics,
metastability, and decision making. PLoS computational biology 4, e1000072.
Razi, A., Kahan, J., Rees, G., Friston, K.J., 2015. Construct validation of a DCM for resting state
fMRI. Neuroimage 106, 1-14.
Ribeiro, T.L., Copelli, M., Caixeta, F., Belchior, H., Chialvo, D.R., Nicolelis, M.A., Ribeiro, S., 2010.
Spike avalanches exhibit universal dynamics across the sleep-wake cycle. PLoS ONE 5,
e14129.
Rice, J., Ruina, A., 1983. Stability of steady frictional slipping. Journal of applied mechanics 50, 343-
349.
Roberts, D.C., Turcotte, D.L., 1998. Fractality and self-organized criticality of wars. Fractals 6, 351-
357.
Roberts, J.A., Boonstra, T.W., Breakspear, M., 2015. The heavy tail of the human brain. Current
opinion in neurobiology 31, 164-172.
Roberts, J.A., Friston, K.J., Breakspear, M., 2017. Clinical applications of stochastic dynamic models
of the brain, part II: A review. Biological Psychiatry: Cognitive Neuroscience and
Neuroimaging.
Roberts, J.A., Iyer, K.K., Finnigan, S., Vanhatalo, S., Breakspear, M., 2014a. Scale-free bursting in
human cortex following hypoxia at birth. The Journal of Neuroscience 34, 6557-6572.
Roberts, J.A., Iyer, K.K., Vanhatalo, S., Breakspear, M., 2014b. Critical role for resource constraints
in neural models. Frontiers in systems neuroscience 8, 154.
Roberts, J.A., Perry, A., Lord, A.R., Roberts, G., Mitchell, P.B., Smith, R.E., Calamante, F.,
Breakspear, M., 2016. The contribution of geometry to the human connectome. Neuroimage
124, 379-393.
Roberts, J.A., Wallis, G., Breakspear, M., 2013. Fixational eye movements during viewing of
dynamic natural scenes. Frontiers in psychology 4, 797.
Robinson, P., Rennie, C., Rowe, D., 2002. Dynamics of large-scale brain activity in normal arousal
states and epileptic seizures. Physical Review E 65, 041924.
Romani, G.L., Williamson, S.J., Kaufman, L., 1982. Tonotopic organization of the human auditory
cortex. Science 216, 1339-1340.
Rubinov, M., Sporns, O., Thivierge, J.P., Breakspear, M., 2011. Neurobiologically realistic
determinants of self-organized criticality in networks of spiking neurons. PLoS computational
biology 7, e1002038.
38
Criticality in the brain
Cocchi et al. 39
Rubinov, M., Sporns, O., van Leeuwen, C., Breakspear, M., 2009. Symbiotic relationship between
brain structure and dynamics. BMC neuroscience 10, 55.
Ruderman, D.L., Bialek, W., 1994. Statistics of natural images: Scaling in the woods. Physical review
letters 73, 814-817.
Scarpetta, S., de Candia, A., 2013. Neural avalanches at the critical point between replay and non-
replay of spatiotemporal patterns. PLoS One 8, e64162.
Scheffer, M., Bascompte, J., Brock, W.A., Brovkin, V., Carpenter, S.R., Dakos, V., Held, H., Van
Nes, E.H., Rietkerk, M., Sugihara, G., 2009. Early-warning signals for critical transitions.
Nature 461, 53-59.
Schrauwen, B., Büsing, L., Legenstein, R. (2009) On computational power and the order-chaos phase
transition in reservoir computing. In: 22nd Annual conference on Neural Information
Processing Systems (NIPS 2008). pp. 1425-1432. NIPS Foundation.
Schuster, H.G., Plenz, D., Niebur, E., 2014. Criticality in neural systems. John Wiley & Sons.
Scott, G., Fagerholm, E.D., Mutoh, H., Leech, R., Sharp, D.J., Shew, W.L., Knöpfel, T., 2014.
Voltage imaging of waking mouse cortex reveals emergence of critical neuronal dynamics.
The Journal of Neuroscience 34, 16611-16620.
Sethna, J.P., Dahmen, K.A., Myers, C.R., 2001. Crackling noise. Nature 410, 242-250.
Shew, W., 2015. Neuronal Avalanches. Encyclopedia of Computational Neuroscience, 2018-2024.
Shew, W.L., Clawson, W.P., Pobst, J., Karimipanah, Y., Wright, N.C., Wessel, R., 2015. Adaptation
to sensory input tunes visual cortex to criticality. Nature Physics 11, 659-663.
Shew, W.L., Plenz, D., 2013. The functional benefits of criticality in the cortex. The neuroscientist 19,
88-100.
Shew, W.L., Yang, H., Yu, S., Roy, R., Plenz, D., 2011. Information capacity and transmission are
maximized in balanced cortical networks with neuronal avalanches. Journal of neuroscience
31, 55-63.
Shew, W.L., Yang, H.D., Petermann, T., Roy, R., Plenz, D., 2009. Neuronal Avalanches Imply
Maximum Dynamic Range in Cortical Networks at Criticality. Journal of Neuroscience 29,
15595-15600.
Shih, H.-Y., Hsieh, T.-L., Goldenfeld, N., 2015. Ecological collapse and the emergence of travelling
waves at the onset of shear turbulence. Nature Physics.
Shriki, O., Alstott, J., Carver, F., Holroyd, T., Henson, R.N., Smith, M.L., Coppola, R., Bullmore, E.,
Plenz, D., 2013. Neuronal avalanches in the resting MEG of the human brain. Journal of
Neuroscience 33, 7079-7090.
Sreekumar, V., Dennis, S., Doxas, I., 2016. The Episodic Nature of Experience: A Dynamical
Systems Analysis. Cognitive Science.
Sreekumar, V., Dennis, S., Doxas, I., Zhuang, Y., Belkin, M., 2014. The geometry and dynamics of
lifelogs: discovering the organizational principles of human experience. PloS one 9, e97166.
Stam, C.J., de Bruin, E.A., 2004. Scale-free dynamics of global functional connectivity in the human
brain. Human brain mapping 22, 97-109.
Stanley, H.E., 1987. Introduction to phase transitions and critical phenomena. Introduction to Phase
Transitions and Critical Phenomena, by H Eugene Stanley, pp. 336. Foreword by H Eugene
Stanley. Oxford University Press, Jul 1987. ISBN-10: 0195053168. ISBN-13: 9780195053166
1.
Stanley, H.E., 1999. Scaling, universality, and renormalization: Three pillars of modern critical
phenomena. Reviews of modern physics 71, S358.
Stepp, N., Plenz, D., Srinivasa, N., 2015. Synaptic plasticity enables adaptive self-tuning critical
networks. PLoS computational biology 11, e1004043.
Stevens, S.S., 1975. Psychophysics. Transaction Publishers.
Stewart, C.V., Plenz, D., 2006. Inverted-U profile of dopamine–NMDA-mediated spontaneous
avalanche recurrence in superficial layers of rat prefrontal cortex. The Journal of
neuroscience 26, 8148-8159.
Stramaglia, S., Pellicoro, M., Angelini, L., Amico, E., Aerts, H., Cortés, J., Laureys, S., Marinazzo,
the Human Connectome. arXiv preprint
Ising Model on
D., 2015. Conserved
arXiv:1509.02697.
39
Criticality in the brain
Cocchi et al. 40
Suffczynski, P., Kalitzin, S., Da Silva, F.L., 2004. Dynamics of non-convulsive epileptic phenomena
modeled by a bistable neuronal network. Neuroscience 126, 467-484.
Tagliazucchi, E., Balenzuela, P., Fraiman, D., Chialvo, D.R., 2012. Criticality in large-scale brain
FMRI dynamics unveiled by a novel point process analysis. Frontiers in physiology 3, 15.
Tagliazucchi, E., Chialvo, D.R., Siniatchkin, M., Amico, E., Brichant, J.-F., Bonhomme, V.,
Noirhomme, Q., Laufs, H., Laureys, S., 2016. Large-scale signatures of unconsciousness are
consistent with a departure from critical dynamics. Journal of The Royal Society Interface 13,
20151027.
Thornton, T.L., Gilden, D.L., 2005. Provenance of correlations in psychological data. Psychonomic
Bulletin & Review 12, 409-441.
Tognoli, E., Kelso, J.A., 2014. The metastable brain. Neuron 81, 35-48.
Tononi, G., Sporns, O., Edelman, G.M., 1994. A measure for brain complexity: relating functional
segregation and integration in the nervous system. Proceedings of the National Academy of
Sciences 91, 5033-5037.
Touboul, J., Destexhe, A., 2010. Can power-law scaling and neuronal avalanches arise from
stochastic dynamics. PLoS ONE 5, e8982.
Touboul, J., Destexhe, A., 2015. Power-law statistics and universal scaling in the absence of
criticality. arXiv preprint arXiv:1503.08033.
Van de Ville, D., Britz, J., Michel, C.M., 2010. EEG microstate sequences in healthy humans at rest
reveal scale-free dynamics. Proceedings of the National Academy of Sciences of the United
States of America 107, 18179-18184.
Van Kessenich, L.M., de Arcangelis, L., Herrmann, H., 2016. Synaptic plasticity and neuronal
refractory time cause scaling behaviour of neuronal avalanches. Scientific reports 6, 32071.
Van Orden, G.C., Holden, J.G., Turvey, M.T., 2005. Human cognition and 1/f scaling. Journal of
Experimental Psychology: General 134, 117.
Vano, J., Wildenberg, J., Anderson, M., Noel, J., Sprott, J., 2006. Chaos in low-dimensional Lotka–
Volterra models of competition. Nonlinearity 19, 2391.
Vasa, F., Shanahan, M., Hellyer, P.J., Scott, G., Cabral, J., Leech, R., 2015. Effects of lesions on
synchrony and metastability in cortical networks. Neuroimage.
Virkar, Y.S., Shew, W.L., Restrepo, J.G., Ott, E., 2016. Feedback control stabilization of critical
dynamics via resource transport on multilayer networks: How glia enable learning dynamics
in the brain. Physical Review E 94, 042310.
Vojta, T., Avella, A., Mancini, F. (2013) Phases and phase transitions in disordered quantum systems.
In: AIP Conference Proceedings. pp. 188-247. AIP.
Vuong, Q.H., 1989. Likelihood ratio tests for model selection and non-nested hypotheses.
Econometrica: Journal of the Econometric Society, 307-333.
Wagenmakers, E.-J., Farrell, S., Ratcliff, R., 2004. Estimation and interpretation of 1/fα noise in
human cognition. Psychonomic Bulletin & Review 11, 579-615.
Wendling, F., Bartolomei, F., Bellanger, J., Chauvel, P., 2002. Epileptic fast activity can be explained
by a model of impaired GABAergic dendritic inhibition. European Journal of Neuroscience
15, 1499-1508.
Wendling, F., Hernandez, A., Bellanger, J.-J., Chauvel, P., Bartolomei, F., 2005. Interictal to ictal
transition in human temporal lobe epilepsy: insights from a computational model of
intracerebral EEG. Journal of Clinical Neurophysiology 22, 343.
Whitfield-Gabrieli, S., Thermenos, H.W., Milanovic, S., Tsuang, M.T., Faraone, S.V., McCarley,
R.W., Shenton, M.E., Green, A.I., Nieto-Castanon, A., LaViolette, P., Wojcik, J., Gabrieli,
J.D., Seidman, L.J., 2009. Hyperactivity and hyperconnectivity of the default network in
schizophrenia and in first-degree relatives of persons with schizophrenia. Proceedings of the
National Academy of Sciences of the United States of America 106, 1279-1284.
Yang, C.-N., Lee, T.-D., 1952. Statistical theory of equations of state and phase transitions. I. Theory
of condensation. Physical Review 87, 404.
Yang, D.-P., Zhou, H.-J., Zhou, C., 2017. Co-emergence of multi-scale cortical activities of irregular
firing, oscillations and avalanches achieves cost-efficient information capacity. PLoS
computational biology 13, e1005384.
40
Criticality in the brain
Cocchi et al. 41
Yang, G.J., Murray, J.D., Repovs, G., Cole, M.W., Savic, A., Glasser, M.F., Pittenger, C., Krystal,
J.H., Wang, X.-J., Pearlson, G.D., 2014. Altered global brain signal in schizophrenia.
Proceedings of the National Academy of Sciences 111, 7438-7443.
Yang, H.D., Shew, W.L., Roy, R., Plenz, D., 2012. Maximal Variability of Phase Synchrony in
Cortical Networks with Neuronal Avalanches. Journal of Neuroscience 32, 1061-1072.
Yeo, B.T., Krienen, F.M., Sepulcre, J., Sabuncu, M.R., Lashkari, D., Hollinshead, M., Roffman, J.L.,
Smoller, J.W., Zollei, L., Polimeni, J.R., Fischl, B., Liu, H., Buckner, R.L., 2011. The
organization of the human cerebral cortex estimated by intrinsic functional connectivity. J
Neurophysiol 106, 1125-1165.
Zalesky, A., Fornito, A., Cocchi, L., Gollo, L.L., Breakspear, M., 2014. Time-resolved resting-state
brain networks. Proceedings of the National Academy of Sciences of the United States of
America 111, 10341-10346.
Zapperi, S., Castellano, C., Colaiori, F., Durin, G., 2005. Signature of effective mass in crackling-
noise asymmetry. Nature Physics 1, 46-49.
Zapperi, S., Lauritsen, K.B., Stanley, H.E., 1995. Self-organized branching processes: mean-field
theory for avalanches. Physical review letters 75, 4071.
Zhigalov, A., Arnulfo, G., Nobili, L., Palva, S., Palva, J.M., 2017. Modular co-organization of
functional connectivity and scale-free dynamics in the human brain. Network Neuroscience.
41
Criticality in the brain
Cocchi et al. 42
Box 1: Ten-Point Summary
1. Criticality arises when a system is close to dynamic instability and is reflected by scale-free
temporal and spatial fluctuations
2. Critical temporal fluctuations (crackling noise) occur in simple systems close to a bifurcation
3. Critical spatiotemporal fluctuations (avalanches) occur in complex systems close to a phase
transition
4. Crackling noise and avalanches have now been observed in a wide variety of neuronal recordings,
at different scales, in different species, and in health and disease
5. Computational models suggest a host of adaptive benefits of criticality, including maximum
dynamic range, optimal information capacity, storage and transmission and selective enhancement of
weak inputs
6. Resting-state EEG and fMRI data show evidence of critical dynamics
7. The onset of a specific cognitive function may reflect the stabilization of a particular subcritical
state under the influence of sustained attention
8. Mounting evidence and models suggest that several neurological disorders such as epilepsies and
neonatal encephalopathy reflect bifurcations and phase transitions to pathological states
9. Novel insights into neuropsychiatric disorders such as schizophrenia and melancholia might also be
obtained by leveraging the tools of criticality, although this currently remains somewhat speculative
10. While the application of criticality to neuroscience is an exciting field, progress needs to proceed
with due caution, using appropriate methods, considering alternative complex processes and using
computational models in partnership with data analysis
42
Criticality in the brain
Box 2: Glossary
Cocchi et al. 43
The attractor of a dynamical system is the set of all points traversed once initial transients have
passed. Attractors can be fixed points (with steady state solutions), limit cycles (periodic) or chaotic
(deterministic but dynamically unstable and aperiodic).
The basin of attraction of an attractor is the set of all initial conditions that eventually flow onto that
attractor.
A system can have more than one attractor even if all its parameters are fixed. Such a system is said to
be multistable. Such systems will also have multiple basins of attraction, separated by basin
boundaries. Bistability occurs in a multistable system which has exactly two attractors.
A metastable system does not have any attractors. It instead has a series of saddles (fixed points with
attracting and repelling subspaces) that are linked into a complex (heteroclinic) cycle. A metastable
system will jump endlessly from the neighborhood of one saddle to another. Metastable systems are
also called winnerless competition.
An attractor that only changes slightly (and smoothly) when its underlying parameters are changes is
said to be structurally stable. If the topology (shape) of the attractor fundamentally changes then the
attractor is said to be unstable – or, alternatively undergo a bifurcation. The value of the parameter at
which that discontinuous change occurs is said to be a critical or bifurcation point.
A system consisting of many interacting components may exhibit a sudden change in state in the
presence of a slowly tuned control parameter (such as temperature). Such a transition is called a
phase transition and typically separates an ordered from a disordered state. Technically, a phase
transition corresponds to a discontinuity in the thermodynamic free energy of a system.
Criticality occurs when a system is poised at the point of a dynamic instability. Because of this,
microscopic fluctuations are not damped but instead appear at all scales of the system. This yields
power-law fluctuations in the temporal domain ("crackling noise") and the spatiotemporal domain
("avalanches"). A critical system will show scaling laws, such that a single (universal) function can
map the shape of fluctuations at any scale into those at the scale above (or below).
Bifurcations may be super-critical (when stable oscillations appear above the critical point) or sub-
critical (when a zone of bistability occurs below the critical point). Only supercritical bifurcations can
yield power law (critical) fluctuations: Sub-critical bifurcations lead to multistable switches that occur
on a characteristic time scale.
Phase transitions may be continuous (second order) or discontinuous (first order). Mathematically,
these are equivalent to a super- and sub-critical bifurcations, respectively. As with bifurcations,
critical power law fluctuations only occur in the neighbourhood of a continuous phase transition.
Discontinuous phase transitions can yield multistable switching or complex mixtures of states (such as
in boiling water).
43
Criticality in the brain
Cocchi et al. 44
−𝑘+1
)
for the same critical exponent k.
Some systems need to be externally tuned by a control parameter close to their critical point. In other
systems, criticality will emerge from many initial parameter values, usually due to plasticity and
memory.
The order parameter of a complex system is a macroscopic observable such as the magnetic field of a
ferromagnet. A non-zero order parameter arises in the ordered state of a system, in the supercritical
(or active) phase. In the subcritical phase, the order parameter remains at zero even with the addition
of energy. Such a state is called an absorbing state.
A power law exists between two variables x and y if they obey the relationship: 𝑦 ∝ 𝑥−𝑘. If the power
law arises in the setting of criticality, then the constant k is called the critical exponent. A system
shows power law behaviour if the probability density function of its fluctuations obeys 𝑓(𝑥) ∝
𝑥−𝑘 for all values of x greater than some minimum cut-off 𝑥𝑚𝑖𝑛. A power law probability distribution
is also called a Pareto distribution. The corresponding cumulative distribution obeys 𝑃𝑟𝑜𝑏(𝑋 > 𝑥) ∝
1 − (𝑥
⁄
𝑥𝑚𝑖𝑛
A system is said to be scale-free when it doesn't have a characteristic time or length scale. Scale-free
systems show a power law probability distribution over several orders of magnitude with an exponent
k that is less than 2. Correspondingly, the variance of a scale-free system is only bounded by the
system size.
A complex system shows evidence of slowing down when it is close to a critical point – that is, the
time scale of its fluctuations (the characteristic return to the mean) slow down, changing from a fast
(exponential) process to a slow power law.
A system has an exponential distribution when its probability density function is given by 𝑓(𝑥) ∝
𝑒−𝑥/𝐿. The system has a single time scale, corresponding to the constant exponent L.
Any distribution whose probability distribution drops off more slowly than an exponential distribution
is said to be heavy-tailed. The Pareto distribution 𝑓(𝑥) ∝ 𝑥−𝑘 is a classic heavy tailed distribution but
log-normal and stretched exponential (Weibull) distributions are also heavy-tailed.
Many physical systems close to a critical point do have some weak damping that acts on the largest
fluctuations: These systems show an exponentially-truncated power law, 𝑓(𝑥) ∝ 𝑥−𝑘𝑒−𝑥𝑙.
44
|
1903.10626 | 1 | 1903 | 2019-03-25T23:06:22 | The role of physiological complexity changes in resting-state EEG in clinical effectiveness of rTMS and tDCS in treatments of resistant depression | [
"q-bio.NC"
] | The present literature about possible mechanisms behind the effectivity of noninvasive electromagnetic stimulation in major depressive disorder (MDD) is not very rich. Despite extensive research in applications for clinical practice, the exact effects are yet not clear. We are comparing our previous results about the complexity changes induced by repetitive Transcranial Magnetic Stimulation (rTMS), and transcranial direct current stimulation (tDCS) which are known to modulate neural dynamics. Also, we are reviewing different biomarkers of complexity changes connected to depression, and how they change with the stimulation. TDCS is low-intensity TES, known to have polarity specific effects (neuromodulatory effects), and rTMS is inducing an electric field in the tissue circumstantially via Faraday's law. Both nonlinear modalities of electromagnetic stimulation may affect the levels of physiological complexity in the brain. We also compare the changes of complexity in electroencephalogram (EEG) and electrocardiogram (ECG), as potential future predictors of therapy outcome. | q-bio.NC | q-bio | The role of physiological complexity changes in resting-state
EEG in clinical effectiveness of rTMS and tDCS in
treatments of resistant depression
1Department for General Physiology and Biophysics, University of Belgrade, Belgrade, Serbia
2Amsterdam Health and Technology Institute, HealthInc, Amsterdam, the Netherlands
Milena Čukić1,2, PhD
Abstract
The present literature about possible mechanisms behind the effectivity of noninvasive
electromagnetic stimulation in major depressive disorder (MDD) is not very rich. Despite
extensive research in applications for clinical practice, the exact effects are yet not clear. We are
comparing our previous results about the complexity changes induced by TMS, and tDCS which
are known to modulate neural dynamics. Also, we are reviewing different biomarkers of
complexity changes connected to depression, and how they change with the stimulation. TDCS is
low-intensity TES, known to have polarity specific effects (neuromodulatory effects), and rTMS
is inducing an electric field in the tissue circumstantially via Faraday's law. Both nonlinear
modalities of electromagnetic stimulation may affect the levels of physiological complexity in the
brain. We also compare the changes of complexity in electroencephalogram (EEG) and
electrocardiogram (ECG), as potential future predictors of therapy outcome.
Keywords: Transcranial direct current stimulation (tDCS), repetitive Transcranial Magnetic Stimulation
(rTMS), Electroencephalogram (EEG), Variability Heart Rate (VHR), treatment-resistant depression,
Anodal stimulation, Cathodal stimulation, plasticity, Theta burst stimulation (TBS), Neuromodulation,
Neurostimulation, Rehabilitation, Major Depressive Disorder (MDD)
Introduction
The course of depression is very often recurrent and can become chronic with highly likely
relapse rates within one year of remission from 35% to 80% (Eaton et al., 2008; Fekadu et al.,
2009). Usual treatments for depression are pharmacological and psychological, but Rush and
colleagues report that one third fail to reach remission (Rush et al., 2006). Non-invasive
electromagnetic modalities of stimulation, like transcranial magnetic stimulation (TMS) and
transcranial electric stimulation (TES) are alternatives to this usual practice. We are going to
briefly review several application modalities of TMS and low-intensity modality of TES, namely
tDCS. Here we want to focus on elucidating the effectiveness of two electromagnetic treatment
considered 'novel.' My putting it under the quotation marks is intentional because transcranial
magnetic stimulation was launched as a commercial stimulator 1985 (Prof. Antony Baker, 1985)
and it cannot be still novel after almost 35 years of research and diagnostic applications. Similar
reasoning applies to tDCS which is low-intensity stimulation modality of transcranial electrical
stimulation (TES), but still, some details important for an understanding of its mechanisms are
missing. Like Opitz put it in recent research, the important point is in interpretability of TES effects
(Opitz et al., 2018): 'if electric fields are delivered inconsistently, but effects are observed
nevertheless, the results are more difficult to interpret because effect could be driven by other
incidentally affected brain regions.'
TMS was initially introduced as a non-invasive tool for investigating and mapping cortical
functioning and connectivity (Barker et al., 1985). TMS primarily use a strong magnetic field to
induce an electric field in the surface layers of the brain (cortex) and without pain characteristic to
ECT, initiate optimally focused activation of neural structures. One of its modalities used in
psychiatry is repetitive TMS (rTMS). While standard TMS induce single (or sometimes) paired
pulses, repetitive TMS delivers repeated pulses showed to enable more extended modulation of
neural activity. It can be slow (if repetitions are up to 1Hz/one per second) or fast (with 10Hz or
higher frequencies). Fast rTMS have excitatory, while slow rTMS applications have inhibitory
effects (Rosa and Lisanby, 2012). It is interesting to note that rTMS was finally confirmed for
reimbursements in a treatment-resisting depression just last year (2018) in the Netherlands, 34
years from its début in medicine (US Food and Drug Administration- FDA approved the first rTMS
device for treatment of MDD in 2008). The basis for rTMS treatment is a poor response to at least
one pharmacological agent in the current episode (O'Reardon et al., 2007). We are going to
elucidate on that question 'how long it takes for a scientifically proven innovation to be translated
to clinical practice?' The scientific literature about a safe and useful application of TMS is so rich
that it is hard to believe that it took for so long. The reason for using TMS in treatments of
depression came from a demonstration of functional impairments in prefrontal cortices and limbic
regions (Atkinson et al., 2014) and very well based frontal asymmetry in alpha (Allen et al., 2004).
There are several different modalities of rTMS use. Since high-frequency stimulation can be
uncomfortable during the initial stimulation period, low-frequency rTMS was introduced to
minimize headaches and scalp discomfort; it is also lowering the risk for developing seizures
(Rossi et al., 2009). When the stimulation is ongoing over the left and right dorsolateral prefrontal
cortex-DLPFC (simultaneously, or stimulation one side after another), it is called Bilateral
application of rTMS. Conca and colleagues argue that this kind of application may reinstate any
imbalance in prefrontal neural activity (Conca et al., 2002). Fitzgerald showed that likelihood for
a clinical response increases by providing both types of stimulation (Fitzgerald et al., 2006). Deep
TMS (dTMS) can stimulate larger brain volumes and deeper structures (Roth et al., 2007) which
is maybe more relevant to the pathophysiology of depression (Costafreda et al., 20013;
Kwaasteniet et al., 2013; Atkinson et al., 2014). Theta burst stimulation (TBS) (Huang et al., 2005)
utilize high and low frequencies in the same stimulation train; it delivers three bursts at high
frequency (50Hz) with an inter-burst interval of 5Hz (theta range) frequency. Two different
approaches to the application of TBS exists: continuous and intermittent (cTBS and iTBS). cTBS
delivers 300 to 600 pulses without interruption, and iTBS delivers 30 pulses (trains) every 10s
during 190s (600 pulses in total) (Chuang et al., 2015). While cTBS reduces cortical excitability,
iTBS increases it; they are mimicking the process of long term potentiation (LTP) and long term
depression (LTD) (Stagg and Nitsche, 2016; Huang et al., 2005). LTP allows for modulation of
synaptic strength that stabilizes for days, months, or even years and has therefore been postulated
as a likely candidate for memory formation in the brain (Anderson and Lomo, 1966; Bliss and
Lomo, 1973). LTD has been studied extensively in the hippocampus and refer to highly specific
processes; synaptic plasticity with very similar properties has been demonstrated in the neocortex
and is therefore commonly referred to as LTP-like plasticity (Stagg and Nitsche, 2016). The main
advantage of TMS is the reduced time of treatment in comparison to all other rTMS modalities.
For conventional rTMS it takes 20-45 min, while for TMS it is typically less than 5 min; also it
operates at 80% of the resting motor threshold (MT) and may be more comfortable than rTMS.
Another modality of stimulation synchronized TMS (sTMS), is a new treatment paradigm which
delivers stimulation synchronized to an individual's alpha frequency (Jin and Phillips, 2014). It is
based on rTMS ability to make entrainment of oscillatory activity to the programmed frequency
of stimulation with the aim of restoring normal oscillatory activity (Leuchter et al., 2013). It does
not cause neural depolarization; therefore it may cause fewer adverse effects.
In opposition to all modalities mentioned above of TMS, tDCS which is low-intensity modality
of TES has specific advantage of use; its low cost and portability are opening the possibilities of
use out of the clinical setup. tDCS is not triggering action potentials directly (unlike TMS), and it
is shown that it modulates cortical excitability by shifting neural membrane resting potential
(Nitsche et al., 2008). TDCS effects can outlast the stimulation period up to two hours (Nitsche,
2001) and even show significant effects for more extended periods. In our recent research, we
showed that tDCS is able of causing the shift in the brain-state (Čukić et al., 2018) based on EEG
recordings reconstruction in phase-space, for at least half an hour. The primary advantage of tDCS
compared to any TMS is its ease of administration, portability, much less recorded adverse events
(AE) and lower price. Mutz and colleagues aimed at examining the efficacy and acceptability of
non-invasive brain stimulation in adult unipolar and bipolar depression (Mutz et al., 2018). They
found the most substantial evidence for high-frequency rTMS over left DLPFC, followed by low-
frequency rTMS over the right DLPFC and bilateral rTMS. Intermittent TBS provides a potential
advance in terms of reduced treatment duration. Authors also stated that tDCS is a potential
treatment for non-resistant depression which has demonstrated efficacy in terms of response as
well as remission (Mutz et al., 2018). Tuomas Neuvonen and colleagues started making their
anytime anywhere tDCS applications in Soma already. They elaborated on interindividual
differences as one of the major sources of misinterpretation (Fonteneau et al., 2018)
Meta-analyses on the efficacy of rTMS and tDCS in treatment-resistant depression
In the recent tDCS review study, authors discussed the interaction of induced electric field
(EF) with tissue, and on electroporation and concluded that the intensities needed for
electroporation remain orders of magnitude above tDCS (Antal et al., 2017). Typical current in
tDCS montage is 1-2 mA (0.03-2mA/cm2 current to electrode area ratios depending on the
electrode size) which results in cortical EF strengths up to 0.4-0.8 V/m (Ruffini et al., 2013b). Both
the applied current and the resting brain EFs are ~1000 times lower than those for pulsed
stimulation used in electroconvulsive therapy (Alam et al., 2016) and are considered to be below
the required intensity for evoking an action potential in resting cell (Radman et al., 2009). They
are still capable of modifying the spontaneous firing rates of neurons, and also of inducing
plasticity processes in neural networks (Ranieri et al., 2012; Jackson et al., 2016). Neural networks
generate their own characteristic EFs in intracellular spaces, and this kind of stimulus can influence
their molecular and structural changes (Frohlich and McCormick, 2010). Based on the work of
Mutanen and Ilmoniemi on the induced brain-shift in an application of TMS, we later aimed at
elucidating the effect of tDCS by utilization of the same method. Frohlich and McCormick
detected the alteration of transmembrane potential for 0.5-1.3 mV, with the intensity of the induced
electric field of 2-4 mV/m. Both Miranda and Saturnino predicted values of the maximum EF in
the cortex of realistic head model are between 0.2 V/m and 0.5 V/m (if the current is 1mA)
(Miranda et al., 2013; Saturnino et al., 2015). According to Nelson and Nunneley, tDCS induced
energy in the cortex is 0.1 mW/kg, for the current of 1mA, a conductivity of 0.4 S/m and the
median value of EF 0.5 V/m (Nelson and Nunneley, 1998). Antal and other experts concluded in
their 2017 guidelines that those values caused by tDCS are safe (Antal et al., 2017). Wagner
reviewed similar details important to understand the electromagnetic interaction involved from
previous literature (Wagner et al., 2007). In case of TMS, a magnetic field (~1-4T pulsed over
0.125-1ms dependent of parameters) induces an electric field that drives currents in the brain of a
magnitude approximately 5.13x10-8 A/m in the cortex per 1A/s steady-state source. Currents are
carried by free charges and ions (known as ohmic or resistive currents, or volume conduction) or
through the depolarization of dipoles in the tissue layers (Wagner et al., 2007).
Fregni and colleagues were searching for predictors of antidepressant response in clinical trials
of transcranial magnetic stimulation (Fregni et al., 2006). At the time they did their research, it
was already known that rTMS has a significant antidepressant effect, the results were
heterogeneous. They hypothesized that individual patients' characteristics might contribute to such
heterogeneity in a sample comprising of 195 patients. Results showed that age and treatment
refractoriness were significant negative predictors of depression improvement when adjusted to
other significant predictors. The findings of Mutz and colleagues (2018) are in line because they
also found that younger people are more probable to be responders. Another study on the
effectivity 'Transcranial direct current stimulation in treatment-resistant depression: a randomized
double-blind, placebo-controlled study' was published in 2012 (Palm et al., 2012). They concluded
that 'anodal tDCS, applied for two weeks, was not superior to placebo treatment in patients with
treatment-resistant depression. However, secondary outcome measures are pointing to a positive
effect of tDCS on emotions'. The authors recommended modified and improved tDCS protocols
to be carried out in controlled pilot trials to develop tDCS towards an effective antidepressant
intervention in therapy-resistant depression. Berlim and colleagues published two studies on the
same topic in 2013 and 2014 (Berlim et al., 2013, 2014). The first one was about the efficiency of
tDCS, and the other on the efficiency of rTMS. They first concluded that the clinical utility of
tDCS as a treatment for MDD remains unclear when clinically relevant outcomes such as response
and remission rates are considered.
Moreover, they recommended future studies on larger samples. In their 2014 study, they
concluded that 'Meta-analyses have shown that high-frequency (HF) repetitive transcranial
magnetic stimulation (rTMS) has antidepressant properties when compared with sham rTMS.'
However, its overall response and remission rates in major depression (MD) remain unclear
(Berlim et al., 2014). Bennabi and his colleagues concluded in their 2015 study that 'tDCS did not
induce the clinically relevant antidepressant effect in active and sham stimulation groups. There
was no impact on psychomotor and neuropsychological functioning' (Bennabi et al., 2015). So, as
Fregni pointed out, the results are still heterogeneous.
Brunoni and colleagues aimed to assess tDCS efficacy and to explore individual response
predictors in their 2016 study (Brunoni et al., 2016). Since tDCS was extensively investigated for
the treatments, particularly of major depression, they performed another review study by
utilization of individual patient data (contrary to aggregate data approach in many review studies
before). The current literature confirms that tDCS can induce neuromodulatory changes in cortical
activity, and essential for depression, to ameliorate depressive symptoms (Brunoni et al., 2012).
Its antidepressant effects are based on previous findings of hypoactivity of the left dorsolateral
prefrontal cortex (DLPFC). Brunoni concentrated on providing precise estimates of tDCS efficacy
based on depression improvement, response and remission rates. They tried to identify variables
associated with tDCS efficacy (Brunoni et al., 2016). They also used several up-to-date software
tools for detecting the bias, to re-analyze the original data and so on. After careful literature
research and selection (among 153 studies), their review study comprises of six studies, only (they
included studies which are double-blinded, sham-controlled, reporting on a number of repetitions,
dosage, medication, and other detailed data). However, on end among those six studies were three
(half of all selected) studies published by the same group which is also a bias (Loo et al., 2010;
Loo et al., 2012; Brunoni et al., 2013). The primary dependent variable was clinical response since
the authors claim that it has better performance over remission which presented heterogeneity of
cut-offs. They used a number needed to treat (NNT) as a measurement which illustrates clinical
intervention effectiveness; the higher the NNT, the less effective the intervention. The authors
found significant effect sizes of the efficacy of active vs. sham tDCS in terms of depression
improvement response and remission, with overall response and remission rates of active tDCS of
34% and 23% respectively (NNT of 7 and 9) (Brunoni et al., 2016). The authors further compared
their own meta-analysis study (on six studies, half of which are their own) with a Cochrane meta-
analysis assessing the efficacy of antidepressant drug treatments in primary care (Arrol et al., 2009)
finding their results are in line.
Brunoni and colleagues also claim their findings to be in the same range as another meta-
analysis for depression which utilized active vs. sham repetitive transcranial magnetic stimulation
(rTMS) that found a response and remission NNTs of 6 and eight respectively. So, for treatment
response and remission rTMS meta-analysis study(Berlim et al, 2014) found absolute rates of
29.3% and 18.6%, while tDCS meta-analysis found 34% and 23% respectively. The only problem
here was that out of 289 patients included in tDCS study there were just 147 patients with an active
stimulation, and in rTMS study, the sample is almost ten-fold; n=1371. From the point of statistical
learning, the other sample would be expected to give us reliable results, while the first sample for
this type of comparison is considered to be very modest.
Brunoni and colleagues also recommended two potential applications of tDCS in treatments
for depression: in primary care settings and as a non-pharmacological, neuromodulatory therapy
for depression (Brunoni et al., 2016). In another meta-analytical study performed by Antal et al.
(2017) the authors commented on limitations of Brunoni study, and used a much larger sample,
covering the publications and data collected from approximately 8000 people. They concluded
that tDCS is considered safe, giving guidelines for avoiding Adverse Events (AE) as well as
questionnaires for further research (designed to prevent all previously reported AEs).
In their meta-analysis comprising of 56 previously published studies, Mutz and colleagues
found about actual effectivity of different modalities of TMS used for treatment-resistant
depression (Mutz et al., 2018). They reported that high-frequency rTMS (over left DLPFC) was
associated with improved rates of response as well as remission in comparison with sham
treatment. Only one of the reviewed studies recruited bipolar depression patients (Nahas et al.,
2003) but with demonstrated antidepressant efficacy. They also found that low-frequency rTMS
over the right DLPFC was associated with significantly higher response and remission rates than
sham stimulation. Low-frequency rTMS over lDLPFC did not show significant improvement
(there were no significant differences in response rates compared to sham). Bilateral rTMS
demonstrated significant improvement in response but not remission rates compared to sham.
When dTMS is concerned, it also showed significant improvement (for both response and
remission rates), but response rates were marginally higher relative to sham. Neither response nor
remission rates for sTMS were significantly higher than a sham. Intermittent TBS over the left
DLPFC was associated with a fivefold improvement in response rates compared to sham. No
evidence was found of improvement in case of cTBS over rDLPFC and bilateral TBS. In case of
tDCS, they confirmed significant improvement in both response and remission rates in comparison
with sham; their analysis suggested tDCS to be effective only in patients with non-treatment
resistant depression (Mutz et al., 2018). In contrast to most rTMS trials, tDCS showed a potential
initial therapeutic option for depression.
The evidence that rTMS and tDCS are activating deeper structures
A study performed by Li and colleagues provides a valuable insight into mechanisms by which
tDCS may modulate cognitive function and also has implications for the design of future
stimulation studies (Li et al., 2018). They showed that the effects of tDCS on brain activity are
dependent on the cognitive state. Both anodal and cathodal tDCS produced widespread BOLD
changes in brain areas anatomically remote from the critical area being stimulated. Those results
are in line with Sven Bestmann's findings, which first demonstrated that even with the sub-
threshold stimulation of TMS, deep structures are activated (Bestmann et al., 2003 and 2004). It
may come as a surprise, but we are looking here at the evidence that both tDCS and rTMS are
affecting other structures than intended. It may be that it is in connection with the finding that
severe depression does have the problem with functional connectivity, due to the deep white matter
tracts significant in a fronto-limbic system (Kwaasteniet et al., 2013; Kim et al., 2013). The focus
of a majority of researchers was for so long at defining the polarity specific and other noticed
effects, but the problem is that both methods are noninvasive and therefore only direct
measurement can be a solution for this riddle.
Opitz et al (Opitz et al., 2015, 2016, 2018) and Alekseichuk (Alekseichuk et al., 2018) applied
the robust physical approach in elucidating this complex interaction between the induced electrical
field and brain tissue. Opitz performed the direct measurement from implanted electrodes (in direct
empirical validation of TES) in patients who were operated in a procedure intended to implant a
grid of stimulation electrodes for their epilepsy, and compared the results with the cohort of 25
healthy people (neurotypical individuals) from connectome project. The motivation for their work
was the lack of exact measurements and the problem with variability between individuals. Opitz
and colleagues combined direct intracranial measurements of electric fields generated by TES in
surgical epilepsy patients with computational modeling (Opitz et al., 2018). They directly validated
the computational models and identified key parameters needed for accurate model predictions.
They also derived practical guidelines for a reliable application of TES in terms of the precision
of electrode placement needed to achieve a desired electric field distribution. When comparing the
measured and predicted values of induced electrical fields, they encounter the problem of
conductivity constants values. They are reported in the literature with great variety, so further
research is needed (Opitz et al., 2018). To establish standards in this field would be of significant
importance to increase the reliability of future stimulation protocols. When they were testing the
model developed for the study they found that maximum strength of the induced field predicted
by the model was larger than that they directly measured. The mean electric field strength was
0.058 mV/mm for the measurement results, while the mean electric field strength over the best
fitting stimulation results was 0.100 mV/mm (range 0.071-0.122 mV/mm) (Opitz et al., 2018). In
another patient with a different implanted grid of electrodes, the trend was in the opposite direction;
mean measured field strength was 0.115 mV/mm, while predicted was 0.060 mV/mm.
The authors concluded that computational models using current standard conductivities
slightly misestimate the measured field estimate strength, suggesting the need for individual
adjustments of conductivities (Huang et al., 2017). They also demonstrated the importance of
factors like the accuracy of the skull or surgical component modeling; other factors as gyral folding
and CSF thickness showed to have a profound effect on TES induced fields (Opitz et al., 2015).
They urge other researchers to direct their efforts to directly measure conductivities to improve
realistic modeling, or even include more tissue types (Aydin et al., 2014). It is based on their
finding that the variation of the position of an electrode may maximally vary to 1cm, due to folding
in the brain surface (electricity concentrate differently on sharp and flat surfaces, depending on
geometry) in order to provide efficient stimulation. In their work, it is also confirmed how
important it is to accurately model the shape of electrodes for realistic field calculations (Saturnino
et al., 2015). In another study (Alekseichuk et al., 2018) they systematically evaluated the induced
electrical fields and analyzed their relationship to brain anatomy (TMS and TES induced fields
were compared in finite element method, in a mouse, capuchin monkey, and human model). The
theoretical explanation and final calculation of the induced electrical field during the stimulation
was necessary for those who are further developing computational models.
Results from our previous research about the changes of complexity induced by the
application of TMS and tDCS
In my Doctoral dissertation, and before that in my magisterial thesis, I explored the parameters
influencing the strength of stimulation (Čukić, 2006) and the changes in complexity due to TMS
(Čukić, 2011). I am going to report here some of our results which can help understand the changes
that both TMS and tDCS can induce, based on co-recording of the electrophysiological signal.
Now, if we want to understand how a stimulus is acting upon a system, we can observe the levels
of complexity before the stimulation, and after the stimulation and by comparing those levels we
can conclude about the influence. This is precisely what we did with two mentioned modalities of
stimulation. In our work from 2011/12, we compared single pulse TMS induced complexity
changes with the complexity changes induced by peripheral electrical stimulation (the same target
muscle was FDI), and we showed that single pulse TMS is able to decrease the complexity of the
system (Čukić et al., 2013). In the first set of results, we aimed at examining the changes in surface
electromyogram (EMG) recorded from the first dorsal interosseous muscle (FDI) which was the
target muscle in TMS protocol. In that protocol, we were also changing the level of voluntary
contraction to test how the amount of present force in the muscle can interfere with the effect of
stimulation. On the first graph from those results (Figure 1) we can see surface EMG recorded in
that experiment and also magnified epochs of that signal we used for further fractal analysis.
Figure 1: The raw electromyogram signal recorded from the target muscle (FDI) in a TMS stimulation protocol. The
epochs for analysis extracted from the trace 'before' and 'after' the stimulation are magnified. They were used for
further analysis to probe for the complexity changes induced by the stimulation (Čukić et al., 2013).
After comparative classical spectral and fractal analysis, we found the significant changes in the
epochs recorded in participants (all healthy persons) before and after the stimulation. The
following two illustrations are showing both the spectral and fractal result of the analysis.
Figure 2: The distributions of frequnecies (spectra) before (left) and after (right) stimulation in the surface EMG signal;
the change induced by the stimulus are not immediately obvious (Čukić et al., 2010).
Figure 3: Amplitude spectra of surface EMG before (left) and after (right) the TMS stimulation. The difference in
complexity levels can be seen with a naked eye (Ćukić et al., 2010).
It can be seen from both graphs that the drop in complexity of the signal is apparent. The level of
complexity in a healthy system (being a brain, a heart or a muscle) is showing systems' ability to
adapt. When an influence out of the dynamical system impact the system, the change can be
observed via changes in complexity (or irregularity). In this case, it can be seen that after the
application of TMS, the complexity of surface myogram reduced. One can see that on the
following figures: 4 and 5. The first is showing a histogram, constructed from fractal dimensions
calculated from sections of EMG signal, showing the distribution of its values. The red line
compared to black dotted one is showing how those FDs calculated from epochs of EMG before
and after the stimulation are distributed. Those representing 'after' sections are shorter, meaning
that the facilitation happened during this stimulation. Another figure (4) is representing the mean
values of calculated FD (from EMG) showing a significant difference between them.
Figure 4: The changes in Higuchi fractal dimension calculated from surface EMG epochs before (red) and after the
stimulation. The distribution is slightly shifted to the left, illustrating the shortened latencies in the response after the
stimulation (Čukić, 2006).
Figure 5: The average values of Higuchi Fractal dimension before (dotted line) and after the stimulation; it can be
seen that the complexity of signal dropped after the stimulation (Čukić, 2012).
Several years after that experiment, our colleagues from Finland published a study which
examined the change TMS induced in brain state (in phase-space) of 10 healthy participants
(Mutanen et al., 2013). They applied recurrence plot analysis on EEG concurrently recorded with
TMS, and they also stimulated motor cortex (the same target muscle as we did in our experiment).
This is why I believe that our results are in a way comparable and compatible. They showed that
even some time after the stimulation is over, the system (brain) stayed in a highly improbable and
higher energy state in comparison to the resting state before the stimulation. The following diagram
is illustrating the process they described in their work (Mutanen et al., 2013), on our graphical
representation similar to theirs (Figure 6).
Figure 6: A) The blue arrow represents the primary current source, a flow of ions in synapses in the motor cortex. The
dashed blue line represents returning volume current. Jp is the primary current density (induced by the stimulation).
B) This schematic representation shows how TMS affects brain-state. The green and red curves are pre- and post-
TMS trajectories (of a brain state). The spontaneous activity (green) occupy one region in state-space. After the
stimulation, it occupies another region (red) which is in a higher energy state. Gradually, over time, the system will
return to lower energy and more probable state. The projections of those two states are in EEG signal space, i and j
standing for two different channels. In the signal, space trajectories are measured only at discrete time points (dotted
projections, red and green).
The green line is representing the trajectory of a system before the stimulation, and red line the trajectory in another
part of phase space after the TMS. Both trajectories have the projections on the plane (dots in the plane Ei Ej), which
are the samples of recorded EEG (Ilmoniemi and Kičić, 2010). Jp and Jv are the current densities illustrating the
ongoing process during the stimulation; Jp is causing the movement of charges in the tissue (it appears due to the
induced electrical field in the tissue), and Jv is volume conductor density current which is a counterpart process to the
first one. In essence, they showed that due to the stimulation the system is staying in elevated state (the brain-shift is
confirmed by the measures of Global recurrence, MSS and SV) long enough so we can detect it.
Next thing we did, was to test the same methodology, but on another type of stimulation, tDCS.
We are presenting here part of the results which are dealing with complexity changes. On figure
6, it can be seen how HFD and SampEn detected the changes in complexity after tDCS stimulation.
We know that this kind of stimulation is very different, due to the much lower level of energy
delivered to the tissue, and we know that it is considered to have just modulatory character. In this
research (in 2015) we tried to test the changes in complexity before and after the stimulation
('before' was considered a resting state since we did not have sham to compare to, and after was
immediately after-t1, or a half an hour after-t2). However, if you can detect changes that survive
more than half an hour later, does that imply early plastic-like changes? Unfortunately, we were
not able to confidently interpret the results, because contrary to our previous knowledge about just
modulatory effect of tDCS we detected temporary decrease and after that increase of complexity
measured by Higuchi fractal dimension and Sample entropy.
What is more, it looked as at some combinations there is no difference between cathodal and
anodal stimulation, in terms of complexity changes. Due to the small sample, we opted to apply
another methodology set, but are working on repeating with the same method in presently ongoing
research. Nitsche showed much earlier that induced changes in system survive more than two
hours (Nitsche, 2001). It would be interesting to see whether the complexity stays on a different
level after a couple of days, or a week or even longer.
Figure 7: Comparison of complexity changes before and after anodal (left) and cathodal (right) stimulation,
represented with two calculated measures Higuchi fractal dimension and sample entropy. The epochs extracted from
EEG traces (on ten electrodes) were before the simulation (t0), immediately after the stimulus (t1) and half an hour
after the stimulation (t2) (Čukić, 2017).
After all those examples, we can say that both TMS and tDCS are capable of changing the
complexity of the system (brain). Although the mechanism and power of their stimulation are
different even more than order (power) of the value of the induced field, some similarities can be
drawn from our results. If we know that in depression an elevated complexity on cortex is
observed, and that majority of researchers agree that this may serve as a biomarker for depression,
is it too big a stretch if we think that maybe both rTMS and tDCS can help because they are both
diminishing that complexity? If a significant number of people suffering from depression is
reporting amelioration of symptoms (let us remember that remission is defined as 'symptoms
becoming bearable') they might profit from the ability of electromagnetic stimulation to at least
temporarily, decrease the problematic elevated excitability in their DLPFC and elsewhere. We
have to stress here, that as well as in ECT, both tDCS and rTMS are requiring maintenance therapy
since the benefit reported by patients is lasting for a limited timeframe.
To conclude, further research is needed, on a much larger sample.
Are rTMS and tDCS capable of changing the complexity of the brain?
In our previous work, we were interested in complexity changes in the human brain after
electromagnetic stimulation (Čukić et al., 2001, 2008, 2013, 2018). The logic behind this is that
the complex system (and the brain is one of the most complex dynamical systems that we know)
inherently exhibit certain levels of complexity with a possible purpose of accommodation to
internal or external changes of parameters which affect the conditions and surroundings of the
system is operating in. We have to be aware that we can observe this only indirectly. We are
analyzing an EEG as a product of the complex dynamical system (EEG is a composite signal
comprising of individual electric signals as a consequence of activation of different neurons and
neuronal groups). An electrical signal generated by neural networks when superimposed give us
the record from every point we positioned an electrode on the scalp. Of course, we can say that
EEG is an electrical representation of brain functioning, but what it is showing us are the voltages
picked up from the brain surface (cortex). We can only speculate how that signal represents the
deeper structures activities. Despite the extensive use in research and treatment, our knowledge
about exact mechanisms by which tDCS is influencing different structures in the brain is limited.
Lucia Li and her colleagues examined how transcranial direct current stimulation modulates brain
network function (Li et al., 2018). They used MRI for simultaneous recording whit tDCS. From
resting state recordings, the main effect of tDCS was to accentuate default mode network (DMN)
activation and salience network (SN) deactivation.
In contrast, during task performance, tDCS increased SN activation (Li et al., 2018). In the
absence of a task, the main effect of anodal tDCS was more pronounced, whereas cathodal tDCS
had a more significant effect during task performance. Cathodal tDCS also accentuated the within
DMN connectivity associated with the performance. There were minimal main effects of
stimulation on network connectivity. These results demonstrate that right inferior frontal gyrus
tDCS can modulate the activity and functional connectivity of large-scale brain networks involved
in cognitive function, in a brain state and polarity dependent manner. Bestmann showed that with
sub-threshold stimulation, MRI captured the activation of the auditory system, transversal and
superior temporal girys, inferior colliculus and mediate geniculate nucleus, which can be
understood as change induced by synaptic transmission (induction of changes of excitability in
sensory-motor areas, M1/S1) (Bestmann et al., 2003).
When we apply any artificial influence (like artificially induced electric field in proximity of
sensitive neural or muscle tissue) which we know is efficient in inducing action potential (or
smaller changes in electrical sense) in the tissue it could be observed through the lens of nonlinear
analysis of composite signal (i.e. EEG or EMG)(Čukić, 2006, 2011 and 2013). The levels of
complexity induced by this artificial (un-natural) stimulus are changed due to the electrical nature
of the phenomena (Čukić, 2011). If we are using TMS we are inducing an artificial electrical field
(via Faraday's law) which in turn force electrical charges to move through the tissue (Wagner et
al., 2008; Ridding and Rothwell, 2007; Thickbroom, 2007; Bestmann, 2008).
When we try to compare those two stimulations (rTMS and tDCS), the data from the literature
are very diverse. We must say that the fundamental biophysical, or physics research followed the
advent of TMS just shortly (Paton and Amassian, 1987). Harris and Miniussi elucidated the effects
of stimulation in cognitive experimenting focusing on the mechanism of action of induced
electrical field on cognitive performances (Harris and Miniussi, 2008), and above mentioned
authors explored those details in another, more physiological or biophysical level (like Bestmann,
2008). Both electromagnetic techniques are evolving, and we have to invent better, safer and
deeper penetrating noninvasive solutions for depressive disorders.
We would like once again to return to the connection between the electromagnetic stimulation
in cases of depression and dynamic of heart rhythm (ECG); we already elaborated on a connection
between variability heart rate and response on ECT, or potential forecasting of responders based
on nonlinear measures. Researchers who explored the connection of VHR and outcome of therapy
concluded that '…low baseline HRV is associated with rapid relapse of depression after ECT.
Both high baseline HRV and increasing HRV predict a sustained outcome' (Karpyak et al., 2004).
Other researchers in the field also noticed that the low level of VHR corresponds with the severity
of disease and that VHR often increases when a patient reaches the remission (Bozkurt et al, 2013).
So, without repeating a summary of many physiological complexity studies in this particular task,
the conclusion is that the dynamics of the heart is characteristically aberrated in persons diagnosed
with bipolar or unipolar depression, and also in burnout syndrome. Lower levels of complexity in
heart rhythm are also indicating a potential risk of developing cardiovascular diseases in people
with depression.
In comparison with EEG analysis results in depression, we can say that the detection from two
physiological signals (EEG and ECG) has the opposite trend, for physiological reasons. In EEG
studies, an elevated complexity in depressive patients can be observed and measured (and used as
potential neuromarker). In ECG studies, a lower level of nonlinear measures (as well as classical
ones) and generally lower complexity can be observed in depressives. However, both of them can
be used as useful biomarkers to early detect and monitor the developments in every single case.
The difference is that recording of EEG requires a visit to the clinic, while the recording of ECG
is already possible with portable monitoring devices. A similar parallel exists between the
utilization of rTMS and tDCS; the latter can be used in primary care and even at home, which is a
much more accessible option to patients and their families. Like DST was used to differentiate
melancholic depression from other less severe depression forms (Shorter and Fink, 2010), so
fractal and nonlinear measures of EEG (or ECG) can serve as useful biomarkers necessary for the
improvement of accuracy of clinical diagnostic and treatment.
Let us return for a moment to one of the results from our 2016 pilot study; we showed that
there is a difference (which can be measured) between the EEG complexity of patients diagnosed
with depression who are in acute episode and those who already are in remission.
Counterintuitively, those who remitted, exhibited higher levels of complexity, in comparison with
participants who were still in exacerbation (Čukić et al., 2018). This phenomenon still needs
further exploration. We are currently working on collecting more data, because we believe that
even an opposite trend could be present if we screen patients later, after the remission (after
possible neural reorganization). Some results from studies dealing with the effect of mindfulness-
based cognitive therapy (MBCT) showed that during eight weeks of treatment (or training), the
connection between the insula and ventromedial prefrontal cortex were uncoupled, as a marker of
improvement measurable within the fronto-limbic system (Davidson et al., 2004; Slagther et al.,
2007). It would be interesting to explore further how applications of different measures of
complexity on both electrophysiological signals (EEG or ECG) can improve mental healthcare
practice. They could become a valuable tool in decision making for clinicians.
When the interaction between electromagnetic stimulation and medication is concerned, we
know much more about that, than about the changes of complexity or irregularity, due to a much
richer body of evidence in existing literature. For example, among other guidelines, it is
recommended in potential tDCS use (Antal et al., 2017) that specific clinical presentation (for
example bipolar depressive disorder) requires some mood stabilizers, and not some of the drugs
which might attenuate the anodal effect of tDCS. I will add here one anecdotal example as an
illustration of that interaction. In the beginning of our project on Mental Healthcare institution,
during our conversation about the possibilities of detecting epileptic foci by utilization of a graph
theory method based on high density EEG, a colleague mentioned, that it maybe correspond with
the fact that in the last five years many patients who are suffering from depression are actually
having excellent reaction on anticonvulsants. Is it possible that they have a focus in their limbic
system? We discarded that idea as highly unrealistic, but after the finalizing results of several
projects on that topic, I think that they are reacting well because of their elevated excitability,
induced by their decreased functional connectivity in frontolimbic system. The same can apply to
the effectiveness of rTMS and tDCS in cases of treatment-resisting depression; maybe they are
effective because of their already demonstrated ability to decrease the levels of complexity in their
systems? We also know that the solution is alas, of a temporary nature. As patients reacting well
on electroconvulsive therapy need to have later maintenance ECT, it is also required in other forms
of electromagnetic stimulation. Both rTMS and tDCS last for a certain amount of time. They are
not capable of resolving the problem, and they can help improve the patient's state for a while. The
effect is also individual; whatever mode of stimulation we consider, we see that intricate technical
details (like the position of electrodes, the physical characteristics of a stimulator), or individual
differences (like the thickness of a cortex, or a bone, or different conductivity of a tissue) play their
role in final effect of the treatment. Another advantage of the therapeutic use of electromagnetic
stimulation is, that opposite to any medication they do not have side effects; or at least they
certainly cannot be (or are so far away of being) life-threatening.
Hence, this kind of detection (by utilization of nonlinear measures of electrophysiological
signals) are, according to our opinion, one of the prerequisites of a personal medicine in mood
disorders.
Conclusions
Asgdsfg
Acknowledgements
We want to express our gratitude to our colleagues from Italy Professor Carlo Miniussi, Debora Brignani,
Ph.D. and Maria Concetta Pelliciari, Ph.D. for sharing their EEG recordings (data are from the Pellicciari et al.,
2013. paper) with us and for all valuable discussions.
References
Eaton, W.W. et al. (2008). Population-based study of first onset and chronicity in major depressive disorder.
Arch. Gen. Psychiatry, 65, 513-520.
Fekadu A., et al., (2009) What happens to patients with treatment-resistant depression? A systematic review
of medium to long term outcome studies. Journal of Affective Disorders, 116, 4-11.Rush et al., 2006
Barker, A.T., Jalinous, R., Freeston, I.L. (1985). Non-invasive magnetic stimulation of human motor cortex.
Lancet, 325, 1106-1107.
Opitz, A., Paulus, W., Will, S., Antunes, A., Thielscher, A. (2015). Determinants of the electric field during
transcranial direct current stimulation. Neuroimage, 109, 140 -- 150. doi: 10.1016/j.neuroimage.2015.01.033
Opitz, A., Fox, M.D., Craddock, R.C., Colcombe, S., Milham, M.P., 2016b. An integrated framework for
targeting
functional networks via
transcranial magnetic
stimulation. Neuroimage 127, 86 -- 96.
https://doi.org/10.1016/j.neuroimage.2015.11.040.29.
Opitz, A., Yeagle, E., Thielscher, A., Schroeder, C., Mehta, A. D, Milham, M.P. (2018). On the importance
of precise electrode placement for targeted transcranial electric stimulation. Neuroimage, 181, 560-567. doi:
10.1016/j.neuroimage.2018.07.027.
Rosa, M. A., Lisanby, S. H. (2012). Somatic treatment for mood disorders. Neuropsychopharmacology, 37,
102-116.
O'Reardon, J.P., et al. (2007). Efficacy and safety of transcranial magnetic stimulation in the acute treatment
of major depression: a multisite randomized controlled trial. Biological Psychiatry, 62, 1208-1216.
Attkinson, L., Sankar, A., Adamas, T.M., Fu, C.H. (2014). Recent advances in neuroimaging of mood
disorders:structural and functional neural correlates of depression, changes with therapy, and potential for
clinical biomarkers. Current Treatment Options in Psychiatry, 1, 278-293.Allen, J. J. B., Urry, H. L., Hitt, S. K.,
Coan, J. A. (2004). The stability of resting frontal electroencephalographic asymmetry in depression.
Psychophysiology, 41(2), 269 -- 280. http://dx.doi.org/10.1111/j.1469-8986.2003.00149.x.
Rossi, S., Hallet, M., Rossini, P.M., Pascual-Leone, A., Group S.o.T.C. (2009). Safety, ethical consideration
and application guidelines for the use of transcranial magnetic stimulation in clinical practice and research. Clin.
Neurophysiology, 120, 2008-2039.
Conca, A., et al. (2002). Combining high and low frequencies in rTMS antidepressive treatment: preliminary
results. Hum. Psychopharmacol.: Clin. Exp. 17, 353-356.
Fertonani, A., and Miniussi, C. (2017). Transcranial electrical stimulation: what we know and do not know
about mechanisms. Neuroscientist, 23, 109 -- 123. doi: 10.1177/1073858416631966
Fitzgerald P.B., Borwn T.L., Marston NAU, Daskalakis ZJ, de Castella A, Kulkarni J. (2003) Transcranial
magnetic stimulation in the treatment of depression: a double-blind, placebo-controlled trial. Arch Gen
Psychiatry 60: 1002-1008.
Fitzgerald, P.B., et al., (2006). A randomized, controlled trial of sequential bilateral repetitive transcranial
magnetic stimulation for treatment-resistant depression. American Journal of Psychiatry, 163, 88-94.
Roth, Y., Amir, A., Levkovitz, Y., Zangen, A. (2007), Three-dimensional distribution of the electric field
induced in the brain by transcranial magnetic stimulation using figure-8 and deep H-coils. J. Clin. Neurophysiol,
24, 31-38.
Costafreda, S.G. et al. (2013). Modulation of amygdala response and connectivity in depression by serotonin
transporter polymorphism and diagnosis. Journal of Affective Disorders, 150, 96-103.
de Kwaasteniet, B., Ruhe, E., Caan, M., et al. (2013). Relation Between Structural and Functional
Connectivity
in Major
Depressive
Disorder.
Biological
Psychiatry,
74(1),
40 -- 47.
http://dx.doi.org/10.1016/j.biopsych.2012.12.024.
Huang, Y.Z., Edwards, M.J., Rounis, E., Bhatia, K.P., Rothwel, J.C. (2005). Theta burst stimulation of the
human motor cortex. Neuron, 45, 201-206.
Chung, S., Hoy, K., Fitzgerald, P. (2015). Theta-burst stimulation: a new for of TMS treatment for
depression? Depress. Anxiety, 32, 182-192.
Stagg, C. J., Nitsche, M. A. (2011). Physiological Basis of Transcranial Direct Current Stimulation. The
Neuroscientist, 17(1), 37-53.
Anderson P, Lomo T. 1966. Mode of activation of hippocampal pyramidal cells by excitatory synapses on
dendrites. Exp Brain Res 2(3):247 -- 60.
Bliss TV, Lomo T. 1973. Long-lasting potentiation of synaptic transmission in the dentate area of the
anaesthetized rabbit following stimulation of the perforant path. J Physiol 232(2):331 -- 56.
Jin, Y., Phillips, B. (2014). a pilot study of the use of EEG-based synchronized Transcranial Magnetic
Stimulation (sTMS) for treatment of major depression. BMC Psychiatry, 14, 1.
Lefaucheur, J. P., Andre-Obadia, N., Antal, A., Ayache, S. S., Baeken, C., Benninger, D. H., et al. (2014).
Evidence-based guidelines on the therapeutic use of repetitive transcranial magnetic stimulation (rTMS).
Clinical Neurophysiology, 125, 2150-206.
Nitsche, M. A., Cohen, L. G., Wassermann, E. M., Priori, A., Lang, N., Antal, A., Paulus, W., Hummel, F.,
Boggio, P.S., Fregni, F., Pascual-Leone, A. (2008). Transcranial direct current stimulation: State of the art 2008.
Brain Stimulation, 1(3), 206-23. doi: 10.1016/j.brs.2008.06.004. Epub 2008 Jul 1.
Nitsche, M. A., Fricke, K., Henschke, U., Schlitterlau, A., Liebetanz, D., Land, N., Hening, S., Terau, F.,
Paulus, W. (2003). Pharmacological modulation of cortical excitability shifts induced by transcranial direct
current stimulation in humans. Journal of Neurophysiology, 553, 293-301.
Mutz J., Edgcumble D.R., Brunoni A.R., Fu C.H.Y. (2018) Efficacy and acceptability of non-invasive brain
stimulation for the treatment of adult unipolar and bipolar depression: A systematic review and meta-analysis of
randomized sham-controlled trials. Neuroscience and Biobehavioral Reviews 92 (2018) 291-303.
Fonteneau C. et al., (2018) Frontal Transcranial Direct Current Stimulation Induces Dopamine Release in
the Ventral Striatum in Human. Cerebral Cortex, Volume 28, Issue 7, July 2018, Pages 2636 -- 2646,
https://doi.org/10.1093/cercor/bhy093
Antal A., Alekseichuk I., Bikson M. et al., (2017) Low intensity transcranial electric stimulation: Safety,
ethical, legal regulatory and application guidelines. Clin Neurophysiol. 2017 Sep; 128(9): 1774 -- 1809.Ruffini et
al., 2013b
Alam M, Truong DQ, Khadka N, Bikson M. 2016. Spatial and polarity precision of concentric high-
definition transcranial direct current stimulation (HD-tDCS). Physics in Medicine and Biology 61:4506 -- 4521.
doi: 10.1088/0031-9155/61/12/4506, PMID: 27223853
Radman T., Datta A., and Bikson M. (2009) One-dimensional representation of a neuron in a uniform
electric field. Conf Proc IEEE Eng Med Biol Soc. 2009 ; 2009: 6481 -- 6484. doi:10.1109/IEMBS.2009.5333586.
Ranieri F, Podda MV, Riccardi E, Frisullo G, Dileone M, Profice P, Pilato F, Di Lazzaro V, Grassi C.(2012).
Modulation of LTP at rat hippocampal CA3-CA1 synapses by direct current stimulation. Journal of
Neurophysiology 107:1868 -- 1880. doi: 10.1152/jn.00319.2011, PMID: 22236710
Frohlich F, McCormick DA. (2010). Endogenous electric fields may guide neocortical network activity.
Neuron 67:129 -- 143. doi: 10.1016/j.neuron.2010.06.005, PMID: 20624597
Miranda PC, Pajevic S, Pierpaoli C, Hallett M, Basser PJ. (2001). The distribution of currents induced in
the brain by magnetic stimulation: a finite element analysis incorporating DT-MRI-derived conductivity data.
Proceedings of the International Society for Magnetic Resonance in Medicine 9.
Miranda, P. C., Lomarev, M., Hallet, M. (2006). Modeling the current distribution during transcranial direct
current stimulation. Clinical Neurophysiology, doi:10.1016/j.clinph.2006.04.009
Nitsche MA, Paulus W. (2001). Sustained excitability elevations induced by transcranial DC motor cortex
stimulation in humans. Neurology 57:1899 -- 1901. doi: 10.1212/WNL.57.10.1899
Saturnino, G. B., Antunes, A., Thielscher, A. (2015). On the importance of electrode parameters for shaping
electric field patterns generated by tDCS. Neuroimage, 120,25-35.
Pelliciari, M. C., Brignani, D., Miniussi, C. (2013). Excitability modulation of the motor system induced by
transcranial direct current stimulation: a multimodal approach. Neuroimage, 83, 569-580.
Nelson D.A. and Nunneley S.A. (1998) Brain temperature and limits on transcranial cooling in humans:
quantitative modeling results. Eur J Appl Physiol (1998) 78: 353. https://doi.org/10.1007/s004210050431
Wagner T., Valero-Cabre A., and Pascual-Leone A. (2007) Noninvasive Human Brain Stimulation. Annu.
Rev. Biomed. Eng. 2007. 9:527 -- 65.
Fregni, F., Marcolin, M.A., Mychowski, M., Amaz, R., Hasey, G., Rumi, D.O., et al. (2006). Predictors of
antidepressant response in clinical trials of transcranial magnetic stimulation. International Journal of
Neuropharmacology, 9, 641-54.
Palm, U., Schiler, C., Fintescu, Z., Obermeier, M., Keeser, D., Reisinger, E. et al. (2012). Transcranial direct
current stimulation in treatment resistant depression: a randomized double-blind, placebo-controlled study. Brain
stimulation, 5, 242-51.
Berlim, M.T., Van den Eynde, F., Daskalakis, Z.J. (2013). Clinical utility of transcranial direct current
stimulation (tDCS) for treating major depression: a systematic review and meta-analysis of randomized, double-
blinded and sham-controlled trials. J Psychiatr Res, 47, 1-7.
Berlim, M.T., van der Eynde, F., Tovar-Perdomo, S., Daskalakis, Z.J. (2014). Response, remission and drop-
out rates following high-frequency repetitive transcranial magnetic stimulation 9rTMS) for treating major
depression: a systematic review and meta-analysis of randomized, double-blind and sham-controlled trials.
Psychol Med, 44, 225-39.
Bennabi, D., Nicolier, M., Monnin, J., Tio, G., Vandel, P. et al. (2015). Pilot study of feasibility of the effect
of treatment with tdcs in patietis suffering from treatment-resistant depression treated with escitalopram. Clinical
Neurophysiology, 126, 1185-1189.
Brunoni A.R. et al., (2016) Transcranial direct current stimulationfor acute major depressive episodes: meta-
analysis of individual patient data. The British Journal of Psychiatry 2016, 208,1-10. Doi: 10.1192/bjp.bp.115-
164715.
Brunoni, A. R., Nitsche, M. A., Bolognini, N., Bikson, M., Wagner, T., Merabet, L. et al. (2012). Clinical
research with transcranial direct current stimulation (tDCS): challenges and future direction. Brain Stimulation,
5, 175-195.
Loo C.L. et al., (2012) Transcranial direct current stimulation for depression; 3-wkke randomized, sham-
controlled trial. Br. J. Psychiatry 2012; 200:52-9.
Loo C.K. et al., (2010). A double-blind sham-controlled trial of transcranial direct current stimulation for
the treatment of depression. Int J Neuropsychpharmacol. 2010; 13: 61-9.
Arroll, B., Elley, C.R., Fishman, T., Goodyear-Smith, F.A., Kenealy, T., Balashki, G., et al. (2009).
Antidepressants versus placebo for depression in primary care. Cochrane Database Syst Rev, 3,
CD007954.Nahas et al., 2003
Li, L. M., Violante, I. R., Leech, R., Ross, E., Hampshire, A., Opitz, A., Rothwell, J. C., Carmichael, D. W,
Sharp, D. J. (2018). Brain state and polarity dependent modulation of brain networks by transcranial current
stimulation. Human Brain Mapping, 1-12. DOI: 10.1002/hbm.24420.
Bestmann, S., Baudewig, J., Siebner, H.R., Rothwell, J.C., Frahm, J. (2003). Is functional magnetic
resonance imaging capable of mapping transcranial magnetic cortex stimulation? Supplements to Clinical
Neurophysiology, 56, 55-62.
Bestmann, S., Baudewig, J., Siebner, H.R., Rothwell, J.C., Frahm, J. (2003). Subthreshold high-frequency
TMS of human primary motor cortex modulates interconnected frontal motor areas as detected by interleaved
fMRI-TMS. Neuroimage, 20, 1685-1696. DOI: 10.1016/j.neuroimage.2003.07.028
Bestmann, S., Baudewig, J., Siebner, H. R., Rothwell, J. C., Frahm, J. (2004). Functional MRI of the
immediate impact of transcranial magnetic stimulation on cortical and subcortical motor circuits. European
Journal of Neuroscience, 19, 1950-1962.
de Kwaasteniet, B., Ruhe, E., Caan, M., et al. (2013). Relation Between Structural and Functional
Connectivity
in Major
Depressive
Disorder.
Biological
Psychiatry,
74(1),
40 -- 47.
http://dx.doi.org/10.1016/j.biopsych.2012.12.024.
Kim D, Bolbecker AR, Howell J et al. Disturbed resting state EEG synchronization in bipolar disorder: a
graph-theoretic analysis (2013). NeuroImage Clin. 2013; 2: 414 -- 423. DOI: 10.1016/j.nicl.2013.03.007.
Alekseichuk I., Mantell K., Shirinpour S., Opitz A. (2018) Comparative Modeling of Transcranial Magnetic
and Electric Stimulation in Mouse, Monkey, and Human. bioRxiv preprint first posted online Oct. 13, 2018; doi:
http://dx.doi.org/10.1101/442426.
Huang, Y., Liu, A. A., Lafon, B., Friedman, D., Dayan, M., Wang, X., Bikson, M., Doyle, W.K., Devinsky,
O, Parra, L.C. (2017) Measurements and models of electroc fields in the in vivo human brain during transcranial
electric stimulation. eLife, 6:e18834. DOI: 10.7554/eLife.18834
Aydin, U., Vorwerk, J., Kupper, P., Heers, M., Kugel, H., Galka, A., Hamid, L., Wellmer, J., et al. (2014).
Combining MEG and EEG for the reconstruction of epileptic activity using a calibration realistic volume
conductor model. PlosOne, 9.
Saturnino, G. B., Antunes, A., Thielscher, A. (2015). On the importance of electrode parameters for shaping
electric field patterns generated by tDCS. Neuroimage, 120,25-35.
Čukić et al., (2018). EEG machine learning with Higuchi's fractal dimension and Sample Entropy as features
for detection of depressive disorder. https://arxiv.org/abs/1803.05985
Čukić, M., Kolar, D., Kajkut, S. (2001). Primena Transkranijalne magnetne stimulacije u tretmanu depresije.
[Application of Transcranial Magnetic Stimulation in treatments of Major Depression]. ME2.2, ETRAN 2001,
pp. 231-234.
Čukić, M., Kalauzi, A., Ljubisavljević, M. (2006). Uticaj rastojanja stimulacionog kalema i voljne aktivacije
na latencije mišićnih odgovora na Transkranijalnu magnetnu stimulaciju [The Influence of distance of
stimulation coil and voluntary activation on MEP latencies in TMS study]. ETRAN 2006, Beograd, book 3 ME2,
str. 252-255.
Čukić, M., Ljubisavljević, M., Kalauzi, A. (2009): Analysis of probability density distributions of TMS
induced
physiological
parameters, Clinical
neurophysiology, Vol
120.March
2009.
pg
e129.doi:10.1016/j.clinph.2008.12.013
Čukić, M., Kalauzi, A., Ilić, T., Mišković, M., Ljubisavljević, M. (2009). The influence of coil-skull distance
on transcranial magnetic stimulation motor evoked responses, Experimental Brain Research, 192, 1, 53-60.
Čukić, M., Filipović, S., Ljubisavljević, M. (2010). Transcranial magnetic stimulation effects on tremor in
Primary writing tremor patients. First International Workshop on "Synaptic Plasticity: from bench to bed side"
Taormina, Sicily, Italy, April 28.-May 1. 2010.
Čukić, M., Platiša, M., Ljubisavljević, M., Kostić, V. (2012). Complexity Changes in TMS Induced Surface
EMG. CCS2012, 4th International Interdisciplinary chaos symposium on chaos and complex systems, April 29-
May 02. 2012, Antalya, Turkey.
Čukić, M., Kalauzi, A., Ljubisavljevic, M., Jorgovanović, N., Kostić, V. (2012). The Evolution of
Complexity in Transcranial Magnetic Stimulation induced surface EMG: a possible illustration of plasticity-like
changes. The 16th Congress of the European Federation of Neurological Societies - EFNS 2012, Stockholm,
September 8-11. 2012.
Čukić, M., Oommen, J., Mutavdzic, D., Jorgovanovic, N., Ljubisavljevic, M. (2013). The effect of single-
pulse transcranial magnetic stimulation and peripheral nerve stimulation on complexity of EMG signal: fractal
analysis. Experimental Brain Research, 228(1), 97-104, DOI: 10.1007/s00221-013-3541-1.
Čukić, M. (2006). The Influence of configuration and geometric characteristics of induced electrical field
induced by transcranial magnetic stimulation on basic physiological parameters of excitability of motor cortex
of man. Magisterium thesis, defended on March 16th 2006. Department for Physiology and Biophysics,
University of Belgrade, School of Biology, Belgrade, Serbia.
Mutanen, T., Nieminen, J. O., Ilmoniemi, R. (2013). TMS-evoked changes in brain-state dynamics
quantified by using EEG data. Frontiers in Human Neuroscience, 25 April, 2013 Volume7, Article 155, doi:
10.3389/fnhum.2013.00155.
Ilmoniemi, R., Kičić, D. (2010). Methodology for combined TMS and EEG. Brain Topogr. 22, 233-
248.Nitsche, 2001
Čukić, M., Pokrajac, D., Stokić, M., Simić, S., Radivojević, V., Ljubisavljević, M. (2018). EEG machine
learning with Higuchi fractal dimension and Sample Entropy as features for successful detection of depression.
Arxive.org/Preprint at Cornell repository for Statistics/Machine learning https://arxiv.org/abs/1803.05985.
Thielscher A, Antunes A, Saturnino GB. (2015). Field modeling for transcranial magnetic stimulation: a
useful tool to understand the physiological effects of TMS? Engineering in Medicine and Biology Society
(EMBC), 2015, 37th Annual
International Conference of
the
IEEE:IEEE. p. 222 -- 225. doi:
10.1109/EMBC.2015.7318340
Bestmann S.(2008) The physiological basis of transcranial magnetic stimulation. Trends in Cognitive
Sciences. 12: 81-3. PMID 18243042 DOI: 10.1016/j.tics.2007.12.002
Amassian, V. E., Cracco, R. Q., Maccabee, P. J. (1989). Focal stimulation of human cerebral cortex with
the magnetic coil: a comparison with electrical stimulation. Electroencephalogr. Clin Neurophysiol, 74, 401-
416.
Amassian, V. E., Eberle, L., Maccabee, P. J., Cracco, R. Q. (1992). Modelling magnetic coil excitation of
human cerebral cortex with a peripheral nerve immersed in a brain-shaped volume conductor: the significance
of fiber bending in excitation, Electroencephalogr Clin Neurophysiol 85, 291 -- 301.
Amassian, V. E., Stewart, M., Quirk, Q. J. (1987). Physiological basis of motor effects of a transient
stimulus to the cerebral cortex. Neurosurgery, 20, 74-93.
Wagner S, Rampersad SM, Aydin U , Vorwerk J, Oostendorp TF, Neuling T, Herrmann CS, Stegeman DF,
Wolters CH. (2014). Investigation of tDCS volume conduction effects in a highly realistic head model. Journal
of Neural Engineering 11:016002. doi: 10.1088/1741-2560/11/1/016002, PMID: 24310982
Harris J.A., Clifford C.W.G. and Miniussi C. (2008) The Functional Effect of Transcranial Magnetic
Stimulation: Signal Suppression or Neural Noise Generation? MIT 2018, Journal of Cognitive Neuroscience
20:4, pp. 734 -- 740.
Karpyak, V. M., Rasmussen, K. G., Hammill, S. C., Mrazek, D. A.(2004). Changes in heart rate variability
in response to treatment with electroconvulsive therapy. .Journal of ECT 20(2):81-8.
Bozkurt A, Barcin C, Isintas M, Ak M, Erdem M, Ozmenler KN. (2013) Changes in heart rate variability
before and after ECT in the treatment of resistant major depressive disorder. Isr J Psychiatry Relat Sci.
2013;50(1):40-6.
Shorter Edward and Fink Max (2010) Endocrine Psychiatry: Solving the Riddle of Melancholia. Oxford
University Press, 2010.
Davidson, R. J. (2004). What does the prefrontal cortex "do" in affect: perspectives on frontal EEG
asymmetry
research?
Biological
Psychology,
67(1 -- 2),
219 -- 234.
http://dx.doi.org/10.1016/j.biopsycho.2004.03.008.
Slagter HA, Lutz A, Greischar LL, Francis AD, Nieuwenhuis S, Davis JM, Davidson RJ.(2007) Mental
training affects distribution of limited brain resources. Plos Biology. 5: e138. PMID 17488185 DOI:
10.1371/journal.pbio.0050138
|
1609.05730 | 1 | 1609 | 2016-09-19T13:56:01 | Multi-contact synapses for stable networks: a spike-timing dependent model of dendritic spine plasticity and turnover | [
"q-bio.NC"
] | Excitatory synaptic connections in the adult neocortex consist of multiple synaptic contacts, almost exclusively formed on dendritic spines. Changes of dendritic spine shape and volume, a correlate of synaptic strength, can be tracked in vivo for weeks. Here, we present a combined model of spike-timing dependent dendritic spine plasticity and turnover that explains the steady state multi-contact configuration of synapses in adult neocortical networks. In this model, many presynaptic neurons compete to make strong synaptic connections onto postsynaptic neurons, while the synaptic contacts comprising each connection cooperate via postsynaptic firing. We demonstrate that the model is consistent with experimentally observed long-term dendritic spine dynamics under steady-state and lesion induced conditions, and show that cooperation of multiple synaptic contacts is crucial for stable, long-term synaptic memories. In simulations of a simplified network of barrel cortex, our plasticity rule reproduces whisker-trimming induced rewiring of thalamo-cortical and recurrent synaptic connectivity on realistic time scales. | q-bio.NC | q-bio |
Multi-contact synapses for stable networks: a spike-timing dependent
model of dendritic spine plasticity and turnover
Moritz Deger1,2, Alexander Seeholzer1, Wulfram Gerstner1
1: School of Computer and Communication Sciences and School of Life Sciences, Brain Mind Institute,
École Polytechnique Fédérale de Lausanne, 1015 Lausanne EPFL, Switzerland
2: Institute for Zoology, Faculty of Mathematics and Natural Sciences,
University of Cologne, 50674 Cologne, Germany
Abstract
Excitatory synaptic connections in the adult neocortex consist of multiple synaptic contacts, almost
exclusively formed on dendritic spines. Changes of dendritic spine shape and volume, a correlate of
synaptic strength, can be tracked in vivo for weeks. Here, we present a combined model of spike-timing
dependent dendritic spine plasticity and turnover that explains the steady state multi-contact con g-
uration of synapses in adult neocortical networks.
In this model, many presynaptic neurons compete
to make strong synaptic connections onto postsynaptic neurons, while the synaptic contacts compris-
ing each connection cooperate via postsynaptic ring. We demonstrate that the model is consistent
with experimentally observed long-term dendritic spine dynamics under steady-state and lesion induced
conditions, and show that cooperation of multiple synaptic contacts is crucial for stable, long-term
synaptic memories. In simulations of a simpli ed network of barrel cortex, our plasticity rule reproduces
whisker-trimming induced rewiring of thalamo-cortical and recurrent synaptic connectivity on realistic
time scales.
Excitatory synapses on pyramidal cells of the mammalian neocortex are almost exclusively made onto
postsynaptic dendritic spines (see [1, 2]). Dendritic spines are small protrusions of the dendrite that vary in
size and shape, and are subject to ongoing plasticity in both the developing and the adult brain [3, 4, 5, 6, 7,
8]. A central role of dendritic spines may be that the compartment of the spine allows for localized, synapse-
speci c plasticity mechanisms [1, 2]. Furthermore, plastic dendritic spines may allow neurons to select the
subset of inputs from nearby axons of candidate presynaptic cells [9], so as to optimally process stimuli
under the constraints of limited total tissue volume available for synaptic connections [10, 11]. Although
dendritic spines show considerable volatility and may be formed and eliminated within a couple of days [12],
some spines are maintained over long periods of time [12] and may support lifelong memories [13]. During
learning, dendritic spines in the neocortex change selectively, and if potentiated spines are selectively shrunk
by optogenetic means, learning is disrupted [14].
Time-lapse imaging of changes of spines on time-scales of seconds, hours and days to weeks [15, 12,
16, 4, 5, 17, 18] has shown that spine removal is tightly linked to the dynamics of the spine volume and
to synaptic e cacy. Despite the ongoing dendritic spine turnover response patterns of neuronal networks
can be long-term stable [19]. While past modeling studies of synaptic consolidation indicate how memories
could be stored beyond the rst hour after plasticity induction [20, 21, 22, 23], spike-timing dependent
plasticity (STDP) based models of synaptic plasticity [24, 25, 26] have not yet been linked quantitatively to
dendritic spine plasticity and turnover [5], lifetime of synaptic contacts, and observed statistics of synaptic
contact numbers [27, 28].
Synapses between pyramidal neurons in the somatosensory cortex of the rat are sparse and the distribution
of the number of synaptic contacts for pairs of pre- and postsynaptic neurons has a characteristic bimodal
shape [27, 28, 29]: most presynaptic neurons make no contact at all with a given postsynaptic neuron,
but those that do are likely to establish 4 or more contacts [27]. However, the number of potential
1
synaptic contacts between a pair of neurons, based on the proximity of axons and dendrites in reconstructed
neuronal morphologies, is always unimodally distributed [28, 30, 29] with a most likely value of 1 and an
estimated mean value signi cantly greater (Fig. 1D) than the mean number of actual contacts observed
in experiments. This puzzling observation led to the hypothesis that the observed dierence between
potential and actual contact numbers cannot arise from independent synapse formation, but instead requires
a mechanism of cooperation between synaptic contacts onto the same neuron [28]. We wondered whether
the cooperativity between synaptic contacts would necessarily imply a novel biochemical mechanism or
whether known principles of STDP are su cient to explain the bimodal distribution of contact numbers in
sparsely connected, stable neocortical networks.
We propose a unifying approach to dendritic spine formation, plasticity and removal, in which STDP of
single synaptic contacts provides the crucial cooperative mechanism regulating structural plasticity. Synaptic
contacts connecting a neuron A to another neuron B share the same pre- and postsynaptic neuronal activity:
presynaptic action potentials (spikes) cause Glutamate release and eventually leave a biochemical trace at
each dendritic spine, while information about postsynaptic spikes may reach the same dendritic spines by
means of backpropagating action potentials [31, 32].Since in some preparations spine plasticity [33, 34] and
spine formation [18] are visible within less than a minute after stimulation, structural spine plasticity cannot
be assumed to be slower than STDP, in contrast to previous models [35, 36, 37, 38, 39, 40].
Here, we present a quantitative model that links continuous synaptic plasticity of dendritic spines to
discrete structural plasticity of synaptic contact formation and removal. We show that our model reproduces
experimentally observed steady-state properties of dendritic spine plasticity and turnover in the somatosen-
sory cortex of adult rodents. While existing theoretical models of the bimodal distribution of synaptic
contacts in networks are algorithmic [28, 30], our model suggests that STDP is su cient to generate the
bimodal distribution observed in experiments. Our model predicts that a synaptic contact that is part
of a connection consisting of multiple synaptic contacts is more stable than an isolated synaptic contact.
Moreover, a recurrent thalamo-cortical network model exhibits long-term stable structure despite ongoing
plasticity, and lesion-induced rewiring on time scales of days and weeks.
A combined model of STDP and structural plasticity
In a rst set of simulated experiments, we model the plasticity of the synapses from N = 1000 presynaptic
neurons onto a single postsynaptic neuron (Fig. 1A). All presynaptic model neurons re as independent
Poisson processes with a rate of 5/s. A connection from a presynaptic neuron j to the postsynaptic cell
may consist of several synaptic contacts k, with 1 ≤ k ≤ nj , where nj is the number of potential contacts
[28, 30]. Each of these contacts is described by a unit-less weight wjk . The total weight of the connection
from a presynaptic neuron j is given by the sum of the weights wjk over all its contacts k, wj =Pnj
k=1 wjk .
The contact weight wjk in our model describes how much the speci c contact k contributes to ring the
postsynaptic cell, and can be viewed as representing the dendritic spine volume which is in turn strongly
correlated to the AMPA receptor content in the postsynaptic density [5].
Our plasticity model (Fig. 1B) describes the temporal dynamics of the weight wjk of a synaptic contact
by the dierential equation
(1)
wjk (t) = F (Sj , Spost, zjk ) − αwjk (t),
t
d d
where F is a functional of the pre- and postsynaptic spike trains Sj and Spost, respectively, and of stochastic
synaptic transmission zjk at the contact. The parameter α > 0 describes a slow decay of the weight.
The model (1) is a local, spike-timing dependent plasticity rule that we imagine to be realized by the
biophysics of dendritic spines in combination with that of the presynaptic terminal.
In our model pre- or
postsynaptic spikes each leave an exponentially decaying trace (rjk and rpost, respectively) at the synaptic
contact (dendritic spine), with a short time constant τ (Fig. 1C). A third lter trace Cjk with a longer
time constant τslow tracks the product of rjk · rpost as a measure of the correlation of pre- and postsynaptic
spike trains. A fourth lter trace Rpost (with time constant τslow) tracks the postsynaptic ring rate. The
combination of these traces gives rise to the STDP model summarized by F in Eq. (1), see Methods for
Results
2
Figure 1: Model overview. A: A postsynaptic neuron (post) receives synaptic connections from several
excitatory neurons (pre). Each connection consists of multiple synaptic contacts. The synaptic e cacy
(weight) of contact k in connection j
is denoted as wjk . The sum of contact weights wjk is the total
weight wj of this synaptic connection. B: Individual contact weights wjk take continuous, positive values,
and change in time according to spike-timing dependent plasticity. A small weight wjk corresponds to
a dendritic spine with a small volume, or a thin spine, and a large weight wjk corresponds to a large, or
mushroom-shaped dendritic spine. If wjk is greater than 0 we call the contact an actual contact. In contrast,
if at any time wjk becomes 0, the contact is pruned and its weight is kept xed at wjk = 0. A contact with
wjk = 0 is called a potential (but inactive) contact. It may be transformed into an actual contact by setting
wjk to a positive value at random times, with a rate λc (creation rate). C: Components of the plasticity
model (top to bottom): presynaptic spike train Sj ; transmitted spike train Sjk at the contact (black, random
synaptic failures occur, indicated by black boxes), and its ltered trace rjk (green); postsynaptic spike train
Spost (black) and its ltered trace rpost (red); product (correlation) term rjk · rpost composed of pre- and
postsynaptic trace (blue); low-pass ltered trace Cjk of the product term rjk·rpost. Synaptic contact plasticity
is determined by Cjk . D: Reference distributions of the number of actual (red) [27] and potential (blue) [28]
synaptic contacts for pairs of neurons in the adult somatosensory cortex (recurrent connections of layer 5
pyramidal neurons, truncated to nj ≤ 10 and renormalized). The steady state distribution generated by our
model is shown in green (data pooled over 150 days of simulation); the distribution of potential contacts in
the model is matched to the blue line.
3
further details. Mathematical analysis shows that this plasticity rule establishes a target level for both ring
rate and spike-timing correlations (see Model analysis in Methods and Supplementary Information).
As a result of ongoing STDP, synaptic model weights can increase or decrease. As soon as a contact
weight wjk (t) decreases to zero in the simulation, its dynamics cease and it turns into a potential but
inactive contact. However, with a rate of λc = 0.019/day, each potential contact (wjk = 0) may randomly
be transformed into an active contact. In such an event the weight wjk of the newly created contact is set
to a low, non-zero value wc. After creation the contact weight remains at the value wc for the duration of
the period of grace [41] of τgp = 15 min, after which its dynamics again follow (1) (see Fig. 1B).
In our model, the number of potential synaptic contacts nj from a presynaptic neuron j to the post-
synaptic model neuron lies between 1 to 10, and is distributed as estimated for recurrent connections of
layer 5 pyramidal neurons in rat somatosensory cortex [28], so that the blue line in Fig. 1D represents the
distribution of potential contacts nj
in both model and experiment. The model is insensitive to the exact
maximum value of nj , and values of nj up to 20 [28] would give equivalent results, albeit at increased
computational cost.
In the following, we describe how our plasticity model turns potential contacts into
stable weights through the mechanism of STDP-mediated cooperation.
Steady state synapse dynamics under STDP-mediated cooperation
After an initial transient at the beginning of the simulation, the synaptic contacts uctuate around a steady
state. Sample trajectories of all contacts from a single presynaptic neuron to the postsynaptic cell are shown
in Fig. 2A-B. All initial contacts of this presynaptic neuron remain over the course of 150 days of simulation,
occasionally new contacts are formed, which mature or are quickly removed (see Fig. 2B). The ring rate of
the postsynaptic neuron, as well as the total synaptic weight and the average number of contacts are tightly
controlled by the plasticity rule (Fig. 2C and Methods). The average number of active contacts hNcontactsi,
corresponding to the spine density, is stationary in the model consistent with experiments [42, 12, 43, 5].
Within one day, the relative change of the contact weight ranges from −100% to 500% (Fig. 2D) but the
relative uctuations decrease with increasing contact weight, consistent with long-term time-lapse imaging
data of dendritic spine volume in vitro [4]. We quantify the statistics of contact weight changes by computing
the mean and standard deviation of the change within one day, as a function of the weight. Consistent
with experimental measurements of dendritic spine volume [4] the mean contact change (Fig. 2E, black)
is positive for an intermediate range of contact weights (spine volumes) and becomes negative for large
weights (large volumes), and the standard deviation of the changes (Fig. 2E, red) is rather homogeneous
for all weights (volumes) but increases slightly for very large weights. The mean change of contacts of very
small weight is negative in our model, in contrast to previous experiments [4], but this dierence might be
due to experimental di culties of observing very small spines (see Discussion).
As synaptic connections in our model consist of several synaptic contacts, we asked whether there is a
systematic relation between synaptic weight and the number of contacts in the steady state (Fig. 2F). Indeed
the synaptic weight is strongly correlated with the number of active contacts: Strong synaptic connections
consist of at least ve synaptic contacts, and connections with less than three contacts are physiologically
weak (Fig. 2F, red, left axis). Moreover, individual contacts in connections made of more than 3 active
contacts tend to be stronger than those made of less than 3 (Fig. 2F, blue, right axis).
We further assess the statistical properties of the steady state by multiple measures. First, the distribution
of active synaptic contact numbers per connection (Fig. 3A) is bimodal, in line with experimental ndings
[27, 28], cf. Fig. 1D. Second, the turnover ratio of synaptic contacts in the model is 0.176 ± 0.018/day
(mean ± STD) consistent with the values found experimentally in somatosensory cortex [12, 5]. Third, a
stable distribution of contacts weights (Fig. 3B) is formed.
The probability of a synaptic contact to survive for 8 consecutive days, irrespective of whether the contact
is newly created, weak or strong, depends strongly on the number of active contacts in the connection that
the contact is part of (Fig. 3C). In other words, contacts within a connection cooperate and stabilize each
other. Tracking of individual synaptic contacts that existed at t = 0 (Fig. 3D) reveals a time course of the
number of surviving synapses that is qualitatively consistent with long-term in-vivo imaging data of dendritic
spines from mouse neocortex [12, 44, 43].
In particular, connections that consist of several contacts are
stable for long periods of time (here 150 days) (Fig. 3E inset). The pruned connections are almost exclusively
connections that consist of less than four contacts and have relatively small total weight (Fig. 3E). Finally,
4
Figure 2: Dynamics of synaptic contacts in the steady state. A: Example synaptic connection number
j = 7, colored lines each correspond to a contact weight wjk over time. New contacts ( lled circles) are
created with a weight given by the lower dashed line. Long-term stable contacts uctuate around the upper
dashed line, which is the xed point of w∗/5 predicted by theory (see Supplementary Information). B:
Zoom into A at the time of creation and pruning of two transient synaptic contacts. C: Firing rate Rpost of
postsynaptic neuron (black), total synaptic input w =PjPk wjk summed over all presynaptic neurons and
contacts (red, unit-less) and average number of active contacts per synaptic connection (green) over time,
each sampled in intervals of 6 hours. D: Relative changes of synaptic contact e cacy ∆wjk within one
day, each dot corresponds to the change of one contact in one day, data pooled over all contacts and 150
days of simulation (sampled in intervals of 2 days). The horizontal line of dots at ordinate value −1 is due
to contacts that were removed within one day. E: Weight dependence of contact changes within one day.
Mean µ and standard deviation σ of the change of contact weight ∆wjk within one day, as a function of the
contact weight wjk . Both µ and σ were estimated from the data shown in D, grouped into wjk -intervals as
indicated (solid lines). Estimates of µ and σ from less than 5 samples were discarded. Error bars denote the
standard error of the mean (SEM). For large contact weights σ increases signi cantly (∗ indicates p < 0.1
in Welsh's two-sided t-test). F: Total synaptic connection weight wj =Pk wjk (left axis, red) and contact
weight wjk (right axis, blue) as a function of the number of actual contacts (wjk > 0), averaged across all
synaptic connections. Error bars denote the standard deviation (STD).
5
Figure 3: Statistics of synaptic contacts in the steady state. A: Histogram of the number of active
contacts (wjk > 0) of a connection, across all connections j. B: Histogram of contact weights wjk , across
all connections j and contacts k. The distribution at day 0 is not signi cantly dierent from day 100 (two-
sample Kolmogorov-Smirnov test (2sKS), p-value 0.910). C: Fraction of synaptic contacts that survive for
8 days, as a function of the number of actual contacts in the connection, for newly created contacts (gray
bars), and existing contacts of weak or strong weight wjk (colors, see legend) in units of 10−3. Survival
fractions are estimated separately per day and then averaged over days. Estimates from less than 5 samples
are discarded; error bars denote SEM. D: Fraction of surviving actual synaptic contacts that were present at
time t = 0 (black: model) in comparison to published experimental dendritic spine survival data in mouse
(Hlt '05 S1: somatosensory cortex L5B, age 6 months (exponential decay t) [12]; Hlt '05 VC: visual cortex
L5B, age 3-6 months (exponential decay t) [12]; Lws '15 AC: auditory cortex [43]). E: Histogram of the
total weights wj of the connections present at t = 0 (black) and of the surviving connections at day 150
(green).
Inset: Histogram of the number of contacts of the connections. The removed connections have
small total weight and number of contacts; strong connections with many contacts are rarely removed. Note
that new contacts created in the meantime are not considered in this analysis. F: Lifetime of entire synaptic
connections (dots) during the course of the simulation, as a function of the total connection weight.
6
the lifetime of synaptic connections increases strongly with the total weight of the connection (Fig. 3F),
indicating that strong connections are protected against spine turnover by mutual cooperation of contacts.
Presynaptic lesions lead to increased formation of persistent contacts
Having established that the steady state of our model has characteristics consistent with experimental
data, we investigate whether the model can explain experimentally observed transient dynamics of dendritic
spines in response to changes of neuronal input. Experimentally, such input changes may be due to trimming
whiskers [42, 16], lesions of the retina [44] or occlusion of one of the eyes [45]. Here we model an abstract
lesion experiment in which a fraction of the input neurons suddenly cease to re (Fig. 4A).
In our lesion model, after ablating 50% of those presynaptic neurons that have active synaptic contacts,
we observe a loss of 50% of the synaptic contacts on the postsynaptic neuron after 30 minutes (Fig. 4B).
However, this loss is rapidly compensated such that the ring rate and total weight (summed over all
presynaptic neurons and all synaptic contacts) are hardly changed throughout the process (Fig. 4B,C). The
compensation occurs on two dierent time scales. First, on the time scale of 10 - 30 minutes, existing
synaptic contacts are up-regulated from a pre-lesion value between (3.3 ± 1.3) · 10−3 to a value of (6.6 ±
1.3) · 10−3 measured 30 minutes after the lesion (Fig. 4D). Second, on the slow time scale of 10-30
days after the lesion, the number of presynaptic neurons without postsynaptic contact decreases from
886 pre-lesion to 810 thirty days after lesion while the number of presynaptic neurons with one, two, or
three contact points transiently increases (Fig. 4E) suggesting that the plasticity rule 'tests' new connection
patterns. Competition between synaptic contact points from dierent presynaptic neurons and simultaneous
cooperation of synaptic contact arising from the same presynaptic neurons leads to pruning or strengthening,
so that after 99 days the distribution of contact numbers and synaptic weights is again very similar (but
not yet completely identical) to the distributions pre-lesion (Fig. 4E). The gradual recovery of the contact
numbers is due to an elevated probability of newly formed contacts to survive and become long-term stable
compared to the control condition (Fig. 4G). In a simulated lesion experiment where 20 percent of the active
contacts are removed (instead of 50 percent in the simulations so far), 14.6 percent of newly created contacts
survive for 8 days or more, consistent with experimental results on dendritic spines in the somatosensory
cortex after whisker trimming [16] and signi cantly above the 7.7 percent of surviving contacts in the control
condition (Fig. 4G).
Synaptic stability and reliability are due to contact multiplicity
To investigate the role of multiple synaptic contacts in our model, we compare its dynamics to the same
model restricted to single-contact connections (nj = 1 for all j).
In the multi-contact model above, with
N = 1000 presynaptic neurons there are 4633 potential synaptic contacts in total because every presynaptic
neuron makes 4.633 potential contacts on average (Fig. 1D, blue line). To maintain the total number of
potential synaptic contacts in the system with single-contact connections we therefore increase the number
of presynaptic neurons to N = 4633, keeping all other parameters the same. Similar to the full model,
the single-contact model exhibits steady state dynamics (Fig. 5A) in which postsynaptic ring rate, total
weight and synaptic contact number are tightly regulated (data not shown), and the distribution of contact
weights is stable over time (Fig. 5B). However, in contrast to the multi-contact model (Fig. 3), in the
single-contact model the temporal dynamics of synaptic contact survival do not indicate the presence of a
subgroup of stable connections (Fig. 5C), and all connections are prone to random pruning, irrespective of
whether they have a small or large synaptic weight (Fig. 5D,E). Thus, multiple synaptic contacts are crucial
for the long-term stability of strong synaptic connections, through mutual cooperation of the contacts from
the same presynaptic neuron: in multi-contact synaptic connections contacts support each other by exciting
the postsynaptic neuron more reliably despite random synaptic transmission failures.
A theoretical analysis further explains the crucial role of contact multiplicity in synaptic transmission.
We characterize the delity of transmission of a presynaptic spike by the ratio of the (trial-averaged) mean
and standard deviation of the evoked postsynaptic potential (PSP), which we call the signal-to-noise ratio
(see Methods).
In our model, failures of synaptic transmission occur randomly at each synaptic contact,
which causes variability of the PSP. If a synaptic connection consists of several contacts, a presynaptic spike
is more reliably transmitted than in the case of a single contact (Fig. 5F). If 10 contacts are considered to be
7
Figure 4: Rewiring in response to input lesion. A: Schematic of the simulated lesion experiment.
A substantial fraction of the presynaptic neurons that have actual synaptic contacts (wjk > 0) onto the
postsynaptic cell are ablated (set to very low ring rate of 0.1/s; each connected neuron is ablated with
probability plesion = 0.5; unconnected neurons are unaected) at t = 0 (vertical dashes in B-C). B-C: Firing
rate Rpost of postsynaptic neuron (black), total synaptic input w =Pj,k wjk summed over all presynaptic
neurons and contacts (red, unit-less) and average number of contacts per synaptic connection (green) over
time. (cf. Fig. 2C). After the lesion the average number of contacts (green) quickly drops by about 50%
(B), and gradually recovers towards the steady state afterwards (C). D-E: Histograms of contact weights
wjk (top) and of the number of active contacts (bottom), across all connections and contacts, before and
after the lesion. Half of the actual synaptic contacts are removed within 30 minutes after the lesion (D,
bottom, green), while the remaining half of the contacts potentiates (D, top, green). Within 99 days the
distributions gradually recover (E). F: Example synaptic connection from presynaptic neuron with index
j = 18, each colored line corresponds to a contact weight wjk over time, as in Fig. 2A. G: Fraction of
newly created persistent contacts (which survive for at least 8 days as de ned in [16]), control case (blue,
cf. Fig. 2) versus lesion model (red). Two lesion models are shown (red bars), marked with their respective
value of plesion. For comparison, data from mouse whisker trimming experiments [16] is shown (open bars,
error bars denote STD over observed cells). A smaller lesion with plesion = 0.2 is in better agreement with
the experimental data.
8
Figure 5: Single-contact synaptic connections are not long-term stable and have less delity. A-E:
Simulation of the plasticity model with only one potential contact per connection (nj = 1 for all j). A:
Example synaptic connection number j = 1392, blue line corresponds to contact weight wj1 over time, as
in Fig. 2A. B: Histogram of contact weights wj1, across all connections j. A steady state is maintained for
100 days, just as in the case of multiple potential contacts (cf. Fig. 3B). C: Fraction of surviving actual
synaptic contacts (solid line) that were present at time t = 0 in comparison to the multi-contact model
(dashed line) of Fig. 3D. D: Histogram of the weights wj = wjk of the connections present at t = 0 (black)
and of the surviving connections at day 100 (green). Connections of small and large weights are removed
unspeci cally. E: Lifetime of entire synaptic connections during the course of the simulation, as a function
of the total connection weight. F: Theoretical signal to noise ratio (SNR) of the postsynaptic potential
(13) in response to a presynaptic spike, in connections with multiple actual contacts and stochastic synaptic
failures (probability pf). Dashed line indicates the number of actual contacts for which a SNR of 80% of
the maximum is achieved, if 10 contacts is considered to be the maximum.
9
the maximum number of active contacts connecting a pair of neurons due to geometrical constraints of the
tissue, about 6 contacts give a signal to noise ratio of 80% of the maximum value achievable (irrespective
of the synaptic failure rate pf). Note that this number coincides with the typical peak of the model
contact numbers in the steady state (cf. Fig. 3A). Indeed, high reliability and low variability of synaptic
transmission in layer 5 neurons of the somatosensory cortex of the rat have been observed in glutamate
uncaging experiments in acute slices in vitro, and were attributed to the likely presence of multiple synaptic
contacts per connection [46].
Network simulations explain reinnervation of input-deprived cortical barrels
We wondered whether the structural plasticity rule discussed above would allows us to make predictions
of structural changes in a large recurrent network of excitatory and inhibitory neurons. We use a network
architecture (Fig. 6A) inspired by rodent barrel cortex with three strongly connected cortical populations
(representing the columns corresponding to dierent barrels) of excitatory neurons (exc 1-3), each one
preferentially innervated by a thalamic population (tha) that conveys sensory input from one of three
whiskers (whi) by strong connections.
In the model, connections between dierent cortical populations
and from thalamic populations to non-preferred cortical barrel columns are random and weaker on average
(Fig. 6D top left), but all excitatory connections whether strong or weak are subject to the same spike-timing
dependent structural plasticity model. A short movement ( ick) of the whisker is represented in the model
by a small increase in the ring rate of the thalamic neurons which results in turn in a modest increase of the
ring rate of the corresponding cortical population (Fig. 6B) riding on top of a spontaneous network activity
of about 5Hz (Fig. 6D bottom left). After an initial transient of 7 days of simulated time, we followed the
mean connection weights of synapses from tha 3 to exc 3, from tha 2 to exc 3, from exc 3 to exc 3, and
from exc 2 to exc 2 during 3 days of simulated time and found no changes (Fig 6C, top), indicating that
the average connectivity pattern is globally robust during ongoing activity and random whisker stimulation,
despite the fact that the structural plasticity rule is always active.
After 10 days of simulated time, the whisker corresponding to barrel 3 is trimmed. Whisker lesion is
modeled by (i) absence of whisker icks in tha 3 while stimulation of tha 1 and tha 2 continues; and (ii)
an exponential decrease of the ring rate of the corresponding thalamic population with a time constant of
5 min to a new baseline level of 0.1Hz. We found that the spiking activity of the network after the lesion
remains asynchronous and irregular (Fig. 6B). Within 3 days after the lesion, the recurrent connectivity of
the barrel column 3 has increased (Fig. 6C, center) consistent with a recent experiment [47]. The average
weight of connections within barrel column 2 is hardly aected by the lesion, but that within barrel column
3 changes substantially. The lateral connections from excitatory neurons of barrel column 2 to 3 and vice
versa increase on average. Similarly, the average connection weight of the non-preferred pathway from tha
2 to exc 3 increases whereas the connections in the preferred pathway disappear after the lesion (Fig. 6C,
center). We followed the synaptic changes for a total of 50 days after the lesion. Further changes were
smaller in magnitude but indicate that the recurrent network slowly continues to reorganize itself into a new
connectivity pattern. In particular, the average connectivity from excitatory neurons in barrel column 3 to
those in 2 continue to increase.
The increase in the average connection weight of the non-preferred pathway (from tha 2 to exc 3)
during the rst three days after the lesion could indicate a small increase of all the existing connections, or
a stronger increase of a subset of the feed-forward connections. A careful look at the detailed connectivity
pattern in Fig. 6D indicates that the latter is the case. Indeed, a subset of neurons located in the excitatory
population corresponding to barrel column 3 has become sensitive to stimulation of whisker 2. We rst used
a clustering algorithm to identify this subset of neurons and then reordered, and color-coded, the excitatory
neurons in barrel column 3 according to their responsiveness (yellow shade: newly responsive to whisker
1; red shade: newly responsive to whisker 2; green shade: responsiveness unchanged). The relevance of a
reorganization into subsets manifests itself in four dierent, but consistent ways: (i) the subset of red-shaded
neurons in population exc 3 responds more strongly to stimulation of whisker 2 than the average of neurons
in exc 3 and nearly as strongly as neurons in the barrel column 2 even though the duration of the response
is a bit shorter (Fig. 6E, bottom); (ii) the same subset of red-shaded neurons in population exc 3 has a
larger fraction of strong lateral connections to excitatory neurons in barrel column 2 than other neurons in
exc 3; (iii) the same subset of red-shaded neurons in population exc 3 receives a larger fraction of strong
10
Figure 6: Structural plasticity in a thalamo-cortical network. A: Schematic of the model. Thalamic
neurons (tha) convey sensory input from whiskers (whi) to the recurrent cortical network. Each tha pop-
ulation (squares) projects to one of three cortical barrels (circles) of excitatory (exc) neurons. Cortical
inhibitory (inh) neurons connect randomly to all barrels. Exc to exc synapses (dotted arrows) are modeled
by the structural plasticity model, all other synapses (solid arrows) are static. Plastic exc connections are
initialized as follows: Thin arrows wij = 0; thick arrows wij = w∗ ( xed point weight, see Supplementary
Information).
Inh synapses have a constant weight and no synaptic failures. Tha neurons re as Poisson
processes but increase their ring rate transiently if the corresponding whisker is icked, which happens ran-
domly with rate 1/s. After 10 days of simulation whisker 3 is trimmed (lesion), modeled as a progressive
loss of ring of tha 3 neurons (white cross). B: Spike raster plot around time of lesion (dashed vertical
line), colors as in A. If a whisker is icked (whi, triangles), the corresponding tha and exc populations
respond. C: Relative changes of average connection weights h∆wiji/hwiji between populations. Changes
before (top), around time of lesion (center) and long after (bottom). Strong changes during the rst three
days post lesion (center) are followed by a slow restructuring process of exc 3 over the following 47 days
(bottom). D (top): Exc synaptic connection weights (greyscale, wij ) just before lesion (left) and 50 days
after lesion (right). The comparison of the connection weight matrix before and after lesion shows that
loss of tha input to exc 3 caused selective rewiring. Exc 3 neurons (401-600) have been clustered in both
graphs according to the inputs they receive at simulation end (see Methods, assignments are indicated by
shading). D (bottom): Average spike response of exc 2 (black) and exc 3 (blue) neurons in response to
whisker 2 icks, just before trimming of whisker 3 (left) and at simulation end (right) (averaged over 60min
of recording). Red and yellow colors denote subgroups of exc 3 identi ed by clustering. Dashed lines: mean
ring rate in recording episode.
11
Figure 7: Restructuring can be predicted from initial responses / eect of bounded weights. A:
Spike rate in response to whisker 2 icks for each neuron of exc 3, before lesion (black dots) and 50 days
post (colored squares), for the thalamo-cortical network simulation shown in Fig. 6. Neurons are ordered
according to their pre-lesion response (black). Rates are estimated by counting spikes in a time window of
25ms after each whisker ick. The 100 exc 3 neurons that initially respond strongest to whi 2 are more
likely to increase their response to this whisker through structural plasticity (dashed horizontal lines with
error bars show mean response 50 days post ± SEM), and more likely to participate in the cluster that is
most strongly innervated by this whisker (red; colors correspond to the cluster assignments of the neurons
in Fig. 6D, right). B: Rewiring in response to input lesion (as in Fig. 4) with an upper bound of the contact
weight so that contact weights cannot grow stronger than twice the xed point value (wjk ≤ 6.4· 10−3). In
this case postsynaptic rate and total weight initially decrease in response to the lesion, and recover on a time
scale of about 10 hours as new contacts are formed. Simulation data are averaged over consecutive time
windows of 1 hour. C: For comparison, the data of the lesion simulation without upper bound of Fig. 4B,C
is replotted as in (B). Here the remaining contacts quickly (within less than 1 hour) compensate for the loss
of input by strongly increasing their weights, so that the postsynaptic rate shows no visible change.
lateral connections from excitatory neurons in barrel column 2 than from neurons in barrel column 1; (iv)
the same subset of red-shaded neurons in population exc 3 receives a larger fraction of strong feed-forward
connections from thalamic neurons in group 2 than other neurons in exc 3. Taken together, these four
observations suggest that the subset of red-shaded neurons in barrel column 3 has been integrated in the
information processing stream of barrel column 2. The same observations can be repeated for the yellow-
shaded subgroup of neurons in barrel column 3, except that these neurons have been integrated into barrel
column 1.
In both cases the integration has been made possible by structural plasticity triggered by the
lesion.
Our simulation results of point (i) above are consistent with experience-dependent receptive eld plastic-
ity found experimentally in pyramidal neurons in mouse somatosensory cortex 3-4 days [42] or 20 days [48]
after whisker trimming, where neurons that were part of a deprived barrel became responsive to the rst-
order surrounding whisker, in particular the subset of neurons located in the border region to the neighboring
barrel [48]. Observations (ii) - (iv) listed above are predictions of our model. Note that in our simulations,
both barrel columns exc 1 and exc 2 are rst-order surrounding columns of the deprived column (exc 3)
since we have not introduced any distance-dependent connectivity. We emphasize that the parameters of
the structural plasticity algorithm are kept xed throughout the simulation, be it before, during, or after the
lesion: First the network connectivity was stationary before the lesion, then it changed signi cantly during
3 days after the lesion, and nally settled into a new state (Fig. 6C) while structural plasticity has always
been 'turned on'. Indeed, individual synaptic contact points continue to grow or disappear even during the
phases where the coarse connectivity pattern remains unchanged; cf. Fig. 3.
Discussion
In this study we linked structural dynamics of synaptic contacts in neuronal networks to spike-timing de-
pendent plasticity (STDP). The implicit coupling between synaptic contacts onto the same postsynaptic
neuron through backpropagating action potentials is su cient to make synaptic contacts from one presy-
naptic neuron compete with contacts of other presynaptic neurons and cooperate with contacts of the same
12
presynaptic neuron. The resulting high-dimensional non-linear dynamical system has a steady state with
properties consistent with those of excitatory synapses in sensory cortices in terms of (i) a bimodal distri-
bution of contact numbers per synaptic connection; (ii) dierences in lifetime between strong and weak
connections; and (iii) turnover of dendritic spines.
Our model makes at least eight novel predictions.
A) A synaptic contact with a given weight is more stable if it is part of a group of four or ve synaptic
contacts arising from the same presynaptic neuron than if it is isolated or part of a group of only two synaptic
contacts (Fig. 3C).
B) The combination of the known result that strong contacts are more stable than small ones (cf. Fig 2D
and [4, 12]) with point A, yields the prediction that connections with larger total synaptic weight (summed
over all synaptic contacts) are more stable than small ones (Fig. 3F). This could be measured by correlating
the survival time of a synaptic contact with the total EPSP amplitude of the connection.
C) We predict a substantial fraction of synaptic connections that have only one active contact (see
Fig. 3A). These synaptic contacts, however, are small and quickly removed (see Fig. 2F). Since weight, PSP
amplitude, volume, and size are tightly correlated [5], these weak synapses might escape electrophysiological
or visual detection with standard methods which could explain the dierences to experimental reports
[27, 42], but might be detectable using sub-diraction resolution imaging [49, 40].
D) After a lesion, the number of connections consisting of exactly two synaptic contacts increases
transiently. While this might be expected since the process of building a novel synaptic connection with
ve contacts has to pass through a transient state with only two contacts, the prediction is that only about
a quarter of the two-contact connections actually stabilize to a multi-contact synapse, while the majority
disappears again. As a consequence, the number of presynaptic neurons without a connection transiently
decreases after a lesion before it increases again to a stable value (Fig. 4D and E).
E) After whisker trimming, the subset of neurons of the deprived cortical column which will be integrated
in the signal processing of an adjacent whisker will establish stronger incoming and lateral projections to
the cortical column in which they become integrated than other neurons in the deprived column (Fig. 6D).
F) The subset of those neurons of the deprived cortical column that are integrated into a new whisker
processing stream do not have to be physical neighbors but can be identi ed as those who, before the
trimming, had already a stronger response to the adjacent whisker (Fig. 7A).
G) The fact that, given a connection, the number of synaptic contacts per pre-post pair in experiments
peaks around a nite number (e.g. ve contacts) well below the maximal number of potential contacts
suggests a novel principle of synaptic plasticity (learning rule) which normalizes the amount of pre-post
correlations. This opens a gateway to a new class of learning rules in unsupervised learning which do not
maximize second-order correlations (as done by Oja's rule [50]) but normalize these. While experimentally
known homeostasis has focused on the normalization of mean ring rate [51] we suggest that it also is worth
while to study a normalization of correlations.
H) Our learning rule assumes that synaptic contact plasticity strongly depends on local traces of cor-
relations that must be processed in the dendritic spines. Therefore, we expect experimental manipulations
of biophysical activity traces, such as the calcium concentration, to crucially in uence the development of
dendritic spines.
In line with this hypothesis, eects of local calcium manipulations on dendritic lopodia
have indeed been observed [52]. A recent study on in Drosophila larvae further emphasizes the role of
presynaptic neurotransmission for the maturation of synaptic terminals [53].
We combined a speci c choice of an STDP model with a point neuron model, but several extensions
are possible. First, the synaptic plasticity rule used in this paper could also produce branch speci c synaptic
plasticity and steady-state con gurations [54], by choosing a more complex neuron model with non-linear
dendritic branches. Branch speci city of the spine dynamics could further be increased by using a voltage-
triggered rule [55] instead of a pure STDP paradigm. Second, while the speci c choice of STDP rule used
in this paper shows, at low frequencies, a symmetric pre-post and post-pre learning window for LTP we can
extend our formalism to more realistic STDP models [56, 57]. Third, we chose a learning rule where the
evolution of the synaptic contact weights is formally unbounded, but it is straightforward to also include an
upper bound. In our model, an explicit upper bound is not necessary (but see end of this section) because
synaptic depression limits further growth of contact weights as soon as the total weight of a connection
becomes strong, or individual synaptic contacts get large (see Eq. 7 in Methods and Fig. 2E) [58]. For
13
strong weights our learning rule eectively turns from a Hebbian rule to an anti-Hebbian rule, in the sense
that a further increase in correlations of pre- and postsynaptic ring leads to a shrinkage of synaptic weight.
Similar principles might explain why, depending on experimental preparations, variable STDP rules have been
reported in experiments [59]. Fourth, whereas for our recurrent network simulation of Fig. 6 we have used
leaky integrate-and- re neurons, for the mathematical analysis of the system dynamics we chose to describe
the activity of the postsynaptic neuron by a Poisson process with a linearly modulated rate. However,
because the activity of the postsynaptic model neuron in our simulations uctuates only in a limited range
(see Fig. 2C, black), we expect very similar mathematical results in the case of a non-linear neuron linearized
around the operating point.
Previous structural plasticity models of lesion-induced rewiring assumed a homeostasis of the postsy-
naptic ring rate [37] or spike-timing dependent structural plasticity [35, 38]. Both of these approaches,
however, are restricted to structural changes, and do not consider the combined dynamics of continuous
weights and discrete structural modi cations, which underlie the concurrent up-regulation of weights in
response to the lesion in our model (Fig. 4B). Moreover, previous studies did not consider the interplay
of synaptic competition between connections, and cooperation within connections, because they either fo-
cused on a single presynaptic neuron with multiple contacts [36, 39], multiple presynaptic neurons with
single contacts each [35, 38, 60], or non-competitive, purely homeostatic synaptic dynamics [37, 61].
Previous non-structural models of lesion-induced synaptic plasticity highlighted the time scale of home-
ostasis over hours or days [62].
In our model, lesion-induced changes of synaptic contact weights occur
rather quickly. While the average synaptic contact number recovers slowly within about 20 days, within
less than 10 minutes after the lesion the contact weights from spared presynaptic neurons are upregulated
to compensate for the loss of input (see Figures 4B-D and 6C). In apparent contrast, homeostatic synaptic
scaling in response to experimental blocking of postsynaptic ring occurs on the time scale of hours [51] or
in response sensory deprivation (likely to correspond to a reduction in presynaptic ring) on the time scale
of days [62]. In principle, this discrepancy in the speed of the readjustments of the synaptic weights might
be alleviated by low-pass ltering the model's weight changes (1) on a time scale of several hours before
applying them to the contacts, but that would make plasticity in general unrealistically slow.
Instead we
propose to extend the model by a combination of hard bounds [63] and multiplicative interaction of Hebbian
and explicit homeostatic processes [62]. With hard bounds set to two times the xed point weight, the
stationary distributions and results of Figs. 1-3 remain unchanged while the recovery of the ring rate after
a lesion is slower since individual synaptic contacts cannot grow beyond the hard bound (see Fig 7B). A
more complete model could combine the hard bounds with the essential features of the model of Toyoizumi
et al [62]. More generally, the interactions of our rule with manifold plasticity mechanisms such as short-
term plasticity [23], inhibitory plasticity [64], homeostasis [51, 37, 62], or intrinsic excitability [65, 60] pose
interesting challenges for future research.
Methods
Spike trains and failures of synaptic transmission.
j , t2
j ), where {t1
The activity of the presynaptic neurons 1 ≤ j ≤ N in our model is described by the spike trains Sj (t) =
Pm δ(t−tm
j , ...} are the spike times of neuron j. All presynaptic spike trains are generated by
independent Poisson processes with a constant ring rate of νpre = 5/s. Not all presynaptic spikes, however,
are transmitted at each synaptic contact that connects neuron j with the postsynaptic neuron. Some spikes
are not transmitted due to random transmission failures [66]. Therefore the spike train transmitted at the
contact k is given by
Sjk (t) = Xm
δ(t − tm
j ) zjk (tm
j ) .
(2)
The spike train Sjk (t) = Sj (t) · zjk (t) diers from the presynaptic spike train Sj through multiplication
with independent Bernoulli random variables zjk (t) ∈ {0, 1} that describe the stochastic failures of synaptic
transmission. In our model, synaptic failures occur randomly and independently with a probability of pf =
0.5, so zjk (tm
j ) is 1 (successful transmission) with probability 1 − pf = 0.5. The postsynaptic neuron, in
14
turn, emits the spike train Spost(t) =Pm δ(t − tm
on the postsynaptic neuron model are given below.
post), where {t1
post, t2
post, ...} are the spike times. Details
Slow and fast traces of pre- and postsynaptic activity.
At each synaptic contact, traces of pre- and postsynaptic activity are formed, which are illustrated in Fig. 1C.
The variables rjk (t) and rpost(t) describe low-pass lters of the pre- and postsynaptic activities Sjk (t) and
Spost(t), de ned by the dierential equations
(3)
(4)
rjk (t) = −rjk (t) + Sjk (t) ,
t
rpost(t) = −rpost(t) + Spost(t) ,
d d
τ
t
d d
τ
with a time constant τ = 20 ms, as is typical for an excitatory postsynaptic potential (EPSP) or an STDP
window function.
As an estimate of the correlations of pre- and postsynaptic ring on a slower time-scale, each synaptic
contact computes Cjk , as well as a slow trace of the postsynaptic activity Rpost. These traces are de ned
by the dierential equations
(5)
(6)
Cjk (t) = −Cjk (t) + rjk (t) · rpost(t)
t
Rpost(t) = −Rpost(t) + Spost(t)
d d
τslow
t
d d
τslow
with a time-constant of τslow = 1 min. The choice of Eqs. (3), (4) and (5) leads to a symmetric STDP
window, but the formalism can be extended to more standard STDP with long-term potentiation and
depression by using more traces with dierent time constants.
Synaptic contact plasticity.
Synaptic contacts in our model follow a variant of spike-timing dependent plasticity (STDP), see also Fig. 1.
Each contact is described by its e cacy (weight) wjk , which is the (unit-less) amplitude of the excitatory
postsynaptic potential that the contact k evokes upon arrival of an action potential at the presynaptic
terminal and in the absence of synaptic failure. Synaptic contacts evolve according to a local STDP rule
4 C 2
jk (t) − apost
4 R 4
post(t) − αwjk (t),
(7)
4 = 0.07506 · 10−6 s3, apost
4 = 0.02016 · 10−6 s3 and
with the parameters acorr
α = 2 · 10−6 s−1 (see Supplementary Information S1.2 for details on parameter values). As soon as the
dynamics (7) lead to a contact of weight wjk (t) ≤ 0 its weight is set to 0 and the dynamics cease (inactive
contacts, see also Fig. 1B).
Each potential but inactive contact (weight wjk = 0) may be created again at random times tc according
to a Poisson process with rate λc = 0.019/day. In such an event, the weight is set to wjk (tc) = wc. Further
details on the choice of wc are described below. As suggested by previous works [41, 35] in our model
newly created contacts rst pass through a period of grace of duration τgp = 15 min, during which the
weight is xed to wc. After the period of grace has passed (for t ≥ tc + τgp), the weight dynamics again
follow Eq. (7). In the event of synaptic contact creation (at tc) the internal state variables rjk (tc), rpost(tc),
Cjk (tc) and Rpost(tc) of the contact are each initialized to zero. The period of grace serves as a protected
time interval for these variables to equilibrate to the current system state i.e. to obtain a good estimate
of the present pre- and postsynaptic spike rates and correlations.
wjk (t) = acorr
2 Cjk (t) − acorr
2 = 1.94569 · 10−6 s, acorr
t
d d
Model of postsynaptic activity.
For the simulation of the recurrent network in Fig. 6, the pre- and postsynaptic activity was generated
by leaky integrate-and- re neuron models. However, to perform a mathematical analysis of the model's
15
dynamics, we assume a minimal model of the postsynaptic neuron for the remainder of the study.
In the
absence of synaptic input from the presynaptic neurons, the postsynaptic neuron res with a baseline ring
rate λ0 = 1/s (as a Poisson process). We further assume that synaptic inputs cause transient increases of
the ring rate on the typical time-scale τ of an EPSP. We thus model the dynamics of the postsynaptic
neuron's ring rate λ(t) as
N
X
j=1
njXk=1
wjk (t)Sjk (t − d) ,
(8)
λ(t) = −(λ(t) − λ0) +
t
d d
τ
where the second term sums all inputs across all nj synaptic contacts over all N presynaptic neurons, and
d = 1ms denotes the synaptic transmission delay.
Potential synaptic contacts.
Each presynaptic neuron j may be connected to the postsynaptic neuron by several synaptic contacts, up
to a maximum of nj , 1 ≤ k ≤ nj , cf. Fig. 1A. The number nj of potential contacts of a connection
is random, with a probability distribution (Fig. 1D, blue line) estimated in [28] for synapses connecting
Layer 5 pyramidal neurons within a maximum distance of 50 µm in rat barrel cortex. For computational
reasons we limited n to a maximum of 10 and renormalized the distribution P(n). Accordingly each value of
1 ≤ n ≤ 10 should appear (in expectation) P(n) · N times, where N is the number of presynaptic neurons.
In the model we randomly assigned P(n) · N presynaptic neurons to each potential contact number n in
order to exactly reproduce P(n), the reported distribution of contact numbers per connection, except for
Fig. 5 (single contact case) and Fig. 6 (recurrent network).
Simulation of the plasticity model.
We performed simulations of the full system using the Neural Simulation Tool (NEST) [67]. We implemented
the model as a new multi-contact synapse object, using analytical integration of Eq. (7) as described in
Supplementary Information S1.4. Spikes were restricted to a simulation time grid with a step size of ∆t =
1 ms. The complete system state (wjk , rjk , rpost, Cjk , Rpost, λ) was recorded in intervals of 5 min. The source
code of our plasticity model will be made available on-line as part of NEST (as 'stdp_spl_synapse_hom')
upon acceptance of this manuscript for publication; referees can get access by request to the journal editor.
To be able to compare the model dynamics to the adult networks of the referenced experimental studies,
we initially simulated the model until a stable synaptic con guration was reached. This steady-state of the
system, which is the state at t = 0 in Figs. 2 to 4, was obtained by simulating for 100 days after initialization
at the theoretically determined xed point. The initial state was set to wjk = w∗/5 for 1 ≤ j ≤ 100 and
1 ≤ k ≤ 5, and wjk = 0 else (see details on the xed point and de nition of w∗ below). However, because
the numbering of presynaptic neurons in the model is random and arbitrary, it is possible that initially not
all presynaptic neurons 1 ≤ j ≤ 100 have at least nj ≥ 5 potential contacts. Therefore, for any presynaptic
neuron 1 ≤ j ≤ 100 for which nj < 5, we looked for another neuron j0 > 100 with nj0 ≥ 5, and exchanged
nj and nj0. This procedure allowed us to initialize the system to the theoretical equilibrium state. We also
checked that if the system is initialized unconnected (wjk = 0 for all j, k), fully connected, or randomly
connected, the postsynaptic rate, the total weight, and the contact numbers approach a similar steady state
(data not shown).
(9)
wjk (t)i ≈ acorr
2
hCjk (t)i − acorr
4
hCjk (t)i2 − apost
4
hRposti4 − αwjk ,
t
d d
h
We derive the expected dynamics of the weight of a single synaptic contact under this model. Since dendritic
spine plasticity is a slow process compared to the dynamics of action potentials and synaptic transmission, we
take the average, denoted as h·i, of Eq. (7) over realizations of the spike trains (S) and synaptic transmission
failures (z). We obtain
Expected dynamics of synaptic contacts.
16
which is approximate because squaring and averaging of the terms Cjk and Rpost have been interchanged.
The terms in Eq. (9) can be evaluated as (see Supplementary Information S1.1 for the derivation)
hRposti = hSposti = hλi = λ0 + νpre(1 − pf)
hCjki = hrjk rposti = hrjk λi = νpre(1 − pf) ·
" 1
2τ
e−d/τ (pfwjk + (1 − pf)
N
j=1
wjk
njXk=1
X
wjl ) + hRposti# .
njXl=1
(10)
(11)
Eq. (10) establishes that the postsynaptic rate is determined by the sum of all synaptic weights. Since
we assume pre- and postsynaptic neurons to be of the same type, we have chosen the parameters of the
plasticity rule such that this rate is on average hRposti ≈ νpost = νpre. Our simulations show that this value is
tightly maintained. This implies that the sum of weightsPjPk wjk is normalized [68], cf. Eq. (10). Indeed
previous theoretical work [58] has shown that terms like −R 4
post in our learning rule lead to a normalization
of the total weight.
According to Eq. (11), for small transmission failure probability pf → 0 the dynamics of the contact wjk
(9) is dominated by the total weight wj =Pk wjk . For large pf → 1, the evolution of wjk is independent
of wj ; so increasing the failure probability pf gradually decouples the dynamics of the contacts.
Inserting
Eqs. (10) and (11) into Eq. (9) yields a closed, non-linear system of ordinary dierential dynamical equations
in wjk (t), with 1 ≤ k ≤ nj and 1 ≤ j ≤ N.
In Supplementary Information we describe how the xed points of the expected contact dynamics can
be analyzed and used to calibrate the system (S1.2), and how prototypical trajectories of newly created
contacts can be constructed (S1.3).
Cooperation and competition.
Eqs. (9)-(11) enable us to illustrate the process of cooperation and competition, closely linked to the
stabilization of the postsynaptic rate and correlations. The postsynaptic rate does not depend on the weight
of any speci c synaptic contact, but only on the total input, summed over all weights and contacts; cf.
Eq. 10. By contrast, Eq. (11) depends not only on the total input via the rate Rpost, but in addition also
on the individual weight wjk and the total weight wj = Pl wjl arising from the same presynaptic neuron.
To study competition, let us consider a uniform state and suppose that all correlations Cjk = c and all
momentary weights wjk = w for all j, k are small but positive. With α (cid:28) 1 in Eq. (9), the dominant
evolution is therefore an increase of all weights, driven by the term acorr
c. However, as the weights increase,
the ring rate does so as well and therefore the term hRposti4 eventually stops further growth. This is the
essential step of ring rate stabilization. For the same ring rate hRposti, some weights will grow further at
the expense of others inducing competition via the instability of the uniform state, just as in other models
[50]. The instability is caused by a positive feedback loop between hdwjk (t)/dti on the left-hand side of
Eq. (9) and wjk on the right-hand side of Eq. (11). Going beyond standard plasticity models, Eq. (11)
shows that correlations Cjk driving the contact weight wjk increase not only proportionally to this speci c
contact, but also increase with the weight of other contactsPl wjl from the same presynaptic neuron. The
positive dependence gives rise to cooperation between contacts arising from the same neuron. The optimal
hCjki2 in Eq. (9). The
amount of correlation, and hence cooperation, however, is limited by the term −acorr
interplay of Eqs. (9)-(11) therefore stabilizes the ring rate, or total input PjPk wjk , as well as the total
amount of correlations in active contacts, or total weight Pk wjk , of those presynaptic neurons that have
at least one active contact.
2
4
Contact creation.
With a rate of λc each potential but inactive contact (wjk = 0) may be randomly transformed into an active
contact. In such an event, called creation, its weight wjk is set to a low, non-zero value wc, and after the
period of grace has elapsed, its dynamics follow Eq. (7) (see Fig. 1B). We have adjusted the value of λc to
t the experimentally measured turnover ratio of dendritic spines as follows.
17
The turnover ratio, de ned as TOR = (#created + #removed)/(2 · #total · day), was found to be
0.154/day in the adult somatosensory cortex [12], where #created, #removed and #total are the numbers
of spines that were created, removed, and the total number observed in one day, respectively. In a steady
state we additionally expect that #created = #removed, which implies TOR = #created/(#total · day).
Now assume that the model
is in a steady-state with a distribution of actual contacts as in Fig. 1D,
red line. We derive the creation rate λc loosely from this distribution, by assuming that approximately
10% of synaptic connections have active contacts, and these have about 5 active contacts. Counting the
active synaptic contacts in this case, one would observe about #total ≈ 10% · N · 5 = 0.5 · N. On
the other hand, the total number of potential contacts in our model is #potential ≈ 4.6 · N (Fig. 1D,
blue, shows the histogram of nj across connections). Thus, the number of inactive contacts that are
available for creation is #potential − #total. Hence, the expected number of creations is #created =
(#potential − #total) · λc · 1 day = (4.6 − 0.5) · N · λc · 1 day.
Inserting #created and #total into the
expression for TOR above, and solving for λc, we obtain λc = TOR · 0.5/4.1 = 0.019/day.
To better understand the eects of multiple synaptic contacts in presence of stochastic transmission, let us
analyze the postsynaptic response. To this aim, we force the presynaptic neuron j to emit an additional
spike at time tpre. For convenience we neglect the synaptic transmission delay here (d → 0), which has no
eect on the following reasoning. Assuming constant weights on the short time-scale of synaptic signaling
τ , and by averaging Eq. (S2) over all other presynaptic spikes and their synaptic transmission, we obtain
the transient postsynaptic response
wjk zjk (tpre),
(12)
njXk=1
θ(t − tpre)e−(t−tpre)/τ
1 τ
Lj (t tpre) = νpost +
where we also inserted hRposti ≈ νpost. Note that Lj (t tpre) is still a stochastic quantity due to stochastic
synaptic transmission zjk (tpre) at the contacts k. We may obtain the mean spike-triggered response by
averaging over the remaining stochasticity of zjk
Signal-to-noise ratio of synaptic responses.
θ(t − tpre)e−(t−tpre)/τ (1 − pf)wj .
1 τ
hLj (t tpre)i = νpost +
Thus the average response only depends on the total synaptic weight wj , and not on the con guration of
contact weights wjk . Similarly, the variance of the response can be derived as
var[Lj (t tpre)] = pf(1 − pf)
1
τ 2 θ(t − tpre)e−2(t−tpre)/τ
njXk=1
w 2
jk .
Here we see that the contact con guration wjk determines the variance of the postsynaptic response. To
further understand these properties, consider a synaptic weight wj that is made of nj contacts of weight
j /nj . For this
jk = w 2
k=1 w 2
case we evaluate the signal-to-noise ratio of the synaptic response as
wjk = wj /nj . Then the sum of squared weights term in var[Lj (t)] becomes Pnj
· nj .
=s 1 − pf
pf
SNRj = hLj (t tpre)i − νpost
qvar[Lj (t tpre)]
(13)
Therefore, in presence of synaptic transmission failures, multiple synaptic contacts increase the signal-to-
noise ratio of synaptic transmission, proportional to the square root of the number of contacts. Previously,
a related result has been found numerically for the mutual information of synaptic inputs and neural outputs
via multiple contacts [69].
Recurrent network model
Here we describe the thalamo-cortical network model presented in Fig. 6. The recurrent cortical network
consists of NB = 3 barrels of NE = 200 excitatory (exc) neurons each, and NI = 200 inhibitory (inh)
18
neurons that connect without preference to all the barrels (random connections, see details below). All
cortical neurons are modeled as leaky integrate-and- re (LIF) neurons with alpha-function shaped postsy-
naptic currents ('iaf_psc_alpha' neuron model in NEST simulator). The parameters of the LIF neurons
are: membrane time constant τLIF = 20.6ms, reset and resting potential −70mV, action potential threshold
−55mV, synaptic time constant 2ms, refractory period 2ms. There are NB · NE + NI = 800 cortical neurons
in total. All synaptic delays are 1ms.
Each barrel of excitatory neurons is further innervated by thalamic inputs that convey information from
the whiskers. There are NT = 100 thalamic input neurons (tha) per barrel. All NB · NT = 300 tha neurons
are modeled as excitatory linear Poisson neurons according to Eq. 8, with a baseline ring rate of λ0 = 4.5/s.
Each group of thalamic neurons modulates its ring rate in response to icks of the corresponding whiskers
(whi). The sequence of whisker icks is stochastic and described by Poisson processes S1, S2, S3 with rate
νwhi = 1/s. The connection weights from whi to tha are chosen as wwhi = 0.5. We assume full connectivity
to the respective tha populations, such that each tha neuron responds to each ick of the corresponding
whisker. In response to a whisker ick, tha neurons of the receiving population increase their ring rate λ(t)
transiently by wwhi/τ = 25/s, and subsequently their rate decays back to λ0 quickly with time constant τ
(cf. Eq. (8)).
The structural plasticity model (7) describes all exc to exc connections, both thalamo-cortical and intra-
cortical, and is continuously active (except for the rst hour after simulation); autapses are excluded. All
connections (both from tha to exc and exc to exc) have potential synaptic contacts, but initially the network
is connected in a whisker-speci c manner as depicted in Fig. 6A, see also below. Because the recurrent
network contains many postsynaptic neurons, we need two indices to name a synaptic connection. A contact
weight here is denoted by wijk instead of wjk above, where i denotes the postsynaptic and j denotes the
presynaptic neuron, and k the contact. Accordingly, the plasticity rule (7) here reads
(14)
wijk (t) = acorr
2 Cijk (t) − acorr
4 C 2
ijk (t) − apost
4 R 4
i (t) − αwijk (t),
t
d d
and the total weight of a synaptic connection is wij (t) =Pnij
k=1 wijk (t), where nij is the number of potential
contacts for the connection from j to i. For simplicity all nij are drawn from the probability distribution
P(nij ) (Fig. 1D, blue), irrespective of which group (exc or tha) the neurons i and j belong to.
Because the postsynaptic neurons here are LIF neurons, synaptic e cacies wijk have to be expressed in
units of the PSP (in contrast, above wjk is a unit-less quantity). To match the impulse response function
of the LIF neurons receiving an input spike with the xed point weight w∗ to the response of the linear
Poisson neurons used above, we scale the synaptic weights as wijk = γ · wijk , with γ = 62.82 mV, leading
to a typical EPSP amplitude of γw∗ = 1.01 mV. Substituting wijk into Eq. (14) implies that, to maintain
the same plasticity dynamics as above, the parameters of the learning rule have to be rescaled according
to acorr
, w0 7→ γw0. We further inject additional Poisson
excitatory and inhibitory input spikes to all LIF neurons, with synaptic weights γw∗ (excitation) and −4γw∗
(inhibition). Exc neurons receive input rates 1519.2/s (excitation) and 328.3/s (inhibition), inh neurons
receive 1391.0/s (excitation) and 351.2/s (inhibition). The scaling factor γ and the Poisson process input
rates were numerically optimized to match the dynamics of the LIF neuron model to that of the linear
Poisson neuron model used above. All other parameters take the same values as before. Note that the
network parameters here are chosen such that the system operates approximately at the xed point of the
plasticity dynamics analyzed in Fig. S1.
7→ γapost
7→ γacorr
7→ γacorr
, apost
4
, acorr
4
2
2
4
4
Apart from its (NB[NT + NE])(NB[NE − 1]) ≈ 5.4 · 105 plastic excitatory connections our network
also has static synapses. These we set as follows. We choose a connection probability of pconn = 1/3.
Each excitatory neuron receives pconnNI synapses from randomly chosen inh neurons with a xed weight of
−(1 − pf)γgw∗, with g = 2.5. Each inhibitory neuron also receives this amount of inhibitory synapses, and
pconnNE excitatory synapses from randomly chosen exc neurons from each of the NB cortical barrels, with
a xed weight of (1 − pf)γw∗ (these synapses have no transmission failures, therefore the weight is scaled
down by the expected transmission rate (1 − pf) of the plastic, stochastic ones).
We initialize the plastic synapses at the theoretically derived xed point w∗, with an expected total of
100 active input connections with 5 contacts each per neuron. So, for each exc-exc and tha-exc connection,
we set wij (0) = γw∗qij if neuron i and neuron j belong to the same barrel, where qij is a Bernoulli random
number that is 1 with probability pconn and 0 else (in this way, we get (NT + NE)pconn = 100 incoming
19
connections per exc neuron in expectation). If i and j are part of dierent barrels, we set wij (0) = 0. If there
are ve or more potential contacts (nij ≥ 5) in connection i, j, we set wijk (0) = wij (0)/5 for 1 ≤ k ≤ 5
and wijk (0) = 0 for k ≥ 5.
If there are less contacts (nij < 5) but wij (0) > 0, we set wij (0) = 0 and
look for a connection i0, j0 that connects the same two groups, has wi0j0(0) = 0 and (ni0j0 ≥ 5), and we set
wi0j0(0) = γw∗ for this connection instead. In Fig. 6D neurons of barrel column exc 3 are ordered according
to labels obtained by clustering. We used feature agglomeration based on Ward's hierarchical clustering [70]
to assign one of three cluster labels to the vector {wi1, wi2, ...} of connection weights (at simulation end)
from any exc neuron onto each exc 3 neuron i.
Simulations were performed using NEST [67] as described above, except for: (i) contact weights were not
recorded, merely the total connection weights wij (t) and the number of active contacts of each connection;
(ii) the system state was recorded only every 60min instead of every 5min above; (iii) we simulated using
44 computing cores in parallel instead of a single core.
Acknowledgments:
Research was supported by the European Union Seventh Framework Program (FP7) under grant agreement
no. 604102 (Human Brain Project, M.D.) and by the Swiss National Science Foundation (200020_147200,
A.S.).
References
[1] Sheng M, Hoogenraad CC (2007) The postsynaptic architecture of excitatory synapses: a more quan-
titative view. Annu Rev Biochem 76: 823847.
[2] Yuste R (2011) Dendritic spines and distributed circuits. Neuron 71: 772781.
[3] Bonhoeer T, Yuste R (2002) Spine motility. phenomenology, mechanisms, and function. Neuron 35:
10191027.
[4] Yasumatsu N, Matsuzaki M, Miyazaki T, Noguchi J, Kasai H (2008) Principles of long-term dynamics
of dendritic spines. J Neurosci 28: 1359213608.
[5] Holtmaat A, Svoboda K (2009) Experience-dependent structural synaptic plasticity in the mammalian
brain. Nat Rev Neurosci 10: 647658.
[6] Kasai H, Fukuda M, Watanabe S, Hayashi-Takagi A, Noguchi J (2010) Structural dynamics of dendritic
spines in memory and cognition. Trends Neurosci 33: 121129.
[7] Loewenstein Y, Kuras A, Rumpel S (2011) Multiplicative dynamics underlie the emergence of the
log-normal distribution of spine sizes in the neocortex in vivo. J Neurosci 31: 94819488.
[8] Sala C, Segal M (2014) Dendritic spines: the locus of structural and functional plasticity. Physiol Rev
94: 141188.
[9] Kalisman N, Silberberg G, Markram H (2005) The neocortical microcircuit as a tabula rasa. Proc Natl
Acad Sci U S A 102: 880885.
[10] Stepanyants A, Hof PR, Chklovskii DB (2002) Geometry and structural plasticity of synaptic connec-
tivity. Neuron 34: 275288.
[11] Kappel D, Habenschuss S, Legenstein R, Maass W (2015) Network plasticity as bayesian inference.
PLoS Comput Biol 11: e1004485.
[12] Holtmaat AJGD, Trachtenberg JT, Wilbrecht L, Shepherd GM, Zhang X, et al. (2005) Transient and
persistent dendritic spines in the neocortex in vivo. Neuron 45: 279291.
20
[13] Grutzendler J, Kasthuri N, Gan WB (2002) Long-term dendritic spine stability in the adult cortex.
Nature 420: 812816.
[14] Hayashi-Takagi A, Yagishita S, Nakamura M, Shirai F, Wu YI, et al. (2015) Labelling and optical
erasure of synaptic memory traces in the motor cortex. Nature 525: 333338.
[15] Matsuzaki M, Honkura N, Ellis-Davies GCR, Kasai H (2004) Structural basis of long-term potentiation
in single dendritic spines. Nature 429: 761766.
[16] Holtmaat A, Wilbrecht L, Knott GW, Welker E, Svoboda K (2006) Experience-dependent and cell-
type-speci c spine growth in the neocortex. Nature 441: 979983.
[17] Zito K, Scheuss V, Knott G, Hill T, Svoboda K (2009) Rapid functional maturation of nascent dendritic
spines. Neuron 61: 247258.
[18] Kwon HB, Sabatini BL (2011) Glutamate induces de novo growth of functional spines in developing
cortex. Nature 474: 100104.
[19] Minerbi A, Kahana R, Goldfeld L, Kaufman M, Marom S, et al. (2009) Long-term relationships between
synaptic tenacity, synaptic remodeling, and network activity. PLoS Biol 7: e1000136.
[20] Fusi S, Drew PJ, Abbott LF (2005) Cascade models of synaptically stored memories. Neuron 45:
599611.
[21] Clopath C, Ziegler L, Vasilaki E, Büsing L, Gerstner W (2008) Tag-trigger-consolidation: a model of
early and late long-term-potentiation and depression. PLoS Comput Biol 4: e1000248.
[22] Roxin A, Fusi S (2013) E cient partitioning of memory systems and its importance for memory con-
solidation. PLoS Comput Biol 9: e1003146.
[23] Zenke F, Agnes EJ, Gerstner W (2015) Diverse synaptic plasticity mechanisms orchestrated to form
and retrieve memories in spiking neural networks. Nat Commun 6: 6922.
[24] Kempter R, Gerstner W, van Hemmen JL (1999) Hebbian learning and spiking neurons. Phys Rev E
59.
[25] Abbott LF, Nelson SB (2000) Synaptic plasticity: taming the beast. Nat Neurosci 3: 11781183.
[26] van Rossum MC, Bi GQ, Turrigiano GG (2000) Stable hebbian learning from spike timing-dependent
plasticity. J Neurosci 20: 88128821.
[27] Markram H, Lübke J, Frotscher M, Roth A, Sakmann B (1997) Physiology and anatomy of synaptic
connections between thick tufted pyramidal neurones in the developing rat neocortex. J Physiol 500 (
Pt 2): 409440.
[28] Fares T, Stepanyants A (2009) Cooperative synapse formation in the neocortex. Proc Natl Acad Sci
U S A 106: 1646316468.
[29] Markram H, Muller E, Ramaswamy S, Reimann MW, Abdellah M, et al. (2015) Reconstruction and
simulation of neocortical microcircuitry. Cell 163: 456492.
[30] Reimann MW, King JG, Muller EB, Ramaswamy S, Markram H (2015) An algorithm to predict the
connectome of neural microcircuits. Front Comput Neurosci 9: 120.
[31] Stuart G, Schiller J, Sakmann B (1997) Action potential initiation and propagation in rat neocortical
pyramidal neurons. J Physiol 505 ( Pt 3): 617632.
[32] Markram H, Lübke J, Frotscher M, Sakmann B (1997) Regulation of synaptic e cacy by coincidence
of postsynaptic aps and epsps. Science 275: 213215.
21
[33] Petersen CC, Malenka RC, Nicoll RA, Hop eld JJ (1998) All-or-none potentiation at ca3-ca1 synapses.
Proc Natl Acad Sci U S A 95: 47324737.
[34] O'Connor DH, Wittenberg GM, Wang SSH (2007) Timing and contributions of pre-synaptic and post-
synaptic parameter changes during unitary plasticity events at ca3-ca1 synapses. Synapse 61: 664678.
[35] Helias M, Rotter S, Gewaltig MO, Diesmann M (2008) Structural plasticity controlled by calcium based
correlation detection. Front Comput Neurosci 2: 7.
[36] Deger M, Helias M, Rotter S, Diesmann M (2012) Spike-timing dependence of structural plasticity
explains cooperative synapse formation in the neocortex. PLoS Comput Biol 8: e1002689.
[37] Butz M, van Ooyen A (2013) A simple rule for dendritic spine and axonal bouton formation can account
for cortical reorganization after focal retinal lesions. PLoS Comput Biol 9: e1003259.
[38] Vlachos A, Helias M, Becker D, Diesmann M, Deller T (2013) Nmda-receptor inhibition increases spine
stability of denervated mouse dentate granule cells and accelerates spine density recovery following
entorhinal denervation in vitro. Neurobiol Dis 59: 267276.
[39] Fauth M, Wörgötter F, Tetzla C (2015) The formation of multi-synaptic connections by the inter-
action of synaptic and structural plasticity and their functional consequences. PLoS Comput Biol 11:
e1004031.
[40] Attardo A, Fitzgerald JE, Schnitzer MJ (2015) Impermanence of dendritic spines in live adult ca1
hippocampus. Nature 523: 592596.
[41] Le Bé JV, Markram H (2006) Spontaneous and evoked synaptic rewiring in the neonatal neocortex.
Proc Natl Acad Sci U S A 103: 1321413219.
[42] Trachtenberg JT, Chen BE, Knott GW, Feng G, Sanes JR, et al. (2002) Long-term in vivo imaging of
experience-dependent synaptic plasticity in adult cortex. Nature 420: 788794.
[43] Loewenstein Y, Yanover U, Rumpel S (2015) Predicting the dynamics of network connectivity in the
neocortex. J Neurosci 35: 1253512544.
[44] Keck T, Mrsic-Flogel TD, Vaz Afonso M, Eysel UT, Bonhoeer T, et al. (2008) Massive restructuring of
neuronal circuits during functional reorganization of adult visual cortex. Nat Neurosci 11: 11621167.
[45] Rose T, Jaepel J, Hübener M, Bonhoeer T (2016) Cell-speci c restoration of stimulus preference after
monocular deprivation in the visual cortex. Science 352: 13191322.
[46] Nawrot MP, Schnepel P, Aertsen A, Boucsein C (2009) Precisely timed signal transmission in neocortical
networks with reliable intermediate-range projections. Front Neural Circuits 3: 1.
[47] Albieri G, Barnes SJ, de Celis Alonso B, Cheetham CEJ, Edwards CE, et al. (2015) Rapid bidirectional
reorganization of cortical microcircuits. Cereb Cortex 25: 30253035.
[48] Wilbrecht L, Holtmaat A, Wright N, Fox K, Svoboda K (2010) Structural plasticity underlies experience-
dependent functional plasticity of cortical circuits. J Neurosci 30: 49274932.
[49] Nägerl UV, Willig KI, Hein B, Hell SW, Bonhoeer T (2008) Live-cell imaging of dendritic spines by
sted microscopy. Proc Natl Acad Sci U S A 105: 1898218987.
[50] Oja E (1982) A simpli ed neuron model as a principal component analyzer. J Math Biol 15: 267273.
[51] Ibata K, Sun Q, Turrigiano GG (2008) Rapid synaptic scaling induced by changes in postsynaptic ring.
Neuron 57: 819826.
[52] Lohmann C, Finski A, Bonhoeer T (2005) Local calcium transients regulate the spontaneous motility
of dendritic lopodia. Nat Neurosci 8: 305312.
22
[53] Choi BJ, Imlach WL, Jiao W, Wolfram V, Wu Y, et al. (2014) Miniature neurotransmission regulates
drosophila synaptic structural maturation. Neuron 82: 618634.
[54] Druckmann S, Feng L, Lee B, Yook C, Zhao T, et al. (2014) Structured synaptic connectivity between
hippocampal regions. Neuron 81: 629640.
[55] Clopath C, Büsing L, Vasilaki E, Gerstner W (2010) Connectivity re ects coding: a model of voltage-
based stdp with homeostasis. Nat Neurosci 13: 344352.
[56] P ster JP, Gerstner W (2006) Triplets of spikes in a model of spike timing-dependent plasticity. J
Neurosci 26: 96739682.
[57] Senn W, Markram H, Tsodyks M (2001) An algorithm for modifying neurotransmitter release probability
based on pre- and postsynaptic spike timing. Neural Comput 13: 3567.
[58] Zenke F, Hennequin G, Gerstner W (2013) Synaptic plasticity in neural networks needs homeostasis
with a fast rate detector. PLoS Comput Biol 9: e1003330.
[59] Sjöström PJ, Turrigiano GG, Nelson SB (2001) Rate, timing, and cooperativity jointly determine
cortical synaptic plasticity. Neuron 32: 11491164.
[60] Miner D, Triesch J (2016) Plasticity-driven self-organization under topological constraints accounts for
non-random features of cortical synaptic wiring. PLoS Comput Biol 12: e1004759.
[61] Diaz-Pier S, Naveau M, Butz-Ostendorf M, Morrison A (2016) Automatic generation of connectivity
for large-scale neuronal network models through structural plasticity. Front Neuroanat 10: 57.
[62] Toyoizumi T, Kaneko M, Stryker MP, Miller KD (2014) Modeling the dynamic interaction of hebbian
and homeostatic plasticity. Neuron 84: 497510.
[63] Morrison A, Diesmann M, Gerstner W (2008) Phenomenological models of synaptic plasticity based
on spike timing. Biol Cybern 98: 459478.
[64] Vogels TP, Sprekeler H, Zenke F, Clopath C, Gerstner W (2011) Inhibitory plasticity balances excitation
and inhibition in sensory pathways and memory networks. Science 334: 15691573.
[65] Daoudal G, Debanne D (2003) Long-term plasticity of intrinsic excitability: learning rules and mecha-
nisms. Learn Mem 10: 456465.
[66] Hessler NA, Shirke AM, Malinow R (1993) The probability of transmitter release at a mammalian
central synapse. Nature 366: 569572.
[67] Bos H, Morrison A, Peyser A, Hahne J, Helias M, et al. (2015). Nest 2.10.0. doi:10.5281/zenodo.44222.
URL http://dx.doi.org/10.5281/zenodo.44222.
[68] Kempter R, Gerstner W, van Hemmen JL (2001) Intrinsic stabilization of output rates by spike-based
hebbian learning. Neural Comput 13: 27092741.
[69] Zador A (1998) Impact of synaptic unreliability on the information transmitted by spiking neurons. J
Neurophysiol 79: 12191229.
[70] Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, et al. (2011) Scikit-learn: Machine
learning in Python. Journal of Machine Learning Research 12: 28252830.
23
Supplementary Information for Multi-contact synapses for stable
networks: a spike-timing dependent model of dendritic spine plasticity
and turnover
Moritz Deger1,2, Alexander Seeholzer1, Wulfram Gerstner1
1: School of Computer and Communication Sciences and School of Life Sciences, Brain Mind Institute,
École Polytechnique Fédérale de Lausanne, 1015 Lausanne EPFL, Switzerland
2: Institute for Zoology, Faculty of Mathematics and Natural Sciences,
University of Cologne, 50674 Cologne, Germany
To derive Eqs. (10) and (11) in the main text, we rst write the solution of the dierential equations of the
traces de ned in Eqs. (3) and (4) of the main text,
(S1)
e−(t−s)/τ S·(s)ds ,
t
-∞
1 τ
r·(t) =
where · stands for either jk or post. So both rjk and rpost are low-pass ltered versions of presynaptic (Sjk )
and postsynaptic activity (Spost), respectively, with time constant τ . Similarly, the expressions Cjk (t) and
Rpost(t) in Eq. (7) of the main text denote low-pass ltered traces (with time-constant τslow) of pre-post
correlations and the postsynaptic spike train, respectively. Because presynaptic ring and synaptic failures
are modeled as independent random variables, taking the expectation value of rjk (S1) with respect to the
realizations of Sjk yields hrjki = hSjki = νpre(1 − pf).
For simplicity, we have assumed that the traces r·(t) as well as the postsynaptic rate λ(t) all evolve on
the common time scale τ . The postsynaptic activity λ(t) can then, by solving Eq. (8) of the main text and
inserting (S1), be written as
S1 Supplementary Methods
S1.1 Derivation of pre-post correlations.
wjk (t)rjk (t − d),
(S2)
e−(t−s)/τ
N
X
j=1
njXk=1
wjk (s)Sjk (s − d)ds
t
-∞
njXk=1
X
1 τ
j=1
N
λ(t) = λ0 +
= λ0 +
where in the second step we assumed that wjk does not change much on the (fast) time-scale of τ . Since
the expectation of rjk is νpre(1 − pf) we obtain Eq. (10) of the main text.
We arrive at the expression hrjk λi, which evaluates to
Analogously, we compute Eq. (11) of the main text by inserting the explicit solutions of the traces (S1).
ds e−(t−s)/τ 1
τ t
-∞
du e−(t−u)/τXi,l
wilhSjk (s)Sil (u − d)i + hSjk (s)iλ0 .
(S3)
t
-∞
1 τ
hrjk λi =
Here we insert
hSjk (s)Sil (u)i = hSj (s)zjk (s)Si (u)zil (u)i = hSj (s)Si (u)ihzjk (s)zil (u)i
= Cji (s, u)[δkl (1 − pf) + (1 − δkl )(1 − pf)2] + ν2
pre(1 − pf)2 ,
1
where we substituted the covariance function of the spike trains j and i, Cji (s, u) = hSj (s)Si (u)i −
hSj (s)ihSi (u)i. For the Poisson processes of our model we have Cji (s, u) = νpreδij δ(s − u). Here δij
denotes the Kronecker symbol and δ(·) denotes the Dirac delta function. When summed over i, l together
with wil , the expression above becomes
X
i,l
wilhSjk (s)Sil (u)i =Xi,l
wilCji (s, u)[δkl (1 − pf) + (1 − δkl )(1 − pf)2] + ν2
pre(1 − pf)2Xi
[pfwik + (1 − pf)wi ]Cji (s, u) + νpre[hRposti − λ0]!
= (1 − pf) Xi
wi
(S4)
(S5)
= νpre(1 − pf) [δ(s − u) (pfwjk + (1 − pf)wj ) + hRposti − λ0] ,
where we inserted the covariance function Cji for independent Poisson process inputs in the last step. By
insertion of (S5) into (S3) the last equality of Eq. (11) of the main text follows.
Note that our theory can as well describe the more general case in which presynaptic spike trains
Sj are non-Poisson, but have stationary covariance function Cji (s, u) = Cji (s − u) and stationary mean
hSj (u)i = νpre. When we insert (S4) instead of (S5) into (S3) we obtain the more general expression for
the expected correlation term at the synaptic contact
hCjki = hrjk λi = (1 − pf)Xi
with the eective spike train covariance
[pfwik + (1 − pf)wi ] ¯Cji + νpre(1 − pf)hRposti ,
(S6)
¯
Cji = τ−2 ∞
0 ∞
0
e−(u+s)/τCji (u − s + d) du ds .
Eq. (S6) shows that the plasticity model is generally sensitive to spike-timing correlations of neurons all
across the network.
S1.2 Mathematical analysis of system dynamics.
As our simulations show, the learning rule maintains a constant postsynaptic ring rate νpost by tight
regulation of the total input weight. Thus the actual degrees of freedom of the multi-connection multi-
contact system reside in the con guration of which of the connections wj are strong or weak, and through
which number of contacts wjk this strength is achieved. To understand these dynamics, we consider the
is
expected change of the weight of a single contact dhwjki/dt assuming the total input w = PN
constant and given by Eq. (10) in the main text. This analysis is displayed in Fig. S1.
With the parameters of the learning rule (Eq. (7) of the main text) our mathematical analysis shows
that the steady state has the following properties: (a) There is a stable xed point of Eq. (9) (marked by the
circle in Fig. S1), such that if only 10% of the synaptic connections are active (consistent with [1, 2]), with
total weight wj = w∗ each, the postsynaptic ring rate of νpost = νpre = 5/s is achieved. From Eq. (10) in
the main text follows that this xed point weight has to be w∗ = (νpost−λ0)/[(1−pf)νpre·10%·N] = 0.016.
The creation weight wc (cross in Fig. S1) was set arbitrarily to 15% of the xed point weight of a contact,
wc = 0.15 w∗/5 (our results are insensitive against the exact value). (b): Contacts only have a stable xed
point of wjk in connections that contain at least three active contacts (dash-dotted lines in Fig. S1A). In
connections with less than three active contacts (dashed lines in Fig. S1A), dhwjki/dt is generally smaller
than 0, so that these contacts are removed eventually. If (b) is ful lled, contacts in connections with about
5 contacts, as observed experimentally in [1], are stable.
j=1 wj
To calibrate our plasticity model (Eq. (7) of the main text) we used a numerical optimization procedure
to nd parameters acorr
and α that ful ll these constraints. As the con guration of xed points
of Eq. 9 of the main text is insensitive to a common rescaling of these parameters, we manually adjusted
the scale of the parameters to achieve a good agreement of the temporal dynamics of the contacts (cf.
Fig. 2A,D,E in the main text) with the experimental references.
, apost
, acorr
2
4
4
2
Figure S1: Expected dynamics of synaptic contacts. A: Positive or negative change (red or blue
color, Eq. 9) of a synaptic contact as a function of its weight wjk (horizontal axis) and the total weight
of the connection from presynaptic neuron j (wj , vertical axis), under the assumption that the sum of all
weights of all presynaptic neurons is implicitly normalized via Eq. (10). Straight lines mark synapses where
wj
is a multiple of wjk , i.e. synapses consisting of 1 and 2 contacts (dashed), or 3, 5 and 10 contacts
(dash-dotted). The circle marks the stable xed point (w∗/5, w∗) of the dynamics (for 5 actual contacts),
the white cross marks the point of a newly created contact in a previously inactive connection (wc, wc).
The black curve marks combinations of wjk and wj that have zero expected change (wjk nullcline). B-C:
Expected trajectories after a perturbation. We assume that in a synaptic connection j with nj contacts,
nj − 1 contacts have identical values while a single contact wjk is perturbed. B: In a synaptic connection
with ve contacts ( xed point at wj = w∗, circle), four contacts have the same weight w∗/5 while the
contact wjk is perturbed by a small amount w∆. Its trajectory (red line) starts at (w∗/5 + w∆, w∗ + w∆)
and evolves back towards the xed point (circle). C: Five contacts have existed for a long time and each
take a value wjl = w∗/5 consistent with the xed point (circle), when spontaneously a new contact wjk is
created at its creation value wc. The trajectory of the new contact (red line) starts at (wc, w∗ + wc) and
evolves towards the new xed point for 6 contacts. The previously existing contacts also evolve towards the
new xed point (green line).
S1.3 Expected trajectories in response to perturbations.
The expected dynamics (Eq. (9) in the main text) allow us to predict trajectories of (wjk , wj ). In Fig. S1B
we show that, in expectation, if a contact weight wjk in an active connection at steady-state is perturbed,
the perturbation decays. In Fig. S1C, we display what happens when a new contact is created at the steady
state in an active connection with ve contacts. We nd that the new contact as well as the ve existing
contacts move to a new xed point corresponding to six stable contacts. As we have seen above, a synaptic
connetion consisting of three or more contacts is stable in expectation. In Fig. S2 we consider the creation
of synaptic contacts in an inactive connection, in which no other active contacts exist. Both for a single
newly created contact (Fig. S2A) and for two simultaneously created new contacts (Fig. S2B), the new
contacts are expected to approach zero. A phase plane analysis of the dynamics of two contacts (Fig. S2C)
further con rms that, in expectation, all connections of only two contacts are eventually removed because
of the absence of stable xed points. However, because of a region on the phase plane where the dynamics
are slow, a connection with two contacts has a lifetime that is su ciently long so that occasionally a third
contact may be added. Still, due to the low creation rate we expect this event to be rare. These theoretical
considerations suggest that the system is indeed calibrated to have a steady state that is qualitatively
consistent with the experimental contact number distributions, in the sense that stable connections have at
least three synaptic contacts.
S1.4 Analytical integration of synaptic dynamics.
As we show here, the dynamics of the plasticity model can be integrated analytically in the absence of
spikes. This allows us to iterate the dynamics from spike to spike precisely and e ciently, and to implement
it as an event-driven algorithm in NEST [3].
If for all t ∈ (t0, t1) there are no transmitted pre- nor postsynaptic spikes, Sjk (t) = Spost(t) = 0, then
the traces de ned in Eqs. (3), (4) and (6) in the main text decay exponentially as rjk (t) = rjk (t0) exp(−(t −
t0)/τ ), rpost(t) = rpost(t0) exp(−(t − t0)/τ ) and Rpost(t) = Rpost(t0) exp(−(t − t0)/τslow), respectively.
3
ai ebi (t−t0) ,
(S8)
7
X
i=1
1 c
wjk (t) =
Figure S2: Transient synaptic contacts in inactive connections. A: New contact in a connection without
other actual contacts. The new contact approaches wjk = 0 and is removed (red line). B: Simultaneous
creation of two new contacts in a connection without other contacts. Both of the contacts identically
approach wjk = 0 and are removed (red line). C: Phase plane analysis of the two contacts weights wjk
and wjl for a connection made of 2 contacts (k 6= l). The nullclines of wjk (green) and wjl (red) do not
intersect, therefore the system does not have a xed point, but there is a region where the nullclines are
close to each other where the ow is rather slow. Colored arrows denotes the gradient d/dt(wjk , wjl ) on the
nullclines (arrow length corresponds to gradient magnitude in arbitrary units). Black streamlines indicate
direction of ow of the dynamics, thickness of arrows denotes speed. Visibly, all trajectories converge to
(0, 0), yet the ghosts of an unstable xed point in the region where nullclines are close slows the dynamics
su ciently to allow for the occasional creation of additional contacts.
Inserting these functions into Eq. (5) of the main text and solving the dierential equation yields
Cjk (t) = Cjk (t0)e−(t−t0)/τslow + rjk (t0)rpost(t0)
e−2(t−t0)/τ − e−(t−t0)/τslow
1 − 2τslow/τ
.
(S7)
Similarly also wjk (t) can be solved for analytically. We insert the solution for Rpost(t) and Eq. (S7)
and solve Eq. (7) of the main text. The resulting analytical expression for wjk (t) is a generalized Dirichlet
polynomial given by
with the parameters
a1 = 2acorr
4
rjk (t0)rpost(t0)τ 2(−4 + ατ )(−2 + ατ )τslow(−(Cjk,post(t0)τ ) +
rjk (t0)rpost(t0)τ + 2Cjk,post(t0)τslow)(−4 + ατslow)(−2 + ατslow)(−1 + ατslow) ,
a2 = acorr
2
(−4 + ατ )(−2 + ατ )(−(rjk (t0)rpost(t0)τ ) +
Cjk,post(t0)(τ − 2τslow))(τ − 2τslow)τslow(−4 + ατslow) ·
(−2 + ατslow)(−2τslow + τ (−1 + ατslow)) ,
a3 = −(acorr
4
(−4 + ατ )(−2 + ατ )τslow(−(Cjk,post(t0)τ ) + rjk (t0)rpost(t0)τ +
2Cjk,post(t0)τslow)(−4 + ατslow)(−1 + ατslow)(−2τslow + τ (−1 + ατslow))) ,
a4 = −(apost
4 R 4
post(t0)(−4 + ατ )(−2 + ατ )(τ − 2τslow)2τslow ·
(−2 + ατslow)(−1 + ατslow)(−2τslow + τ (−1 + ατslow))) ,
a5 = −(acorr
4
r 2
jk (t0)r 2
post(t0)τ 3(−2 + ατ )(−4 + ατslow) ·
(−2 + ατslow)(−1 + ατslow)(−2τslow + τ (−1 + ατslow))) ,
4
a6 = acorr
2
rjk (t0)rpost(t0)τ 2(−4 + ατ )(τ − 2τslow)(−4 + ατslow) ·
(−2 + ατslow)(−1 + ατslow)(−2τslow + τ (−1 + ατslow)) ,
a7 = (τ − 2τslow)2(wjk (t0)(−4 + ατ )(−2 + ατ )(−4 + ατslow)(−2 + ατslow) ·
2
(−1 + ατslow)(−2τslow + τ (−1 + ατslow)) + acorr
(−2 + ατslow)(rjk (t0)rpost(t0)τ + Cjk,post(t0)(2 − ατ )τslow) ·
(−2τslow + τ (−1 + ατslow)) + (−2 + ατ )(−1 + ατslow)(apost
(−4 + ατ )τslow(−2 + ατslow)(−τ − 2τslow + ατ τslow) + acorr
(2r 2
2rjk (t0)rpost(t0))τ (−4 + ατ )τslow + C 2
post(t0)τ 2 − Cjk,post(t0)(Cjk,post(t0) +
jk (t0)r 2
4
jk,post(t0)(−4 + ατ )(−2 + ατ )τ 2
slow))) ,
(−4 + ατ )(−4 + ατslow) ·
4 R 4
post(t0) ·
(−4 + ατslow) ·
b1 = −(1/τslow + 2/τ ),
−2/τ ,
and
b7 = −α,
b2 = −1/τslow,
b3 = −2/τslow,
b4 = −4/τslow,
b5 = −4/τ ,
b6 =
c = (−4 + ατ )(−2 + ατ )(τ − 2τslow)2(−4 + ατslow) ·
(−2 + ατslow)(−1 + ατslow)(−2τslow + τ (−1 + ατslow)) .
We can thus integrate the dynamics of the plasticity model exactly between any two spikes by evaluating
Eq. (S8).
At any time t at which there is a spike (Sjk (t) 6= 0 or Spost(t) 6= 0), we merely have to increment the
traces: rjk (t) ← rjk (t) + 1/τ , rpost(t) ← rpost(t) + 1/τ , or Rpost(t) ← Rpost(t) + 1/τslow, respectively. Then
the solutions of the dierential equations for the subsequent inter-spike-interval are as given above, with
t0 ← t.
When iterating the time-evolution of wjk from spike to spike by Eq. (S8), however, we need to make
sure that wjk (t) does not cross the zero line at any intermediate t ∈ (t0, t1) (zero-crossings are important
because they cause pruning of the contact). Therefore we apply Theorem 4.7 of [4], which provides a
criterion to rule out zero-crossings of wjk (t) for any t > t0.
If zero crossings cannot be ruled out by this
criterion, we evaluate Eq. (S8) for all t in the interval (t0, t1), on the simulation time grid ∆t, to guarantee
that no zero-crossings of wjk were missed.
References
[1] Markram H, Lübke J, Frotscher M, Roth A, Sakmann B (1997) Physiology and anatomy of synaptic
connections between thick tufted pyramidal neurones in the developing rat neocortex. J Physiol 500 (
Pt 2): 409440.
[2] Fares T, Stepanyants A (2009) Cooperative synapse formation in the neocortex. Proc Natl Acad Sci U
S A 106: 1646316468.
[3] Bos H, Morrison A, Peyser A, Hahne J, Helias M, et al. (2015). Nest 2.10.0. doi:10.5281/zenodo.44222.
URL http://dx.doi.org/10.5281/zenodo.44222.
[4] Jameson GJO (2006) Counting zeros of generalised polynomials: Descartes rule of signs and laguerres
extensions. Math Gazette 90: 223-234.
5
|
1906.09290 | 1 | 1906 | 2019-06-21T18:46:09 | Brain state stability during working memory is explained by network control theory, modulated by dopamine D1/D2 receptor function, and diminished in schizophrenia | [
"q-bio.NC",
"physics.bio-ph"
] | Dynamical brain state transitions are critical for flexible working memory but the network mechanisms are incompletely understood. Here, we show that working memory entails brainwide switching between activity states. The stability of states relates to dopamine D1 receptor gene expression while state transitions are influenced by D2 receptor expression and pharmacological modulation. Schizophrenia patients show altered network control properties, including a more diverse energy landscape and decreased stability of working memory representations. | q-bio.NC | q-bio | Title Page
Brain state stability during working memory is explained by network control theory,
modulated by dopamine D1/D2 receptor function, and diminished in schizophrenia
Urs Braun1,2* MD, Anais Harneit1 MSc, Giulio Pergola3 PhD, Tommaso Menara4 MSc, Axel
Schaefer5 PhD, Richard F. Betzel6 PhD, Zhenxiang Zang1 MSc, Janina I. Schweiger1 MD,
Kristina Schwarz1 MSc, Junfang Chen1 MSc, Giuseppe Blasi3 MD PhD, Alessandro Bertolino3
MD PhD, Daniel Durstewitz7 PhD, Fabio Pasqualetti4 PhD, Emanuel Schwarz1 PhD, Andreas
Meyer-Lindenberg1 MD, Danielle S. Bassett2,8# PhD, Heike Tost1# MD
1 Department of Psychiatry and Psychotherapy, Central Institute of Mental Health, Medical
Faculty Mannheim, University of Heidelberg, Mannheim, Germany
2 Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, USA
3 Department of Basic Medical Science, Neuroscience, and Sense Organs, University of Bari
Aldo Moro, 70124 Bari, Italy
4 Mechanical Engineering Department, University of California at Riverside, Riverside, CA,
USA
5 Bender Institute of Neuroimaging, Justus Liebig University Giessen & Center for Mind,
Brain and Behavior, University of Marburg and Justus Liebig University Giessen, Giessen,
Germany
6 Department of Psychological and Brain Sciences, Indiana University, Bloomington, IN, USA
7 Department of Theoretical Neuroscience, Central Institute of Mental Health, Medical Faculty
Mannheim/Heidelberg University, 68159 Mannheim, Germany
8 Department of Psychiatry, Department of Neurology, Department of Physics & Astronomy,
and Department of Electrical & Systems Engineering, University of Pennsylvania,
Philadelphia, PA, USA
# these authors contributed equally
Corresponding Author: Urs Braun M.D., Department of Psychiatry and Psychotherapy,
Central Institute of Mental Health, Medical Faculty Mannheim, University of Heidelberg, J5,
68159 Mannheim, Germany; Tel.: +49 621 1703 6519; e-mail: [email protected]
Key Words: network control theory, executive function, working memory, schizophrenia,
dopamine
Short title: Control over brain states during working memory impacted by dopamine function.
Word Count (unreferenced Abstract): 72/70
Word Count (Article Body): 1488/1500
Number of Figures: 2/3
1
Number of References: 20/20
This paper contains Supplementary Materials.
2
Dynamical brain state transitions are critical for flexible working memory but the network
mechanisms are incompletely understood. Here, we show that working memory entails brain-
wide switching between activity states. The stability of states relates to dopamine D1
receptor gene expression while state transitions are influenced by D2 receptor expression
and pharmacological modulation. Schizophrenia patients show altered network control
properties, including a more diverse energy landscape and decreased stability of working
memory representations.
Working memory is an essential part of executive cognition depending on prefrontal neurons
functionally modulated through dopamine D1 and D2 receptor activation (1-3). The dual-state
theory of prefrontal dopamine function links the differential activation of dopamine receptors
to two discrete dynamical regimes: a D1-dominated state with a high energy barrier favoring
robust maintenance of cognitive representations and a D2-dominated state with a flattened
energy landscape enabling flexible switching between states (4). Recent accounts extend the
idea of dopamine's impact on working memory from a local prefrontal to a brain-wide network
perspective (5, 6), but the underlying neural dynamics and brain-wide interactions have
remained unclear.
Network control theory (NCT) can be used to model brain network dynamics as a function of
interconnecting white matter tracts and regional control energy (7). Based on the
connectome, NCT can be used to examine the landscape of brain activity states: that is,
which states within a dynamic scheme would the system have difficulty accessing, and more
importantly, which regions need to be influenced (and to what extent) to make those states
accessible (8). Specifically, to quantify accessibility, we approximate brain dynamics locally
by a simple linear dynamical system, 𝑥(𝑡)=𝑨𝑥(𝑡)+𝑩𝑢(𝑡), where x (t) is the brain state
inferred from functional magnetic resonance imaging (fMRI), A is a structural connectome
inferred from DTI data, u is the control input, and B is a matrix describing which regions enact
3
control. To investigate, based on this conception, how the brain transitions between different
cognitive states, we defined states as individual brain activity patterns related to a working
memory condition (2-back) and to an attention control condition requiring motor response (0-
back) in a sample of 178 healthy individuals undergoing fMRI (Fig. S1; Online Methods).
Further, we obtained individual structural connectomes from white matter by DTI fiber
tracking, and computed the optimal control energy necessary to drive the dynamical system
from the 0-back activity pattern to the 2-back pattern, or vice versa (Fig. S2).
We defined the stability of both brain states as the inverse energy necessary to revisit that
state, where the energy, loosely, is defined as the average size of the control signals u(t)
needed to instantiate a specific trajectory in the dynamical system as defined above (see Eq.
3 & 5 in Online Methods). As expected, the cognitively more demanding 2-back state was
less stable (i.e., required higher energy for maintenance) than the control state (Fig. 1a;
repeated measures ANOVA: main effect of 0- vs. 2-back stability: F(1,173) = 66.80, p <
0.001, see Online Methods for details on all analyses). Further, the stability of the 2-back
state was significantly associated with working memory accuracy (Fig. 1b; b = 0.274, p =
0.006), suggesting that more stable 2-back network representations support higher working
memory performance. We next investigated how the brain flexibly changes its activity pattern
between states. Transitioning into the cognitively more demanding 2-back state required
more control energy than the opposite transition (Fig. 1c; repeated measures ANOVA:
F(1,174) = 27.98, p = 0.001). Other analyses suggested that prefrontal and parietal cortices
steer both types of transitions, while default mode areas are preferentially important for the
switch to the more cognitively demanding state (Fig. 1d; Online Methods). These results are
in line with the assumed role of frontal-parietal circuits in steering brain dynamics (9) and
shifting brain connectivity patterns (10); they also emphasize the importance of the
coordinated behavior of brain systems commonly displaying deactivations during demanding
cognitive tasks (11).
4
Following from the dual-state theory of network function, the stability of task-related brain
states should be related to prefrontal D1 receptor status. To estimate individual prefrontal D1
receptor expression, we utilized methods relating prefrontal cortex D1 and D2 receptor
expression to genetic variation in their co-expression partner (Online Methods), thereby
enabling us to predict individual dopamine receptor expression levels from genotype data
across the whole genome (12, 13). We found that D1 (but not the D2) expression-related
gene score predicted stability of both states (Fig. 2a; 0-back: b = 0.184, p = 0.034; 2-back: b
= 0.242, p = 0.007, Online Methods), in line with the assumed role of D1-related signaling in
maintaining stable activity patterns during task performance (4, 14).
Independent of stability, switching between different activity representations should relate to
dopamine D2 receptor function. Indeed, when controlling for stability as a nuisance covariate
in the regression model, the control energy of both state transitions could be predicted by the
D2 (but not the D1) receptor expression gene score (Fig. 2b; 0- to 2-back: b = -0.076, p =
0.037; and trending for 2- to 0-back: b = -0.134, p = 0.068, Online Methods). This finding is
particularly interesting, as it suggests that the function of D1 and D2 receptors are
differentially, but cooperatively, involved in steering brain dynamics between different activity
patterns, in line with previous research on D1 and D2 functioning in prefrontal circuits (4, 15).
Our results thus far support the notion that the brain is a dynamical system in which the
stability of a state is substantially defined by cognitive effort and modulated by D1 receptor
expression, while transitions between states depend primarily on D2 receptor expression. If
true, such a system should be sensitive to dopaminergic manipulation, and interference with
D2-related signaling should reduce the brain's ability to control its optimal trajectories, i.e.
increase the control energy needed when switching between states. To test these
hypotheses, we investigated an independent sample of healthy controls (n=16, Table S2)
receiving 400 mg Amisulpride, a selective D2 receptor antagonist, in a randomized, placebo-
controlled, double-blind pharmacological fMRI study. As expected, we observed that greater
control energy was needed for transitions under D2 receptor blockade (Fig. 2c; repeated
5
measures ANOVA with drug and transition as within-subject factors; main effect of drug:
F(1,10) = 7.27, p = 0.022; drug-by-condition interaction: F(1,10) = 0.42, p = 0.665). We
observed no effect on the stability of states; that is, the inverse control energy required to
stabilize a current state (main effect of drug: F(1,8) = 0.715, p = 0.422, Table S3).
Dopamine dysfunction, working memory deficits, and alterations in brain network
organization are hallmarks of schizophrenia (16-19). We therefore tested for differences in
the state stability and in the ability to control state transitions between schizophrenia patients
and a healthy control sample balanced for age, sex, performance, head motion, and
premorbid IQ (see Table S1). Stability in schizophrenia patients was reduced for the
cognitively demanding working memory state (F(1,98) = 6.43, p = 0.013), but not for the
control condition (F(1,98) = 0.052, p = 0.840, Table S3). Control energy needed for the 0- to
2-back transition was significantly higher in schizophrenia (Fig. 2d; F(1,98) = 5.238, p =
0.024), while the opposite transition showed no significant group difference (ANOVA: F(1,98)
= 0.620, p = 0.433, Table S3), in line with clinical observations that D2 blockade does not
ameliorate cognitive symptoms in schizophrenia (20). These results suggest that the brain
energy landscape is more diverse in schizophrenia, making the system more difficult to steer
appropriately. To further strengthen this notion, we estimated the variability in suboptimal
(higher energy) trajectories connecting different of cognitive states (Online Methods). We
expected that in a diversified energy landscape, the variation of trajectories around the
minimum-energy trajectory should be larger, implying that small perturbations may have a
more substantial impact. In line with our hypothesis, we found that the variability in such
perturbed trajectories was indeed increased in schizophrenia (rm-ANOVA: main effect of
group: F(1,98) = 4.789, p = 0.031, Online Methods).
Several aspects of our work require special consideration. Firstly, to relate brain dynamics to
cognitive function, we focus on discrete brain states where each state is summarized by a
single brain activation patterns rather than linear combination of multiple brain activity
patterns. Secondly, although we could demonstrate a link between brain dynamics,
6
measured by means of control energy, and predicted prefrontal dopamine receptor
expression, the link is indirect and requires confirmation by direct measurements. Thirdly, we
cannot exclude the possibility that disorder severity, duration, symptoms or medication may
have influenced network dynamics in schizophrenia patients, although our supplemental
analyses do not support this conclusion (Online Methods). Finally, while the sample sizes of
our pharmacological and patient study are rather small, we were able to show comparable
effects of dopaminergic manipulation on control properties using a second (Online Methods),
further supporting the validity of the underlying rationale.
In summary, our data demonstrate the utility of network control theory for the non-invasive
investigation of the mechanistic underpinnings of (altered) brain states and their transitions
during cognition. Our data suggest that engagement of working memory involves brain-wide
switching between activity states and that the steering of these network dynamics is
differentially, but cooperatively, influenced by dopamine D1 and D2 receptor function.
Moreover, we show that schizophrenia patients show reduced controllability and stability of
working memory network dynamics, consistent with the idea of an altered functional
architecture and energy landscape of cognitive brain networks.
7
Acknowledgements
The authors thank all individuals who have supported our work by participating in our studies.
There was no involvement by the funding bodies at any stage of the study. We thank Oliver
Grimm, Leila Haddad, Michael Schneider, Natalie Hess, Sarah Plier and Petya Vicheva for
valuable research assistance. The authors thank Jason Kim and Lorenzo Caciagli for
valuable feedback on the manuscript.
U.B. acknowledges grant support by the German Research Foundation (DFG, grant BR
5951/1-1). H.T. acknowledges grant support by the German Research Foundation (DFG,
Collaborative Research Center SFB 1158 subproject B04, Collaborative Research Center
TRR 265 subproject A04, GRK 2350 project B2, grant TO 539/3-1) and German Federal
Ministry of Education and Research (BMBF, grants 01EF1803A project WP3, 01GQ1102).
AML acknowledges grant support by the German Research Foundation (DFG, Collaborative
Research Center SFB 1158 subproject B09, Collaborative Research Center TRR 265
subproject S02, grant ME 1591/4-1) and German Federal Ministry of Education and Research
(BMBF, grants 01EF1803A, 01ZX1314G, 01GQ1003B), European Union's Seventh
Framework Programme (FP7, grants 602450, 602805, 115300 and HEALTH-F2-2010-
241909, Innovative Medicines Initiative Joint Undertaking (IMI, grant 115008) and Ministry of
Science, Research and the Arts of the State of Baden-Wuerttemberg, Germany (MWK, grant
42-04HV.MED(16)/16/1). DSB and RBF would like to acknowledge support from the John D.
and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Army Research
Laboratory and the Army Research Office through contract numbers W911NF-10-2-0022 and
W911NF-14-1-0679, the National Institute of Health (2-R01-DC-009209-11, 1R01HD086888-
01, R01-MH107235, R01-MH107703, and R21-M MH-106799), the Office of Naval Research,
and the National Science Foundation (BCS-1441502, CAREER PHY-1554488, and BCS-
1631550). E.S.
gratefully
acknowledges
grant
support
by
the Deutsche
Forschungsgemeinschaft, DFG (SCHW 1768/1-1). X.L.Z. is a Ph.D. scholarship awardee of
the Chinese Scholarship Council. DD acknowledges grant support by the German Research
8
Foundation (DFG, Du 354/10-1). G.P. has received funding from the European Union's
Horizon 2020 research and innovation program under the Marie Skłodowska-Curie No.
798181: "IdentiFication of brain deveLopmental gene co-expression netwOrks to Understand
RIsk for SchizopHrenia" (FLOURISH).
The content of this paper is solely the responsibility of the authors and does not necessarily
represent the official views of any of the funding agencies
9
Financial disclosures
A.M.-L. has received consultant fees from Blueprint Partnership, Boehringer Ingelheim,
Daimler und Benz Stiftung, Elsevier, F. Hoffmann-La Roche, ICARE Schizophrenia, K. G.
Jebsen Foundation, L.E.K Consulting, Lundbeck International Foundation (LINF), R.
Adamczak, Roche Pharma, Science Foundation, Synapsis Foundation -- Alzheimer
Research Switzerland, System Analytics, and has received lectures including travel fees
from Boehringer Ingelheim, Fama Public Relations, Institut d'investigacions Biomèdiques
August Pi i Sunyer (IDIBAPS), Janssen-Cilag, Klinikum Christophsbad, Göppingen, Lilly
Deutschland, Luzerner Psychiatrie, LVR Klinikum Düsseldorf, LWL PsychiatrieVerbund
Westfalen-Lippe, Otsuka Pharmaceuticals, Reunions i Ciencia S. L., Spanish Society of
Psychiatry, Südwestrundfunk Fernsehen, Stern TV, and Vitos Klinikum Kurhessen. A.B. has
received consultant fees from Biogen and speaker fees from Lundbeck, Otsuka, Recordati,
and Angelini.
The remaining authors reported no biomedical financial interests of potential conflicts of
interest.
10
References
Kim JZ, Soffer JM, Kahn AE, Vettel JM, Pasqualetti F, Bassett DS (2018): Role of Graph
Cole MW, Reynolds JR, Power JD, Repovs G, Anticevic A, Braver TS (2013): Multi-task
Durstewitz D, Seamans JK (2008): The dual-state theory of prefrontal cortex dopamine
Arnsten AF (2011): Catecholamine influences on dorsolateral prefrontal cortical networks.
Roffman JL, Tanner AS, Eryilmaz H, Rodriguez-Thompson A, Silverstein NJ, Ho NF, et al.
Betzel RF, Gu S, Medaglia JD, Pasqualetti F, Bassett DS (2016): Optimally controlling the
Ferenczi EA, Zalocusky KA, Liston C, Grosenick L, Warden MR, Amatya D, et al. (2016):
Meyer-Lindenberg A, Kohn PD, Kolachana B, Kippenhan S, McInerney-Leo A, Nussbaum R, et
Goldman-Rakic PS (1995): Cellular basis of working memory. Neuron. 14:477-485.
Ott T, Jacob SN, Nieder A (2014): Dopamine receptors differentially enhance rule coding in
1.
2.
primate prefrontal cortex neurons. Neuron. 84:1317-1328.
3.
al. (2005): Midbrain dopamine and prefrontal function in humans: interaction and modulation by
COMT genotype. Nat Neurosci. 8:594-596.
4.
function with relevance to catechol-o-methyltransferase genotypes and schizophrenia. Biol
Psychiatry. 64:739-749.
5.
Biol Psychiatry. 69:e89-99.
6.
(2016): Dopamine D1 signaling organizes network dynamics underlying working memory. Sci Adv.
2:e1501672.
7.
Architecture in Controlling Dynamical Networks with Applications to Neural Systems. Nat Phys.
14:91-98.
8.
human connectome: the role of network topology. Sci Rep. 6:30770.
9.
Prefrontal cortical regulation of brainwide circuit dynamics and reward-related behavior. Science.
351:aac9698.
10.
connectivity reveals flexible hubs for adaptive task control. Nat Neurosci. 16:1348-1355.
11.
Greicius MD, Krasnow B, Reiss AL, Menon V (2003): Functional connectivity in the resting
brain: a network analysis of the default mode hypothesis. Proceedings of the National Academy of
Sciences of the United States of America. 100:253-258.
12.
Fazio L, Pergola G, Papalino M, Di Carlo P, Monda A, Gelao B, et al. (2018): Transcriptomic
context of DRD1 is associated with prefrontal activity and behavior during working memory. Proc
Natl Acad Sci U S A. 115:5582-5587.
13.
expression network and a related polygenic index predict imaging, behavioral and clinical
phenotypes linked to schizophrenia. Transl Psychiatry. 7:e1006.
14.
Dopaminergic modulation of distracter-resistance and prefrontal delay period signal.
Psychopharmacology. 232:1061-1070.
15.
differential D1 versus D2 dopamine receptor regulation of inhibition in prefrontal cortex. J Neurosci.
24:10652-10659.
16.
common pathway. Schizophrenia bulletin. 35:549-562.
17.
CNTRICS and schizophrenia. Biol Psychiatry. 64:11-17.
18.
Tost H, Alam T, Meyer-Lindenberg A (2010): Dopamine and psychosis: theory,
pathomechanisms and intermediate phenotypes. Neurosci Biobehav Rev. 34:689-700.
19.
Braun U, Schafer A, Bassett DS, Rausch F, Schweiger JI, Bilek E, et al. (2016): Dynamic brain
network reconfiguration as a potential schizophrenia genetic risk mechanism modulated by NMDA
receptor function. Proc Natl Acad Sci U S A. 113:12568-12573.
20. Millan MJ, Fone K, Steckler T, Horan WP (2014): Negative symptoms of schizophrenia: clinical
characteristics, pathophysiological substrates, experimental models and prospects for improved
Bloemendaal M, van Schouwenburg MR, Miyakawa A, Aarts E, D'Esposito M, Cools R (2015):
Pergola G, Di Carlo P, D'Ambrosio E, Gelao B, Fazio L, Papalino M, et al. (2017): DRD2 co-
Trantham-Davidson H, Neely LC, Lavin A, Seamans JK (2004): Mechanisms underlying
Howes OD, Kapur S (2009): The dopamine hypothesis of schizophrenia: version III--the final
Barch DM, Smith E (2008): The cognitive neuroscience of working memory: relevance to
11
treatment. European neuropsychopharmacology : the journal of the European College of
Neuropsychopharmacology. 24:645-692.
12
Figures
Figure 1:
Controllability and stability of brain dynamics during working memory
A) The stability of the 2-back state reflecting working memory activity is lower than that of the
0-back state reflecting motor and basic attention control activity (F(1,173) = 66.80, p <
0.001). Red lines indicate mean values and boxes indicate one standard deviation of the
mean. B) Associations of 2-back stability with working memory performance (accuracy: b =
0.274, p = 0.006; covarying for age, sex, and mean activity). C) Steering brain dynamics from
the control condition to the working memory condition requires more control energy than vice
versa (F(1,174) = 27.98, p < 0.001). D) Unique and common sets of brain regions contribute
most to the transition from 0-back to 2-back and the transition from 2-back to 0-back
transitions, respectively. For illustrative purposes, we projected the computed control impact
13
of each brain region (Online Methods) for the respective transitions on a 3D structural
template, displaying the 20% highest for each transition.
Figure 2:
Dopamine receptor expression and pharmacological modulation impact whole brain
dynamics
A) Genetic scores predicting DRD1 expression in prefrontal regions positively predict stability
of both brain states (0-back: b = 0.184, p = 0.034; 2-back: b = 0.242 p = 0.007; age, sex,
mean brain state activity, first 5 genetic PCA components as covariates of non-interest). B)
Genetic scores predicting DRD2 expression in prefrontal regions negatively predict control
energy for both brain state transitions (0-back to 2-back: b = -0.076, p = 0.037; and trend
wise for 2-back to 0-back: b = -0.134, p = 0.068; age, sex, mean brain activity difference, first
14
5 genetic PCA components, stability of 0-back and 2-back as covariates of non-interest). C)
Amisulpride increases control energy for transitions in comparison to placebo (main effect of
drug: F(1,10) = 7.27, p = 0.022; interaction drug by condition: F(1,10) = 0.42, p = 0.665,
activity difference, drug order, and sex as covariates of non-interest). Black lines indicate
mean values and boxes indicate one standard deviation of the mean. D) Schizophrenia
patients need more control energy when transitioning into the working memory condition than
matched healthy controls (F(1,98) = 5.238, p = 0.024, age, sex, tSNR and mean activity as
covariates of non-interest), but not vice versa.
15
ONLINE METHODS
Urs Braun et al. "Brain state stability during working memory is explained by network control
theory, modulated by dopamine D1/D2 receptor function, and diminished in schizophrenia"
6.1.
6.2.
6.3.
7.1.
7.2.
2.1.
2.2.
1. Participants and study design ....................................................................................................... 2
2. Data acquisition .............................................................................................................................. 3
fMRI ...................................................................................................................................................... 3
Diffusion Tensor Imaging (DTI) ............................................................................................................... 4
3. Atlas construction .......................................................................................................................... 4
4. Connectome construction .............................................................................................................. 5
5. Brain state definition ...................................................................................................................... 5
6. Network control theory ................................................................................................................... 6
Optimal control theory framework ......................................................................................................... 6
Control energy, stability, and impact ..................................................................................................... 7
Suboptimal trajectories ......................................................................................................................... 8
7. Gene based polygenic co-expression indices .............................................................................. 9
Genotyping, imputation and quality control ........................................................................................... 9
Polygenic co-expression index calculation ............................................................................................ 9
8. Statistical inference ...................................................................................................................... 11
9. Control analyses ........................................................................................................................... 12
Null models of structural brain networks .............................................................................................. 12
Null models of spatial activity patterns ................................................................................................ 12
Robustness to choice of parcellation scheme ..................................................................................... 12
Robustness to choice of edge definition .............................................................................................. 13
Impact of medication and duration of illness on control properties ...................................................... 13
Pharmacological validation using Risperidone ..................................................................................... 13
Null results for gene score and imaging associations .......................................................................... 14
Suboptimal trajectories ....................................................................................................................... 14
Supplementary figures ............................................................................................................. 15
10.1. Figure S1: N-back task design ............................................................................................................ 15
10.2. Figure S2: Network control theory concept and methods .................................................................... 16
Supplementary tables ............................................................................................................... 17
11.1. Table S1: Characteristics of the healthy control and patient samples .................................................. 17
11.2. Table S2: Sample characteristics of the pharmacological study .......................................................... 18
11.3. Table S3: Statistical details for the main findings and replication analyses .......................................... 19
11.4. Table S4: List of brain regions included in the extended Glasser parcellation ...................................... 21
Supplementary references ....................................................................................................... 24
9.1.
9.2.
9.3.
9.4.
9.5.
9.6.
9.7.
9.8.
10.
11.
12.
1
1.
Participants and study design
All participants provided written informed consent for protocols approved by the Institutional Review Board
of the medical faculty in Mannheim. For the first study including healthy controls and patients with
schizophrenia, a total of 202 subjects (178 healthy controls, 24 schizophrenia patients) were included (see
Table S1). General exclusion criteria in controls included the presence of a lifetime history of psychiatric,
neurological, or significant general medical illness, pregnancy, a history of head trauma, and current alcohol
or drug abuse. The patients were recruited from the Department of Psychiatry and Psychotherapy at the
Central Institute of Mental Health in Mannheim and via local advertisements. A trained psychiatrist or
psychologist verified the diagnosis of schizophrenia based on ICD-10 criteria.
Neuropsychological characterization of healthy controls included the trail-making-test B (TMT-B) (1) and the
German multiple-choice vocabulary intelligence test (MWT-B) (2) as a measure of premorbid IQ. Clinical
characterization included the assessment of current symptom severity using the Positive and Negative
Symptom Scale (PANSS) (3), Beck's Depression Inventory (4), measures of global functioning in daily life
(global assessment of function (GAF)) and current antipsychotic medication dosage (converted into
chlorpromazine dose equivalents (CPZE)).
For the second, pharmacological intervention study, 17 healthy individuals completed a subject- and
observer-blind, placebo-controlled, randomized three-period cross-over study (see Table S2). Exclusion
criteria included a regular consumption of drugs or history of drug or alcohol abuse; systolic blood pressure
(SBP) greater than 140 or less than 90 mm Hg, and diastolic blood pressure (DBP) greater than 90 or less
than 50 mm Hg; notable resting bradycardia (heart rate (HR) <40 bpm) or tachycardia (HR >90 bpm); use of
any medication or herbal remedies taken within 14 days prior to randomization into the study or 5 times the
elimination half-life of the medication, clinically significant abnormalities in laboratory test results (including
hepatic and renal panels, complete blood count, chemistry panel and urinalysis): a history or presence of
clinically significant ECG abnormalities (e.g. PQ/PR interval >210 ms, QTcF >450 ms) or cardiovascular
disease (e.g. cardiac insufficiency, coronary artery disease, cardiomyopathy, hypokalemia, congestive heart
failure, family history of congenital long QT syndrome, family history of sudden death); any personal or familial
2
history of seizures, epilepsy or other convulsive condition, previous significant head trauma, or other factors
predisposing to seizures; disorders of the central nervous system, cerebrovascular events, Parkinson's
disease, migraine, depression, bipolar disorder, anxiety, any other psychiatric disorders or behavioral
disturbances; regular smoking (>5 cigarettes, >3 pipe-fulls, >3 cigars per day); habitual caffeine consumption
of more than 400 mg/d (approximately 4 cups of coffee or equivalent); a history or evidence of any clinically
significant endocrinological, hepatic, renal, autoimmune, pulmonary, gastrointestinal, urogenital, oncological,
hematological or any other disease; or a body mass index (BMI) of over 30 or below 22.
Participants were invited for a fixed interval of 7 days with each scanning session taking place at
approximately the same time of day. On each of three scanning visits, individuals either received a single
oral dose of 400 mg Amisulpride, 3 mg Risperidone or Placebo. MRI scanning took place 2 hours after drug
administration, with the N-back paradigm commencing approximately 10 min after the start of the scan.
One subject was excluded from the analysis due to an excessive body-mass index (BMI > 30).
2.
Data acquisition
2.1.
fMRI
For the first study, BOLD fMRI was performed on a 3T Siemens Trio (Erlangen, Germany) in Mannheim,
Germany. Prior to the acquisition of functional images, a high-resolution T1-weighted 3D MRI sequence
was conducted (MPRAGE, slice thickness=1.0 mm, FoV = 256 mm, TR = 1570ms, TE = 2.75 ms, TI =
800ms, α = 15°). Subsequently, functional data was acquired during performance of the N-back paradigm
using an echo-planar imaging (EPI) sequence with the following scanning parameters: TR/ TE = 2000/30
ms, α =80°, 28 axial slices (slice-thickness = 4 mm + 1 mm gap), descending acquisition, FoV = 192 mm,
acquisition matrix = 64 x 64 128 volumes.
For the second study, BOLD fMRI was performed on a 3T Siemens Trio Scanner (Erlangen, Germany)
using an echo-planar imaging (EPI) sequence with the following parameters: TR = 1790 ms, TE = 28 ms,
34 axial slices per volume, voxel size = 3 x 3 x 3 mm, 1 mm gap, 192 x 192 mm field of view, 76° flip angle,
descending acquisition.
3
The visual N-back paradigm is a well-established and reliable working memory task consisting of a high
memory load (2-back) and an attention control condition requiring motor response (0-back) (5-7).
Specifically, a diamond-shaped stimulus containing a number from 1 to 4 was presented every 2 seconds
(see Figure S1). In the 0-back condition, subjects were required to press the button on the response box
corresponding to the number currently displayed on the presentation screen. In the 2-back condition,
subjects were required to press the button on the response box corresponding to the number presented
two stimuli before the number currently displayed on the presentation screen. Stimuli were presented in
alternating blocks of either 0-back or 2-back conditions. In each condition block, 14 stimuli were presented.
Each condition block was repeated 4 times. Task performance was measured by accuracy (defined as the
percent of correct answers) and reaction time (defined as the time span between stimulus onset and button
press) for each condition separately.
2.2. Diffusion Tensor Imaging (DTI)
DTI data were acquired by using spin echo EPI sequences with the following parameters: TR 14000ms, TE
86ms, 2mm slice thickness, 60 non-collinear directions, b-value 1000 s/mm2, 1 b0 image, FOV 256 mm.
3.
Atlas construction
To combine structural and functional brain imaging data, we first constructed a brain atlas that equally well
respects functional and anatomical features. We transformed a recently published multimodal atlas (8) into
a volumetric format by projecting its FreeSurfer pial cortex coordinates into standard MNI space. A grey
matter prior probability map (thresholded at 0.3) as provided by SPM was used to define relevant voxels.
Voxels were labeled by choosing the closest label with maximum distance of 4 mm. Since the published
multimodal atlas does not cover all subcortical regions of interest (e.g. amygdala, thalamus), we
complemented it with subcortical structures from the Harvard-Oxford atlas as implemented in FSL (9).
Combining the two atlases resulted in 374 regions that covered cortical and subcortical structures. A full list
of regions included in the combined atlas can be found in Supplemental Table S3.
4
4.
Connectome construction
For the DTI data, the following preprocessing steps were performed with standard routines implemented in
the software package FSL (9): i) correction of the diffusion images for head motion and eddy currents by
affine registration to a reference (b0) image, ii) extraction of non-brain tissues (10), and iii) linear diffusion
tensor fitting. After estimation of the diffusion tensor, we performed deterministic whole-brain fiber tracking
as implemented in DSI Studio using a modified FACT algorithm (11). For each subject, 1,000,000
streamlines were initiated. Streamlines with a length of less than 10 mm were removed (12). For the
construction of structural connectivity matrices, the brain was parcellated into 374 regions. To map these
parcellations into subject space, we applied a nonlinear registration implemented in DSI Studio. We
estimated the structural connectivity between any two regions of the atlas by using the mean fractional
anisotropy values between respective brain regions. This procedure resulted in a weighted adjacency
matrix A whose entries Aij reflect the strength of structural connectivity between two brain regions.
5.
Brain state definition
Because we were interested in investigating how the brain controls and transitions between global brain
states underlying circumscribed cognitive processes (such as those supporting working memory, attention,
and motor behavior), we defined brain states as stationary patterns of activity during execution of these
processes. It is important to note that the temporal resolution of fMRI and the design of the N-back task
limits the investigation of differential cognitive processes contributing to memory performance on fast time
scales. Therefore, we cannot investigate the detailed temporal dynamics of working memory processes.
However, our simplified design allows us to extract meaningful brain states elicited by a controlled cognitive
process and therefore enables us to relate brain dynamics to cognitive function.
Specifically, we defined individual brain states as spatial patterns of beta estimates associated with activity
across brain regions of interest during both conditions of the N-back task (13). For that purpose, after
standard preprocessing procedures in SPM12/8 (including realignment to the mean image, slice time
correction, spatial normalization into standard stereotactic space, resampling to 3 mm isotropic voxels, and
smoothing with an 8 mm full-width at half-maximum Gaussian Kernel) (5, 6, 14), we estimated standard
5
first-level general linear models for the N-back task, separately for each individual. Except for the SPM
version, both studies followed the same preprocessing procedure. These GLM models included regressors
for the 0-back and 2-back conditions of interest, as well as the 6 motion parameters as regressors of non-
interest. To define the brain activity pattern associated with each condition of the task, we extracted GLM
(beta) parameter estimates for the 0-back and 2-back conditions separately (13) and we averaged them
across all voxels in each of the 374 regions without applying any threshold. This procedure yielded a 374
by 1 vector for each condition (x0-back, x2-back) per subject representing how strongly the BOLD response in
back), respectively. These vectors (x0-back, x2-back) defined the final and target brain states for the following
each brain region was associated with the working memory (2-back) or the motor control condition (0-
network control analyses.
6.
Network control theory
6.1. Optimal control theory framework
To model the transition between 0-back and 2-back brain states, we used the framework of optimal control,
following prior work (15-17) implemented in MATLAB. Based on individual brain states X=[x1,…xn] (in our
case simplified to n = 2 states: 0-back and 2-back, see above) and a structural brain network A for each
(1) ẋ(t) = A x(t) + B u(t)
subject, we approximated the local brain dynamics by a linear continuous-time equation:
to model the flow among task-related brain activity states. In the model, x(t) is the state of the system at
time t, A is the wiring diagram of the underlying network, B denotes an input matrix defining the control
nodes, and u(t) is the time-dependent control signal (17, 18). Note that while the initial state (x0) and the
target state (xT) are empirically defined, any states of the system at other times are virtual intermediate
steps in the trajectories of the state-space model. The problem of finding an optimal control energy u* that
induces a trajectory from an initial state x0 to a target state xT reduces to the problem of finding an optimal
solution to the minimization problem of the corresponding Hamiltonian:
(2) min[H(p, x, u, t) = xTx + ñuTu + pT (Ax + Bu)] (15).
6
For simplicity, we set the input matrix B = IN´N, the identity matrix, allowing all brain regions to be
independent controllers (16, 18).
6.2. Control energy, stability, and impact
Control energy for each node ki, i = 1…m (m = total number of brain nodes), was defined as
(3) 𝐸𝑘𝑖=∫ D𝑢∗𝑘𝑖(𝑡)D2𝑑𝑡
𝑇𝑡=0
,
i.e. the squared integral over time of energy input that the node has to exhibit to facilitate the transitions
from the initial state to the target state (15, 16, 18). While the neurobiological foundations of control
properties in the brain are not yet well understood, control energy can be interpreted as the effort of a brain
region needed to steer the activity pattern of itself and its connected brain regions into the desired final
activation state, for example by tuning their internal firing or activity patterns by recurrent inhibitory
connections. Accordingly, the total control energy for the entire brain was defined as the sum of all control
energy across all nodes
(4) 𝐸=∑ 𝐸KL
MNOP
,
yielding one value for each transition per subject. To ensure a normal distribution of metric values for
subsequent statistical testing, we applied a logarithmic transformation (base 10) to the control energy(16).
From the control energy, we can also obtain the control stability and control impact. Stability was defined as
(5) 𝑆=
P
RSTUV(WXYZX[),
i.e. the inverse control energy needed to maintain a state, or in other words, the control energy needed to
go, e.g., from 2-back to 2-back (18). To investigate the influence that a single brain region has on the entire
system's dynamics during state trajectories, we computed the control impact of each node by iteratively
removing one brain region from the network and re-computing the change in control energy (15).
7
6.3. Suboptimal trajectories
To investigate the energy landscape surrounding the minimum energy (in this sense 'optimal') trajectories,
we quantified the variability of suboptimal trajectories by adding subtle random perturbation to the minimum
energy trajectories over 200 iterations. Note that we employed a discrete-time dynamical system rather
than a continuous-time system for these analyses, as discrete-time systems are computationally more
tractable. To discretize our linear continuous-time system, we employed the following transformations
(6) 𝑨] ≜ 𝑒𝑨'a and
(7) 𝑩]≜ c∫ 𝑒d('aef)
'ag
𝑑𝜏i𝑩,
where 𝑨] and 𝑩] are the corresponding structural matrix and control input matrix in a discrete time system
and Ts is the sampling time. As there is no prescriptive way to choose Ts, we estimated 𝑇m= PPg 𝑅𝑇, where
RT is the rise time of its fastest mode, i.e. the time that the system requires to go from 10% to 90% of its
fastest step response. In the resulting discrete-time dynamical system
(8) 𝑥 (𝑡+1)= 𝑨]𝑥(𝑡)+𝐵s𝑢s(𝑡),
at each discrete time step t, we applied a principal component analysis (PCA) to the cloud of suboptimal
points to reduce dimensionality. Because the previous simulations show that the first PCA component
explains more than 90% of variance, we continued by using the first component as a summary measure of
suboptimal trajectories. To estimate the relative distance between suboptimal and optimal trajectories, we
computed the percentage of variation of the maximum distance of the projected suboptimal points on the
first principal component from the optimal one. This procedure resulted in a normalized measure giving the
percentage of deviations for the 200 suboptimal trajectories from the optimal trajectory. Due to the heuristic
nature of the algorithm applying a random perturbation, results can vary from run to run. Therefore, to
increase replicability of our results, we repeated our analysis 10 times per subject (10*200 = 2,000
suboptimal trajectories per subject).
Because DTI data was not acquired during the pharmacological intervention study, we used the average
connectivity matrix across all healthy subjects from study 1 to model the transition between individual brain
states in the pharmacological intervention study.
8
7.
Gene based polygenic co-expression indices
7.1. Genotyping, imputation and quality control
In this study, we used human GWAS data of 63 healthy subjects who were genotyped using HumanHap
610 and 660w Quad BeadChips. For all subjects, standard quality control (QC) and imputation were
performed using the Gimpute pipeline (Chen, Lippold et al. 2018) and the following established QC steps
were applied. Step 1: Determine the number of male subjects for each heterozygous SNP of the
chromosome 23 and remove SNPs whose number is larger than 5% of the number of male samples. Step
2: Determine the number of heterozygous SNPs on X chromosome for each male sample, and remove
samples that have the number of heterozygous SNPs larger than 10. Step 3: Remove SNPs with missing
genotyping rate > 5% before sample removal. Step 4: Exclude samples with missingness >= 0.02. Step 5:
Exclude samples with autosomal heterozygosity deviation Fhet >= 0.2. Step 6: Remove SNPs with the
proportion of missing genotyping > 2% after sample removal. Step 7: Remove SNPs if the Hardy-Weinberg
equilibrium exact test P-value was < 1 x 10-6. Step 8: Principal component analysis (PCA) was applied to
detect population outliers. Imputation was carried out using IMPUTE2/SHAPEIT (19-21), with a European
reference panel for each study sample in each 3 Mb segment of the genome. This imputation reference set
is from the full 1000 Genome Project dataset (August 2012, 30,069,288 variants, release "v3.macGT1").
The length of the buffer region is set to be 500 kb on either side of each segment. All other parameters
were set to default values implemented in IMPUTE2. After imputation, SNPs with high imputation quality
(INFO >= 0.6) and successfully imputed in >= 20 samples were retained. From the final well-imputed
dataset with 63 subjects, we extracted 8 SNPs for DRD2 and 13 SNPs for DRD1 (22-24). For the
subsequent genetic imaging analyses, we only used subjects for whom both data modalities (DTI and fMRI
data) were available.
7.2. Polygenic co-expression index calculation
Previous publications have shown that gene sets defined using co-expression networks and selected for their
association with the genes DRD1 and DRD2 provided replicable predictions of n-back-related brain activity
and behavioral indices (23, 25-27). Weighted Gene Co-expression Network Analysis [WGCNA (28)] applied
9
on the Braincloud dataset (N=199) of post-mortem DLPFC gene expression (29) identified 67 non-
overlapping sets of genes based on their expression pattern. The co-expression gene sets including DRD1
and DRD2 were summarized into Polygenic Co-expression Indices (PCIs) based on SNPs that predicted co-
expression of these genes (called co-expression quantitative trait loci, or co-eQTLs).
PCIs are a proxy for the assessment of the genetic component of gene transcription co-regulation and are
computed as a weighted average of the effect of all genotypes of an individual among those selected in the
data mining study as co-eQTLs. The effect of individual SNPs is computed as the difference between the
gene co-expression distribution of minor allele carriers (heterozygotes and homozygotes) and that of major
allele homozygotes, using common tools from signal detection theory (30). Genotype weights, therefore,
represent the deviation in gene co-expression from a reference distribution and are not constrained by
allele dose. For each genotype of each SNP we computed an index, called A', proportional to the
expression of the gene of interest (DRD1 or DRD2) within its co-expression module. The A' index is less
dependent than d' on the assumption of a normal distribution of gene expression in each genotypic
population (25). Both PCI-based predictions were significantly replicated in an independent post-mortem
dataset, while controlling for ethnicity. The translational effect of these two scores on brain activity during n-
back has been assessed and replicated across multiple samples, which combined amount to approximately
600 participants (22-24).
It is important to note that these dopamine-related genetic effects are large in magnitude compared to
those estimated by polygenic risk score approaches that focus on epidemiological data, rather than on
molecular processes. The DRD2-PCI we developed (23) yielded an effect size f = 0.30 in our n-back
discovery sample (required sample size to obtain 80% power with α = 0.05 and covariates as in the current
work: N = 71). Results were replicated in an independent fMRI dataset collected at a different institution
with f = 0.20 (required sample size computed as above: N = 156). Our follow-up work on the DRD2-PCI
(26) considered two datasets of 50 individuals each and yielded a minimum effect size f = 0.28 (in the
replication sample; required sample size computed as above: N = 81). The DRD1-PCI was also tested in
two independent samples (25), yielding a minimum effect size f = 0.37 (in the replication sample; required
sample size computed as above: N = 46). Taken together, these published results show that the effects of
these polygenic indices on n-back activity in the prefrontal cortex are relatively large, with sizes ranging
between 0.20 and 0.37 and with required samples ranging from 46 to 156 individuals. Importantly, the
10
DRD2-PCI was also tested in a small sample of 29 patients with SCZ and yielded results consistent with
the effects discovered in healthy controls (23). Although the required sample sizes were computed based
on the top cluster, it should be borne in mind that the technique we used in this work employs the entire
brain, and therefore (i) is not subject to correction for multiple comparisons, as reflected in the uncorrected
alpha used for the power calculations and (ii) benefits from the greatest possible amount of information
about brain states.
8.
Statistical inference
Statistical inference was performed using the Statistical Package for the Social Sciences 24 (IBM SPSS
Statistics for Windows, Version 24). All statistical comparisons were performed while controlling for age and
sex and were tested two-sided. Because we were interested in control properties of brain state transitions
independent of differences in the amount of mere activation, we controlled for the respective parameters
reflecting the individual differences in activations in all analyses involving control properties. In particular,
for all analyses involving stability measures, we additionally controlled for the average brain energy defined
as the GLM parameter estimates over all regions in the 0-back and 2-back conditions. As the control
energy of a transition depends highly on the absolute difference in energy between its initial state and its
final state (i.e. in our case the brain-wide activation difference between both task conditions) and as we
were interested in the unique control properties independent of traditional activation differences, we
additionally controlled for the difference in the mean baseline energy in each analysis involving control
energy. Baseline energy was defined as the absolute average difference in the unthresholded GLM
parameter estimates over all regions between 0-back and 2-back conditions. As the same arguments apply
to stability measures, we controlled for the respective baseline energy of each condition in each analysis
involving stability.
In all analyses involving polygenic scores for D1-/D2-expression, we further used the first 5 principal
components from the PCA on the linkage-disequilibrium pruned set of autosomal SNPs to control for
population stratification. Differences between control energy/stability of both conditions were assessed
using repeated-measures ANOVA with condition as a within-subject factor. Drug effects were modeled
using a repeated measures ANOVA with drug and condition as within-subject factors. As healthy control
11
and schizophrenia patients differed in one of the quantified DTI imaging quality parameters (tSNR, see
Table S1), we also controlled for tSNR in each analysis involving both groups.
9.
Control analyses
9.1. Null models of structural brain networks
To study the impact of structural brain networks on control properties, we repeated the computation of
control energy using a randomized null model of the individuals' structural brain networks that preserves
the average weight, the strength distribution and the degree distribution. Null models were created using
the null_model_und_sign function as implemented in the Brain Connectivity Toolbox
(https://sites.google.com/site/bctnet/). In line with our expectation, control energy increased significantly for
randomized networks in both groups (repeated measures ANOVA with null_model_vs_data and transition
as within-subject factors, HC: main effect of null_model_vs_data, F(1,174) = 38.284, p < 0.001; SZ: F(1,20)
= 3.561, p = 0.074), suggesting that human brain structural networks are in some form optimized to control
brain state transitions, but schizophrenia patients tend to have less optimal, more random structural brain
networks.
9.2. Null models of spatial activity patterns
To study the impact of the spatial distribution of activity patterns on control properties, we repeated the
computation of control energy and spatially randomized individuals' brain activation patterns. In line with
our expectation, control energy increased significantly for randomized networks in both groups (repeated
measures ANOVA with model_vs_data and transition as within-subject factors, HC: main effect of
model_vs_data, F(1,174) = 6.995, p = 0.009; SZ: F(1,20) = 0.019, p < 0.0893), suggesting that the spatial
distribution of brain activity patterns is important for minimizing control effort.
9.3. Robustness to choice of parcellation scheme
To demonstrate the robustness of the results to our choice of parcellation scheme, we repeated our
analysis using a recently published functionally defined atlas comprising a similar number of areas (31).
Specifically, we used the "Gordon" template (31) consisting of 333 regions that are functionally derived from
12
resting-state connectivity analyses. Data were reprocessed using the same pipeline as for the main
analysis and all parameters were kept identical in the subsequent analysis. Notably, we replicated all main
results (see Table S4), indicating that our reported findings are robust to the choice of parcellation scheme.
9.4. Robustness to choice of edge definition
To demonstrate the robustness of our results to our selection of connectivity measure, we repeated our
analysis using the number of streamlines normalized by the respective size of the regions to construct
structural connectivity matrices (15). All parameters were kept identical in the subsequent analysis. All main
results could be replicated (see Table S4), indicating that our findings are robust to the choice of edge
definition.
9.5.
Impact of medication and duration of illness on control properties
In patients, the potential relationship between control energy and stability, antipsychotic drug dose
(expressed in chlorpromazine equivalents (CPZE), n=20), and clinical parameters (illness duration, illness
severity as indexed by global functioning (GAF) and Positive and Negative Symptom Scale (PANSS)) were
explored using Pearson correlation. Neither the control energy for the 0-back to 2-back transition nor the
opposite transition or the stability of either state were significantly associated with CPZE (N = 20, 0- to 2-
back: r = 0.078, p = 0.767; 2- to 0-back: r = 0.320, p = 0.210; 0- back stability: r = 0.150, p = 0.564; 2- back
stability: r = 0.096, p = 0.713), with illness duration (N = 23, 0- to 2-back: r = 0.017, p = 0.937; 2- to 0-back:
r = -0.226, p = 0.299; 0- back stability: r = 0.110, p = 0.644; 2- back stability: r = 0.281, p = 0.230), or with
GAF (N = 24, 0- to 2-back: r = -0.086, p = 0.690; 2- to 0-back: r = -0.254, p = 0.230; 0- back stability: r = -
0.135, p = 0.570; 2- back stability: r = 0.066, p = 0.793).
9.6. Pharmacological validation using Risperidone
To demonstrate the robustness of our pharmacological intervention of dopaminergic signaling, we
additionally analyzed the data of the Risperidone condition in the same subjects. Risperidone also
preferentially targets D2 receptors, but also affects D1, adrenergic, serotoninergic and histaminergic
pathways. Using the same models and covariates as in the main analysis, we detected a trend-wise
increase in control energy needed for both transitions (repeated measures ANOVA with drug and transition
as within-person factors; main effect of drug: F(1,10) = 3.490, p = 0.091; drug-by-condition interaction:
13
F(1,10) = 0.238, p = 0.636; activity difference, drug order, and sex as covariates of no interest), but no
effect on stability (F(1,8) = 0.105, p = 0.334; mean brain activity, sex, and drug order as covariates of no
interest). Although these results showed only trend-wise significance, likely due to the lower D2-specificity
of Risperidone, the detected pattern was conserved across drugs, validating the proposed underlying
concepts.
9.7. Null results for gene score and imaging associations
As mentioned in the main text, D1 receptor expression-related gene scores predicted stability of both states
(0-back: b = 0.184, p = 0.034; 2-back: b = 0.242 p = 0.007), but not D2 receptor expression-related gene
scores (0-back: b = 0.153, p = 0.109; 2-back: b = -0.01 p = 0.924). In turn, the control energy of both state
transitions could be predicted by the D2 receptor expression-related score (0- to 2-back: b = -0.076, p =
0.037; and trending for 2- to 0-back: b = -0.134, p = 0.068), but not by the D1 receptor expression-related
gene score (0- to 2-back: b = -0.037, p = 0.324; 2- to 0-back: b = -0.06, p = 0.418).
9.8. Suboptimal trajectories
As mentioned in the main text, the variability in suboptimal trajectories was greater in schizophrenia (rm-
ANOVA: main effect of group: F(1,98) = 4.789, p = 0.031, controlling for age, sex, DTI tSNR ,brain state
energy difference). These results remained significant after additionally accounting for the stability of both
states and for the control energy of both transitions (rm-ANOVA: main effect of group: F(1,95) = 11.2, p =
0.001).
14
10. Supplementary figures
10.1. Figure S1: N-back task design
Design of the N-back task: Stimuli were presented in blocks of either 0-back (left) or 2-back (right)
conditions. There was no additional control or resting condition. In the 0-back condition, subjects were
instructed to press the button on the response box corresponding to the number currently displayed on the
presentation screen. Here, the red numbers next to the screen images on the left indicate correct
responses. In the 2-back condition, subjects were instructed to press the button on the response box
corresponding to the number presented two stimuli before the number currently displayed on the
presentation screen. Here, the red numbers next to the screen images on the right indicate the correct
responses. Each condition block lasted 30 seconds and was repeated four times in an interleaved manner
as shown on the bottom of the figure.
15
10.2. Figure S2: Network control theory concept and methods
A summary of the methods to assess brain dynamics using network control theory. We used a multimodal
atlas, and applied it to both diffusion tensor imaging data to obtain a structural connectome, and to
functional magnetic resonance imaging data to obtain activation patterns during a 0-back motor and
attention control task and during a 2-back working memory task. Finally, we use network control theory to
explain transitions between 0-back and 2-back states based on the underlying structural connectome.
16
11. Supplementary tables
11.1. Table S1: Characteristics of the healthy control and patient samples
Healthy
controls
(n = 178)
33.05 ± 10.98
13.66 ± 2.41
30.74 ± 3.84
n.a
n.a
n.a
n.a.
Demographic
information
Age (year)
Sex (male / female) 93 / 85
Years of education
Psychological
assessments
MWTB
PANSS positive
PANSS negative
BDI
Years of illness
fMRI task
performance
Accuracy (%)
Reaction time (ms)
Imaging quality
parameters
fMRI: mean frame-
wise displacement
(mm)
DTI: mean absolute
root-mean-square
displacement (mm)
DTI: tSNR
1.27 ± 0.74
5.61 ± 0.45
0.15 ± 0.06
68.75 ± 19.33
80.10 ± 18.35
496.15 ± 279.50 589.21 ± 286.80 627.06 ± 306.99
65.54 ± 19.79
Matched
controls
(n = 80)
35.49 ± 10.55
46 / 34
13.65 ± 2.73
30.32 ± 4.83
n.a.
n.a
n.a.
n.a.
Schizophrenia
patients
(n = 24)
32.25 ± 10.33
18 / 6
11.68 ± 1.45
29.13 ±3.27
12.50 ± 6.76
15.17 ± 6.76
12.42 ± 7.71
10.22 ± 9.32
t or χ²
value
P value
1.32
2.39
2.72
1.11
-
-
-
-
0.70
-0.6
0.188
0.122
0.008
0.272
-
-
-
-
0.479
0.553
0.18 ± 0.07
0.20 ± 0.09
-1.11
0.270
1.37 ± 0.89
1.34 ± 0.59
0.18
0.860
5.52 ± 0.49
5.26 ± 0.49
2.21
0.028
Abbreviations: MWTB = Mehrfach Wortschatz Intelligenztest B, a German multiple-choice vocabulary
intelligence test as a measure of premorbid IQ; PANSS = positive and negative symptom scale; BDI =
Beck's depression inventory; DTI = diffusion tensor imaging; tSNR temporal signal-to-noise.
17
11.2. Table S2: Sample characteristics of the pharmacological study
Healthy control (n = 16)
t value P value
Placebo
Amisulpride
Demographic
information
Age (year)
Sex (male /
female)
fMRI task
performance *
26.63 ± 5.34
8 / 8
Accuracy (%)
89.32 ± 10.00
85.68 ± 11.66
1.27
0.223
Reaction time (ms) 347.42 ± 104.26 365.40 ± 142.54 -0.582
0.569
Head motion
parameters
0.124 ± 0.03
0.126 ± 0.04
-0.301
0.767
fMRI: Mean frame-
wise displacement
(mm)
18
11.3. Table S3: Statistical details for the main findings and replication analyses
Glasser
count
Gordon
FA
p-val
df
F or t
p-val
df
F or t
p-val
(stand
beta)
Result
FA
F or t
(stand
beta)
df
General properties of
control
stability 0 back > 2 back
T02 > T20
stability 2-back -> accuracy
stability 2-back -> RT
Differential relation to D1
and D2 expression
D1 -> stability 0 back
D1 -> stability 2 back
D2 -> stability 0 back
D2 -> stability 2 back
D2 -> T02
D2 -> T20
D1 -> T02
D1 -> T20
Drug effect on transition
energy
Pla vs. Ami
Pla vs. Ris
Drug effect on stability
Pla vs. Ami
Pla vs. Ris
Schizophrenia
stability 2-back: HC vs. SZ
stability 0-back: HC vs. SZ
T02: HC vs. SZ
T20: HC vs. SZ
1,173
1,174
1,10
1,10
1,8
1,8
1,98
1,98
1,98
1,98
66.80 < 0.001 1,173
27.98 <0.001 1,174
-2.78
(0.274)
-1.94
(-0.192)
0.006
0.054
2.18
(0.184)
2.78
(0.242)
1.629
(0.153)
-0.095
(-0.010)
-2.14
(-0.076)
-1.87
(-0.134)
-0.996
(-0.037)
-0.817
(-0.06)
7.272
3.49
0.715
1.057
6.436
0.041
5.238
0.620
0.034
0.007
0.109
0.924
0.037
0.068
0.324
0.418
0.022
0.091
0.422
0.334
0.013
0.840
0.024
0.433
1,10
1,10
1,8
1,8
1,98
1,98
1,98
1,98
19
60.50
19.73
2.70
(0.252)
-1.94
(-0.182)
2.212
(0.190)
2.978
(0.270)
1.963
(0.185)
0071
(0.008)
-2.33
(-0.09)
-1.83
(-0.134)
-1.066
(-0.045)
-0.650
(-0.049)
4.954
3.797
0.698
2.941
6.552
0.105
6.414
0.534
<0.001 1,173
<0.001 1,174
0.008
0.054
0.031
0.004
0.055
0.944
0.023
0.073
0.291
0.519
0.428
0.125
0.05 1,10
0.08 1,10
1,8
1,8
0.012 1,98
0.746 1,98
0.013 1,98
0.467 1,98
(stand
beta)
56.846
21.706
2.58
(-0.239)
-1.95
(-0.181)
3.10
(0.270)
3.39
(0.307)
0.146
(0.147)
0.174
(0.019)
-2.07
(-0.07)
-2.16
(-0.152)
-0.832
(-0.029)
-0.376
(-0.028)
8.839
3.014
0.013
0.358
4.951
0.327
5.070
0.275
<0.001
<0.001
0.011
0.053
0.003
0.001
0.149
0.862
0.043
0.036
0.414
0.709
0.014
0.113
0.913
0.566
0.028
0.569
0.027
0.601
Abbreviations: FA = structural edge weight defined as mean fraction anisotropy of a track connecting two
regions; Count = structural edge weight defined as track count connecting two regions; Gordon = resting-
state defined atlas with 333 regions; T02 = control energy for the transition 0 to 2-back; T20 = control
energy for the transition 2 to 0-back; -> = predicts in a regression model; Q = modularity estimate; Pla =
Placebo; Ami = Amisulpride; Ris = Risperidone; HC = healthy control; SZ = schizophrenia patients.
20
11.4. Table S4: List of brain regions included in the extended Glasser parcellation
# Brain Region Label
1 Glasser_L_V1
2 Glasser_L_MST
3 Glasser_L_V6
4 Glasser_L_V2
5 Glasser_L_V3
6 Glasser_L_V4
7 Glasser_L_V8
8 Glasser_L_4
9 Glasser_L_3b
10 Glasser_L_FEF
11 Glasser_L_PEF
12 Glasser_L_55b
13 Glasser_L_V3A
14 Glasser_L_RSC
15 Glasser_L_POS2
16 Glasser_L_V7
17 Glasser_L_IPS1
18 Glasser_L_FFC
19 Glasser_L_V3B
20 Glasser_L_LO1
21 Glasser_L_LO2
22 Glasser_L_PIT
23 Glasser_L_MT
24 Glasser_L_A1
25 Glasser_L_PSL
26 Glasser_L_SFL
27 Glasser_L_PCV
28 Glasser_L_STV
29 Glasser_L_7Pm
30 Glasser_L_7m
31 Glasser_L_POS1
32 Glasser_L_23d
33 Glasser_L_v23ab
34 Glasser_L_d23ab
35 Glasser_L_31pv
36 Glasser_L_5m
37 Glasser_L_5mv
38 Glasser_L_23c
39 Glasser_L_5L
40 Glasser_L_24dd
41 Glasser_L_24dv
42 Glasser_L_7AL
43 Glasser_L_SCEF
44 Glasser_L_6ma
45 Glasser_L_7Am
46 Glasser_L_7PL
47 Glasser_L_7PC
48 Glasser_L_LIPv
49 Glasser_L_VIP
50 Glasser_L_MIP
51 Glasser_L_1
52 Glasser_L_2
53 Glasser_L_3a
54 Glasser_L_6d
55 Glasser_L_6mp
56 Glasser_L_6v
57 Glasser_L_p24pr
58 Glasser_L_33pr
59 Glasser_L_a24pr
60 Glasser_L_p32pr
61 Glasser_L_a24
62 Glasser_L_d32
63 Glasser_L_8BM
64 Glasser_L_p32
65 Glasser_L_10r
66 Glasser_L_47m
67 Glasser_L_8Av
68 Glasser_L_8Ad
69 Glasser_L_9m
70 Glasser_L_8BL
71 Glasser_L_9p
72 Glasser_L_10d
73 Glasser_L_8C
74 Glasser_L_44
75 Glasser_L_45
76 Glasser_L_47l
77 Glasser_L_a47r
78 Glasser_L_6r
79 Glasser_L_IFJa
80 Glasser_L_IFJp
81 Glasser_L_IFSp
82 Glasser_L_IFSa
83 Glasser_L_p9-46v
84 Glasser_L_46
85 Glasser_L_a9-46v
86 Glasser_L_9-46d
87 Glasser_L_9a
88 Glasser_L_10v
89 Glasser_L_a10p
90 Glasser_L_10pp
91 Glasser_L_11l
92 Glasser_L_13l
93 Glasser_L_OFC
94 Glasser_L_47s
95 Glasser_L_LIPd
96 Glasser_L_6a
97 Glasser_L_i6-8
98 Glasser_L_s6-8
99 Glasser_L_43
100 Glasser_L_OP4
101 Glasser_L_OP1
102 Glasser_L_OP2-3
103 Glasser_L_52
104 Glasser_L_RI
105 Glasser_L_PFcm
106 Glasser_L_PoI2
107 Glasser_L_TA2
108 Glasser_L_FOP4
109 Glasser_L_MI
110 Glasser_L_Pir
111 Glasser_L_AVI
112 Glasser_L_AAIC
113 Glasser_L_FOP1
114 Glasser_L_FOP3
115 Glasser_L_FOP2
116 Glasser_L_PFt
117 Glasser_L_AIP
118 Glasser_L_EC
119 Glasser_L_PreS
120 Glasser_L_H
121 Glasser_L_ProS
122 Glasser_L_PeEc
123 Glasser_L_STGa
124 Glasser_L_PBelt
125 Glasser_L_A5
126 Glasser_L_PHA1
127 Glasser_L_PHA3
128 Glasser_L_STSda
129 Glasser_L_STSdp
130 Glasser_L_STSvp
131 Glasser_L_TGd
132 Glasser_L_TE1a
133 Glasser_L_TE1p
134 Glasser_L_TE2a
135 Glasser_L_TF
136 Glasser_L_TE2p
137 Glasser_L_PHT
138 Glasser_L_PH
139 Glasser_L_TPOJ1
140 Glasser_L_TPOJ2
141 Glasser_L_TPOJ3
142 Glasser_L_DVT
143 Glasser_L_PGp
144 Glasser_L_IP2
145 Glasser_L_IP1
146 Glasser_L_IP0
147 Glasser_L_PFop
148 Glasser_L_PF
149 Glasser_L_PFm
21
150 Glasser_L_PGi
151 Glasser_L_PGs
152 Glasser_L_V6A
153 Glasser_L_VMV1
154 Glasser_L_VMV3
155 Glasser_L_PHA2
156 Glasser_L_V4t
157 Glasser_L_FST
158 Glasser_L_V3CD
159 Glasser_L_LO3
160 Glasser_L_VMV2
161 Glasser_L_31pd
162 Glasser_L_31a
163 Glasser_L_VVC
164 Glasser_L_25
165 Glasser_L_s32
166 Glasser_L_pOFC
167 Glasser_L_PoI1
168 Glasser_L_Ig
169 Glasser_L_FOP5
170 Glasser_L_p10p
171 Glasser_L_p47r
172 Glasser_L_TGv
173 Glasser_L_MBelt
174 Glasser_L_LBelt
175 Glasser_L_A4
176 Glasser_L_STSva
177 Glasser_L_TE1m
178 Glasser_L_PI
179 Glasser_L_a32pr
180 Glasser_L_p24
181 Glasser_R_V1
182 Glasser_R_MST
183 Glasser_R_V6
184 Glasser_R_V2
185 Glasser_R_V3
186 Glasser_R_V4
187 Glasser_R_V8
188 Glasser_R_4
189 Glasser_R_3b
190 Glasser_R_FEF
191 Glasser_R_PEF
192 Glasser_R_55b
193 Glasser_R_V3A
194 Glasser_R_RSC
195 Glasser_R_POS2
196 Glasser_R_V7
197 Glasser_R_IPS1
198 Glasser_R_FFC
199 Glasser_R_V3B
200 Glasser_R_LO1
201 Glasser_R_LO2
202 Glasser_R_PIT
203 Glasser_R_MT
204 Glasser_R_A1
205 Glasser_R_PSL
206 Glasser_R_SFL
207 Glasser_R_PCV
208 Glasser_R_STV
209 Glasser_R_7Pm
210 Glasser_R_7m
211 Glasser_R_POS1
212 Glasser_R_23d
213 Glasser_R_v23ab
214 Glasser_R_d23ab
215 Glasser_R_31pv
216 Glasser_R_5m
217 Glasser_R_5mv
218 Glasser_R_23c
219 Glasser_R_5L
220 Glasser_R_24dd
221 Glasser_R_24dv
222 Glasser_R_7AL
223 Glasser_R_SCEF
224 Glasser_R_6ma
225 Glasser_R_7Am
226 Glasser_R_7PL
227 Glasser_R_7PC
228 Glasser_R_LIPv
229 Glasser_R_VIP
230 Glasser_R_MIP
231 Glasser_R_1
232 Glasser_R_2
233 Glasser_R_3a
234 Glasser_R_6d
235 Glasser_R_6mp
236 Glasser_R_6v
237 Glasser_R_p24pr
238 Glasser_R_33pr
239 Glasser_R_a24pr
240 Glasser_R_p32pr
241 Glasser_R_a24
242 Glasser_R_d32
243 Glasser_R_8BM
244 Glasser_R_p32
245 Glasser_R_10r
246 Glasser_R_47m
247 Glasser_R_8Av
248 Glasser_R_8Ad
249 Glasser_R_9m
22
250 Glasser_R_8BL
251 Glasser_R_9p
252 Glasser_R_10d
253 Glasser_R_8C
254 Glasser_R_44
255 Glasser_R_45
256 Glasser_R_47l
257 Glasser_R_a47r
258 Glasser_R_6r
259 Glasser_R_IFJa
260 Glasser_R_IFJp
261 Glasser_R_IFSp
262 Glasser_R_IFSa
263 Glasser_R_p9-46v
264 Glasser_R_46
265 Glasser_R_a9-46v
266 Glasser_R_9-46d
267 Glasser_R_9a
268 Glasser_R_10v
269 Glasser_R_a10p
270 Glasser_R_10pp
271 Glasser_R_11l
272 Glasser_R_13l
273 Glasser_R_OFC
274 Glasser_R_47s
275 Glasser_R_LIPd
276 Glasser_R_6a
277 Glasser_R_i6-8
278 Glasser_R_s6-8
279 Glasser_R_43
280 Glasser_R_OP4
281 Glasser_R_OP1
282 Glasser_R_OP2-3
283 Glasser_R_52
284 Glasser_R_RI
285 Glasser_R_PFcm
286 Glasser_R_PoI2
287 Glasser_R_TA2
288 Glasser_R_FOP4
289 Glasser_R_MI
290 Glasser_R_Pir
291 Glasser_R_AVI
292 Glasser_R_AAIC
293 Glasser_R_FOP1
294 Glasser_R_FOP3
295 Glasser_R_FOP2
296 Glasser_R_PFt
297 Glasser_R_AIP
298 Glasser_R_EC
299 Glasser_R_PreS
300 Glasser_R_H
301 Glasser_R_ProS
302 Glasser_R_PeEc
303 Glasser_R_STGa
304 Glasser_R_PBelt
305 Glasser_R_A5
306 Glasser_R_PHA1
307 Glasser_R_PHA3
308 Glasser_R_STSda
309 Glasser_R_STSdp
310 Glasser_R_STSvp
311 Glasser_R_TGd
312 Glasser_R_TE1a
313 Glasser_R_TE1p
314 Glasser_R_TE2a
315 Glasser_R_TF
316 Glasser_R_TE2p
317 Glasser_R_PHT
318 Glasser_R_PH
319 Glasser_R_TPOJ1
320 Glasser_R_TPOJ2
321 Glasser_R_TPOJ3
322 Glasser_R_DVT
323 Glasser_R_PGp
324 Glasser_R_IP2
325 Glasser_R_IP1
326 Glasser_R_IP0
327 Glasser_R_PFop
328 Glasser_R_PF
329 Glasser_R_PFm
330 Glasser_R_PGi
331 Glasser_R_PGs
332 Glasser_R_V6A
333 Glasser_R_VMV1
334 Glasser_R_VMV3
335 Glasser_R_PHA2
336 Glasser_R_V4t
337 Glasser_R_FST
338 Glasser_R_V3CD
339 Glasser_R_LO3
340 Glasser_R_VMV2
341 Glasser_R_31pd
342 Glasser_R_31a
343 Glasser_R_VVC
344 Glasser_R_25
345 Glasser_R_s32
346 Glasser_R_pOFC
347 Glasser_R_PoI1
348 Glasser_R_Ig
349 Glasser_R_FOP5
350 Glasser_R_p10p
351 Glasser_R_p47r
352 Glasser_R_TGv
353 Glasser_R_MBelt
354 Glasser_R_LBelt
355 Glasser_R_A4
356 Glasser_R_STSva
357 Glasser_R_TE1m
358 Glasser_R_PI
359 Glasser_R_a32pr
360 Glasser_R_p24
504 HO_Left_Thalamus
505 HO_Left_Caudate
506 HO_Left_Putamen
507 HO_Left_Pallidum
509 HO_Left_Hippocampus
510 HO_Left_Amygdala
511 HO_Left_Accumbens
515 HO_Right_Thalamus
516 HO_Right_Caudate
517 HO_Right_Putamen
518 HO_Right_Pallidum
519 HO_Right_Hippocampus
520 HO_Right_Amygdala
521 HO_Right_Accumbens
23
12. Supplementary references
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
Tombaugh TN (2004) Trail Making Test A and B: normative data stratified by age and education.
Archives of clinical neuropsychology 19(2):203-214.
Lehrl S, Triebig G, & Fischer B (1995) Multiple choice vocabulary test MWT as a valid and short test to
estimate premorbid intelligence. Acta Neurologica Scandinavica 91(5):335-345.
Kay SR, Fiszbein A, & Opler LA (1987) The positive and negative syndrome scale (PANSS) for
schizophrenia. Schizophrenia bulletin 13(2):261-276.
Beck AT, Ward CH, Mendelson M, Mock J, & Erbaugh J (1961) An inventory for measuring depression.
Arch Gen Psychiatry 4:561-571.
Braun U, et al. (2015) Dynamic reconfiguration of frontal brain networks during executive cognition in
humans. Proc Natl Acad Sci U S A 112(37):11678-11683.
Plichta MM, et al. (2012) Test-retest reliability of evoked BOLD signals from a cognitive-emotive fMRI
test battery. Neuroimage 60(3):1746-1758.
Callicott JH, et al. (1999) Physiological characteristics of capacity constraints in working memory as
revealed by functional MRI. Cerebral Cortex 9(1):20-26.
Glasser MF, et al. (2016) A multi-modal parcellation of human cerebral cortex. Nature 536(7615):171-
178.
Smith SM, et al. (2004) Advances in functional and structural MR image analysis and implementation
as FSL. Neuroimage 23 Suppl 1:S208-219.
Smith SM (2002) Fast robust automated brain extraction. Hum Brain Mapp 17(3):143-155.
Yeh FC, Verstynen TD, Wang Y, Fernandez-Miranda JC, & Tseng WY (2013) Deterministic diffusion
fiber tracking improved by quantitative anisotropy. PLoS One 8(11):e80713.
Baum GL, et al. (2018) The impact of in-scanner head motion on structural connectivity derived from
diffusion MRI. Neuroimage 173:275-286.
Chung MH, et al. (2018) Individual differences in rate of acquiring stable neural representations of tasks
in fMRI. PLoS One 13(11):e0207352.
Braun U, et al. (2016) Dynamic brain network reconfiguration as a potential schizophrenia genetic risk
mechanism modulated by NMDA receptor function. Proc Natl Acad Sci U S A 113(44):12568-12573.
Betzel RF, Gu S, Medaglia JD, Pasqualetti F, & Bassett DS (2016) Optimally controlling the human
connectome: the role of network topology. Sci Rep 6:30770.
Gu S, et al. (2017) Optimal trajectories of brain state transitions. Neuroimage 148:305-317.
Kim JZ, et al. (2018) Role of Graph Architecture in Controlling Dynamical Networks with Applications to
Neural Systems. Nat Phys 14:91-98.
Cornblath EJ, et al. (2018) Context-dependent architecture of brain state dynamics is explained by
white matter connectivity and theories of network control. arXiv preprint arXiv:1809.02849.
Howie B, Fuchsberger C, Stephens M, Marchini J, & Abecasis GR (2012) Fast and accurate genotype
imputation in genome-wide association studies through pre-phasing. Nat Genet 44(8):955-959.
Howie BN, Donnelly P, & Marchini J (2009) A flexible and accurate genotype imputation method for the
next generation of genome-wide association studies. PLoS genetics 5(6):e1000529.
Delaneau O, Zagury JF, & Marchini J (2013) Improved whole-chromosome phasing for disease and
population genetic studies. Nature methods 10(1):5-6.
Fazio L, et al. (2018) Transcriptomic context of DRD1 is associated with prefrontal activity and behavior
during working memory. Proc Natl Acad Sci U S A 115(21):5582-5587.
Pergola G, et al. (2017) DRD2 co-expression network and a related polygenic index predict imaging,
behavioral and clinical phenotypes linked to schizophrenia. Transl Psychiatry 7(1):e1006.
Selvaggi P, et al. (2019) Genetic Variation of a DRD2 Co-expression Network is Associated with
Changes in Prefrontal Function After D2 Receptors Stimulation. Cereb Cortex 29(3):1162-1173.
Fazio L, et al. (2018) Transcriptomic context of DRD1 is associated with prefrontal activity and behavior
during working memory. Proceedings of the National Academy of Sciences of the United States of
America.
Selvaggi P, et al. (2018) Genetic Variation of a DRD2 Co-expression Network is Associated with
Changes in Prefrontal Function After D2 Receptors Stimulation. Cerebral cortex.
24
27.
28.
29.
30.
31.
Pergola G, et al. (2018) Prefrontal co-expression of schizophrenia risk genes is associated with
treatment response in patients. bioRxiv.
Zhang B & Horvath S (2005) A general framework for weighted gene co-expression network analysis.
Statistical applications in genetics and molecular biology 4:Article17.
Colantuoni C, et al. (2011) Temporal dynamics and genetic control of transcription in the human
prefrontal cortex. Nature 478(7370):519-523.
Pergola G, et al. (2016) Combined effect of genetic variants in the GluN2B coding gene (GRIN2B) on
prefrontal function during working memory performance. Psychological medicine 46(6):1135-1150.
Gordon EM, et al. (2016) Generation and Evaluation of a Cortical Area Parcellation from Resting-State
Correlations. Cereb Cortex 26(1):288-303.
25
|
1807.02126 | 1 | 1807 | 2018-07-05T18:03:19 | Cognitive chimera states in human brain networks | [
"q-bio.NC",
"math.DS",
"physics.data-an"
] | The human brain is a complex dynamical system that gives rise to cognition through spatiotemporal patterns of coherent and incoherent activity between brain regions. As different regions dynamically interact to perform cognitive tasks, variable patterns of partial synchrony can be observed, forming chimera states. We propose that the emergence of such states plays a fundamental role in the cognitive organization of the brain, and present a novel cognitively-informed, chimera-based framework to explore how large-scale brain architecture affects brain dynamics and function. Using personalized brain network models, we systematically study how regional brain stimulation produces different patterns of synchronization across predefined cognitive systems. We then analyze these emergent patterns within our novel framework to understand the impact of subject-specific and region-specific structural variability on brain dynamics. Our results suggest a classification of cognitive systems into four groups with differing levels of subject and regional variability that reflect their different functional roles. | q-bio.NC | q-bio | Cognitive chimera states in human brain networks
July 9, 2018
Kanika Bansal1,2,3,*, Javier O. Garcia1,4, Steven H. Tompson1,4, Timothy Verstynen5, Jean
M. Vettel1,4,6, Sarah F. Muldoon3,7,*
1 Human Sciences, US Army Research Laboratory, Aberdeen Proving Ground, MD 21005, USA
2 Department of Biomedical Engineering, Columbia University, New York, NY 10027, USA
3 Mathematics Department, University at Buffalo – SUNY, Buffalo, NY 14260, USA
4 Department of Biomedical Engineering, University of Pennsylvania, Philadelphia, PA 19104, USA
5 Department of Psychology, Carnegie Mellon University, Pittsburgh, PA 15213, USA
6 Department of Psychological and Brain Sciences, University of California, Santa Barbara, CA 93106 , USA
7 CDSE Program and Neuroscience Program, University at Buffalo – SUNY, Buffalo, NY 14260, USA
*Correspondence should be addressed to KB ([email protected]) and SFM ([email protected]).
8
1
0
2
l
u
J
5
]
.
C
N
o
i
b
-
q
[
1
v
6
2
1
2
0
.
7
0
8
1
:
v
i
X
r
a
1
Abstract
The human brain is a complex dynamical system that gives rise to cognition through spatiotemporal
patterns of coherent and incoherent activity between brain regions. As different regions dynamically
interact to perform cognitive tasks, variable patterns of partial synchrony can be observed, forming
chimera states. We propose that the emergence of such states plays a fundamental role in the cognitive
organization of the brain, and present a novel cognitively-informed, chimera-based framework to explore
how large-scale brain architecture affects brain dynamics and function. Using personalized brain network
models, we systematically study how regional brain stimulation produces different patterns of synchro-
nization across predefined cognitive systems. We then analyze these emergent patterns within our novel
framework to understand the impact of subject-specific and region-specific structural variability on brain
dynamics. Our results suggest a classification of cognitive systems into four groups with differing levels
of subject and regional variability that reflect their different functional roles.
2
Introduction
Rhythmic behavior is ubiquitous in complex systems, and a diverse set of research has examined how in-
teracting system elements come together to form synchronized, coherent behavior across domains that span
biological [1, 2], social [3], and engineered [4] settings. However, the emergence of complete system-wide
synchronization might not always provide the best description of system dynamics. In many systems, states
of partial synchrony have been observed, where a system organizes in separate domains of synchronized
elements [5, 6, 7]. This is particularly true in the human brain, where patterns of neurophysiological activity
evolve rapidly, showing transient domains of synchronization across subsets of brain regions [8, 9, 5]. In the
last decade, the rise of network neuroscience approaches [10, 11] have demonstrated a foundational role for
partial synchrony among separate cognitive sub-networks, where the underlying architecture of the brain
ensures efficient integration of sensory input with stored knowledge while also segregating task-irrelevant in-
formation to support cognition [12, 13]. However, the fundamental principles and constraints that subserve
the intricate timing and specificity of these time-evolving patterns of synchrony are not well understood [14].
The dynamical systems framework of chimera states offers a powerful tool to study the evolution of
coherent and incoherent dynamics in oscillating systems such as the brain. A chimera state emerges when
a system of oscillators evolves into two subsets of mutually coherent and incoherent populations [15, 16].
Although chimera states represent a natural link between coherent and incoherent dynamics [6, 17, 18],
they were initially only explored analytically and their relationship to physical systems was unknown. In
fact, it wasn't until almost a decade after their theoretical discovery that chimera states were demonstrated
experimentally in opto-electronic [19], and chemical [20] oscillators, finally establishing their connection with
real-world systems. Recently, the chimera framework has been extended to include states of multi-chimera
[21, 22], traveling chimera [23], amplitude chimera, and chimera death [24], demonstrating the versatility of
the framework to describe critical system dynamics.
Because of its natural ability to describe patterns of partial synchronization, the chimera framework
has an intuitive utility for augmenting our understanding of the brain. Patterns of synchronization between
cognitive systems are thought to form the basis of cognition, and the interplay of synchrony among subsets of
brain regions has been shown to underlie sleep-wake states [25], variability in task performance [12], and the
continuum between healthy and disease states [26, 27]. Indeed, recent work has speculated that similarities
exist between chimera states and brain dynamics during unihemispheric sleep [28], the transition to a seizure
state in epilepsy [29], and coordinated finger tapping exercises [30]. Fundamentally, these dynamics are the
result of complex interactions between neuronal populations and are often modeled using networks of coupled
oscillators [31]. As a result, despite the intuitive similarities between chimera states and brain dynamics,
much of the work relating chimera dynamics to neuroscience thus far has focused on understanding chimera
states at the level of neuronal networks, using mathematical modeling of networks of individual neurons with
fewer elements and/or simplified connection topologies [21, 22, 32, 33]. Only recently have neuronal models
been used to examine the possibility of the emergence of chimera-like states within large-scale brain networks
derived from two well-characterized animal brains – C. elegans [34] and the cat cortex [35]. However, even in
these instances, the network connectivities were modified for simplicity. Thus, there remains a gap between
studies of chimera states and applications to large-scale functional patterns of brain activity thought to
underlie cognition. This largely reflects the computational complexity of modeling whole-brain dynamics
and identifying an informative, yet simplified, model of cognitive processing.
Here, we bridge this gap by presenting a novel cognitively-informed, chimera-based framework combined
with in silico experiments, where we leverage the existence of a core set of predefined cognitive systems
that constitute the functional organization of the brain [36, 37]. Consequently, our novel framework keeps
the computational complexity of the analysis to a minimum while providing the unique ability to connect
chimera states as an underlying basis for cognition. Using personalized brain network models (BNMs), we
study cognitive system-level patterns of synchrony that emerge across 76 brain regions within 9 cognitive
3
systems as the result of regional brain stimulation. Our analysis focuses on how brain architecture relates
to the frequency and types of dynamical patterns produced after stimulation. More specifically, we aim
to answer two questions: (i) do patterns of synchronization observed for each cognitive system depend on
what region was stimulated (region-specific effects), and (ii) does structural variability between participants
decrease the consistency of patterns observed for each cognitive system (subject-specific effects)?
From our in silico experiments, we observe different patterns of synchronization that can be classified
into three dynamical states: (i) a coherent state of global synchrony, (ii) a chimera state with coexisting
domains of synchrony and de-synchrony, and (iii) a metastable state with an absence of any large-scale stable
synchrony. Our results demonstrate rich diversity in the states produced across all nine cognitive systems,
including variability in patterns based on both region-specific and subject-specific structural variability.
Critically, all nine cognitive systems give rise to chimera states and this likely reflects the foundational role
that partial synchrony serves in large-scale brain function. Neuronal dynamics must concurrently segregate
specialized processing while integrating localized functions for coordinated, cohesive cognitive performance.
Our novel chimera-based framework provides an avenue to study how dynamical states give rise to variability
in cognitive performance, providing the first approach that can uncover the link between chimera states and
cognitive system functions that subserve human behavior.
Results
We first build subject-specific BNMs using anatomical connectivity derived from diffusion spectrum imaging
data of 30 healthy individuals [38]. As illustrated in Fig. 1a, we parcellate the brain into 76 regions (network
nodes) and define weighted network edges based on structural connectivity between brain regions. The
dynamics of each brain region are modeled using Wilson-Cowan oscillators (WCOs), a biologically inspired,
nonlinear mean-field model of a small population of neurons [39]. The coupling between regions is derived
based on the unique structural connectivity of each individual (Fig. S1). When multiple WCOs are coupled,
the resulting patterns of synchronization are highly dependent upon the topology of the coupling, ensuring
that these BNMs are highly sensitive to individual variability in the underlying anatomical connectivity
[36, 40, 41]. As illustrated in Fig. 1b and Fig. S2, we then perform in silico experiments by systematically
applying computational regional (nodal) stimulation to the BNM and assessing the resulting patterns of
synchronization that emerge based on subject-specific and region-specific variation in structural connectivity.
Classification of brain states using cognitive systems.
Traditionally, one would measure system-wide synchronization by calculating a measure such as the Ku-
ramoto order parameter within the entire network of oscillators [42]. However, in order to study how the
stimulation of different brain regions drives brain function, we instead focus on the relationship between
patterns of synchronization between large-scale cognitive systems. Thus, based on previous research [36, 37],
we assigned each of the 76 brain regions into one of nine cognitive systems (Fig. S3). Each cognitive system
is defined by regions that coactivate in support of a generalized class of cognitive functions. This delin-
eation includes several sensory-motor related systems, including auditory (Aud), visual (V), and motor and
somatosensory (MS) systems, as well as the ventral temporal association system (VT) that encapsulates
regions involved in knowledge representation. Several of the systems are involved in functional roles that
are generic across cognitive performance, including the attention system (Att), the medial default mode sys-
tem (mDM), and two systems associated with cognitive control, cingulo-opercular (CP) and frontoparietal
(FP) systems. Finally, we include the subcortical system (SC) that consists of the regions responsible for
autonomic and primal functions.
In each in silico experiment, we stimulate a brain region in each subject-specific BNM and then compute
cognitive-system-level synchronization. We calculate a cognitive-system-based Kuramoto order parameter
4
Figure 1: Design of the in silico experiments.
(a) We construct personalized brain network models by
estimating white matter anatomical connectivity of the brain using diffusion spectrum imaging. This con-
nectivity is combined with a brain parcellation scheme with 76 cortical and subcortical regions to obtain large
scale connectivity map of the regional brain volume. These regions constitute the nodes of the structural
brain network, whose dynamics are simulated by nonlinear Wilson-Cowan oscillators, coupled through the
structural connectivity map of a given individual (see Methods). (b) In the resulting data-driven models of
the spatiotemporal dynamics of the brain, each brain region is systematically stimulated across a cohort of
30 subjects. The spread of the stimulation is measured through synchronization within the brain network.
5
ρci,cj that measures the amount of synchrony among all oscillators (regions) within two cognitive systems ci
and cj, and obtain a cognitive-system-based synchronization matrix as shown in Fig. 2a. In this synchro-
nization matrix we define two cognitive systems, ci and cj, to be synchronized if ρci,cj exceeds a threshold
value (ρT h). In Fig. 2a we chose ρT h = 0.8 to define three dynamical states observed in this study: (i)
a coherent state, where all cognitive systems are synchronized; (ii) a cognitive chimera state, where some
cognitive systems form a synchronized cluster (yellow) while the other systems remain incoherent (blue);
and (iii) a metastable state, where no stable synchrony between cognitive systems is observed.
Next, we compare our cognitive-system-based analysis with two traditional measures of synchronization:
(1) the classical Kuramoto order parameter [42] calculated across all 76 oscillators (regions) that captures
the level of global synchrony in the network, and (2) the chimera-index [43, 34] that describes the closeness of
the state with an ideal chimera state (see Methods). In Fig. 2b, we show how the three dynamical states (co-
herent, chimera, and metastable) observed after stimulation of different brain regions in different individuals
are distributed in the global synchrony and chimera-index parameter space. We observe a clear separation
of these states, and as expected, global synchrony decreases from the coherent to chimera to metastable
state. Thus, a cognitively-informed, systems-based classification of dynamical states is comparable to the
traditional measures of estimating synchrony within a network. It is also robust across a range of threshold
values (ρT h) as shown in Figures S4 and S5.
We also examine how these two traditional metrics of synchronization relate to the connectivity properties
of the node (region) itself. As seen in Fig. 2c, the level of global synchrony is positively correlated with
the degree of the region being stimulated (r = 0.81; p = 1.8 × 10−124). Stimulation of a network hub
(highly connected brain region) is therefore more likely to produce a coherent state, while stimulation of
a non-hub is more likely to result in either a chimera or metastable state. Interestingly, Fig. 2d reveals
that the chimera-index shows a weaker and negative correlation with degree (r = −0.57; p = 7.4 × 10−195).
This relationship indicates the ability of moderately connected brain regions to produce a variety of spatially
distinct synchronization patterns as a result of stimulation. These results not only demonstrate that chimera
states emerge among large-scale cognitive systems but also reveal that the variable structural connectivity
of regions within cognitive systems can drive the whole brain into diverse synchronization patterns.
Variable brain states emerge from stimulation of different brain regions.
Given that the type of brain state that emerges as result of stimulation is related to the local connectivity of
the region, we next asked if there was also a spatial relationship between the location of the stimulated region
and the type of dynamical state produced. In Fig. 3, brain regions are depicted as an orb, and their sizes
denote the normalized occurrence of a given state, that they produce upon stimulation, calculated across
subjects. Regions that produce coherent states (network hubs) are distributed more closely to the center of
the brain (3a), while regions that produce the opposite extreme, metastable states, are distributed along the
periphery of the brain (3c). Interestingly, regions that produce chimera states are located between the two
(3b). This asymmetry in the distribution of states highlights the differences in the structural organization of
the brain. Regions in the center of the brain that produce a coherent state (e.g., subcortical regions such as
the hippocampus, thalamus) can play a global cognitive role and facilitate communication between spatially
separated brain regions. Conversely, regions located in the periphery produce metastable states consistent
with the notion that local and/or specialized computations take place in the cortex and could reflect the fact
that the structure of the human brain has evolved to produce complex cognitive abilities [44].
Due to the diversity in the distribution of dynamical states across spatially distributed regions, we inves-
tigate the relative contribution of the nine cognitive systems in producing each dynamical state after regional
stimulation. As shown in Fig. 4a, coherent states are produced predominantly by regional stimulation within
subcortical and medial default mode systems, reflecting that many of their constituent regions are network
hubs.
6
Figure 2: Emergence of dynamical states within a cognitively-informed framework. (a) Cognitive system-
level synchronization matrices whose entries denote the extent of synchronization between cognitive systems
for coherent, chimera and metastable states. (b) The global synchronization and chimera-index of the system
after stimulation of each brain region across all subjects. Within this traditional framework, our cognitively-
defined states can be identified to have distinguishable characteristics. A coherent state shows high global
synchronization value and low chimera-index (red). Both chimera (yellow) and metastable (blue) states
show lower global synchronization. A chimera state can have either a higher global synchronization or a
higher chimera-index than metastable states. (c)-(d) Origin of these states follows the connectivity of the
stimulated brain region. (c) The global synchronization measure is positively correlated with the weighted
degree of brain regions (r = 0.81, p = 1.8 × 10−124), indicating that the network hubs are more likely to
produce a coherent state. (d) The Chimera-index is weakly and negatively correlated with the weighted
degree (r = −0.57, p = 7.4 × 10−195), indicating that stimulation of lower degree nodes is more likely to
produce an ideal chimera state with half of the population synchronized and the other half de-synchronized.
7
Figure 3: Distributed origin of dynamical states. Separate depictions of brain network regions (nodes) that
produce each of the three dynamical states: (a) coherent, (b) chimera, and (c) metastable. Both axial and
coronal views of the brain are presented to visualize the anatomical location of the node in left (LH) and
right (RH) hemispheres. Node size represents the normalized occurrence of a given state upon stimulation
of the node across all the subjects. Variability in node sizes imply that different states can emerge upon the
stimulation of a single node across individuals. A coherent state is significantly more likely to be originated
from nodes within the center of the brain. (b) Chimera states are likely to be originated by the stimulation of
nodes that are relatively equally distributed within the brain; however, these nodes may vary in the spatial
patterns of synchronization that they produce. (c) A metastable state is more likely to originate from the
stimulation of peripheral brain regions.
8
Figure 4: Contribution of cognitive systems to dynamical states. (a) Coherent states are likely to result when
nodes within medial default mode and subcortical systems are stimulated. (b) Chimera states emerge upon
the stimulation of nodes that are equally distributed across all the cognitive systems. (c) Metastable states
frequently occur after stimulation of nodes within auditory, cingulo-opercular, frontoparietal, and ventral
temporal association systems. This distribution indicates the dominance of a particular type of cognitive
role within the nodes of different cognitive systems.
9
There is also system specificity for metastable states which are preferentially produced by four systems
(Fig. 4c). Two of those systems: cingulo-opercular and forntoparietal, are associated with cognitive control
and proposed to be complementary systems that may often need to process task-relevant information concur-
rently. Their dominance in producing metastable states likely reflects the fact that these systems can work
in seclusion without co-activating a large part of the brain (facilitating parallel processing) and that they
are flexible and not constrained by their structural connectivity. Both the auditory and the ventral temporal
association systems also contribute significantly to metastable states. These two systems are both predomi-
nantly located in the temporal lobe, an area of the brain associated with knowledge representation, so their
functional roles may also frequently require working in seclusion in support of higher order perception.
Although coherent and metastable states are dominantly produced by specific cognitive systems, all nine
systems give rise to chimera states (Fig. 4b). In fact, chimera states are more likely to occur than either
coherent or metastable states (Figures 3b and 2b) and encompass a variety of different spatial patterns of
coexisting coherent and incoherent behavior. This likely reflects the foundational role of partial synchrony
in large-scale brain function. Cognitive tasks constantly require the intricate balance between segregated
and integrated neural processing [12]. The robust occurrence of chimera states following stimulation to each
of the nine cognitive systems reflects the complexity and flexibility of the brains underlying architecture to
support diverse processing requirements.
Structural variability influences observed dynamic states.
We next characterize the spatial patterns that comprise the chimera states to understand which cognitive
systems are synchronized and de-synchronized following stimulation to each region. The results are organized
by stimulation of brain regions within a cognitive system in Fig. 5a-i, where each row is a possible pattern of
synchronization within the nine cognitive systems. The rows are organized by the frequency with which the
pattern was observed (listed to the right of the row) after stimulation to all regions within that system and
across all individuals in the study. For each pattern, systems that are part of the synchronized population
after stimulation are shown in orange, and systems that are part of the de-synchronized population are shown
in white. Consequently, coherent states are demarcated by a fully orange row, metastable states by a fully
white row, and chimera states by a mixed pattern of coloring. For each cognitive system (Fig. 5a-i), we
present the prevalent patterns observed and these results illustrate what systems are likely to synchronize
after stimulation.
Aligned with their complementary roles in cognitive control, the cingulo-opercular and frontoparietal
systems continue to show similarity in their patterns of synchronization. The dominant pattern after stimu-
lation to regions in either system is a metastable state, occurring 45% for cingulo-opercular system (Fig. 5c)
and 50% for the frontoparietal system (Fig. 5d). Similarly, the auditory system also produces a metastable
state 50% of the time (Fig. 5b). For all three of these systems, the second most frequent state is the opposite
extreme, a coherent state (20% for auditory and 11% for cingulo-opercular) or nearly coherent state (7%
for frontoparietal). Thus, these three systems show diversity in the types of dynamical states that they are
capable of producing: sometimes stimulation of the system produces a metastable (segregated) state, while
other times, stimulation of the system drives the brain to a coherent (integrated) state.
The ventral temporal association system also produces a metastable state as its most prevalent pattern
(Fig. 5h), but unlike in the three previously discussed systems, this state is produced less frequently (16%),
and the system also produces a much larger variety of prevalent patterns of synchronization (10 unique
patterns). The only other system to show this high level of diversity in its produced patterns is the motor
and somatosensory system (10 patterns, Fig. 5f). In both systems, multiple patterns of chimera states are
observed. This likely reflects the ubiquitous need for neural processing related to both action coordination
(motor and somatosensory) and higher order perception (ventral temporal association) to be integrated with
the processing occurring in other systems within the brain [45, 46].
10
Figure 5: Patterns of synchronization and cognitive chimera states.
(a)-(i) Prevalent patterns (with an
occurrence of at least 3%) that emerge as the regions within different cognitive systems are stimulated
across all subjects. Each panel represents stimulation of regions within a particular cognitive system. Each
row represents one pattern of synchronization, and each column represents the state of a cognitive system.
Cognitive systems that belong to the synchronized population are colored orange and cognitive systems
that remain desynchronized are colored white. Thus, a fully orange or white row represents a coherent or
metastable state respectively. Chimera states show different patterns of coloring depending on the cognitive
systems that are recruited to the synchronized group. Different rows of patterns are stacked based on
their relative occurrences (mentioned on the right side). To summarize the observed patterns, (j) shows
the probability with which different cognitive systems can be synchronized when the regions within a given
system are stimulated across subjects (shown along the vertical axis).
11
A coherent state occurs most frequently for the attention (36%, Fig. 5a), default mode (33%, Fig. 5e),
and subcortical systems (43%, Fig. 5g). The visual system is also similar, though the coherent state is the
second most prevalent (19%) with a nearly coherent state (all but auditory) as its dominant pattern (22%).
Overall, these four systems are less dynamically diverse since their stimulation largely results in chimera
states with high synchrony. All of these systems serve fundamental functional roles to rapidly respond to the
external environment, and their dynamical patterns reflect this need to efficiently integrate this information
with other cognitive systems.
For all systems, the most common state following regional stimulation is a chimera state with high
synchrony, emphasizing the importance of partial synchrony for all of the diverse functional roles provided
by large-scale cognitive systems. Visual inspection of the patterns suggests that the systems most likely to
belong to the desynchronized population of a chimera state are the three systems that predominantly produce
a metastable state (auditory, cingulo-opercular, and frontoparietal systems). This effect is quantified in the
synchronization probability plot (Fig. 5j). Systems with a high probability of synchronization have dark
colors, while systems that are unlikely to be synchronized as a result of stimulation to a specific cognitive
system are shown in light colors. While the auditory, cingulo-opercular, and frontoparietal systems systems
are all unlikely to synchronize with other systems, the attention system, in contrast, is highly likely to be part
of the synchronized population following stimulation to any system. We also observe that the stimulation of
a region within a particular system does not necessarily induce synchronization within that system, which
is particularly the case with the auditory, cingulo-opercular, and frontoparietal systems.
Collectively, these results reveal the power of this approach to characterize large-scale system effects
after regional stimulation. In Fig. S6, we further describe how different nodes in a system contribute to
observed patterns by spatially mapping the probability of a given pattern onto the brain. Taken together,
these results highlight that stimulation within each system gives rise to multiple patterns, similar patterns
can emerge from spatially different regions, and within a system, there can be a special distribution of states
across brain regions. This likely arises from individual differences in the structural connectivity between the
participants in the study or differences in the structural connectivity of the regions themselves within each
system (or combination of the two). Consequently, we introduce a new metric to assess the contribution of
subject-specific and region-specific variability on the observed patterns.
Dissociation of subject-specific and region-specific variability
In our final analysis, we compute a measure called tenacity, which we defined to quantify the level of similarity
between a set of observed patterns (see Methods), to assess how structural variability, either between subjects
or between regions, influenced the patterns observed in Fig. 5. To differentiate these two potential sources
of variability, we separately compute a subject tenacity and a region tenacity score (see Methods). When
tenacity is calculated across subjects (subject tenacity), it measures the similarity of different patterns
produced across individuals by stimulating the same brain region. A cognitive system's subject tenacity
is then the average subject tenacity across regions within the system. When tenacity is calculated across
brain regions (region tenacity), it measures the similarity of different patterns produced by stimulation of
different brain regions within a given cognitive system in a single individual. This value is then averaged over
all individuals. Consequently, a high value of tenacity indicates a high similarity between synchronization
patterns produced by stimulation across individuals (subject tenacity) or brain regions (region tenacity).
In Fig. 6, we plot each cognitive system based on its score for subject and region tenacity. Additionally,
we group cognitive systems by applying a clustering algorithm (see Methods), and the color and shape of a
region's icon reflects its group assignment. We identify four distinct groups of systems that are characteris-
tically different from each other in terms of their location in the tenacity space and also in their cognitive
roles. To better delineate the four groups, we partition the tenacity space based on the level of subject
and node tenacity. We define two levels of subject tenacity: variable and stable, while the node tenacity is
12
Figure 6: Classification of cognitive systems based on pattern tenacity. To estimate the consistency of
emergent patterns of synchronization within cognitive systems, we constructed a measure called tenacity that
estimates the similarity between a set of patterns. Within a cognitive system, we calculate tenacity across
two dimensions: patterns that are produced after stimulating each region across subjects (subject tenacity),
and patterns that are produced after stimulating different regions of the system within each subject (region
tenacity). In the parameter space constructed along these dimensions, we can cluster cognitive systems into
4 groups that suggest a 2x3 partitioning of the tenacity space. This partitioning allows us to dissociate
subject-specific and region-specific variability in observed patterns.
13
partitioned into three levels: diverse, flexible, and consistent.
Across the subject tenacity dimension, we observe four individually-variable systems that demonstrate the
largest variability in patterns. These systems include the frontoparietal and cingulo-opercular, the cognitive
control systems that have previously been shown to have large individual variability [47], and ventral temporal
association and motor and somatosensory systems that show learning-dependent changes [48, 45, 49]. The
remaining five systems are classified as individually-stable with high subject tenacity scores. The default
mode, subcortical, and attention systems have previously been found preserved across individuals as well
as across species [50], whereas the auditory and visual systems support fundamental perceptual processing
[51, 52].
Across the region tenacity dimension, we observe three levels of tenacity. Congitively-consistent systems
include the attention, subcortical, and visual systems, and stimulation to regions within these systems give
rise to similar patterns of synchrony and de-synchrony. The auditory emerges as the sole cognitively-diverse
system. This reflects the starkly different patterns that arise after stimulation of the regions within this
system. For example, stimulation to the superior temporal region results in high synchrony (both coherent
and chimera) states, while stimulation to the transverse temporal region leads to a metastable state (see
Fig. S6). Thus, the different local connectivity patterns of regions within this system produce immense
diversity in the resulting synchrony patterns upon their stimulation. The remaining five cognitive systems
are classified as cognitively-flexible, indicating that stimulation of regions within these systems produced
variable patterns of synchrony.
Overall, the tenacity scores both confirm and extend our knowledge of brain structure-function relation-
ships. The variability in subject tenacity among the systems reflects known differences in system stability
between individuals, and confirms that variability in chimera patterns captures these coarse differences
among the systems. On the other hand, the spread of region tenacity scores captures the diversity in the
functional roles that the regions within a system serve across diverse cognitive tasks. Cognitively-consistent
systems can largely be involved in the core sensory processing and associative learning, whereas the variabil-
ity of patterns within cognitively-flexible systems may enable them to serve diverse cognitive roles, relying
on stimulation of each constituent as a means to synchronize and integrate with different cognitive systems
to support particular cognitive demand or task-relevant processing.
Discussion
Using a novel, chimera-based framework, we explored the dynamical states that emerge across large-scale
cognitive systems following the spread of a targeted regional stimulation. We identified three distinct dy-
namical states – coherent, chimera, and metastable – that arise as a function of the structural connectivity
of the stimulated regions. A core result across all analyses is the variety in frequency and distribution of the
observed dynamical states. Chimera states are the most pervasive state to emerge following regional stimu-
lation. This likely reflects the foundational role that partial synchrony serves in large-scale brain function to
enable the intricate balance between segregated and integrated neural processing. Furthermore, the diver-
sity in these patterns captured both subject-specific and region-specific variability in structural connectivity.
Based on its sensitivity to these different sources of variability, our novel chimera-based framework shows
immense promise to better understand individual differences, relating patterns to performance, as well as
system constraints that underlie how to drive the brain to different task-relevant states.
14
Prevalence of chimera states across cognitive systems enables segregation and
integration in brain dynamics
The brain is a complex dynamical system that must integrate information across spatially-distributed, seg-
regated regions that serve specialized functional roles [53]. Neuroscience research therefore attempts to
understand how the brain creates selective synchrony across subsets of task-relevant regions to enable rapid
and adaptive cognitive processing [54, 55]. Recently, network neuroscience approaches have identified sets
of brain regions that form cognitive systems during rest and task states [56]. By studying the interactions
of these cognitive systems, functional analyses have identified the importance of (a) integrated states where
the connections are stronger between cognitive systems[12] and (b) segregated states where connections
are weaker between cognitive systems and are likely to be stronger within. The relative level of func-
tional integration vs segregation of cognitive systems has important consequences for cognitive performance.
Highly segregated systems enable efficient computations in local, functionally-specialized brain regions, while
strongly integrated systems provide rapid consolidation of information across systems necessary for coordi-
nated, cohesive performance of complex tasks [12, 13, 10]. These critical brain states are captured in the
chimera framework as metastable (segregated) states and coherent (integrated) states, and perhaps the most
critically for brain function, the chimera state that describes partial synchrony across subsets of cognitive
systems.
All nine systems give rise to a chimera state following stimulation, suggesting that all cognitive systems
can drive the brain to partial synchrony in support of their functional roles in cognition. Our results augment
a burgeoning literature on what brain dynamics support rapid shifts between more segregated or integrated
brain states. Previous work has found that functionally segregated states tend to involve shorter, local
connections [57], while integration largely relies on the global influence of subcortical regions and cortical
hubs that have many diverse connections to other brain regions [58]. Collectively, our results demonstrate
that our novel chimera-framework can investigate critical cognitive states where a balance between integration
and segregation is required for adaptive cognition.
Chimera framework reveals subject-specific and region-specific variability in
brain connectivity through the analyses of emergent dynamical states
We found that a coherent state is likely to be produced by the stimulation of regions in the medial default
mode and subcortical systems. This reflects the propensity of these systems to contain regional hubs, and the
prevalent emergence of a coherent state reflects their functional roles to bridge spatially disperse regions and
facilitate global brain communication. Interestingly, these systems also have high subject tenacity, indicating
robust occurrence across the 30 individuals. This could reflect that the subcortical and medial default mode
systems provide a fundamental, constant pillar of brain organization, which when disrupted, could lead to
impairments in global brain function. Previous research has shown that network hubs are often found to be
impacted in neurological disorders such as schizophrenia [59] and Alzheimer's disease [60]. These disorders
are associated with network-wide deficits in brain function [26], which is consistent with our finding that
cognitive systems that produce coherent states also contain network hubs.
Conversely, metastable states are preferentially produced by four systems with more sparse structural
connectivity: two systems associated with cognitive control (cingulo-opercular and frontoparietal systems)
and two systems associated with intricate sensory, object, and language representations (auditory and ventral
temporal association systems). These systems all have functional roles that frequently require working in
seclusion from other specialized processing in the brain. Interestingly, the three systems that are the most
unlikely be synchronized upon stimulation are also the ones that are most likely to produce a metastable
state: the auditory, cingulo-opercular, and frontoparietal systems. The cingulo-opercular and forntoparietal
systems are associated with cognitive control, and they are proposed as complementary systems that are
15
specialized for guiding successful task performance at different timescales: the cingulo-opercular system for
maintaining task-related goals across trials, and the frontoparietal systems for trial-by-trial control [61].
These functional roles may often need to occur concurrently, and their production of metastable states could
indicate that these systems can work in seclusion without co-activating a large part of the brain, an attribute
that facilitates parallel processing.
While our chimera framework revealed stable features of brain architecture, it also captured cognitive
systems where between subject variability leads to variety in frequency and type of synchronization patterns:
cingulo-opercular, frontoparietal, ventral temporal association, and motor systems. These four systems with
low subject tenacity are associated with higher cognitive functions where an individuals experience and
knowledge are likely captured by variability in their structural connectivity [62]. Our results demonstrated
that frontoparietal and cingulo-opercular systems exhibit the strongest individual variability, and this mirrors
recent results that showed cognitive control systems have weaker within-subject variability and greater
between-subject variability relative to sensory processing systems [47]. Our results also demonstrated that the
ventral temporal association and motor/somatosensory systems show an especially high number of prevalent
patterns with no single dominant pattern, and this may reflect their roles in learning and development-related
changes [49, 63].
Methodological Considerations and Future Directions
Our model is only sensitive to functional relationships that are induced through structural connections, so
the observed dynamical states and patterns are only constrained by the anatomical structure of the network.
In reality, neuronal activity patterns that are observed in brain using different functional measurement
techniques, such as fMRI, EEG, MEG, and PET, are a result of a complex neurophysiological activity that
develops on top of the structural connectivity infrastructure. Thus, the actual patterns of brain activity that
are observed across functional modalities may come from the simultaneous activation of different brain regions
via multiple input sources and therefore might differ from the patterns observed in our in silico experiments.
Here, the emergence of a coherent pattern would imply that a node can, in principle, communicate with
all of the spatially distributed regions within the brain; however, the actual nodes that it communicates
with may vary between different tasks according to the specific cognitive demands of the task. Likewise, the
emergence of a metastable state in our model may not reflect total de-synchrony, but just synchronization at
a smaller population level that our framework simulates, requiring future models to study dynamical states at
a finer spatial resolution. Despite these limitations, our approach is sensitive to variability in region-specific
and subject-specific brain connectivity, and it can be used to answer fundamental questions concerning the
cognitive organization of the human brain necessary for quantifying meaningful individual differences in brain
architecture, supporting individualized medicine, performance enhancement technologies, and personalized
stimulation protocols for treatment and/or augmentation.
Conclusion
By employing the cognitive system framework [37], our novel chimera-based approach keeps the computa-
tional complexity of the analysis to a minimum while confirming the existence of chimera states in the large-
scale cognitive systems in the human brain, as predicted from low-level, small scale models [21, 22, 32, 33].
The partial synchrony observed in a chimera state has a natural link to the well-documented role of functional
segregation and integration of cognitive systems thought to support cognition [54, 14], and the approach cap-
tures robust system differences for those that are largely stable across people as well as those that capture
individual training and expertise. Thus, our approach provides a rich opportunity to study how dynamical
states give rise to variability in cognitive performance, providing the first link between how chimera states
may subserve human behavior.
16
Methods
Human diffusion spectrum imaging data acquisition and pre-processing.
Diffusion MRI analysis was performed on the 30 individual participant scans previously reported elsewhere
[64]. Twenty male and ten female subjects were recruited from Pittsburgh and the Army Research Laboratory
in Aberdeen, Maryland. All subjects were neurologically healthy, with no history of either head trauma or
neurological or psychiatric illness. Subject ages ranged from 21 to 45 years of age at the time of scanning
(mean age of 31 years) and four were left-handed (2 male, 2 female). All participants signed an informed
consent approved by Carnegie Mellon University and conforming with the Declaration of Helsinki and were
financially remunerated for their participation.
Macroscopic white matter pathways were imaged using a Diffusion Spectrum Imaging (DSI) aquisition
sequence on a Siemens Verio 3T MRI system located at the Scientific Imaging & Brain Research Center
(SIBR) at Carnegie Mellon University using a 32-channel head coil. A total of 257-direction were sampled
using a twice-refocused spin-echo sequence (51 slices, TR = 9.916s, TE = 157ms, 2.4 x 2.4 x 2.4mm voxels, 231
x 231mm FoV, and b-max = 5,000s/mm2). Diffusion data were reconstructed using q-space diffeomorphic
reconstruction (QSDR;[65]) with a diffusion sampling length ratio of 1.25 and a 2mm output resolution.
Construction of individual structural brain networks.
Whole-brain structural connectivity matrices were constructed for each subject using a bootstrapping ap-
proach. To minimize the impact of bias in the tractography parameter scheme on streamline generation,
whole-brain fiber tractography [66] was performed 100 times for each participant, generating 250,000 stream-
lines per iteration. Across the 100 iterations, values were randomly sampled for QA-based fiber termination
thresholds (0.01-0.10), turning angle thresholds (40 -80 ), and smoothing (50%-80%), while constant values
were used for step size (1mm) and min/max fiber length thresholds (10mm/400mm). On each iteration a
binary connectivity matrix was generated, where an edge between two regions of interest was considered
present if 5% or more of streamlines generated were found to connect them. The probability of observing
a connection was estimated by calculating the frequency of detecting an edge across all 100 iterations. The
region of interest were determined using cortical components of the Desikan-Killiany atlas and subcorti-
cal components of the Harvard-Oxford Subcortical atlas.
In the resulting weighted matrices, connection
strengths were normalized by the sum of the regional brain volumes, and these normalized matrices were
used as the structural representations of individual brains.
All analysis was performed using DSI Studio (http://dsi-studio.labsolver.org/) and Matlab (MathWorks,
Inc.; Natick, MA, USA).
Data-driven network model of brain dynamics.
In our data-driven network model, regional brain dynamics are given by Wilson-Cowan oscillators [39, 36, 40].
In this biologically motivated neural mass model, the fraction of excitatory and inhibitory neurons active at
time t in the ith brain region are denoted by Ei(t) and Ii(t) respectively, and their temporal dynamics are
given by:
τ
dEi
dt
= −Ei(t) + (SEm − Ei(t))SE
c1Ei(t) − c2Ii(t) + c5
AijEj(t − τ ij
d ) + Pi(t)
+ σwi(t),
(1)
τ
dIi
dt
= −Ii(t) + (SIm − Ii(t))SI
c3Ei(t) − c4Ii(t) + c6
AijIj(t − τ ij
d )
+ σvi(t),
(2)
(cid:16)
(cid:16)
(cid:88)
(cid:88)
j
(cid:17)
(cid:17)
j
17
where
SE,I (x) =
1
1 + e(−aE,I (x−θE,I )
−
1
1 + eaE,I θE,I
.
(3)
Aij is an element of the subject-specific coupling matrix A whose value is the connection strength between
brain regions i and j as determined from diffusion spectrum imaging as described above. The global strength
of coupling between brain regions is tuned by excitatory and inhibitory coupling parameters c5 and c6 re-
spectively. In this case c6 = c5/4. Pi(t) represents the external stimulation to excitatory state activity and
was used to perform computational stimulation experiments. The parameter τ ij
d represents the communi-
cation delay between regions i and j. If the spatial distance between regions i and j is dij, τ ij
d = dij/td,
where td = 10m/s is the signal transmission velocity. We added noise as an input to the system through the
parameters wi(t) and vi(t) which are derived from a normal distribution with σ = 10−5. Other constants in
the model are biologically derived: c1 = 16, c2 = 12, c3 = 15, c4 = 3, aE = 1.3, aI = 2, θE = 4, θI = 3.7,
τ = 8 as described in references [39, 36, 40]. To numerically simulate the dynamics of the system, we used
a second order Runge Kutta method with step size 0.1 with initial conditions (Ei(0), Ii(0) = 0.1, 0.1).
Targeted stimulation.
The model was optimized for each individual to allow a regime of maximum dynamical sensitivity. This was
done by choosing a global coupling parameter, c5, such that the system was just below the critical transition
point to the excited state (see Fig. S1). Regional stimulation was achieved by applying a constant external
input Pi = 1.15 to a single region and perturbing its dynamics (Fig. S2). As the dynamics evolve, the
stimulation spreads within the brain through the network connectivity of the stimulated node.
Cognitive systems.
We assigned each brain region to one of 9 cognitive systems: attention, auditory, cingulo-opercular, fron-
toparietal, medial default mode, motor and somatosensory, subcortical, ventral temporal association and
visual. This node assignment is described in Table S1 and is similar to the one used by Muldoon et al. [36].
The distribution of brain regions within cognitive systems is shown in Fig. S3.
Calculation of synchronization within cognitively-informed framework.
We used the standard order parameter ρ to estimate the extent of synchronization after a targeted regional
stimulation within the brain networks. This measure was proposed by Kuramoto for the estimation of
coherence in a population of Kuramoto phase oscillators [42]. In this case, the instantaneous order parameter
at a given time t it is defined as
ρN (t)eiΦ(t) =
1
N
eiφj (t),
N(cid:88)
j=1
where φj is the phase of the jth oscillator at time t and is given by
φj(t) = tan−1 Ij(t)
Ej(t)
.
Here, N = 76 is the number of oscillators in the system. In order to estimate the global synchronization in
the system, one needs to average the instantaneous order parameter for a sufficiently long period of time,
ρN =< ρN (t) >t .
(6)
18
(4)
(5)
We used simulated activity over 1 s to estimate the average order parameter.
Within our cognitively-informed framework, we measured the synchronization between all pairs of cog-
nitive systems following a regional stimulation. This was done by calculating an order parameter for the
combined oscillator population of a pair of cognitive systems. For cognitive systems ci and cj, this order
parameter is given by
where
ρci,cj (t)eiΘ(t) =
1
Nci + Ncj
k∈(ci∪cj )
ρci,cj =< ρci,cj (t) >t,
(cid:88)
eiφk(t).
(7)
(8)
Here, Nci and Ncj represent the number of oscillators (brain regions or nodes) within cognitive systems ci
and cj respectively.
This analysis resulted in synchronization matrices, as shown in Fig. 2a, whose entries represent the extent
of synchronization between cognitive systems. These matrices were used in order to identify the dynamical
cognitive state that emerged as a result of regional stimulation.
Chimera-index.
We calculated the chimera-index (C) as described in [43, 34] as a measure of the normalized average variation
in order parameter within cognitive systems over time. For ci ∈ [c1, c2, ..., cM ] (M = 9 is the total number
of systems),
where
σch(t) =
1
M − 1
(ρci (t)− < ρc(t) >M )2.
< σch(t) >t
,
C =
CM ax
M(cid:88)
i=1
(9)
(10)
In this case, CM ax = 5/36 is a normalization factor and represents the maximum value of variability in the
order parameter in an ideal chimera state where the network organizes such that the half of its population is
completely synchronized and half is completely desynchronized [43]. The instantaneous quantity < ρc(t) >M
measures the synchronization of cognitive systems, averaged over all systems at a given time t.
Robust detection of the patterns of synchronization.
In order to robustly identify emergent cognitive patterns, we first obtained a binarized synchronization matrix
(s) such that, sij = 1 if systems i and j are identified synchronized, and sij = 0 otherwise. We defined
two cognitive systems i and j to be synchronized if ρci,cj ≥ ρT h, where ρT h represents a synchronization
threshold. For the results discussed in the main text we used ρT h = 0.8 [43] (as indicated in Fig. 2a).
In principle, one can directly use such binarized synchronization matrices in order to classify the emergent
states and patterns. However, we performed community detection on these binarized matrices. This method
clusters the group of synchronized systems into a single community whereas desynchronized systems remain
separate communities.
In case of a coherent state, we observed only one community, and in case of a
metastable state we observed nine separate communities, each representing a cognitive system. For chimera
states, communities with different distributions of cognitive systems emerged. Thus, applying the community
detection algorithm not only allowed us to robustly classify the emergent dynamical states, but also let us
separate various spatially distributed patterns of chimera states. Community detection was performed using
19
modularity maximization through the generalized Louvain algorithm [67]. For community detection, the
value of the resolution parameter was varied between 0.8 to 0.95 and a consensus was run to determine the
community structure [68].
Pattern tenacity.
A pattern in our analysis describes if the given cognitive system falls into the synchronized population or
remains desynchronized (Fig. 5). In order to calculate similarity between patterns, we defined tenacity for
a set of observed patterns as follows
p(cid:88)
(cid:0) 1
M(cid:88)
(cid:1),
δc
i,j
T =
1
p(p − 1)
i,j=1
M
c=1
(11)
where p is the number of patterns in the set. δc
i,j = 1 if the cognitive system c falls into the same state
of either synchrony or desynchrony in patterns i and j, and = 0 otherwise. We calculated the tenacity of
cognitive systems in two dimensions: across subjects for a given brain region within the system (subject
tenacity), and across regions of the system in a given subject (region tenacity). For subject tenacity of a
given cognitive system, p constitutes the patterns that are produced by a given node for all the subjects
and equals the number of subjects i.e. 30. For region tenacity of a given cognitive system, p constitutes
the patterns that all the nodes for a given cognitive system produce for a given subject, and the value of p
varies between systems. Thus for each cognitive system we obtained two distributions of tenacity, one for
each subject and region tenacity.
Clustering of cognitive systems using tenacity features.
In the subject-node tenacity parameter space, we grouped cognitive systems into clusters using the k-means
algorithm and silhouette analysis. We used k = 3, 4, 5, and 6 and identified the stable clustering that
maximizes similarity within clusters and dissimilarity across clusters. For k = 4 we observed an optimized
clustering. The corresponding silhouette plot is shown in Fig. S7.
Rendering of brain images.
BrainNet Viewer was used to perform spatial mapping onto brain images [69].
References
[1] Winfree, A. T. Biological rhythms and the behavior of populations of coupled oscillators. J. Theoret.
Biol 16, 15–42 (1967).
[2] Glass, L. Synchronization and rhythmic processes in physiology. Nature 410, 277–284 (2001).
[3] Jia, T., Wang, D. & Szymanski, B. K. Quantifying patterns of research-interest evolution. Nature
Human Behaviour 1, 0078 (2017).
[4] Strogatz, S. H., Abrams, D. M., McRobie, A., Eckhardt, B. & Ott, E. Crowd synchrony on the
millennium bridge. Nature 438, 43–44 (2005).
[5] Bertolero, M. A., Yeo, B. T. T. & D'Esposito, M. The modular and integrative functional architecture
of the human brain. Proceedings of the National Academy of Sciences 112, E6798–E6807 (2015).
20
[6] Motter, A. E. Nonlinear dynamics: Spontaneous synchrony breaking. Nature Physics 6, 164–165 (2010).
[7] Pecora, L. M., Sorrentino, F., Hagerstrom, A. M., Murphy, T. E. & Roy, R. Cluster synchronization
and isolated desynchronization in complex networks with symmetries. Nature Communications 5, 4079
(2014).
[8] Gray, C. M. Synchronous oscillations in neuronal systems: Mechanisms and functions. Journal of
Computational Neuroscience 1, 11–38 (1994).
[9] Lachaux, J. P., Rodriguez, E., Martinerie, J. & Varela, F. J. Measuring phase synchrony in brain
signals. Hum. Brain Mapp. 8, 194–208 (1999).
[10] Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis of structural and
functional systems. Nature Reviews Neuroscience 186–198 (2009).
[11] Bassett, D. S. & Sporns, O. Network neuroscience. Nature Neuroscience 20, 353–364 (2017).
[12] Shine, J. M. et al. The Dynamics of Functional Brain Networks: Integrated Network States during
Cognitive Task Performance. Neuron 92, 544–554 (2016).
[13] Vatansever, D., Menon, D. K. & Stamatakis, E. A. Default mode contributions to automated information
processing. Proceedings of the National Academy of Sciences of the United States of America 114,
12821–12826 (2017).
[14] Deco, G., Tononi, G., Boly, M. & Kringelbach, M. L. Rethinking segregation and integration: Contri-
butions of whole-brain modelling. Nature Reviews Neuroscience 16, 430–439 (2015).
[15] Kuramoto, Y. & Battogtokh, D. Coexistence of coherence and incoherence in nonlocally coupled phase
oscillators. Nonlinear Phenomena in Complex Systems 5, 380–385 (2002).
[16] Abrams, D. M. & Strogatz, S. H. Chimera states for coupled oscillators. Phys. Rev. Lett. 93, 174102
(2004).
[17] Omelchenko, O. E., Maistrenko, Y. L. & Tass, P. A. Chimera states: The natural link between coherence
and incoherence. Phys. Rev. Lett. 100, 044105 (2008).
[18] Laing, C. R. The dynamics of chimera states in heterogeneous kuramoto networks. Physica D: Nonlinear
Phenomena 238, 1569 – 1588 (2009).
[19] Hagerstrom, A. M. et al. Experimental observation of chimeras in coupled-map lattices. Nature Phys.
8, 658–661 (2012).
[20] Tinsley, M. R., Nkomo, S. & Showalter, K. Chimera and phase-cluster states in populations of coupled
chemical oscillators. Nature Physics 8, 662–665 (2012).
[21] Omelchenko, I., Omel'chenko, O. E., Hovel, P. & Scholl, E. When nonlocal coupling between oscillators
becomes stronger: Patched synchrony or multichimera states. Phys. Rev. Lett. 110, 224101 (2013).
[22] Vllings, A., Hizanidis, J., Omelchenko, I. & Hovel, P. Clustered chimera states in systems of type-i
excitability. New Journal of Physics 16, 123039 (2014).
[23] Xie, J., Knobloch, E. & Kao, H.-C. Multicluster and traveling chimera states in nonlocal phase-coupled
oscillators. Phys. Rev. E 90, 022919 (2014).
21
[24] Zakharova, A., Kapeller, M. & Scholl, E. Amplitude chimeras and chimera death in dynamical networks.
Journal of Physics: Conference Series 727, 012018 (2016).
[25] Killgore, W. D. S. Effects of sleep deprivation on cognition. Progress in Brain Research 185, 105129
(2010).
[26] Menon, V. Large-scale brain networks and psychopathology: A unifying triple network model. Trends
in Cognitive Sciences 15, 483–506 (2011).
[27] Fornito, A., Zalesky, A. & Breakspear, M. The connectomics of brain disorders. Nature Reviews
Neuroscience 16, 159–172 (2015).
[28] Panaggio, M. J. & Abrams, D. M. Chimera states: Coexistence of coherence and incoherence in networks
of coupled oscillators. Nonlinearity 28, R67 (2015).
[29] Andrzejak, R. G., Rummel, C., Mormann, F. & Schindler, K. All together now: Analogies between
chimera state collapses and epileptic seizures. Scientific reports 6, 23000 (2016).
[30] Tognoli, E. & Kelso, J. A. S. The metastable brain. Neuron 81, 35–48 (2014).
[31] Ashwin, P., Coombes, S. & Nicks, R. Mathematical frameworks for oscillatory network dynamics in
neuroscience. The Journal of Mathematical Neuroscience 6, 2 (2016).
[32] Hizanidis, J., Kanas, V. G., Bezerianos, A. & Bountis, T. Chimera states in networks of nonlocally
coupled hindmarshrose neuron models. International Journal of Bifurcation and Chaos 24, 1450030
(2014).
[33] Bera, B. K., Ghosh, D. & Lakshmanan, M. Chimera states in bursting neurons. Phys. Rev. E 93,
012205 (2016).
[34] Hizanidis, J., Kouvaris, N. E., Gorka, Z.-L., D´ıaz-Guilera, A. & Antonopoulos, C. G. Chimera-like
States in Modular Neural Networks. Scientific Reports 6, 19845 (2016). 1510.00286.
[35] Santos, M. et al. Chimera-like states in a neuronal network model of the cat brain. Chaos, Solitons and
Fractals 101, 86 – 91 (2017).
[36] Muldoon, S. F. et al. Stimulation-based control of dynamic brain networks. PLoS Comput Biol 12,
e1005076 (2016).
[37] Power, J. D. et al. Functional network organization of the human brain. Neuron 72, 665–678 (2011).
[38] Yeh, F.-C. et al. Quantifying Differences and Similarities in Whole-Brain White Matter Architecture
Using Local Connectome Fingerprints. PLoS Comput Biol 12 (2016).
[39] Wilson, H. R. & Cowan, J. D. Excitatory and inhibitory interactions in localized populations of model
neurons. Biophys J 12, 1–24 (1972).
[40] Bansal, K., Medaglia, J. D., Bassett, D. S., Vettel, J. M. & Muldoon, S. F. Data-driven brain network
models predict individual variability in behavior Preprint. Available from: arXiv:1802.08747v1.
Cited 27 April 2018. (2018).
[41] Bansal, K., Nakuci, J. & Muldoon, S. F. Personalized brain network models for assessing structure-
function relationships. Current Opinion in Neurobiology 52, 42 – 47 (2018).
22
[42] Kuramoto, Y. Self-entrainment of a population of coupled non-linear oscillators. In Araki, H. (ed.)
Lecture Notes in Physics, International Symposium on Mathematical Problems in Theoretical Physics,
420 (Springer-Verlag, New York, 1975).
[43] Shanahan, M. Metastable chimera states in community-structured oscillator networks. Chaos: An
Interdisciplinary Journal of Nonlinear Science 20, 013108 (2010).
[44] Dehaene, S., Cohen, L., Morais, J. & Kolinsky, R. Illiterate to literate: behavioural and cerebral changes
induced by reading acquisition. Nature Reviews Neuroscience 16, 234–244 (2015).
[45] Sheinberg, D. L. & Logothetis, N. K. Noticing familiar objects in real world scenes: the role of temporal
cortical neurons in natural vision. J Neurosci 21, 134050 (2001).
[46] Grill-Spector, K., Kourtzi, Z. & Kanwisher, N. The lateral occipital complex and its role in object
recognition. Vision Res 41, 140922 (2001).
[47] Gratton, C. et al. Functional Brain Networks Are Dominated by Stable Group and Individual Factors,
Not Cognitive or Daily Variation. Neuron 98, 439–452.e5 (2018).
[48] Muraskin, J. et al. Fusing multiple neuroimaging modalities to assess group differences in perception
action coupling. Proceedings of the IEEE 105, 83100 (2017).
[49] Scholz, J., Klein, M. C., Behrens, T. E. J. & Johansen-Berg, H. Training induces changes in white-matter
architecture. Nature Neuroscience 12, 13701371 (2009).
[50] Margulies, D. S. et al. Precuneus shares intrinsic functional architecture in humans and monkeys.
Proceedings of the National Academy of Sciences 106, 20069–20074 (2009).
[51] Kaskan, P. M. et al. Peripheral variability and central constancy in mammalian visual system evolution.
Proceedings. Biological sciences 272, 91–100 (2005).
[52] Stevens, C. F. An evolutionary scaling law for the primate visual system and its basis in cortical
function. Nature 411, 193–195 (2001).
[53] Alivisatos, A. P. et al. The brain activity map project and the challenge of functional connectomics.
Neuron 74, 970–4 (2012).
[54] Menon, V. Large-scale brain networks in cognition: emerging methods and principles. Trends Cogn Sci
14, 277–290 (2010).
[55] Gollo, L. L., Roberts, J. A. & Cocchi, L. Mapping how local perturbations influence systems-level brain
dynamics. NeuroImage 160, 97–112 (2017).
[56] Smith, S. M. et al. Correspondence of the brains functional architecture during activation and rest.
Proceedings of the National Academy of Sciences 106, 1304013045 (2009).
[57] Li´egeois, R. et al. Cerebral functional connectivity periodically (de)synchronizes with anatomical con-
straints. Brain Structure and Function 221, 2985–2997 (2016).
[58] Shine, J. M., Aburn, M. J., Breakspear, M. & Poldrack, R. A. The modulation of neural gain facilitates
a transition between functional segregation and integration in the brain. eLife 7 (2018).
[59] Klauser, P. et al. White matter disruptions in schizophrenia are spatially widespread and topologically
converge on brain network hubs. Schizophrenia Bulletin 43, 425–435 (2017).
23
[60] Buckner, R. L. et al. Cortical hubs revealed by intrinsic functional connectivity: Mapping, assessment
of stability, and relation to alzheimer's disease. Journal of Neuroscience 29, 1860–1873 (2009).
[61] Cocchi, L., Zalesky, A., Fornito, A. & Mattingley, J. B. Dynamic cooperation and competition between
brain systems during cognitive control. Trends in Cognitive Sciences 17, 493–501 (2013).
[62] Powell, M. A., Garcia, J. O., Yeh, F.-C., Vettel, J. M. & Verstynen, T. Local connectome phenotypes
predict social, health, and cognitive factors. Network Neuroscience 2, 86–105 (2018).
[63] Kahn, A. E. et al. Structural pathways supporting swift acquisition of new visuomotor skills. Cerebral
Cortex 27, 173–184 (2017).
[64] Verstynen, T. D. The organization and dynamics of corticostriatal pathways link the medial orbitofrontal
cortex to future behavioral responses. Journal of neurophysiology 112, 2457–2469 (2014).
[65] Yeh, F.-C. & Tseng, W.-Y. I. Ntu-90: a high angular resolution brain atlas constructed by q-space
diffeomorphic reconstruction. Neuroimage 58, 91–99 (2011).
[66] Yeh, F.-C., Verstynen, T. D., Wang, Y., Fern´andez-Miranda, J. C. & Tseng, W.-Y. I. Deterministic
diffusion fiber tracking improved by quantitative anisotropy. PloS one 8, e80713 (2013).
[67] Mucha, P. J., Richardson, T., Macon, K., Porter, M. A. & Onnela, J.-P. Community structure in
time-dependent, multiscale, and multiplex networks. Science 328, 876–878 (2010).
[68] Bassett, D. S. et al. Robust detection of dynamic community structure in networks. Chaos: An
Interdisciplinary Journal of Nonlinear Science 23, 013142 (2013).
[69] Xia, M., Wang, J. & He, Y. BrainNet Viewer: A Network Visualization Tool for Human Brain Con-
nectomics. PLoS ONE 8, e68910 (2013).
Acknowledgments
This work is supported by the Army Research Laboratory through contract # W911NF-10-2-0022 and
W911NF-16-2-0158 from the U.S. Army research office. The authors acknowledge Gregory Lieberman for
the pre-processing of the brain anatomical connectivity data, and Rajarshi Roy and Joseph D. Hart for a
useful discussion on chimera states. The content is solely the responsibility of the authors and does not
necessarily represent the official views of the funding agency.
Author contributions
KB conceived the idea; KB, JMV, and SFM designed the research; TV and JMV collected the data; SFM
developed the model; KB implemented the model and led the analysis; all authors contributed analysis ideas
and wrote the paper.
24
Figure S1: Optimizing personalized brain network models. In our model, coupling between different brain
regions can be tuned by the global coupling parameter,c5. We optimized this parameter for each individual
to place the brain in a dynamical state that is maximally sensitive to a perturbation (stimulation). To
find this optimal value for each individual, we varied c5 between 100 and 1500 in steps of 10 and measured
average excitatory activity (Ei) of each brain region (i) for a time period of 1 s. Each simulation was started
with (Ei , Ii) = (0.1, 0.1), and before starting the measurement, we let the system evolve for 1 s in order
to eliminate initial transients. (a) As we varied the value of c5, we observed a sudden transition in systems
behavior to an excited state. (b) The value of c5 at which the transition is observed typically varies between
individuals. The transition value signifies the ease with which a given brain can be excited. In order to apply
a computational stimulation within the personalized model, we fixed the value of c5 just below its transition
value.
25
Figure S2: Targeted regional stimulation. (a) To stimulate a particular brain region within the brain network
model of a given individual, we applied a constant external input, Pi = 1.15. In our model, as we switch
on the external input, the stimulated brain region changes its dynamics from a (b) stable fixed point to (c)
a limit cycle (oscillatory motion). The effect of this stimulation on the other regions within the brain is
measured through the resulting patterns of synchronization as discussed in the main text.
26
Figure S3: Distribution of brain volume within cognitive systems. Spatial mappings showing the distribution
of the nine cognitive systems within the brain. The identification of regions is given in Table S1. L and R
denote the left and right hemisphere respectively.
27
Figure S4: Effect of changing synchronization threshold on the distribution of states. (a)-(d) Separation
of three different states in the parameter space of global synchrony and chimera-index for different synchro-
nization threshold values (ρT h). Using different values of ρT h in our analysis, we obtain results that are
qualitatively similar to the results discussed in Fig. 2b. We observe three different states i.e. coherent,
chimera and metastable states, for all four threshold values. These states can be clearly separated in the
parameter space of global synchrony and chimera-index. However, a higher threshold value (e.g. 0.9), which
signifies strictness in identifying a population to be synchronized, allows a lower number of states to be clas-
sified as a coherent state, as opposed to a lower threshold value (e.g. 0.7), which is more relaxed. Conversely,
number of observed metastable states increases with increasing threshold values.
28
Figure S5: Likelihood of the emergence of dynamical states across cognitive systems. Contribution of
different cognitive systems towards producing a given cognitive states is depicted for (a) ρT h = 0.75 and (b)
ρT h = 0.85. These figures corroborate our findings presented in Fig. 4. We observe that a coherent state
dominantly originates when the regions in subcortical and medial default mode systems are stimulated. A
metastable state is likely to originate upon stimulation of the regions within cingulo-opercular, frontoparietal,
auditory and ventral temporal association systems. Chimera states do not show the dominance of any
particular cognitive system, regions from different cognitive systems are relatively equally likely to produce
of chimera state upon stimulation.
29
30
31
Figure S6: Normalized contribution of brain regions to the prevalent patterns of synchronization. In Fig.
5 of the main text we show the prevalent patterns of synchronization that we observed after systematic
regional stimulation within the brain across 30 individuals. Here, we show the brain regions (nodes) that
contributed in producing these patterns within each cognitive system. Each sub-figure represents a cognitive
system and the first image depicts all the brain regions within the system.
In the following images, we
show only the brain regions (nodes) that produced the corresponding patterns in Fig. 5. The size of the
node represents the normalized occurrence for the given pattern. In panel (b), in order to help the reader
differentiate superior and transverse regions, we mark them with S and T respectively. From this figure it
can be clearly observed that different nodes within the same cognitive system can produce different patterns.
Normalized occurrences for different patterns also vary across patterns and across cognitive systems.
32
Figure S7: Clustering of cognitive systems using patterns tenacity. As described in the main text, in the
parameter space of subject and node tenacity we grouped cognitive systems into clusters using the k-means
algorithm and silhouette analysis. For k = 4 we observed an optimized clustering (Fig. 6). Here we show
the corresponding silhouette plot using squared Euclidean distance of the data points from the centroid of
the cluster to which they were assigned.
33
S.N. Name of the brain region Cognitive assignment
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Lateral orbitofrontal
Superior parietal
Superior temporal
Transverse temporal
Caudal anterior cingulate
Pars opercularis
Pars orbitalis
Rostral anterior cingulate
Rostral middle frontal
Supramarginal
Caudal middle frontal
Inferior parietal
Medial orbitofrontal
Pars triangularis
Insula
Isthmus cingulate
Posterior cingulate
Precuneus
Superior frontal
Att
Att
Aud
Aud
CP
CP
CP
CP
CP
CP
FP
FP
FP
FP
FP
mDM
mDM
mDM
mDM
Paracentral
Postcentral
Precentral
Thalamus
Caudate
Putamen
Pallidum
S.N. Name of the brain region Cognitive assignment
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
MS
MS
MS
SC
SC
SC
SC
SC
SC
SC
VT
VT
VT
VT
VT
V
V
V
V
Inferior temporal
Middle temporal
Parahippocampal
Amygdala
Accumbens
Entorhinal
Fusiform
Hippocampus
Cuneus
Lateral occipital
Lingual
Pericalcarine
Table S1: Assignment of brain regions to cognitive systems. Each node in the brain network is assigned to a
predefined cognitive system. System assignments are the same for regions in both left and right hemisphere.
In Figure S3, we show the distribution of regional brain volume for each cognitive system.
34
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.